Upload 13 files

src/extract_features.py

@@ -0,0 +1,54 @@
+import os
+import sys
+import pandas as pd
+import json
+from tqdm import tqdm 
+from PIL import Image
+import torch
+from multiprocessing import Pool
+import h5py
+from transformers import logging
+from transformers import CLIPFeatureExtractor, CLIPVisionModel
+logging.set_verbosity_error()
+
+data_dir = 'data/images/'
+features_dir = 'features/'
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+encoder_name = 'openai/clip-vit-base-patch32'
+feature_extractor = CLIPFeatureExtractor.from_pretrained(encoder_name) 
+clip_encoder = CLIPVisionModel.from_pretrained(encoder_name).to(device)
+
+annotations = json.load(open('data/dataset_coco.json'))['images']
+
+def load_data():
+    data = {'train': [], 'val': []}
+
+    for item in annotations:
+        file_name = item['filename'].split('_')[-1]
+        if item['split'] == 'train' or item['split'] == 'restval':
+            data['train'].append({'file_name': file_name, 'cocoid': item['cocoid']})
+        elif item['split'] == 'val':
+            data['val'].append({'file_name': file_name, 'cocoid': item['cocoid']})
+    return data
+
+def encode_split(data, split):
+    df = pd.DataFrame(data[split])
+
+    bs = 256
+    h5py_file = h5py.File(features_dir + '{}.hdf5'.format(split), 'w')
+    for idx in tqdm(range(0, len(df), bs)):
+        cocoids = df['cocoid'][idx:idx + bs]
+        file_names = df['file_name'][idx:idx + bs]
+        images = [Image.open(data_dir + file_name).convert("RGB") for file_name in file_names]
+        with torch.no_grad(): 
+            pixel_values = feature_extractor(images, return_tensors='pt').pixel_values.to(device)
+            encodings = clip_encoder(pixel_values=pixel_values).last_hidden_state.cpu().numpy() 
+        for cocoid, encoding in zip(cocoids, encodings):
+            h5py_file.create_dataset(str(cocoid), (50, 768), data=encoding)
+
+data = load_data()
+
+encode_split(data, 'train')
+encode_split(data, 'val')

src/gpt2.py

@@ -0,0 +1,167 @@
+# coding=utf-8
+# Copyright 2018 The OpenAI Team Authors and HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PyTorch OpenAI GPT-2 model."""
+
+import math
+import os
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from packaging import version
+from torch import nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+
+from transformers.models.gpt2.modeling_gpt2 import load_tf_weights_in_gpt2, GPT2LMHeadModel, GPT2MLP, GPT2Attention, GPT2Block, GPT2Model 
+
+from transformers.activations import ACT2FN
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPastAndCrossAttentions,
+    CausalLMOutputWithCrossAttentions,
+    SequenceClassifierOutputWithPast,
+    TokenClassifierOutput,
+)
+from transformers.modeling_utils import PreTrainedModel, SequenceSummary
+from transformers.pytorch_utils import Conv1D, find_pruneable_heads_and_indices, prune_conv1d_layer
+from transformers.utils import (
+    ModelOutput,
+    logging,
+)
+from transformers.utils.model_parallel_utils import assert_device_map, get_device_map
+from transformers.models.gpt2.configuration_gpt2 import GPT2Config
+
+
+if version.parse(torch.__version__) >= version.parse("1.6"):
+    is_amp_available = True
+    from torch.cuda.amp import autocast
+else:
+    is_amp_available = False
+
+    
+class ThisGPT2Config(GPT2Config):
+    model_type = "this_gpt2"
+
+    def __init__(
+        self,
+        cross_attention_reduce_factor = 1,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.cross_attention_reduce_factor = cross_attention_reduce_factor
+        
+class ThisGPT2Attention(GPT2Attention):
+    def __init__(self, config, is_cross_attention=False, layer_idx=None):
+        super().__init__(config, is_cross_attention, layer_idx)
+
+        #print("this gpt2")
+
+        #print("self.is_cross_attention = is_cross_attention", self.is_cross_attention, is_cross_attention)
+        
+        self.cross_attention_reduce_factor = config.cross_attention_reduce_factor
+        
+        if self.is_cross_attention:
+            self.c_attn = Conv1D(int(2 / self.cross_attention_reduce_factor * self.embed_dim), 
+                                                                                  self.embed_dim) 
+            self.q_attn = Conv1D(int(self.embed_dim / self.cross_attention_reduce_factor), self.embed_dim)
+            self.c_proj = Conv1D(self.embed_dim, int(self.embed_dim / self.cross_attention_reduce_factor))
+        else:
+            self.c_attn = Conv1D(3 * self.embed_dim, self.embed_dim)
+            self.c_proj = Conv1D(self.embed_dim, self.embed_dim)
+
+    def forward(
+        self,
+        hidden_states: Optional[Tuple[torch.FloatTensor]],
+        layer_past: Optional[Tuple[torch.Tensor]] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = False,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[Union[torch.Tensor, Tuple[torch.Tensor]], ...]:
+        if encoder_hidden_states is not None:
+            if not hasattr(self, "q_attn"):
+                raise ValueError(
+                    "If class is used as cross attention, the weights `q_attn` have to be defined. "
+                    "Please make sure to instantiate class with `GPT2Attention(..., is_cross_attention=True)`."
+                )
+            split_size = int(self.split_size / self.cross_attention_reduce_factor)
+            head_dim = int(self.head_dim / self.cross_attention_reduce_factor)
+
+            query = self.q_attn(hidden_states)
+            key, value = self.c_attn(encoder_hidden_states).split(split_size, dim=2)
+            attention_mask = encoder_attention_mask
+
+            query = self._split_heads(query, self.num_heads, head_dim)
+            key = self._split_heads(key, self.num_heads, head_dim)
+            value = self._split_heads(value, self.num_heads, head_dim)
+        else:
+            query, key, value = self.c_attn(hidden_states).split(self.split_size, dim=2)
+        
+            query = self._split_heads(query, self.num_heads, self.head_dim)
+            key = self._split_heads(key, self.num_heads, self.head_dim)
+            value = self._split_heads(value, self.num_heads, self.head_dim)
+
+        if layer_past is not None:
+            past_key, past_value = layer_past
+            key = torch.cat((past_key, key), dim=-2)
+            value = torch.cat((past_value, value), dim=-2)
+
+        if use_cache is True:
+            present = (key, value)
+        else:
+            present = None
+
+        if self.reorder_and_upcast_attn:
+            attn_output, attn_weights = self._upcast_and_reordered_attn(query, key, value, attention_mask, head_mask)
+        else:
+            attn_output, attn_weights = self._attn(query, key, value, attention_mask, head_mask)
+        
+        attn_output = self._merge_heads(attn_output, self.num_heads, int(self.head_dim / self.cross_attention_reduce_factor))
+        attn_output = self.c_proj(attn_output)
+        attn_output = self.resid_dropout(attn_output)
+
+        outputs = (attn_output, present)
+        if output_attentions:
+            outputs += (attn_weights,)
+
+        return outputs  # a, present, (attentions)
+
+
+class ThisGPT2Block(GPT2Block):
+    def __init__(self, config, layer_idx=None):
+        super().__init__(config, layer_idx)
+        hidden_size = config.hidden_size
+
+        if config.add_cross_attention:
+            self.crossattention = ThisGPT2Attention(config, is_cross_attention=True, layer_idx=layer_idx)
+            self.ln_cross_attn = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
+
+class ThisGPT2Model(GPT2Model):
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.h = nn.ModuleList([ThisGPT2Block(config, layer_idx=i) for i in range(config.num_hidden_layers)])
+
+
+class ThisGPT2LMHeadModel(GPT2LMHeadModel):
+    config_class = ThisGPT2Config
+    
+    def __init__(self, config):
+        super().__init__(config)
+        self.transformer = ThisGPT2Model(config)
+

src/opt.py

@@ -0,0 +1,696 @@
+# coding=utf-8
+# Copyright 2018 The OpenAI Team Authors and HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PyTorch OpenAI GPT-2 model."""
+
+import math
+import os
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+import random
+
+import torch
+import torch.utils.checkpoint
+from packaging import version
+from torch import nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+
+from transformers.models.gpt2.modeling_gpt2 import load_tf_weights_in_gpt2, GPT2LMHeadModel, GPT2MLP, GPT2Attention, GPT2Block, GPT2Model 
+
+from transformers.models.opt.modeling_opt import OPTForCausalLM, OPTAttention, OPTDecoderLayer, OPTModel, OPTDecoder 
+
+from transformers.activations import ACT2FN
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPastAndCrossAttentions,
+    CausalLMOutputWithCrossAttentions,
+    SequenceClassifierOutputWithPast,
+    TokenClassifierOutput,
+)
+
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPast,
+    CausalLMOutputWithPast,
+    QuestionAnsweringModelOutput,
+    SequenceClassifierOutputWithPast,
+)
+
+from transformers.modeling_utils import PreTrainedModel, SequenceSummary
+from transformers.pytorch_utils import Conv1D, find_pruneable_heads_and_indices, prune_conv1d_layer
+from transformers.utils import (
+    ModelOutput,
+    logging,
+)
+from transformers.utils.model_parallel_utils import assert_device_map, get_device_map
+from transformers.models.opt.configuration_opt import OPTConfig
+
+
+if version.parse(torch.__version__) >= version.parse("1.6"):
+    is_amp_available = True
+    from torch.cuda.amp import autocast
+else:
+    is_amp_available = False
+
+    
+class ThisOPTConfig(OPTConfig):
+    model_type = "this_opt"
+
+    def __init__(
+        self,
+        cross_attention_reduce_factor = 1,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.cross_attention_reduce_factor = cross_attention_reduce_factor
+        
+
+class ThisOPTAttention(OPTAttention):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(
+        self,
+        embed_dim,
+        num_heads,
+        dropout = 0.0,
+        is_decoder = False,
+        bias = True,
+        config=None,
+        is_cross_attention=False,
+    ):
+        super().__init__(embed_dim,num_heads, dropout,is_decoder,bias)
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.head_dim = embed_dim // num_heads
+
+        if (self.head_dim * num_heads) != self.embed_dim:
+            raise ValueError(
+                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim}"
+                f" and `num_heads`: {num_heads})."
+            )
+        self.scaling = self.head_dim**-0.5
+        self.is_decoder = is_decoder
+
+        self.cross_attention_reduce_factor = config.cross_attention_reduce_factor
+        self.head_dim = int(self.head_dim / self.cross_attention_reduce_factor)
+
+
+        if is_cross_attention:
+            #print("self", int(embed_dim / self.cross_attention_reduce_factor))
+            self.k_proj = nn.Linear(768, int(embed_dim / self.cross_attention_reduce_factor), bias=bias)
+            #print("self.k_proj",self.k_proj)
+            self.v_proj = nn.Linear(768, int(embed_dim / self.cross_attention_reduce_factor), bias=bias)
+            self.q_proj = nn.Linear(embed_dim, int(embed_dim / self.cross_attention_reduce_factor), bias=bias)
+            self.out_proj = nn.Linear(int(embed_dim / self.cross_attention_reduce_factor),embed_dim, bias=bias)
+
+            self.embed_dim=int(embed_dim / self.cross_attention_reduce_factor)
+        else:
+            self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+            self.v_proj = nn.Linear(embed_dim, embed_dim , bias=bias)
+            self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+            self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+
+
+    def forward(
+        self,
+        hidden_states,
+        key_value_states = None,
+        past_key_value = None,
+        attention_mask = None,
+        layer_head_mask = None,
+        output_attentions= False,
+    ):
+        """Input shape: Batch x Time x Channel"""
+
+        # if key_value_states are provided this layer is used as a cross-attention layer
+        # for the decoder
+        is_cross_attention = key_value_states is not None
+
+        bsz, tgt_len, _ = hidden_states.size()
+
+        # get query proj
+        query_states = self.q_proj(hidden_states) * self.scaling
+        # get key, value proj
+        if is_cross_attention and past_key_value is not None:
+            # reuse k,v, cross_attentions
+            key_states = past_key_value[0]
+            value_states = past_key_value[1]
+        elif is_cross_attention:
+            # cross_attentions
+            key_states = self._shape(self.k_proj(key_value_states), -1, bsz)
+            value_states = self._shape(self.v_proj(key_value_states), -1, bsz)
+        elif past_key_value is not None:
+            # reuse k, v, self_attention
+            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+            key_states = torch.cat([past_key_value[0], key_states], dim=2)
+            value_states = torch.cat([past_key_value[1], value_states], dim=2)
+        else:
+            # self_attention
+            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+
+        if self.is_decoder:
+            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+            past_key_value = (key_states, value_states)
+
+        proj_shape = (bsz * self.num_heads, -1, self.head_dim)
+        query_states = self._shape(query_states, tgt_len, bsz).view(*proj_shape)
+        key_states = key_states.view(*proj_shape)
+        value_states = value_states.view(*proj_shape)
+
+        src_len = key_states.size(1)
+        attn_weights = torch.bmm(query_states, key_states.transpose(1, 2))
+
+        if attn_weights.size() != (bsz * self.num_heads, tgt_len, src_len):
+            raise ValueError(
+                f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is"
+                f" {attn_weights.size()}"
+            )
+
+        if attention_mask is not None:
+            if attention_mask.size() != (bsz, 1, tgt_len, src_len):
+                raise ValueError(
+                    f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}"
+                )
+            attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attention_mask
+            attn_weights = torch.max(attn_weights, torch.tensor(torch.finfo(attn_weights.dtype).min))
+            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
+
+        # upcast to fp32 if the weights are in fp16. Please see https://github.com/huggingface/transformers/pull/17437
+        if attn_weights.dtype == torch.float16:
+            attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(torch.float16)
+        else:
+            attn_weights = nn.functional.softmax(attn_weights, dim=-1)
+
+        if layer_head_mask is not None:
+            if layer_head_mask.size() != (self.num_heads,):
+                raise ValueError(
+                    f"Head mask for a single layer should be of size {(self.num_heads,)}, but is"
+                    f" {layer_head_mask.size()}"
+                )
+            attn_weights = layer_head_mask.view(1, -1, 1, 1) * attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
+            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
+
+        if output_attentions:
+            # this operation is a bit awkward, but it's required to
+            # make sure that attn_weights keeps its gradient.
+            # In order to do so, attn_weights have to be reshaped
+            # twice and have to be reused in the following
+            attn_weights_reshaped = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
+            attn_weights = attn_weights_reshaped.view(bsz * self.num_heads, tgt_len, src_len)
+        else:
+            attn_weights_reshaped = None
+
+        attn_probs = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
+
+        attn_output = torch.bmm(attn_probs, value_states)
+
+        if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
+                f" {attn_output.size()}"
+            )
+
+        attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
+        attn_output = attn_output.transpose(1, 2)
+
+        # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
+        # partitioned aross GPUs when using tensor-parallelism.
+        attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
+
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output, attn_weights_reshaped, past_key_value
+
+
+class ThisOPTDecoderLayer(OPTDecoderLayer):
+    def __init__(self, config):
+        super().__init__(config)
+
+        if config.add_cross_attention:
+            self.encoder_attn = ThisOPTAttention(
+                embed_dim=self.embed_dim,
+                num_heads=config.num_attention_heads,
+                dropout=config.attention_dropout,
+                is_decoder=True,
+                #bias=config.enable_bias,
+                config=config,
+                is_cross_attention=True
+            )
+            self.encoder_attn_layer_norm = nn.LayerNorm(self.embed_dim,elementwise_affine=config.layer_norm_elementwise_affine)
+
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask= None,
+        encoder_hidden_states = None,
+        encoder_attention_mask = None,
+        layer_head_mask = None,
+        cross_attn_head_mask = None,
+        output_attentions = False,
+        use_cache = False,
+        past_key_value = None,
+    ):
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            layer_head_mask (`torch.FloatTensor`, *optional*): mask for attention heads in a given layer of size
+                `(encoder_attention_heads,)`.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            use_cache (`bool`, *optional*):
+                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
+                (see `past_key_values`).
+            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
+        """
+
+        residual = hidden_states
+
+        # 125m, 1.7B, ..., 175B applies layer norm BEFORE attention
+        if self.do_layer_norm_before:
+            hidden_states = self.self_attn_layer_norm(hidden_states)
+
+        # Self Attention
+        hidden_states, self_attn_weights, present_key_value = self.self_attn(
+            hidden_states=hidden_states,
+            past_key_value=past_key_value,
+            attention_mask=attention_mask,
+            layer_head_mask=layer_head_mask,
+            output_attentions=output_attentions,
+        )
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+
+
+        # Cross-Attention Block
+        cross_attn_present_key_value = None
+        cross_attn_weights = None
+        if encoder_hidden_states is not None:
+            residual = hidden_states
+            hidden_states = self.encoder_attn_layer_norm(hidden_states)
+
+            # cross_attn cached key/values tuple is at positions 3,4 of present_key_value tuple
+            cross_attn_past_key_value = past_key_value[-2:] if past_key_value is not None else None
+            hidden_states, cross_attn_weights, cross_attn_present_key_value = self.encoder_attn(
+                hidden_states=hidden_states,
+                key_value_states=encoder_hidden_states,
+                attention_mask=encoder_attention_mask,
+                layer_head_mask=cross_attn_head_mask,
+                past_key_value=cross_attn_past_key_value,
+                output_attentions=output_attentions,
+            )
+            hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+            hidden_states = residual + hidden_states
+
+            # add cross-attn to positions 3,4 of present_key_value tuple
+            present_key_value = present_key_value + cross_attn_present_key_value
+
+
+
+        # 350m applies layer norm AFTER attention
+        if not self.do_layer_norm_before:
+            hidden_states = self.self_attn_layer_norm(hidden_states)
+
+        # Fully Connected
+        hidden_states_shape = hidden_states.shape
+        hidden_states = hidden_states.reshape(-1, hidden_states.size(-1))
+        residual = hidden_states
+
+        # 125m, 1.7B, ..., 175B applies layer norm BEFORE attention
+        if self.do_layer_norm_before:
+            hidden_states = self.final_layer_norm(hidden_states)
+
+        hidden_states = self.fc1(hidden_states)
+        hidden_states = self.activation_fn(hidden_states)
+
+        hidden_states = self.fc2(hidden_states)
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+
+        hidden_states = (residual + hidden_states).view(hidden_states_shape)
+
+        # 350m applies layer norm AFTER attention
+        if not self.do_layer_norm_before:
+            hidden_states = self.final_layer_norm(hidden_states)
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        return outputs
+
+class ThisOPTDecoder(OPTDecoder):
+    def __init__(self, config):
+        super().__init__(config)
+        self.layers = nn.ModuleList([ThisOPTDecoderLayer(config) for _ in range(config.num_hidden_layers)])
+
+
+    def forward(
+        self,
+        input_ids = None,
+        attention_mask = None,
+        encoder_hidden_states=None,
+        encoder_attention_mask = None,
+        head_mask = None,
+        cross_attn_head_mask = None,
+        past_key_values = None,
+        inputs_embeds = None,
+        use_cache = None,
+        output_attentions = None,
+        output_hidden_states = None,
+        return_dict = None,
+    ):
+        r"""
+        Args:
+            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
+                provide it.
+                Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+                [`PreTrainedTokenizer.__call__`] for details.
+                [What are input IDs?](../glossary#input-ids)
+            attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+                - 1 for tokens that are **not masked**,
+                - 0 for tokens that are **masked**.
+                [What are attention masks?](../glossary#attention-mask)
+            head_mask (`torch.Tensor` of shape `(num_hidden_layers, num_attention_heads)`, *optional*):
+                Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:
+                - 1 indicates the head is **not masked**,
+                - 0 indicates the head is **masked**.
+            past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+                Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
+                shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of
+                Contains pre-computed hidden-states (key and values in the self-attention blocks and in the
+                cross-attention blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
+                If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
+                that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
+                all `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+                than the model's internal embedding lookup matrix.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            output_hidden_states (`bool`, *optional*):
+                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
+                for more detail.
+            return_dict (`bool`, *optional*):
+                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+            input_ids = input_ids.view(-1, input_shape[-1])
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
+
+        past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        # embed positions
+        if attention_mask is None:
+            attention_mask = torch.ones(inputs_embeds.shape[:2], dtype=torch.bool, device=inputs_embeds.device)
+        pos_embeds = self.embed_positions(attention_mask, past_key_values_length)
+
+        attention_mask = self._prepare_decoder_attention_mask(
+            attention_mask, input_shape, inputs_embeds, past_key_values_length
+        )
+
+        if self.project_in is not None:
+            inputs_embeds = self.project_in(inputs_embeds)
+
+        hidden_states = inputs_embeds + pos_embeds
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        next_decoder_cache = () if use_cache else None
+        all_cross_attentions = () if (output_attentions and encoder_hidden_states is not None) else None
+
+
+        # check if head_mask has a correct number of layers specified if desired
+        for attn_mask, mask_name in zip([head_mask], ["head_mask"]):
+            if attn_mask is not None:
+                if attn_mask.size()[0] != (len(self.layers)):
+                    raise ValueError(
+                        f"The `{mask_name}` should be specified for {len(self.layers)} layers, but it is for"
+                        f" {head_mask.size()[0]}."
+                    )
+
+        for idx, decoder_layer in enumerate(self.layers):
+            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            dropout_probability = random.uniform(0, 1)
+            if self.training and (dropout_probability < self.layerdrop):
+                continue
+
+            past_key_value = past_key_values[idx] if past_key_values is not None else None
+
+            if self.gradient_checkpointing and self.training:
+
+                if use_cache:
+                    logger.warning(
+                        "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                    )
+                    use_cache = False
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        # None for past_key_value
+                        return module(*inputs, output_attentions, None)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(decoder_layer),
+                    hidden_states,
+                    attention_mask,
+                    head_mask[idx] if head_mask is not None else None,
+                    None,
+                )
+            else:
+
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    encoder_attention_mask=encoder_attention_mask,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attn_head_mask=cross_attn_head_mask[idx] if cross_attn_head_mask is not None else None,
+                    attention_mask=attention_mask,
+                    layer_head_mask=(head_mask[idx] if head_mask is not None else None),
+                    past_key_value=past_key_value,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                )
+
+            hidden_states = layer_outputs[0]
+
+            if use_cache:
+                next_decoder_cache += (layer_outputs[2 if output_attentions else 1],)
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+        if self.final_layer_norm is not None:
+            hidden_states = self.final_layer_norm(hidden_states)
+
+        if self.project_out is not None:
+            hidden_states = self.project_out(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+            if encoder_hidden_states is not None:
+                all_cross_attentions += (layer_outputs[2],)
+
+        next_cache = next_decoder_cache if use_cache else None
+        if not return_dict:
+            return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
+        return BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+            cross_attentions=all_cross_attentions,
+        )
+        
+
+
+
+class ThisOPTModel(OPTModel):
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.decoder = ThisOPTDecoder(config)
+
+    def forward(
+        self,
+        input_ids = None,
+        attention_mask = None,
+        encoder_hidden_states=None,
+        encoder_attention_mask = None,
+        head_mask = None,
+        cross_attn_head_mask = None,
+        past_key_values = None,
+        inputs_embeds = None,
+        use_cache = None,
+        output_attentions = None,
+        output_hidden_states = None,
+        return_dict = None,
+    ):
+
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if encoder_hidden_states is not None and encoder_attention_mask is not None:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            encoder_attention_mask = _expand_mask(encoder_attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1])
+
+
+        # decoder outputs consists of (dec_features, past_key_value, dec_hidden, dec_attn)
+        decoder_outputs = self.decoder(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            cross_attn_head_mask=(
+                        cross_attn_head_mask[idx] if cross_attn_head_mask is not None else None
+            ),
+            head_mask=head_mask,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+
+        if not return_dict:
+            return decoder_outputs
+
+        return BaseModelOutputWithPast(
+            last_hidden_state=decoder_outputs.last_hidden_state,
+            past_key_values=decoder_outputs.past_key_values,
+            hidden_states=decoder_outputs.hidden_states,
+            attentions=decoder_outputs.attentions,
+        )
+
+class ThisOPTForCausalLM(OPTForCausalLM):
+    config_class = ThisOPTConfig
+    
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = ThisOPTModel(config)
+
+
+    def forward(
+        self,
+        input_ids = None,
+        attention_mask = None,
+        encoder_hidden_states=None,
+        encoder_attention_mask = None,
+        head_mask = None,
+        cross_attn_head_mask = None,
+        past_key_values = None,
+        inputs_embeds = None,
+        labels = None,
+        use_cache = None,
+        output_attentions = None,
+        output_hidden_states = None,
+        return_dict = None,
+    ) :
+
+
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs = self.model.decoder(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask = encoder_attention_mask,
+            cross_attn_head_mask = cross_attn_head_mask,
+            head_mask=head_mask,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        logits = self.lm_head(outputs[0]).contiguous()
+
+        loss = None
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(shift_logits.view(-1, self.config.vocab_size), shift_labels.view(-1))
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithCrossAttentions(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            cross_attentions=outputs.cross_attentions,
+        )
+    

src/retrieve_caps.py

@@ -0,0 +1,145 @@
+import json
+from tqdm import tqdm
+from transformers import AutoTokenizer
+import clip
+import torch
+import faiss
+import os
+import numpy as np
+from PIL import Image
+from PIL import ImageFile
+ImageFile.LOAD_TRUNCATED_IMAGES = True
+
+def load_coco_data(coco_data_path):
+    """We load in all images and only the train captions."""
+
+    annotations = json.load(open(coco_data_path))['images']
+    images = []
+    captions = []
+    for item in annotations:
+        if item['split'] == 'restval':
+             item['split'] = 'train'
+        if item['split'] == 'train':
+            for sentence in item['sentences']:
+                captions.append({'image_id': item['cocoid'],  'caption': ' '.join(sentence['tokens'])})
+        images.append({'image_id': item['cocoid'], 'file_name': item['filename'].split('_')[-1]})
+ 
+    return images, captions
+
+def filter_captions(data):
+
+    decoder_name = 'gpt2'
+    tokenizer = AutoTokenizer.from_pretrained(decoder_name)
+    bs = 512
+
+    image_ids = [d['image_id'] for d in data]
+    caps = [d['caption'] for d in data]
+    encodings = []
+    for idx in range(0, len(data), bs):
+        encodings += tokenizer.batch_encode_plus(caps[idx:idx+bs], return_tensors='np')['input_ids'].tolist()
+    
+    filtered_image_ids, filtered_captions = [], []
+
+    assert len(image_ids) == len(caps) and len(caps) == len(encodings)
+    for image_id, cap, encoding in zip(image_ids, caps, encodings):
+        if len(encoding) <= 25:
+            filtered_image_ids.append(image_id)
+            filtered_captions.append(cap)
+
+    return filtered_image_ids, filtered_captions
+
+def encode_captions(captions, model, device):
+
+    bs = 256
+    encoded_captions = []
+
+    for idx in tqdm(range(0, len(captions), bs)):
+        with torch.no_grad():
+            input_ids = clip.tokenize(captions[idx:idx+bs]).to(device)
+            encoded_captions.append(model.encode_text(input_ids).cpu().numpy())
+
+    encoded_captions = np.concatenate(encoded_captions)
+
+    return encoded_captions
+
+def encode_images(images, image_path, model, feature_extractor, device):
+
+    image_ids = [i['image_id'] for i in images]
+    
+    bs = 64	
+    image_features = []
+    
+    for idx in tqdm(range(0, len(images), bs)):
+        image_input = [feature_extractor(Image.open(os.path.join(image_path, i['file_name'])))
+                                                                    for i in images[idx:idx+bs]]
+        with torch.no_grad():
+            image_features.append(model.encode_image(torch.tensor(np.stack(image_input)).to(device)).cpu().numpy())
+
+    image_features = np.concatenate(image_features)
+
+    return image_ids, image_features
+
+def get_nns(captions, images, k=15):
+    xq = images.astype(np.float32)
+    xb = captions.astype(np.float32)
+    faiss.normalize_L2(xb)
+    index = faiss.IndexFlatIP(xb.shape[1])
+    index.add(xb)
+    faiss.normalize_L2(xq)
+    D, I = index.search(xq, k) 
+
+    return index, I
+
+def filter_nns(nns, xb_image_ids, captions, xq_image_ids):
+    """ We filter out nearest neighbors which are actual captions for the query image, keeping 7 neighbors per image."""
+    retrieved_captions = {}
+    for nns_list, image_id in zip(nns, xq_image_ids):
+        good_nns = []
+        for nn in zip(nns_list):
+            if xb_image_ids[nn] == image_id:
+                continue
+            good_nns.append(captions[nn])
+            if len(good_nns) == 7:
+                break
+        assert len(good_nns) == 7
+        retrieved_captions[image_id] = good_nns
+    return retrieved_captions
+ 
+def main(): 
+
+    coco_data_path = 'data/dataset_coco.json' # path to Karpathy splits downloaded from Kaggle
+    image_path = 'data/images/'
+    
+    print('Loading data')
+    images, captions = load_coco_data(coco_data_path)
+
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    clip_model, feature_extractor = clip.load("RN50x64", device=device)
+
+    print('Filtering captions')    
+    xb_image_ids, captions = filter_captions(captions)
+
+    print('Encoding captions')
+    encoded_captions = encode_captions(captions, clip_model, device)
+    
+    print('Encoding images')
+    xq_image_ids, encoded_images = encode_images(images, image_path, clip_model, feature_extractor, device)
+    
+    print('Retrieving neighbors')
+    index, nns = get_nns(encoded_captions, encoded_images)
+    retrieved_caps = filter_nns(nns, xb_image_ids, captions, xq_image_ids)
+
+    print('Writing files')
+    faiss.write_index(index, "datastore/coco_index")
+    json.dump(captions, open('datastore/coco_index_captions.json', 'w'))
+
+    json.dump(retrieved_caps, open('data/retrieved_caps_resnet50x64.json', 'w'))
+
+if __name__ == '__main__':
+    main()
+
+
+
+
+    
+

src/retrieve_caps2.py

@@ -0,0 +1,178 @@
+import sys
+import json
+import os.path
+import logging
+import argparse
+from tqdm import tqdm
+import numpy as np
+import torch
+import torch.backends.cudnn as cudnn
+import clip
+from collections import defaultdict
+from PIL import Image
+import faiss 
+import os
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+cudnn.benchmark = True 
+torch.manual_seed(0)
+if torch.cuda.is_available():
+    torch.cuda.manual_seed(0)
+
+import gc
+
+
+
+class ClipRetrieval():
+    def __init__(self, index_name):
+        self.datastore = faiss.read_index(index_name)
+        #self.datastore.nprobe=25
+    
+    def get_nns(self, query_img, k=20):
+        #get k nearest image
+        D, I = self.datastore.search(query_img, k)
+        return D, I[:,:k]
+
+
+class EvalDataset():
+
+    def __init__(self, dataset_splits, images_dir,  images_names, clip_retrieval_processor, eval_split="val_images"):
+        super().__init__()
+
+        with open(dataset_splits) as f:
+            self.split = json.load(f)
+
+        self.split = self.split[eval_split]
+        self.images_dir= images_dir
+
+        with open(args.images_names) as f:
+            self.images_names = json.load(f)
+
+        self.clip_retrieval_processor = clip_retrieval_processor
+
+    def __getitem__(self, i):
+        coco_id = self.split[i]
+
+        image_filename= self.images_dir+self.images_names[coco_id]
+        img_open = Image.open(image_filename).copy()
+        img = np.array(img_open)
+        if len(img.shape) ==2 or img.shape[-1]!=3: #convert grey or CMYK to RGB
+            img_open = img_open.convert('RGB')
+        gc.collect()
+
+        print("img_open",np.array(img_open).shape)
+    
+        #inputs_features_retrieval = self.clip_retrieval_processor(img_open).unsqueeze(0)
+        return self.clip_retrieval_processor(img_open).unsqueeze(0), coco_id
+
+    def __len__(self):
+        return len(self.split)
+
+
+def evaluate(args):
+
+    #load data of the datastore (i.e., captions)
+    with open(args.index_captions) as f:
+        data_datastore = json.load(f)
+
+    datastore = ClipRetrieval(args.datastore_path)
+    datastore_name = args.datastore_path.split("/")[-1]
+    
+    #load clip to encode the images that we want to retrieve captions for
+    clip_retrieval_model, clip_retrieval_feature_extractor = clip.load("RN50x64", device=device)
+    clip_retrieval_model.eval()
+    #data_loader to get images that we want to retrieve captions for
+    data_loader = torch.utils.data.DataLoader(
+        EvalDataset(
+            args.dataset_splits, 
+            args.images_dir,
+            args.images_names,
+            clip_retrieval_feature_extractor, 
+            args.split), 
+        batch_size=1, 
+        shuffle=True,  
+        num_workers=1, 
+        pin_memory=True
+    )
+
+    print("device",device)
+    nearest_caps={}
+    for data in tqdm(data_loader):
+            
+        inputs_features_retrieval, coco_id = data
+        coco_id = coco_id[0]
+
+        #normalize images to retrieve (since datastore has also normalized captions)
+        inputs_features_retrieval = inputs_features_retrieval.to(device)
+        image_retrieval_features = clip_retrieval_model.encode_image(inputs_features_retrieval[0])  
+        image_retrieval_features /= image_retrieval_features.norm(dim=-1, keepdim=True)
+        image_retrieval_features=image_retrieval_features.detach().cpu().numpy().astype(np.float32)
+        
+        print("inputs_features_retrieval",inputs_features_retrieval.size())
+        print("image_retrieval_features",image_retrieval_features.shape)
+
+        D, nearest_ids=datastore.get_nns(image_retrieval_features, k=5)
+        print("D size", D.shape)
+        print("nea", nearest_ids.shape)
+        gc.collect()
+
+        #Since at inference batch is 1
+        D=D[0]
+        nearest_ids=nearest_ids[0]
+
+        list_of_similar_caps=defaultdict(list)
+        for index in range(len(nearest_ids)):
+            nearest_id = str(nearest_ids[index])            
+            nearest_cap=data_datastore[nearest_id]
+    
+            if len(nearest_cap.split()) > args.max_caption_len:
+                print("retrieve cap too big" )
+                continue
+
+            #distance=D[index]
+            #list_of_similar_caps[datastore_name].append((nearest_cap, str(distance)))
+            #list_of_similar_caps[datastore_name].append(nearest_cap)
+
+        #nearest_caps[str(coco_id)]=list_of_similar_caps
+
+
+    #save results
+    outputs_dir = os.path.join(args.output_path, "retrieved_caps")
+    if not os.path.exists(outputs_dir):
+        os.makedirs(outputs_dir)
+
+    data_name=dataset_splits.split("/")[-1]
+
+    name = "nearest_caps_"+data_name +"_w_"+datastore_name + "_"+ args.split
+    results_output_file_name = os.path.join(outputs_dir, name + ".json")
+    json.dump(nearest_caps, open(results_output_file_name, "w"))
+
+
+
+def check_args(args):
+    parser = argparse.ArgumentParser()
+    
+    #Info of the dataset to evaluate on (vizwiz, flick30k, msr-vtt)
+    parser.add_argument("--images_dir",help="Folder where the preprocessed image data is located", default="data/vizwiz/images")
+    parser.add_argument("--dataset_splits",help="File containing the dataset splits", default="data/vizwiz/dataset_splits.json")
+    parser.add_argument("--images_names",help="File containing the images names per id", default="data/vizwiz/images_names.json")
+    parser.add_argument("--split", default="val_images", choices=["val_images", "test_images"])
+    parser.add_argument("--max-caption-len", type=int, default=25)
+
+    #Which datastore to use (web, human)
+    parser.add_argument("--datastore_path", type=str, default="datastore2/vizwiz/vizwiz")
+    parser.add_argument("--index_captions",
+                        help="File containing the captions of the datastore per id", default="datastore2/vizwiz/vizwiz.json")
+    parser.add_argument("--output-path",help="Folder where to store outputs", default="eval_vizwiz_with_datastore_from_vizwiz.json")          
+
+    parsed_args = parser.parse_args(args)
+    return parsed_args
+
+
+if __name__ == "__main__":
+    args = check_args(sys.argv[1:])
+    logging.basicConfig(
+            format='%(levelname)s: %(message)s', level=logging.INFO)
+
+    logging.info(args)
+    evaluate(args)
+

src/template.txt

@@ -0,0 +1,5 @@
+Similar images show
+
+||
+
+This image shows 

src/utils.py

@@ -0,0 +1,131 @@
+from torch.utils.data import Dataset
+from PIL import Image
+import torch
+import json
+import h5py
+import bisect
+
+CAPTION_LENGTH = 25
+SIMPLE_PREFIX = "This image shows "
+
+def prep_strings(text, tokenizer, template=None, retrieved_caps=None, k=None, is_test=False, max_length=None):
+
+    if is_test:
+        padding = False
+        truncation = False
+    else:
+        padding = True 
+        truncation = True
+    
+    if retrieved_caps is not None:
+        infix = '\n\n'.join(retrieved_caps[:k]) + '.'
+        prefix = template.replace('||', infix)
+    else:
+        prefix = SIMPLE_PREFIX
+
+    prefix_ids = tokenizer.encode(prefix)
+    len_prefix = len(prefix_ids)
+
+    text_ids = tokenizer.encode(text, add_special_tokens=False)
+    if truncation:
+        text_ids = text_ids[:CAPTION_LENGTH]
+    input_ids = prefix_ids + text_ids if not is_test else prefix_ids
+
+    # we ignore the prefix (minus one as the first subtoken in the prefix is not predicted)
+    label_ids = [-100] * (len_prefix - 1) + text_ids + [tokenizer.eos_token_id] 
+    if padding:
+        input_ids += [tokenizer.pad_token_id] * (max_length - len(input_ids))
+        label_ids += [-100] * (max_length - len(label_ids))
+    
+    if is_test:
+        return input_ids
+    else:  
+        return input_ids, label_ids
+
+def postprocess_preds(pred, tokenizer):
+    pred = pred.split(SIMPLE_PREFIX)[-1]
+    pred = pred.replace(tokenizer.pad_token, '')
+    if pred.startswith(tokenizer.bos_token):
+        pred = pred[len(tokenizer.bos_token):]
+    if pred.endswith(tokenizer.eos_token):
+        pred = pred[:-len(tokenizer.eos_token)]
+    return pred
+
+class TrainDataset(Dataset):
+    def __init__(self, df, features_path, tokenizer, rag=False, template_path=None, k=None, max_caption_length=25):
+        self.df = df
+        self.tokenizer = tokenizer
+        self.features = h5py.File(features_path, 'r')
+
+        if rag:
+            self.template = open(template_path).read().strip() + ' '
+            self.max_target_length = (max_caption_length  # target caption
+                                     + max_caption_length * k # retrieved captions
+                                     + len(tokenizer.encode(self.template)) # template
+                                     + len(tokenizer.encode('\n\n')) * (k-1) # separator between captions
+                                     )
+            assert k is not None 
+            self.k = k
+        self.rag = rag
+
+    def __len__(self):
+        return len(self.df)
+
+    def __getitem__(self, idx):
+        text = self.df['text'][idx]
+        if self.rag: 
+            caps = self.df['caps'][idx]
+            decoder_input_ids, labels = prep_strings(text, self.tokenizer, template=self.template,
+                                                     retrieved_caps=caps, k=self.k, max_length=self.max_target_length)
+        else:
+            decoder_input_ids, labels = prep_strings(text, self.tokenizer, max_length=self.max_target_length)
+        # load precomputed features
+        encoder_outputs = self.features[self.df['cocoid'][idx]][()]
+        encoding = {"encoder_outputs": torch.tensor(encoder_outputs), 
+                    "decoder_input_ids": torch.tensor(decoder_input_ids),
+                    "labels": torch.tensor(labels)}
+
+        return encoding
+
+
+def load_data_for_training(annot_path, caps_path=None):
+    annotations = json.load(open(annot_path))['images']
+    if caps_path is not None:
+        retrieved_caps = json.load(open(caps_path))
+    data = {'train': [], 'val': []}
+
+    for item in annotations:
+        file_name = item['filename'].split('_')[-1]
+        if caps_path is not None:
+            caps = retrieved_caps[str(item['cocoid'])]
+        else:
+            caps = None
+        samples = []
+        for sentence in item['sentences']:
+            samples.append({'file_name': file_name, 'cocoid': str(item['cocoid']), 'caps': caps, 'text': ' '.join(sentence['tokens'])})
+        if item['split'] == 'train' or item['split'] == 'restval':
+            data['train'] += samples
+        elif item['split'] == 'val':
+            data['val'] += samples
+    return data 
+
+def load_data_for_inference(annot_path, caps_path=None):
+    annotations = json.load(open(annot_path))['images']
+    if caps_path is not None:
+        retrieved_caps = json.load(open(caps_path))
+    data = {'test': [], 'val': []}
+
+    for item in annotations:
+        file_name = item['filename'].split('_')[-1]
+        if caps_path is not None:
+            caps = retrieved_caps[str(item['cocoid'])]
+        else:
+            caps = None
+        image = {'file_name': file_name, 'caps': caps, 'image_id': str(item['cocoid'])}
+        if item['split'] == 'test':
+            data['test'].append(image)
+        elif item['split'] == 'val':
+            data['val'].append(image)
+
+    return data      
+

src/utils_generate_retrieved_caps.py

@@ -0,0 +1,135 @@
+from torch.utils.data import Dataset
+from PIL import Image
+import torch
+import json
+import h5py
+import bisect
+
+CAPTION_LENGTH = 25
+SIMPLE_PREFIX = "This image shows "
+
+def prep_strings(text, tokenizer, template=None, retrieved_caps=None, k=None, is_test=False, max_length=None):
+
+    if is_test:
+        padding = False
+        truncation = False
+    else:
+        padding = True 
+        truncation = True
+    
+    if retrieved_caps is not None:
+        infix = '\n\n'.join(retrieved_caps[:k]) + '.'
+        prefix = template.replace('||', infix)
+    else:
+        prefix = SIMPLE_PREFIX
+
+    prefix_ids = tokenizer.encode(prefix)
+    len_prefix = len(prefix_ids)
+
+    text_ids = tokenizer.encode(text, add_special_tokens=False)
+    if truncation:
+        text_ids = text_ids[:CAPTION_LENGTH]
+    input_ids = prefix_ids + text_ids if not is_test else prefix_ids
+
+    # we ignore the prefix (minus one as the first subtoken in the prefix is not predicted)
+    label_ids = [-100] * (len_prefix - 1) + text_ids + [tokenizer.eos_token_id] 
+    if padding:
+        input_ids += [tokenizer.pad_token_id] * (max_length - len(input_ids))
+        label_ids += [-100] * (max_length - len(label_ids))
+    
+    if is_test:
+        return input_ids
+    else:  
+        return input_ids, label_ids
+
+def postprocess_preds(pred, tokenizer):
+    pred = pred.split(SIMPLE_PREFIX)[-1]
+    pred = pred.replace(tokenizer.pad_token, '')
+    if pred.startswith(tokenizer.bos_token):
+        pred = pred[len(tokenizer.bos_token):]
+    if pred.endswith(tokenizer.eos_token):
+        pred = pred[:-len(tokenizer.eos_token)]
+    return pred
+
+class TrainDataset(Dataset):
+    def __init__(self, df, features_path, tokenizer, rag=False, template_path=None, k=None, max_caption_length=25):
+        self.df = df
+        self.tokenizer = tokenizer
+        self.features = h5py.File(features_path, 'r')
+
+        if rag:
+            self.template = open(template_path).read().strip() + ' '
+            self.max_target_length = (max_caption_length  # target caption
+                                     + max_caption_length * k # retrieved captions
+                                     + len(tokenizer.encode(self.template)) # template
+                                     + len(tokenizer.encode('\n\n')) * (k-1) # separator between captions
+                                     )
+            assert k is not None 
+            self.k = k
+        self.rag = rag
+
+    def __len__(self):
+        return len(self.df)
+
+    def __getitem__(self, idx):
+        text = self.df['text'][idx]
+        if self.rag: 
+            caps = self.df['caps'][idx]
+            decoder_input_ids, labels = prep_strings(text, self.tokenizer, template=self.template,
+                                                     retrieved_caps=caps, k=self.k, max_length=self.max_target_length)
+        else:
+            decoder_input_ids, labels = prep_strings(text, self.tokenizer, max_length=self.max_target_length)
+        # load precomputed features
+        encoder_outputs = self.features[self.df['cocoid'][idx]][()]
+        encoding = {"encoder_outputs": torch.tensor(encoder_outputs), 
+                    "decoder_input_ids": torch.tensor(decoder_input_ids),
+                    "labels": torch.tensor(labels)}
+
+        return encoding
+
+
+def load_data_for_training(annot_path, caps_path=None):
+    annotations = json.load(open(annot_path))['images']
+    if caps_path is not None:
+        retrieved_caps = json.load(open(caps_path))
+    data = {'train': [], 'val': []}
+
+    for item in annotations:
+        file_name = item['filename'].split('_')[-1]
+        caps = retrieved_caps[str(item['cocoid'])]
+        
+        samples = []
+        for sentence in item['sentences']:
+            print("how are the retrieved caps", caps + ' '.join(sentence['tokens']))
+
+            samples.append({'file_name': file_name, 'cocoid': str(item['cocoid']), 'caps': None, 'text': " ".join(caps) + ' '.join(sentence['tokens'])})
+        if item['split'] == 'train' or item['split'] == 'restval':
+            data['train'] += samples
+        elif item['split'] == 'val':
+            data['val'] += samples
+    return data 
+
+
+
+
+
+def load_data_for_inference(annot_path, caps_path=None):
+    annotations = json.load(open(annot_path))['images']
+    if caps_path is not None:
+        retrieved_caps = json.load(open(caps_path))
+    data = {'test': [], 'val': []}
+
+    for item in annotations:
+        file_name = item['filename'].split('_')[-1]
+        if caps_path is not None:
+            caps = retrieved_caps[str(item['cocoid'])]
+        else:
+            caps = None
+        image = {'file_name': file_name, 'caps': caps, 'image_id': str(item['cocoid'])}
+        if item['split'] == 'test':
+            data['test'].append(image)
+        elif item['split'] == 'val':
+            data['val'].append(image)
+
+    return data      
+

src/vision_encoder_decoder.py

@@ -0,0 +1,560 @@
+# coding=utf-8
+# Copyright 2021 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Classes to support Vision-Encoder-Text-Decoder architectures"""
+import timeit
+
+from typing import Optional
+
+import torch
+from torch import nn
+from torch.nn import CrossEntropyLoss
+from transformers.configuration_utils import PretrainedConfig
+from transformers.modeling_outputs import BaseModelOutput, Seq2SeqLMOutput
+from transformers.modeling_utils import PreTrainedModel
+#from transformers.utils import add_start_docstrings, add_start_docstrings_to_model_forward, logging, replace_return_docstrings
+from transformers.utils import logging
+from transformers.models.auto.configuration_auto import AutoConfig
+from transformers.models.auto.modeling_auto import AutoModel, AutoModelForCausalLM
+from transformers.models.vision_encoder_decoder.configuration_vision_encoder_decoder import VisionEncoderDecoderConfig
+import inspect
+
+from .gpt2 import ThisGPT2LMHeadModel
+from .gpt2 import ThisGPT2Config
+from .xglm import ThisXGLMForCausalLM
+from .xglm import ThisXGLMConfig
+from .opt import ThisOPTForCausalLM
+from .opt import ThisOPTConfig
+
+# Copied from transformers.models.encoder_decoder.modeling_encoder_decoder.shift_tokens_right
+def shift_tokens_right(input_ids: torch.Tensor, pad_token_id: int, decoder_start_token_id: int):
+    """
+    Shift input ids one token to the right.
+    """
+    shifted_input_ids = input_ids.new_zeros(input_ids.shape)
+    shifted_input_ids[:, 1:] = input_ids[:, :-1].clone()
+    if decoder_start_token_id is None:
+        raise ValueError("Make sure to set the decoder_start_token_id attribute of the model's configuration.")
+    shifted_input_ids[:, 0] = decoder_start_token_id
+
+    if pad_token_id is None:
+        raise ValueError("Make sure to set the pad_token_id attribute of the model's configuration.")
+    # replace possible -100 values in labels by `pad_token_id`
+    shifted_input_ids.masked_fill_(shifted_input_ids == -100, pad_token_id)
+
+    return shifted_input_ids
+
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "SmallCapConfig"
+
+VISION_ENCODER_DECODER_START_DOCSTRING = r"""
+    This class can be used to initialize an image-to-text-sequence model with any pretrained vision autoencoding model
+    as the encoder and any pretrained text autoregressive model as the decoder. The encoder is loaded via
+    [`~AutoModel.from_pretrained`] function and the decoder is loaded via [`~AutoModelForCausalLM.from_pretrained`]
+    function. Cross-attention layers are automatically added to the decoder and should be fine-tuned on a downstream
+    generative task, like image captioning.
+
+    The effectiveness of initializing sequence-to-sequence models with pretrained checkpoints for sequence generation
+    tasks was shown in [Leveraging Pre-trained Checkpoints for Sequence Generation
+    Tasks](https://arxiv.org/abs/1907.12461) by Sascha Rothe, Shashi Narayan, Aliaksei Severyn. Michael Matena, Yanqi
+    Zhou, Wei Li, Peter J. Liu.
+
+    Additionally, in [TrOCR: Transformer-based Optical Character Recognition with Pre-trained
+    Models](https://arxiv.org/abs/2109.10282) it is shown how leveraging large pretrained vision models for optical
+    character recognition (OCR) yields a significant performance improvement.
+
+    After such a Vision-Encoder-Text-Decoder model has been trained/fine-tuned, it can be saved/loaded just like any
+    other models (see the examples for more information).
+
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`VisionEncoderDecoderConfig`]): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+VISION_ENCODER_DECODER_INPUTS_DOCSTRING = r"""
+    Args:
+        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            Pixel values. Pixel values can be obtained using a feature extractor (e.g. if you use ViT as the encoder,
+            you should use [`ViTFeatureExtractor`]). See [`ViTFeatureExtractor.__call__`] for details.
+        decoder_input_ids (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*):
+            Indices of decoder input sequence tokens in the vocabulary.
+
+            Indices can be obtained using [`PreTrainedTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+
+            If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
+            `past_key_values`).
+
+            For training, `decoder_input_ids` are automatically created by the model by shifting the `labels` to the
+            right, replacing -100 by the `pad_token_id` and prepending them with the `decoder_start_token_id`.
+        decoder_attention_mask (`torch.BoolTensor` of shape `(batch_size, target_sequence_length)`, *optional*):
+            Default behavior: generate a tensor that ignores pad tokens in `decoder_input_ids`. Causal mask will also
+            be used by default.
+        encoder_outputs (`tuple(torch.FloatTensor)`, *optional*):
+            This tuple must consist of (`last_hidden_state`, *optional*: `hidden_states`, *optional*: `attentions`)
+            `last_hidden_state` (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`) is a tensor
+            of hidden-states at the output of the last layer of the encoder. Used in the cross-attention of the
+            decoder.
+        past_key_values (`tuple(tuple(torch.FloatTensor))` of length `config.n_layers` with each tuple having 4 tensors of shape `(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
+            Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
+
+            If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
+            don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
+            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+        decoder_inputs_embeds (`torch.FloatTensor` of shape `(batch_size, target_sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `decoder_input_ids` you can choose to directly pass an embedded
+            representation. This is useful if you want more control over how to convert `decoder_input_ids` indices
+            into associated vectors than the model's internal embedding lookup matrix.
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss for the decoder. Indices should be in `[-100, 0,
+            ..., config.vocab_size]` (see `input_ids` docstring) Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            If set to `True`, the model will return a [`~utils.Seq2SeqLMOutput`] instead of a plain tuple.
+        kwargs: (*optional*) Remaining dictionary of keyword arguments. Keyword arguments come in two flavors:
+
+            - Without a prefix which will be input as `**encoder_kwargs` for the encoder forward function.
+            - With a *decoder_* prefix which will be input as `**decoder_kwargs` for the decoder forward function.
+"""
+
+class SmallCapConfig(VisionEncoderDecoderConfig):
+    model_type = "smallcap"
+
+    def __init__(
+        self,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        
+
+class SmallCap(PreTrainedModel):
+    r"""
+    [`VisionEncoderDecoderModel`] is a generic model class that will be instantiated as a transformer architecture with
+    one of the base vision model classes of the library as encoder and another one as decoder when created with the
+    :meth*~transformers.AutoModel.from_pretrained* class method for the encoder and
+    :meth*~transformers.AutoModelForCausalLM.from_pretrained* class method for the decoder.
+    """
+    config_class = SmallCapConfig
+    base_model_prefix = "smallcap"
+    main_input_name = "pixel_values"
+
+    def __init__(
+        self,
+        config: Optional[PretrainedConfig] = None,
+        encoder: Optional[PreTrainedModel] = None,
+        decoder: Optional[PreTrainedModel] = None,
+    ):
+        if config is None and (encoder is None or decoder is None):
+            raise ValueError("Either a configuration or an encoder and a decoder has to be provided.")
+        if config is None:
+            config = SmallCapConfig.from_encoder_decoder_configs(encoder.config, decoder.config)
+        else:
+            if not isinstance(config, self.config_class):
+                raise ValueError(f"Config: {config} has to be of type {self.config_class}")
+
+        if config.decoder.cross_attention_hidden_size is not None:
+            if config.decoder.cross_attention_hidden_size != config.encoder.hidden_size:
+                raise ValueError(
+                    "If `cross_attention_hidden_size` is specified in the decoder's configuration, it has to be equal#"
+                    f" to the encoder's `hidden_size`. Got {config.decoder.cross_attention_hidden_size} for"
+                    f" `config.decoder.cross_attention_hidden_size` and {config.encoder.hidden_size} for"
+                    " `config.encoder.hidden_size`."
+                )
+
+        # initialize with config
+        # make sure input & output embeddings is not tied
+        config.tie_word_embeddings = False
+        super().__init__(config)
+
+        if encoder is None:
+            encoder = AutoModel.from_config(config.encoder)
+
+        if decoder is None:
+            decoder = AutoModelForCausalLM.from_config(config.decoder)
+
+        self.encoder = encoder.vision_model
+        self.encoder.main_input_name = 'pixel_values'
+        self.decoder = decoder
+        # make sure that the individual model's config refers to the shared config
+        # so that the updates to the config will be synced
+        self.encoder.config = self.config.encoder
+        self.decoder.config = self.config.decoder
+
+    def get_encoder(self):
+        return self.encoder
+
+    def get_decoder(self):
+        return self.decoder
+
+    def get_output_embeddings(self):
+        return self.decoder.get_output_embeddings()
+
+    def set_output_embeddings(self, new_embeddings):
+        return self.decoder.set_output_embeddings(new_embeddings)
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        # At the moment fast initialization is not supported for composite models
+        if kwargs.get("_fast_init", False):
+            logger.warning(
+                "Fast initialization is currently not supported for VisionEncoderDecoderModel. "
+                "Falling back to slow initialization..."
+            )
+        kwargs["_fast_init"] = False
+        return super().from_pretrained(*args, **kwargs)
+
+    @classmethod
+    def from_encoder_decoder_pretrained(
+        cls,
+        encoder_pretrained_model_name_or_path: str = None,
+        decoder_pretrained_model_name_or_path: str = None,
+        cross_attention_reduce_factor: int = None,
+        *model_args,
+        **kwargs
+    ) -> PreTrainedModel:
+        r"""
+        Instantiate an encoder and a decoder from one or two base classes of the library from pretrained model
+        checkpoints.
+
+
+        The model is set in evaluation mode by default using `model.eval()` (Dropout modules are deactivated). To train
+        the model, you need to first set it back in training mode with `model.train()`.
+
+        Params:
+            encoder_pretrained_model_name_or_path (`str`, *optional*):
+                Information necessary to initiate the image encoder. Can be either:
+
+                    - A string, the *model id* of a pretrained model hosted inside a model repo on huggingface.co. An
+                      example is `google/vit-base-patch16-224-in21k`.
+                    - A path to a *directory* containing model weights saved using
+                      [`~PreTrainedModel.save_pretrained`], e.g., `./my_model_directory/`.
+                    - A path or url to a *tensorflow index checkpoint file* (e.g, `./tf_model/model.ckpt.index`). In
+                      this case, `from_tf` should be set to `True` and a configuration object should be provided as
+                      `config` argument. This loading path is slower than converting the TensorFlow checkpoint in a
+                      PyTorch model using the provided conversion scripts and loading the PyTorch model afterwards.
+
+            decoder_pretrained_model_name_or_path (`str`, *optional*, defaults to `None`):
+                Information necessary to initiate the text decoder. Can be either:
+
+                    - A string, the *model id* of a pretrained model hosted inside a model repo on huggingface.co.
+                      Valid model ids can be located at the root-level, like `bert-base-uncased`, or namespaced under a
+                      user or organization name, like `dbmdz/bert-base-german-cased`.
+                    - A path to a *directory* containing model weights saved using
+                      [`~PreTrainedModel.save_pretrained`], e.g., `./my_model_directory/`.
+                    - A path or url to a *tensorflow index checkpoint file* (e.g, `./tf_model/model.ckpt.index`). In
+                      this case, `from_tf` should be set to `True` and a configuration object should be provided as
+                      `config` argument. This loading path is slower than converting the TensorFlow checkpoint in a
+                      PyTorch model using the provided conversion scripts and loading the PyTorch model afterwards.
+
+            model_args (remaining positional arguments, *optional*):
+                All remaning positional arguments will be passed to the underlying model's `__init__` method.
+
+            kwargs (remaining dictionary of keyword arguments, *optional*):
+                Can be used to update the configuration object (after it being loaded) and initiate the model (e.g.,
+                `output_attentions=True`).
+
+                - To update the encoder configuration, use the prefix *encoder_* for each configuration parameter.
+                - To update the decoder configuration, use the prefix *decoder_* for each configuration parameter.
+                - To update the parent model configuration, do not use a prefix for each configuration parameter.
+
+                Behaves differently depending on whether a `config` is provided or automatically loaded.
+
+        Example:
+
+        ```python
+        >>> from transformers import VisionEncoderDecoderModel
+
+        >>> # initialize a vit-bert from a pretrained ViT and a pretrained BERT model. Note that the cross-attention layers will be randomly initialized
+        >>> model = VisionEncoderDecoderModel.from_encoder_decoder_pretrained(
+        ...     "google/vit-base-patch16-224-in21k", "bert-base-uncased"
+        ... )
+        >>> # saving model after fine-tuning
+        >>> model.save_pretrained("./vit-bert")
+        >>> # load fine-tuned model
+        >>> model = VisionEncoderDecoderModel.from_pretrained("./vit-bert")
+        ```"""
+
+        kwargs_encoder = {
+            argument[len("encoder_") :]: value for argument, value in kwargs.items() if argument.startswith("encoder_")
+        }
+
+        kwargs_decoder = {
+            argument[len("decoder_") :]: value for argument, value in kwargs.items() if argument.startswith("decoder_")
+        }
+
+        # remove encoder, decoder kwargs from kwargs
+        for key in kwargs_encoder.keys():
+            del kwargs["encoder_" + key]
+        for key in kwargs_decoder.keys():
+            del kwargs["decoder_" + key]
+
+        # Load and initialize the encoder and decoder
+        # The distinction between encoder and decoder at the model level is made
+        # by the value of the flag `is_decoder` that we need to set correctly.
+        encoder = kwargs_encoder.pop("model", None)
+        if encoder is None:
+            if encoder_pretrained_model_name_or_path is None:
+                raise ValueError(
+                    "If `encoder_model` is not defined as an argument, a `encoder_pretrained_model_name_or_path` has "
+                    "to be defined."
+                )
+
+            if "config" not in kwargs_encoder:
+                encoder_config, kwargs_encoder = AutoConfig.from_pretrained(
+                    encoder_pretrained_model_name_or_path, **kwargs_encoder, return_unused_kwargs=True
+                )
+
+                if encoder_config.is_decoder is True or encoder_config.add_cross_attention is True:
+                    logger.info(
+                        f"Initializing {encoder_pretrained_model_name_or_path} as a encoder model "
+                        "from a decoder model. Cross-attention and casual mask are disabled."
+                    )
+                    encoder_config.is_decoder = False
+                    encoder_config.add_cross_attention = False
+
+                kwargs_encoder["config"] = encoder_config
+
+            encoder = AutoModel.from_pretrained(encoder_pretrained_model_name_or_path, *model_args, **kwargs_encoder)
+
+        decoder = kwargs_decoder.pop("model", None)
+        if decoder is None:
+            if decoder_pretrained_model_name_or_path is None:
+                raise ValueError(
+                    "If `decoder_model` is not defined as an argument, a `decoder_pretrained_model_name_or_path` has "
+                    "to be defined."
+                )
+
+            if "config" not in kwargs_decoder:
+                if "xglm" in decoder_pretrained_model_name_or_path:
+                    decoder_config, kwargs_decoder = ThisXGLMConfig.from_pretrained(
+                        decoder_pretrained_model_name_or_path, **kwargs_decoder, return_unused_kwargs=True
+                    )
+
+                elif "opt" in decoder_pretrained_model_name_or_path:
+                    decoder_config, kwargs_decoder = ThisOPTConfig.from_pretrained(
+                        decoder_pretrained_model_name_or_path, **kwargs_decoder, return_unused_kwargs=True
+                    )
+
+                else:
+                    decoder_config, kwargs_decoder = ThisGPT2Config.from_pretrained(
+                        decoder_pretrained_model_name_or_path, **kwargs_decoder, return_unused_kwargs=True
+                    )
+
+                if decoder_config.is_decoder is False or decoder_config.add_cross_attention is False:
+                    logger.info(
+                        f"Initializing {decoder_pretrained_model_name_or_path} as a decoder model. Cross attention"
+                        f" layers are added to {decoder_pretrained_model_name_or_path} and randomly initialized if"
+                        f" {decoder_pretrained_model_name_or_path}'s architecture allows for cross attention layers."
+                    )
+                    decoder_config.is_decoder = True
+                    decoder_config.add_cross_attention = True
+                decoder_config.encoder_hidden_size = encoder.config.vision_config.hidden_size
+                decoder_config.cross_attention_reduce_factor = cross_attention_reduce_factor
+                kwargs_decoder["config"] = decoder_config
+            
+            if kwargs_decoder["config"].is_decoder is False or kwargs_decoder["config"].add_cross_attention is False:
+                logger.warning(
+                    f"Decoder model {decoder_pretrained_model_name_or_path} is not initialized as a decoder. "
+                    f"In order to initialize {decoder_pretrained_model_name_or_path} as a decoder, "
+                    "make sure that the attributes `is_decoder` and `add_cross_attention` of `decoder_config` "
+                    "passed to `.from_encoder_decoder_pretrained(...)` are set to `True` or do not pass a "
+                    "`decoder_config` to `.from_encoder_decoder_pretrained(...)`"
+                )
+            
+            #decoder = AutoModelForCausalLM.from_pretrained(decoder_pretrained_model_name_or_path, **kwargs_decoder)
+            if "xglm" in decoder_pretrained_model_name_or_path:
+                decoder = ThisXGLMForCausalLM.from_pretrained(decoder_pretrained_model_name_or_path, **kwargs_decoder)
+
+            elif "opt" in decoder_pretrained_model_name_or_path:
+                decoder = ThisOPTForCausalLM.from_pretrained(decoder_pretrained_model_name_or_path, **kwargs_decoder)
+            else:
+                decoder = ThisGPT2LMHeadModel.from_pretrained(decoder_pretrained_model_name_or_path, **kwargs_decoder)
+
+        # instantiate config with corresponding kwargs
+        config = SmallCapConfig.from_encoder_decoder_configs(encoder.config, decoder.config, **kwargs)
+
+        # make sure input & output embeddings is not tied
+        config.tie_word_embeddings = False
+        return cls(encoder=encoder, decoder=decoder, config=config)
+
+    def forward(
+        self,
+        pixel_values=None,
+        decoder_input_ids=None,
+        decoder_attention_mask=None,
+        encoder_outputs=None,
+        past_key_values=None,
+        decoder_inputs_embeds=None,
+        labels=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        **kwargs,
+    ):
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from transformers import TrOCRProcessor, VisionEncoderDecoderModel
+        >>> import requests
+        >>> from PIL import Image
+        >>> import torch
+
+        >>> processor = TrOCRProcessor.from_pretrained("microsoft/trocr-base-handwritten")
+        >>> model = VisionEncoderDecoderModel.from_pretrained("microsoft/trocr-base-handwritten")
+
+        >>> # load image from the IAM dataset
+        >>> url = "https://fki.tic.heia-fr.ch/static/img/a01-122-02.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw).convert("RGB")
+
+        >>> # training
+        >>> model.config.decoder_start_token_id = processor.tokenizer.cls_token_id
+        >>> model.config.pad_token_id = processor.tokenizer.pad_token_id
+        >>> model.config.vocab_size = model.config.decoder.vocab_size
+
+        >>> pixel_values = processor(image, return_tensors="pt").pixel_values
+        >>> text = "hello world"
+        >>> labels = processor.tokenizer(text, return_tensors="pt").input_ids
+        >>> outputs = model(pixel_values=pixel_values, labels=labels)
+        >>> loss = outputs.loss
+
+        >>> # inference (generation)
+        >>> generated_ids = model.generate(pixel_values)
+        >>> generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
+        ```"""
+ 
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        kwargs_encoder = {argument: value for argument, value in kwargs.items() if not argument.startswith("decoder_")}
+
+        kwargs_decoder = {
+            argument[len("decoder_") :]: value for argument, value in kwargs.items() if argument.startswith("decoder_")
+        }
+        if encoder_outputs is None:
+            if pixel_values is None:
+                raise ValueError("You have to specify pixel_values")
+            
+            encoder_outputs = self.encoder(
+                pixel_values=pixel_values,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+                **kwargs_encoder,
+            )
+        elif isinstance(encoder_outputs, tuple):
+            encoder_outputs = BaseModelOutput(*encoder_outputs)
+        else:
+            encoder_outputs = BaseModelOutput(encoder_outputs, None)
+
+        encoder_hidden_states = encoder_outputs[0]
+
+        # else:
+        encoder_attention_mask = None
+        if (labels is not None) and (decoder_input_ids is None and decoder_inputs_embeds is None):
+            decoder_input_ids = shift_tokens_right(
+                labels, self.config.pad_token_id, self.config.decoder_start_token_id
+            )
+        
+        # Decode
+        decoder_outputs = self.decoder(
+            input_ids=decoder_input_ids,
+            attention_mask=decoder_attention_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            inputs_embeds=decoder_inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            use_cache=use_cache,
+            past_key_values=past_key_values,
+            return_dict=return_dict,
+            **kwargs_decoder,
+        )
+
+        # Compute loss independent from decoder (as some shift the logits inside them)
+        loss = None
+        if labels is not None:
+            logits = decoder_outputs.logits if return_dict else decoder_outputs[0]
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(logits.reshape(-1, self.decoder.config.vocab_size), labels.view(-1))
+        
+        if not return_dict:
+            if loss is not None:
+                return (loss,) + decoder_outputs + encoder_outputs
+            else:
+                return decoder_outputs + encoder_outputs
+
+        return Seq2SeqLMOutput(
+            loss=loss,
+            logits=decoder_outputs.logits,
+            past_key_values=decoder_outputs.past_key_values,
+            decoder_hidden_states=decoder_outputs.hidden_states,
+            decoder_attentions=decoder_outputs.attentions,
+            cross_attentions=decoder_outputs.cross_attentions,
+            encoder_last_hidden_state=encoder_outputs.last_hidden_state,
+            encoder_hidden_states=encoder_outputs.hidden_states,
+            encoder_attentions=encoder_outputs.attentions,
+        )
+
+    def prepare_decoder_input_ids_from_labels(self, labels: torch.Tensor):
+        return shift_tokens_right(labels, self.config.pad_token_id, self.config.decoder_start_token_id)
+
+    def prepare_inputs_for_generation(
+        self, input_ids, past=None, attention_mask=None, use_cache=None, encoder_outputs=None, **kwargs
+    ):
+        decoder_inputs = self.decoder.prepare_inputs_for_generation(input_ids, past=past)
+        decoder_attention_mask = decoder_inputs["attention_mask"] if "attention_mask" in decoder_inputs else None
+        input_dict = {
+            "attention_mask": attention_mask,
+            "decoder_attention_mask": decoder_attention_mask,
+            "decoder_input_ids": decoder_inputs["input_ids"],
+            "encoder_outputs": encoder_outputs,
+            "past_key_values": decoder_inputs["past_key_values"],
+            "use_cache": use_cache,
+        }
+        return input_dict
+
+    def resize_token_embeddings(self, *args, **kwargs):
+        raise NotImplementedError(
+            "Resizing the embedding layers via the VisionEncoderDecoderModel directly is not supported.Please use the"
+            " respective methods of the wrapped decoder object (model.decoder.resize_token_embeddings(...))"
+        )
+
+    def _reorder_cache(self, past, beam_idx):
+        # apply decoder cache reordering here
+        return self.decoder._reorder_cache(past, beam_idx)

src/xglm.py

@@ -0,0 +1,269 @@
+# coding=utf-8
+# Copyright 2018 The OpenAI Team Authors and HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PyTorch OpenAI GPT-2 model."""
+
+import math
+import os
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from packaging import version
+from torch import nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+
+from transformers.models.gpt2.modeling_gpt2 import load_tf_weights_in_gpt2, GPT2LMHeadModel, GPT2MLP, GPT2Attention, GPT2Block, GPT2Model 
+
+from transformers.models.xglm.modeling_xglm import XGLMForCausalLM, XGLMAttention, XGLMDecoderLayer, XGLMModel 
+
+from transformers.activations import ACT2FN
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPastAndCrossAttentions,
+    CausalLMOutputWithCrossAttentions,
+    SequenceClassifierOutputWithPast,
+    TokenClassifierOutput,
+)
+from transformers.modeling_utils import PreTrainedModel, SequenceSummary
+from transformers.pytorch_utils import Conv1D, find_pruneable_heads_and_indices, prune_conv1d_layer
+from transformers.utils import (
+    ModelOutput,
+    logging,
+)
+from transformers.utils.model_parallel_utils import assert_device_map, get_device_map
+from transformers.models.xglm.configuration_xglm import XGLMConfig
+
+
+if version.parse(torch.__version__) >= version.parse("1.6"):
+    is_amp_available = True
+    from torch.cuda.amp import autocast
+else:
+    is_amp_available = False
+
+    
+class ThisXGLMConfig(XGLMConfig):
+    model_type = "this_xglm"
+
+    def __init__(
+        self,
+        cross_attention_reduce_factor = 1,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.cross_attention_reduce_factor = cross_attention_reduce_factor
+        
+
+class ThisXGLMAttention(XGLMAttention):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(
+        self,
+        embed_dim,
+        num_heads,
+        dropout= 0.0,
+        is_decoder= False,
+        bias= True,
+        config=None,
+        is_cross_attention=False,
+    ):
+        super().__init__(embed_dim,num_heads, dropout,is_decoder,bias)
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.head_dim = embed_dim // num_heads
+
+        if (self.head_dim * num_heads) != self.embed_dim:
+            raise ValueError(
+                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim}"
+                f" and `num_heads`: {num_heads})."
+            )
+        self.scaling = self.head_dim**-0.5
+        self.is_decoder = is_decoder
+
+        self.cross_attention_reduce_factor = config.cross_attention_reduce_factor
+        self.head_dim = int(self.head_dim / self.cross_attention_reduce_factor)
+
+
+        if is_cross_attention:
+            #print("self", int(embed_dim / self.cross_attention_reduce_factor))
+            self.k_proj = nn.Linear(768, int(embed_dim / self.cross_attention_reduce_factor), bias=bias)
+            #print("self.k_proj",self.k_proj)
+            self.v_proj = nn.Linear(768, int(embed_dim / self.cross_attention_reduce_factor), bias=bias)
+            self.q_proj = nn.Linear(embed_dim, int(embed_dim / self.cross_attention_reduce_factor), bias=bias)
+            self.out_proj = nn.Linear(int(embed_dim / self.cross_attention_reduce_factor),embed_dim, bias=bias)
+
+            self.embed_dim=int(embed_dim / self.cross_attention_reduce_factor)
+        else:
+            self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+            self.v_proj = nn.Linear(embed_dim, embed_dim , bias=bias)
+            self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+            self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+
+    def forward(
+        self,
+        hidden_states,
+        key_value_states,
+        past_key_value,
+        attention_mask,
+        layer_head_mask,
+        output_attentions,
+    ):
+        """Input shape: Batch x Time x Channel"""
+
+        # if key_value_states are provided this layer is used as a cross-attention layer
+        # for the decoder
+        is_cross_attention = key_value_states is not None
+
+        bsz, tgt_len, _ = hidden_states.size()
+
+        # get query proj
+        query_states = self.q_proj(hidden_states) * self.scaling
+        # get key, value proj
+        if is_cross_attention and past_key_value is not None:
+            # reuse k,v, cross_attentions
+            key_states = past_key_value[0]
+            value_states = past_key_value[1]
+        elif is_cross_attention:
+            # cross_attentions
+            #print("key_value_states",key_value_states.size())
+            #print("self.k_proj(key_value_states)",self.k_proj(key_value_states).size())
+                
+            key_states = self._shape(self.k_proj(key_value_states), -1, bsz)
+            value_states = self._shape(self.v_proj(key_value_states), -1, bsz)
+        elif past_key_value is not None:
+            # reuse k, v, self_attention
+            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+            key_states = torch.cat([past_key_value[0], key_states], dim=2)
+            value_states = torch.cat([past_key_value[1], value_states], dim=2)
+        else:
+            # self_attention
+            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+
+        if self.is_decoder:
+            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+            past_key_value = (key_states, value_states)
+
+        proj_shape = (bsz * self.num_heads, -1, self.head_dim)
+        query_states = self._shape(query_states, tgt_len, bsz).view(*proj_shape)
+        key_states = key_states.view(*proj_shape)
+        value_states = value_states.view(*proj_shape)
+
+        src_len = key_states.size(1)
+        attn_weights = torch.bmm(query_states, key_states.transpose(1, 2))
+
+        if attn_weights.size() != (bsz * self.num_heads, tgt_len, src_len):
+            raise ValueError(
+                f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is"
+                f" {attn_weights.size()}"
+            )
+
+        if attention_mask is not None:
+            if attention_mask.size() != (bsz, 1, tgt_len, src_len):
+                raise ValueError(
+                    f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}"
+                )
+            attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attention_mask
+            attn_weights = torch.max(attn_weights, torch.tensor(torch.finfo(attn_weights.dtype).min))
+            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
+
+        # upcast to fp32 if the weights are in fp16. Please see https://github.com/huggingface/transformers/pull/17437
+        if attn_weights.dtype == torch.float16:
+            attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(torch.float16)
+        else:
+            attn_weights = nn.functional.softmax(attn_weights, dim=-1)
+
+        if layer_head_mask is not None:
+            if layer_head_mask.size() != (self.num_heads,):
+                raise ValueError(
+                    f"Head mask for a single layer should be of size {(self.num_heads,)}, but is"
+                    f" {layer_head_mask.size()}"
+                )
+            attn_weights = layer_head_mask.view(1, -1, 1, 1) * attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
+            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
+
+        if output_attentions:
+            # this operation is a bit awkward, but it's required to
+            # make sure that attn_weights keeps its gradient.
+            # In order to do so, attn_weights have to be reshaped
+            # twice and have to be reused in the following
+            attn_weights_reshaped = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
+            attn_weights = attn_weights_reshaped.view(bsz * self.num_heads, tgt_len, src_len)
+        else:
+            attn_weights_reshaped = None
+
+        attn_probs = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
+
+        attn_output = torch.bmm(attn_probs, value_states)
+
+        if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
+                f" {attn_output.size()}"
+            )
+
+        #print("boraaa self.head_dim",self.head_dim)
+        attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
+        
+        #print("attn_output bef",attn_output.size())
+        attn_output = attn_output.transpose(1, 2)
+        #print("attn_output",attn_output.size())
+        # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
+        # partitioned aross GPUs when using tensor-parallelism.
+        attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
+
+        #print("attn_output",attn_output.size())
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output, attn_weights_reshaped, past_key_value
+
+
+class ThisXGLMDecoderLayer(XGLMDecoderLayer):
+    def __init__(self, config):
+        super().__init__(config)
+
+        if config.add_cross_attention:
+            print("add cross")
+            self.encoder_attn = ThisXGLMAttention(
+                embed_dim=self.embed_dim,
+                num_heads=config.attention_heads,
+                dropout=config.attention_dropout,
+                is_decoder=True,
+                config=config,
+                is_cross_attention=True
+            )
+            self.encoder_attn_layer_norm = nn.LayerNorm(self.embed_dim)
+
+class ThisXGLMModel(XGLMModel):
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.layers = nn.ModuleList([ThisXGLMDecoderLayer(config) for _ in range(config.num_layers)])
+
+class ThisXGLMForCausalLM(XGLMForCausalLM):
+    config_class = ThisXGLMConfig
+    
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = ThisXGLMModel(config)
+