-
Notifications
You must be signed in to change notification settings - Fork 0
/
generator_init_inject.py
203 lines (161 loc) · 6.71 KB
/
generator_init_inject.py
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
import torch
import torchvision
from torchvision import transforms
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch import Tensor
from Inception import inception_v3
from test_init_inject import get_test
from decoder_init_inject import get_decoder
torch.set_default_tensor_type('torch.cuda.FloatTensor')
from encoder import get_encoder
# from autoencoder import *
#from decoder import get_decoder
from utils import *
from create_vocabulary import *
from tokenization import *
import numpy as np
from tqdm import tqdm
import time
torch.manual_seed(12321)
torch.cuda.manual_seed(12321)
torch.cuda.manual_seed_all(12321)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
val_captions, val_image_names = load_mscoco_annotations_val()
train_captions, train_image_names = load_mscoco_annotations_train(root='Resized')
voc = Voc(name="Vocabulary")
train_normalized_captions = normalizeAllCaptions(train_captions)
print()
print("Creating Vocabulary...")
for caption in tqdm(train_normalized_captions):
voc.addCaption(caption=caption)
voc.trim(min_count=77)
tokenized_train_captions = tokenize(voc, train_normalized_captions)
voc_size = len(voc.index2word)
core_img_path = []
core_img_capt_tokens = []
for img_path, img_cap in zip(train_image_names, tokenized_train_captions):
if 15 >= len(img_cap) >= 9:
core_img_path.append(img_path)
core_img_capt_tokens.append(img_cap)
len_captions = np.zeros((70,), dtype=int)
for caption in tqdm(tokenized_train_captions):
len_captions[len(caption)] += 1
print(len_captions)
train_captions_tokens = torch.LongTensor(pad_sequences(core_img_capt_tokens))
print(train_captions_tokens.shape)
train_image_names = core_img_path
voc.save_vocabulary()
batch_size = 128 # 128
hidden_size = 2048
output_size = voc_size # num words
embed_size = 128
feature_size = 2048
num_layers = 1
preprocess = transforms.Compose([
transforms.Resize(299), # 299
transforms.CenterCrop(299),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
# encoder = get_encoder()
encoder = inception_v3(pretrained=True)
encoder = encoder.to(device)
decoder = get_decoder(output_size=output_size)
decoder = decoder.to(device)
def train_step(tokens_tensor, image_tensor, decoder, decoder_optimizer, encoder, criterion):
decoder_optimizer.zero_grad()
# encoder_optimizer.zero_grad()
batch_size = tokens_tensor.size(0)
sequence_length = tokens_tensor.size(1)
loss = 0
#with torch.no_grad():
# feature_tensor = encoder(image_tensor.to(device))
feature_tensor = image_tensor.to(device)
feature_tensor = feature_tensor.unsqueeze(0)
decoder_hidden = torch.zeros(num_layers, batch_size, hidden_size).to(device)
for i in range(num_layers):
decoder_hidden[i] = feature_tensor
# decoder_hidden = feature_tensor.unsqueeze(0)
for seq in range(sequence_length - 1):
input_tensor = tokens_tensor[:, seq]
input_tensor = input_tensor.unsqueeze(0)
input_tensor = input_tensor.type(torch.LongTensor).to(device)
output = tokens_tensor[:, seq + 1].type(torch.LongTensor).to(device)
decoder_output, decoder_hidden = decoder(input_tensor, decoder_hidden)
loss += criterion(decoder_output, output)
loss.backward()
decoder_optimizer.step()
return loss.item() / sequence_length
def generate_caption(image_tensor, encoder, decoder, max_len=30):
image_tensor = image_tensor.unsqueeze(0)
#with torch.no_grad():
# feature = encoder(image_tensor.to(device))
"Does not work..."
feature = image_tensor.to(device)
feature = feature.unsqueeze(0)
decoder_hidden = torch.zeros(num_layers, 1, hidden_size).to(device)
for i in range(num_layers):
decoder_hidden[i] = feature
# input = torch.tensor(voc.word2index["soc"]).type(torch.LongTensor).to(device)
decoder_input = torch.ones(1, 1).type(torch.LongTensor).to(device)
caption = ""
for i in range(max_len):
out, decoder_hidden = decoder(decoder_input, decoder_hidden)
out = out.argmax(dim=1)
caption += voc.index2word[int(out)] + " "
decoder_input = out.unsqueeze(0)
print(caption)
def train(encoder, decoder, batch_size=128, n_iters=40):
plot_losses = []
print_loss_total = 0 # Reset every print_every
plot_loss_total = 0 # Reset every plot_every
encoder = encoder.to(device)
decoder = decoder.to(device)
decoder_optimizer = optim.Adam(decoder.parameters())
compare_loss = 99999999
criterion = nn.NLLLoss()
batch_index = 0
data_sample = train_captions_tokens.shape[0]
image_batch = torch.zeros(batch_size, 2048)
#image_batch = torch.zeros(batch_size, 3, 299, 299)
caption_batch = torch.zeros(batch_size, 16)
for iters in tqdm(range(1, n_iters + 1)):
for ids, im_path in tqdm(enumerate(train_image_names)):
encoder.eval()
decoder.train()
#im = load_image(im_path)
#im = gray_to_rgb(im)
#im = preprocess(im)
im_path = im_path.split("Resized/train2017/")[1]
image_id = im_path[:-4]
image = torch.load(f"inception_v3_features/features_{image_id}.pt")
batch_index = ids % batch_size
image_batch[batch_index] = image
caption_batch[batch_index] = train_captions_tokens[ids]
if ids == len(train_image_names) - 1:
loss = train_step(tokens_tensor=caption_batch[:batch_index + 1],
image_tensor=image_batch[:batch_index + 1], decoder=decoder,
decoder_optimizer=decoder_optimizer, encoder=encoder, criterion=criterion)
print_loss_total += loss
batch_index = 0
break
if batch_index == batch_size - 1:
loss = train_step(tokens_tensor=caption_batch, image_tensor=image_batch, decoder=decoder,
decoder_optimizer=decoder_optimizer, encoder=encoder, criterion=criterion)
print_loss_total += loss
if ids % (batch_size + 20) == 0:
print_loss_avg = print_loss_total / (batch_size + 20)
print_loss_total = 0
print(print_loss_avg)
print()
encoder.eval()
decoder.eval()
generate_caption(image_batch[0], encoder=encoder, decoder=decoder)
torch.save(decoder, f"init_inject_1_layer_{iters}.pth")
# torch.save(encoder, "best_encoder.pth")
get_test(iters)
train(encoder=encoder, decoder=decoder, batch_size=batch_size)