modeling_MERT.py
| 1 | """ |
| 2 | MERT model definition. |
| 3 | We largely adapt codes from: |
| 4 | 1. https://github.com/huggingface/transformers/blob/main/src/transformers/models/hubert/modeling_hubert.py |
| 5 | 2. https://github.com/facebookresearch/fairseq/blob/main/fairseq/models/wav2vec/wav2vec2.py |
| 6 | """ |
| 7 | |
| 8 | from typing import Optional, Tuple, Union |
| 9 | from transformers.modeling_outputs import BaseModelOutput |
| 10 | import torch |
| 11 | from torch import nn |
| 12 | |
| 13 | from transformers.models.hubert.modeling_hubert import ( |
| 14 | HubertFeatureEncoder, |
| 15 | HubertModel, |
| 16 | HubertEncoderStableLayerNorm, |
| 17 | HubertEncoder, |
| 18 | HubertEncoderLayer, |
| 19 | HubertPositionalConvEmbedding, |
| 20 | HubertAttention, |
| 21 | HubertFeedForward, |
| 22 | ) |
| 23 | |
| 24 | try: |
| 25 | from nnAudio import features as nnAudioFeatures |
| 26 | NNAUDIO_INSTALLED=True |
| 27 | except: |
| 28 | print("WARNING: feature_extractor_cqt requires the libray 'nnAudio'") |
| 29 | NNAUDIO_INSTALLED=False |
| 30 | |
| 31 | from .configuration_MERT import MERTConfig |
| 32 | |
| 33 | class MERTFeatureProjection(nn.Module): |
| 34 | def __init__(self, config): |
| 35 | super().__init__() |
| 36 | self.feat_proj_layer_norm = config.feat_proj_layer_norm |
| 37 | self.feature_extractor_cqt = config.feature_extractor_cqt |
| 38 | |
| 39 | if self.feature_extractor_cqt: |
| 40 | # v3 concat features |
| 41 | self.feature_dimension = config.conv_dim[-1] + config.feature_extractor_cqt_bins |
| 42 | print(f"feature dimention: {self.feature_dimension}") |
| 43 | else: |
| 44 | self.feature_dimension = config.conv_dim[-1] |
| 45 | if self.feat_proj_layer_norm: |
| 46 | self.layer_norm = nn.LayerNorm(self.feature_dimension, eps=config.layer_norm_eps) |
| 47 | self.projection = nn.Linear(self.feature_dimension, config.hidden_size) |
| 48 | self.dropout = nn.Dropout(config.feat_proj_dropout) |
| 49 | |
| 50 | def forward(self, hidden_states): |
| 51 | # non-projected hidden states are needed for quantization |
| 52 | if self.feat_proj_layer_norm: |
| 53 | hidden_states = self.layer_norm(hidden_states) |
| 54 | hidden_states = self.projection(hidden_states) |
| 55 | hidden_states = self.dropout(hidden_states) |
| 56 | return hidden_states |
| 57 | |
| 58 | class MERTModel(HubertModel): |
| 59 | # overwrite config class |
| 60 | config_class = MERTConfig |
| 61 | base_model_prefix = "mert_model" |
| 62 | def __init__( |
| 63 | self, |
| 64 | config: MERTConfig, |
| 65 | ) -> None: |
| 66 | """ |
| 67 | initialize the with the grandparent method HubertPreTrainedModel.__init__() |
| 68 | and modify the HuBERTModel.__init__() |
| 69 | """ |
| 70 | super(HubertModel, self).__init__(config) |
| 71 | |
| 72 | self.config = config |
| 73 | |
| 74 | self.feature_extractor = HubertFeatureEncoder(config) |
| 75 | self.feature_projection = MERTFeatureProjection(config) # replace Feature Projection for introcuing new feature |
| 76 | |
| 77 | if self.config.feature_extractor_cqt: |
| 78 | assert NNAUDIO_INSTALLED, "ERROR: feature_extractor_cqt requires the libray 'nnAudio', try after `pip install nnAudio` " |
| 79 | print('initializing cqt extractor for MERT') |
| 80 | self.feature_extractor_cqt = nnAudioFeatures.cqt.CQT(sr=self.config.sample_rate, hop_length=self.config.sample_rate//50, fmin=32.7, |
| 81 | fmax=None, n_bins=self.config.feature_extractor_cqt_bins, bins_per_octave=self.config.feature_extractor_cqt_bins//7, |
| 82 | filter_scale=1, norm=1, window='hann', center=True, |
| 83 | pad_mode='constant', trainable=False, |
| 84 | output_format='Magnitude', verbose=True) |
| 85 | |
| 86 | if config.mask_time_prob > 0.0 or config.mask_feature_prob > 0.0: |
| 87 | self.masked_spec_embed = nn.Parameter(torch.FloatTensor(config.hidden_size).uniform_()) |
| 88 | |
| 89 | |
| 90 | if config.do_stable_layer_norm: |
| 91 | assert not config.deepnorm, "must use post-layer_norm with deepnorm" |
| 92 | self.encoder = HubertEncoderStableLayerNorm(config) |
| 93 | else: |
| 94 | if config.deepnorm: |
| 95 | self.encoder = HubertEncoder_extend(config) |
| 96 | else: |
| 97 | self.encoder = HubertEncoder(config) |
| 98 | |
| 99 | # Initialize weights and apply final processing |
| 100 | self.post_init() |
| 101 | |
| 102 | def forward(self, input_values: Optional[torch.Tensor], attention_mask: Optional[torch.Tensor] = None, mask_time_indices: Optional[torch.FloatTensor] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None) -> Union[Tuple, BaseModelOutput]: |
| 103 | |
| 104 | # return super().forward(input_values, attention_mask, mask_time_indices, output_attentions, output_hidden_states, return_dict) |
| 105 | |
| 106 | output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions |
| 107 | output_hidden_states = ( |
| 108 | output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states |
| 109 | ) |
| 110 | return_dict = return_dict if return_dict is not None else self.config.use_return_dict |
| 111 | |
| 112 | extract_features = self.feature_extractor(input_values) |
| 113 | extract_features = extract_features.transpose(1, 2) |
| 114 | |
| 115 | # add additional cqt features for transformer input |
| 116 | if self.config.feature_extractor_cqt: |
| 117 | features_cqt = self.feature_extractor_cqt(input_values).transpose(1, 2) |
| 118 | features_cqt = features_cqt[:,:extract_features.shape[1],:] # align shape |
| 119 | # # v2 |
| 120 | # features_cqt = self.post_cqt_feature_proj(features_cqt) |
| 121 | # extract_features = self.feature_projection.layer_norm(extract_features) + self.feature_projection.layer_norm(features_cqt) #v2 |
| 122 | # v3 |
| 123 | extract_features = torch.cat([extract_features,features_cqt], 2) |
| 124 | |
| 125 | if attention_mask is not None: |
| 126 | # compute reduced attention_mask corresponding to feature vectors |
| 127 | attention_mask = self._get_feature_vector_attention_mask(extract_features.shape[1], attention_mask) |
| 128 | |
| 129 | hidden_states = self.feature_projection(extract_features) |
| 130 | hidden_states = self._mask_hidden_states(hidden_states, mask_time_indices=mask_time_indices) |
| 131 | |
| 132 | encoder_outputs = self.encoder( |
| 133 | hidden_states, |
| 134 | attention_mask=attention_mask, |
| 135 | output_attentions=output_attentions, |
| 136 | output_hidden_states=output_hidden_states, |
| 137 | return_dict=return_dict, |
| 138 | ) |
| 139 | |
| 140 | hidden_states = encoder_outputs[0] # take last_hidden from encoder output |
| 141 | |
| 142 | if not return_dict: |
| 143 | return (hidden_states,) + encoder_outputs[1:] |
| 144 | |
| 145 | return BaseModelOutput( |
| 146 | last_hidden_state=hidden_states, |
| 147 | hidden_states=encoder_outputs.hidden_states, |
| 148 | attentions=encoder_outputs.attentions, |
| 149 | ) |
| 150 | |
| 151 | |
| 152 | class HubertEncoder_extend(HubertEncoder): |
| 153 | def __init__(self, config): |
| 154 | # super().__init__() |
| 155 | # call nn module initialization |
| 156 | nn.Module.__init__(self) |
| 157 | # super(HubertEncoder_extend, self).__init__() |
| 158 | |
| 159 | self.config = config |
| 160 | self.pos_conv_embed = HubertPositionalConvEmbedding(config) |
| 161 | self.layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) |
| 162 | self.dropout = nn.Dropout(config.hidden_dropout) |
| 163 | |
| 164 | |
| 165 | self.layers = nn.ModuleList([HubertEncoderLayerExtend(config) for _ in range(config.num_hidden_layers)]) |
| 166 | |
| 167 | self.gradient_checkpointing = False |
| 168 | |
| 169 | if config.deepnorm: |
| 170 | import math |
| 171 | init_scale = math.pow(8.0 * config.num_hidden_layers, 0.25) |
| 172 | for name, p in self.named_parameters(): |
| 173 | if ( |
| 174 | "feed_forward.intermediate_dense" in name |
| 175 | or "feed_forward.output_dense" in name |
| 176 | or "out_proj" in name |
| 177 | or "v_proj" in name |
| 178 | ): |
| 179 | p.data.div_(init_scale) |
| 180 | |
| 181 | class HubertEncoderLayerExtend(HubertEncoderLayer): |
| 182 | def __init__(self, config): |
| 183 | nn.Module.__init__(self) |
| 184 | # super(HubertEncoderLayerExtend, self).__init__() |
| 185 | if config.attention_relax > 0 : |
| 186 | self.attention = HubertAttention_extend( |
| 187 | embed_dim=config.hidden_size, |
| 188 | num_heads=config.num_attention_heads, |
| 189 | dropout=config.attention_dropout, |
| 190 | is_decoder=False, |
| 191 | attention_relax=config.attention_relax, |
| 192 | ) |
| 193 | else: |
| 194 | self.attention = HubertAttention( |
| 195 | embed_dim=config.hidden_size, |
| 196 | num_heads=config.num_attention_heads, |
| 197 | dropout=config.attention_dropout, |
| 198 | is_decoder=False, |
| 199 | ) |
| 200 | self.dropout = nn.Dropout(config.hidden_dropout) |
| 201 | self.layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) |
| 202 | self.feed_forward = HubertFeedForward(config) |
| 203 | self.final_layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) |
| 204 | |
| 205 | if config.deepnorm: |
| 206 | import math |
| 207 | self.residual_alpha = math.pow(2.0 * config.num_hidden_layers, 0.25) |
| 208 | else: |
| 209 | self.residual_alpha = 1.0 |
| 210 | |
| 211 | def residual_connection(self, x, residual): |
| 212 | ''' |
| 213 | residual: input before f() |
| 214 | x: output of f(residual) |
| 215 | ''' |
| 216 | return residual * self.residual_alpha + x |
| 217 | |
| 218 | def forward(self, hidden_states, attention_mask=None, output_attentions=False): |
| 219 | attn_residual = hidden_states |
| 220 | hidden_states, attn_weights, _ = self.attention( |
| 221 | hidden_states, attention_mask=attention_mask, output_attentions=output_attentions |
| 222 | ) |
| 223 | hidden_states = self.dropout(hidden_states) |
| 224 | |
| 225 | # hidden_states = attn_residual + hidden_states |
| 226 | hidden_states = self.residual_connection(hidden_states, attn_residual) |
| 227 | |
| 228 | hidden_states = self.layer_norm(hidden_states) |
| 229 | |
| 230 | # hidden_states = hidden_states + self.feed_forward(hidden_states) |
| 231 | ffn_residual = hidden_states |
| 232 | hidden_states = self.feed_forward(hidden_states) |
| 233 | hidden_states = self.residual_connection(hidden_states, ffn_residual) |
| 234 | |
| 235 | hidden_states = self.final_layer_norm(hidden_states) |
| 236 | |
| 237 | outputs = (hidden_states,) |
| 238 | |
| 239 | if output_attentions: |
| 240 | outputs += (attn_weights,) |
| 241 | |
| 242 | return outputs |
| 243 | |
| 244 | |
| 245 | class HubertAttention_extend(nn.Module): |
| 246 | def __init__( |
| 247 | self, |
| 248 | embed_dim: int, |
| 249 | num_heads: int, |
| 250 | dropout: float = 0.0, |
| 251 | is_decoder: bool = False, |
| 252 | bias: bool = True, |
| 253 | attention_relax: float = -1.0, |
| 254 | ): |
| 255 | super().__init__() |
| 256 | # nn.Module.__init__(self) |
| 257 | self.embed_dim = embed_dim |
| 258 | self.num_heads = num_heads |
| 259 | self.dropout = dropout |
| 260 | self.head_dim = embed_dim // num_heads |
| 261 | |
| 262 | if (self.head_dim * num_heads) != self.embed_dim: |
| 263 | raise ValueError( |
| 264 | f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim}" |
| 265 | f" and `num_heads`: {num_heads})." |
| 266 | ) |
| 267 | self.scaling = self.head_dim**-0.5 |
| 268 | self.is_decoder = is_decoder |
| 269 | |
| 270 | self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias) |
| 271 | self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias) |
| 272 | self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias) |
| 273 | self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias) |
| 274 | |
| 275 | if attention_relax > 0: |
| 276 | self.attention_relax = attention_relax |
| 277 | |
| 278 | def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int): |
| 279 | return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous() |
| 280 | |
| 281 | def forward( |
| 282 | self, |
| 283 | hidden_states: torch.Tensor, |
| 284 | key_value_states: Optional[torch.Tensor] = None, |
| 285 | past_key_value: Optional[Tuple[torch.Tensor]] = None, |
| 286 | attention_mask: Optional[torch.Tensor] = None, |
| 287 | layer_head_mask: Optional[torch.Tensor] = None, |
| 288 | output_attentions: bool = False, |
| 289 | ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: |
| 290 | """Input shape: Batch x Time x Channel""" |
| 291 | |
| 292 | # if key_value_states are provided this layer is used as a cross-attention layer |
| 293 | # for the decoder |
| 294 | is_cross_attention = key_value_states is not None |
| 295 | |
| 296 | bsz, tgt_len, _ = hidden_states.size() |
| 297 | |
| 298 | # get query proj |
| 299 | query_states = self.q_proj(hidden_states) * self.scaling |
| 300 | # get key, value proj |
| 301 | # `past_key_value[0].shape[2] == key_value_states.shape[1]` |
| 302 | # is checking that the `sequence_length` of the `past_key_value` is the same as |
| 303 | # the provided `key_value_states` to support prefix tuning |
| 304 | if ( |
| 305 | is_cross_attention |
| 306 | and past_key_value is not None |
| 307 | and past_key_value[0].shape[2] == key_value_states.shape[1] |
| 308 | ): |
| 309 | # reuse k,v, cross_attentions |
| 310 | key_states = past_key_value[0] |
| 311 | value_states = past_key_value[1] |
| 312 | elif is_cross_attention: |
| 313 | # cross_attentions |
| 314 | key_states = self._shape(self.k_proj(key_value_states), -1, bsz) |
| 315 | value_states = self._shape(self.v_proj(key_value_states), -1, bsz) |
| 316 | elif past_key_value is not None: |
| 317 | # reuse k, v, self_attention |
| 318 | key_states = self._shape(self.k_proj(hidden_states), -1, bsz) |
| 319 | value_states = self._shape(self.v_proj(hidden_states), -1, bsz) |
| 320 | key_states = torch.cat([past_key_value[0], key_states], dim=2) |
| 321 | value_states = torch.cat([past_key_value[1], value_states], dim=2) |
| 322 | else: |
| 323 | # self_attention |
| 324 | key_states = self._shape(self.k_proj(hidden_states), -1, bsz) |
| 325 | value_states = self._shape(self.v_proj(hidden_states), -1, bsz) |
| 326 | |
| 327 | if self.is_decoder: |
| 328 | # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states. |
| 329 | # Further calls to cross_attention layer can then reuse all cross-attention |
| 330 | # key/value_states (first "if" case) |
| 331 | # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of |
| 332 | # all previous decoder key/value_states. Further calls to uni-directional self-attention |
| 333 | # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case) |
| 334 | # if encoder bi-directional self-attention `past_key_value` is always `None` |
| 335 | past_key_value = (key_states, value_states) |
| 336 | |
| 337 | proj_shape = (bsz * self.num_heads, -1, self.head_dim) |
| 338 | query_states = self._shape(query_states, tgt_len, bsz).view(*proj_shape) |
| 339 | key_states = key_states.view(*proj_shape) |
| 340 | value_states = value_states.view(*proj_shape) |
| 341 | |
| 342 | src_len = key_states.size(1) |
| 343 | attn_weights = torch.bmm(query_states, key_states.transpose(1, 2)) |
| 344 | |
| 345 | if attn_weights.size() != (bsz * self.num_heads, tgt_len, src_len): |
| 346 | raise ValueError( |
| 347 | f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is" |
| 348 | f" {attn_weights.size()}" |
| 349 | ) |
| 350 | |
| 351 | if attention_mask is not None: |
| 352 | if attention_mask.size() != (bsz, 1, tgt_len, src_len): |
| 353 | raise ValueError( |
| 354 | f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}" |
| 355 | ) |
| 356 | attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attention_mask |
| 357 | attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len) |
| 358 | |
| 359 | if self.attention_relax > 0: |
| 360 | # => (bsz, self.num_heads, tgt_len, src_len) |
| 361 | # attn_weights_relax = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)/self.attention_relax |
| 362 | # => (bsz*self.num_heads, tgt_len, src_len) |
| 363 | attn_weights_relax = attn_weights / self.attention_relax |
| 364 | |
| 365 | # => (bsz* self.num_heads, tgt_len, 1) |
| 366 | attn_max_relax = torch.max(attn_weights_relax, dim=-1, keepdim=False).unsqueeze(2) |
| 367 | attn_weights = (attn_weights_relax - attn_max_relax) * self.attention_relax |
| 368 | |
| 369 | attn_weights = nn.functional.softmax(attn_weights, dim=-1) |
| 370 | |
| 371 | if layer_head_mask is not None: |
| 372 | if layer_head_mask.size() != (self.num_heads,): |
| 373 | raise ValueError( |
| 374 | f"Head mask for a single layer should be of size {(self.num_heads,)}, but is" |
| 375 | f" {layer_head_mask.size()}" |
| 376 | ) |
| 377 | attn_weights = layer_head_mask.view(1, -1, 1, 1) * attn_weights.view(bsz, self.num_heads, tgt_len, src_len) |
| 378 | attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len) |
| 379 | |
| 380 | if output_attentions: |
| 381 | # this operation is a bit awkward, but it's required to |
| 382 | # make sure that attn_weights keeps its gradient. |
| 383 | # In order to do so, attn_weights have to be reshaped |
| 384 | # twice and have to be reused in the following |
| 385 | attn_weights_reshaped = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) |
| 386 | attn_weights = attn_weights_reshaped.view(bsz * self.num_heads, tgt_len, src_len) |
| 387 | else: |
| 388 | attn_weights_reshaped = None |
| 389 | |
| 390 | attn_probs = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training) |
| 391 | |
| 392 | attn_output = torch.bmm(attn_probs, value_states) |
| 393 | |
| 394 | if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim): |
| 395 | raise ValueError( |
| 396 | f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is" |
| 397 | f" {attn_output.size()}" |
| 398 | ) |
| 399 | |
| 400 | attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim) |
| 401 | attn_output = attn_output.transpose(1, 2) |
| 402 | |
| 403 | # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be |
| 404 | # partitioned aross GPUs when using tensor-parallelism. |
| 405 | attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim) |
| 406 | |
| 407 | attn_output = self.out_proj(attn_output) |
| 408 | |
| 409 | return attn_output, attn_weights_reshaped, past_key_value |
| 410 | |