Spaces:
Running
Running
File size: 8,855 Bytes
29f689c |
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 |
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
class SAREncoder(nn.Module):
def __init__(self,
enc_bi_rnn=False,
enc_drop_rnn=0.1,
in_channels=512,
d_enc=512,
**kwargs):
super().__init__()
# LSTM Encoder
if enc_bi_rnn:
bidirectional = True
else:
bidirectional = False
hidden_size = d_enc
self.rnn_encoder = nn.LSTM(input_size=in_channels,
hidden_size=hidden_size,
num_layers=2,
dropout=enc_drop_rnn,
bidirectional=bidirectional,
batch_first=True)
# global feature transformation
encoder_rnn_out_size = hidden_size * (int(enc_bi_rnn) + 1)
self.linear = nn.Linear(encoder_rnn_out_size, encoder_rnn_out_size)
def forward(self, feat):
h_feat = feat.shape[2]
feat_v = F.max_pool2d(feat,
kernel_size=(h_feat, 1),
stride=1,
padding=0)
feat_v = feat_v.squeeze(2)
feat_v = feat_v.permute(0, 2, 1).contiguous() # bsz * W * C
holistic_feat = self.rnn_encoder(feat_v)[0] # bsz * T * hidden_size
valid_hf = holistic_feat[:, -1, :] # bsz * hidden_size
holistic_feat = self.linear(valid_hf) # bsz * C
return holistic_feat
class SARDecoder(nn.Module):
def __init__(self,
in_channels,
out_channels,
max_len=25,
enc_bi_rnn=False,
enc_drop_rnn=0.1,
dec_bi_rnn=False,
dec_drop_rnn=0.0,
pred_dropout=0.1,
pred_concat=True,
mask=True,
use_lstm=True,
**kwargs):
super(SARDecoder, self).__init__()
self.num_classes = out_channels
self.start_idx = out_channels - 2
self.padding_idx = out_channels - 1
self.end_idx = 0
self.max_seq_len = max_len + 1
self.pred_concat = pred_concat
self.mask = mask
enc_dim = in_channels
d = in_channels
embedding_dim = in_channels
dec_dim = in_channels
self.use_lstm = use_lstm
if use_lstm:
# encoder module
self.encoder = SAREncoder(enc_bi_rnn=enc_bi_rnn,
enc_drop_rnn=enc_drop_rnn,
in_channels=in_channels,
d_enc=enc_dim)
# decoder module
# 2D attention layer
self.conv1x1_1 = nn.Linear(dec_dim, d)
self.conv3x3_1 = nn.Conv2d(in_channels,
d,
kernel_size=3,
stride=1,
padding=1)
self.conv1x1_2 = nn.Linear(d, 1)
# Decoder input embedding
self.embedding = nn.Embedding(self.num_classes,
embedding_dim,
padding_idx=self.padding_idx)
self.rnndecoder = nn.LSTM(input_size=embedding_dim,
hidden_size=dec_dim,
num_layers=2,
dropout=dec_drop_rnn,
bidirectional=dec_bi_rnn,
batch_first=True)
# Prediction layer
self.pred_dropout = nn.Dropout(pred_dropout)
if pred_concat:
fc_in_channel = in_channels + in_channels + dec_dim
else:
fc_in_channel = in_channels
self.prediction = nn.Linear(fc_in_channel, self.num_classes)
self.softmax = nn.Softmax(dim=-1)
def _2d_attation(self, feat, tokens, data, training):
Hidden_state = self.rnndecoder(tokens)[0]
attn_query = self.conv1x1_1(Hidden_state)
bsz, seq_len, _ = attn_query.size()
attn_query = attn_query.unsqueeze(-1).unsqueeze(-1)
# bsz * seq_len+1 * attn_size * 1 * 1
attn_key = self.conv3x3_1(feat).unsqueeze(1)
# bsz * 1 * attn_size * h * w
attn_weight = torch.tanh(torch.add(attn_key, attn_query, alpha=1))
attn_weight = attn_weight.permute(0, 1, 3, 4, 2).contiguous()
attn_weight = self.conv1x1_2(attn_weight)
_, T, h, w, c = attn_weight.size()
if self.mask:
valid_ratios = data[-1]
# cal mask of attention weight
attn_mask = torch.zeros_like(attn_weight)
for i, valid_ratio in enumerate(valid_ratios):
valid_width = min(w, math.ceil(w * valid_ratio))
attn_mask[i, :, :, valid_width:, :] = 1
attn_weight = attn_weight.masked_fill(attn_mask.bool(),
float('-inf'))
attn_weight = attn_weight.view(bsz, T, -1)
attn_weight = F.softmax(attn_weight, dim=-1)
attn_weight = attn_weight.view(bsz, T, h, w,
c).permute(0, 1, 4, 2, 3).contiguous()
# bsz, T, 1, h, w
# bsz, 1, f_c ,h, w
attn_feat = torch.sum(torch.mul(feat.unsqueeze(1), attn_weight),
(3, 4),
keepdim=False)
return [Hidden_state, attn_feat]
def forward_train(self, feat, holistic_feat, data):
max_len = data[1].max()
label = data[0][:, :1 + max_len] # label
label_embedding = self.embedding(label)
holistic_feat = holistic_feat.unsqueeze(1)
tokens = torch.cat((holistic_feat, label_embedding), dim=1)
Hidden_state, attn_feat = self._2d_attation(feat,
tokens,
data,
training=self.training)
bsz, seq_len, f_c = Hidden_state.size()
# linear transformation
if self.pred_concat:
f_c = holistic_feat.size(-1)
holistic_feat = holistic_feat.expand(bsz, seq_len, f_c)
preds = self.prediction(
torch.cat((Hidden_state, attn_feat, holistic_feat), 2))
else:
preds = self.prediction(attn_feat)
# bsz * (seq_len + 1) * num_classes
preds = self.pred_dropout(preds)
return preds[:, 1:, :]
def forward_test(self, feat, holistic_feat, data=None):
bsz = feat.shape[0]
seq_len = self.max_seq_len
holistic_feat = holistic_feat.unsqueeze(1)
tokens = torch.full((bsz, ),
self.start_idx,
device=feat.device,
dtype=torch.long)
outputs = []
tokens = self.embedding(tokens)
tokens = tokens.unsqueeze(1).expand(-1, seq_len, -1)
tokens = torch.cat((holistic_feat, tokens), dim=1)
for i in range(1, seq_len + 1):
Hidden_state, attn_feat = self._2d_attation(feat,
tokens,
data=data,
training=self.training)
if self.pred_concat:
f_c = holistic_feat.size(-1)
holistic_feat = holistic_feat.expand(bsz, seq_len + 1, f_c)
preds = self.prediction(
torch.cat((Hidden_state, attn_feat, holistic_feat), 2))
else:
preds = self.prediction(attn_feat)
# bsz * (seq_len + 1) * num_classes
char_output = preds[:, i, :]
char_output = F.softmax(char_output, -1)
outputs.append(char_output)
_, max_idx = torch.max(char_output, dim=1, keepdim=False)
char_embedding = self.embedding(max_idx)
if (i < seq_len):
tokens[:, i + 1, :] = char_embedding
if (tokens == self.end_idx).any(dim=-1).all():
break
outputs = torch.stack(outputs, 1)
return outputs
def forward(self, feat, data=None):
if self.use_lstm:
holistic_feat = self.encoder(feat) # bsz c
else:
holistic_feat = F.adaptive_avg_pool2d(feat, (1, 1)).squeeze()
if self.training:
preds = self.forward_train(feat, holistic_feat, data=data)
else:
preds = self.forward_test(feat, holistic_feat, data=data)
# (bsz, seq_len, num_classes)
return preds
|