model for audio manipulation detection

app.py

@@ -1,4 +1,8 @@
 import gradio as gr
+from manipulate_model.utils import get_config_and_model
+
+
+manpulate_config, manipulate_model = get_config_and_model()
 
 def process(filepath):
     return filepath

manipulate_model/decoder/aasist/aasist.py

@@ -0,0 +1,617 @@
+import random
+from typing import Union
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch import Tensor
+
+# import fairseq
+
+
+___author__ = "Hemlata Tak"
+__email__ = "[email protected]"
+
+############################
+## FOR fine-tuned SSL MODEL
+############################
+
+
+# class SSLModel(nn.Module):
+#     def __init__(self, device):
+#         super(SSLModel, self).__init__()
+
+#         cp_path = "xlsr2_300m.pt"  # Change the pre-trained XLSR model path.
+#         model, cfg, task = fairseq.checkpoint_utils.load_model_ensemble_and_task(
+#             [cp_path]
+#         )
+#         self.model = model[0]
+#         self.device = device
+#         self.out_dim = 1024
+#         return
+
+#     def extract_feat(self, input_data):
+
+#         # put the model to GPU if it not there
+#         if (
+#             next(self.model.parameters()).device != input_data.device
+#             or next(self.model.parameters()).dtype != input_data.dtype
+#         ):
+#             self.model.to(input_data.device, dtype=input_data.dtype)
+#             self.model.train()
+
+#         if True:
+#             # input should be in shape (batch, length)
+#             if input_data.ndim == 3:
+#                 input_tmp = input_data[:, :, 0]
+#             else:
+#                 input_tmp = input_data
+
+#             # [batch, length, dim]
+#             emb = self.model(input_tmp, mask=False, features_only=True)["x"]
+#         return emb
+
+
+# ---------AASIST back-end------------------------#
+""" Jee-weon Jung, Hee-Soo Heo, Hemlata Tak, Hye-jin Shim, Joon Son Chung, Bong-Jin Lee, Ha-Jin Yu and Nicholas Evans. 
+    AASIST: Audio Anti-Spoofing Using Integrated Spectro-Temporal Graph Attention Networks. 
+    In Proc. ICASSP 2022, pp: 6367--6371."""
+
+
+class GraphAttentionLayer(nn.Module):
+    def __init__(self, in_dim, out_dim, **kwargs):
+        super().__init__()
+
+        # attention map
+        self.att_proj = nn.Linear(in_dim, out_dim)
+        self.att_weight = self._init_new_params(out_dim, 1)
+
+        # project
+        self.proj_with_att = nn.Linear(in_dim, out_dim)
+        self.proj_without_att = nn.Linear(in_dim, out_dim)
+
+        # batch norm
+        self.bn = nn.BatchNorm1d(out_dim)
+
+        # dropout for inputs
+        self.input_drop = nn.Dropout(p=0.2)
+
+        # activate
+        self.act = nn.SELU(inplace=True)
+
+        # temperature
+        self.temp = 1.0
+        if "temperature" in kwargs:
+            self.temp = kwargs["temperature"]
+
+    def forward(self, x):
+        """
+        x   :(#bs, #node, #dim)
+        """
+        # apply input dropout
+        x = self.input_drop(x)
+
+        # derive attention map
+        att_map = self._derive_att_map(x)
+
+        # projection
+        x = self._project(x, att_map)
+
+        # apply batch norm
+        x = self._apply_BN(x)
+        x = self.act(x)
+        return x
+
+    def _pairwise_mul_nodes(self, x):
+        """
+        Calculates pairwise multiplication of nodes.
+        - for attention map
+        x           :(#bs, #node, #dim)
+        out_shape   :(#bs, #node, #node, #dim)
+        """
+
+        nb_nodes = x.size(1)
+        x = x.unsqueeze(2).expand(-1, -1, nb_nodes, -1)
+        x_mirror = x.transpose(1, 2)
+
+        return x * x_mirror
+
+    def _derive_att_map(self, x):
+        """
+        x           :(#bs, #node, #dim)
+        out_shape   :(#bs, #node, #node, 1)
+        """
+        att_map = self._pairwise_mul_nodes(x)
+        # size: (#bs, #node, #node, #dim_out)
+        att_map = torch.tanh(self.att_proj(att_map))
+        # size: (#bs, #node, #node, 1)
+        att_map = torch.matmul(att_map, self.att_weight)
+
+        # apply temperature
+        att_map = att_map / self.temp
+
+        att_map = F.softmax(att_map, dim=-2)
+
+        return att_map
+
+    def _project(self, x, att_map):
+        x1 = self.proj_with_att(torch.matmul(att_map.squeeze(-1), x))
+        x2 = self.proj_without_att(x)
+
+        return x1 + x2
+
+    def _apply_BN(self, x):
+        org_size = x.size()
+        x = x.view(-1, org_size[-1])
+        x = self.bn(x)
+        x = x.view(org_size)
+
+        return x
+
+    def _init_new_params(self, *size):
+        out = nn.Parameter(torch.FloatTensor(*size))
+        nn.init.xavier_normal_(out)
+        return out
+
+
+class HtrgGraphAttentionLayer(nn.Module):
+    def __init__(self, in_dim, out_dim, **kwargs):
+        super().__init__()
+
+        self.proj_type1 = nn.Linear(in_dim, in_dim)
+        self.proj_type2 = nn.Linear(in_dim, in_dim)
+
+        # attention map
+        self.att_proj = nn.Linear(in_dim, out_dim)
+        self.att_projM = nn.Linear(in_dim, out_dim)
+
+        self.att_weight11 = self._init_new_params(out_dim, 1)
+        self.att_weight22 = self._init_new_params(out_dim, 1)
+        self.att_weight12 = self._init_new_params(out_dim, 1)
+        self.att_weightM = self._init_new_params(out_dim, 1)
+
+        # project
+        self.proj_with_att = nn.Linear(in_dim, out_dim)
+        self.proj_without_att = nn.Linear(in_dim, out_dim)
+
+        self.proj_with_attM = nn.Linear(in_dim, out_dim)
+        self.proj_without_attM = nn.Linear(in_dim, out_dim)
+
+        # batch norm
+        self.bn = nn.BatchNorm1d(out_dim)
+
+        # dropout for inputs
+        self.input_drop = nn.Dropout(p=0.2)
+
+        # activate
+        self.act = nn.SELU(inplace=True)
+
+        # temperature
+        self.temp = 1.0
+        if "temperature" in kwargs:
+            self.temp = kwargs["temperature"]
+
+    def forward(self, x1, x2, master=None):
+        """
+        x1  :(#bs, #node, #dim)
+        x2  :(#bs, #node, #dim)
+        """
+        # print('x1',x1.shape)
+        # print('x2',x2.shape)
+        num_type1 = x1.size(1)
+        num_type2 = x2.size(1)
+        # print('num_type1',num_type1)
+        # print('num_type2',num_type2)
+        x1 = self.proj_type1(x1)
+        # print('proj_type1',x1.shape)
+        x2 = self.proj_type2(x2)
+        # print('proj_type2',x2.shape)
+        x = torch.cat([x1, x2], dim=1)
+        # print('Concat x1 and x2',x.shape)
+
+        if master is None:
+            master = torch.mean(x, dim=1, keepdim=True)
+            # print('master',master.shape)
+        # apply input dropout
+        x = self.input_drop(x)
+
+        # derive attention map
+        att_map = self._derive_att_map(x, num_type1, num_type2)
+        # print('master',master.shape)
+        # directional edge for master node
+        master = self._update_master(x, master)
+        # print('master',master.shape)
+        # projection
+        x = self._project(x, att_map)
+        # print('proj x',x.shape)
+        # apply batch norm
+        x = self._apply_BN(x)
+        x = self.act(x)
+
+        x1 = x.narrow(1, 0, num_type1)
+        # print('x1',x1.shape)
+        x2 = x.narrow(1, num_type1, num_type2)
+        # print('x2',x2.shape)
+        return x1, x2, master
+
+    def _update_master(self, x, master):
+
+        att_map = self._derive_att_map_master(x, master)
+        master = self._project_master(x, master, att_map)
+
+        return master
+
+    def _pairwise_mul_nodes(self, x):
+        """
+        Calculates pairwise multiplication of nodes.
+        - for attention map
+        x           :(#bs, #node, #dim)
+        out_shape   :(#bs, #node, #node, #dim)
+        """
+
+        nb_nodes = x.size(1)
+        x = x.unsqueeze(2).expand(-1, -1, nb_nodes, -1)
+        x_mirror = x.transpose(1, 2)
+
+        return x * x_mirror
+
+    def _derive_att_map_master(self, x, master):
+        """
+        x           :(#bs, #node, #dim)
+        out_shape   :(#bs, #node, #node, 1)
+        """
+        att_map = x * master
+        att_map = torch.tanh(self.att_projM(att_map))
+
+        att_map = torch.matmul(att_map, self.att_weightM)
+
+        # apply temperature
+        att_map = att_map / self.temp
+
+        att_map = F.softmax(att_map, dim=-2)
+
+        return att_map
+
+    def _derive_att_map(self, x, num_type1, num_type2):
+        """
+        x           :(#bs, #node, #dim)
+        out_shape   :(#bs, #node, #node, 1)
+        """
+        att_map = self._pairwise_mul_nodes(x)
+        # size: (#bs, #node, #node, #dim_out)
+        att_map = torch.tanh(self.att_proj(att_map))
+        # size: (#bs, #node, #node, 1)
+
+        att_board = torch.zeros_like(att_map[:, :, :, 0]).unsqueeze(-1)
+
+        att_board[:, :num_type1, :num_type1, :] = torch.matmul(
+            att_map[:, :num_type1, :num_type1, :], self.att_weight11
+        )
+        att_board[:, num_type1:, num_type1:, :] = torch.matmul(
+            att_map[:, num_type1:, num_type1:, :], self.att_weight22
+        )
+        att_board[:, :num_type1, num_type1:, :] = torch.matmul(
+            att_map[:, :num_type1, num_type1:, :], self.att_weight12
+        )
+        att_board[:, num_type1:, :num_type1, :] = torch.matmul(
+            att_map[:, num_type1:, :num_type1, :], self.att_weight12
+        )
+
+        att_map = att_board
+
+        # apply temperature
+        att_map = att_map / self.temp
+
+        att_map = F.softmax(att_map, dim=-2)
+
+        return att_map
+
+    def _project(self, x, att_map):
+        x1 = self.proj_with_att(torch.matmul(att_map.squeeze(-1), x))
+        x2 = self.proj_without_att(x)
+
+        return x1 + x2
+
+    def _project_master(self, x, master, att_map):
+
+        x1 = self.proj_with_attM(torch.matmul(att_map.squeeze(-1).unsqueeze(1), x))
+        x2 = self.proj_without_attM(master)
+
+        return x1 + x2
+
+    def _apply_BN(self, x):
+        org_size = x.size()
+        x = x.view(-1, org_size[-1])
+        x = self.bn(x)
+        x = x.view(org_size)
+
+        return x
+
+    def _init_new_params(self, *size):
+        out = nn.Parameter(torch.FloatTensor(*size))
+        nn.init.xavier_normal_(out)
+        return out
+
+
+class GraphPool(nn.Module):
+    def __init__(self, k: float, in_dim: int, p: Union[float, int]):
+        super().__init__()
+        self.k = k
+        self.sigmoid = nn.Sigmoid()
+        self.proj = nn.Linear(in_dim, 1)
+        self.drop = nn.Dropout(p=p) if p > 0 else nn.Identity()
+        self.in_dim = in_dim
+
+    def forward(self, h):
+        Z = self.drop(h)
+        weights = self.proj(Z)
+        scores = self.sigmoid(weights)
+        new_h = self.top_k_graph(scores, h, self.k)
+
+        return new_h
+
+    def top_k_graph(self, scores, h, k):
+        """
+        args
+        =====
+        scores: attention-based weights (#bs, #node, 1)
+        h: graph data (#bs, #node, #dim)
+        k: ratio of remaining nodes, (float)
+        returns
+        =====
+        h: graph pool applied data (#bs, #node', #dim)
+        """
+        _, n_nodes, n_feat = h.size()
+        n_nodes = max(int(n_nodes * k), 1)
+        _, idx = torch.topk(scores, n_nodes, dim=1)
+        idx = idx.expand(-1, -1, n_feat)
+
+        h = h * scores
+        h = torch.gather(h, 1, idx)
+
+        return h
+
+
+class Residual_block(nn.Module):
+    def __init__(self, nb_filts, first=False):
+        super().__init__()
+        self.first = first
+
+        if not self.first:
+            self.bn1 = nn.BatchNorm2d(num_features=nb_filts[0])
+        self.conv1 = nn.Conv2d(
+            in_channels=nb_filts[0],
+            out_channels=nb_filts[1],
+            kernel_size=(2, 3),
+            padding=(1, 1),
+            stride=1,
+        )
+        self.selu = nn.SELU(inplace=True)
+
+        self.bn2 = nn.BatchNorm2d(num_features=nb_filts[1])
+        self.conv2 = nn.Conv2d(
+            in_channels=nb_filts[1],
+            out_channels=nb_filts[1],
+            kernel_size=(2, 3),
+            padding=(0, 1),
+            stride=1,
+        )
+
+        if nb_filts[0] != nb_filts[1]:
+            self.downsample = True
+            self.conv_downsample = nn.Conv2d(
+                in_channels=nb_filts[0],
+                out_channels=nb_filts[1],
+                padding=(0, 1),
+                kernel_size=(1, 3),
+                stride=1,
+            )
+
+        else:
+            self.downsample = False
+
+    def forward(self, x):
+        identity = x
+        if not self.first:
+            out = self.bn1(x)
+            out = self.selu(out)
+        else:
+            out = x
+
+        # print('out',out.shape)
+        out = self.conv1(x)
+
+        # print('aft conv1 out',out.shape)
+        out = self.bn2(out)
+        out = self.selu(out)
+        # print('out',out.shape)
+        out = self.conv2(out)
+        # print('conv2 out',out.shape)
+
+        if self.downsample:
+            identity = self.conv_downsample(identity)
+
+        out += identity
+        # out = self.mp(out)
+        return out
+
+
+class AASIST(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+
+        # AASIST parameters
+        filts = [128, [1, 32], [32, 32], [32, 64], [64, 64]]
+        gat_dims = [64, 32]
+        pool_ratios = [0.5, 0.5, 0.5, 0.5]
+        temperatures = [2.0, 2.0, 100.0, 100.0]
+
+        ####
+        # create network wav2vec 2.0
+        ####
+        # self.ssl_model = SSLModel(self.device)
+        self.LL = nn.Linear(self.config.model.decoder.encoding_dim, 128)
+
+        self.first_bn = nn.BatchNorm2d(num_features=1)
+        self.first_bn1 = nn.BatchNorm2d(num_features=64)
+        self.drop = nn.Dropout(0.5, inplace=True)
+        self.drop_way = nn.Dropout(0.2, inplace=True)
+        self.selu = nn.SELU(inplace=True)
+
+        # RawNet2 encoder
+        self.encoder = nn.Sequential(
+            nn.Sequential(Residual_block(nb_filts=filts[1], first=True)),
+            nn.Sequential(Residual_block(nb_filts=filts[2])),
+            nn.Sequential(Residual_block(nb_filts=filts[3])),
+            nn.Sequential(Residual_block(nb_filts=filts[4])),
+            nn.Sequential(Residual_block(nb_filts=filts[4])),
+            nn.Sequential(Residual_block(nb_filts=filts[4])),
+        )
+
+        self.attention = nn.Sequential(
+            nn.Conv2d(64, 128, kernel_size=(1, 1)),
+            nn.SELU(inplace=True),
+            nn.BatchNorm2d(128),
+            nn.Conv2d(128, 64, kernel_size=(1, 1)),
+        )
+        # position encoding
+        self.pos_S = nn.Parameter(torch.randn(1, 42, filts[-1][-1]))
+
+        self.master1 = nn.Parameter(torch.randn(1, 1, gat_dims[0]))
+        self.master2 = nn.Parameter(torch.randn(1, 1, gat_dims[0]))
+
+        # Graph module
+        self.GAT_layer_S = GraphAttentionLayer(
+            filts[-1][-1], gat_dims[0], temperature=temperatures[0]
+        )
+        self.GAT_layer_T = GraphAttentionLayer(
+            filts[-1][-1], gat_dims[0], temperature=temperatures[1]
+        )
+        # HS-GAL layer
+        self.HtrgGAT_layer_ST11 = HtrgGraphAttentionLayer(
+            gat_dims[0], gat_dims[1], temperature=temperatures[2]
+        )
+        self.HtrgGAT_layer_ST12 = HtrgGraphAttentionLayer(
+            gat_dims[1], gat_dims[1], temperature=temperatures[2]
+        )
+        self.HtrgGAT_layer_ST21 = HtrgGraphAttentionLayer(
+            gat_dims[0], gat_dims[1], temperature=temperatures[2]
+        )
+        self.HtrgGAT_layer_ST22 = HtrgGraphAttentionLayer(
+            gat_dims[1], gat_dims[1], temperature=temperatures[2]
+        )
+
+        # Graph pooling layers
+        self.pool_S = GraphPool(pool_ratios[0], gat_dims[0], 0.3)
+        self.pool_T = GraphPool(pool_ratios[1], gat_dims[0], 0.3)
+        self.pool_hS1 = GraphPool(pool_ratios[2], gat_dims[1], 0.3)
+        self.pool_hT1 = GraphPool(pool_ratios[2], gat_dims[1], 0.3)
+
+        self.pool_hS2 = GraphPool(pool_ratios[2], gat_dims[1], 0.3)
+        self.pool_hT2 = GraphPool(pool_ratios[2], gat_dims[1], 0.3)
+
+        if self.config.model.task == "audio-video":
+            self.out_layer = nn.Linear(5 * gat_dims[1], 4)
+        else:
+            self.out_layer = nn.Linear(5 * gat_dims[1], 2)
+
+    def forward(self, x):
+        # -------pre-trained Wav2vec model fine tunning ------------------------##
+        # x_ssl_feat = self.ssl_model.extract_feat(x.squeeze(-1))
+        x = self.LL(x)  # (bs,frame_number,feat_out_dim)
+
+        # post-processing on front-end features
+        x = x.transpose(1, 2)  # (bs,feat_out_dim,frame_number)
+        x = x.unsqueeze(dim=1)  # add channel
+        x = F.max_pool2d(x, (3, 3))
+        x = self.first_bn(x)
+        x = self.selu(x)
+
+        # RawNet2-based encoder
+        x = self.encoder(x)
+        x = self.first_bn1(x)
+        x = self.selu(x)
+
+        w = self.attention(x)
+
+        # ------------SA for spectral feature-------------#
+        w1 = F.softmax(w, dim=-1)
+        m = torch.sum(x * w1, dim=-1)
+        e_S = m.transpose(1, 2) + self.pos_S
+
+        # graph module layer
+        gat_S = self.GAT_layer_S(e_S)
+        out_S = self.pool_S(gat_S)  # (#bs, #node, #dim)
+
+        # ------------SA for temporal feature-------------#
+        w2 = F.softmax(w, dim=-2)
+        m1 = torch.sum(x * w2, dim=-2)
+
+        e_T = m1.transpose(1, 2)
+
+        # graph module layer
+        gat_T = self.GAT_layer_T(e_T)
+        out_T = self.pool_T(gat_T)
+
+        # learnable master node
+        master1 = self.master1.expand(x.size(0), -1, -1)
+        master2 = self.master2.expand(x.size(0), -1, -1)
+
+        # inference 1
+        out_T1, out_S1, master1 = self.HtrgGAT_layer_ST11(
+            out_T, out_S, master=self.master1
+        )
+
+        out_S1 = self.pool_hS1(out_S1)
+        out_T1 = self.pool_hT1(out_T1)
+
+        out_T_aug, out_S_aug, master_aug = self.HtrgGAT_layer_ST12(
+            out_T1, out_S1, master=master1
+        )
+        out_T1 = out_T1 + out_T_aug
+        out_S1 = out_S1 + out_S_aug
+        master1 = master1 + master_aug
+
+        # inference 2
+        out_T2, out_S2, master2 = self.HtrgGAT_layer_ST21(
+            out_T, out_S, master=self.master2
+        )
+        out_S2 = self.pool_hS2(out_S2)
+        out_T2 = self.pool_hT2(out_T2)
+
+        out_T_aug, out_S_aug, master_aug = self.HtrgGAT_layer_ST22(
+            out_T2, out_S2, master=master2
+        )
+        out_T2 = out_T2 + out_T_aug
+        out_S2 = out_S2 + out_S_aug
+        master2 = master2 + master_aug
+
+        out_T1 = self.drop_way(out_T1)
+        out_T2 = self.drop_way(out_T2)
+        out_S1 = self.drop_way(out_S1)
+        out_S2 = self.drop_way(out_S2)
+        master1 = self.drop_way(master1)
+        master2 = self.drop_way(master2)
+
+        out_T = torch.max(out_T1, out_T2)
+        out_S = torch.max(out_S1, out_S2)
+        master = torch.max(master1, master2)
+
+        # Readout operation
+        T_max, _ = torch.max(torch.abs(out_T), dim=1)
+        T_avg = torch.mean(out_T, dim=1)
+
+        S_max, _ = torch.max(torch.abs(out_S), dim=1)
+        S_avg = torch.mean(out_S, dim=1)
+
+        last_hidden = torch.cat([T_max, T_avg, S_max, S_avg, master.squeeze(1)], dim=1)
+
+        last_hidden = self.drop(last_hidden)
+        output = self.out_layer(last_hidden)
+
+        if self.config.model.task == "audio-video":
+            output = output.view(-1, 2, 2)
+
+        return output

manipulate_model/decoder/decoder.py

@@ -0,0 +1,20 @@
+import torch
+import torch.nn as nn
+
+
+class Decoder(nn.Module):
+    def __init__(self, config):
+        super(Decoder, self).__init__()
+        self.config = config
+
+        self.decoder = None
+
+        if config.model.decoder.name.lower() == "aasist":
+            from manipulate_model.decoder.aasist.aasist import AASIST
+
+            self.decoder = AASIST(config)
+        else:
+            raise ValueError("Invalid decoder name")
+
+    def forward(self, x):
+        return self.decoder(x)

manipulate_model/demo-model/audio/config.yaml

@@ -0,0 +1,44 @@
+model:
+  task: audio
+  encoder:
+    name: wavlm
+    version: base
+    pretrained: true
+    pretrained_path: manipulate_model/encoder_checkpoints/wavlm/WavLM-Base+.pt
+    output_layer: 3
+  encoder_freeze: false
+  decoder:
+    name: aasist
+    version: default
+    output_size: 2
+  online_encoding: true
+data:
+  name: av1m
+  train_parts: all
+  val_parts: all
+  test_parts: all
+  train_size: -1
+  val_size: -1
+  test_size: -1
+  shape:
+  - 3
+  - 224
+  - 224
+  sr: 16000
+  fps: 25
+  center_transition: true
+  window_size: 4
+  sliding_window: false
+train:
+  num_workers: 16
+  batch_size: 64
+  num_epochs: 15
+  optimizer: adam
+  scheduler: step
+  lr: 0.0001
+  step_size: 1
+  gamma: 0.1
+  loss: bce
+  log_interval: 100
+  shuffle: true
+debug: false

manipulate_model/encoder/encoder.py

@@ -0,0 +1,40 @@
+import torch
+import torch.nn as nn
+
+from torch.nn.functional import pad
+from collections import OrderedDict
+
+
+class Encoder(nn.Module):
+    def __init__(self, config):
+        super(Encoder, self).__init__()
+        self.config = config
+
+        self.encoder = None
+        self.succeeding_layers = None
+
+        # AUDIO
+        if self.config.model.task == "audio":
+            if self.config.model.encoder.name.lower() == "wavlm":
+                from manipulate_model.encoder.wavlm.WavLM import WavLM, WavLMConfig
+
+                ckpt = torch.load(
+                    config.model.encoder.pretrained_path, map_location="cpu"
+                )
+                cfg = WavLMConfig(ckpt["cfg"])
+                self.encoder = WavLM(cfg)
+
+
+    def forward(self, x):
+        if self.config.model.encoder.name.lower() == "wavlm":
+            return self.encoder(x, output_layer=self.config.model.encoder.output_layer)
+        elif self.config.model.encoder.name.lower() == "videomamba":
+            return self.encoder(x)
+
+        return self.encoder(x)
+
+    def get_encoding_dim(self):
+        return self.encoder.get_encoding_dim()
+
+    def get_temporal_dim(self):
+        return self.encoder.get_temporal_dim(window_size=self.config.data.window_size)

manipulate_model/encoder/wavlm/WavLM.py

@@ -0,0 +1,795 @@
+# --------------------------------------------------------
+# WavLM: Large-Scale Self-Supervised  Pre-training  for Full Stack Speech Processing (https://arxiv.org/abs/2110.13900.pdf)
+# Github source: https://github.com/microsoft/unilm/tree/master/wavlm
+# Copyright (c) 2021 Microsoft
+# Licensed under The MIT License [see LICENSE for details]
+# Based on fairseq code bases
+# https://github.com/pytorch/fairseq
+# --------------------------------------------------------
+
+import math
+import logging
+from typing import List, Optional, Tuple
+
+import numpy as np
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.nn import LayerNorm
+from manipulate_model.encoder.wavlm.modules import (
+    Fp32GroupNorm,
+    Fp32LayerNorm,
+    GradMultiply,
+    MultiheadAttention,
+    SamePad,
+    init_bert_params,
+    get_activation_fn,
+    TransposeLast,
+    GLU_Linear,
+)
+
+logger = logging.getLogger(__name__)
+
+
+def compute_mask_indices(
+    shape: Tuple[int, int],
+    padding_mask: Optional[torch.Tensor],
+    mask_prob: float,
+    mask_length: int,
+    mask_type: str = "static",
+    mask_other: float = 0.0,
+    min_masks: int = 0,
+    no_overlap: bool = False,
+    min_space: int = 0,
+) -> np.ndarray:
+    """
+    Computes random mask spans for a given shape
+
+    Args:
+        shape: the the shape for which to compute masks.
+            should be of size 2 where first element is batch size and 2nd is timesteps
+        padding_mask: optional padding mask of the same size as shape, which will prevent masking padded elements
+        mask_prob: probability for each token to be chosen as start of the span to be masked. this will be multiplied by
+            number of timesteps divided by length of mask span to mask approximately this percentage of all elements.
+            however due to overlaps, the actual number will be smaller (unless no_overlap is True)
+        mask_type: how to compute mask lengths
+            static = fixed size
+            uniform = sample from uniform distribution [mask_other, mask_length*2]
+            normal = sample from normal distribution with mean mask_length and stdev mask_other. mask is min 1 element
+            poisson = sample from possion distribution with lambda = mask length
+        min_masks: minimum number of masked spans
+        no_overlap: if false, will switch to an alternative recursive algorithm that prevents spans from overlapping
+        min_space: only used if no_overlap is True, this is how many elements to keep unmasked between spans
+    """
+
+    bsz, all_sz = shape
+    mask = np.full((bsz, all_sz), False)
+
+    all_num_mask = int(
+        # add a random number for probabilistic rounding
+        mask_prob * all_sz / float(mask_length)
+        + np.random.rand()
+    )
+
+    all_num_mask = max(min_masks, all_num_mask)
+
+    mask_idcs = []
+    for i in range(bsz):
+        if padding_mask is not None:
+            sz = all_sz - padding_mask[i].long().sum().item()
+            num_mask = int(
+                # add a random number for probabilistic rounding
+                mask_prob * sz / float(mask_length)
+                + np.random.rand()
+            )
+            num_mask = max(min_masks, num_mask)
+        else:
+            sz = all_sz
+            num_mask = all_num_mask
+
+        if mask_type == "static":
+            lengths = np.full(num_mask, mask_length)
+        elif mask_type == "uniform":
+            lengths = np.random.randint(mask_other, mask_length * 2 + 1, size=num_mask)
+        elif mask_type == "normal":
+            lengths = np.random.normal(mask_length, mask_other, size=num_mask)
+            lengths = [max(1, int(round(x))) for x in lengths]
+        elif mask_type == "poisson":
+            lengths = np.random.poisson(mask_length, size=num_mask)
+            lengths = [int(round(x)) for x in lengths]
+        else:
+            raise Exception("unknown mask selection " + mask_type)
+
+        if sum(lengths) == 0:
+            lengths[0] = min(mask_length, sz - 1)
+
+        if no_overlap:
+            mask_idc = []
+
+            def arrange(s, e, length, keep_length):
+                span_start = np.random.randint(s, e - length)
+                mask_idc.extend(span_start + i for i in range(length))
+
+                new_parts = []
+                if span_start - s - min_space >= keep_length:
+                    new_parts.append((s, span_start - min_space + 1))
+                if e - span_start - keep_length - min_space > keep_length:
+                    new_parts.append((span_start + length + min_space, e))
+                return new_parts
+
+            parts = [(0, sz)]
+            min_length = min(lengths)
+            for length in sorted(lengths, reverse=True):
+                lens = np.fromiter(
+                    (e - s if e - s >= length + min_space else 0 for s, e in parts),
+                    np.int,
+                )
+                l_sum = np.sum(lens)
+                if l_sum == 0:
+                    break
+                probs = lens / np.sum(lens)
+                c = np.random.choice(len(parts), p=probs)
+                s, e = parts.pop(c)
+                parts.extend(arrange(s, e, length, min_length))
+            mask_idc = np.asarray(mask_idc)
+        else:
+            min_len = min(lengths)
+            if sz - min_len <= num_mask:
+                min_len = sz - num_mask - 1
+
+            mask_idc = np.random.choice(sz - min_len, num_mask, replace=False)
+
+            mask_idc = np.asarray(
+                [
+                    mask_idc[j] + offset
+                    for j in range(len(mask_idc))
+                    for offset in range(lengths[j])
+                ]
+            )
+
+        mask_idcs.append(np.unique(mask_idc[mask_idc < sz]))
+
+    min_len = min([len(m) for m in mask_idcs])
+    for i, mask_idc in enumerate(mask_idcs):
+        if len(mask_idc) > min_len:
+            mask_idc = np.random.choice(mask_idc, min_len, replace=False)
+        mask[i, mask_idc] = True
+
+    return mask
+
+
+class WavLMConfig:
+    def __init__(self, cfg=None):
+        self.extractor_mode: str = (
+            "default"  # mode for feature extractor. default has a single group norm with d groups in the first conv block, whereas layer_norm has layer norms in every block (meant to use with normalize=True)
+        )
+        self.encoder_layers: int = 12  # num encoder layers in the transformer
+
+        self.encoder_embed_dim: int = 768  # encoder embedding dimension
+        self.encoder_ffn_embed_dim: int = 3072  # encoder embedding dimension for FFN
+        self.encoder_attention_heads: int = 12  # num encoder attention heads
+        self.activation_fn: str = "gelu"  # activation function to use
+
+        self.layer_norm_first: bool = False  # apply layernorm first in the transformer
+        self.conv_feature_layers: str = (
+            "[(512,10,5)] + [(512,3,2)] * 4 + [(512,2,2)] * 2"  # string describing convolutional feature extraction layers in form of a python list that contains [(dim, kernel_size, stride), ...]
+        )
+        self.conv_bias: bool = False  # include bias in conv encoder
+        self.feature_grad_mult: float = (
+            1.0  # multiply feature extractor var grads by this
+        )
+
+        self.normalize: bool = (
+            False  # normalize input to have 0 mean and unit variance during training
+        )
+
+        # dropouts
+        self.dropout: float = 0.1  # dropout probability for the transformer
+        self.attention_dropout: float = 0.1  # dropout probability for attention weights
+        self.activation_dropout: float = (
+            0.0  # dropout probability after activation in FFN
+        )
+        self.encoder_layerdrop: float = (
+            0.0  # probability of dropping a tarnsformer layer
+        )
+        self.dropout_input: float = (
+            0.0  # dropout to apply to the input (after feat extr)
+        )
+        self.dropout_features: float = (
+            0.0  # dropout to apply to the features (after feat extr)
+        )
+
+        # masking
+        self.mask_length: int = 10  # mask length
+        self.mask_prob: float = 0.65  # probability of replacing a token with mask
+        self.mask_selection: str = "static"  # how to choose mask length
+        self.mask_other: float = (
+            0  # secondary mask argument (used for more complex distributions), see help in compute_mask_indicesh
+        )
+        self.no_mask_overlap: bool = False  # whether to allow masks to overlap
+        self.mask_min_space: int = (
+            1  # min space between spans (if no overlap is enabled)
+        )
+
+        # channel masking
+        self.mask_channel_length: int = 10  # length of the mask for features (channels)
+        self.mask_channel_prob: float = 0.0  # probability of replacing a feature with 0
+        self.mask_channel_selection: str = (
+            "static"  # how to choose mask length for channel masking
+        )
+        self.mask_channel_other: float = (
+            0  # secondary mask argument (used for more complex distributions), see help in compute_mask_indices
+        )
+        self.no_mask_channel_overlap: bool = (
+            False  # whether to allow channel masks to overlap
+        )
+        self.mask_channel_min_space: int = (
+            1  # min space between spans (if no overlap is enabled)
+        )
+
+        # positional embeddings
+        self.conv_pos: int = (
+            128  # number of filters for convolutional positional embeddings
+        )
+        self.conv_pos_groups: int = (
+            16  # number of groups for convolutional positional embedding
+        )
+
+        # relative position embedding
+        self.relative_position_embedding: bool = (
+            False  # apply relative position embedding
+        )
+        self.num_buckets: int = 320  # number of buckets for relative position embedding
+        self.max_distance: int = (
+            1280  # maximum distance for relative position embedding
+        )
+        self.gru_rel_pos: bool = False  # apply gated relative position embedding
+
+        if cfg is not None:
+            self.update(cfg)
+
+    def update(self, cfg: dict):
+        self.__dict__.update(cfg)
+
+
+class WavLM(nn.Module):
+    def __init__(
+        self,
+        cfg: WavLMConfig,
+    ) -> None:
+        super().__init__()
+        logger.info(f"WavLM Config: {cfg.__dict__}")
+
+        self.cfg = cfg
+        feature_enc_layers = eval(cfg.conv_feature_layers)
+        self.embed = feature_enc_layers[-1][0]
+
+        self.feature_extractor = ConvFeatureExtractionModel(
+            conv_layers=feature_enc_layers,
+            dropout=0.0,
+            mode=cfg.extractor_mode,
+            conv_bias=cfg.conv_bias,
+        )
+
+        self.post_extract_proj = (
+            nn.Linear(self.embed, cfg.encoder_embed_dim)
+            if self.embed != cfg.encoder_embed_dim
+            else None
+        )
+
+        self.mask_prob = cfg.mask_prob
+        self.mask_selection = cfg.mask_selection
+        self.mask_other = cfg.mask_other
+        self.mask_length = cfg.mask_length
+        self.no_mask_overlap = cfg.no_mask_overlap
+        self.mask_min_space = cfg.mask_min_space
+
+        self.mask_channel_prob = cfg.mask_channel_prob
+        self.mask_channel_selection = cfg.mask_channel_selection
+        self.mask_channel_other = cfg.mask_channel_other
+        self.mask_channel_length = cfg.mask_channel_length
+        self.no_mask_channel_overlap = cfg.no_mask_channel_overlap
+        self.mask_channel_min_space = cfg.mask_channel_min_space
+
+        self.dropout_input = nn.Dropout(cfg.dropout_input)
+        self.dropout_features = nn.Dropout(cfg.dropout_features)
+
+        self.feature_grad_mult = cfg.feature_grad_mult
+
+        self.mask_emb = nn.Parameter(
+            torch.FloatTensor(cfg.encoder_embed_dim).uniform_()
+        )
+
+        self.encoder = TransformerEncoder(cfg)
+        self.layer_norm = LayerNorm(self.embed)
+
+    def apply_mask(self, x, padding_mask):
+        B, T, C = x.shape
+        if self.mask_prob > 0:
+            mask_indices = compute_mask_indices(
+                (B, T),
+                padding_mask,
+                self.mask_prob,
+                self.mask_length,
+                self.mask_selection,
+                self.mask_other,
+                min_masks=2,
+                no_overlap=self.no_mask_overlap,
+                min_space=self.mask_min_space,
+            )
+            mask_indices = torch.from_numpy(mask_indices).to(x.device)
+            x[mask_indices] = self.mask_emb
+        else:
+            mask_indices = None
+
+        if self.mask_channel_prob > 0:
+            mask_channel_indices = compute_mask_indices(
+                (B, C),
+                None,
+                self.mask_channel_prob,
+                self.mask_channel_length,
+                self.mask_channel_selection,
+                self.mask_channel_other,
+                no_overlap=self.no_mask_channel_overlap,
+                min_space=self.mask_channel_min_space,
+            )
+            mask_channel_indices = (
+                torch.from_numpy(mask_channel_indices)
+                .to(x.device)
+                .unsqueeze(1)
+                .expand(-1, T, -1)
+            )
+            x[mask_channel_indices] = 0
+
+        return x, mask_indices
+
+    def forward_padding_mask(
+        self,
+        features: torch.Tensor,
+        padding_mask: torch.Tensor,
+    ) -> torch.Tensor:
+        extra = padding_mask.size(1) % features.size(1)
+        if extra > 0:
+            padding_mask = padding_mask[:, :-extra]
+        padding_mask = padding_mask.view(padding_mask.size(0), features.size(1), -1)
+        padding_mask = padding_mask.all(-1)
+        return padding_mask
+
+    def extract_features(
+        self,
+        source: torch.Tensor,
+        padding_mask: Optional[torch.Tensor] = None,
+        mask: bool = False,
+        ret_conv: bool = False,
+        output_layer: Optional[int] = None,
+        ret_layer_results: bool = False,
+    ):
+
+        if self.feature_grad_mult > 0:
+            features = self.feature_extractor(source)
+            if self.feature_grad_mult != 1.0:
+                features = GradMultiply.apply(features, self.feature_grad_mult)
+        else:
+            with torch.no_grad():
+                features = self.feature_extractor(source)
+
+        features = features.transpose(1, 2)
+        features = self.layer_norm(features)
+
+        if padding_mask is not None:
+            padding_mask = self.forward_padding_mask(features, padding_mask)
+
+        if self.post_extract_proj is not None:
+            features = self.post_extract_proj(features)
+
+        features = self.dropout_input(features)
+
+        if mask:
+            x, mask_indices = self.apply_mask(features, padding_mask)
+        else:
+            x = features
+
+        # feature: (B, T, D), float
+        # target: (B, T), long
+        # x: (B, T, D), float
+        # padding_mask: (B, T), bool
+        # mask_indices: (B, T), bool
+        x, layer_results = self.encoder(
+            x,
+            padding_mask=padding_mask,
+            layer=None if output_layer is None else output_layer - 1,
+        )
+
+        res = {
+            "x": x,
+            "padding_mask": padding_mask,
+            "features": features,
+            "layer_results": layer_results,
+        }
+
+        feature = res["features"] if ret_conv else res["x"]
+        if ret_layer_results:
+            feature = (feature, res["layer_results"])
+        return feature, res["padding_mask"]
+
+    def forward(self, x, output_layer=None):
+        return self.extract_features(x, output_layer=output_layer)[0]
+
+    def get_encoding_dim(self):
+        return self.cfg.encoder_embed_dim
+
+    def get_temporal_dim(self, window_size):
+        return 2 * window_size - 1
+
+
+class ConvFeatureExtractionModel(nn.Module):
+    def __init__(
+        self,
+        conv_layers: List[Tuple[int, int, int]],
+        dropout: float = 0.0,
+        mode: str = "default",
+        conv_bias: bool = False,
+        conv_type: str = "default",
+    ):
+        super().__init__()
+
+        assert mode in {"default", "layer_norm"}
+
+        def block(
+            n_in,
+            n_out,
+            k,
+            stride,
+            is_layer_norm=False,
+            is_group_norm=False,
+            conv_bias=False,
+        ):
+            def make_conv():
+                conv = nn.Conv1d(n_in, n_out, k, stride=stride, bias=conv_bias)
+                nn.init.kaiming_normal_(conv.weight)
+                return conv
+
+            assert (
+                is_layer_norm and is_group_norm
+            ) == False, "layer norm and group norm are exclusive"
+
+            if is_layer_norm:
+                return nn.Sequential(
+                    make_conv(),
+                    nn.Dropout(p=dropout),
+                    nn.Sequential(
+                        TransposeLast(),
+                        Fp32LayerNorm(dim, elementwise_affine=True),
+                        TransposeLast(),
+                    ),
+                    nn.GELU(),
+                )
+            elif is_group_norm:
+                return nn.Sequential(
+                    make_conv(),
+                    nn.Dropout(p=dropout),
+                    Fp32GroupNorm(dim, dim, affine=True),
+                    nn.GELU(),
+                )
+            else:
+                return nn.Sequential(make_conv(), nn.Dropout(p=dropout), nn.GELU())
+
+        self.conv_type = conv_type
+        if self.conv_type == "default":
+            in_d = 1
+            self.conv_layers = nn.ModuleList()
+            for i, cl in enumerate(conv_layers):
+                assert len(cl) == 3, "invalid conv definition: " + str(cl)
+                (dim, k, stride) = cl
+
+                self.conv_layers.append(
+                    block(
+                        in_d,
+                        dim,
+                        k,
+                        stride,
+                        is_layer_norm=mode == "layer_norm",
+                        is_group_norm=mode == "default" and i == 0,
+                        conv_bias=conv_bias,
+                    )
+                )
+                in_d = dim
+        elif self.conv_type == "conv2d":
+            in_d = 1
+            self.conv_layers = nn.ModuleList()
+            for i, cl in enumerate(conv_layers):
+                assert len(cl) == 3
+                (dim, k, stride) = cl
+
+                self.conv_layers.append(torch.nn.Conv2d(in_d, dim, k, stride))
+                self.conv_layers.append(torch.nn.ReLU())
+                in_d = dim
+        elif self.conv_type == "custom":
+            in_d = 1
+            idim = 80
+            self.conv_layers = nn.ModuleList()
+            for i, cl in enumerate(conv_layers):
+                assert len(cl) == 3
+                (dim, k, stride) = cl
+                self.conv_layers.append(
+                    torch.nn.Conv2d(in_d, dim, k, stride, padding=1)
+                )
+                self.conv_layers.append(torch.nn.LayerNorm([dim, idim]))
+                self.conv_layers.append(torch.nn.ReLU())
+                in_d = dim
+                if (i + 1) % 2 == 0:
+                    self.conv_layers.append(
+                        torch.nn.MaxPool2d(2, stride=2, ceil_mode=True)
+                    )
+                    idim = int(math.ceil(idim / 2))
+        else:
+            pass
+
+    def forward(self, x, mask=None):
+
+        # BxT -> BxCxT
+        x = x.unsqueeze(1)
+        if self.conv_type == "custom":
+            for conv in self.conv_layers:
+                if isinstance(conv, nn.LayerNorm):
+                    x = x.transpose(1, 2)
+                    x = conv(x).transpose(1, 2)
+                else:
+                    x = conv(x)
+            x = x.transpose(2, 3).contiguous()
+            x = x.view(x.size(0), -1, x.size(-1))
+        else:
+            for conv in self.conv_layers:
+                x = conv(x)
+            if self.conv_type == "conv2d":
+                b, c, t, f = x.size()
+                x = x.transpose(2, 3).contiguous().view(b, c * f, t)
+        return x
+
+
+class TransformerEncoder(nn.Module):
+    def __init__(self, args):
+        super().__init__()
+
+        self.dropout = args.dropout
+        self.embedding_dim = args.encoder_embed_dim
+
+        self.pos_conv = nn.Conv1d(
+            self.embedding_dim,
+            self.embedding_dim,
+            kernel_size=args.conv_pos,
+            padding=args.conv_pos // 2,
+            groups=args.conv_pos_groups,
+        )
+        dropout = 0
+        std = math.sqrt((4 * (1.0 - dropout)) / (args.conv_pos * self.embedding_dim))
+        nn.init.normal_(self.pos_conv.weight, mean=0, std=std)
+        nn.init.constant_(self.pos_conv.bias, 0)
+
+        self.pos_conv = nn.utils.weight_norm(self.pos_conv, name="weight", dim=2)
+        self.pos_conv = nn.Sequential(self.pos_conv, SamePad(args.conv_pos), nn.GELU())
+
+        if hasattr(args, "relative_position_embedding"):
+            self.relative_position_embedding = args.relative_position_embedding
+            self.num_buckets = args.num_buckets
+            self.max_distance = args.max_distance
+        else:
+            self.relative_position_embedding = False
+            self.num_buckets = 0
+            self.max_distance = 0
+
+        self.layers = nn.ModuleList(
+            [
+                TransformerSentenceEncoderLayer(
+                    embedding_dim=self.embedding_dim,
+                    ffn_embedding_dim=args.encoder_ffn_embed_dim,
+                    num_attention_heads=args.encoder_attention_heads,
+                    dropout=self.dropout,
+                    attention_dropout=args.attention_dropout,
+                    activation_dropout=args.activation_dropout,
+                    activation_fn=args.activation_fn,
+                    layer_norm_first=args.layer_norm_first,
+                    has_relative_attention_bias=(
+                        self.relative_position_embedding and i == 0
+                    ),
+                    num_buckets=self.num_buckets,
+                    max_distance=self.max_distance,
+                    gru_rel_pos=args.gru_rel_pos,
+                )
+                for i in range(args.encoder_layers)
+            ]
+        )
+
+        self.layer_norm_first = args.layer_norm_first
+        self.layer_norm = LayerNorm(self.embedding_dim)
+        self.layerdrop = args.encoder_layerdrop
+
+        self.apply(init_bert_params)
+
+    def forward(self, x, padding_mask=None, streaming_mask=None, layer=None):
+        x, layer_results = self.extract_features(x, padding_mask, streaming_mask, layer)
+
+        if self.layer_norm_first and layer is None:
+            x = self.layer_norm(x)
+
+        return x, layer_results
+
+    def extract_features(
+        self, x, padding_mask=None, streaming_mask=None, tgt_layer=None
+    ):
+
+        if padding_mask is not None:
+            x[padding_mask] = 0
+
+        x_conv = self.pos_conv(x.transpose(1, 2))
+        x_conv = x_conv.transpose(1, 2)
+        x = x + x_conv
+
+        if not self.layer_norm_first:
+            x = self.layer_norm(x)
+
+        x = F.dropout(x, p=self.dropout, training=self.training)
+
+        # B x T x C -> T x B x C
+        x = x.transpose(0, 1)
+
+        layer_results = []
+        z = None
+        if tgt_layer is not None:
+            layer_results.append((x, z))
+        r = None
+        pos_bias = None
+        for i, layer in enumerate(self.layers):
+            dropout_probability = np.random.random()
+            if not self.training or (dropout_probability > self.layerdrop):
+                x, z, pos_bias = layer(
+                    x,
+                    self_attn_padding_mask=padding_mask,
+                    need_weights=False,
+                    self_attn_mask=streaming_mask,
+                    pos_bias=pos_bias,
+                )
+            if tgt_layer is not None:
+                layer_results.append((x, z))
+            if i == tgt_layer:
+                r = x
+                break
+
+        if r is not None:
+            x = r
+
+        # T x B x C -> B x T x C
+        x = x.transpose(0, 1)
+
+        return x, layer_results
+
+
+class TransformerSentenceEncoderLayer(nn.Module):
+    """
+    Implements a Transformer Encoder Layer used in BERT/XLM style pre-trained
+    models.
+    """
+
+    def __init__(
+        self,
+        embedding_dim: float = 768,
+        ffn_embedding_dim: float = 3072,
+        num_attention_heads: float = 8,
+        dropout: float = 0.1,
+        attention_dropout: float = 0.1,
+        activation_dropout: float = 0.1,
+        activation_fn: str = "relu",
+        layer_norm_first: bool = False,
+        has_relative_attention_bias: bool = False,
+        num_buckets: int = 0,
+        max_distance: int = 0,
+        rescale_init: bool = False,
+        gru_rel_pos: bool = False,
+    ) -> None:
+
+        super().__init__()
+        # Initialize parameters
+        self.embedding_dim = embedding_dim
+        self.dropout = dropout
+        self.activation_dropout = activation_dropout
+
+        # Initialize blocks
+        self.activation_name = activation_fn
+        self.activation_fn = get_activation_fn(activation_fn)
+        self.self_attn = MultiheadAttention(
+            self.embedding_dim,
+            num_attention_heads,
+            dropout=attention_dropout,
+            self_attention=True,
+            has_relative_attention_bias=has_relative_attention_bias,
+            num_buckets=num_buckets,
+            max_distance=max_distance,
+            rescale_init=rescale_init,
+            gru_rel_pos=gru_rel_pos,
+        )
+
+        self.dropout1 = nn.Dropout(dropout)
+        self.dropout2 = nn.Dropout(self.activation_dropout)
+        self.dropout3 = nn.Dropout(dropout)
+
+        self.layer_norm_first = layer_norm_first
+
+        # layer norm associated with the self attention layer
+        self.self_attn_layer_norm = LayerNorm(self.embedding_dim)
+
+        if self.activation_name == "glu":
+            self.fc1 = GLU_Linear(self.embedding_dim, ffn_embedding_dim, "swish")
+        else:
+            self.fc1 = nn.Linear(self.embedding_dim, ffn_embedding_dim)
+        self.fc2 = nn.Linear(ffn_embedding_dim, self.embedding_dim)
+
+        # layer norm associated with the position wise feed-forward NN
+        self.final_layer_norm = LayerNorm(self.embedding_dim)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        self_attn_mask: torch.Tensor = None,
+        self_attn_padding_mask: torch.Tensor = None,
+        need_weights: bool = False,
+        pos_bias=None,
+    ):
+        """
+        LayerNorm is applied either before or after the self-attention/ffn
+        modules similar to the original Transformer imlementation.
+        """
+        residual = x
+
+        if self.layer_norm_first:
+            x = self.self_attn_layer_norm(x)
+            x, attn, pos_bias = self.self_attn(
+                query=x,
+                key=x,
+                value=x,
+                key_padding_mask=self_attn_padding_mask,
+                need_weights=False,
+                attn_mask=self_attn_mask,
+                position_bias=pos_bias,
+            )
+            x = self.dropout1(x)
+            x = residual + x
+
+            residual = x
+            x = self.final_layer_norm(x)
+            if self.activation_name == "glu":
+                x = self.fc1(x)
+            else:
+                x = self.activation_fn(self.fc1(x))
+            x = self.dropout2(x)
+            x = self.fc2(x)
+            x = self.dropout3(x)
+            x = residual + x
+        else:
+            x, attn, pos_bias = self.self_attn(
+                query=x,
+                key=x,
+                value=x,
+                key_padding_mask=self_attn_padding_mask,
+                need_weights=need_weights,
+                attn_mask=self_attn_mask,
+                position_bias=pos_bias,
+            )
+
+            x = self.dropout1(x)
+            x = residual + x
+
+            x = self.self_attn_layer_norm(x)
+
+            residual = x
+            if self.activation_name == "glu":
+                x = self.fc1(x)
+            else:
+                x = self.activation_fn(self.fc1(x))
+            x = self.dropout2(x)
+            x = self.fc2(x)
+            x = self.dropout3(x)
+            x = residual + x
+            x = self.final_layer_norm(x)
+
+        return x, attn, pos_bias

manipulate_model/encoder/wavlm/modules.py

@@ -0,0 +1,827 @@
+# --------------------------------------------------------
+# WavLM: Large-Scale Self-Supervised  Pre-training  for Full Stack Speech Processing (https://arxiv.org/abs/2110.13900.pdf)
+# Github source: https://github.com/microsoft/unilm/tree/master/wavlm
+# Copyright (c) 2021 Microsoft
+# Licensed under The MIT License [see LICENSE for details]
+# Based on fairseq code bases
+# https://github.com/pytorch/fairseq
+# --------------------------------------------------------
+
+import math
+import warnings
+from typing import Dict, Optional, Tuple
+import torch
+from torch import Tensor, nn
+from torch.nn import Parameter
+import torch.nn.functional as F
+
+
+class TransposeLast(nn.Module):
+    def __init__(self, deconstruct_idx=None):
+        super().__init__()
+        self.deconstruct_idx = deconstruct_idx
+
+    def forward(self, x):
+        if self.deconstruct_idx is not None:
+            x = x[self.deconstruct_idx]
+        return x.transpose(-2, -1)
+
+
+class Fp32LayerNorm(nn.LayerNorm):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+    def forward(self, input):
+        output = F.layer_norm(
+            input.float(),
+            self.normalized_shape,
+            self.weight.float() if self.weight is not None else None,
+            self.bias.float() if self.bias is not None else None,
+            self.eps,
+        )
+        return output.type_as(input)
+
+
+class Fp32GroupNorm(nn.GroupNorm):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+    def forward(self, input):
+        output = F.group_norm(
+            input.float(),
+            self.num_groups,
+            self.weight.float() if self.weight is not None else None,
+            self.bias.float() if self.bias is not None else None,
+            self.eps,
+        )
+        return output.type_as(input)
+
+
+class GradMultiply(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, x, scale):
+        ctx.scale = scale
+        res = x.new(x)
+        return res
+
+    @staticmethod
+    def backward(ctx, grad):
+        return grad * ctx.scale, None
+
+
+class SamePad(nn.Module):
+    def __init__(self, kernel_size, causal=False):
+        super().__init__()
+        if causal:
+            self.remove = kernel_size - 1
+        else:
+            self.remove = 1 if kernel_size % 2 == 0 else 0
+
+    def forward(self, x):
+        if self.remove > 0:
+            x = x[:, :, : -self.remove]
+        return x
+
+
+class Swish(nn.Module):
+    """Swish function
+    """
+
+    def __init__(self):
+        """Construct an MultiHeadedAttention object."""
+        super(Swish, self).__init__()
+        self.act = torch.nn.Sigmoid()
+
+    def forward(self, x):
+        return x * self.act(x)
+
+
+class GLU_Linear(nn.Module):
+    def __init__(self, input_dim, output_dim, glu_type="sigmoid", bias_in_glu=True):
+        super(GLU_Linear, self).__init__()
+
+        self.glu_type = glu_type
+        self.output_dim = output_dim
+
+        if glu_type == "sigmoid":
+            self.glu_act = torch.nn.Sigmoid()
+        elif glu_type == "swish":
+            self.glu_act = Swish()
+        elif glu_type == "relu":
+            self.glu_act = torch.nn.ReLU()
+        elif glu_type == "gelu":
+            self.glu_act = torch.nn.GELU()
+
+        if bias_in_glu:
+            self.linear = nn.Linear(input_dim, output_dim * 2, True)
+        else:
+            self.linear = nn.Linear(input_dim, output_dim * 2, False)
+
+    def forward(self, x):
+        # to be consistent with GLU_Linear, we assume the input always has the #channel (#dim) in the last dimension of the tensor, so need to switch the dimension first for 1D-Conv case
+        x = self.linear(x)
+
+        if self.glu_type == "bilinear":
+            x = (x[:, :, 0:self.output_dim] * x[:, :, self.output_dim:self.output_dim * 2])
+        else:
+            x = (x[:, :, 0:self.output_dim] * self.glu_act(x[:, :, self.output_dim:self.output_dim * 2]))
+
+        return x
+
+
+def gelu_accurate(x):
+    if not hasattr(gelu_accurate, "_a"):
+        gelu_accurate._a = math.sqrt(2 / math.pi)
+    return (
+        0.5 * x * (1 + torch.tanh(gelu_accurate._a * (x + 0.044715 * torch.pow(x, 3))))
+    )
+
+
+def gelu(x: torch.Tensor) -> torch.Tensor:
+    return torch.nn.functional.gelu(x.float()).type_as(x)
+
+
+def get_activation_fn(activation: str):
+    """Returns the activation function corresponding to `activation`"""
+
+    if activation == "relu":
+        return F.relu
+    elif activation == "gelu":
+        return gelu
+    elif activation == "gelu_fast":
+        warnings.warn(
+            "--activation-fn=gelu_fast has been renamed to gelu_accurate"
+        )
+        return gelu_accurate
+    elif activation == "gelu_accurate":
+        return gelu_accurate
+    elif activation == "tanh":
+        return torch.tanh
+    elif activation == "linear":
+        return lambda x: x
+    elif activation == "glu":
+        return lambda x: x
+    else:
+        raise RuntimeError("--activation-fn {} not supported".format(activation))
+
+
+def init_bert_params(module):
+    """
+    Initialize the weights specific to the BERT Model.
+    This overrides the default initializations depending on the specified arguments.
+        1. If normal_init_linear_weights is set then weights of linear
+           layer will be initialized using the normal distribution and
+           bais will be set to the specified value.
+        2. If normal_init_embed_weights is set then weights of embedding
+           layer will be initialized using the normal distribution.
+        3. If normal_init_proj_weights is set then weights of
+           in_project_weight for MultiHeadAttention initialized using
+           the normal distribution (to be validated).
+    """
+
+    def normal_(data):
+        # with FSDP, module params will be on CUDA, so we cast them back to CPU
+        # so that the RNG is consistent with and without FSDP
+        data.copy_(
+            data.cpu().normal_(mean=0.0, std=0.02).to(data.device)
+        )
+
+    if isinstance(module, nn.Linear):
+        normal_(module.weight.data)
+        if module.bias is not None:
+            module.bias.data.zero_()
+    if isinstance(module, nn.Embedding):
+        normal_(module.weight.data)
+        if module.padding_idx is not None:
+            module.weight.data[module.padding_idx].zero_()
+    if isinstance(module, MultiheadAttention):
+        normal_(module.q_proj.weight.data)
+        normal_(module.k_proj.weight.data)
+        normal_(module.v_proj.weight.data)
+
+
+def quant_noise(module, p, block_size):
+    """
+    Wraps modules and applies quantization noise to the weights for
+    subsequent quantization with Iterative Product Quantization as
+    described in "Training with Quantization Noise for Extreme Model Compression"
+
+    Args:
+        - module: nn.Module
+        - p: amount of Quantization Noise
+        - block_size: size of the blocks for subsequent quantization with iPQ
+
+    Remarks:
+        - Module weights must have the right sizes wrt the block size
+        - Only Linear, Embedding and Conv2d modules are supported for the moment
+        - For more detail on how to quantize by blocks with convolutional weights,
+          see "And the Bit Goes Down: Revisiting the Quantization of Neural Networks"
+        - We implement the simplest form of noise here as stated in the paper
+          which consists in randomly dropping blocks
+    """
+
+    # if no quantization noise, don't register hook
+    if p <= 0:
+        return module
+
+    # supported modules
+    assert isinstance(module, (nn.Linear, nn.Embedding, nn.Conv2d))
+
+    # test whether module.weight has the right sizes wrt block_size
+    is_conv = module.weight.ndim == 4
+
+    # 2D matrix
+    if not is_conv:
+        assert (
+            module.weight.size(1) % block_size == 0
+        ), "Input features must be a multiple of block sizes"
+
+    # 4D matrix
+    else:
+        # 1x1 convolutions
+        if module.kernel_size == (1, 1):
+            assert (
+                module.in_channels % block_size == 0
+            ), "Input channels must be a multiple of block sizes"
+        # regular convolutions
+        else:
+            k = module.kernel_size[0] * module.kernel_size[1]
+            assert k % block_size == 0, "Kernel size must be a multiple of block size"
+
+    def _forward_pre_hook(mod, input):
+        # no noise for evaluation
+        if mod.training:
+            if not is_conv:
+                # gather weight and sizes
+                weight = mod.weight
+                in_features = weight.size(1)
+                out_features = weight.size(0)
+
+                # split weight matrix into blocks and randomly drop selected blocks
+                mask = torch.zeros(
+                    in_features // block_size * out_features, device=weight.device
+                )
+                mask.bernoulli_(p)
+                mask = mask.repeat_interleave(block_size, -1).view(-1, in_features)
+
+            else:
+                # gather weight and sizes
+                weight = mod.weight
+                in_channels = mod.in_channels
+                out_channels = mod.out_channels
+
+                # split weight matrix into blocks and randomly drop selected blocks
+                if mod.kernel_size == (1, 1):
+                    mask = torch.zeros(
+                        int(in_channels // block_size * out_channels),
+                        device=weight.device,
+                    )
+                    mask.bernoulli_(p)
+                    mask = mask.repeat_interleave(block_size, -1).view(-1, in_channels)
+                else:
+                    mask = torch.zeros(
+                        weight.size(0), weight.size(1), device=weight.device
+                    )
+                    mask.bernoulli_(p)
+                    mask = (
+                        mask.unsqueeze(2)
+                        .unsqueeze(3)
+                        .repeat(1, 1, mod.kernel_size[0], mod.kernel_size[1])
+                    )
+
+            # scale weights and apply mask
+            mask = mask.to(
+                torch.bool
+            )  # x.bool() is not currently supported in TorchScript
+            s = 1 / (1 - p)
+            mod.weight.data = s * weight.masked_fill(mask, 0)
+
+    module.register_forward_pre_hook(_forward_pre_hook)
+    return module
+
+
+class MultiheadAttention(nn.Module):
+    """Multi-headed attention.
+
+    See "Attention Is All You Need" for more details.
+    """
+
+    def __init__(
+            self,
+            embed_dim,
+            num_heads,
+            kdim=None,
+            vdim=None,
+            dropout=0.0,
+            bias=True,
+            add_bias_kv=False,
+            add_zero_attn=False,
+            self_attention=False,
+            encoder_decoder_attention=False,
+            q_noise=0.0,
+            qn_block_size=8,
+            has_relative_attention_bias=False,
+            num_buckets=32,
+            max_distance=128,
+            gru_rel_pos=False,
+            rescale_init=False,
+    ):
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.kdim = kdim if kdim is not None else embed_dim
+        self.vdim = vdim if vdim is not None else embed_dim
+        self.qkv_same_dim = self.kdim == embed_dim and self.vdim == embed_dim
+
+        self.num_heads = num_heads
+        self.dropout_module = nn.Dropout(dropout)
+
+        self.has_relative_attention_bias = has_relative_attention_bias
+        self.num_buckets = num_buckets
+        self.max_distance = max_distance
+        if self.has_relative_attention_bias:
+            self.relative_attention_bias = nn.Embedding(num_buckets, num_heads)
+
+        self.head_dim = embed_dim // num_heads
+        self.q_head_dim = self.head_dim
+        self.k_head_dim = self.head_dim
+        assert (
+                self.head_dim * num_heads == self.embed_dim
+        ), "embed_dim must be divisible by num_heads"
+        self.scaling = self.head_dim ** -0.5
+
+        self.self_attention = self_attention
+        self.encoder_decoder_attention = encoder_decoder_attention
+
+        assert not self.self_attention or self.qkv_same_dim, (
+            "Self-attention requires query, key and " "value to be of the same size"
+        )
+
+        k_bias = True
+        if rescale_init:
+            k_bias = False
+
+        k_embed_dim = embed_dim
+        q_embed_dim = embed_dim
+
+        self.k_proj = quant_noise(
+            nn.Linear(self.kdim, k_embed_dim, bias=k_bias), q_noise, qn_block_size
+        )
+        self.v_proj = quant_noise(
+            nn.Linear(self.vdim, embed_dim, bias=bias), q_noise, qn_block_size
+        )
+        self.q_proj = quant_noise(
+            nn.Linear(embed_dim, q_embed_dim, bias=bias), q_noise, qn_block_size
+        )
+
+        self.out_proj = quant_noise(
+            nn.Linear(embed_dim, embed_dim, bias=bias), q_noise, qn_block_size
+        )
+
+        if add_bias_kv:
+            self.bias_k = Parameter(torch.Tensor(1, 1, embed_dim))
+            self.bias_v = Parameter(torch.Tensor(1, 1, embed_dim))
+        else:
+            self.bias_k = self.bias_v = None
+
+        self.add_zero_attn = add_zero_attn
+
+        self.gru_rel_pos = gru_rel_pos
+        if self.gru_rel_pos:
+            self.grep_linear = nn.Linear(self.q_head_dim, 8)
+            self.grep_a = nn.Parameter(torch.ones(1, num_heads, 1, 1))
+
+        self.reset_parameters()
+
+    def reset_parameters(self):
+        if self.qkv_same_dim:
+            # Empirically observed the convergence to be much better with
+            # the scaled initialization
+            nn.init.xavier_uniform_(self.k_proj.weight, gain=1 / math.sqrt(2))
+            nn.init.xavier_uniform_(self.v_proj.weight, gain=1 / math.sqrt(2))
+            nn.init.xavier_uniform_(self.q_proj.weight, gain=1 / math.sqrt(2))
+        else:
+            nn.init.xavier_uniform_(self.k_proj.weight)
+            nn.init.xavier_uniform_(self.v_proj.weight)
+            nn.init.xavier_uniform_(self.q_proj.weight)
+
+        nn.init.xavier_uniform_(self.out_proj.weight)
+        if self.out_proj.bias is not None:
+            nn.init.constant_(self.out_proj.bias, 0.0)
+        if self.bias_k is not None:
+            nn.init.xavier_normal_(self.bias_k)
+        if self.bias_v is not None:
+            nn.init.xavier_normal_(self.bias_v)
+        if self.has_relative_attention_bias:
+            nn.init.xavier_normal_(self.relative_attention_bias.weight)
+
+    def _relative_positions_bucket(self, relative_positions, bidirectional=True):
+        num_buckets = self.num_buckets
+        max_distance = self.max_distance
+        relative_buckets = 0
+
+        if bidirectional:
+            num_buckets = num_buckets // 2
+            relative_buckets += (relative_positions > 0).to(torch.long) * num_buckets
+            relative_positions = torch.abs(relative_positions)
+        else:
+            relative_positions = -torch.min(relative_positions, torch.zeros_like(relative_positions))
+
+        max_exact = num_buckets // 2
+        is_small = relative_positions < max_exact
+
+        relative_postion_if_large = max_exact + (
+                torch.log(relative_positions.float() / max_exact)
+                / math.log(max_distance / max_exact)
+                * (num_buckets - max_exact)
+        ).to(torch.long)
+        relative_postion_if_large = torch.min(
+            relative_postion_if_large, torch.full_like(relative_postion_if_large, num_buckets - 1)
+        )
+
+        relative_buckets += torch.where(is_small, relative_positions, relative_postion_if_large)
+        return relative_buckets
+
+    def compute_bias(self, query_length, key_length):
+        context_position = torch.arange(query_length, dtype=torch.long)[:, None]
+        memory_position = torch.arange(key_length, dtype=torch.long)[None, :]
+        relative_position = memory_position - context_position
+        relative_position_bucket = self._relative_positions_bucket(
+            relative_position,
+            bidirectional=True
+        )
+        relative_position_bucket = relative_position_bucket.to(self.relative_attention_bias.weight.device)
+        values = self.relative_attention_bias(relative_position_bucket)
+        values = values.permute([2, 0, 1])
+        return values
+
+    def forward(
+            self,
+            query,
+            key: Optional[Tensor],
+            value: Optional[Tensor],
+            key_padding_mask: Optional[Tensor] = None,
+            incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None,
+            need_weights: bool = True,
+            static_kv: bool = False,
+            attn_mask: Optional[Tensor] = None,
+            before_softmax: bool = False,
+            need_head_weights: bool = False,
+            position_bias: Optional[Tensor] = None
+    ) -> Tuple[Tensor, Optional[Tensor], Optional[Tensor]]:
+        """Input shape: Time x Batch x Channel
+
+        Args:
+            key_padding_mask (ByteTensor, optional): mask to exclude
+                keys that are pads, of shape `(batch, src_len)`, where
+                padding elements are indicated by 1s.
+            need_weights (bool, optional): return the attention weights,
+                averaged over heads (default: False).
+            attn_mask (ByteTensor, optional): typically used to
+                implement causal attention, where the mask prevents the
+                attention from looking forward in time (default: None).
+            before_softmax (bool, optional): return the raw attention
+                weights and values before the attention softmax.
+            need_head_weights (bool, optional): return the attention
+                weights for each head. Implies *need_weights*. Default:
+                return the average attention weights over all heads.
+        """
+        if need_head_weights:
+            need_weights = True
+
+        is_tpu = query.device.type == "xla"
+
+        tgt_len, bsz, embed_dim = query.size()
+        src_len = tgt_len
+        assert embed_dim == self.embed_dim
+        assert list(query.size()) == [tgt_len, bsz, embed_dim]
+        if key is not None:
+            src_len, key_bsz, _ = key.size()
+            if not torch.jit.is_scripting():
+                assert key_bsz == bsz
+                assert value is not None
+                assert src_len, bsz == value.shape[:2]
+
+        if self.has_relative_attention_bias and position_bias is None:
+            position_bias = self.compute_bias(tgt_len, src_len)
+            position_bias = position_bias.unsqueeze(0).repeat(bsz, 1, 1, 1).view(bsz * self.num_heads, tgt_len, src_len)
+
+        if (
+                not is_tpu  # don't use PyTorch version on TPUs
+                and incremental_state is None
+                and not static_kv
+                # A workaround for quantization to work. Otherwise JIT compilation
+                # treats bias in linear module as method.
+                and not torch.jit.is_scripting()
+                and self.q_head_dim == self.head_dim
+        ):
+            assert key is not None and value is not None
+            assert attn_mask is None
+
+            attn_mask_rel_pos = None
+            if position_bias is not None:
+                attn_mask_rel_pos = position_bias
+                if self.gru_rel_pos:
+                    query_layer = query.transpose(0, 1)
+                    new_x_shape = query_layer.size()[:-1] + (self.num_heads, -1)
+                    query_layer = query_layer.view(*new_x_shape)
+                    query_layer = query_layer.permute(0, 2, 1, 3)
+                    _B, _H, _L, __ = query_layer.size()
+
+                    gate_a, gate_b = torch.sigmoid(self.grep_linear(query_layer).view(
+                        _B, _H, _L, 2, 4).sum(-1, keepdim=False)).chunk(2, dim=-1)
+                    gate_a_1 = gate_a * (gate_b * self.grep_a - 1.0) + 2.0
+                    attn_mask_rel_pos = gate_a_1.view(bsz * self.num_heads, -1, 1) * position_bias
+
+                attn_mask_rel_pos = attn_mask_rel_pos.view((-1, tgt_len, tgt_len))
+            k_proj_bias = self.k_proj.bias
+            if k_proj_bias is None:
+                k_proj_bias = torch.zeros_like(self.q_proj.bias)
+
+            x, attn = F.multi_head_attention_forward(
+                query,
+                key,
+                value,
+                self.embed_dim,
+                self.num_heads,
+                torch.empty([0]),
+                torch.cat((self.q_proj.bias, self.k_proj.bias, self.v_proj.bias)),
+                self.bias_k,
+                self.bias_v,
+                self.add_zero_attn,
+                self.dropout_module.p,
+                self.out_proj.weight,
+                self.out_proj.bias,
+                self.training,
+                # self.training or self.dropout_module.apply_during_inference,
+                key_padding_mask,
+                need_weights,
+                attn_mask_rel_pos,
+                use_separate_proj_weight=True,
+                q_proj_weight=self.q_proj.weight,
+                k_proj_weight=self.k_proj.weight,
+                v_proj_weight=self.v_proj.weight,
+            )
+            return x, attn, position_bias
+
+        if incremental_state is not None:
+            saved_state = self._get_input_buffer(incremental_state)
+            if saved_state is not None and "prev_key" in saved_state:
+                # previous time steps are cached - no need to recompute
+                # key and value if they are static
+                if static_kv:
+                    assert self.encoder_decoder_attention and not self.self_attention
+                    key = value = None
+        else:
+            saved_state = None
+
+        if self.self_attention:
+            q = self.q_proj(query)
+            k = self.k_proj(query)
+            v = self.v_proj(query)
+        elif self.encoder_decoder_attention:
+            # encoder-decoder attention
+            q = self.q_proj(query)
+            if key is None:
+                assert value is None
+                k = v = None
+            else:
+                k = self.k_proj(key)
+                v = self.v_proj(key)
+
+        else:
+            assert key is not None and value is not None
+            q = self.q_proj(query)
+            k = self.k_proj(key)
+            v = self.v_proj(value)
+        q *= self.scaling
+
+        if self.bias_k is not None:
+            assert self.bias_v is not None
+            k = torch.cat([k, self.bias_k.repeat(1, bsz, 1)])
+            v = torch.cat([v, self.bias_v.repeat(1, bsz, 1)])
+            if attn_mask is not None:
+                attn_mask = torch.cat(
+                    [attn_mask, attn_mask.new_zeros(attn_mask.size(0), 1)], dim=1
+                )
+            if key_padding_mask is not None:
+                key_padding_mask = torch.cat(
+                    [
+                        key_padding_mask,
+                        key_padding_mask.new_zeros(key_padding_mask.size(0), 1),
+                    ],
+                    dim=1,
+                )
+
+        q = (
+            q.contiguous()
+                .view(tgt_len, bsz * self.num_heads, self.q_head_dim)
+                .transpose(0, 1)
+        )
+        if k is not None:
+            k = (
+                k.contiguous()
+                    .view(-1, bsz * self.num_heads, self.k_head_dim)
+                    .transpose(0, 1)
+            )
+        if v is not None:
+            v = (
+                v.contiguous()
+                    .view(-1, bsz * self.num_heads, self.head_dim)
+                    .transpose(0, 1)
+            )
+
+        if saved_state is not None:
+            # saved states are stored with shape (bsz, num_heads, seq_len, head_dim)
+            if "prev_key" in saved_state:
+                _prev_key = saved_state["prev_key"]
+                assert _prev_key is not None
+                prev_key = _prev_key.view(bsz * self.num_heads, -1, self.head_dim)
+                if static_kv:
+                    k = prev_key
+                else:
+                    assert k is not None
+                    k = torch.cat([prev_key, k], dim=1)
+                src_len = k.size(1)
+            if "prev_value" in saved_state:
+                _prev_value = saved_state["prev_value"]
+                assert _prev_value is not None
+                prev_value = _prev_value.view(bsz * self.num_heads, -1, self.head_dim)
+                if static_kv:
+                    v = prev_value
+                else:
+                    assert v is not None
+                    v = torch.cat([prev_value, v], dim=1)
+            prev_key_padding_mask: Optional[Tensor] = None
+            if "prev_key_padding_mask" in saved_state:
+                prev_key_padding_mask = saved_state["prev_key_padding_mask"]
+            assert k is not None and v is not None
+            key_padding_mask = MultiheadAttention._append_prev_key_padding_mask(
+                key_padding_mask=key_padding_mask,
+                prev_key_padding_mask=prev_key_padding_mask,
+                batch_size=bsz,
+                src_len=k.size(1),
+                static_kv=static_kv,
+            )
+
+            saved_state["prev_key"] = k.view(bsz, self.num_heads, -1, self.head_dim)
+            saved_state["prev_value"] = v.view(bsz, self.num_heads, -1, self.head_dim)
+            saved_state["prev_key_padding_mask"] = key_padding_mask
+            # In this branch incremental_state is never None
+            assert incremental_state is not None
+            incremental_state = self._set_input_buffer(incremental_state, saved_state)
+        assert k is not None
+        assert k.size(1) == src_len
+
+        # This is part of a workaround to get around fork/join parallelism
+        # not supporting Optional types.
+        if key_padding_mask is not None and key_padding_mask.dim() == 0:
+            key_padding_mask = None
+
+        if key_padding_mask is not None:
+            assert key_padding_mask.size(0) == bsz
+            assert key_padding_mask.size(1) == src_len
+
+        if self.add_zero_attn:
+            assert v is not None
+            src_len += 1
+            k = torch.cat([k, k.new_zeros((k.size(0), 1) + k.size()[2:])], dim=1)
+            v = torch.cat([v, v.new_zeros((v.size(0), 1) + v.size()[2:])], dim=1)
+            if attn_mask is not None:
+                attn_mask = torch.cat(
+                    [attn_mask, attn_mask.new_zeros(attn_mask.size(0), 1)], dim=1
+                )
+            if key_padding_mask is not None:
+                key_padding_mask = torch.cat(
+                    [
+                        key_padding_mask,
+                        torch.zeros(key_padding_mask.size(0), 1).type_as(
+                            key_padding_mask
+                        ),
+                    ],
+                    dim=1,
+                )
+
+        attn_weights = torch.bmm(q, k.transpose(1, 2))
+        attn_weights = self.apply_sparse_mask(attn_weights, tgt_len, src_len, bsz)
+
+        assert list(attn_weights.size()) == [bsz * self.num_heads, tgt_len, src_len]
+
+        if attn_mask is not None:
+            attn_mask = attn_mask.unsqueeze(0)
+            attn_weights += attn_mask
+
+        if key_padding_mask is not None:
+            # don't attend to padding symbols
+            attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
+            if not is_tpu:
+                attn_weights = attn_weights.masked_fill(
+                    key_padding_mask.unsqueeze(1).unsqueeze(2).to(torch.bool),
+                    float("-inf"),
+                )
+            else:
+                attn_weights = attn_weights.transpose(0, 2)
+                attn_weights = attn_weights.masked_fill(key_padding_mask, float("-inf"))
+                attn_weights = attn_weights.transpose(0, 2)
+            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
+
+        if before_softmax:
+            return attn_weights, v, position_bias
+
+        if position_bias is not None:
+            if self.gru_rel_pos == 1:
+                query_layer = q.view(bsz, self.num_heads, tgt_len, self.q_head_dim)
+                _B, _H, _L, __ = query_layer.size()
+                gate_a, gate_b = torch.sigmoid(self.grep_linear(query_layer).view(
+                    _B, _H, _L, 2, 4).sum(-1, keepdim=False)).chunk(2, dim=-1)
+                gate_a_1 = gate_a * (gate_b * self.grep_a - 1.0) + 2.0
+                position_bias = gate_a_1.view(bsz * self.num_heads, -1, 1) * position_bias
+
+            position_bias = position_bias.view(attn_weights.size())
+
+            attn_weights = attn_weights + position_bias
+
+        attn_weights_float = F.softmax(
+            attn_weights, dim=-1
+        )
+        attn_weights = attn_weights_float.type_as(attn_weights)
+        attn_probs = self.dropout_module(attn_weights)
+
+        assert v is not None
+        attn = torch.bmm(attn_probs, v)
+        assert list(attn.size()) == [bsz * self.num_heads, tgt_len, self.head_dim]
+        attn = attn.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim)
+        attn = self.out_proj(attn)
+        attn_weights: Optional[Tensor] = None
+        if need_weights:
+            attn_weights = attn_weights_float.view(
+                bsz, self.num_heads, tgt_len, src_len
+            ).transpose(1, 0)
+            if not need_head_weights:
+                # average attention weights over heads
+                attn_weights = attn_weights.mean(dim=0)
+
+        return attn, attn_weights, position_bias
+
+    @staticmethod
+    def _append_prev_key_padding_mask(
+            key_padding_mask: Optional[Tensor],
+            prev_key_padding_mask: Optional[Tensor],
+            batch_size: int,
+            src_len: int,
+            static_kv: bool,
+    ) -> Optional[Tensor]:
+        # saved key padding masks have shape (bsz, seq_len)
+        if prev_key_padding_mask is not None and static_kv:
+            new_key_padding_mask = prev_key_padding_mask
+        elif prev_key_padding_mask is not None and key_padding_mask is not None:
+            new_key_padding_mask = torch.cat(
+                [prev_key_padding_mask.float(), key_padding_mask.float()], dim=1
+            )
+        # During incremental decoding, as the padding token enters and
+        # leaves the frame, there will be a time when prev or current
+        # is None
+        elif prev_key_padding_mask is not None:
+            if src_len > prev_key_padding_mask.size(1):
+                filler = torch.zeros(
+                    (batch_size, src_len - prev_key_padding_mask.size(1)),
+                    device=prev_key_padding_mask.device,
+                )
+                new_key_padding_mask = torch.cat(
+                    [prev_key_padding_mask.float(), filler.float()], dim=1
+                )
+            else:
+                new_key_padding_mask = prev_key_padding_mask.float()
+        elif key_padding_mask is not None:
+            if src_len > key_padding_mask.size(1):
+                filler = torch.zeros(
+                    (batch_size, src_len - key_padding_mask.size(1)),
+                    device=key_padding_mask.device,
+                )
+                new_key_padding_mask = torch.cat(
+                    [filler.float(), key_padding_mask.float()], dim=1
+                )
+            else:
+                new_key_padding_mask = key_padding_mask.float()
+        else:
+            new_key_padding_mask = prev_key_padding_mask
+        return new_key_padding_mask
+
+    def _get_input_buffer(
+            self, incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]]
+    ) -> Dict[str, Optional[Tensor]]:
+        result = self.get_incremental_state(incremental_state, "attn_state")
+        if result is not None:
+            return result
+        else:
+            empty_result: Dict[str, Optional[Tensor]] = {}
+            return empty_result
+
+    def _set_input_buffer(
+            self,
+            incremental_state: Dict[str, Dict[str, Optional[Tensor]]],
+            buffer: Dict[str, Optional[Tensor]],
+    ):
+        return self.set_incremental_state(incremental_state, "attn_state", buffer)
+
+    def apply_sparse_mask(self, attn_weights, tgt_len: int, src_len: int, bsz: int):
+        return attn_weights

manipulate_model/model.py

@@ -0,0 +1,25 @@
+import torch
+import torch.nn as nn
+
+from manipulate_model.encoder.encoder import Encoder
+from manipulate_model.decoder.decoder import Decoder
+
+
+class Model(nn.Module):
+    def __init__(self, config):
+        super(Model, self).__init__()
+        self.config = config
+
+        self.encoder = Encoder(self.config)
+        self.config.model.decoder.temporal_dim = self.encoder.get_temporal_dim()
+        self.config.model.decoder.encoding_dim = self.encoder.get_encoding_dim()
+        self.decoder = Decoder(self.config)
+
+    def forward(self, x):
+        if self.config.model.encoder_freeze:
+            with torch.no_grad():
+                x = self.encoder(x)
+        else:
+            x = self.encoder(x)
+        x = self.decoder(x)
+        return x

manipulate_model/utils.py

@@ -0,0 +1,139 @@
+import os
+import torch
+import torchaudio
+import numpy as np
+from omegaconf import OmegaConf
+from torchvision.io import read_video
+from torch.nn.functional import pad, normalize, softmax
+
+from manipulate_model.model import Model
+
+
+
+def get_config_and_model(model_root="manipulate_model/demo-model/audio"):
+    config_path = os.path.join(model_root, "config.yaml")
+    config = OmegaConf.load(config_path)
+    if isinstance(config.model.encoder, str):
+        config.model.encoder = OmegaConf.load(config.model.encoder)
+    if isinstance(config.model.decoder, str):
+        config.model.decoder = OmegaConf.load(config.model.decoder)
+
+    model = Model(config)
+    #weights = torch.load(os.path.join(model_root, "weights.pt"))
+    #model.load_state_dict(weights["model_state_dict"])
+
+    return config, model
+
+
+def load_audio(file_path, config):
+    # Load audio
+    # Parameters
+    # ----------
+    # file_path : str
+    #     Path to audio file
+    # Returns
+    # -------
+    # torch.Tensor
+
+    audio = None
+
+    if file_path.endswith(".wav") or file_path.endswith(".flac"):
+        audio, sample_rate = torchaudio.load(file_path)
+    elif file_path.endswith(".mp3"):
+        pass
+    elif file_path.endswith(".mp4"):
+        #_, audio, _ = read_video(file_path)
+        pass
+
+    return preprocess_audio(audio, config)
+
+
+def preprocess_audio(audio, config, step_size=1):
+    # Preprocess audio
+    # Parameters
+    # ----------
+    # audio : torch.Tensor
+    #     Audio signal
+    # config : OmegaConf
+    #     Configuration object
+    # Returns
+    # -------
+    # torch.Tensor : Normalized audio signal
+
+    window_size = config.data.window_size
+    sr = config.data.sr
+    fps = config.data.fps
+
+    audio_len = audio.shape[1]
+    step_size = step_size * (sr // fps)
+    window_size = window_size * (sr // fps)
+    audio = pad(audio, (window_size, window_size), "constant", 0)
+
+    sliced_audio = []
+
+    for i in range(0, audio_len + window_size, step_size):
+        audio_slice = audio[:, i : i + window_size]
+
+        if audio_slice.shape[1] < window_size:
+            audio_slice = pad(
+                audio_slice, (0, window_size - audio_slice.shape[1]), "constant", 0
+            )
+
+        audio_slice = normalize(audio_slice, dim=1)
+        sliced_audio.append(audio_slice)
+
+    sliced_audio = torch.stack(sliced_audio).squeeze()
+
+    return sliced_audio
+
+
+def infere(model, x, config, device="cpu", bs=8):
+    print(x)
+    model.eval()
+
+    x = load_audio(x, config)
+
+    # Inference (x is a stack of windows)
+    frame_predictions = []
+
+    with torch.no_grad():
+        n_iter = x.shape[0]
+
+        for i in range(0, n_iter, bs):
+            input_batch = x[i: i + bs]
+            input_batch = input_batch.to(device)
+
+            output = softmax(model(input_batch), dim=1)
+            frame_predictions.append(output.cpu().numpy())
+
+        frame_predictions = np.concatenate(frame_predictions, axis=0)[:,0]
+
+
+    return frame_predictions
+
+def convert_frame_predictions_to_timestamps(frame_predictions, fps, window_size):
+    # Convert frame predictions to timestamps
+    # Parameters
+    # ----------
+    # frame_predictions : np.ndarray
+    #     Frame predictions
+    # fps : int
+    #     Frames per second
+    # Returns
+    # -------
+    # np.ndarray : Timestamps
+
+    frame_predictions = (
+        frame_predictions[
+            int(window_size / 2) : -int(window_size / 2), 0
+        ]  # removes the padding, does not consider step size as of now
+        .round()
+        .astype(int)
+    )
+    timestamps = []
+
+    for i, frame_prediction in enumerate(frame_predictions):
+        if frame_prediction == 1:
+            timestamps.append(i / fps)
+
+    return timestamps