def build_models():
# build model ############################################################
text_encoder = BERT_RNN_ENCODER(dataset.n_words,
nhidden=cfg.TEXT.EMBEDDING_DIM)
image_encoder = BERT_CNN_ENCODER_RNN_DECODER(cfg.TEXT.EMBEDDING_DIM,
cfg.CNN_RNN.HIDDEN_DIM,
dataset.n_words,
rec_unit=cfg.RNN_TYPE)
labels = Variable(torch.LongTensor(range(batch_size)))
start_epoch = 0
if cfg.TRAIN.NET_E != '':
state_dict = torch.load(cfg.TRAIN.NET_E)
text_encoder.load_state_dict(state_dict)
print('Load ', cfg.TRAIN.NET_E)
#
name = cfg.TRAIN.NET_E.replace('text_encoder', 'image_encoder')
state_dict = torch.load(name)
image_encoder.load_state_dict(state_dict)
print('Load ', name)
istart = cfg.TRAIN.NET_E.rfind('_') + 8
iend = cfg.TRAIN.NET_E.rfind('.')
start_epoch = cfg.TRAIN.NET_E[istart:iend]
start_epoch = int(start_epoch) + 1
print('start_epoch', start_epoch)
if cfg.CUDA:
text_encoder = text_encoder.cuda()
image_encoder = image_encoder.cuda()
labels = labels.cuda()
return text_encoder, image_encoder, labels, start_epoch
def build_models():
# build model ############################################################
#cfg.LOCAL_PRETRAINED = False
if(cfg.LOCAL_PRETRAINED):
tokenizer = tokenization.FullTokenizer(vocab_file=cfg.BERT_ENCODER.VOCAB, do_lower_case=True)
vocab_size = len(tokenizer.vocab)
#vocab_size = 3770
#vocab_size = 4000
else:
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
vocab_size = len(tokenizer.vocab)
#vocab_size = 30522
#text_encoder = BERT_RNN_ENCODER(dataset.n_words, nhidden=cfg.TEXT.EMBEDDING_DIM)
#image_encoder = BERT_CNN_ENCODER_RNN_DECODER(cfg.TEXT.EMBEDDING_DIM, cfg.CNN_RNN.HIDDEN_DIM, dataset.n_words, rec_unit=cfg.RNN_TYPE)
text_encoder = BERT_RNN_ENCODER(vocab_size, nhidden=cfg.TEXT.EMBEDDING_DIM)
image_encoder = BERT_CNN_ENCODER_RNN_DECODER(cfg.TEXT.EMBEDDING_DIM, cfg.CNN_RNN.HIDDEN_DIM, vocab_size, rec_unit=cfg.RNN_TYPE)
labels = Variable(torch.LongTensor(range(batch_size)))
start_epoch = 0
if cfg.TRAIN.NET_E != '':
state_dict = torch.load(cfg.TRAIN.NET_E)
text_encoder.load_state_dict(state_dict)
print('Load ', cfg.TRAIN.NET_E)
#
name = cfg.TRAIN.NET_E.replace('text_encoder', 'image_encoder')
state_dict = torch.load(name)
image_encoder.load_state_dict(state_dict)
print('Load ', name)
istart = cfg.TRAIN.NET_E.rfind('_') + 8
iend = cfg.TRAIN.NET_E.rfind('.')
start_epoch = cfg.TRAIN.NET_E[istart:iend]
start_epoch = int(start_epoch) + 1
print('start_epoch', start_epoch)
if cfg.CUDA:
text_encoder = text_encoder.cuda()
image_encoder = image_encoder.cuda()
labels = labels.cuda()
return text_encoder, image_encoder, labels, start_epoch
def gen_example(self, data_dic):
if cfg.TRAIN.NET_G == '':
print('Error: the path for morels is not found!')
else:
# Build and load the generator
text_encoder = \
BERT_RNN_ENCODER(self.n_words, nhidden=cfg.TEXT.EMBEDDING_DIM)
state_dict = \
torch.load(cfg.TRAIN.NET_E, map_location=lambda storage, loc: storage)
text_encoder.load_state_dict(state_dict)
print('Load text encoder from:', cfg.TRAIN.NET_E)
text_encoder = text_encoder.cuda()
text_encoder.eval()
# the path to save generated images
if cfg.GAN.B_DCGAN:
netG = G_DCGAN()
else:
netG = G_NET()
s_tmp = cfg.TRAIN.NET_G[:cfg.TRAIN.NET_G.rfind('.pth')]
model_dir = cfg.TRAIN.NET_G
state_dict = \
torch.load(model_dir, map_location=lambda storage, loc: storage)
netG.load_state_dict(state_dict)
print('Load G from: ', model_dir)
netG.cuda()
netG.eval()
for key in data_dic:
save_dir = '%s/%s' % (s_tmp, key)
mkdir_p(save_dir)
captions, cap_lens, sorted_indices = data_dic[key]
batch_size = captions.shape[0]
nz = cfg.GAN.Z_DIM
captions = Variable(torch.from_numpy(captions), volatile=True)
cap_lens = Variable(torch.from_numpy(cap_lens), volatile=True)
captions = captions.cuda()
cap_lens = cap_lens.cuda()
for i in range(1): # 16
noise = Variable(torch.FloatTensor(batch_size, nz),
volatile=True)
noise = noise.cuda()
#######################################################
# (1) Extract text embeddings
######################################################
hidden = text_encoder.init_hidden(batch_size)
# words_embs: batch_size x nef x seq_len
# sent_emb: batch_size x nef
words_embs, sent_emb = text_encoder(
captions, cap_lens, hidden)
mask = (captions == 0)
#######################################################
# (2) Generate fake images
######################################################
noise.data.normal_(0, 1)
fake_imgs, attention_maps, _, _ = netG(
noise, sent_emb, words_embs, mask)
# G attention
cap_lens_np = cap_lens.cpu().data.numpy()
for j in range(batch_size):
save_name = '%s/%d_s_%d' % (save_dir, i,
sorted_indices[j])
for k in range(len(fake_imgs)):
im = fake_imgs[k][j].data.cpu().numpy()
im = (im + 1.0) * 127.5
im = im.astype(np.uint8)
# print('im', im.shape)
im = np.transpose(im, (1, 2, 0))
# print('im', im.shape)
im = Image.fromarray(im)
fullpath = '%s_g%d.png' % (save_name, k)
im.save(fullpath)
for k in range(len(attention_maps)):
if len(fake_imgs) > 1:
im = fake_imgs[k + 1].detach().cpu()
else:
im = fake_imgs[0].detach().cpu()
attn_maps = attention_maps[k]
att_sze = attn_maps.size(2)
img_set, sentences = \
build_super_images2(im[j].unsqueeze(0),
captions[j].unsqueeze(0),
[cap_lens_np[j]], self.ixtoword,
[attn_maps[j]], att_sze)
if img_set is not None:
im = Image.fromarray(img_set)
fullpath = '%s_a%d.png' % (save_name, k)
im.save(fullpath)
def sampling(self, split_dir):
if cfg.TRAIN.NET_G == '':
print('Error: the path for morels is not found!')
else:
if split_dir == 'test':
split_dir = 'valid'
# Build and load the generator
if cfg.GAN.B_DCGAN:
netG = G_DCGAN()
else:
netG = G_NET()
netG.apply(weights_init)
netG.cuda()
netG.eval()
#
text_encoder = BERT_RNN_ENCODER(self.n_words,
nhidden=cfg.TEXT.EMBEDDING_DIM)
state_dict = \
torch.load(cfg.TRAIN.NET_E, map_location=lambda storage, loc: storage)
text_encoder.load_state_dict(state_dict)
print('Load text encoder from:', cfg.TRAIN.NET_E)
text_encoder = text_encoder.cuda()
text_encoder.eval()
batch_size = self.batch_size
nz = cfg.GAN.Z_DIM
noise = Variable(torch.FloatTensor(batch_size, nz), volatile=True)
noise = noise.cuda()
model_dir = cfg.TRAIN.NET_G
state_dict = \
torch.load(model_dir, map_location=lambda storage, loc: storage)
# state_dict = torch.load(cfg.TRAIN.NET_G)
netG.load_state_dict(state_dict)
print('Load G from: ', model_dir)
# the path to save generated images
s_tmp = model_dir[:model_dir.rfind('.pth')]
save_dir = '%s/%s' % (s_tmp, split_dir)
mkdir_p(save_dir)
cnt = 0
for _ in range(1): # (cfg.TEXT.CAPTIONS_PER_IMAGE):
for step, data in enumerate(self.data_loader, 0):
cnt += batch_size
if step % 100 == 0:
print('step: ', step)
# if step > 50:
# break
imgs, captions, cap_lens, class_ids, keys = prepare_data(
data)
hidden = text_encoder.init_hidden(batch_size)
# words_embs: batch_size x nef x seq_len
# sent_emb: batch_size x nef
words_embs, sent_emb = text_encoder(
captions, cap_lens, hidden)
words_embs, sent_emb = words_embs.detach(
), sent_emb.detach()
mask = (captions == 0)
num_words = words_embs.size(2)
if mask.size(1) > num_words:
mask = mask[:, :num_words]
#######################################################
# (2) Generate fake images
######################################################
noise.data.normal_(0, 1)
fake_imgs, _, _, _ = netG(noise, sent_emb, words_embs,
mask)
for j in range(batch_size):
s_tmp = '%s/single/%s' % (save_dir, keys[j])
folder = s_tmp[:s_tmp.rfind('/')]
if not os.path.isdir(folder):
print('Make a new folder: ', folder)
mkdir_p(folder)
k = -1
# for k in range(len(fake_imgs)):
im = fake_imgs[k][j].data.cpu().numpy()
# [-1, 1] --> [0, 255]
im = (im + 1.0) * 127.5
im = im.astype(np.uint8)
im = np.transpose(im, (1, 2, 0))
im = Image.fromarray(im)
fullpath = '%s_s%d.png' % (s_tmp, k)
im.save(fullpath)
def build_models(self):
# ###################encoders######################################## #
if cfg.TRAIN.NET_E == '':
print('Error: no pretrained text-image encoders')
return
image_encoder = BERT_CNN_ENCODER_RNN_DECODER(cfg.TEXT.EMBEDDING_DIM,
cfg.CNN_RNN.HIDDEN_DIM,
self.n_words,
rec_unit=cfg.RNN_TYPE)
img_encoder_path = cfg.TRAIN.NET_E.replace('text_encoder',
'image_encoder')
state_dict = \
torch.load(img_encoder_path, map_location=lambda storage, loc: storage)
image_encoder.load_state_dict(state_dict)
for p in image_encoder.parameters():
p.requires_grad = False
print('Load image encoder from:', img_encoder_path)
# image_encoder.eval()
text_encoder = \
BERT_RNN_ENCODER(self.n_words, nhidden=cfg.TEXT.EMBEDDING_DIM)
state_dict = \
torch.load(cfg.TRAIN.NET_E,
map_location=lambda storage, loc: storage)
text_encoder.load_state_dict(state_dict)
for p in text_encoder.parameters():
p.requires_grad = False
print('Load text encoder from:', cfg.TRAIN.NET_E)
text_encoder.eval()
# #######################generator and discriminators############## #
netsD = []
if cfg.GAN.B_DCGAN:
if cfg.TREE.BRANCH_NUM == 1:
from model import D_NET64 as D_NET
elif cfg.TREE.BRANCH_NUM == 2:
from model import D_NET128 as D_NET
else: # cfg.TREE.BRANCH_NUM == 3:
from model import D_NET256 as D_NET
# TODO: elif cfg.TREE.BRANCH_NUM > 3:
netG = G_DCGAN()
netsD = [D_NET(b_jcu=False)]
else:
from model import D_NET64, D_NET128, D_NET256
netG = G_NET()
if cfg.TREE.BRANCH_NUM > 0:
netsD.append(D_NET64())
if cfg.TREE.BRANCH_NUM > 1:
netsD.append(D_NET128())
if cfg.TREE.BRANCH_NUM > 2:
netsD.append(D_NET256())
# TODO: if cfg.TREE.BRANCH_NUM > 3:
netG.apply(weights_init)
# print(netG)
for i in range(len(netsD)):
netsD[i].apply(weights_init)
# print(netsD[i])
print('# of netsD', len(netsD))
#
epoch = 0
if cfg.TRAIN.NET_G != '':
state_dict = \
torch.load(cfg.TRAIN.NET_G, map_location=lambda storage, loc: storage)
netG.load_state_dict(state_dict)
print('Load G from: ', cfg.TRAIN.NET_G)
istart = cfg.TRAIN.NET_G.rfind('_') + 1
iend = cfg.TRAIN.NET_G.rfind('.')
epoch = cfg.TRAIN.NET_G[istart:iend]
epoch = int(epoch) + 1
if cfg.TRAIN.B_NET_D:
Gname = cfg.TRAIN.NET_G
for i in range(len(netsD)):
s_tmp = Gname[:Gname.rfind('/')]
Dname = '%s/netD%d.pth' % (s_tmp, i)
print('Load D from: ', Dname)
state_dict = \
torch.load(Dname, map_location=lambda storage, loc: storage)
netsD[i].load_state_dict(state_dict)
# ########################################################### #
if cfg.CUDA:
text_encoder = text_encoder.cuda()
image_encoder = image_encoder.cuda()
netG.cuda()
for i in range(len(netsD)):
netsD[i].cuda()
return [text_encoder, image_encoder, netG, netsD, epoch]
def generate_fake_images_with_incremental_noise(self, data_dic, sizeim):
global text_encoder_path, net_G_path
print(os.getcwd(), os.path.join(os.getcwd(), text_encoder_path))
text_encoder_path = os.path.join(os.getcwd(), text_encoder_path)
net_G_path = os.path.join(os.getcwd(), net_G_path)
# Build and load the generator
#####################################
## load the encoder #
#####################################
print('Loading text encoder from:', text_encoder_path)
text_encoder = \
BERT_RNN_ENCODER(self.n_words, nhidden=cfg.TEXT.EMBEDDING_DIM)
state_dict = \
torch.load(text_encoder_path,
map_location=lambda storage, loc: storage)
text_encoder.load_state_dict(state_dict)
print('Loaded text encoder from:', text_encoder_path)
text_encoder.eval()
text_encoder = text_encoder.cuda()
netG = G_NET()
######################################
## load the generator #
######################################
state_dict = \
torch.load(net_G_path, map_location=lambda storage, loc: storage)
netG.load_state_dict(state_dict)
print('Load Generator from: ', net_G_path)
s_tmp = net_G_path[:net_G_path.rfind('.pth')]
netG.cuda()
netG.eval()
for key in data_dic:
save_dir = '%s/%s' % ('res', key)
mkdir_p(save_dir)
captions, cap_lens, sorted_indices = data_dic[key]
batch_size = captions.shape[0]
nz = cfg.GAN.Z_DIM
captions = Variable(torch.from_numpy(captions))
cap_lens = Variable(torch.from_numpy(cap_lens))
captions = captions.cuda()
cap_lens = cap_lens.cuda()
base_noise = Variable(torch.FloatTensor(batch_size, nz))
base_noise = base_noise.cuda()
for i in range(sizeim): # number of images to be created
noise = base_noise.clone()
noise[0][i %
100] = base_noise[0][i % 100] + torch.mean(base_noise)
#######################################################
# (1) Extract text embeddings
######################################################
hidden = text_encoder.init_hidden(batch_size)
# words_embs: batch_size x nef x seq_len
# sent_emb: batch_size x nef
words_embs, sent_emb = text_encoder(captions, cap_lens, hidden)
mask = (captions == 0)
#######################################################
# (2) Generate fake images
######################################################
noise.data.normal_(0, 1)
fake_imgs, attention_maps, _, _ = netG(noise, sent_emb,
words_embs, mask)
im = fake_imgs[2].squeeze(0).data.cpu().numpy()
im = (im + 1.0) * 127.5
im = im.astype(np.uint8)
# print('im', im.shape)
im = np.transpose(im, (1, 2, 0))
# print('im', im.shape)
im = Image.fromarray(im)
fullpath = os.path.join(save_dir, '{0}.png'.format(i))
im.save(fullpath)
def generate_fake_im(self, data_dic):
global text_encoder_path, net_G_path
# Build and load the generator
#####################################
## load the encoder #
#####################################
text_encoder = \
BERT_RNN_ENCODER(self.n_words, nhidden=cfg.TEXT.EMBEDDING_DIM)
state_dict = \
torch.load(text_encoder_path,
map_location=lambda storage, loc: storage)
text_encoder.load_state_dict(state_dict)
print('Loaded text encoder from:', text_encoder_path)
text_encoder.eval()
text_encoder = text_encoder.cuda()
netG = G_NET()
######################################
## load the generator #
######################################
state_dict = \
torch.load(net_G_path, map_location=lambda storage, loc: storage)
netG.load_state_dict(state_dict)
print('Load Generator from: ', net_G_path)
s_tmp = net_G_path[:net_G_path.rfind('.pth')]
netG.cuda()
netG.eval()
for key in data_dic:
save_dir = '%s/%s' % (s_tmp, key)
mkdir_p(save_dir)
captions, cap_lens, sorted_indices = data_dic[key]
batch_size = captions.shape[0]
nz = cfg.GAN.Z_DIM
captions = Variable(torch.from_numpy(captions))
cap_lens = Variable(torch.from_numpy(cap_lens))
captions = captions.cuda()
cap_lens = cap_lens.cuda()
for i in range(1): # 16
noise = Variable(torch.FloatTensor(batch_size, nz))
noise = noise.cuda()
#######################################################
# (1) Extract text embeddings
######################################################
hidden = text_encoder.init_hidden(batch_size)
# words_embs: batch_size x nef x seq_len
# sent_emb: batch_size x nef
words_embs, sent_emb = text_encoder(captions, cap_lens, hidden)
mask = (captions == 0)
#######################################################
# (2) Generate fake images
######################################################
noise.data.normal_(0, 1)
fake_imgs, attention_maps, _, _ = netG(noise, sent_emb,
words_embs, mask)
return fake_imgs, attention_maps