File size: 6,490 Bytes
0015cb8 |
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 |
import math
import torch
import torch.nn as nn
from transformers import PreTrainedModel
from typing import List
from .config import LidirlCNNConfig
def torch_max_no_pads(model_out, lengths):
indices = torch.arange(model_out.size(1)).to(model_out.device)
mask = (indices < lengths.view(-1, 1)).unsqueeze(-1).expand(model_out.size())
model_out = torch.where(mask, model_out, torch.tensor(-1e9))
max_pool = torch.max(model_out, 1)[0]
return max_pool
class TransposeModule(nn.Module):
def __init__(self):
super().__init__()
def forward(self, x):
return x.transpose(1, 2)
class ProjectionLayer(nn.Module):
"""
Noise-aware labels layer or traditional linear projection
"""
def __init__(self,
hidden_dim : int,
label_size : int,
montecarlo_layer : bool = False):
super().__init__()
self.montecarlo_layer = montecarlo_layer
if montecarlo_layer:
self.proj = MCSoftmaxDenseFA(hidden_dim, label_size, 1, logits_only=True)
else:
# eventually I'm considering making this a variable size classifier so it's a sequence of layers even though it's just one
self.projs = [
nn.Linear(hidden_dim, label_size)
]
self.proj = nn.Sequential(*self.projs)
self.init_layer()
def forward(self, x):
return self.proj(x)
def init_layer(self, pi : float = 0.01):
"""
Initialize the final classification layer so all predictions are close to 0
"""
if self.montecarlo_layer:
self.proj.init_layer(pi)
return
for layer in self.proj.modules():
if isinstance(layer, nn.Linear):
nn.init.normal_(layer.weight, mean=0.0, std=0.01)
if layer.bias is not None:
# Bias: -log((1 - pi) / pi)
bias_value = -math.log((1 - pi) / pi)
nn.init.constant_(layer.bias, bias_value)
class ConvolutionalBlock(
nn.Module,
):
"""
Convolutional block
https://jonathanbgn.com/2021/09/30/illustrated-wav2vec-2.html
"""
def __init__(self,
embed_dim : int,
channels : List[int],
kernels : List[int],
strides : List[int]):
super(ConvolutionalBlock, self).__init__()
layers = []
self.embed_dim = embed_dim
input_dimension = embed_dim
for channel, kernel, stride in zip(channels, kernels, strides):
next_layer = nn.Conv1d(
in_channels = input_dimension,
out_channels = channel,
kernel_size = kernel,
stride = stride,
padding = 'valid', # we handle the padding ourselves in the forward function
)
input_dimension = channel
layers.append(TransposeModule())
layers.append(next_layer)
layers.append(TransposeModule())
layers.append(nn.LayerNorm(channel, elementwise_affine=True))
layers.append(nn.GELU())
layers.append(nn.Dropout(0.1))
self.model = nn.Sequential(*layers)
self.output_dim = channels[-1]
self.min_length = 1
for kernel, stride in zip(kernels[::-1], strides[::-1]):
self.min_length = ((self.min_length - 1) * stride) + kernel
def forward(self, inputs, lengths):
# this is our padding trick instead of consistent padding
if inputs.size(1) < self.min_length:
pads = torch.zeros((inputs.size(0), self.min_length - inputs.size(1), self.embed_dim), device=inputs.device)
inputs = torch.cat((inputs, pads), dim=1)
outputs = self.model(inputs)
for layer_i in range(1, len(self.model), 6):
lengths = torch.floor(((lengths - self.model[layer_i].kernel_size[0]) / self.model[layer_i].stride[0]) + 1).to(lengths.device, dtype=torch.long)
lengths[lengths < 1] = 1
return outputs, lengths
class LidirlCNN(PreTrainedModel):
"""
Defines the Lidirl CNN MODEL
"""
config_class = LidirlCNNConfig
def __init__(self, config):
super().__init__(config)
self.encoder = ConvolutionalBlock(config.embed_dim, config.channels, config.kernels, config.strides)
self.embed_layer = nn.Embedding(config.vocab_size, config.embed_dim)
self.proj = ProjectionLayer(self.encoder.output_dim, config.label_size, config.montecarlo_layer)
self.label_size = config.label_size
self.max_length = config.max_length
self.multilabel = config.multilabel
self.monte_carlo = config.montecarlo_layer
self.labels = ["" for _ in config.labels]
for key, value in config.labels.items():
self.labels[value] = key
def forward(self, inputs, lengths):
inputs = inputs[:, :self.max_length]
lengths = lengths.clamp(max=self.max_length)
embeddings = self.embed_layer(inputs)
encoding, lengths = self.encoder(embeddings, lengths=lengths)
max_pool = torch_max_no_pads(encoding, lengths)
projection = self.proj(max_pool)
return projection
def __call__(self, inputs, lengths):
# this is inference only model
with torch.no_grad():
logits = self.forward(inputs, lengths)
if self.multilabel:
probs = torch.sigmoid(logits)
else:
probs = torch.softmax(logits, dim=-1)
return probs
def predict(self, inputs, lengths, threshold=0.5, top_k=None):
probs = self.__call__(inputs, lengths)
if top_k is not None and top_k > 0:
top_k_preds = torch.topk(probs, top_k, dim=1)
pred_labels = []
for pred, prob in zip(top_k_preds.indices, top_k_preds.values):
pred_labels.append([(self.labels[p.item()], pr.item()) for (p, pr) in zip(pred, prob)])
return pred_labels
if self.multilabel:
batch_idx, label_idx = torch.where(probs > threshold)
output = [[] for _ in range(len(inputs))]
for batch, label in zip(batch_idx, label_idx):
label_string = self.labels
output[batch.item()].append(
(self.labels[label.item()], probs[batch, label])
)
return output
|