Spaces:
Sleeping
Sleeping
Create trainning.py
Browse files- trainning.py +301 -0
trainning.py
ADDED
@@ -0,0 +1,301 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import string
|
3 |
+
import torch.nn.functional as F
|
4 |
+
import torch.nn as nn
|
5 |
+
import torchvision.models as models
|
6 |
+
|
7 |
+
def decoder(indices, vocab):
|
8 |
+
tokens = [vocab.lookup_token(idx) for idx in indices]
|
9 |
+
words = []
|
10 |
+
current_word = []
|
11 |
+
for token in tokens:
|
12 |
+
if len(token) == 1 and token in string.ascii_lowercase:
|
13 |
+
current_word.append(token)
|
14 |
+
else:
|
15 |
+
if current_word:
|
16 |
+
words.append("".join(current_word))
|
17 |
+
current_word = []
|
18 |
+
words.append(token)
|
19 |
+
|
20 |
+
if current_word:
|
21 |
+
words.append(" "+"".join(current_word))
|
22 |
+
|
23 |
+
return "".join(words)
|
24 |
+
|
25 |
+
def beam_search_caption(model, images, vocab, decoder, device="cpu",
|
26 |
+
start_token="<sos>", end_token="<eos>",
|
27 |
+
beam_width=3, max_seq_length=100):
|
28 |
+
"""
|
29 |
+
Generates captions for images using beam search.
|
30 |
+
|
31 |
+
Args:
|
32 |
+
model (ImgCap): The image captioning model.
|
33 |
+
images (torch.Tensor): Batch of images.
|
34 |
+
vocab (Vocab): Vocabulary object.
|
35 |
+
decoder (function): Function to decode indices to words.
|
36 |
+
device (str): Device to perform computation on.
|
37 |
+
start_token (str): Start-of-sequence token.
|
38 |
+
end_token (str): End-of-sequence token.
|
39 |
+
beam_width (int): Number of beams to keep.
|
40 |
+
max_seq_length (int): Maximum length of the generated caption.
|
41 |
+
|
42 |
+
Returns:
|
43 |
+
list: Generated captions for each image in the batch.
|
44 |
+
"""
|
45 |
+
model.eval()
|
46 |
+
|
47 |
+
with torch.no_grad():
|
48 |
+
start_index = vocab[start_token]
|
49 |
+
end_index = vocab[end_token]
|
50 |
+
images = images.to(device)
|
51 |
+
batch_size = images.size(0)
|
52 |
+
|
53 |
+
# Ensure batch_size is 1 for beam search (one image at a time)
|
54 |
+
if batch_size != 1:
|
55 |
+
raise ValueError("Beam search currently supports batch_size=1.")
|
56 |
+
|
57 |
+
cnn_feature = model.cnn(images) # Shape: (1, 1024)
|
58 |
+
lstm_input = model.lstm.projection(cnn_feature).unsqueeze(1) # Shape: (1, 1, 1024)
|
59 |
+
state = None # Initial LSTM state
|
60 |
+
|
61 |
+
# Initialize the beam with the start token
|
62 |
+
sequences = [([start_index], 0.0, lstm_input, state)] # List of tuples: (sequence, score, input, state)
|
63 |
+
|
64 |
+
completed_sequences = []
|
65 |
+
|
66 |
+
for _ in range(max_seq_length):
|
67 |
+
all_candidates = []
|
68 |
+
|
69 |
+
# Iterate over all current sequences in the beam
|
70 |
+
for seq, score, lstm_input, state in sequences:
|
71 |
+
# If the last token is the end token, add the sequence to completed_sequences
|
72 |
+
if seq[-1] == end_index:
|
73 |
+
completed_sequences.append((seq, score))
|
74 |
+
continue
|
75 |
+
|
76 |
+
# Pass the current input and state through the LSTM
|
77 |
+
lstm_out, state_new = model.lstm.lstm(lstm_input, state) # lstm_out: (1, 1, 1024)
|
78 |
+
|
79 |
+
# Pass the LSTM output through the fully connected layer to get logits
|
80 |
+
output = model.lstm.fc(lstm_out.squeeze(1)) # Shape: (1, vocab_size)
|
81 |
+
|
82 |
+
# Compute log probabilities
|
83 |
+
log_probs = F.log_softmax(output, dim=1) # Shape: (1, vocab_size)
|
84 |
+
|
85 |
+
# Get the top beam_width tokens and their log probabilities
|
86 |
+
top_log_probs, top_indices = log_probs.topk(beam_width, dim=1) # Each of shape: (1, beam_width)
|
87 |
+
|
88 |
+
# Iterate over the top tokens to create new candidate sequences
|
89 |
+
for i in range(beam_width):
|
90 |
+
token = top_indices[0, i].item()
|
91 |
+
token_log_prob = top_log_probs[0, i].item()
|
92 |
+
|
93 |
+
# Create a new sequence by appending the current token
|
94 |
+
new_seq = seq + [token]
|
95 |
+
new_score = score + token_log_prob
|
96 |
+
|
97 |
+
# Get the embedding of the new token
|
98 |
+
token_tensor = torch.tensor([token], device=device)
|
99 |
+
new_lstm_input = model.lstm.embedding(token_tensor).unsqueeze(1) # Shape: (1, 1, 1024)
|
100 |
+
|
101 |
+
# Clone the new state to ensure each beam has its own state
|
102 |
+
if state_new is not None:
|
103 |
+
new_state = (state_new[0].clone(), state_new[1].clone())
|
104 |
+
else:
|
105 |
+
new_state = None
|
106 |
+
|
107 |
+
# Add the new candidate to all_candidates
|
108 |
+
all_candidates.append((new_seq, new_score, new_lstm_input, new_state))
|
109 |
+
|
110 |
+
# If no candidates are left to process, break out of the loop
|
111 |
+
if not all_candidates:
|
112 |
+
break
|
113 |
+
|
114 |
+
# Sort all candidates by score in descending order
|
115 |
+
ordered = sorted(all_candidates, key=lambda tup: tup[1], reverse=True)
|
116 |
+
|
117 |
+
# Select the top beam_width sequences to form the new beam
|
118 |
+
sequences = ordered[:beam_width]
|
119 |
+
|
120 |
+
# If enough completed sequences are found, stop early
|
121 |
+
if len(completed_sequences) >= beam_width:
|
122 |
+
break
|
123 |
+
|
124 |
+
# If no sequences have completed, use the current sequences
|
125 |
+
if len(completed_sequences) == 0:
|
126 |
+
completed_sequences = sequences
|
127 |
+
|
128 |
+
# Select the sequence with the highest score
|
129 |
+
best_seq, best_score = max(completed_sequences, key=lambda x: x[1])
|
130 |
+
|
131 |
+
if best_seq[0] == start_index:
|
132 |
+
best_seq = best_seq[1:]
|
133 |
+
|
134 |
+
best_caption = decoder(best_seq, vocab)
|
135 |
+
|
136 |
+
return best_caption
|
137 |
+
|
138 |
+
|
139 |
+
def generate_caption(model, images, vocab, decoder, device="cpu", start_token="<sos>", end_token="<eos>", max_seq_length=100, top_k=2):
|
140 |
+
model.eval()
|
141 |
+
|
142 |
+
with torch.no_grad():
|
143 |
+
start_index = vocab[start_token]
|
144 |
+
end_index = vocab[end_token]
|
145 |
+
images = images.to(device)
|
146 |
+
batch_size = images.size(0)
|
147 |
+
|
148 |
+
end_token_appear = {i: False for i in range(batch_size)}
|
149 |
+
captions = [[] for _ in range(batch_size)]
|
150 |
+
|
151 |
+
cnn_feature = model.cnn(images)
|
152 |
+
lstm_input = model.lstm.projection(cnn_feature).unsqueeze(1) # (B, 1, hidden_size)
|
153 |
+
|
154 |
+
state = None
|
155 |
+
|
156 |
+
for i in range(max_seq_length):
|
157 |
+
lstm_out, state = model.lstm.lstm(lstm_input, state)
|
158 |
+
output = model.lstm.fc(lstm_out.squeeze(1))
|
159 |
+
|
160 |
+
top_k_probs, top_k_indices = torch.topk(F.softmax(output, dim=1), top_k, dim=1)
|
161 |
+
top_k_probs = top_k_probs / torch.sum(top_k_probs, dim=1, keepdim=True)
|
162 |
+
top_k_samples = torch.multinomial(top_k_probs, 1).squeeze()
|
163 |
+
|
164 |
+
predicted_word_indices = top_k_indices[range(batch_size), top_k_samples]
|
165 |
+
|
166 |
+
lstm_input = model.lstm.embedding(predicted_word_indices).unsqueeze(1) # (B, 1, hidden_size)
|
167 |
+
|
168 |
+
for j in range(batch_size):
|
169 |
+
if end_token_appear[j]:
|
170 |
+
continue
|
171 |
+
|
172 |
+
word = vocab.lookup_token(predicted_word_indices[j].item())
|
173 |
+
if word == end_token:
|
174 |
+
end_token_appear[j] = True
|
175 |
+
|
176 |
+
captions[j].append(predicted_word_indices[j].item())
|
177 |
+
|
178 |
+
captions = [decoder(caption, vocab) for caption in captions]
|
179 |
+
|
180 |
+
return captions
|
181 |
+
|
182 |
+
|
183 |
+
|
184 |
+
|
185 |
+
class ResNet50(nn.Module):
|
186 |
+
def __init__(self):
|
187 |
+
super(ResNet50, self).__init__()
|
188 |
+
self.ResNet50 = models.resnet50(weights=models.ResNet50_Weights.DEFAULT)
|
189 |
+
|
190 |
+
self.ResNet50.fc = nn.Sequential(
|
191 |
+
nn.Linear(2048, 1024),
|
192 |
+
nn.ReLU(),
|
193 |
+
nn.Dropout(0.5),
|
194 |
+
nn.Linear(1024, 1024),
|
195 |
+
nn.ReLU(),
|
196 |
+
)
|
197 |
+
|
198 |
+
for k,v in self.ResNet50.named_parameters(recurse=True):
|
199 |
+
if 'fc' in k:
|
200 |
+
v.requires_grad = True
|
201 |
+
else:
|
202 |
+
v.requires_grad = False
|
203 |
+
|
204 |
+
def forward(self,x):
|
205 |
+
return self.ResNet50(x)
|
206 |
+
|
207 |
+
## lSTM (Decoder)
|
208 |
+
|
209 |
+
class lstm(nn.Module):
|
210 |
+
def __init__(self, input_size, hidden_size, number_layers, embedding_dim, vocab_size):
|
211 |
+
super(lstm, self).__init__()
|
212 |
+
|
213 |
+
self.input_size = input_size
|
214 |
+
self.hidden_size = hidden_size
|
215 |
+
self.number_layers = number_layers
|
216 |
+
self.embedding_dim = embedding_dim
|
217 |
+
self.vocab_size = vocab_size
|
218 |
+
|
219 |
+
self.embedding = nn.Embedding(vocab_size, hidden_size)
|
220 |
+
self.projection = nn.Linear(input_size, hidden_size)
|
221 |
+
self.relu = nn.ReLU()
|
222 |
+
|
223 |
+
self.lstm = nn.LSTM(
|
224 |
+
input_size=hidden_size,
|
225 |
+
hidden_size=hidden_size,
|
226 |
+
num_layers=number_layers,
|
227 |
+
dropout=0.5,
|
228 |
+
batch_first=True,
|
229 |
+
)
|
230 |
+
|
231 |
+
self.fc = nn.Linear(hidden_size, vocab_size)
|
232 |
+
|
233 |
+
def forward(self, x, captions):
|
234 |
+
projected_image = self.projection(x).unsqueeze(dim=1)
|
235 |
+
embeddings = self.embedding(captions[:, :-1])
|
236 |
+
|
237 |
+
# Concatenate the image feature as frist step with word embeddings
|
238 |
+
lstm_input = torch.cat((projected_image, embeddings), dim=1)
|
239 |
+
# print(torch.all(projected_image[:, 0, :] == lstm_input[:, 0, :])) # check
|
240 |
+
|
241 |
+
lstm_out, _ = self.lstm(lstm_input)
|
242 |
+
logits = self.fc(lstm_out)
|
243 |
+
|
244 |
+
return logits
|
245 |
+
|
246 |
+
## ImgCap
|
247 |
+
|
248 |
+
class ImgCap(nn.Module):
|
249 |
+
def __init__(self, cnn_feature_size, lstm_hidden_size, num_layers, vocab_size, embedding_dim):
|
250 |
+
super(ImgCap, self).__init__()
|
251 |
+
|
252 |
+
self.cnn = ResNet50()
|
253 |
+
|
254 |
+
self.lstm = lstm(input_size=cnn_feature_size,
|
255 |
+
hidden_size=lstm_hidden_size,
|
256 |
+
number_layers=num_layers,
|
257 |
+
embedding_dim=embedding_dim,
|
258 |
+
vocab_size=vocab_size)
|
259 |
+
|
260 |
+
def forward(self, images, captions):
|
261 |
+
cnn_features = self.cnn(images)
|
262 |
+
output = self.lstm(cnn_features, captions)
|
263 |
+
return output
|
264 |
+
|
265 |
+
def generate_caption(self, images, vocab, decoder, device="cpu", start_token="<sos>", end_token="<eos>", max_seq_length=100):
|
266 |
+
self.eval()
|
267 |
+
|
268 |
+
with torch.no_grad():
|
269 |
+
start_index = vocab[start_token]
|
270 |
+
end_index = vocab[end_token]
|
271 |
+
images = images.to(device)
|
272 |
+
batch_size = images.size(0)
|
273 |
+
|
274 |
+
end_token_appear = {i: False for i in range(batch_size)}
|
275 |
+
captions = [[] for _ in range(batch_size)]
|
276 |
+
|
277 |
+
cnn_feature = self.cnn(images)
|
278 |
+
lstm_input = self.lstm.projection(cnn_feature).unsqueeze(1) # (B, 1, hidden_size)
|
279 |
+
|
280 |
+
state = None
|
281 |
+
|
282 |
+
for i in range(max_seq_length):
|
283 |
+
lstm_out, state = self.lstm.lstm(lstm_input, state)
|
284 |
+
output = self.lstm.fc(lstm_out.squeeze(1))
|
285 |
+
predicted_word_indices = torch.argmax(output, dim=1)
|
286 |
+
lstm_input = self.lstm.embedding(predicted_word_indices).unsqueeze(1) # (B, 1, hidden_size)
|
287 |
+
|
288 |
+
for j in range(batch_size):
|
289 |
+
if end_token_appear[j]:
|
290 |
+
continue
|
291 |
+
|
292 |
+
word = vocab.lookup_token(predicted_word_indices[j].item())
|
293 |
+
if word == end_token:
|
294 |
+
end_token_appear[j] = True
|
295 |
+
|
296 |
+
captions[j].append(predicted_word_indices[j].item())
|
297 |
+
|
298 |
+
captions = [decoder(caption) for caption in captions]
|
299 |
+
|
300 |
+
return captions
|
301 |
+
|