def __init__(self, num_answers):
super().__init__()
# Build LXRT encoder
self.lxrt_encoder = LXRTEncoder(args, max_seq_length=MAX_VQA_LENGTH)
hid_dim = self.lxrt_encoder.dim
#Build Decoder with Attention
self.decoder = DecoderWithAttention(attention_dim=hid_dim,
embed_dim=hid_dim,
decoder_dim=hid_dim,
vocab_size=vocab_size,
features_dim=hid_dim,
dropout=0.5)
# VQA Answer heads
self.logit_fc = nn.Sequential(nn.Linear(hid_dim, hid_dim * 2), GeLU(),
BertLayerNorm(hid_dim * 2, eps=1e-12),
nn.Linear(hid_dim * 2, num_answers))
self.logit_fc.apply(self.lxrt_encoder.model.init_bert_weights)
self.lstm = nn.LSTM(input_size=hid_dim,
hidden_size=hid_dim,
num_layers=1,
batch_first=True)
self.linear = nn.Linear(hid_dim,
vocab_size) #vocab size of bert is 30000
def saveToBento(checkpoint_path, word_map_file):
# Define model
# IMP: make sure the hyperparameters for the model is the same as
# that in training.
emb_dim = 512
attention_dim = 512
decoder_dim = 512
dropout = 0.5
with open(word_map_file, 'r') as j:
word_map = json.load(j)
vocab_size = len(word_map)
decoder = DecoderWithAttention(
attention_dim=attention_dim,
embed_dim=emb_dim,
decoder_dim=decoder_dim,
vocab_size=vocab_size,
dropout=dropout,
)
encoder = Encoder()
# Initialize model
checkpoint = torch.load(checkpoint_path)
decoder_state = checkpoint['decoder_state']
encoder_state = checkpoint['encoder_state']
decoder.load_state_dict(decoder_state)
encoder.load_state_dict(encoder_state)
# Add model to BentoML
bento_svc = ImageCaptioner()
bento_svc.pack('encoder', encoder)
bento_svc.pack('decoder', decoder)
# save bento service
saved_path = bento_svc.save()
# save word_map_file to bento container
dest_file = saved_path + '/ImageCaptioner/word_map.json'
shutil.copy(word_map_file, dest_file)
def main(imgurl):
# Load word map (word2ix)
with open('input_files/WORDMAP.json', 'r') as j:
word_map = json.load(j)
rev_word_map = {v: k for k, v in word_map.items()} # ix2word
# Load model
decoder = DecoderWithAttention(attention_dim=attention_dim,
embed_dim=emb_dim,
decoder_dim=decoder_dim,
vocab_size=len(word_map),
dropout=dropout)
decoder_optimizer = torch.optim.Adam(params=filter(
lambda p: p.requires_grad, decoder.parameters()),
lr=decoder_lr)
encoder = Encoder()
encoder.fine_tune(fine_tune_encoder)
encoder_optimizer = torch.optim.Adam(
params=filter(lambda p: p.requires_grad, encoder.parameters()),
lr=encoder_lr) if fine_tune_encoder else None
decoder.load_state_dict(
torch.load('output_files/BEST_checkpoint_decoder.pth.tar'))
encoder.load_state_dict(
torch.load('output_files/BEST_checkpoint_encoder.pth.tar'))
decoder = decoder.to(device)
decoder.eval()
encoder = encoder.to(device)
encoder.eval()
# Encode, decode with attention and beam search
seq, alphas = caption_image_beam_search(encoder,
decoder,
imgurl,
word_map,
beam_size=5)
alphas = torch.FloatTensor(alphas)
# Visualize caption and attention of best sequence
# visualize_att(img, seq, alphas, rev_word_map, args.smooth)
words = [rev_word_map[ind] for ind in seq]
caption = ' '.join(words[1:-1])
visualize_att(imgurl, seq, alphas, rev_word_map)
def main(args):
"""
Training and validation.
"""
with open(args.word_map_file, 'rb') as f:
word_map = pickle.load(f)
#make choice wich ecoder to use
encoder = Encoder(input_feature_dim=args.input_feature_dim,
encoder_hidden_dim=args.encoder_hidden_dim,
encoder_layer=args.encoder_layer,
rnn_unit=args.rnn_unit,
use_gpu=args.CUDA,
dropout_rate=args.dropout_rate
)
#encoder = EncoderT(input_feature_dim=args.input_feature_dim,
# encoder_hidden_dim=args.encoder_hidden_dim,
# encoder_layer=args.encoder_layer,
# rnn_unit=args.rnn_unit,
# use_gpu=args.CUDA,
# dropout_rate=args.dropout,
# nhead=args.nhead
# )
decoder = DecoderWithAttention(attention_dim=args.attention_dim,
embed_dim=args.emb_dim,
decoder_dim=args.decoder_dim,
vocab_size=len(word_map),
dropout=args.dropout)
if args.resume:
encoder.load_state_dict(torch.load(args.encoder_path))
decoder.load_state_dict(torch.load(args.decoder_path))
encoder_parameter = [p for p in encoder.parameters() if p.requires_grad] # selecting every parameter.
decoder_parameter = [p for p in decoder.parameters() if p.requires_grad]
encoder_optimizer = torch.optim.Adam(encoder_parameter,lr=args.decoder_lr) #Adam selected
decoder_optimizer = torch.optim.Adam(decoder_parameter,lr=args.decoder_lr)
if args.CUDA:
decoder = decoder.cuda()
encoder = encoder.cuda()
if args.CUDA:
criterion = nn.CrossEntropyLoss().cuda()
else:
criterion = nn.CrossEntropyLoss() #gewoon naar cpu dan
train_loader = torch.utils.data.DataLoader(
CaptionDataset(args.data_path, split='TRAIN'),
batch_size=args.batch_size, shuffle=True, num_workers=args.workers, collate_fn=pad_collate_train,pin_memory=True)
val_loader = torch.utils.data.DataLoader(
CaptionDataset(args.data_path,split='VAL'),
batch_size=args.batch_size, shuffle=False, num_workers=args.workers,collate_fn=pad_collate_train, pin_memory=True)
# Epochs
best_bleu4 = 0
for epoch in range(args.start_epoch, args.epochs):
losst = train(train_loader=train_loader, ## deze los is de trainit_weight loss!
encoder = encoder,
decoder=decoder,
criterion=criterion,
encoder_optimizer=encoder_optimizer,
decoder_optimizer=decoder_optimizer,
epoch=epoch,args=args)
# One epoch's validation
if epoch%1==0:
lossv = validate(val_loader=val_loader,
encoder=encoder,
decoder=decoder,
criterion=criterion,
best_bleu=best_bleu4,
args=args)
info = 'LOSST - {losst:.4f}, LOSSv - {lossv:.4f}\n'.format(
losst=losst,
lossv=lossv)
with open(dev, "a") as f: ## de los moet ook voor de validation
f.write(info)
f.write("\n")
#Selecteren op basis van Bleu gaat als volgt:
#print('BLEU4: ' + bleu4)
#print('best_bleu4 '+ best_bleu4)
#if bleu4>best_bleu4:
if epoch %3 ==0:
save_checkpoint(epoch, encoder, decoder, encoder_optimizer,
decoder_optimizer, lossv)
attention_dim = 512
emb_dim = 512
decoder_dim = 512
dropout = 0.5
data_name = 'coco_5_cap_per_img_5_min_word_freq'
# Read word map
word_map_file = os.path.join('/scratch/scratch2/adsue/caption_dataset',
'WORDMAP_' + data_name + '.json')
with open(word_map_file, 'r') as j:
word_map = json.load(j)
decoder = DecoderWithAttention(attention_dim=attention_dim,
embed_dim=emb_dim,
decoder_dim=decoder_dim,
vocab_size=len(word_map),
dropout=dropout)
decoder.load_state_dict(
torch.load('/scratch/scratch2/adsue/pretrained/decoder_dict.pkl'))
decoder = decoder.to(device)
decoder.eval()
encoder = Encoder()
encoder.load_state_dict(
torch.load('/scratch/scratch2/adsue/pretrained/encoder_dict.pkl'))
encoder = encoder.to(device)
encoder.eval()
##########################################################################################################################
def main():
"""
Training and validation.
"""
global best_bleu4, epochs_since_improvement, checkpoint, start_epoch, fine_tune_encoder, data_name, word_map
# Read word map (w2i)
word_map_file = os.path.join(data_folder, 'WORDMAP_' + data_name + '.json')
with open(word_map_file, 'r') as j:
word_map = json.load(j)
# Initialize / load checkpoint
if checkpoint is None:
decoder = DecoderWithAttention(attention_dim=attention_dim,
embed_dim=emb_dim,
decoder_dim=decoder_dim,
vocab_size=len(word_map),
dropout=dropout)
decoder_optimizer = torch.optim.Adam(params=filter(
lambda p: p.requires_grad, decoder.parameters()),
lr=decoder_lr)
encoder = Encoder()
encoder.fine_tune(fine_tune_encoder)
encoder_optimizer = torch.optim.Adam(
params=filter(lambda p: p.requires_grad, encoder.parameters()),
lr=encoder_lr) if fine_tune_encoder else None
else:
checkpoint = torch.load(checkpoint)
start_epoch = checkpoint['epoch'] + 1
epochs_since_improvement = checkpoint['epochs_since_improvement']
best_bleu4 = checkpoint['bleu-4']
decoder = checkpoint['decoder']
decoder_optimizer = checkpoint['decoder_optimizer']
encoder = checkpoint['encoder']
encoder_optimizer = checkpoint['encoder_optimizer']
if fine_tune_encoder is True and encoder_optimizer is None:
encoder.fine_tune(fine_tune_encoder)
encoder_optimizer = torch.optim.Adam(params=filter(
lambda p: p.requires_grad, encoder.parameters()),
lr=encoder_lr)
# Move to GPU, if available
decoder = decoder.to(device)
encoder = encoder.to(device)
# Loss function
criterion = nn.CrossEntropyLoss().to(device)
# Custom dataloaders (This page details the preprocessing or transformation we need to perform –
# pixel values must be in the range [0,1] and we must then normalize the image by the mean and standard
# deviation of the ImageNet images' RGB channels.)
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
train_loader = torch.utils.data.DataLoader(CaptionDataset(
data_folder,
data_name,
'TRAIN',
transform=transforms.Compose([normalize])),
batch_size=batch_size,
shuffle=True,
num_workers=workers,
pin_memory=True)
val_loader = torch.utils.data.DataLoader(CaptionDataset(
data_folder,
data_name,
'VAL',
transform=transforms.Compose([normalize])),
batch_size=batch_size,
shuffle=True,
num_workers=workers,
pin_memory=True)
initial_time = time.time()
print("Initial time", initial_time)
# Epochs
for epoch in range(start_epoch, epochs):
print("Starting epoch ", epoch)
# Decay learning rate if there is no improvement for 8 consecutive epochs, and terminate training after 20
if epochs_since_improvement == 20:
break
if epochs_since_improvement > 0 and epochs_since_improvement % 8 == 0:
adjust_learning_rate(decoder_optimizer, 0.8)
if fine_tune_encoder:
adjust_learning_rate(encoder_optimizer, 0.8)
# One epoch's training
train(train_loader=train_loader,
encoder=encoder,
decoder=decoder,
criterion=criterion,
encoder_optimizer=encoder_optimizer,
decoder_optimizer=decoder_optimizer,
epoch=epoch,
initial_time=initial_time)
# One epoch's validation
recent_bleu4 = validate(val_loader=val_loader,
encoder=encoder,
decoder=decoder,
criterion=criterion)
# Check if there was an improvement
is_best = recent_bleu4 > best_bleu4
best_bleu4 = max(recent_bleu4, best_bleu4)
if not is_best:
epochs_since_improvement += 1
print("\nEpochs since last improvement: %d\n" %
(epochs_since_improvement, ))
else:
epochs_since_improvement = 0
# Save checkpoint
save_checkpoint(data_name, epoch, epochs_since_improvement, encoder,
decoder, encoder_optimizer, decoder_optimizer,
recent_bleu4, is_best)
def main(args):
"""
Training and validation.
"""
global best_bleu4, epochs_since_improvement, checkpoint, start_epoch, fine_tune_encoder, data_name, word_map
data_folder = '{:s}_folder'.format(args.dataset) # folder with data files saved by create_input_files.py
data_name = '{:s}_5_cap_per_img_5_min_word_freq'.format(args.dataset) # base name shared by data files
# Read word map
word_map_file = os.path.join(data_folder, 'WORDMAP_' + data_name + '.json')
with open(word_map_file, 'r') as j:
word_map = json.load(j)
# Initialize / load checkpoint
if checkpoint is None:
decoder = DecoderWithAttention(attention_dim=attention_dim,
embed_dim=emb_dim,
decoder_dim=decoder_dim,
vocab_size=len(word_map),
dropout=dropout,
adaptive_att=args.adaptive)
decoder_optimizer = torch.optim.Adam(params=filter(lambda p: p.requires_grad, decoder.parameters()),
lr=decoder_lr)
if args.adaptive:
encoder = Adaptive_Encoder(encoded_image_size=14,
embed_dim=emb_dim,
decoder_dim=decoder_dim)
else:
encoder = Encoder()
encoder.fine_tune(fine_tune_encoder)
encoder_optimizer = torch.optim.Adam(params=filter(lambda p: p.requires_grad, encoder.parameters()),
lr=encoder_lr) if fine_tune_encoder else None
else:
checkpoint = torch.load(checkpoint)
start_epoch = checkpoint['epoch'] + 1
epochs_since_improvement = checkpoint['epochs_since_improvement']
best_bleu4 = checkpoint['bleu-4']
decoder = checkpoint['decoder']
decoder_optimizer = checkpoint['decoder_optimizer']
encoder = checkpoint['encoder']
encoder_optimizer = checkpoint['encoder_optimizer']
if fine_tune_encoder is True and encoder_optimizer is None:
encoder.fine_tune(fine_tune_encoder)
encoder_optimizer = torch.optim.Adam(params=filter(lambda p: p.requires_grad, encoder.parameters()),
lr=encoder_lr)
# Move to GPU, if available
decoder = decoder.to(device)
encoder = encoder.to(device)
# Loss function
criterion = nn.CrossEntropyLoss().to(device)
# Custom dataloaders
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
train_loader = torch.utils.data.DataLoader(
CaptionDataset(data_folder, data_name, 'TRAIN', transform=transforms.Compose([normalize])),
batch_size=batch_size, shuffle=True, pin_memory=False)
val_loader = torch.utils.data.DataLoader(
CaptionDataset(data_folder, data_name, 'VAL', transform=transforms.Compose([normalize])),
batch_size=batch_size, shuffle=True, pin_memory=False)
# Epochs
for epoch in range(start_epoch, epochs):
# Decay learning rate if there is no improvement for 8 consecutive epochs, and terminate training after 20
if epochs_since_improvement == 20:
break
if epochs_since_improvement > 0 and epochs_since_improvement % 8 == 0:
adjust_learning_rate(decoder_optimizer, 0.8)
if fine_tune_encoder:
adjust_learning_rate(encoder_optimizer, 0.8)
# One epoch's training
train(train_loader=train_loader,
encoder=encoder,
decoder=decoder,
criterion=criterion,
encoder_optimizer=encoder_optimizer,
decoder_optimizer=decoder_optimizer,
epoch=epoch)
# One epoch's validation
recent_bleu4 = validate(val_loader=val_loader,
encoder=encoder,
decoder=decoder,
criterion=criterion)
# Check if there was an improvement
is_best = recent_bleu4 > best_bleu4
best_bleu4 = max(recent_bleu4, best_bleu4)
if not is_best:
epochs_since_improvement += 1
print("\nEpochs since last improvement: %d\n" % (epochs_since_improvement,))
else:
epochs_since_improvement = 0
# Save checkpoint
save_checkpoint(data_name, epoch, epochs_since_improvement, encoder, decoder, encoder_optimizer,
decoder_optimizer, recent_bleu4, is_best)
def main():
"""
Training and validation.
"""
parser = argparse.ArgumentParser()
parser.add_argument(
"--data_folder",
default='data/',
type=str,
help="folder with data files saved by create_input_files.py")
parser.add_argument("--data_name",
default='coco_5_cap_per_img_5_min_word_freq',
type=str,
help="base name shared by data files")
parser.add_argument("--output_dir",
default='saved_models/',
type=str,
help="path to save checkpoints")
parser.add_argument("--checkpoint",
default=None,
type=str,
help="path to checkpoint")
parser.add_argument("--emb_dim",
default=512,
type=int,
help="dimension of word embeddings")
parser.add_argument("--attention_dim",
default=512,
type=int,
help="dimension of attention linear layers")
parser.add_argument("--decoder_dim",
default=512,
type=int,
help="dimension of decoder RNN")
parser.add_argument("--dropout",
default=0.5,
type=float,
help="dimension of word embeddings")
parser.add_argument("--start_epoch", default=0, type=int)
parser.add_argument(
"--epochs",
default=120,
type=int,
help=
"number of epochs to train for (if early stopping is not triggered)")
parser.add_argument("--batch_size",
default=128,
type=int,
help="batch size for training and testing")
parser.add_argument("--workers",
default=8,
type=int,
help="num of workers for data-loading")
parser.add_argument("--encoder_lr", default=1e-4, type=float)
parser.add_argument("--decoder_lr", default=5e-4, type=float)
parser.add_argument("--grad_clip",
default=5,
type=float,
help="clip gradients at an absolute value of")
parser.add_argument(
"--alpha_c",
default=1,
type=int,
help=
"regularization parameter for 'doubly stochastic attention', as in the paper"
)
parser.add_argument(
"--print_freq",
default=100,
type=int,
help="print training/validation stats every __ batches")
parser.add_argument("--fine_tune_encoder",
action='store_true',
help="Whether to finetune the encoder")
args = parser.parse_args()
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
args.device = device
best_bleu4 = 0
epochs_since_improvement = 0
# Read word map
word_map_file = os.path.join(args.data_folder,
'WORDMAP_' + args.data_name + '.json')
with open(word_map_file, 'r') as j:
word_map = json.load(j)
# Initialize / load checkpoint
if args.checkpoint is None:
decoder = DecoderWithAttention(attention_dim=args.attention_dim,
embed_dim=args.emb_dim,
decoder_dim=args.decoder_dim,
vocab_size=len(word_map),
dropout=args.dropout)
decoder_optimizer = torch.optim.Adam(params=filter(
lambda p: p.requires_grad, decoder.parameters()),
lr=args.decoder_lr)
encoder = Encoder()
encoder.fine_tune(args.fine_tune_encoder)
encoder_optimizer = torch.optim.Adam(
params=filter(lambda p: p.requires_grad, encoder.parameters()),
lr=args.encoder_lr) if args.fine_tune_encoder else None
else:
checkpoint = torch.load(args.checkpoint)
args.start_epoch = checkpoint['epoch'] + 1
epochs_since_improvement = checkpoint['epochs_since_improvement']
best_bleu4 = checkpoint['bleu-4']
decoder = checkpoint['decoder']
decoder_optimizer = checkpoint['decoder_optimizer']
encoder = checkpoint['encoder']
encoder_optimizer = checkpoint['encoder_optimizer']
if args.fine_tune_encoder is True and encoder_optimizer is None:
encoder.fine_tune(args.fine_tune_encoder)
encoder_optimizer = torch.optim.Adam(params=filter(
lambda p: p.requires_grad, encoder.parameters()),
lr=args.encoder_lr)
# Move to GPU, if available
decoder = decoder.to(args.device)
encoder = encoder.to(args.device)
# Loss function
criterion = nn.CrossEntropyLoss(ignore_index=0).to(args.device)
# Custom dataloaders
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
train_loader = torch.utils.data.DataLoader(CaptionDataset(
args.data_folder,
args.data_name,
'TRAIN',
transform=transforms.Compose([normalize])),
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
print(f'train dataset length {len(train_loader)}')
val_loader = torch.utils.data.DataLoader(CaptionDataset(
args.data_folder,
args.data_name,
'VAL',
transform=transforms.Compose([normalize])),
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
print(f'val dataset length {len(val_loader)}')
# Epochs
for epoch in range(args.start_epoch, args.epochs):
# Decay learning rate if there is no improvement for 8 consecutive epochs, and terminate training after 20
if epochs_since_improvement == 20:
break
if epochs_since_improvement > 0 and epochs_since_improvement % 8 == 0:
adjust_learning_rate(decoder_optimizer, 0.8)
if args.fine_tune_encoder:
adjust_learning_rate(encoder_optimizer, 0.8)
# One epoch's training
train(train_loader=train_loader,
encoder=encoder,
decoder=decoder,
criterion=criterion,
encoder_optimizer=encoder_optimizer,
decoder_optimizer=decoder_optimizer,
epoch=epoch,
args=args)
# One epoch's validation
recent_bleu4 = validate(val_loader=val_loader,
encoder=encoder,
decoder=decoder,
criterion=criterion,
word_map=word_map,
args=args)
# Check if there was an improvement
is_best = recent_bleu4 > best_bleu4
best_bleu4 = max(recent_bleu4, best_bleu4)
if not is_best:
epochs_since_improvement += 1
print("\nEpochs since last improvement: %d\n" %
(epochs_since_improvement, ))
else:
epochs_since_improvement = 0
# Save checkpoint
save_checkpoint(args.data_name, args.output_dir, epoch,
epochs_since_improvement, encoder, decoder,
encoder_optimizer, decoder_optimizer, recent_bleu4,
is_best)
def main(checkpoint, tienet):
"""
Training and validation.
"""
global best_bleu4, epochs_since_improvement, start_epoch, fine_tune_encoder, data_name, word_map
if checkpoint:
dest_dir = checkpoint
checkpoint = os.path.join(
dest_dir,
'checkpoint_mimiccxr_1_cap_per_img_5_min_word_freq.pth.tar'
) # path to checkpoint, None if none
else:
dest_dir = os.path.join(
'/data/medg/misc/liuguanx/TieNet/models',
datetime.datetime.now().strftime('%Y-%m-%d-%H%M%S-%f'))
os.makedirs(dest_dir)
checkpoint = None
# Read word map
word_map_file = os.path.join(data_folder, 'WORDMAP_' + data_name + '.json')
with open(word_map_file, 'r') as j:
word_map = json.load(j)
# Initialize / load checkpoint
if checkpoint is None:
decoder = DecoderWithAttention(attention_dim=attention_dim,
embed_dim=emb_dim,
decoder_dim=decoder_dim,
vocab_size=len(word_map),
dropout=dropout)
decoder_optimizer = torch.optim.Adam(params=filter(
lambda p: p.requires_grad, decoder.parameters()),
lr=decoder_lr)
encoder = Encoder()
encoder.fine_tune(fine_tune_encoder)
encoder_optimizer = torch.optim.Adam(
params=filter(lambda p: p.requires_grad, encoder.parameters()),
lr=encoder_lr) if fine_tune_encoder else None
if (tienet):
jointlearner = JointLearning(num_global_att=num_global_att,
s=s,
decoder_dim=decoder_dim,
label_size=label_size)
jointlearner_optimizer = torch.optim.Adam(params=filter(
lambda p: p.requires_grad, jointlearner.parameters()),
lr=jointlearning_lr)
else:
jointlearner = None
jointlearner_optimizer = None
else:
checkpoint = torch.load(checkpoint)
print('checkpoint loaded')
start_epoch = checkpoint['epoch'] + 1
epochs_since_improvement = checkpoint['epochs_since_improvement']
best_bleu4 = checkpoint['best_bleu']
decoder = checkpoint['decoder']
decoder_optimizer = checkpoint['decoder_optimizer']
encoder = checkpoint['encoder']
encoder_optimizer = checkpoint['encoder_optimizer']
jointlearner = checkpoint['jointlearner']
jointlearner_optimizer = checkpoint['jointlearner_optimizer']
if fine_tune_encoder is True and encoder_optimizer is None:
encoder.fine_tune(fine_tune_encoder)
encoder_optimizer = torch.optim.Adam(params=filter(
lambda p: p.requires_grad, encoder.parameters()),
lr=encoder_lr)
# Move to GPU, if available
if torch.cuda.device_count() > 1:
print('Using', torch.cuda.device_count(), 'GPUs')
# decoder = nn.DataParallel(decoder)
encoder = nn.DataParallel(encoder, device_ids=[1])
if tienet:
jointlearner = nn.DataParallel(jointlearner, device_ids=[1])
decoder = decoder.to(device)
encoder = encoder.to(device)
if tienet:
jointlearner = jointlearner.to(device)
# Loss function
criterion_R = nn.CrossEntropyLoss().to(device)
criterion_C = nn.BCEWithLogitsLoss().to(device)
# Custom dataloaders
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
train_loader = torch.utils.data.DataLoader(CaptionDataset(
data_folder,
data_name,
'TRAIN',
transform=transforms.Compose([normalize])),
batch_size=batch_size,
shuffle=True,
num_workers=workers,
pin_memory=True)
val_loader = torch.utils.data.DataLoader(CaptionDataset(
data_folder,
data_name,
'VAL',
transform=transforms.Compose([normalize])),
batch_size=batch_size,
shuffle=True,
num_workers=workers,
pin_memory=True)
# Epochs
for epoch in range(start_epoch, epochs):
# Decay learning rate if there is no improvement for 8 consecutive epochs, and terminate training after 20
if epochs_since_improvement == 20:
break
if epochs_since_improvement > 0 and epochs_since_improvement % 8 == 0:
adjust_learning_rate(decoder_optimizer, 0.8)
if tienet:
adjust_learning_rate(jointlearner_optimizer, 0.8)
if fine_tune_encoder:
adjust_learning_rate(encoder_optimizer, 0.8)
# One epoch's training
train(train_loader=train_loader,
encoder=encoder,
decoder=decoder,
jointlearner=jointlearner,
criterion_R=criterion_R,
criterion_C=criterion_C,
encoder_optimizer=encoder_optimizer,
decoder_optimizer=decoder_optimizer,
jointlearner_optimizer=jointlearner_optimizer,
epoch=epoch,
dest_dir=dest_dir,
tienet=tienet)
# One epoch's validation
recent_bleu4 = validate(val_loader=val_loader,
encoder=encoder,
decoder=decoder,
jointlearner=jointlearner,
criterion_R=criterion_R,
criterion_C=criterion_C,
tienet=tienet)
# Check if there was an improvement
is_best = recent_bleu4 > best_bleu4
best_bleu4 = max(recent_bleu4, best_bleu4)
if not is_best:
epochs_since_improvement += 1
print("\nEpochs since last improvement: %d\n" %
(epochs_since_improvement, ))
else:
epochs_since_improvement = 0
# Save checkpoint
save_checkpoint(data_name, epoch, epochs_since_improvement, encoder,
decoder, jointlearner, encoder_optimizer,
decoder_optimizer, jointlearner_optimizer,
recent_bleu4, best_bleu4, is_best, dest_dir)
def train(args):
cfg_from_file(args.cfg)
cfg.WORKERS = args.num_workers
pprint.pprint(cfg)
# set the seed manually
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
# define outputer
outputer_train = Outputer(args.output_dir, cfg.IMAGETEXT.PRINT_EVERY,
cfg.IMAGETEXT.SAVE_EVERY)
outputer_val = Outputer(args.output_dir, cfg.IMAGETEXT.PRINT_EVERY,
cfg.IMAGETEXT.SAVE_EVERY)
# define the dataset
split_dir, bshuffle = 'train', True
# Get data loader
imsize = cfg.TREE.BASE_SIZE * (2**(cfg.TREE.BRANCH_NUM - 1))
train_transform = transforms.Compose([
transforms.Scale(int(imsize * 76 / 64)),
transforms.RandomCrop(imsize),
])
val_transform = transforms.Compose([
transforms.Scale(int(imsize * 76 / 64)),
transforms.CenterCrop(imsize),
])
if args.dataset == 'bird':
train_dataset = ImageTextDataset(args.data_dir,
split_dir,
transform=train_transform,
sample_type='train')
val_dataset = ImageTextDataset(args.data_dir,
'val',
transform=val_transform,
sample_type='val')
elif args.dataset == 'coco':
train_dataset = CaptionDataset(args.data_dir,
split_dir,
transform=train_transform,
sample_type='train',
coco_data_json=args.coco_data_json)
val_dataset = CaptionDataset(args.data_dir,
'val',
transform=val_transform,
sample_type='val',
coco_data_json=args.coco_data_json)
else:
raise NotImplementedError
train_dataloader = torch.utils.data.DataLoader(
train_dataset,
batch_size=cfg.IMAGETEXT.BATCH_SIZE,
shuffle=bshuffle,
num_workers=int(cfg.WORKERS))
val_dataloader = torch.utils.data.DataLoader(
val_dataset,
batch_size=cfg.IMAGETEXT.BATCH_SIZE,
shuffle=False,
num_workers=1)
# define the model and optimizer
if args.raw_checkpoint != '':
encoder, decoder = load_raw_checkpoint(args.raw_checkpoint)
else:
encoder = Encoder()
decoder = DecoderWithAttention(
attention_dim=cfg.IMAGETEXT.ATTENTION_DIM,
embed_dim=cfg.IMAGETEXT.EMBED_DIM,
decoder_dim=cfg.IMAGETEXT.DECODER_DIM,
vocab_size=train_dataset.n_words)
# load checkpoint
if cfg.IMAGETEXT.CHECKPOINT != '':
outputer_val.log("load model from: {}".format(
cfg.IMAGETEXT.CHECKPOINT))
encoder, decoder = load_checkpoint(encoder, decoder,
cfg.IMAGETEXT.CHECKPOINT)
encoder.fine_tune(False)
# to cuda
encoder = encoder.cuda()
decoder = decoder.cuda()
loss_func = torch.nn.CrossEntropyLoss()
if args.eval: # eval only
outputer_val.log("only eval the model...")
assert cfg.IMAGETEXT.CHECKPOINT != ''
val_rtn_dict, outputer_val = validate_one_epoch(
0, val_dataloader, encoder, decoder, loss_func, outputer_val)
outputer_val.log("\n[valid]: {}\n".format(dict2str(val_rtn_dict)))
return
# define optimizer
optimizer_encoder = torch.optim.Adam(encoder.parameters(),
lr=cfg.IMAGETEXT.ENCODER_LR)
optimizer_decoder = torch.optim.Adam(decoder.parameters(),
lr=cfg.IMAGETEXT.DECODER_LR)
encoder_lr_scheduler = torch.optim.lr_scheduler.StepLR(
optimizer_encoder, step_size=10, gamma=cfg.IMAGETEXT.LR_GAMMA)
decoder_lr_scheduler = torch.optim.lr_scheduler.StepLR(
optimizer_decoder, step_size=10, gamma=cfg.IMAGETEXT.LR_GAMMA)
print("train the model...")
for epoch_idx in range(cfg.IMAGETEXT.EPOCH):
# val_rtn_dict, outputer_val = validate_one_epoch(epoch_idx, val_dataloader, encoder,
# decoder, loss_func, outputer_val)
# outputer_val.log("\n[valid] epoch: {}, {}".format(epoch_idx, dict2str(val_rtn_dict)))
train_rtn_dict, outputer_train = train_one_epoch(
epoch_idx, train_dataloader, encoder, decoder, optimizer_encoder,
optimizer_decoder, loss_func, outputer_train)
# adjust lr scheduler
encoder_lr_scheduler.step()
decoder_lr_scheduler.step()
outputer_train.log("\n[train] epoch: {}, {}\n".format(
epoch_idx, dict2str(train_rtn_dict)))
val_rtn_dict, outputer_val = validate_one_epoch(
epoch_idx, val_dataloader, encoder, decoder, loss_func,
outputer_val)
outputer_val.log("\n[valid] epoch: {}, {}\n".format(
epoch_idx, dict2str(val_rtn_dict)))
outputer_val.save_step({
"encoder": encoder.state_dict(),
"decoder": decoder.state_dict()
})
outputer_val.save({
"encoder": encoder.state_dict(),
"decoder": decoder.state_dict()
})
def main():
"""
Training and validation.
"""
global best_bleu4, epochs_since_improvement, checkpoint, start_epoch, fine_tune_encoder, data_name, word_map, rev_word_map
# Read word map
word_map_file = os.path.join(data_folder, 'WORDMAP_' + data_name + '.json')
with open(word_map_file, 'r') as j:
word_map = json.load(j)
rev_word_map = {v: k for k, v in word_map.items()}
# Initialize / load checkpoint
if checkpoint is None:
decoder = DecoderWithAttention(attention_dim=attention_dim,
embed_dim=emb_dim,
decoder_dim=decoder_dim,
vocab_size=len(word_map),
dropout=dropout)
pretrained_embs, pretrained_embs_dim = load_embeddings(
'/home/Iwamura/datasets/datasets/GloVe/glove.6B.300d.txt',
word_map)
assert pretrained_embs_dim == decoder.embed_dim
decoder.load_pretrained_embeddings(pretrained_embs)
decoder.fine_tune_embeddings(True)
decoder_optimizer = torch.optim.Adam(params=filter(
lambda p: p.requires_grad, decoder.parameters()),
lr=decoder_lr)
encoder = Encoder()
encoder_opt = Encoder()
encoder.fine_tune(fine_tune_encoder)
encoder_opt.fine_tune(fine_tune_encoder_opt)
encoder_optimizer = torch.optim.Adam(
params=filter(lambda p: p.requires_grad, encoder.parameters()),
lr=encoder_lr) if fine_tune_encoder else None
encoder_optimizer_opt = torch.optim.Adam(
params=filter(lambda p: p.requires_grad, encoder_opt.parameters()),
lr=encoder_opt_lr) if fine_tune_encoder_opt else None
else:
checkpoint = torch.load(checkpoint)
start_epoch = checkpoint['epoch'] + 1
epochs_since_improvement = checkpoint['epochs_since_improvement']
best_bleu4 = checkpoint['bleu-4']
decoder = checkpoint['decoder']
decoder_optimizer = checkpoint['decoder_optimizer']
encoder_opt = checkpoint['encoder_opt']
encoder_optimizer_opt = checkpoint['encoder_optimizer_opt']
# if fine_tune_encoder is True and encoder_optimizer is None and encoder_optimizer_opt is None
if fine_tune_encoder_opt is True and encoder_optimizer_opt is None:
encoder_opt.fine_tune(fine_tune_encoder_opt)
encoder_optimizer_opt = torch.optim.Adam(params=filter(
lambda p: p.requires_grad, encoder_opt.parameters()),
lr=encoder_opt_lr)
# Move to GPU, if available
decoder = decoder.to(device)
encoder_opt = encoder_opt.to(device)
# Loss function
criterion = nn.CrossEntropyLoss().to(device)
# Custom dataloaders
normalize_opt = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
train_loader = torch.utils.data.DataLoader(CaptionDataset(
data_folder,
data_name,
'TRAIN',
transform=transforms.Compose([normalize_opt])),
batch_size=batch_size,
shuffle=True,
num_workers=workers,
pin_memory=True)
val_loader = torch.utils.data.DataLoader(CaptionDataset(
data_folder,
data_name,
'VAL',
transform=transforms.Compose([normalize_opt])),
batch_size=batch_size,
shuffle=True,
num_workers=workers,
pin_memory=True)
# Epochs
for epoch in range(start_epoch, epochs):
# Decay learning rate if there is no improvement for 8 consecutive epochs, and terminate training after 20
if epochs_since_improvement == 10:
break
if epoch > 0 and epoch % 4 == 0:
adjust_learning_rate(decoder_optimizer, 0.8)
if fine_tune_encoder_opt:
adjust_learning_rate(encoder_optimizer_opt, 0.8)
# One epoch's training
train(train_loader=train_loader,
encoder_opt=encoder_opt,
decoder=decoder,
criterion=criterion,
encoder_optimizer_opt=encoder_optimizer_opt,
decoder_optimizer=decoder_optimizer,
epoch=epoch)
# One epoch's validation
recent_bleu4 = validate(val_loader=val_loader,
encoder_opt=encoder_opt,
decoder=decoder,
criterion=criterion)
# Check if there was an improvement
is_best = recent_bleu4 > best_bleu4
best_bleu4 = max(recent_bleu4, best_bleu4)
if not is_best:
epochs_since_improvement += 1
print("\nEpochs since last improvement: %d\n" %
(epochs_since_improvement, ))
else:
epochs_since_improvement = 0
# Save checkpoint
save_checkpoint(data_name, epoch, epochs_since_improvement,
encoder_opt, decoder, encoder_optimizer_opt,
decoder_optimizer, recent_bleu4, is_best)
def main():
"""
Training and validation.
"""
global best_bleu4, epochs_since_improvement, checkpoint, start_epoch, fine_tune_encoder, data_name, word_map
# Read word map
word_map_file = os.path.join(data_folder, "WORDMAP_" + data_name + ".json")
with open(word_map_file, "r") as j:
word_map = json.load(j)
# Initialize / load checkpoint
if checkpoint is None:
decoder = DecoderWithAttention(
attention_dim=attention_dim,
embed_dim=emb_dim,
decoder_dim=decoder_dim,
vocab_size=len(word_map),
dropout=dropout,
)
decoder_optimizer = torch.optim.Adam(
params=filter(lambda p: p.requires_grad, decoder.parameters()),
lr=decoder_lr,
)
encoder = Encoder()
encoder.fine_tune(fine_tune_encoder)
encoder_optimizer = (torch.optim.Adam(
params=filter(lambda p: p.requires_grad, encoder.parameters()),
lr=encoder_lr,
) if fine_tune_encoder else None)
else:
checkpoint = torch.load(checkpoint)
start_epoch = checkpoint["epoch"] + 1
epochs_since_improvement = checkpoint["epochs_since_improvement"]
best_bleu4 = checkpoint["bleu-4"]
decoder = checkpoint["decoder"]
decoder_optimizer = checkpoint["decoder_optimizer"]
encoder = checkpoint["encoder"]
encoder_optimizer = checkpoint["encoder_optimizer"]
if fine_tune_encoder is True and encoder_optimizer is None:
encoder.fine_tune(fine_tune_encoder)
encoder_optimizer = torch.optim.Adam(
params=filter(lambda p: p.requires_grad, encoder.parameters()),
lr=encoder_lr,
)
# Move to GPU, if available
decoder = decoder.to(device)
encoder = encoder.to(device)
# Loss function
criterion = nn.CrossEntropyLoss().to(device)
# Custom dataloaders
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
train_loader = torch.utils.data.DataLoader(
CaptionDataset(data_folder,
data_name,
"TRAIN",
transform=transforms.Compose([normalize])),
batch_size=batch_size,
shuffle=True,
num_workers=workers,
pin_memory=True,
)
val_loader = torch.utils.data.DataLoader(
CaptionDataset(data_folder,
data_name,
"VAL",
transform=transforms.Compose([normalize])),
batch_size=batch_size,
shuffle=True,
num_workers=workers,
pin_memory=True,
)
# Epochs
for epoch in range(start_epoch, epochs):
# Decay learning rate if there is no improvement for 8 consecutive epochs, and terminate training after 20
if epochs_since_improvement == 20:
break
if epochs_since_improvement > 0 and epochs_since_improvement % 8 == 0:
adjust_learning_rate(decoder_optimizer, 0.8)
if fine_tune_encoder:
adjust_learning_rate(encoder_optimizer, 0.8)
# One epoch's training
train(
train_loader=train_loader,
encoder=encoder,
decoder=decoder,
criterion=criterion,
encoder_optimizer=encoder_optimizer,
decoder_optimizer=decoder_optimizer,
epoch=epoch,
)
# One epoch's validation
recent_bleu4 = validate(val_loader=val_loader,
encoder=encoder,
decoder=decoder,
criterion=criterion)
# Check if there was an improvement
is_best = recent_bleu4 > best_bleu4
best_bleu4 = max(recent_bleu4, best_bleu4)
if not is_best:
epochs_since_improvement += 1
print("\nEpochs since last improvement: %d\n" %
(epochs_since_improvement, ))
else:
epochs_since_improvement = 0
# Save checkpoint
save_checkpoint(
data_name,
epoch,
epochs_since_improvement,
encoder,
decoder,
encoder_optimizer,
decoder_optimizer,
recent_bleu4,
is_best,
)
def main():
"""
Training and validation.
"""
global best_bleu4, epochs_since_improvement, checkpoint, start_epoch, fine_tune_encoder, data_name, word_map
# Read word map
#word_map_file = os.path.join(data_folder, 'WORDMAP_' + data_name + '.json')
#with open(word_map_file, 'r') as j:
# word_map = json.load(j)
with open("/content/image_captioning/Image-Captioning-Codebase/vocab.pkl",
"rb") as f:
vocab = pickle.load(f)
word_map = vocab.word2idx
# Initialize / load checkpoint
if checkpoint is None:
decoder = DecoderWithAttention(attention_dim=attention_dim,
embed_dim=emb_dim,
decoder_dim=decoder_dim,
vocab_size=len(word_map),
dropout=dropout)
decoder_optimizer = torch.optim.Adam(params=filter(
lambda p: p.requires_grad, decoder.parameters()),
lr=decoder_lr)
encoder = Encoder()
encoder.fine_tune(fine_tune_encoder)
encoder_optimizer = torch.optim.Adam(
params=filter(lambda p: p.requires_grad, encoder.parameters()),
lr=encoder_lr) if fine_tune_encoder else None
else:
checkpoint = torch.load(checkpoint)
start_epoch = checkpoint['epoch'] + 1
epochs_since_improvement = checkpoint['epochs_since_improvement']
best_bleu4 = checkpoint['bleu-4']
decoder = checkpoint['decoder']
decoder_optimizer = checkpoint['decoder_optimizer']
encoder = checkpoint['encoder']
encoder_optimizer = checkpoint['encoder_optimizer']
if fine_tune_encoder is True and encoder_optimizer is None:
encoder.fine_tune(fine_tune_encoder)
encoder_optimizer = torch.optim.Adam(params=filter(
lambda p: p.requires_grad, encoder.parameters()),
lr=encoder_lr)
# Move to GPU, if available
decoder = decoder.to(device)
encoder = encoder.to(device)
# Loss function
criterion = nn.CrossEntropyLoss().to(device)
# Custom dataloaders
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
transform_train = transforms.Compose(
[ # smaller edge of image resized to 256
transforms.Resize(
(224, 224)), # get 224x224 crop from random location
transforms.RandomHorizontalFlip(
), # horizontally flip image with probability=0.5
transforms.ToTensor(), # convert the PIL Image to a tensor
transforms.Normalize(
(0.485, 0.456, 0.406), # normalize image for pre-trained model
(0.229, 0.224, 0.225))
])
"""
train_loader = torch.utils.data.DataLoader(
CaptionDataset(data_folder, data_name, 'TRAIN', transform=transforms.Compose([normalize])),
batch_size=batch_size, shuffle=True, num_workers=workers, pin_memory=True)
val_loader = torch.utils.data.DataLoader(
CaptionDataset(data_folder, data_name, 'VAL', transform=transforms.Compose([normalize])),
batch_size=batch_size, shuffle=True, num_workers=workers, pin_memory=True)
"""
train_loader = torch.utils.data.DataLoader(
Flickr8kDataset(annot_path="/content/", img_path="/content/Flicker8k_Dataset/", \
split="train", transform=transform_train), \
batch_size=batch_size, shuffle=True, num_workers=workers, pin_memory=True)
val_loader = torch.utils.data.DataLoader(
Flickr8kDataset(annot_path="/content/", img_path="/content/Flicker8k_Dataset/", \
split="dev", transform=transform_train), \
batch_size=batch_size, shuffle=True, num_workers=workers, pin_memory=True)
# Epochs
for epoch in range(start_epoch, epochs):
# Decay learning rate if there is no improvement for 8 consecutive epochs, and terminate training after 20
if epochs_since_improvement == 20:
break
if epochs_since_improvement > 0 and epochs_since_improvement % 8 == 0:
adjust_learning_rate(decoder_optimizer, 0.8)
if fine_tune_encoder:
adjust_learning_rate(encoder_optimizer, 0.8)
# One epoch's training
train(train_loader=train_loader,
encoder=encoder,
decoder=decoder,
criterion=criterion,
encoder_optimizer=encoder_optimizer,
decoder_optimizer=decoder_optimizer,
epoch=epoch)
# One epoch's validation
recent_bleu4 = validate(val_loader=val_loader,
encoder=encoder,
decoder=decoder,
criterion=criterion)
# Check if there was an improvement
is_best = recent_bleu4 > best_bleu4
best_bleu4 = max(recent_bleu4, best_bleu4)
if not is_best:
epochs_since_improvement += 1
print("\nEpochs since last improvement: %d\n" %
(epochs_since_improvement, ))
else:
epochs_since_improvement = 0
# Save checkpoint
save_checkpoint(data_name, epoch, epochs_since_improvement, encoder,
decoder, encoder_optimizer, decoder_optimizer,
recent_bleu4, is_best)
def main():
parser = argparse.ArgumentParser(description='caption model')
parser.add_argument('--save_dir',
type=str,
default='logs/tmp',
help='directory of model save')
# 数据集参数
parser.add_argument('--data_folder',
type=str,
default='./datasets/caption_data',
help='caption dataset folder')
parser.add_argument('--data_name',
type=str,
default='flickr8k_5_cap_per_img_5_min_word_freq',
help='dataset name [coco, flickr8k, flickr30k]')
parser.add_argument('--batch_size',
type=int,
default=32,
help='training batch size')
parser.add_argument('--print_freq',
type=int,
default=100,
help='print training state every n times')
parser.add_argument(
'--num_workers',
type=int,
default=0, #8,
help='number of data loader workers ')
parser.add_argument('--epochs',
type=int,
default=120,
help='total training epochs')
parser.add_argument('--grad_clip',
type=float,
default=5.,
help='number of gradient clip')
parser.add_argument('--alpha_c',
type=float,
default=1.,
help='ratio of attention matrix')
parser.add_argument('--encoder_lr',
type=float,
default=1e-4,
help='encoder learning rate')
parser.add_argument('--decoder_lr',
type=float,
default=4e-4,
help='decoder learning rate')
# 模型参数
parser.add_argument('--attention_dim',
type=float,
default=512,
help='dimension of attention')
parser.add_argument('--embed_dim',
type=float,
default=512,
help='dimension of word embedding')
parser.add_argument('--decoder_dim',
type=float,
default=512,
help='dimension of decoder')
#default=2048, help='dimension of decoder')
parser.add_argument('--dropout',
type=float,
default=0.5,
help='rate of dropout')
parser.add_argument('-frz',
'--freeze_encoder',
action='store_true',
help='whether freeze encoder parameters')
args = parser.parse_args()
mkdir_if_missing(args.save_dir)
log_path = os.path.join(args.save_dir, 'log.txt')
with open(log_path, 'w') as f:
f.write('{}\n'.format(args))
# 定义训练集的数据增强操作和验证集的数据增强操作
# 图片的大小都已经 resize 到 256 x 256
# 训练集和验证集都只需要将图片转换成 Tensor，然后用 ImageNet 的 mean 和 std 做标准化
tfms = T.Compose([
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])
train_dataset = CaptionDataset(args.data_folder,
args.data_name,
split='TRAIN',
transform=tfms)
val_dataset = CaptionDataset(args.data_folder,
args.data_name,
split='VAL',
transform=tfms)
train_loader = DataLoader(
train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers) # suffle:打乱 num_workers:数据加载的子进程数量
val_loader = DataLoader(val_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers)
word_map_file = os.path.join(args.data_folder,
'WORDMAP_' + args.data_name + '.json')
with open(word_map_file, 'r') as f:
word_map = json.load(f)
# 初始化模型
encoder = Encoder()
encoder.freeze_params(args.freeze_encoder)
decoder = DecoderWithAttention(attention_dim=args.attention_dim,
embed_dim=args.embed_dim,
decoder_dim=args.decoder_dim,
vocab_size=len(word_map),
dropout=args.dropout)
# 定义 Encoder 和 Decoder 的优化器
encoder_optimizer = torch.optim.Adam(params=filter(
lambda p: p.requires_grad, encoder.parameters()),
lr=args.encoder_lr)
decoder_optimizer = torch.optim.Adam(params=filter(
lambda p: p.requires_grad, decoder.parameters()),
lr=args.decoder_lr)
# 把模型放到 GPU 上
encoder = encoder.to(device)
decoder = decoder.to(device)
criterion = nn.CrossEntropyLoss()
train(args=args,
train_loader=train_loader,
val_loader=val_loader,
encoder=encoder,
decoder=decoder,
criterion=criterion,
encoder_optimizer=encoder_optimizer,
decoder_optimizer=decoder_optimizer,
log_path=log_path)
def fit(t_params, checkpoint=None, m_params=None, logger=None):
# info
data_name = t_params['data_name']
imgs_path = t_params['imgs_path']
df_path = t_params['df_path']
vocab = t_params['vocab']
start_epoch = 0
epochs_since_improvement = 0
best_bleu4 = 0
epochs = t_params['epochs']
batch_size = t_params['batch_size']
workers = t_params['workers']
encoder_lr = t_params['encoder_lr']
decoder_lr = t_params['decoder_lr']
fine_tune_encoder = t_params['fine_tune_encoder']
# pretrained word embeddings
pretrained_embeddings = t_params['pretrained_embeddings']
if pretrained_embeddings:
fine_tune_embeddings = t_params['fine_tune_embeddings']
embeddings_matrix = m_params['embeddings_matrix']
# init / load checkpoint
if checkpoint is None:
# getting hyperparameters
attention_dim = m_params['attention_dim']
embed_dim = m_params['embed_dim']
decoder_dim = m_params['decoder_dim']
encoder_dim = m_params['encoder_dim']
dropout = m_params['dropout']
decoder = DecoderWithAttention(attention_dim=attention_dim,
embed_dim=embed_dim,
decoder_dim=decoder_dim,
encoder_dim=encoder_dim,
vocab_size=len(vocab),
dropout=dropout)
if pretrained_embeddings:
decoder.load_pretrained_embeddings(
torch.tensor(embeddings_matrix, dtype=torch.float32))
decoder.fine_tune_embeddings(fine_tune=fine_tune_embeddings)
decoder_optimizer = torch.optim.RMSprop(params=filter(
lambda p: p.requires_grad, decoder.parameters()),
lr=decoder_lr)
encoder = Encoder()
encoder.fine_tune(fine_tune_encoder)
encoder_optimizer = torch.optim.RMSprop(
params=filter(lambda p: p.requires_grad, encoder.parameters()),
lr=encoder_lr) if fine_tune_encoder else None
# load checkpoint
else:
checkpoint = torch.load(checkpoint)
print('Loaded Checkpoint!!')
start_epoch = checkpoint['epoch'] + 1
print(f"Starting Epoch: {start_epoch}")
epochs_since_improvement = checkpoint['epochs_since_imrovment']
best_bleu4 = checkpoint['bleu-4']
decoder = checkpoint['decoder']
decoder_optimizer = checkpoint['deocder_optimizer']
encoder = checkpoint['encoder']
encoder_optimizer = checkpoint['encoder_optimizer']
if fine_tune_encoder is True and encoder_optimizer is None:
encoder.fine_tune(fine_tune_encoder)
encoder_optimizer = torch.optim.RMSprop(params=filter(
lambda p: p.requires_grad, encoder.parameters()),
lr=encoder_lr)
# Schedulers
decoder_scheduler = ReduceLROnPlateau(decoder_optimizer,
patience=2,
verbose=True)
if fine_tune_encoder:
encoder_scheduler = ReduceLROnPlateau(encoder_optimizer,
patience=2,
verbose=True)
# move to gpu, if available
decoder = decoder.to(device)
encoder = encoder.to(device)
# loss function
criterion = nn.CrossEntropyLoss().to(device)
# dataloaders
transform = transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
])
print('Loading Data')
train_loader, val_loader = get_loaders(batch_size, imgs_path, df_path,
transform, vocab, False, workers)
print('_' * 50)
print('-' * 20, 'Fitting', '-' * 20)
for epoch in range(start_epoch, epochs):
# if epochs_since_improvement > 0 and epochs_since_improvement % 2 == 0:
# adjust_learning_rate(decoder_optimizer, 0.8)
# if fine_tune_encoder:
# adjust_learning_rate(encoder_optimizer, 0.8)
print('_' * 50)
print('-' * 20, 'Training', '-' * 20)
# one epoch of training
epoch_time = AverageMeter()
start_time = time.time()
train(train_loader=train_loader,
encoder=encoder,
decoder=decoder,
criterion=criterion,
encoder_optimizer=encoder_optimizer,
decoder_optimizer=decoder_optimizer,
epoch=epoch,
logger=logger)
epoch_time.update(time.time() - start_time)
print(f"Epoch train time {epoch_time.val:.3f} (epoch_time.avg:.3f)")
# one epoch of validation
epoch_time = AverageMeter()
start_time = time.time()
print('-' * 20, 'Validation', '-' * 20)
b1, b2, b3, recent_bleu4 = validate(val_loader=val_loader,
encoder=encoder,
decoder=decoder,
criterion=criterion,
vocab=vocab,
epoch=epoch,
logger=logger)
epoch_time.update(time.time() - start_time)
# tensorboard
logger.add_scalar(f'b-1/valid', b1, epoch)
logger.add_scalar(f'b-2/valid', b2, epoch)
logger.add_scalar(f'b-3/valid', b3, epoch)
logger.add_scalar(f'b-4/valid', recent_bleu4, epoch)
# logger.add_scalar(f'Meteor/valid', m, epoch)
print(
f"Epoch validation time {epoch_time.val:.3f} (epoch_time.avg:.3f)")
# check for improvement
is_best = recent_bleu4 > best_bleu4
best_bleu4 = max(recent_bleu4, best_bleu4)
if not is_best:
epochs_since_improvement += 1
print(
f'\nEpochs since last improvement: {epochs_since_improvement,}'
)
else:
# reset
epochs_since_improvement = 0
# stop training if no improvement for 5 epochs
if epochs_since_improvement == 5:
print('No improvement for 5 consecutive epochs, terminating...')
break
# learning rate schedular
decoder_scheduler.step(recent_bleu4)
if fine_tune_encoder:
encoder_scheduler.step(recent_bleu4)
save_checkpoint(data_name, epoch, epochs_since_improvement, encoder,
decoder, encoder_optimizer, decoder_optimizer,
recent_bleu4, is_best)
def main():
"""
Training and validation.
"""
global best_bleu4, epochs_since_improvement, checkpoint, start_epoch, fine_tune_encoder, data_name, word_map
best_bleu4 = config.best_bleu4
epochs_since_improvement = config.epochs_since_improvement
checkpoint = config.checkpoint
start_epoch = config.start_epoch
fine_tune_encoder = config.fine_tune_encoder
data_name = config.data_name
checkpoint = config.checkpoint
log_f = open(config.train_log_path, 'a+', encoding='utf-8')
# Read word map
word_map_file = os.path.join(config.data_folder,
'WORDMAP_' + data_name + '.json')
with open(word_map_file, 'r') as j:
word_map = json.load(j)
# Initialize / load checkpoint
if checkpoint is None:
print('no checkpoint, rebuild')
log_f.write('\n\nno checkpoint, rebuild' + '\n')
decoder = DecoderWithAttention(attention_dim=config.attention_dim,
embed_dim=config.emb_dim,
decoder_dim=config.decoder_dim,
vocab_size=len(word_map),
dropout=config.dropout)
decoder_optimizer = torch.optim.Adam(params=filter(
lambda p: p.requires_grad, decoder.parameters()),
lr=config.decoder_lr)
encoder = Encoder()
encoder.fine_tune(fine_tune_encoder)
encoder_optimizer = torch.optim.Adam(
params=filter(lambda p: p.requires_grad, encoder.parameters()),
lr=config.encoder_lr) if fine_tune_encoder else None
else:
print('checkpoint exist,continue.. \n{}'.format(checkpoint))
log_f.write('\n\ncheckpoint exist,continue.. \n{}'.format(checkpoint) +
'\n')
log_f.close()
checkpoint = torch.load(
checkpoint,
map_location=config.device) # map_location=device for cpu
start_epoch = checkpoint['epoch'] + 1
epochs_since_improvement = checkpoint['epochs_since_improvement']
best_bleu4 = checkpoint['bleu-4']
decoder = checkpoint['decoder']
decoder_optimizer = checkpoint['decoder_optimizer']
encoder = checkpoint['encoder']
encoder_optimizer = checkpoint['encoder_optimizer']
if fine_tune_encoder is True and encoder_optimizer is None: # check for fine tuning
encoder.fine_tune(
fine_tune_encoder) # change requires_grad for weights
encoder_optimizer = torch.optim.Adam(params=filter(
lambda p: p.requires_grad, encoder.parameters()),
lr=config.encoder_lr)
# Move to GPU, if available
# decoder = decoder.to(config.device) # no GPU
# encoder = encoder.to(config.device) # no GPU
# Loss function
# criterion = nn.CrossEntropyLoss().to(config.device) # no GPU
criterion = nn.CrossEntropyLoss()
# Custom batch dataloaders
normalize = transforms.Normalize(
mean=[0.485, 0.456, 0.406], # 这里是原ResNet的mean和std
std=[0.229, 0.224, 0.225])
train_loader = torch.utils.data.DataLoader(
CaptionDataset(config.data_folder,
data_name,
'TRAIN',
transform=transforms.Compose(
[normalize])), # CaptionDataset is in datasets.py
batch_size=config.batch_size,
shuffle=True,
num_workers=config.workers,
pin_memory=True)
val_loader = torch.utils.data.DataLoader(CaptionDataset(
config.data_folder,
data_name,
'VAL',
transform=transforms.Compose([normalize])),
batch_size=config.batch_size,
shuffle=True,
num_workers=config.workers,
pin_memory=True)
# Epochs
val_writer = SummaryWriter(
log_dir=config.tensorboard_path + '/val/' +
time.strftime('%m-%d_%H%M', time.localtime())) # for tensorboard
for epoch in range(start_epoch, config.epochs):
# Decay learning rate if there is no improvement for 8 consecutive epochs, and terminate training after 20
if epochs_since_improvement == 20:
break
if epochs_since_improvement > 0 and epochs_since_improvement % 8 == 0:
adjust_learning_rate(decoder_optimizer, 0.8) # utils.py
if fine_tune_encoder:
adjust_learning_rate(encoder_optimizer, 0.8)
# One epoch's training
train(train_loader=train_loader,
encoder=encoder,
decoder=decoder,
criterion=criterion,
encoder_optimizer=encoder_optimizer,
decoder_optimizer=decoder_optimizer,
epoch=epoch)
# One epoch's validation
recent_bleu4 = validate(val_loader=val_loader,
encoder=encoder,
decoder=decoder,
criterion=criterion,
writer=val_writer,
epoch=epoch)
# Check if there was an improvement, check each epoch
is_best = recent_bleu4 > best_bleu4
best_bleu4 = max(recent_bleu4, best_bleu4)
log_f = open(config.train_log_path, 'a+', encoding='utf-8')
if not is_best:
epochs_since_improvement += 1
print("\nEpochs since last improvement: %d\n" %
(epochs_since_improvement, ))
log_f.write("\nEpochs since last improvement: %d\n" %
(epochs_since_improvement, ) + '\n')
else:
epochs_since_improvement = 0
log_f.write('\n')
log_f.close()
# Save checkpoint
save_checkpoint(data_name, epoch, epochs_since_improvement, encoder,
decoder, encoder_optimizer, decoder_optimizer,
recent_bleu4, is_best)
val_writer.close()
def main():
"""
Training and validation.
"""
global best_bleu4, epochs_since_improvement, checkpoint, start_epoch, fine_tune_encoder, data_name, word_map
# Read word map
word_map_file = os.path.join(data_folder, 'WORDMAP_' + data_name + '.json')
with open(word_map_file, 'r') as j:
word_map = json.load(j)
# Initialize / load checkpoint
decoder = Fine_Tune_DecoderWithAttention(attention_dim=attention_dim,
embed_dim=emb_dim,
decoder_dim=decoder_dim,
vocab_size=len(word_map),
dropout=dropout)
val_decoder = DecoderWithAttention(attention_dim=attention_dim,
embed_dim=emb_dim,
decoder_dim=decoder_dim,
vocab_size=len(word_map),
dropout=dropout)
decoder_optimizer = torch.optim.Adam(params=filter(lambda p: p.requires_grad, decoder.parameters()),
lr=decoder_lr)
encoder = Encoder()
encoder.fine_tune(fine_tune_encoder)
encoder_optimizer = torch.optim.Adam(params=filter(lambda p: p.requires_grad, encoder.parameters()),
lr=encoder_lr) if fine_tune_encoder else None
g_remover = RemoveGenderRegion()
if checkpoint is not None:
if is_cpu:
checkpoint = torch.load(checkpoint, map_location='cpu')
else:
checkpoint = torch.load(checkpoint)
start_epoch = checkpoint['epoch'] + 1
epochs_since_improvement = checkpoint['epochs_since_improvement']
best_bleu4 = checkpoint['bleu-4']
decoder = load_parameter(checkpoint['decoder'], decoder)
encoder = load_parameter(checkpoint['encoder'], encoder)
# decoder_optimizer = checkpoint['decoder_optimizer']
decoder_optimizer = load_parameter(checkpoint['decoder_optimizer'], decoder_optimizer)
#encoder_optimizer = checkpoint['encoder_optimizer']
if fine_tune_encoder is True and encoder_optimizer is None:
encoder_optimizer = load_parameter(checkpoint['encoder_optimizer'], encoder_optimizer)
if fine_tune_encoder is True and encoder_optimizer is None:
encoder.fine_tune(fine_tune_encoder)
encoder_optimizer = torch.optim.Adam(params=filter(lambda p: p.requires_grad, encoder.parameters()),
lr=encoder_lr)
if freeze_decoder_lstm:
decoder.freeze_LSTM(freeze=True)
# Move to GPU, if available
decoder = decoder.to(device)
encoder = encoder.to(device)
g_remover = g_remover.to(device)
# Loss function
criterion = nn.CrossEntropyLoss().to(device)
# Custom dataloaders
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
# fix CUDA bug
if not is_cpu:
for state in decoder_optimizer.state.values():
for k, v in state.items():
if isinstance(v, torch.Tensor):
state[k] = v.cuda()
'''
train_loader = torch.utils.data.DataLoader(
CaptionDataset(data_folder, data_name, 'TRAIN', transform=transforms.Compose([normalize])),
batch_size=batch_size, shuffle=True, num_workers=workers, pin_memory=True)
val_loader = torch.utils.data.DataLoader(
CaptionDataset(data_folder, data_name, 'VAL', transform=transforms.Compose([normalize])),
batch_size=batch_size, shuffle=True, num_workers=workers, pin_memory=True)
'''
if not supervised_training:
train_loader = torch.utils.data.DataLoader(
Fine_Tune_CaptionDataset(data_folder, data_name, 'TRAIN', transform=transforms.Compose([normalize])),
batch_size=batch_size, shuffle=False, num_workers=workers, pin_memory=True)
else:
train_loader = torch.utils.data.DataLoader(
Fine_Tune_CaptionDataset_With_Mask(data_folder, data_name, 'TRAIN', transform=transforms.Compose([normalize])),
batch_size=batch_size, shuffle=False, num_workers=workers, pin_memory=True)
val_loader = torch.utils.data.DataLoader(
CaptionDataset(data_folder, data_name, 'VAL', transform=transforms.Compose([normalize])),
batch_size=batch_size, shuffle=True, num_workers=workers, pin_memory=True)
# Epochs
for epoch in range(start_epoch, epochs):
# Decay learning rate if there is no improvement for 8 consecutive epochs, and terminate training after 20
if epochs_since_improvement == 20:
break
if epochs_since_improvement > 0 and epochs_since_improvement % 8 == 0:
adjust_learning_rate(decoder_optimizer, 0.8)
if fine_tune_encoder:
adjust_learning_rate(encoder_optimizer, 0.8)
if not supervised_training:
# One epoch's training
self_guided_fine_tune_train(train_loader=train_loader,
encoder=encoder,
decoder=decoder,
criterion=criterion,
encoder_optimizer=encoder_optimizer,
decoder_optimizer=decoder_optimizer,
g_remover=g_remover,
epoch=epoch)
else:
supervised_guided_fine_tune_train(train_loader=train_loader,
encoder=encoder,
decoder=decoder,
criterion=criterion,
encoder_optimizer=encoder_optimizer,
decoder_optimizer=decoder_optimizer,
g_remover=g_remover,
epoch=epoch)
# One epoch's validation
val_decoder = load_parameter(decoder, val_decoder)
val_decoder = val_decoder.to(device)
recent_bleu4 = validate(val_loader=val_loader,
encoder=encoder,
decoder=val_decoder,
criterion=criterion)
# Check if there was an improvement
is_best = recent_bleu4 > best_bleu4
best_bleu4 = max(recent_bleu4, best_bleu4)
if not is_best:
epochs_since_improvement += 1
print("\nEpochs since last improvement: %d\n" % (epochs_since_improvement,))
else:
epochs_since_improvement = 0
# Save checkpoint
save_checkpoint(data_name, epoch, epochs_since_improvement, encoder, decoder, encoder_optimizer,
decoder_optimizer, recent_bleu4, is_best, checkpoint_savepath)
def main():
"""
Training and validation.
"""
global best_bleu4, epochs_since_improvement, checkpoint, start_epoch, fine_tune_encoder, data_name, word_map
# Read word map
word_map_file = os.path.join(data_folder, 'WORDMAP_' + data_name + '.json')
with open(word_map_file, 'r') as j:
word_map = json.load(j)
# Initialize / load checkpoint
if checkpoint is None:
emb_dim=100 #remove if not usiong pretrained model
decoder = DecoderWithAttention(attention_dim=attention_dim,
embed_dim=emb_dim,
decoder_dim=decoder_dim,
vocab_size=len(word_map),
dropout=dropout)
pretrained_embeddings = decoder.create_pretrained_embedding_matrix(word_map)
decoder.load_pretrained_embeddings(
pretrained_embeddings) # pretrained_embeddings should be of dimensions (len(word_map), emb_dim)
decoder.fine_tune_embeddings(True)
decoder_optimizer = torch.optim.Adam(params=filter(lambda p: p.requires_grad, decoder.parameters()),
lr=decoder_lr)
encoder = Encoder()
encoder.fine_tune(fine_tune_encoder)
encoder_optimizer = torch.optim.Adam(params=filter(lambda p: p.requires_grad, encoder.parameters()),
lr=encoder_lr) if fine_tune_encoder else None
else:
checkpoint = torch.load(checkpoint)
start_epoch = checkpoint['epoch'] + 1
epochs_since_improvement = checkpoint['epochs_since_improvement']
best_bleu4 = checkpoint['bleu-4']
decoder = checkpoint['decoder']
decoder_optimizer = checkpoint['decoder_optimizer']
encoder = checkpoint['encoder']
encoder_optimizer = checkpoint['encoder_optimizer']
if fine_tune_encoder is True and encoder_optimizer is None:
encoder.fine_tune(fine_tune_encoder)
encoder_optimizer = torch.optim.Adam(params=filter(lambda p: p.requires_grad, encoder.parameters()),
lr=encoder_lr)
# Move to GPU, if available
decoder = decoder.to(device)
encoder = encoder.to(device)
# Loss function
criterion = nn.CrossEntropyLoss().to(device)
# Custom dataloaders
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
train_loader = torch.utils.data.DataLoader(
CaptionDataset(data_folder, data_name, 'TRAIN', transform=transforms.Compose([normalize])),
batch_size=batch_size, shuffle=True, num_workers=workers, pin_memory=True)
val_loader = torch.utils.data.DataLoader(
CaptionDataset(data_folder, data_name, 'VAL', transform=transforms.Compose([normalize])),
batch_size=batch_size, shuffle=True, num_workers=workers, pin_memory=True)
# Epochs
for epoch in range(start_epoch, epochs):
# Decay learning rate if there is no improvement for 8 consecutive epochs, and terminate training after 20
if epochs_since_improvement == 20:
break
if epochs_since_improvement > 0 and epochs_since_improvement % 8 == 0:
adjust_learning_rate(decoder_optimizer, 0.8)
if fine_tune_encoder:
adjust_learning_rate(encoder_optimizer, 0.8)
# One epoch's training
train(train_loader=train_loader,
encoder=encoder,
decoder=decoder,
criterion=criterion,
encoder_optimizer=encoder_optimizer,
decoder_optimizer=decoder_optimizer,
epoch=epoch)
# One epoch's validation
recent_bleu4, val_loss_avg, val_accu_avg = validate(val_loader=val_loader,
encoder=encoder,
decoder=decoder,
criterion=criterion)
#write to tensorboard
writer.add_scalar('validation_loss', val_loss_avg, epoch)
writer.add_scalar('validation_accuracy', val_accu_avg, epoch)
writer.add_scalar('validation_bleu4', recent_bleu4, epoch)
# Check if there was an improvement
is_best = recent_bleu4 > best_bleu4
best_bleu4 = max(recent_bleu4, best_bleu4)
if not is_best:
epochs_since_improvement += 1
print("\nEpochs since last improvement: %d\n" % (epochs_since_improvement,))
else:
epochs_since_improvement = 0
# Save checkpoint
print("Saving model to file",ckpt_name.format(epoch, bleu=recent_bleu4, loss=val_loss_avg, acc=val_accu_avg))
save_checkpoint(ckpt_name.format(epoch, bleu=recent_bleu4, loss=val_loss_avg, acc=val_accu_avg),
epoch, epochs_since_improvement, encoder, decoder, encoder_optimizer,
decoder_optimizer, recent_bleu4, is_best)
#close tensorboard writer
writer.close()
def main():
"""
Training and validation.
"""
global best, epochs_since_improvement, checkpoint, start_epoch, fine_tune_encoder
if checkpoint is None:
decoder = DecoderWithAttention(
attention_dim=attention_dim,
embed_dim=emb_dim,
decoder_dim=decoder_dim,
# vocab_size=len(word_map),
vocab_size=2, # X, Y coordinates and use it for regression
dropout=dropout)
decoder_optimizer = torch.optim.Adam(params=filter(
lambda p: p.requires_grad, decoder.parameters()),
lr=decoder_lr)
decoderMulti = network.__dict__['MultiTask'](output_size)
decoderMulti_optimizer = torch.optim.Adam(
decoderMulti.parameters(),
lr=decoderMulti_lr,
weight_decay=decoderMulti_lr_weight_decay)
encoder = Encoder()
encoder_optimizer = torch.optim.Adam(
params=filter(lambda p: p.requires_grad, encoder.parameters()),
lr=encoder_lr) if fine_tune_encoder else None
else:
checkpoint = torch.load(checkpoint)
start_epoch = checkpoint['epoch'] + 1
epochs_since_improvement = checkpoint['epochs_since_improvement']
best = checkpoint['b4']
decoder = checkpoint['decoder']
decoder_optimizer = checkpoint['decoder_optimizer']
encoder = checkpoint['encoder']
encoder_optimizer = checkpoint['encoder_optimizer']
decoderMulti = checkpoint['decoderMulti']
decoderMulti_optimizer = checkpoint['decoderMulti_optimizer']
if fine_tune_encoder is True and encoder_optimizer is None:
encoder.fine_tune(fine_tune_encoder)
encoder_optimizer = torch.optim.Adam(params=filter(
lambda p: p.requires_grad, encoder.parameters()),
lr=encoder_lr)
# Move to GPU, if available
if multiGpu:
decoder = torch.nn.DataParallel(decoder).to(device)
encoder = torch.nn.DataParallel(encoder).to(device)
decoderMulti = torch.nn.DataParallel(decoderMulti).to(device)
else:
decoder = decoder.to(device)
encoder = encoder.to(device)
decoderMulti = decoderMulti.to(device)
# Loss function
criterionBinary = nn.BCELoss().to(device)
criterionMse = nn.MSELoss().to(device)
criterion = [criterionBinary, criterionMse]
# Custom dataloaders
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
train_loader_p = torch.utils.data.DataLoader(fiberDataset_COCO(
data_folder_p, jason_file_p, image_folder, offset_folder,
transforms.Compose([transforms.ToTensor(), normalize]), True),
batch_size=batch_size,
shuffle=True,
num_workers=workers,
pin_memory=True,
drop_last=True)
val_loader_p = torch.utils.data.DataLoader(fiberDataset_COCO(
data_folder_val, jason_file_val, image_folder_val, offset_folder_val,
transforms.Compose([transforms.ToTensor(), normalize]), True),
batch_size=batch_size,
shuffle=True,
num_workers=workers,
pin_memory=True,
drop_last=True)
# Epochs
for epoch in range(start_epoch, epochs):
# Decay learning rate if there is no improvement for 8 consecutive epochs, and terminate training after 20
if epochs_since_improvement == 20:
break
if epochs_since_improvement > 0 and epochs_since_improvement % 8 == 0:
adjust_learning_rate(decoder_optimizer, 0.8)
if fine_tune_encoder:
adjust_learning_rate(encoder_optimizer, 0.8)
train(train_loader=train_loader_p,
encoder=encoder,
decoder=decoder,
decoderMulti=decoderMulti,
criterion=criterion,
encoder_optimizer=encoder_optimizer,
decoder_optimizer=decoder_optimizer,
decoderMulti_optimizer=decoderMulti_optimizer,
epoch=epoch)
recent = validate(val_loader=val_loader_p,
encoder=encoder,
decoder=decoder,
decoderMulti=decoderMulti,
criterion=criterion)
#
# # Check if there was an improvement
is_best = recent < best
best = min(recent, best)
if not is_best:
epochs_since_improvement += 1
print("\nEpochs since last improvement: %d\n" %
(epochs_since_improvement, ))
else:
epochs_since_improvement = 0
epochs_since_improvement = 0
# Save checkpoint
save_checkpoint(save_weights_name, epoch, epochs_since_improvement,
encoder, decoder, encoder_optimizer,
decoder_optimizer, decoderMulti,
decoderMulti_optimizer, best, is_best)
from models import DecoderWithAttention, Encoder
data_name = 'flickr8k_5_cap_per_img_5_min_word_freq'
word_map_file = 'datasets/caption_data/WORDMAP_' + data_name + '.json'
checkpoint = 'logs/tmp/BEST_MODEL.pth.tar'
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
with open(word_map_file, 'r') as f:
word_map = json.load(f)
rev_word_map = {v: k for k, v in word_map.items()}
vocab_size = len(word_map)
# 载入模型
encoder = Encoder()
decoder = DecoderWithAttention(512, 512, 512, vocab_size)
checkpoint = torch.load(checkpoint)
encoder.load_state_dict(checkpoint['encoder'])
decoder.load_state_dict(checkpoint['decoder'])
encoder.to(device)
decoder.to(device)
encoder.eval()
decoder.eval()
preprocess = T.Compose([
T.Resize(size=(256, 256)),
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])
def main():
"""
Training and validation.
"""
global best_bleu4, use_amp, epochs_since_improvement, checkpoint, start_epoch, fine_tune_encoder, data_name, word_map
#use_amp = True
#print("Using amp for mized precision training")
# Read word map
word_map_file = os.path.join(data_folder, 'WORDMAP_' + data_name + '.json')
with open(word_map_file, 'r') as j:
word_map = json.load(j)
# Initialize / load checkpoint
if checkpoint is None:
decoder = DecoderWithAttention(attention_dim=attention_dim,
embed_dim=emb_dim,
decoder_dim=decoder_dim,
vocab_size=len(word_map),
dropout=dropout)
decoder_optimizer = torch.optim.Adam(params=filter(lambda p: p.requires_grad, decoder.parameters()),
lr=decoder_lr)
encoder = Encoder()
encoder.fine_tune(fine_tune_encoder)
encoder_optimizer = torch.optim.Adam(params=filter(lambda p: p.requires_grad, encoder.parameters()),
lr=encoder_lr) if fine_tune_encoder else None
else:
checkpoint = torch.load(checkpoint)
start_epoch = checkpoint['epoch'] + 1
epochs_since_improvement = checkpoint['epochs_since_improvement']
best_bleu4 = checkpoint['bleu-4']
decoder = checkpoint['decoder']
decoder_optimizer = checkpoint['decoder_optimizer']
encoder = checkpoint['encoder']
encoder_optimizer = checkpoint['encoder_optimizer']
if fine_tune_encoder is True and encoder_optimizer is None:
encoder.fine_tune(fine_tune_encoder)
encoder_optimizer = torch.optim.Adam(params=filter(lambda p: p.requires_grad, encoder.parameters()),
lr=encoder_lr)
# Move to GPU, if available
decoder = decoder.to(device)
# use mixed precision training using Nvidia Apex
if use_amp:
decoder, decoder_optimizer = amp.initialize(
decoder, decoder_optimizer, opt_level="O2",
keep_batchnorm_fp32=True, loss_scale="dynamic")
encoder = encoder.to(device)
if not encoder_optimizer:
print("Encoder is not being optimized")
elif use_amp:
encoder, encoder_optimizer = amp.initialize(
encoder, encoder_optimizer, opt_level="O2",
keep_batchnorm_fp32=True, loss_scale="dynamic")
# Loss function
criterion = nn.CrossEntropyLoss().to(device)
# Custom dataloaders
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
train_loader = torch.utils.data.DataLoader(
CaptionDataset(data_folder, data_name, 'TRAIN', transform=transforms.Compose([normalize])),
batch_size=batch_size, shuffle=True, num_workers=workers, pin_memory=True)
val_loader = torch.utils.data.DataLoader(
CaptionDataset(data_folder, data_name, 'VAL', transform=transforms.Compose([normalize])),
batch_size=batch_size, shuffle=True, num_workers=workers, pin_memory=True)
# Epochs
for epoch in range(start_epoch, epochs):
# Decay learning rate if there is no improvement for 8 consecutive epochs, and terminate training after 20
if epochs_since_improvement == 20:
break
if epochs_since_improvement > 0 and epochs_since_improvement % 8 == 0:
adjust_learning_rate(decoder_optimizer, 0.8)
if fine_tune_encoder:
adjust_learning_rate(encoder_optimizer, 0.8)
# One epoch's training
train(train_loader=train_loader,
encoder=encoder,
decoder=decoder,
criterion=criterion,
encoder_optimizer=encoder_optimizer,
decoder_optimizer=decoder_optimizer,
epoch=epoch)
# One epoch's validation
recent_bleu4 = validate(val_loader=val_loader,
encoder=encoder,
decoder=decoder,
criterion=criterion)
# Check if there was an improvement
is_best = recent_bleu4 > best_bleu4
best_bleu4 = max(recent_bleu4, best_bleu4)
if not is_best:
epochs_since_improvement += 1
print("\nEpochs since last improvement: %d\n" % (epochs_since_improvement,))
else:
epochs_since_improvement = 0
# Save checkpoint
save_checkpoint(data_name, epoch, epochs_since_improvement, encoder, decoder, encoder_optimizer,
decoder_optimizer, recent_bleu4, is_best)
def main():
"""
Training and validation.
"""
global best_bleu4, epochs_since_improvement, checkpoint, start_epoch, fine_tune_encoder, data_name, word_map
if dual_encoder: # this is always initialized with pre-trained models:
print("DUAL ENCODER")
if dual_encoder_checkpoint is not None:
print('Loaded Dual Encoder Checkpoint')
dual_branch_checkpoint = torch.load(checkpoint,
map_location='cuda:0')
encoder = dual_branch_checkpoint['encoder']
decoder = dual_branch_checkpoint['decoder']
decoder_optimizer = torch.optim.Adam(params=filter(
lambda p: p.requires_grad, decoder.parameters()),
lr=decoder_lr)
else:
main_branch_checkpoint = torch.load(checkpoint,
map_location='cuda:0')
encoder = DualEncoder(sketch_resnet=sketch_encoder_resnet)
encoder.m_resnet = main_branch_checkpoint['encoder'].resnet
print("Use pre-trained resnet")
# encoder.m_adaptive_pool = main_branch_checkpoint['encoder'].adaptive_pool
decoder = main_branch_checkpoint['decoder']
decoder_optimizer = torch.optim.Adam(params=filter(
lambda p: p.requires_grad, decoder.parameters()),
lr=decoder_lr)
if fine_tune_encoder is True:
print("!!! Will fine tune Encoder !!!")
encoder.fine_tune(fine_tune_encoder)
encoder_optimizer = torch.optim.Adam(params=filter(
lambda p: p.requires_grad, encoder.parameters()),
lr=encoder_lr)
else:
encoder_optimizer = torch.optim.Adam(params=filter(
lambda p: p.requires_grad, encoder.parameters()),
lr=encoder_lr)
else: # following method is for One Encoder architecture
# Initialize / load checkpoint
if checkpoint is None:
decoder = DecoderWithAttention(attention_dim=attention_dim,
embed_dim=emb_dim,
decoder_dim=decoder_dim,
vocab_size=len(word_map),
dropout=dropout)
decoder_optimizer = torch.optim.Adam(params=filter(
lambda p: p.requires_grad, decoder.parameters()),
lr=decoder_lr)
encoder = Encoder(specify_resnet=main_encoder_resnet)
encoder.fine_tune(fine_tune_encoder)
encoder_optimizer = torch.optim.Adam(
params=filter(lambda p: p.requires_grad, encoder.parameters()),
lr=encoder_lr) if fine_tune_encoder else None
else:
checkpoint = torch.load(checkpoint, map_location='cuda:0')
# start_epoch = checkpoint['epoch'] + 1
# epochs_since_improvement = checkpoint['epochs_since_improvement']
# best_bleu4 = checkpoint['bleu-4'] this metric is unfair when we switch to a different domain
decoder = checkpoint['decoder']
# decoder_optimizer = checkpoint['decoder_optimizer']
decoder_optimizer = torch.optim.Adam(params=filter(
lambda p: p.requires_grad, decoder.parameters()),
lr=decoder_lr)
if main_encoder_resnet is not None:
encoder = Encoder(
specify_resnet=main_encoder_resnet
) # specify here so the encoder remove the last 2 layers of resnet
encoder.adaptive_pool = checkpoint['encoder'].adaptive_pool
else:
encoder = checkpoint['encoder']
# encoder_optimizer = checkpoint['encoder_optimizer']
# if fine_tune_encoder is True and encoder_optimizer is None:
if fine_tune_encoder is True:
print("Will fine tune Encoder")
encoder.fine_tune(fine_tune_encoder)
encoder_optimizer = torch.optim.Adam(params=filter(
lambda p: p.requires_grad, encoder.parameters()),
lr=encoder_lr)
else:
encoder_optimizer = torch.optim.Adam(params=filter(
lambda p: p.requires_grad, encoder.parameters()),
lr=encoder_lr)
# Move to GPU, if available
decoder = decoder.to(device)
encoder = encoder.to(device)
# Loss function
criterion = nn.CrossEntropyLoss().to(device)
# Custom dataloaders
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
# data augmention for nycc dataset
augment = transforms.Compose([
transforms.RandomAffine(20, (0.1, 0.1), (0.8, 1.2)),
transforms.RandomHorizontalFlip(p=0.5)
])
train_loader = torch.utils.data.DataLoader(CaptionDataset(
data_folder,
data_name,
'TRAIN',
transform=transforms.Compose([augment, normalize])),
batch_size=batch_size,
shuffle=True,
num_workers=workers,
pin_memory=True)
val_loader = torch.utils.data.DataLoader(CaptionDataset(
data_folder,
data_name,
'VAL',
transform=transforms.Compose([normalize])),
batch_size=batch_size,
shuffle=True,
num_workers=workers,
pin_memory=True)
# Epochs
for epoch in range(start_epoch, epochs):
# Decay learning rate if there is no improvement for 8 consecutive epochs, and terminate training after 20
# if epochs_since_improvement == 40:
# break
# if epochs_since_improvement > 0 and epochs_since_improvement % 8 == 0:
# adjust_learning_rate(decoder_optimizer, 0.8)
# if fine_tune_encoder:
# adjust_learning_rate(encoder_optimizer, 0.8)
# One epoch's training
train(train_loader=train_loader,
encoder=encoder,
decoder=decoder,
criterion=criterion,
encoder_optimizer=encoder_optimizer,
decoder_optimizer=decoder_optimizer,
epoch=epoch)
# One epoch's validation
recent_bleu4 = validate(val_loader=val_loader,
encoder=encoder,
decoder=decoder,
criterion=criterion,
epoch=epoch)
# Check if there was an improvement
is_best = recent_bleu4 > best_bleu4
best_bleu4 = max(recent_bleu4, best_bleu4)
if not is_best:
epochs_since_improvement += 1
print("\nEpochs since last improvement: %d\n" %
(epochs_since_improvement, ))
else:
epochs_since_improvement = 0
# Save checkpoint
print(" *** saving model with bleu score: ", recent_bleu4)
save_checkpoint(data_name, epoch, epochs_since_improvement,
encoder, decoder, encoder_optimizer,
decoder_optimizer, recent_bleu4, is_best)
print(" *** LAST EPOCH saving model with bleu score: ", recent_bleu4)
save_checkpoint(data_name, epoch, epochs_since_improvement, encoder,
decoder, encoder_optimizer, decoder_optimizer,
recent_bleu4, is_best)
def main():
"""
Training and validation.
"""
global best_bleu4, epochs_since_improvement, checkpoint, start_epoch, fine_tune_encoder, data_name, word_map
# Read word map
word_map_file = os.path.join(data_folder, 'WORDMAP_' + data_name + '.json')
with open(word_map_file, 'r') as j:
word_map = json.load(j)
# Initialize / load checkpoint
if checkpoint is None:
decoder = DecoderWithAttention(attention_dim=attention_dim,
embed_dim=emb_dim,
decoder_dim=decoder_dim,
vocab_size=len(word_map),
dropout=dropout)
decoder_optimizer = torch.optim.Adam(params=filter(
lambda p: p.requires_grad, decoder.parameters()),
lr=decoder_lr)
encoder = Encoder()
encoder.fine_tune(fine_tune_encoder)
encoder_optimizer = torch.optim.Adam(
params=filter(lambda p: p.requires_grad, encoder.parameters()),
lr=encoder_lr) if fine_tune_encoder else None
else:
checkpoint = torch.load(checkpoint)
start_epoch = checkpoint['epoch'] + 1
epochs_since_improvement = checkpoint['epochs_since_improvement']
best_bleu4 = checkpoint['bleu-4']
decoder = checkpoint['decoder']
decoder_optimizer = checkpoint['decoder_optimizer']
encoder = checkpoint['encoder']
encoder_optimizer = checkpoint['encoder_optimizer']
if fine_tune_encoder is True and encoder_optimizer is None:
encoder.fine_tune(fine_tune_encoder)
encoder_optimizer = torch.optim.Adam(params=filter(
lambda p: p.requires_grad, encoder.parameters()),
lr=encoder_lr)
# Move to GPU, if available
decoder = decoder.to(device)
encoder = encoder.to(device)
# Loss function
criterion = nn.CrossEntropyLoss().to(device)
train_annotations = COCO(
os.path.join('dataset', 'annotations', 'instances_train2014.json'))
val_annotations = COCO(
os.path.join('dataset', 'annotations', 'instances_val2014.json'))
# Custom dataloaders
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
train_loader = torch.utils.data.DataLoader(
CaptionDatasetFastTextWithReplacementCV(
data_folder,
data_name,
'TRAIN',
transform=transforms.Compose([normalize]),
train_annotations=train_annotations,
val_annotations=val_annotations),
batch_size=batch_size,
shuffle=True,
num_workers=workers,
pin_memory=True,
collate_fn=my_collate)
val_loader = torch.utils.data.DataLoader(CaptionDatasetFastText(
data_folder,
data_name,
'VAL',
transform=transforms.Compose([normalize]),
train_annotations=train_annotations,
val_annotations=val_annotations),
batch_size=batch_size,
shuffle=True,
num_workers=workers,
pin_memory=True,
collate_fn=my_collate)
# Epochs
for epoch in range(start_epoch, epochs):
# Decay learning rate if there is no improvement for 8 consecutive epochs, and terminate training after 20
if epochs_since_improvement == 5:
break
if epochs_since_improvement > 0 and epochs_since_improvement % 8 == 0:
adjust_learning_rate(decoder_optimizer, 0.8)
if fine_tune_encoder:
adjust_learning_rate(encoder_optimizer, 0.8)
# One epoch's training
train(train_loader=train_loader,
encoder=encoder,
decoder=decoder,
criterion=criterion,
encoder_optimizer=encoder_optimizer,
decoder_optimizer=decoder_optimizer,
epoch=epoch)
# One epoch's validation
recent_bleu4 = validate(val_loader=val_loader,
encoder=encoder,
decoder=decoder,
criterion=criterion)
# Check if there was an improvement
is_best = recent_bleu4 > best_bleu4
best_bleu4 = max(recent_bleu4, best_bleu4)
if not is_best:
epochs_since_improvement += 1
print("\nEpochs since last improvement: %d\n" %
(epochs_since_improvement, ))
else:
epochs_since_improvement = 0
# Save checkpoint
save_checkpoint(data_name, epoch, epochs_since_improvement, encoder,
decoder, encoder_optimizer, decoder_optimizer,
recent_bleu4, is_best)
#Show timestamps of the end of each epoch
now = datetime.now()
print("now =", now)
# dd/mm/YY H:M:S
dt_string = now.strftime("%d/%m/%Y %H:%M:%S")
print("date and time = ", dt_string)
def main():
"""
训练和验证
"""
global best_bleu4, epochs_since_improvement, checkpoint, start_epoch, fine_tune_encoder, data_name, word_map
# 读入词典
word_map_file = os.path.join(data_folder, 'WORDMAP_' + data_name + '.json')
with open(word_map_file, 'r') as j:
word_map = json.load(j)
# 初始化/加载模型
if checkpoint is None:
decoder = DecoderWithAttention(hidden_size=hidden_size,
vocab_size=len(word_map),
attention_dim=attention_dim,
embed_size=emb_dim,
dropout=dropout)
decoder_optimizer = torch.optim.Adam(params=decoder.parameters(),
lr=decoder_lr,
betas=(0.8, 0.999))
encoder = Encoder(hidden_size=hidden_size,
embed_size=emb_dim,
dropout=dropout)
# 是否微调
encoder.fine_tune(fine_tune_encoder)
encoder_optimizer = torch.optim.Adam(
params=filter(lambda p: p.requires_grad, encoder.parameters()),
lr=encoder_lr,
betas=(0.8, 0.999)) if fine_tune_encoder else None
else:
#载入checkpoint
checkpoint = torch.load(checkpoint)
start_epoch = checkpoint['epoch'] + 1
epochs_since_improvement = checkpoint['epochs_since_improvement']
best_bleu4 = checkpoint['bleu-4']
decoder = checkpoint['decoder']
decoder_optimizer = checkpoint['decoder_optimizer']
encoder = checkpoint['encoder']
encoder_optimizer = checkpoint['encoder_optimizer']
if fine_tune_encoder is True and encoder_optimizer is None:
# 如果此时要开始微调,需要定义优化器
encoder.fine_tune(fine_tune_encoder)
encoder_optimizer = torch.optim.Adam(params=filter(
lambda p: p.requires_grad, encoder.parameters()),
lr=encoder_lr,
betas=(0.8, 0.999))
# 移动到GPU
decoder = decoder.to(device)
encoder = encoder.to(device)
# Loss function
criterion = nn.CrossEntropyLoss().to(device)
# dataloaders
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]) #ImageNet
# pin_memory = True 驻留内存，不换进换出
train_loader = torch.utils.data.DataLoader(CaptionDataset(
data_folder,
data_name,
'TRAIN',
transform=transforms.Compose([normalize])),
batch_size=batch_size,
shuffle=True,
num_workers=workers,
pin_memory=True)
val_loader = torch.utils.data.DataLoader(CaptionDataset(
data_folder,
data_name,
'VAL',
transform=transforms.Compose([normalize])),
batch_size=batch_size,
shuffle=True,
num_workers=workers,
pin_memory=True)
# Epochs
for epoch in range(start_epoch, epochs):
if epoch > 15:
adjust_learning_rate(decoder_optimizer, epoch)
if fine_tune_encoder:
adjust_learning_rate(encoder_optimizer, epoch)
# Early Stopping if the validation score does not imporive for 6 consecutive epochs
if epochs_since_improvement == 6:
break
# 一个epoch的训练
train(train_loader=train_loader,
encoder=encoder,
decoder=decoder,
criterion=criterion,
encoder_optimizer=encoder_optimizer,
decoder_optimizer=decoder_optimizer,
epoch=epoch,
vocab_size=len(word_map))
# 一个epoch的验证
recent_bleu4 = validate(val_loader=val_loader,
encoder=encoder,
decoder=decoder,
criterion=criterion)
# 检查是否有提升
is_best = recent_bleu4 > best_bleu4
best_bleu4 = max(recent_bleu4, best_bleu4)
if not is_best:
epochs_since_improvement += 1
print("\nEpochs since last improvement: %d\n" %
(epochs_since_improvement, ))
else:
epochs_since_improvement = 0
# 保存模型
save_checkpoint(data_name, epoch, epochs_since_improvement, encoder,
decoder, encoder_optimizer, decoder_optimizer,
recent_bleu4, is_best)
def main():
"""
Training and validation.
"""
# In Python, global keyword allows you to modify the variable outside of the current scope.
# It is used to create a global variable and make changes to the variable in a local context.
'''
The basic rules for global keyword in Python are:
When we create a variable inside a function, it is local by default.
When we define a variable outside of a function, it is global by default. You don't have to use global keyword.
We use global keyword to read and write a global variable inside a function.
Use of global keyword outside a function has no effect.
'''
global best_bleu4, epochs_since_improvement, checkpoint, start_epoch, fine_tune_encoder, data_name, word_map
# Read word map
word_map_file = os.path.join(data_folder, 'WORDMAP_' + data_name + '.json')
with open(word_map_file, 'r') as j:
word_map = json.load(j)
# Initialize / load checkpoint
if checkpoint is None:
decoder = DecoderWithAttention(attention_dim=attention_dim,
embed_dim=emb_dim,
decoder_dim=decoder_dim,
vocab_size=len(word_map),
dropout=dropout)
'''
The filter() method constructs an iterator from elements of an iterable for which a function returns true.
The filter() method takes two parameters:
function - function that tests if elements of an iterable returns true or false
If None, the function defaults to Identity function - which returns false if any elements are false
iterable - iterable which is to be filtered, could be sets, lists, tuples, or containers of any iterators
The filter() method returns an iterator that passed the function check for each element in the iterable.
'''
decoder_optimizer = torch.optim.Adam(params=filter(
lambda p: p.requires_grad, decoder.parameters()),
lr=decoder_lr)
encoder = Encoder()
encoder.fine_tune(fine_tune_encoder)
encoder_optimizer = torch.optim.Adam(
params=filter(lambda p: p.requires_grad, encoder.parameters()),
lr=encoder_lr) if fine_tune_encoder else None
else:
checkpoint = torch.load(checkpoint)
start_epoch = checkpoint['epoch'] + 1
epochs_since_improvement = checkpoint['epochs_since_improvement']
best_bleu4 = checkpoint['bleu-4']
decoder = checkpoint['decoder']
decoder_optimizer = checkpoint['decoder_optimizer']
encoder = checkpoint['encoder']
encoder_optimizer = checkpoint['encoder_optimizer']
if fine_tune_encoder is True and encoder_optimizer is None:
encoder.fine_tune(fine_tune_encoder)
encoder_optimizer = torch.optim.Adam(params=filter(
lambda p: p.requires_grad, encoder.parameters()),
lr=encoder_lr)
# Move to GPU, if available
decoder = decoder.to(device)
encoder = encoder.to(device)
# Loss function
criterion = nn.CrossEntropyLoss().to(device)
# Custom dataloaders
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
train_loader = torch.utils.data.DataLoader(CaptionDataset(
data_folder,
data_name,
'TRAIN',
transform=transforms.Compose([normalize])),
batch_size=batch_size,
shuffle=True,
num_workers=workers,
pin_memory=True)
val_loader = torch.utils.data.DataLoader(CaptionDataset(
data_folder,
data_name,
'VAL',
transform=transforms.Compose([normalize])),
batch_size=batch_size,
shuffle=True,
num_workers=workers,
pin_memory=True)
# Epochs
for epoch in range(start_epoch, epochs):
# If there's no improvement in Bleu score for 20 epochs then stop training
if epochs_since_improvement == 20:
break
# If there's no improvement in Bleu score for 8 epochs lower the lr
if epochs_since_improvement > 0 and epochs_since_improvement % 8 == 0:
adjust_learning_rate(decoder_optimizer, 0.8)
if fine_tune_encoder:
adjust_learning_rate(encoder_optimizer, 0.8)
# One epoch's training
train(train_loader=train_loader,
encoder=encoder,
decoder=decoder,
criterion=criterion,
encoder_optimizer=encoder_optimizer,
decoder_optimizer=decoder_optimizer,
epoch=epoch)
# One epoch's validation
recent_bleu4 = validate(val_loader=val_loader,
encoder=encoder,
decoder=decoder,
criterion=criterion)
# Check if there was an improvement
is_best = recent_bleu4 > best_bleu4
best_bleu4 = max(recent_bleu4, best_bleu4)
if not is_best:
epochs_since_improvement += 1
print("\nEpochs since last improvement: %d\n" %
(epochs_since_improvement, ))
else:
epochs_since_improvement = 0
# Save checkpoint
save_checkpoint(data_name, epoch, epochs_since_improvement, encoder,
decoder, encoder_optimizer, decoder_optimizer,
recent_bleu4, is_best)
def main():
"""
Training and validation.
"""
global best_bleu4, epochs_since_improvement, checkpoint, start_epoch, fine_tune_encoder, data_name, word_map, glove_path, emb_dim, rev_word_map
# Read word map
word_map_file = os.path.join(data_folder, 'WORDMAP_' + data_name + '.json')
with open(word_map_file, 'r') as j:
word_map = json.load(j)
rev_word_map = {v: k for k, v in word_map.items()}
#get glove
vectors = bcolz.open(f'{glove_path}/6B.300.dat')[:]
words = pickle.load(open(f'{glove_path}/6B.300_words.pkl', 'rb'))
word2idx = pickle.load(open(f'{glove_path}/6B.300_idx.pkl', 'rb'))
glove = {w: vectors[word2idx[w]] for w in words}
matrix_len = len(word_map)
weights_matrix = np.zeros((matrix_len, emb_dim))
words_found = 0
for i, word in enumerate(word_map.keys()):
try:
weights_matrix[i] = glove[word]
words_found += 1
except KeyError:
weights_matrix[i] = np.random.normal(scale=0.6, size=(emb_dim, ))
# weights_matrix = np.float64(weights_matrix)
# weights_matrix = torch.from_numpy(weights_matrix)
# pretrained_embedding = weights_matrix.to(dtype=torch.float)
# print(pretrained_embedding.dtype)
# if device.type == 'cpu' :
# pretrained_embedding = torch.FloatTensor(weights_matrix)
# else:
# pretrained_embedding = torch.cuda.FloatTensor(weights_matrix)
pretrained_embedding = torch.FloatTensor(weights_matrix)
# Initialize / load checkpoint
if checkpoint is None:
decoder = DecoderWithAttention(attention_dim=attention_dim,
embed_dim=emb_dim,
decoder_dim=decoder_dim,
vocab_size=len(word_map),
dropout=dropout)
decoder.load_pretrained_embeddings(
pretrained_embedding
) # pretrained_embeddings should be of dimensions (len(word_map), emb_dim)
decoder.fine_tune_embeddings(True) # or False
decoder_optimizer = torch.optim.Adam(params=filter(
lambda p: p.requires_grad, decoder.parameters()),
lr=decoder_lr)
encoder = Encoder()
encoder.fine_tune(fine_tune_encoder)
encoder_optimizer = torch.optim.Adam(
params=filter(lambda p: p.requires_grad, encoder.parameters()),
lr=encoder_lr) if fine_tune_encoder else None
else:
checkpoint = torch.load(checkpoint)
start_epoch = checkpoint['epoch'] + 1
epochs_since_improvement = checkpoint['epochs_since_improvement']
best_bleu4 = checkpoint['bleu-4']
decoder = checkpoint['decoder']
decoder_optimizer = checkpoint['decoder_optimizer']
encoder = checkpoint['encoder']
encoder_optimizer = checkpoint['encoder_optimizer']
if fine_tune_encoder is True and encoder_optimizer is None:
encoder.fine_tune(fine_tune_encoder)
encoder_optimizer = torch.optim.Adam(params=filter(
lambda p: p.requires_grad, encoder.parameters()),
lr=encoder_lr)
# Move to GPU, if available
decoder = decoder.to(device)
encoder = encoder.to(device)
# Loss function
criterion = nn.CrossEntropyLoss().to(device)
# Custom dataloaders
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
train_loader = torch.utils.data.DataLoader(CaptionDataset(
data_folder,
data_name,
'TRAIN',
transform=transforms.Compose([normalize])),
batch_size=batch_size,
shuffle=True,
num_workers=workers,
pin_memory=True)
val_loader = torch.utils.data.DataLoader(CaptionDataset(
data_folder,
data_name,
'VAL',
transform=transforms.Compose([normalize])),
batch_size=batch_size,
shuffle=True,
num_workers=workers,
pin_memory=True)
# Epochs
for epoch in range(start_epoch, epochs):
# Decay learning rate if there is no improvement for 8 consecutive epochs, and terminate training after 20
if epochs_since_improvement == 20:
break
if epochs_since_improvement > 0 and epochs_since_improvement % 8 == 0:
adjust_learning_rate(decoder_optimizer, 0.8)
if fine_tune_encoder:
adjust_learning_rate(encoder_optimizer, 0.8)
# One epoch's training
# train(train_loader=train_loader,
# encoder=encoder,
# decoder=decoder,
# criterion=criterion,
# encoder_optimizer=encoder_optimizer,
# decoder_optimizer=decoder_optimizer,
# epoch=epoch)
# One epoch's validation
recent_bleu4 = validate(val_loader=val_loader,
encoder=encoder,
decoder=decoder,
criterion=criterion)
# Check if there was an improvement
is_best = recent_bleu4 > best_bleu4
best_bleu4 = max(recent_bleu4, best_bleu4)
if not is_best:
epochs_since_improvement += 1
print("\nEpochs since last improvement: %d\n" %
(epochs_since_improvement, ))
else:
epochs_since_improvement = 0
# Save checkpoint
save_checkpoint(data_name, epoch, epochs_since_improvement, encoder,
decoder, encoder_optimizer, decoder_optimizer,
recent_bleu4, is_best)
def main():
"""
Training and validation.
"""
global best_bleu4, epochs_since_improvement, checkpoint, start_epoch,data_name, word_map
# Read word map
word_map_file = os.path.join(data_folder, 'WORDMAP_' + data_name + '.json')
with open(word_map_file, 'r') as j:
word_map = json.load(j)
# Initialize / load checkpoint
if checkpoint is None:
decoder = DecoderWithAttention(attention_dim=attention_dim,
embed_dim=emb_dim,
decoder_dim=decoder_dim,
vocab_size=len(word_map),
dropout=dropout)
decoder_optimizer = torch.optim.Adamax(params=filter(lambda p: p.requires_grad, decoder.parameters()))
else:
checkpoint = torch.load(checkpoint)
start_epoch = checkpoint['epoch'] + 1
epochs_since_improvement = checkpoint['epochs_since_improvement']
best_bleu4 = checkpoint['bleu-4']
decoder = checkpoint['decoder']
decoder_optimizer = checkpoint['decoder_optimizer']
# Move to GPU, if available
decoder = decoder.to(device)
# Loss functions
criterion_ce = nn.CrossEntropyLoss().to(device)
criterion_dis = nn.MultiLabelMarginLoss().to(device)
# Custom dataloaders
train_loader = torch.utils.data.DataLoader(
CaptionDataset(data_folder, data_name, 'TRAIN'),
batch_size=batch_size, shuffle=True, num_workers=workers, pin_memory=True)
val_loader = torch.utils.data.DataLoader(
CaptionDataset(data_folder, data_name, 'VAL'),
batch_size=batch_size, shuffle=True, num_workers=workers, pin_memory=True)
# Epochs
for epoch in range(start_epoch, epochs):
# Decay learning rate if there is no improvement for 8 consecutive epochs, and terminate training after 20
if epochs_since_improvement == 20:
break
if epochs_since_improvement > 0 and epochs_since_improvement % 8 == 0:
adjust_learning_rate(decoder_optimizer, 0.8)
# One epoch's training
train(train_loader=train_loader,
decoder=decoder,
criterion_ce = criterion_ce,
criterion_dis=criterion_dis,
decoder_optimizer=decoder_optimizer,
epoch=epoch)
# One epoch's validation
recent_bleu4 = validate(val_loader=val_loader,
decoder=decoder,
criterion_ce=criterion_ce,
criterion_dis=criterion_dis)
# Check if there was an improvement
is_best = recent_bleu4 > best_bleu4
best_bleu4 = max(recent_bleu4, best_bleu4)
if not is_best:
epochs_since_improvement += 1
print("\nEpochs since last improvement: %d\n" % (epochs_since_improvement,))
else:
epochs_since_improvement = 0
# Save checkpoint
save_checkpoint(data_name, epoch, epochs_since_improvement, decoder,decoder_optimizer, recent_bleu4, is_best)
def main():
"""
Training and validation.
"""
global checkpoint, start_epoch, fine_tune_encoder
# Initialize / load checkpoint
if checkpoint is None:
encoder = Encoder()
print(encoder)
encoder.fine_tune(fine_tune_encoder)
encoder_optimizer = torch.optim.Adam(
params=filter(lambda p: p.requires_grad, encoder.parameters()),
lr=encoder_lr) if fine_tune_encoder else None
decoder = DecoderWithAttention(attention_dim=attention_dim,
embed_dim=emb_dim,
decoder_dim=decoder_dim,
vocab_size=vocab_size,
encoder_dim=encoder_dim,
dropout=dropout)
decoder_optimizer = torch.optim.Adam(params=filter(
lambda p: p.requires_grad, decoder.parameters()),
lr=decoder_lr)
else:
checkpoint = torch.load(checkpoint)
start_epoch = checkpoint['epoch'] + 1
decoder = checkpoint['decoder']
decoder_optimizer = checkpoint['decoder_optimizer']
encoder = checkpoint['encoder']
encoder_optimizer = checkpoint['encoder_optimizer']
if fine_tune_encoder is True and encoder_optimizer is None:
encoder.fine_tune(fine_tune_encoder)
encoder_optimizer = torch.optim.Adam(params=filter(
lambda p: p.requires_grad, encoder.parameters()),
lr=encoder_lr)
# Move to GPU, if available
decoder = decoder.to(device)
encoder = encoder.to(device)
# Loss function
criterion = nn.CrossEntropyLoss().to(device)
# customized dataloader
MyDataset = DualLoadDatasets(imgsz, txt_folder, img_folder, bin_folder,
split, Gfiltersz, Gblursigma)
#drop the last batch since it is not divisible by batchsize
train_loader = torch.utils.data.DataLoader(MyDataset,
batch_size=batch_size,
shuffle=True,
num_workers=workers,
pin_memory=True,
drop_last=True)
# val_loader = torch.utils.data.DataLoader(
# CaptionDataset(data_folder, data_name, 'VAL', transform=transforms.Compose([normalize])),
# batch_size=batch_size, shuffle=True, num_workers=workers, pin_memory=True)
# Save checkpoint
epoch = 0
save_checkpoint(epoch, encoder, decoder, encoder_optimizer,
decoder_optimizer)
print('saving models to models/checkpoint')
# Epochs
for epoch in range(start_epoch, epochs):
#print(image_transforms)
# One epoch's training
train(train_loader=train_loader,
encoder=encoder,
decoder=decoder,
transform=transform,
criterion=criterion,
encoder_optimizer=encoder_optimizer,
decoder_optimizer=decoder_optimizer,
epoch=epoch)
# Save checkpoint
save_checkpoint(epoch, encoder, decoder, encoder_optimizer,
decoder_optimizer)
print('saving models to models/checkpoint')
emb_dim = 512 # dimension of word embeddings
attention_dim = 512 # dimension of attention linear layers
decoder_dim = 512 # dimension of decoder RNN
dropout = 0.1
device = torch.device("cuda" if torch.cuda.is_available() else
"cpu") # sets device for model and PyTorch tensors
encoder_lr = 5e-4 # learning rate for encoder if fine-tuning
decoder_lr = 5e-4 # learning rate for decoder
grad_clip = 5. # clip gradients at an absolute value of
alpha_c = 1. # regularization parameter for 'doubly stochastic attention', as in the paper
fine_tune_encoder = True # fine-tune encoder?
decoder = DecoderWithAttention(attention_dim=attention_dim,
embed_dim=emb_dim,
decoder_dim=decoder_dim,
vocab_size=len(vocab),
dropout=dropout)
decoder.fine_tune_embeddings(True)
decoder_optimizer = torch.optim.Adam(params=filter(lambda p: p.requires_grad,
decoder.parameters()),
lr=decoder_lr)
encoder = Encoder()
encoder.fine_tune(fine_tune_encoder)
encoder_optimizer = torch.optim.Adam(
params=filter(lambda p: p.requires_grad, encoder.parameters()),
lr=encoder_lr) if fine_tune_encoder else None
decoder_sched = torch.optim.lr_scheduler.CosineAnnealingLR(decoder_optimizer,
5,
eta_min=1e-5,
def main():
"""
Training and validation.
"""
global epochs_since_improvement, checkpoint, start_epoch, fine_tune_encoder, data_name, word_map, role_map
#print('reading word map')
# Read word map
word_map_file = os.path.join(data_folder, 'token2id' + '.json')
with open(word_map_file, 'r') as j:
word_map = json.load(j)
#print('reading role map')
role_map_file = os.path.join(data_folder, 'roles2id' + '.json')
with open(role_map_file, 'r') as j:
role_map = json.load(j)
#print('initializing..')
# Initialize / load checkpoint
if checkpoint is None:
decoder = DecoderWithAttention(attention_dim=attention_dim,
embed_dim=emb_dim,
decoder_dim=decoder_dim,
vocab_size=len(word_map),
role_vocab_size=len(role_map),
role_embed_dim=role_dim,
dropout=dropout)
decoder_optimizer = torch.optim.Adam(params=filter(
lambda p: p.requires_grad, decoder.parameters()),
lr=decoder_lr)
encoder = Encoder()
encoder.fine_tune(fine_tune_encoder)
encoder_optimizer = torch.optim.Adam(
params=filter(lambda p: p.requires_grad, encoder.parameters()),
lr=encoder_lr) if fine_tune_encoder else None
else:
checkpoint = torch.load(checkpoint)
start_epoch = checkpoint['epoch'] + 1
epochs_since_improvement = checkpoint['epochs_since_improvement']
#best_bleu4 = checkpoint['bleu-4']
decoder = checkpoint['decoder']
decoder_optimizer = checkpoint['decoder_optimizer']
encoder = checkpoint['encoder']
encoder_optimizer = checkpoint['encoder_optimizer']
if fine_tune_encoder is True and encoder_optimizer is None:
encoder.fine_tune(fine_tune_encoder)
encoder_optimizer = torch.optim.Adam(params=filter(
lambda p: p.requires_grad, encoder.parameters()),
lr=encoder_lr)
# Move to GPU, if available
decoder = decoder.to(device)
encoder = encoder.to(device)
#print('creating encoder/decoder..')
#encoder = nn.DataParallel(encoder,device_ids=[0,1])
#decoder = nn.DataParallel(decoder,device_ids=[0,1])
# Loss function
criterion = nn.CrossEntropyLoss().to(device)
#print('creating dataloader..')
# Custom dataloaders
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
train_loader = torch.utils.data.DataLoader(FrameDataset(
data_folder, 'TRAIN', transform=transforms.Compose([normalize])),
batch_size=batch_size,
shuffle=True,
num_workers=workers,
pin_memory=True)
# val_loader = torch.utils.data.DataLoader(
# FrameDataset(data_folder, 'VAL', transform=transforms.Compose([normalize])),
# batch_size=batch_size, shuffle=True, num_workers=workers, pin_memory=True)
# Epochs
for epoch in range(start_epoch, epochs):
# decay learning rate somehow
# One epoch's training
#print('start training')
train(train_loader=train_loader,
encoder=encoder,
decoder=decoder,
criterion=criterion,
encoder_optimizer=encoder_optimizer,
decoder_optimizer=decoder_optimizer,
epoch=epoch)
print('start validation..')
