def generate(captions, copies):
x = pickle.load(open('data/captions.pickle', 'rb'))
# print(x)
ixtoword = x[2]
wordtoix = x[3]
del x
word_len = len(wordtoix)
text_encoder = RNN_ENCODER(word_len, 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)
text_encoder.eval()
netG = G_NET()
state_dict = torch.load(cfg.TRAIN.NET_G,
map_location=lambda storage, loc: storage)
netG.load_state_dict(state_dict)
# netG.eval()
# seed = 100
# random.seed(seed)
# np.random.seed(seed)
# torch.manual_seed(seed)
# load word vector
captions, cap_lens = vectorize_caption(wordtoix, captions, copies)
n_words = len(wordtoix)
# only one to generate
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)
noise = Variable(torch.FloatTensor(batch_size, nz), volatile=True)
#######################################################
# (1) Extract text embeddings
#######################################################
hidden = text_encoder.init_hidden(batch_size)
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()
# # storing to blob storage
# container_name = "images"
# full_path = "https://attgan.blob.core.windows.net/images/%s"
# prefix = datetime.now().strftime('%Y/%B/%d/%H_%M_%S_%f')
imgs = []
# only look at first one
#j = 0
return imglist(fake_imgs, batch_size)
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()
# load text encoder
text_encoder = 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()
#load image encoder
image_encoder = CNN_ENCODER(cfg.TEXT.EMBEDDING_DIM)
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)
print('Load image encoder from:', img_encoder_path)
image_encoder = image_encoder.cuda()
image_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
R_count = 0
R = np.zeros(30000)
cont = True
for ii in range(11): # (cfg.TEXT.CAPTIONS_PER_IMAGE):
if (cont == False):
break
for step, data in enumerate(self.data_loader, 0):
cnt += batch_size
if (cont == False):
break
if step % 100 == 0:
print('cnt: ', cnt)
# 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, cap_lens)
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_%d.png' % (s_tmp, k, ii)
im.save(fullpath)
_, cnn_code = image_encoder(fake_imgs[-1])
for i in range(batch_size):
mis_captions, mis_captions_len = self.dataset.get_mis_caption(
class_ids[i])
hidden = text_encoder.init_hidden(99)
_, sent_emb_t = text_encoder(mis_captions,
mis_captions_len, hidden)
rnn_code = torch.cat(
(sent_emb[i, :].unsqueeze(0), sent_emb_t), 0)
### cnn_code = 1 * nef
### rnn_code = 100 * nef
scores = torch.mm(cnn_code[i].unsqueeze(0),
rnn_code.transpose(0, 1)) # 1* 100
cnn_code_norm = torch.norm(cnn_code[i].unsqueeze(0),
2,
dim=1,
keepdim=True)
rnn_code_norm = torch.norm(rnn_code,
2,
dim=1,
keepdim=True)
norm = torch.mm(cnn_code_norm,
rnn_code_norm.transpose(0, 1))
scores0 = scores / norm.clamp(min=1e-8)
if torch.argmax(scores0) == 0:
R[R_count] = 1
R_count += 1
if R_count >= 30000:
sum = np.zeros(10)
np.random.shuffle(R)
for i in range(10):
sum[i] = np.average(R[i * 3000:(i + 1) * 3000 - 1])
R_mean = np.average(sum)
R_std = np.std(sum)
print("R mean:{:.4f} std:{:.4f}".format(R_mean, R_std))
cont = False
def gen_example(n_words, wordtoix, ixtoword, model_dir):
'''generate images from example sentences'''
# filepath = 'example_captions.txt'
filepath = 'caption.txt'
data_dic = {}
with open(filepath, "r") as f:
filenames = f.read().split('\n')
captions = []
cap_lens = []
for sent in filenames:
if len(sent) == 0:
continue
sent = sent.replace("\ufffd\ufffd", " ")
tokenizer = RegexpTokenizer(r'\w+')
tokens = tokenizer.tokenize(sent.lower())
if len(tokens) == 0:
print('sentence token == 0 !')
continue
rev = []
for t in tokens:
t = t.encode('ascii', 'ignore').decode('ascii')
if len(t) > 0 and t in wordtoix:
rev.append(wordtoix[t])
captions.append(rev)
cap_lens.append(len(rev))
max_len = np.max(cap_lens)
sorted_indices = np.argsort(cap_lens)[::-1]
cap_lens = np.asarray(cap_lens)
cap_lens = cap_lens[sorted_indices]
cap_array = np.zeros((len(captions), max_len), dtype='int64')
for i in range(len(captions)):
idx = sorted_indices[i]
cap = captions[idx]
c_len = len(cap)
cap_array[i, :c_len] = cap
# key = name[(name.rfind('/') + 1):]
key = 0
data_dic[key] = [cap_array, cap_lens, sorted_indices]
# algo.gen_example(data_dic)
text_encoder = RNN_ENCODER(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.eval()
netG = G_NET()
netG.apply(weights_init)
# netG.cuda()
netG.eval()
state_dict = torch.load(model_dir,
map_location=lambda storage, loc: storage)
netG.load_state_dict(state_dict)
print('Load G from: ', model_dir)
save_dir = 'results'
mkdir_p(save_dir)
for key in data_dic:
captions, cap_lens, sorted_indices = data_dic[key]
batch_size = captions.shape[0]
nz = cfg.GAN.Z_DIM
with torch.no_grad():
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(image_per_caption): # 16
with torch.no_grad():
noise = Variable(torch.FloatTensor(batch_size, nz))
# noise = noise.cuda()
# (1) Extract text embeddings
hidden = text_encoder.init_hidden(batch_size)
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)
cap_lens_np = cap_lens.data.numpy()
for j in range(batch_size):
save_name = '%s/%d_%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]
else:
im = fake_imgs[0]
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]], ixtoword,
[attn_maps[j]], att_sze)
if img_set is not None:
im = Image.fromarray(img_set)
fullpath = '%s_a%d_attention.png' % (save_name, k)
im.save(fullpath)
def gen_samples(self, idx):
text_encoder = 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: {}'.format(cfg.TRAIN.NET_E))
text_encoder = text_encoder.cuda()
text_encoder.eval()
netG = G_NET()
state_dict = torch.load(cfg.TRAIN.NET_G,
map_location=lambda storage, loc: storage)
netG.load_state_dict(state_dict)
print('Load G from: {}'.format(cfg.TRAIN.NET_G))
netG.cuda()
netG.eval()
s_tmp = cfg.TRAIN.NET_G[:cfg.TRAIN.NET_G.rfind('.pth')]
save_dir = '%s/samples' % (s_tmp)
mkdir_p(save_dir)
batch_size = self.batch_size
nz = cfg.GAN.Z_DIM
with torch.no_grad():
noise = Variable(torch.FloatTensor(batch_size, nz))
noise = noise.cuda()
step = 0
data_iter = iter(self.data_loader)
while step < self.num_batches:
data = data_iter.next()
imgs, captions, cap_lens, class_ids, sorted_cap_indices = self.prepare_data(
data)
hidden = text_encoder.init_hidden(batch_size)
words_embs, sent_emb = text_encoder(captions, cap_lens, hidden)
mask = (captions == 0)
num_words = words_embs.size(2)
if mask.size(1) > num_words:
mask = mask[:, :num_words]
for i in range(10):
noise.data.normal_(0, 1)
fake_imgs, attention_maps, _, _ = netG(noise, sent_emb,
words_embs, mask)
cap_lens_np = cap_lens.cpu().data.numpy()
for j in range(batch_size):
right_idx = step * batch_size + sorted_cap_indices[j]
save_name = '%s/%d_s_%d' % (save_dir, i, right_idx)
original_idx = idx[right_idx]
shutil.copyfile(
'/.local/AttnGAN/data/FashionSynthesis/test/original/test128_{}.png'
.format(original_idx + 1),
save_dir + '/test128_{0}_{1}.png'.format(
original_idx + 1, right_idx))
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)
im = np.transpose(im, (1, 2, 0))
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)
step += 1
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
batch_size = 16
text_encoder = \
RNN_ENCODER(self.n_words, nhidden=cfg.TEXT.EMBEDDING_DIM)
print("=======self.n_words: %d", self.n_words)
state_dict = \
torch.load(cfg.TRAIN.NET_E, map_location=lambda storage, loc: storage)
# customed restore text encoder parameters
# ext_encoder.load_state_dict(state_dict)
own_state = text_encoder.state_dict()
for name, param in state_dict.items():
if name not in own_state:
continue
own_state[name] = param
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(text_encoder)
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]
total_time = len(captions)//batch_size
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()
with torch.no_grad():
for i in range(total_time): # 16
noise = Variable(torch.FloatTensor(batch_size, nz))
noise = noise.cuda()
caption_tmp = Variable(torch.from_numpy(captions[i*batch_size:(i+1)*batch_size]))
if i < 3:
print(caption_tmp.data)
cap_len_tmp = Variable(torch.from_numpy(cap_lens[i*batch_size:(i+1)*batch_size]))
caption_tmp = caption_tmp.cuda()
cap_len_tmp = cap_len_tmp.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(caption_tmp, cap_len_tmp, None)
words_embs, sent_emb = words_embs.detach(), sent_emb.detach()
mask = (caption_tmp == 0)
#######################################################
# (2) Generate fake images
######################################################
random.seed(datetime.now())
rnd= random.randint(0,1000)
torch.cuda.manual_seed(rnd)
noise.data.normal_(0, 1)
fake_imgs, attention_maps, _, _, _ = netG(noise, sent_emb, words_embs, mask, caption_tmp, cap_len_tmp)
# G attention
# cap_lens_np = cap_lens.cpu().data.numpy()
cap_lens_np = cap_len_tmp.cpu().data.numpy()
for j in range(batch_size):
save_name = '%s/s_%d' % (save_dir, sorted_indices[i*batch_size+j])
for k in range(len(fake_imgs)):
im = fake_imgs[k][j].data.cpu().numpy()
im = ((im + 1.0) / 2)* 255.0
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)
# save to seperate directory
save_dir2 = '%s/stage_%d' % (save_dir, k)
mkdir_p(save_dir2)
fullpath = '%s/%d_g%d.png' % (save_dir2, sorted_indices[i*batch_size+j], 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),
caption_tmp[j].unsqueeze(0),
[cap_len_tmp[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 sample(self, split_dir, num_samples=25, draw_bbox=False):
from PIL import Image, ImageDraw, ImageFont
import cPickle as pickle
import torchvision
import torchvision.utils as vutils
if cfg.TRAIN.NET_G == '':
print('Error: the path for model NET_G is not found!')
else:
if split_dir == 'test':
split_dir = 'valid'
# Build and load the generator
text_encoder = 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 = cfg.TRAIN.BATCH_SIZE
nz = cfg.GAN.Z_DIM
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 = G_NET()
print('Load G from: ', model_dir)
netG.apply(weights_init)
netG.load_state_dict(state_dict["netG"])
netG.cuda()
netG.eval()
# 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)
#######################################
noise = Variable(torch.FloatTensor(9, nz))
imsize = 256
for step, data in enumerate(self.data_loader, 0):
if step >= num_samples:
break
imgs, captions, cap_lens, class_ids, keys, transformation_matrices, label_one_hot, bbox = \
prepare_data(data, eval=True)
transf_matrices_inv = transformation_matrices[1][0].unsqueeze(0)
label_one_hot = label_one_hot[0].unsqueeze(0)
img = imgs[-1][0]
val_image = img.view(1, 3, imsize, imsize)
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[0].unsqueeze(0).detach(), sent_emb[0].unsqueeze(0).detach()
words_embs = words_embs.repeat(9, 1, 1)
sent_emb = sent_emb.repeat(9, 1)
mask = (captions == 0)
mask = mask[0].unsqueeze(0)
num_words = words_embs.size(2)
if mask.size(1) > num_words:
mask = mask[:, :num_words]
mask = mask.repeat(9, 1)
transf_matrices_inv = transf_matrices_inv.repeat(9, 1, 1, 1)
label_one_hot = label_one_hot.repeat(9, 1, 1)
#######################################################
# (2) Generate fake images
######################################################
noise.data.normal_(0, 1)
inputs = (noise, sent_emb, words_embs, mask, transf_matrices_inv, label_one_hot)
with torch.no_grad():
fake_imgs, _, mu, logvar = nn.parallel.data_parallel(netG, inputs, self.gpus)
data_img = torch.FloatTensor(10, 3, imsize, imsize).fill_(0)
data_img[0] = val_image
data_img[1:10] = fake_imgs[-1]
if draw_bbox:
for idx in range(3):
x, y, w, h = tuple([int(imsize*x) for x in bbox[0, idx]])
w = imsize-1 if w > imsize-1 else w
h = imsize-1 if h > imsize-1 else h
if x <= -1:
break
data_img[:10, :, y, x:x + w] = 1
data_img[:10, :, y:y + h, x] = 1
data_img[:10, :, y+h, x:x + w] = 1
data_img[:10, :, y:y + h, x + w] = 1
# get caption
cap = captions[0].data.cpu().numpy()
sentence = ""
for j in range(len(cap)):
if cap[j] == 0:
break
word = self.ixtoword[cap[j]].encode('ascii', 'ignore').decode('ascii')
sentence += word + " "
sentence = sentence[:-1]
vutils.save_image(data_img, '{}/{}_{}.png'.format(save_dir, sentence, step), normalize=True, nrow=10)
print("Saved {} files to {}".format(step, save_dir))
def sampling(self, split_dir, num_samples=30000):
if cfg.TRAIN.NET_G == '':
logger.error('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.to(cfg.DEVICE)
netG.eval()
#
text_encoder = 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)
text_encoder = text_encoder.to(cfg.DEVICE)
text_encoder.eval()
logger.info('Loaded text encoder from: %s', cfg.TRAIN.NET_E)
batch_size = self.batch_size[0]
nz = cfg.GAN.GLOBAL_Z_DIM
noise = Variable(torch.FloatTensor(batch_size, nz)).to(cfg.DEVICE)
local_noise = Variable(torch.FloatTensor(batch_size, cfg.GAN.LOCAL_Z_DIM)).to(cfg.DEVICE)
model_dir = cfg.TRAIN.NET_G
state_dict = torch.load(model_dir, map_location=lambda storage, loc: storage)
netG.load_state_dict(state_dict["netG"])
max_objects = 10
logger.info('Load G from: %s', model_dir)
# the path to save generated images
s_tmp = model_dir[:model_dir.rfind('.pth')].split("/")[-1]
save_dir = '%s/%s/%s' % ("../output", s_tmp, split_dir)
mkdir_p(save_dir)
logger.info("Saving images to: {}".format(save_dir))
number_batches = num_samples // batch_size
if number_batches < 1:
number_batches = 1
data_iter = iter(self.data_loader)
for step in tqdm(range(number_batches)):
data = data_iter.next()
imgs, captions, cap_lens, class_ids, keys, transformation_matrices, label_one_hot, _ = prepare_data(
data, eval=True)
transf_matrices = transformation_matrices[0]
transf_matrices_inv = transformation_matrices[1]
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)
local_noise.data.normal_(0, 1)
inputs = (noise, local_noise, sent_emb, words_embs, mask, transf_matrices, transf_matrices_inv, label_one_hot, max_objects)
inputs = tuple((inp.to(cfg.DEVICE) if isinstance(inp, torch.Tensor) else inp) for inp in inputs)
with torch.no_grad():
fake_imgs, _, mu, logvar = netG(*inputs)
for batch_idx, j in enumerate(range(batch_size)):
s_tmp = '%s/%s' % (save_dir, keys[j])
folder = s_tmp[:s_tmp.rfind('/')]
if not os.path.isdir(folder):
logger.info('Make a new folder: %s', 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, step*batch_size+batch_idx)
im.save(fullpath)
def embedding(self, split_dir, model):
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)
if cfg.GPU_ID != -1:
netG.cuda()
netG.eval()
#
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)
image_encoder = CNN_ENCODER(cfg.TEXT.EMBEDDING_DIM)
img_encoder_path = cfg.TRAIN.NET_E.replace('text_encoder',
'image_encoder')
print(img_encoder_path)
print('Load image encoder from:', img_encoder_path)
state_dict = \
torch.load(img_encoder_path, map_location=lambda storage, loc: storage)
image_encoder.load_state_dict(state_dict)
if cfg.GPU_ID != -1:
image_encoder = image_encoder.cuda()
image_encoder.eval()
print('Load text encoder from:', cfg.TRAIN.NET_E)
text_encoder = 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)
if cfg.GPU_ID != -1:
text_encoder = text_encoder.cuda()
text_encoder.eval()
batch_size = self.batch_size
nz = cfg.GAN.Z_DIM
with torch.no_grad():
noise = Variable(torch.FloatTensor(batch_size, nz))
if cfg.GPU_ID != -1:
noise = noise.cuda()
# the path to save generated images
save_dir = model_dir[:model_dir.rfind('.pth')]
cnt = 0
# new
if cfg.TRAIN.CLIP_SENTENCODER:
print("Use CLIP SentEncoder for sampling")
img_features = dict()
txt_features = dict()
with torch.no_grad():
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)
imgs, captions, cap_lens, class_ids, keys, texts = 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]
if cfg.TRAIN.CLIP_SENTENCODER:
# random select one paragraph for each training example
sents = []
for idx in range(len(texts)):
sents_per_image = texts[idx].split(
'\n') # new 3/11
if len(sents_per_image) > 1:
sent_ix = np.random.randint(
0,
len(sents_per_image) - 1)
else:
sent_ix = 0
sents.append(sents_per_image[0])
# print('sents: ', sents)
sent = clip.tokenize(sents) # .to(device)
# load clip
#model = torch.jit.load("model.pt").cuda().eval()
sent_input = sent
if cfg.GPU_ID != -1:
sent_input = sent.cuda()
# print("text input", sent_input)
sent_emb_clip = model.encode_text(
sent_input).float()
if CLIP:
sent_emb = sent_emb_clip
#######################################################
# (2) Generate fake images
######################################################
noise.data.normal_(0, 1)
fake_imgs, _, _, _ = netG(noise, sent_emb, words_embs,
mask)
if CLIP:
images = []
for j in range(fake_imgs[-1].shape[0]):
image = fake_imgs[-1][j].cpu().clone()
image = image.squeeze(0)
unloader = transforms.ToPILImage()
image = unloader(image)
image = preprocess(
image.convert("RGB")) # 256*256 -> 224*224
images.append(image)
image_mean = torch.tensor(
[0.48145466, 0.4578275, 0.40821073]).cuda()
image_std = torch.tensor(
[0.26862954, 0.26130258, 0.27577711]).cuda()
image_input = torch.tensor(np.stack(images)).cuda()
image_input -= image_mean[:, None, None]
image_input /= image_std[:, None, None]
cnn_codes = model.encode_image(image_input).float()
else:
region_features, cnn_codes = image_encoder(
fake_imgs[-1])
for j in range(batch_size):
cnn_code = cnn_codes[j]
temp = keys[j].replace('b', '').replace("'", '')
img_features[temp] = cnn_code.cpu().numpy()
txt_features[temp] = sent_emb[j].cpu().numpy()
with open(save_dir + ".pkl", 'wb') as f:
pickle.dump(img_features, f)
with open(save_dir + "_text.pkl", 'wb') as f:
pickle.dump(txt_features, f)
def gen_example(self, data_dic):
if cfg.TRAIN.NET_G == '' or cfg.TRAIN.NET_C == '':
print('Error: the path for main module or DCM is not found!')
else:
# The text encoder
text_encoder = \
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 image encoder
image_encoder = CNN_ENCODER(cfg.TEXT.EMBEDDING_DIM)
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)
print('Load image encoder from:', img_encoder_path)
image_encoder = image_encoder.cuda()
image_encoder.eval()
# The VGG network
VGG = VGGNet()
print("Load the VGG model")
VGG.cuda()
VGG.eval()
# The main module
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()
# The DCM
netDCM = DCM_Net()
if cfg.TRAIN.NET_C != '':
state_dict = \
torch.load(cfg.TRAIN.NET_C, map_location=lambda storage, loc: storage)
netDCM.load_state_dict(state_dict)
print('Load DCM from: ', cfg.TRAIN.NET_C)
netDCM.cuda()
netDCM.eval()
for key in data_dic:
save_dir = '%s/%s' % (s_tmp, key)
mkdir_p(save_dir)
captions, cap_lens, sorted_indices, imgs = 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):
noise = Variable(torch.FloatTensor(batch_size, nz),
volatile=True)
noise = noise.cuda()
#######################################################
# (1) Extract text and image embeddings
######################################################
hidden = text_encoder.init_hidden(batch_size)
# The text embeddings
words_embs, sent_emb = text_encoder(
captions, cap_lens, hidden)
# The image embeddings
region_features, cnn_code = \
image_encoder(imgs[cfg.TREE.BRANCH_NUM - 1].unsqueeze(0))
mask = (captions == 0)
#######################################################
# (2) Modify real images
######################################################
noise.data.normal_(0, 1)
fake_imgs, attention_maps, mu, logvar, h_code, c_code = netG(
noise, sent_emb, words_embs, mask, cnn_code,
region_features)
real_img = imgs[cfg.TREE.BRANCH_NUM - 1].unsqueeze(0)
real_features = VGG(real_img)[0]
fake_img = netDCM(h_code, real_features, sent_emb, words_embs, \
mask, c_code)
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)
im = np.transpose(im, (1, 2, 0))
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)
save_name = '%s/%d_sf_%d' % (save_dir, 1,
sorted_indices[j])
im = fake_img[j].data.cpu().numpy()
im = (im + 1.0) * 127.5
im = im.astype(np.uint8)
im = np.transpose(im, (1, 2, 0))
im = Image.fromarray(im)
fullpath = '%s_SF.png' % (save_name)
im.save(fullpath)
save_name = '%s/%d_s_%d' % (save_dir, 1, 9)
im = imgs[2].data.cpu().numpy()
im = (im + 1.0) * 127.5
im = im.astype(np.uint8)
im = np.transpose(im, (1, 2, 0))
im = Image.fromarray(im)
fullpath = '%s_SR.png' % (save_name)
im.save(fullpath)
def sampling(self, split_dir):
if cfg.TRAIN.NET_G == '' or cfg.TRAIN.NET_C == '':
print('Error: the path for main module or DCM is not found!')
else:
if split_dir == 'test':
split_dir = 'valid'
if cfg.GAN.B_DCGAN:
netG = G_DCGAN()
else:
netG = G_NET()
netG.apply(weights_init)
netG.cuda()
netG.eval()
# The text encoder
text_encoder = 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 image encoder
image_encoder = CNN_ENCODER(cfg.TEXT.EMBEDDING_DIM)
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)
print('Load image encoder from:', img_encoder_path)
image_encoder = image_encoder.cuda()
image_encoder.eval()
# The VGG network
VGG = VGGNet()
print("Load the VGG model")
VGG.cuda()
VGG.eval()
batch_size = self.batch_size
nz = cfg.GAN.Z_DIM
noise = Variable(torch.FloatTensor(batch_size, nz), volatile=True)
noise = noise.cuda()
# The DCM
netDCM = DCM_Net()
if cfg.TRAIN.NET_C != '':
state_dict = \
torch.load(cfg.TRAIN.NET_C, map_location=lambda storage, loc: storage)
netDCM.load_state_dict(state_dict)
print('Load DCM from: ', cfg.TRAIN.NET_C)
netDCM.cuda()
netDCM.eval()
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)
# the path to save modified images
s_tmp = model_dir[:model_dir.rfind('.pth')]
save_dir = '%s/%s' % (s_tmp, split_dir)
mkdir_p(save_dir)
cnt = 0
idx = 0
for _ in range(5): # (cfg.TEXT.CAPTIONS_PER_IMAGE):
for step, data in enumerate(self.data_loader, 0):
cnt += batch_size
if step % 100 == 0:
print('step: ', step)
imgs, captions, cap_lens, class_ids, keys, wrong_caps, \
wrong_caps_len, wrong_cls_id = prepare_data(data)
#######################################################
# (1) Extract text and image embeddings
######################################################
hidden = text_encoder.init_hidden(batch_size)
words_embs, sent_emb = text_encoder(
wrong_caps, wrong_caps_len, hidden)
words_embs, sent_emb = words_embs.detach(
), sent_emb.detach()
mask = (wrong_caps == 0)
num_words = words_embs.size(2)
if mask.size(1) > num_words:
mask = mask[:, :num_words]
region_features, cnn_code = \
image_encoder(imgs[cfg.TREE.BRANCH_NUM - 1])
#######################################################
# (2) Modify real images
######################################################
noise.data.normal_(0, 1)
fake_imgs, attention_maps, mu, logvar, h_code, c_code = netG(
noise, sent_emb, words_embs, mask, cnn_code,
region_features)
real_img = imgs[cfg.TREE.BRANCH_NUM - 1]
real_features = VGG(real_img)[0]
fake_img = netDCM(h_code, real_features, sent_emb, words_embs, \
mask, c_code)
for j in range(batch_size):
s_tmp = '%s/single' % (save_dir)
folder = s_tmp[:s_tmp.rfind('/')]
if not os.path.isdir(folder):
print('Make a new folder: ', folder)
mkdir_p(folder)
k = -1
im = fake_img[j].data.cpu().numpy()
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, idx)
idx = idx + 1
im.save(fullpath)
class Generator:
def __init__(self, caption_file, saveable, cuda=False, profile=False):
# flags
self.cuda = cuda
self.profile = profile
if self.profile:
print('Initializing Generator...')
print('cuda={}\nprofile={}'.format(self.cuda, self.profile))
# load caption indices
x = pickle.load(open(caption_file, 'rb'))
self.ixtoword = x[2]
self.wordtoix = x[3]
del x
# load text encoder
self.text_encoder = RNN_ENCODER(len(self.wordtoix), nhidden=cfg.TEXT.EMBEDDING_DIM)
state_dict = torch.load(cfg.TRAIN.NET_E, map_location=lambda storage, loc: storage)
self.text_encoder.load_state_dict(state_dict)
if self.cuda:
self.text_encoder.cuda()
self.text_encoder.eval()
# load generative model
self.netG = G_NET()
state_dict = torch.load(cfg.TRAIN.NET_G, map_location=lambda storage, loc: storage)
self.netG.load_state_dict(state_dict)
if self.cuda:
self.netG.cuda()
self.netG.eval()
# saveable items -> push to storage
self.saveable = saveable
def vectorize_caption(self, caption, copies):
# create caption vector
tokens = caption.split(' ')
cap_v = []
for t in tokens:
t = t.strip().encode('ascii', 'ignore').decode('ascii')
if len(t) > 0 and t in self.wordtoix:
cap_v.append(self.wordtoix[t])
# expected state for single generation
captions = np.zeros((copies, len(cap_v)))
for i in range(copies):
captions[i,:] = np.array(cap_v)
cap_lens = np.zeros(copies) + len(cap_v)
return captions.astype(int), cap_lens.astype(int), len(self.wordtoix)
def generate(self, caption, copies=2):
# load word vector
captions, cap_lens, n_words = self.vectorize_caption(caption, copies)
# only one to generate
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)
noise = Variable(torch.FloatTensor(batch_size, nz), volatile=True)
if self.cuda:
captions = captions.cuda()
cap_lens = cap_lens.cuda()
noise = noise.cuda()
#######################################################
# (1) Extract text embeddings
#######################################################
hidden = self.text_encoder.init_hidden(batch_size)
words_embs, sent_emb = self.text_encoder(captions, cap_lens, hidden)
mask = (captions == 0)
#######################################################
# (2) Generate fake images
#######################################################
noise.data.normal_(0, 1)
fake_imgs, attention_maps, _, _ = self.netG(noise, sent_emb, words_embs, mask)
# G attention
cap_lens_np = cap_lens.cpu().data.numpy()
# prefix for partitioning images
prefix = datetime.now().strftime('%Y/%B/%d/%H_%M_%S_%f')
urls = []
# only look at first one
for j in range(batch_size):
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)
im = np.transpose(im, (1, 2, 0))
# save using saveable
birdy = 'bird_g{}'.format(k)
if copies > 2:
item = self.saveable.save('{}/{}'.format(prefix, j), birdy, im)
else:
item = self.saveable.save(prefix, birdy, im)
urls.append(item)
if copies == 2:
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:
attnmap = 'attmaps_a{}'.format(k)
item = self.saveable.save(prefix, attnmap, img_set)
urls.append(item)
if copies == 2:
break
return urls
if use_cuda:
noise, fixed_noise = noise.cuda(), fixed_noise.cuda()
for i in range(start_epoch, args.max_epoch):
iterator_train = iter(train_set)
train_step = 0
num_batches = len(train_set)
while train_step < num_batches:
train_step = train_step + 1
data_train = iterator_train.next()
imgs, captions, cap_lens, class_ids, keys = prepare_data(data_train)
hidden = text_encoder.init_hidden(args.batch_size)
# words_embs: batch_size x nef x seq_len
# sent_emb: batch_size x nef
# words_embs: torch.Size([4, 256, WORDS_NUM])
# sent_emb: torch.Size([4, 256])
words_embs, sent_emb = text_encoder(captions, cap_lens, hidden)
words_embs, sent_emb = words_embs.detach(), sent_emb.detach()
text_mask = (captions == 0)
num_words = words_embs.size(2)
if text_mask.size(1) > num_words:
text_mask = text_mask[:, :num_words]
img = imgs[-1]
mask = get_mask()
masked = img * (1. - mask)
def sampling(self, split_dir):
if cfg.TRAIN.NET_G == '':
print('Error: the path for models is not found!')
else:
if split_dir == 'test':
split_dir = 'valid'
if cfg.GAN.B_DCGAN:
netG = G_DCGAN()
else:
netG = G_NET()
netG.apply(weights_init)
netG.cuda()
netG.eval()
#
text_encoder = 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)
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
idx = 0 ###
for _ in range(1):
for step, data in enumerate(self.data_loader, 0):
cnt += batch_size
if step % 100 == 0:
print('step: ', step)
imgs, captions, cap_lens, class_ids, keys, wrong_caps, \
wrong_caps_len, wrong_cls_id = prepare_data(data)
hidden = text_encoder.init_hidden(batch_size)
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]
noise.data.normal_(0, 1)
fake_imgs, _, _, _ = netG(noise, sent_emb, words_embs,
mask)
for j in range(batch_size):
s_tmp = '%s/single' % (save_dir)
folder = s_tmp[:s_tmp.rfind('/')]
if not os.path.isdir(folder):
print('Make a new folder: ', folder)
mkdir_p(folder)
k = -1
im = fake_imgs[k][j].data.cpu().numpy()
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, idx)
idx = idx + 1
im.save(fullpath)
def genDiscOutputs(self, split_dir, num_samples=57140):
if cfg.TRAIN.NET_G == '':
logger.error('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.to(cfg.DEVICE)
netG.eval()
#
text_encoder = RNN_ENCODER(self.n_words,
nhidden=cfg.TEXT.EMBEDDING_DIM) ###HACK
state_dict = torch.load(cfg.TRAIN.NET_E,
map_location=lambda storage, loc: storage)
text_encoder.load_state_dict(state_dict)
text_encoder = text_encoder.to(cfg.DEVICE)
text_encoder.eval()
logger.info('Loaded text encoder from: %s', cfg.TRAIN.NET_E)
batch_size = self.batch_size[0]
nz = cfg.GAN.GLOBAL_Z_DIM
noise = Variable(torch.FloatTensor(batch_size, nz)).to(cfg.DEVICE)
local_noise = Variable(
torch.FloatTensor(batch_size,
cfg.GAN.LOCAL_Z_DIM)).to(cfg.DEVICE)
model_dir = cfg.TRAIN.NET_G
state_dict = torch.load(model_dir,
map_location=lambda storage, loc: storage)
netG.load_state_dict(state_dict["netG"])
for keys in state_dict.keys():
print(keys)
logger.info('Load G from: %s', model_dir)
max_objects = 3
from model import D_NET256
netD = D_NET256()
netD.load_state_dict(state_dict["netD"][2])
netD.eval()
# the path to save generated images
s_tmp = model_dir[:model_dir.rfind('.pth')].split("/")[-1]
save_dir = '%s/%s/%s' % ("../output", s_tmp, split_dir)
mkdir_p(save_dir)
logger.info("Saving images to: {}".format(save_dir))
number_batches = num_samples // batch_size
if number_batches < 1:
number_batches = 1
data_iter = iter(self.data_loader)
real_labels, fake_labels, match_labels = self.prepare_labels()
for step in tqdm(range(number_batches)):
data = data_iter.next()
imgs, captions, cap_lens, class_ids, keys, transformation_matrices, label_one_hot, _ = prepare_data(
data, eval=True)
transf_matrices = transformation_matrices[0]
transf_matrices_inv = transformation_matrices[1]
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)
local_noise.data.normal_(0, 1)
inputs = (noise, local_noise, sent_emb, words_embs, mask,
transf_matrices, transf_matrices_inv, label_one_hot,
max_objects)
inputs = tuple(
(inp.to(cfg.DEVICE) if isinstance(inp, torch.Tensor
) else inp)
for inp in inputs)
with torch.no_grad():
fake_imgs, _, mu, logvar = netG(*inputs)
inputs = (fake_imgs, fake_labels, transf_matrices,
transf_matrices_inv, max_objects)
codes = netsD[-1].partial_forward(*inputs)
def sampling(self, split_dir):
if cfg.TRAIN.NET_G == '':
print('Error: the path for morels is not found!')
else:
netG_list = [
'../models/netG_epoch_50.pth', '../models/netG_epoch_60.pth',
'../models/netG_epoch_70.pth', '../models/netG_epoch_80.pth',
'../models/netG_epoch_90.pth', '../models/netG_epoch_100.pth',
'../models/netG_epoch_110.pth', '../models/netG_epoch_120.pth',
'../models/netG_epoch_130.pth', '../models/netG_epoch_140.pth',
'../models/netG_epoch_150.pth', '../models/netG_epoch_160.pth'
]
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 = 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)
print("LINE==380")
print("-----------------netG------------------------")
print(netG)
print("--------------state-dict---------------------")
#print(state_dict)
netG.load_state_dict(state_dict)
print('Load G from: ', model_dir)
print(
'!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!'
)
# the path to save generated images
s_tmp = model_dir[:model_dir.rfind('.pth')]
save_dir = '%s/%s' % (s_tmp, split_dir)
print('save_dir:', save_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 sampling(self, split_dir, model):
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)
if cfg.GPU_ID != -1:
netG.cuda()
netG.eval()
#
text_encoder = 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)
if cfg.GPU_ID != -1:
text_encoder = text_encoder.cuda()
text_encoder.eval()
batch_size = self.batch_size
nz = cfg.GAN.Z_DIM
with torch.no_grad():
noise = Variable(torch.FloatTensor(batch_size, nz))
if cfg.GPU_ID != -1:
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
#new
if cfg.TRAIN.CLIP_SENTENCODER:
print("Use CLIP SentEncoder for sampling")
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)
#new
imgs, captions, cap_lens, class_ids, keys, texts = 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]
# new
if cfg.TRAIN.CLIP_SENTENCODER:
# random select one paragraph for each training example
sents = []
for idx in range(len(texts)):
sents_per_image = texts[idx].split(
'\n') # new 3/11
if len(sents_per_image) > 1:
sent_ix = np.random.randint(
0,
len(sents_per_image) - 1)
else:
sent_ix = 0
sents.append(sents_per_image[sent_ix])
with open('%s/%s' % (save_dir, 'eval_sents.txt'),
'a+') as f:
f.write(sents_per_image[sent_ix] + '\n')
# print('sents: ', sents)
sent = clip.tokenize(sents) # .to(device)
# load clip
#model = torch.jit.load("model.pt").cuda().eval()
sent_input = sent
if cfg.GPU_ID != -1:
sent_input = sent.cuda()
# print("text input", sent_input)
with torch.no_grad():
sent_emb = model.encode_text(sent_input).float()
#######################################################
# (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/fake/%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)
print('Make a new folder: ', f'{save_dir}/real')
mkdir_p(f'{save_dir}/real')
print('Make a new folder: ', f'{save_dir}/text')
mkdir_p(f'{save_dir}/text')
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)
temp = keys[j].replace('b', '').replace("'", '')
shutil.copy(f"../data/Face/images/{temp}.jpg",
f"{save_dir}/real/")
shutil.copy(f"../data/Face/text/{temp}.txt",
f"{save_dir}/text/")
def calc_mp(self):
if cfg.TRAIN.NET_G == '':
print('Error: the path for main module is not found!')
else:
#if split_dir == 'test':
# split_dir = 'valid'
if cfg.GAN.B_DCGAN:
netG = G_DCGAN()
else:
netG = EncDecNet()
netG.apply(weights_init)
netG.cuda()
netG.eval()
# The text encoder
text_encoder = 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 image encoder
#image_encoder = CNN_dummy()
#print('define image_encoder')
image_encoder = CNN_ENCODER(cfg.TEXT.EMBEDDING_DIM)
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)
print('Load image encoder from:', img_encoder_path)
image_encoder = image_encoder.cuda()
image_encoder.eval()
# The VGG network
VGG = VGG16()
print("Load the VGG model")
#VGG.to(torch.device("cuda:1"))
VGG.cuda()
VGG.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 = os.path.join(cfg.DATA_DIR, 'output', self.args.netG,
'Model', self.args.netG_epoch)
state_dict = \
torch.load(model_dir, map_location=lambda storage, loc: storage)
netG.load_state_dict(state_dict)
print('Load G from: ', model_dir)
# the path to save modified images
cnt = 0
idx = 0
diffs, sims = [], []
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)
imgs, w_imgs, captions, cap_lens, class_ids, keys, wrong_caps, \
wrong_caps_len, wrong_cls_id = prepare_data(data)
#######################################################
# (1) Extract text and image embeddings
######################################################
hidden = text_encoder.init_hidden(batch_size)
words_embs, sent_emb = text_encoder(
wrong_caps, wrong_caps_len, hidden)
words_embs, sent_emb = words_embs.detach(
), sent_emb.detach()
mask = (wrong_caps == 0)
num_words = words_embs.size(2)
if mask.size(1) > num_words:
mask = mask[:, :num_words]
#######################################################
# (2) Modify real images
######################################################
noise.data.normal_(0, 1)
fake_img, mu, logvar = netG(imgs[-1], sent_emb, words_embs,
noise, mask, VGG)
diff = F.l1_loss(fake_img, imgs[-1])
diffs.append(diff.item())
region_code, cnn_code = image_encoder(fake_img)
sim = cosine_similarity(sent_emb, cnn_code)
sim = torch.mean(sim)
sims.append(sim.item())
diff = np.sum(diffs) / len(diffs)
sim = np.sum(sims) / len(sims)
print('diff: %.3f, sim:%.3f' % (diff, sim))
print('MP: %.3f' % ((1 - diff) * sim))
netG_epoch = self.args.netG_epoch[self.args.netG_epoch.find('_') +
1:-4]
print('model_epoch:%s, diff: %.3f, sim:%.3f, MP:%.3f' %
(netG_epoch, np.sum(diffs) / len(diffs),
np.sum(sims) / len(sims), (1 - diff) * sim))
def sampling(self, split_dir):
if cfg.TRAIN.NET_G == '':
print('Error: the path for models is not found!')
else:
if split_dir == 'test':
split_dir = 'valid'
if cfg.GAN.B_DCGAN:
netG = G_DCGAN()
else:
netG = G_NET()
netG.apply(weights_init)
netG.cuda()
netG.eval()
#
text_encoder = 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)
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
idx = 0 ###
avg_ddva = 0
for _ in range(1):
for step, data in enumerate(self.data_loader, 0):
cnt += batch_size
if step % 100 == 0:
print('step: ', step)
captions, cap_lens, imperfect_captions, imperfect_cap_lens, misc = data
# Generate images for human-text ----------------------------------------------------------------
data_human = [captions, cap_lens, misc]
imgs, captions, cap_lens, class_ids, keys, wrong_caps,\
wrong_caps_len, wrong_cls_id= prepare_data(data_human)
hidden = text_encoder.init_hidden(batch_size)
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]
noise.data.normal_(0, 1)
fake_imgs, _, _, _ = netG(noise, sent_emb, words_embs,
mask)
# Generate images for imperfect caption-text-------------------------------------------------------
data_imperfect = [
imperfect_captions, imperfect_cap_lens, misc
]
imgs, imperfect_captions, imperfect_cap_lens, class_ids, imperfect_keys, wrong_caps,\
wrong_caps_len, wrong_cls_id = prepare_data(data_imperfect)
hidden = text_encoder.init_hidden(batch_size)
words_embs, sent_emb = text_encoder(
imperfect_captions, imperfect_cap_lens, hidden)
words_embs, sent_emb = words_embs.detach(
), sent_emb.detach()
mask = (imperfect_captions == 0)
num_words = words_embs.size(2)
if mask.size(1) > num_words:
mask = mask[:, :num_words]
noise.data.normal_(0, 1)
imperfect_fake_imgs, _, _, _ = netG(
noise, sent_emb, words_embs, mask)
# Sort the results by keys to align ----------------------------------------------------------------
keys, captions, cap_lens, fake_imgs, _, _ = sort_by_keys(
keys, captions, cap_lens, fake_imgs, None, None)
imperfect_keys, imperfect_captions, imperfect_cap_lens, imperfect_fake_imgs, true_imgs, _ = \
sort_by_keys(imperfect_keys, imperfect_captions, imperfect_cap_lens, imperfect_fake_imgs,\
imgs, None)
# Shift device for the imgs, target_imgs and imperfect_imgs------------------------------------------------
for i in range(len(imgs)):
imgs[i] = imgs[i].to(secondary_device)
imperfect_fake_imgs[i] = imperfect_fake_imgs[i].to(
secondary_device)
fake_imgs[i] = fake_imgs[i].to(secondary_device)
for j in range(batch_size):
s_tmp = '%s/single' % (save_dir)
folder = s_tmp[:s_tmp.rfind('/')]
if not os.path.isdir(folder):
print('Make a new folder: ', folder)
mkdir_p(folder)
k = -1
im = fake_imgs[k][j].data.cpu().numpy()
im = (im + 1.0) * 127.5
im = im.astype(np.uint8)
im = np.transpose(im, (1, 2, 0))
cap_im = imperfect_fake_imgs[k][j].data.cpu().numpy()
cap_im = (cap_im + 1.0) * 127.5
cap_im = cap_im.astype(np.uint8)
cap_im = np.transpose(cap_im, (1, 2, 0))
# Uncomment to scale true image
true_im = true_imgs[k][j].data.cpu().numpy()
true_im = (true_im + 1.0) * 127.5
true_im = true_im.astype(np.uint8)
true_im = np.transpose(true_im, (1, 2, 0))
# Uncomment to save images.
#true_im = Image.fromarray(true_im)
#fullpath = '%s_true_s%d.png' % (s_tmp, idx)
#true_im.save(fullpath)
im = Image.fromarray(im)
fullpath = '%s_s%d.png' % (s_tmp, idx)
im.save(fullpath)
#cap_im = Image.fromarray(cap_im)
#fullpath = '%s_imperfect_s%d.png' % (s_tmp, idx)
idx = idx + 1
#cap_im.save(fullpath)
neg_ddva = negative_ddva(
imperfect_fake_imgs,
imgs,
fake_imgs,
reduce='mean',
final_only=True).data.cpu().numpy()
avg_ddva += neg_ddva * (-1)
#text_caps = [[self.ixtoword[word] for word in sent if word!=0] for sent in captions.tolist()]
#imperfect_text_caps = [[self.ixtoword[word] for word in sent if word!=0] for sent in
# imperfect_captions.tolist()]
print(step)
avg_ddva = avg_ddva / (step + 1)
print('\n\nAvg_DDVA: ', avg_ddva)
def sampling(self, split_dir):
if cfg.TRAIN.NET_G == '':
print('Error: the path for main module is not found!')
else:
if split_dir == 'test':
split_dir = 'valid'
if cfg.GAN.B_DCGAN:
netG = G_DCGAN()
else:
netG = EncDecNet()
netG.apply(weights_init)
netG.cuda()
netG.eval()
# The text encoder
text_encoder = 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 image encoder
"""
image_encoder = CNN_ENCODER(cfg.TEXT.EMBEDDING_DIM)
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)
print('Load image encoder from:', img_encoder_path)
image_encoder = image_encoder.cuda()
image_encoder.eval()
"""
# The VGG network
VGG = VGG16()
print("Load the VGG model")
VGG.cuda()
VGG.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 = os.path.join(cfg.DATA_DIR, 'output', self.args.netG,
'Model/netG_epoch_600.pth')
state_dict = \
torch.load(model_dir, map_location=lambda storage, loc: storage)
netG.load_state_dict(state_dict)
print('Load G from: ', model_dir)
# the path to save modified images
save_dir_valid = os.path.join(cfg.DATA_DIR, 'output',
self.args.netG, 'valid')
#mkdir_p(save_dir)
cnt = 0
idx = 0
for i in range(5): # (cfg.TEXT.CAPTIONS_PER_IMAGE):
# the path to save modified images
save_dir = os.path.join(save_dir_valid, 'valid_%d' % i)
save_dir_super = os.path.join(save_dir, 'super')
save_dir_single = os.path.join(save_dir, 'single')
mkdir_p(save_dir_super)
mkdir_p(save_dir_single)
for step, data in enumerate(self.data_loader, 0):
cnt += batch_size
if step % 100 == 0:
print('step: ', step)
imgs, w_imgs, captions, cap_lens, class_ids, keys, wrong_caps, \
wrong_caps_len, wrong_cls_id = prepare_data(data)
#######################################################
# (1) Extract text and image embeddings
######################################################
hidden = text_encoder.init_hidden(batch_size)
words_embs, sent_emb = text_encoder(
wrong_caps, wrong_caps_len, hidden)
words_embs, sent_emb = words_embs.detach(
), sent_emb.detach()
mask = (wrong_caps == 0)
num_words = words_embs.size(2)
if mask.size(1) > num_words:
mask = mask[:, :num_words]
#######################################################
# (2) Modify real images
######################################################
noise.data.normal_(0, 1)
fake_img, mu, logvar = netG(imgs[-1], sent_emb, words_embs,
noise, mask, VGG)
img_set = build_images(imgs[-1], fake_img, captions,
wrong_caps, self.ixtoword)
img = Image.fromarray(img_set)
full_path = '%s/super_step%d.png' % (save_dir_super, step)
img.save(full_path)
for j in range(batch_size):
s_tmp = '%s/single' % (save_dir_single)
folder = s_tmp[:s_tmp.rfind('/')]
if not os.path.isdir(folder):
print('Make a new folder: ', folder)
mkdir_p(folder)
k = -1
im = fake_img[j].data.cpu().numpy()
#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, idx)
idx = idx + 1
im.save(fullpath)
def sampling(self, split_dir, num_samples=30000):
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 = 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))
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["netG"])
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 % 10000 == 0:
print('step: ', step)
if step >= num_samples:
break
imgs, captions, cap_lens, class_ids, keys, transformation_matrices, label_one_hot = prepare_data(data)
transf_matrices_inv = transformation_matrices[1]
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)
inputs = (noise, sent_emb, words_embs, mask, transf_matrices_inv, label_one_hot)
with torch.no_grad():
fake_imgs, _, mu, logvar = nn.parallel.data_parallel(netG, inputs, self.gpus)
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 gen_example(self, data_dic):
if cfg.TRAIN.NET_G == '' or cfg.TRAIN.NET_C == '':
print('Error: the path for main module or DCM is not found!')
else:
# The text encoder
text_encoder = \
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 image encoder
"""
image_encoder = CNN_ENCODER(cfg.TEXT.EMBEDDING_DIM)
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)
print('Load image encoder from:', img_encoder_path)
image_encoder = image_encoder.cuda()
image_encoder.eval()
"""
"""
image_encoder = CNN_dummy()
image_encoder = image_encoder.cuda()
image_encoder.eval()
"""
# The VGG network
VGG = VGG16()
print("Load the VGG model")
VGG.cuda()
VGG.eval()
# The main module
if cfg.GAN.B_DCGAN:
netG = G_DCGAN()
else:
netG = EncDecNet()
s_tmp = cfg.TRAIN.NET_G[:cfg.TRAIN.NET_G.rfind('.pth')]
s_tmp = os.path.join(cfg.DATA_DIR, 'output', self.args.netG,
'valid/gen_example')
model_dir = os.path.join(cfg.DATA_DIR, 'output', self.args.netG,
'Model/netG_epoch_8.pth')
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 = nn.DataParallel(netG, device_ids= self.gpus)
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, imgs = 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):
noise = Variable(torch.FloatTensor(batch_size, nz),
volatile=True)
noise = noise.cuda()
#######################################################
# (1) Extract text and image embeddings
######################################################
hidden = text_encoder.init_hidden(batch_size)
# The text embeddings
words_embs, sent_emb = text_encoder(
captions, cap_lens, hidden)
words_embs, sent_emb = words_embs.detach(
), sent_emb.detach()
# The image embeddings
mask = (captions == 0)
#######################################################
# (2) Modify real images
######################################################
noise.data.normal_(0, 1)
imgs_256 = imgs[-1].unsqueeze(0).repeat(
batch_size, 1, 1, 1)
enc_features = VGG(imgs_256)
fake_img, mu, logvar = nn.parallel.data_parallel(
netG, (imgs[-1], sent_emb, words_embs, noise, mask,
enc_features), self.gpus)
cap_lens_np = cap_lens.cpu().data.numpy()
one_imgs = []
for j in range(captions.shape[0]):
font = ImageFont.truetype('./FreeMono.ttf', 20)
canv = Image.new('RGB', (256, 256), (255, 255, 255))
draw = ImageDraw.Draw(canv)
sent = []
for k in range(len(captions[j])):
if (captions[j][k] == 0):
break
word = self.ixtoword[captions[j][k].item()].encode(
'ascii', 'ignore').decode('ascii')
if (k % 2 == 1):
word = word + '\n'
sent.append(word)
fake_sent = ' '.join(sent)
draw.text((0, 0), fake_sent, font=font, fill=(0, 0, 0))
canv_np = np.asarray(canv)
real_im = imgs[-1]
real_im = (real_im + 1) * 127.5
real_im = real_im.cpu().numpy().astype(np.uint8)
real_im = np.transpose(real_im, (1, 2, 0))
fake_im = fake_img[j]
fake_im = (fake_im + 1.0) * 127.5
fake_im = fake_im.detach().cpu().numpy().astype(
np.uint8)
fake_im = np.transpose(fake_im, (1, 2, 0))
one_img = np.concatenate([real_im, canv_np, fake_im],
axis=1)
one_imgs.append(one_img)
img_set = np.concatenate(one_imgs, axis=0)
super_img = Image.fromarray(img_set)
full_path = os.path.join(save_dir, 'super.png')
super_img.save(full_path)
"""
for j in range(5): ## 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)
im = np.transpose(im, (1, 2, 0))
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 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 = \
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)
with torch.no_grad():
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)
for epoch in range(60):
train_epoch(cnn, rnn, train_dataloader, epoch, device)
val_epoch(cnn, rnn, val_dataloader, epoch, device)
else:
rnn.eval()
cnn.eval()
checkpoint = torch.load('saved/checkpoint_50.pth',
map_location='cuda:0')['cnn_state_dict']
cnn.load_state_dict(checkpoint)
checkpoint = torch.load('saved/checkpoint_50.pth',
map_location='cuda:0')['rnn_state_dict']
rnn.load_state_dict(checkpoint)
correct = 0
total = len(val_dataloader.dataset)
for batch_idx, (img, caption, cap_len) in enumerate(val_dataloader):
img = img.to(device)
caption = caption.to(device)
cap_len = cap_len.to(device)
features, img_feature = cnn(img)
batch_size = img.size(0)
hidden = rnn.init_hidden(batch_size)
words_emb, text_feature = rnn(caption, cap_len, hidden)
score = torch.mm(img_feature, text_feature.transpose(1, 0))
score = score.argmax(1)
gt = torch.arange(score.size(0)).to(device)
score = torch.sum(score == gt)
correct += score.item()
print(correct / total)