File size: 10,472 Bytes
fdc1efd |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 |
from warnings import warn
import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers import ASTConfig, ASTModel
class StudentAST(nn.Module):
"""
A student model for audio classification using the AST architecture.
:param n_classes: The number of classes to classify.
:type n_classes: int
:param hidden_size: The number of units in the hidden layers, defaults to 384.
:type hidden_size: int, optional
:param num_heads: The number of attention heads to use, defaults to 6.
:type num_heads: int, optional
"""
def __init__(self, n_classes: int, hidden_size: int = 384, num_heads: int = 6):
super().__init__()
config = ASTConfig(hidden_size=hidden_size, num_attention_heads=num_heads, intermediate_size=hidden_size * 4)
self.base_model = ASTModel(config=config)
self.classifier = StudentClassificationHead(hidden_size, n_classes)
def forward(self, x: torch.Tensor):
"""
Forward pass of the student model.
:param x: The input tensor of shape [batch_size, sequence_length, input_dim].
:type x: torch.Tensor
:return: The output tensor of shape [batch_size, n_classes].
:rtype: torch.Tensor
"""
x = self.base_model(x)[0]
x = self.classifier(x)
return x
class StudentClassificationHead(nn.Module):
"""
A classification head for the student model.
:param emb_size: The size of the embedding.
:type emb_size: int
:param n_classes: The number of classes to classify.
:type n_classes: int
"""
def __init__(self, emb_size: int, n_classes: int):
super().__init__()
self.cls_head = nn.Linear(emb_size, n_classes)
self.dist_head = nn.Linear(emb_size, n_classes)
def forward(self, x: torch.Tensor):
"""
Forward pass of the classification head.
:param x: The input tensor of shape [batch_size, emb_size*2].
:type x: torch.Tensor
:return: The output tensor of shape [batch_size, n_classes].
:rtype: torch.Tensor
"""
x_cls, x_dist = x[:, 0], x[:, 1]
x_cls_head = self.cls_head(x_cls)
x_dist_head = self.dist_head(x_dist)
if self.training:
x = x_cls_head, x_dist_head
else:
x = (x_cls_head + x_dist_head) / 2
return x
class ASTPretrained(nn.Module):
"""
This class implements a PyTorch module for a pre-trained Audio Set Transformer (AST) model
fine-tuned on MIT's dataset for audio event classification.
:param n_classes: The number of classes for audio event classification.
:type n_classes: int
:param dropout: The dropout probability for the fully connected layer, defaults to 0.5.
:type dropout: float, optional
:raises ValueError: If n_classes is not positive.
:raises TypeError: If dropout is not a float or is not between 0 and 1.
:return: The output tensor of shape [batch_size, n_classes] containing the probabilities of each class.
:rtype: torch.Tensor
"""
def __init__(self, n_classes: int, download_weights: bool = True, freeze_body: bool = False, dropout: float = 0.5):
super().__init__()
if download_weights:
self.base_model = ASTModel.from_pretrained("MIT/ast-finetuned-audioset-10-10-0.4593")
else:
config = ASTConfig()
self.base_model = ASTModel(config=config)
if freeze_body:
self.base_model = freeze(self.base_model)
fc_in = self.base_model.config.hidden_size
self.classifier = nn.Sequential(
nn.LayerNorm((fc_in,), eps=1e-12), nn.Dropout(p=dropout), nn.Linear(fc_in, n_classes)
)
def forward(self, x):
"""Passes the input tensor through the pre-trained Audio Set Transformer (AST) model
followed by a fully connected layer.
:param x: The input tensor of shape [batch_size, seq_len, num_features].
:type x: torch.Tensor
:return: The output tensor of shape [batch_size, n_classes] containing the probabilities of each class.
:rtype: torch.Tensor
:raises ValueError: If the shape of x is not [batch_size, seq_len, num_features].
"""
x = self.base_model(x)[1]
x = self.classifier(x)
return x
def layerwise_lr_decay(config, model: ASTModel):
"""
LLRD (Layer-wise Learning Rate Decay) function computes the learning rate for each layer in a deep neural network
using a specific decay rate and a base learning rate for the optimizer.
:param config: A configuration object that contains the parameters required for LLRD.
:type config: Any
:param model: A PyTorch neural network model.
:type model: ASTModel
:raises Warning: If the configuration object does not contain the LLRD parameters.
:return: A dictionary containing the optimizer parameters (parameters, weight decay, and learning rate)
for each layer.
:rtype: dict
"""
try:
config = config.LLRD
except Exception:
warn("No LLRD found in config. Learner will use single lr for whole model.")
return None
lr = config["base_lr"]
weight_decay = config["weight_decay"]
no_decay = ["bias", "layernorm"]
body = ["embeddings", "encoder.layer"]
head_params = [(n, p) for n, p in model.named_parameters() if not any(body_param in n for body_param in body)]
optimizer_grouped_parameters = [
{
"params": [p for n, p in head_params if not any(nd in n for nd in no_decay)],
"weight_decay": weight_decay,
"lr": lr,
},
{
"params": [p for n, p in head_params if any(nd in n for nd in no_decay)],
"weight_decay": 0.0,
"lr": lr,
},
]
# initialize lrs for every layer
layers = [getattr(model.module, config["body"]).embeddings] + list(
getattr(model.module, config["body"]).encoder.layer
)
layers.reverse()
for layer in layers:
lr *= config["lr_decay_rate"]
optimizer_grouped_parameters += [
{
"params": [p for n, p in layer.named_parameters() if not any(nd in n for nd in no_decay)],
"weight_decay": weight_decay,
"lr": lr,
},
{
"params": [p for n, p in layer.named_parameters() if any(nd in n for nd in no_decay)],
"weight_decay": 0.0,
"lr": lr,
},
]
return optimizer_grouped_parameters
def freeze(model: nn.Module):
"""
Freeze function sets the requires_grad attribute to False for all parameters
in the given PyTorch neural network model. This is used to freeze the weights of
the model during training or inference.
:param model: A PyTorch neural network model.
:type model: nn.Module
:return: The same model with requires_grad attribute set to False for all parameters.
:rtype: nn.Module
"""
model.eval()
for param in model.parameters():
param.requires_grad = False
return model
def unfreeze(model: nn.Module):
"""
Unfreeze the model by setting requires_grad to True for all parameters.
:param model: The model to unfreeze.
:type model: nn.Module
:return: The unfrozen model.
:rtype: nn.Module
"""
model.train()
for param in model.parameters():
param.requires_grad = True
return model
def interpolate_params(student: nn.Module, teacher: nn.Module):
"""
Interpolate parameters between two models. This function scales the parameters of the
teacher model to match the shape of the corresponding parameters in the student model
using bilinear interpolation. If the shapes of the parameters in the two models are already the same,
the parameters are unchanged.
:param student: The student model.
:type student: nn.Module
:param teacher: The teacher model.
:type teacher: nn.Module
:return: A dictionary of interpolated parameters for the student model.
:rtype: dict
"""
new_params = {}
# Iterate over the parameters in the first model
for name, param in teacher.base_model.named_parameters():
# Scale the parameter using interpolate if its shape is different from that of the second model
target_param = student.base_model.state_dict()[name]
if param.shape != target_param.shape:
squeeze_count = 0
permuted = False
while param.ndim < 4:
param = param.unsqueeze(0)
squeeze_count += 1
if param.shape[0] > 1:
param = param.permute(1, 2, 3, 0)
target_param = target_param.permute(1, 2, 3, 0)
permuted = True
if target_param.ndim < 2:
target_param = target_param.unsqueeze(0)
scaled_param = F.interpolate(param, size=(target_param.shape[-2:]), mode="bilinear")
while squeeze_count > 0:
scaled_param = scaled_param.squeeze(0)
squeeze_count -= 1
if permuted:
scaled_param = scaled_param.permute(-1, 0, 1, 2)
else:
scaled_param = param
new_params[name] = scaled_param
return new_params
def average_model_weights(model_weights_list):
"""
Compute the average weights of a list of PyTorch models.
:param model_weights_list: A list of file paths to PyTorch model weight files.
:type model_weights_list: List[str]
:raises ValueError: If the input list is empty.
:return: A dictionary containing the average weights of the models.
:rtype: Dict[str, torch.Tensor]
"""
if not model_weights_list:
raise ValueError("The input list cannot be empty.")
num_models = len(model_weights_list)
averaged_weights = {}
# Load the first model weights
state_dict = torch.load(model_weights_list[0])
# Iterate through the remaining models and add their weights to the first model's weights
for i in range(1, num_models):
state_dict_i = torch.load(model_weights_list[i])
for key in state_dict.keys():
state_dict[key] += state_dict_i[key]
# Compute the average of the weights
for key in state_dict.keys():
averaged_weights[key] = state_dict[key] / num_models
return averaged_weights
|