File size: 7,430 Bytes
c8ddb9b |
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 |
"""Module to train the GAN model"""
from typing import Any, Dict
import torch
from src.models.losses import discriminator_loss, generator_loss, kl_loss
from src.models.modules.discriminator import Discriminator
from src.models.modules.generator import Generator
from src.models.modules.image_encoder import InceptionEncoder, VGGEncoder
from src.models.modules.text_encoder import TextEncoder
from src.models.utils import (
define_optimizers,
load_model,
prepare_labels,
save_image_and_caption,
save_model,
save_plot,
)
# pylint: disable=too-many-locals
# pylint: disable=too-many-statements
def train(data_loader: Any, config_dict: Dict[str, Any]) -> None:
"""
Function to train the GAN model
:param data_loader: Data loader for the dataset
:param vocab_len: Length of the vocabulary
:param config_dict: Dictionary containing the configuration parameters
"""
(
Ng, # pylint: disable=invalid-name
D, # pylint: disable=invalid-name
condition_dim,
noise_dim,
lr_config,
batch_size,
device,
epochs,
vocab_len,
ix2word,
output_dir,
snapshot,
const_dict,
) = (
config_dict["Ng"],
config_dict["D"],
config_dict["condition_dim"],
config_dict["noise_dim"],
config_dict["lr_config"],
config_dict["batch_size"],
config_dict["device"],
config_dict["epochs"],
config_dict["vocab_len"],
config_dict["ix2word"],
config_dict["output_dir"],
config_dict["snapshot"],
config_dict["const_dict"],
)
generator = Generator(Ng, D, condition_dim, noise_dim).to(device)
discriminator = Discriminator().to(device)
text_encoder = TextEncoder(vocab_len, D, D // 2).to(device)
image_encoder = InceptionEncoder(D).to(device)
vgg_encoder = VGGEncoder().to(device)
gen_loss = []
disc_loss = []
load_model(generator, discriminator, image_encoder, text_encoder, output_dir)
(
optimizer_g,
optimizer_d,
optimizer_text_encoder,
opt_image_encoder,
) = define_optimizers(
generator, discriminator, image_encoder, text_encoder, lr_config
)
for epoch in range(1, epochs + 1):
for batch_idx, (
images,
correct_capt,
correct_capt_len,
curr_class,
word_labels,
) in enumerate(data_loader):
labels_real, labels_fake, labels_match, fake_word_labels = prepare_labels(
batch_size, word_labels.size(1), device
)
optimizer_d.zero_grad()
optimizer_text_encoder.zero_grad()
noise = torch.randn(batch_size, noise_dim).to(device)
word_emb, sent_emb = text_encoder(correct_capt)
local_incept_feat, global_incept_feat = image_encoder(images)
vgg_feat = vgg_encoder(images)
mask = correct_capt == 0
# Generate Fake Images
fake_imgs, mu_tensor, logvar = generator(
noise,
sent_emb,
word_emb,
global_incept_feat,
local_incept_feat,
vgg_feat,
mask,
)
# Generate Logits for discriminator update
real_discri_feat = discriminator(images)
fake_discri_feat = discriminator(fake_imgs.detach())
logits_discri = {
"fake": {
"uncond": discriminator.logits_uncond(fake_discri_feat),
"cond": discriminator.logits_cond(fake_discri_feat, sent_emb),
},
"real": {
"word_level": discriminator.logits_word_level(
real_discri_feat, word_emb, mask
),
"uncond": discriminator.logits_uncond(real_discri_feat),
"cond": discriminator.logits_cond(real_discri_feat, sent_emb),
},
}
labels_discri = {
"fake": {"word_level": fake_word_labels, "image": labels_fake},
"real": {"word_level": word_labels, "image": labels_real},
}
# Update Discriminator
loss_discri = discriminator_loss(logits_discri, labels_discri)
loss_discri.backward(retain_graph=True)
optimizer_d.step()
optimizer_text_encoder.step()
disc_loss.append(loss_discri.item())
optimizer_g.zero_grad()
opt_image_encoder.zero_grad()
word_emb, sent_emb = text_encoder(correct_capt)
fake_imgs, mu_tensor, logvar = generator(
noise,
sent_emb,
word_emb,
global_incept_feat,
local_incept_feat,
vgg_feat,
mask,
)
local_fake_incept_feat, global_fake_incept_feat = image_encoder(fake_imgs)
vgg_feat_fake = vgg_encoder(fake_imgs)
fake_feat_d = discriminator(fake_imgs)
logits_gen = {
"fake": {
"uncond": discriminator.logits_uncond(fake_feat_d),
"cond": discriminator.logits_cond(fake_feat_d, sent_emb),
}
}
# Update Generator
loss_gen = generator_loss(
logits_gen,
local_fake_incept_feat,
global_fake_incept_feat,
labels_real,
word_emb,
sent_emb,
labels_match,
correct_capt_len,
curr_class,
vgg_feat,
vgg_feat_fake,
const_dict,
)
loss_kl = kl_loss(mu_tensor, logvar)
loss_gen += loss_kl
loss_gen.backward()
optimizer_g.step()
opt_image_encoder.step()
gen_loss.append(loss_gen.item())
if (batch_idx + 1) % 20 == 0:
print(
f"Epoch [{epoch}/{epochs}], Batch [{batch_idx + 1}/{len(data_loader)}],\
Loss D: {loss_discri.item():.4f}, Loss G: {loss_gen.item():.4f}"
)
if (batch_idx + 1) % 50 == 0:
with torch.no_grad():
fake_imgs_act, _, _ = generator(
noise,
sent_emb,
word_emb,
global_incept_feat,
local_incept_feat,
vgg_feat,
mask,
)
save_image_and_caption(
fake_imgs_act,
images,
correct_capt,
ix2word,
batch_idx,
epoch,
output_dir,
)
save_plot(gen_loss, disc_loss, epoch, batch_idx, output_dir)
if epoch % snapshot == 0 and epoch != 0:
save_model(
generator, discriminator, image_encoder, text_encoder, epoch, output_dir
)
save_model(
generator, discriminator, image_encoder, text_encoder, epochs, output_dir
)
|