Delete modeling.py

modeling.py

@@ -1,75 +0,0 @@
-from transformers import CLIPVisionModel
-import torch
-import torch.nn as nn
-import torch.nn.functional as F
-from dataclasses import dataclass
-from rwkv.model import RWKV
-
-class UpdatableRWKV(RWKV):
- def __init__(self, *args, **kwargs):
- super().__init__(*args, **kwargs)
- 
- def update_emb_weight(self, new_value):
- self.w.update({"emb.weight", new_value}) 
-
-@dataclass
-class VisualEncoderConfig:
- n_embd: int = 2048
- vision_tower_name: str = 'openai/clip-vit-large-patch14-336'
- grid_size: int = -1 # -1: no grid pooling, 0: take cls token, 1: global avg pooling, 2, 3, 4, ...: grid pooling
-
-class VisualEncoder(nn.Module):
- def __init__(self, args):
- super().__init__()
- self.args = args
- self.vit = CLIPVisionModel.from_pretrained(args.vision_tower_name)
- self.proj = nn.Linear(self.vit.config.hidden_size, args.n_embd, bias=False)
-
- def encode_images(self, images):
- B, N, C, H, W = images.shape
- images = images.view(B*N, C, H, W)
- image_features = self.vit(images).last_hidden_state
- L, D = image_features.shape[1], image_features.shape[2]
- # rerange [B*N, L, D] -> [B, N, L, D]
- image_features = image_features.view(B, N, L, D)[:, 0, :, :]
- image_features = self.grid_pooling(image_features)
- return self.proj(image_features)
- 
- def grid_pooling(self, image_features):
- if self.args.grid_size == -1: # no grid pooling
- return image_features
- if self.args.grid_size == 0: # take cls token
- return image_features[:, 0:1, :]
- if self.args.grid_size == 1: # global avg pooling
- return image_features.mean(dim=1, keepdim=True)
- cls_features = image_features[:, 0:1, :]
- image_features = image_features[:, 1:, :] #drop cls token
- B, L, D = image_features.shape
- H_or_W = int(L**0.5)
- image_features = image_features.view(B, H_or_W, H_or_W, D)
- grid_stride = H_or_W // self.args.grid_size
- image_features = F.avg_pool2d(image_features.permute(0, 3, 1, 2), 
- padding=0,
- kernel_size=grid_stride, 
- stride=grid_stride)
- image_features = image_features.permute(0, 2, 3, 1).view(B, -1, D)
- return torch.cat((cls_features, image_features), dim=1)
-
-
-class EmbeddingMixer(nn.Module):
- def __init__(self, original_embedding, num_image_embeddings=4096):
- super().__init__()
- image_embedding = torch.zeros(num_image_embeddings, 
- original_embedding.shape[1],
- device=original_embedding.device,
- dtype=original_embedding.dtype)
- self.embedding = torch.cat((original_embedding, image_embedding), dim=0)
- self.image_start_index = len(original_embedding)
-
- def set_image_embeddings(self, image_embeddings):
- assert len(image_embeddings.shape) == 2, "image_embeddings should be 2D"
- end_index = self.image_start_index + image_embeddings.shape[0]
- self.embedding[self.image_start_index:end_index] = image_embeddings
-
- def get_input_embeddings(self):
- return self.embedding