def load_model():
opt = parse_args()
# Load infos
infos = load_infos(opt)
ignore = ["id", "batch_size", "beam_size", "start_from_best"]
for k in vars(infos['opt']).keys():
if k not in ignore:
if k in vars(opt):
assert vars(opt)[k] == vars(
infos['opt'])[k], k + ' option not consistent'
else:
vars(opt).update({k: vars(infos['opt'])[k]
}) # copy over options from model
print(opt)
# Setup the model
model_cnn = models.setup_cnn(opt)
model_cnn.cuda()
model = models.setup(opt)
model.cuda()
# Make sure in the evaluation mode
model_cnn.eval()
model.eval()
ix_to_word = infos['vocab']
return model_cnn, model, ix_to_word, opt
def main():
opt = parse_args()
# make dirs
print(opt.output_dir)
if not os.path.isdir(opt.output_dir):
os.makedirs(opt.output_dir)
# Load infos
infos = load_infos(opt)
ignore = ["id", "batch_size", "beam_size", "start_from_best"]
for k in vars(infos['opt']).keys():
if k not in ignore:
if k in vars(opt):
assert vars(opt)[k] == vars(
infos['opt'])[k], k + ' option not consistent'
else:
vars(opt).update({k: vars(infos['opt'])[k]
}) # copy over options from model
print(opt)
# Setup the model
model_cnn = models.setup_cnn(opt)
model_cnn.cuda()
model = models.setup(opt)
model.cuda()
# Make sure in the evaluation mode
model_cnn.eval()
model.eval()
str_id = ''.join(opt.id.split('_'))
path_zip = opt.output_dir + '/results.zip'
# zipf = zipfile.ZipFile(path_zip, 'w', zipfile.ZIP_DEFLATED)
for dataset in opt.datasets:
loader = DataLoaderRaw({
'folder_path':
os.path.join(opt.image_folder, dataset),
'batch_size':
opt.batch_size
})
loader.ix_to_word = infos['vocab']
# Set sample options
predictions = eval_split(model_cnn, model, loader, vars(opt))
path_json = opt.output_dir + '/captions_' + dataset + '_' + str_id + '_results.json'
json.dump(predictions, open(path_json, 'w'))
def main():
opt = parse_args()
# make dirs
print(opt.output_dir)
if not os.path.isdir(opt.output_dir):
os.makedirs(opt.output_dir)
# Load infos
infos = load_infos(opt)
ignore = [
"id", "batch_size", "beam_size", "start_from_best",
"checkpoint_best_path"
]
for k in vars(infos['opt']).keys():
if k not in ignore:
if k in vars(opt):
assert vars(opt)[k] == vars(
infos['opt'])[k], k + ' option not consistent'
else:
vars(opt).update({k: vars(infos['opt'])[k]
}) # copy over options from model
print(opt)
# Setup the model
model_cnn = models.setup_cnn(opt)
model_cnn.cuda()
model = models.setup(opt)
model.cuda()
# Make sure in the evaluation mode
model_cnn.eval()
model.eval()
save_model_best(model, model_cnn, infos, opt)
loader = DataLoaderRaw({
'folder_path': opt.image_folder,
'batch_size': opt.batch_size
})
loader.ix_to_word = infos['vocab']
# Set sample options
predictions = eval_split(model_cnn, model, loader, vars(opt))
json.dump(predictions, open(opt.output_dir + '/result.json', 'w'))
def train(opt):
loader = DataLoader(opt)
opt.vocab_size = loader.vocab_size
opt.seq_length = loader.seq_length
opt.caption_model = "SCST"
infos = {}
if opt.start_from is not None and len(opt.start_from) > 0:
print("start from %s" % (opt.start_from))
# open old infos and check if models are compatible
with open(os.path.join(opt.start_from,
'infos_' + opt.id + '.pkl')) as f:
infos = cPickle.load(f)
saved_model_opt = infos['opt']
need_be_same = ["caption_model", "rnn_size", "num_layers"]
for checkme in need_be_same:
assert vars(saved_model_opt)[checkme] == vars(
opt
)[checkme], "Command line argument and saved model disagree on '%s' " % checkme
iteration = infos.get('iter', 0)
epoch = infos.get('epoch', 0)
val_result_history = infos.get('val_result_history', {})
loss_history = infos.get('loss_history', {})
lr_history = infos.get('lr_history', {})
loader.iterators = infos.get('iterators', loader.iterators)
if opt.load_best_score == 1:
best_val_score = infos.get('best_val_score', None)
model_cnn = models.setup_cnn(opt)
model_cnn.cuda()
model = models.setup(opt)
model.cuda()
model_cnn.train()
model.train()
fc_expander = utils.FeatExpander(5)
att_expander = utils.FeatExpander(5)
crit = Criterion.LanguageModelCriterion()
crit_reinforce = Criterion.SCSTCriterion()
optimizer = optim.Adam(model.parameters(), lr=opt.learning_rate)
optimizer_cnn = optim.Adam(model_cnn.parameters(),
lr=opt.cnn_learning_rate,
weight_decay=opt.cnn_weight_decay)
# Load the optimizer
if opt.start_from is not None and len(opt.start_from) > 0:
optimizer.load_state_dict(
torch.load(os.path.join(opt.start_from, 'optimizer.pth')))
optimizer_cnn.load_state_dict(
torch.load(os.path.join(opt.start_from, 'optimizer_cnn.pth')))
finetune_cnn_start = False
use_reinforce = False
update_lr_flag = True
while True:
if update_lr_flag:
# Assign the learning rate
if opt.learning_rate_decay_start >= 0 and epoch >= opt.learning_rate_decay_start:
frac = (epoch - opt.learning_rate_decay_start
) // opt.learning_rate_decay_every
decay_factor = opt.learning_rate_decay_rate**frac
opt.current_lr = opt.learning_rate * decay_factor
utils.set_lr(optimizer, opt.current_lr) # set the decayed rate
else:
opt.current_lr = opt.learning_rate
if opt.finetune_cnn_after >= 0 and epoch >= opt.finetune_cnn_after:
for p in model_cnn.parameters():
p.requires_grad = True
model_cnn.train()
finetune_cnn_start = True
else:
for p in model_cnn.parameters():
p.requires_grad = False
model_cnn.eval()
finetune_cnn_start = False
if opt.cnn_learning_rate_decay_start >= 0 and epoch >= opt.cnn_learning_rate_decay_start:
frac = (epoch - opt.cnn_learning_rate_decay_start
) // opt.cnn_learning_rate_decay_every
decay_factor = opt.cnn_learning_rate_decay_rate**frac
opt.current_cnn_lr = opt.cnn_learning_rate * decay_factor
utils.set_lr(optimizer_cnn,
opt.current_cnn_lr) # set the decayed rate
else:
opt.current_cnn_lr = opt.cnn_learning_rate
update_lr_flag = False
start_total = time.time()
data = loader.get_batch('train')
vocab = loader.get_vocab()
images = torch.from_numpy(data['images']).cuda()
images = utils.prepro(images, False)
images = Variable(images, requires_grad=False)
labels = torch.from_numpy(data['labels']).cuda()
labels = Variable(labels, requires_grad=False)
fc_feats, att_feats = model_cnn(images)
fc_feats_ext = fc_expander(fc_feats)
att_feats_ext = att_expander(att_feats)
optimizer.zero_grad()
if opt.finetune_cnn_after >= 0 and epoch >= opt.finetune_cnn_after:
optimizer_cnn.zero_grad()
reward = 0
reward1 = 0
if opt.reinforce_start >= 0 and epoch >= opt.reinforce_start:
use_reinforce = True
output1, seq1 = model(fc_feats_ext, att_feats_ext, labels, "test")
output, seq = model(fc_feats_ext, att_feats_ext, labels, "train")
loss, reward, reward1 = crit_reinforce(output, output1, seq, seq1,
labels, vocab)
loss.backward()
else:
use_reinforce = False
output, _ = model(fc_feats_ext, att_feats_ext, labels, "xent")
loss = crit(output, labels)
loss.backward()
utils.clip_gradient(optimizer, opt.grad_clip)
optimizer.step()
if opt.finetune_cnn_after >= 0 and epoch >= opt.finetune_cnn_after:
utils.clip_gradient(optimizer_cnn, opt.grad_clip)
optimizer_cnn.step()
train_loss = loss.data[0]
print("iter {} (epoch {}), train_loss = {:.3f}, lr = {} lr_cnn = {} finetune_cnn = {} use_reinforce = {} reward = {} reward1 = {} time/batch = {:.3f}" \
.format(iteration, epoch, train_loss, opt.current_lr, opt.current_cnn_lr, finetune_cnn_start, use_reinforce, reward, reward1, time.time() - start_total))
# Update the iteration and epoch
iteration += 1
if data['bounds']['wrapped']:
epoch += 1
update_lr_flag = True
# Write the training loss summary
if (iteration % opt.losses_log_every == 0):
loss_history[iteration] = train_loss
lr_history[iteration] = opt.current_lr
# make evaluation on validation set, and save model
if (iteration % opt.save_checkpoint_every == 0):
# eval model
eval_kwargs = {
'split': 'val',
'dataset': opt.input_json,
'caption_model': 'SCST'
}
eval_kwargs.update(vars(opt))
val_loss, predictions, lang_stats, str_stats = eval_utils.eval_split(
model_cnn, model, crit, loader, eval_kwargs)
if not os.path.exists(opt.eval_result_path):
os.makedirs(opt.eval_result_path)
eval_result_file = os.path.join(opt.eval_result_path,
opt.id + ".csv")
with open(eval_result_file, 'a') as f:
f.write(str_stats + "\n")
predictions_file = os.path.join(opt.eval_result_path,
opt.id + ".json")
with open(predictions_file, 'w') as f:
json.dump(predictions, f)
val_result_history[iteration] = {
'loss': val_loss,
'lang_stats': lang_stats,
'predictions': predictions
}
# Save model if is improving on validation result
if opt.language_eval == 1:
current_score = lang_stats['CIDEr']
else:
current_score = -val_loss
best_flag = False
if True: # if true
if best_val_score is None or current_score > best_val_score:
best_val_score = current_score
best_flag = True
if not os.path.exists(opt.checkpoint_path):
os.makedirs(opt.checkpoint_path)
checkpoint_path = os.path.join(opt.checkpoint_path,
'model.pth')
torch.save(model.state_dict(), checkpoint_path)
print("model saved to {}".format(checkpoint_path))
checkpoint_path_cnn = os.path.join(opt.checkpoint_path,
'model_cnn.pth')
torch.save(model_cnn.state_dict(), checkpoint_path_cnn)
print("model cnn saved to {}".format(checkpoint_path_cnn))
optimizer_path = os.path.join(opt.checkpoint_path,
'optimizer.pth')
torch.save(optimizer.state_dict(), optimizer_path)
print("optimizer saved to {}".format(optimizer_path))
optimizer_path_cnn = os.path.join(opt.checkpoint_path,
'optimizer_cnn.pth')
torch.save(optimizer_cnn.state_dict(), optimizer_path_cnn)
print("optimizer cnn saved to {}".format(optimizer_path_cnn))
infos['iter'] = iteration
infos['epoch'] = epoch
infos['iterators'] = loader.iterators
infos['best_val_score'] = best_val_score
infos['opt'] = opt
infos['val_result_history'] = val_result_history
infos['loss_history'] = loss_history
infos['lr_history'] = lr_history
infos['vocab'] = loader.get_vocab()
info_path = os.path.join(opt.checkpoint_path,
'infos_' + opt.id + '.pkl')
with open(info_path, 'wb') as f:
cPickle.dump(infos, f)
if best_flag:
checkpoint_path = os.path.join(opt.checkpoint_path,
'model_best.pth')
torch.save(model.state_dict(), checkpoint_path)
print("model saved to {}".format(checkpoint_path))
checkpoint_path_cnn = os.path.join(opt.checkpoint_path,
'model_cnn_best.pth')
torch.save(model_cnn.state_dict(), checkpoint_path_cnn)
print("model cnn saved to {}".format(checkpoint_path_cnn))
info_path = os.path.join(opt.checkpoint_path,
'infos_' + opt.id + '_best.pkl')
with open(info_path, 'wb') as f:
cPickle.dump(infos, f)
if epoch >= opt.max_epochs and opt.max_epochs != -1:
break
def main():
opt = parse_args()
# make dirs
print(opt.eval_result_path)
if not os.path.isdir(opt.eval_result_path):
os.makedirs(opt.eval_result_path)
# Load infos
infos = load_infos(opt)
ignore = [
"id", "input_json", "input_h5", "input_anno", "images_root",
"coco_caption_path", "batch_size", "beam_size", "start_from_best",
"eval_result_path"
]
for k in vars(infos['opt']).keys():
if k not in ignore:
if k in vars(opt):
assert vars(opt)[k] == vars(
infos['opt'])[k], k + ' option not consistent'
else:
vars(opt).update({k: vars(infos['opt'])[k]
}) # copy over options from model
# print(opt)
# Setup the model
model_cnn = models.setup_cnn(opt)
model_cnn.cuda()
model = models.setup(opt)
model.cuda()
# Make sure in the evaluation mode
model_cnn.eval()
model.eval()
if models.has_bu(opt.caption_model) or \
models.has_sub_regions(opt.caption_model) or \
models.has_sub_region_bu(opt.caption_model):
loader = DataLoaderThreadBu(opt)
print("DataLoaderThreadBu")
else:
loader = DataLoaderThreadNew(opt)
print("DataLoaderThreadNew")
loader.ix_to_word = infos['vocab']
eval_kwargs = {'split': opt.val_split, 'dataset': opt.input_json}
eval_kwargs.update(vars(opt))
start_beam = 0
total_beam = 20
for beam in range(start_beam, total_beam):
opt.beam_size = beam + 1
eval_kwargs.update(vars(opt))
print("beam_size: " + str(opt.beam_size))
print("start eval ...")
crit = None
val_loss, predictions, lang_stats, str_stats = eval_utils.eval_split(
model_cnn, model, crit, loader, eval_kwargs)
print("end eval ...")
msg = "str_stats = {}".format(str_stats)
print(msg)
save_result(str(opt.beam_size) + "," + str_stats, predictions, opt)
def main():
opt = parse_args()
opt.datasets = opt.datasets.split(',')
opt.ids = opt.ids.split(',')
# make dirs
print(opt.output_dir)
if not os.path.isdir(opt.output_dir):
os.makedirs(opt.output_dir)
print(opt.output_beam_dir)
if not os.path.isdir(opt.output_beam_dir):
os.makedirs(opt.output_beam_dir)
# print(opt)
all_model_cnns = []
all_models = []
for i in range(len(opt.ids)):
# id
opt.id = opt.ids[i]
# Load infos
infos = load_infos(opt)
ignore = ["id", "batch_size", "beam_size", "start_from_best", "input_json",
"input_h5", "input_anno", "images_root", "aic_caption_path", "input_bu"]
for k in vars(infos['opt']).keys():
if k not in ignore:
vars(opt).update({k: vars(infos['opt'])[k]})
opt.relu_type = 0
# Setup the model
model_cnn = models.setup_cnn(opt)
# model_cnn.cuda()
model_cnn = nn.DataParallel(model_cnn.cuda())
model = models.setup(opt)
model.cuda()
# Make sure in the evaluation mode
model_cnn.eval()
model.eval()
all_model_cnns.append(model_cnn)
all_models.append(model)
if opt.eval_type == 0: # local test
print('eval local')
if models.has_bu(opt.caption_model):
loader = DataLoaderThreadBu(opt)
else:
loader = DataLoaderThreadNew(opt)
# Set sample options
predictions, lang_stats, str_stats, beam_vis = eval_split(all_model_cnns, all_models, loader, opt, vars(opt))
save_result(opt.output_dir, str_stats, predictions)
save_beam_vis_result(opt.output_beam_dir, "eval_beam_vis.json", beam_vis)
elif opt.eval_type == 1: # server
print('eval server')
for dataset in opt.datasets:
print(os.path.join(opt.image_folder, dataset))
loader = DataLoaderRaw({'folder_path': os.path.join(opt.image_folder, dataset),
'batch_size': opt.batch_size,
'start': opt.start,
'num': opt.num,
'use_bu_att': opt.use_bu_att,
'input_bu': opt.input_bu,
'bu_size': opt.bu_size,
'bu_feat_size': opt.bu_feat_size})
loader.ix_to_word = infos['vocab']
# Set sample options
predictions, lang_stats, str_stats, beam_vis = eval_split(all_model_cnns, all_models, loader, opt, vars(opt))
path_json = opt.output_dir + '/captions_' + dataset + str(opt.start) + '_ensemble_results.json'
json.dump(predictions, open(path_json, 'w'))
save_beam_vis_result(opt.output_beam_dir, dataset + str(opt.start) + "_beam_size_" + str(opt.beam_size) + "_beam_type_" + str(opt.beam_type) + "_eval_beam_vis.json", beam_vis)
def train(opt):
notifier = notify()
notifier.login()
# init path
if not os.path.exists(opt.eval_result_path):
os.makedirs(opt.eval_result_path)
config_file = os.path.join(opt.eval_result_path, opt.id + '_config.txt')
with open(config_file, 'w') as f:
f.write("{}\n".format(json.dumps(vars(opt), sort_keys=True, indent=2)))
torch.backends.cudnn.benchmark = True
if opt.use_tensorboard:
if opt.tensorboard_type == 0:
board = tensorboard.TensorBoard()
board.start(opt.id, opt.tensorboard_ip, opt.tensorboard_port)
else:
board = trans_client.TransClient()
board.start(opt.id)
print(opt.cnn_model)
loader = get_loader()
opt.vocab_size = loader.vocab_size
opt.seq_length = loader.seq_length
vocab = loader.get_vocab()
opt.vocab = vocab
batch_size = loader.batch_size
infos = get_infos()
infos['vocab'] = vocab
iteration = infos.get('iter', 0)
epoch = infos.get('epoch', 0)
val_result_history = infos.get('val_result_history', {})
loss_history = infos.get('loss_history', {})
lr_history = infos.get('lr_history', {})
finetune_cnn_history = infos.get('finetune_cnn_history', {})
loader.iterators = infos.get('iterators', loader.iterators)
if opt.load_best_score == 1:
best_val_score = infos.get('best_val_score', None)
else:
best_val_score = None
model_cnn = models.setup_cnn(opt)
model_cnn = model_cnn.cuda()
model_cnn = nn.DataParallel(model_cnn)
model = models.setup(opt)
model = model.cuda()
# if models.is_transformer(opt.caption_model) or models.is_ctransformer(opt.caption_model):
# model = nn.DataParallel(model)
train_utils.save_model_conf(model_cnn, model, opt)
update_lr_flag = True
model_cnn.train()
model.train()
fc_expander, att_expander, bu_expander = get_expander()
optimizer = None
optimizer_cnn = None
finetune_cnn_start = False
early_stop_cnt = 0
params = {}
params['model'] = model
params['vocab'] = vocab
# crit_pg, crit_rl, crit_ctc, crit_c, crit_ac, crit
params['crit_pg'] = None
params['crit_rl'] = None
params['crit_ctc'] = None
params['crit_c'] = None
params['crit_ac'] = None
params['crit'] = None
is_eval_start = opt.is_eval_start
if opt.use_auto_learning_rate == 1:
train_process = train_utils.init_train_process()
train_process_index = infos.get('train_process_index', 0)
train_step = train_process[train_process_index]
optimizer_cnn = None
optimizer = None
opt.learning_rate = train_step.learning_rate
opt.cnn_learning_rate = train_step.cnn_learning_rate
opt.finetune_cnn_after = train_step.finetune_cnn_after
while True:
current_score = None
# make evaluation on validation set, and save model
if (iteration > 0 and iteration % opt.save_checkpoint_every == 0 and
not val_result_history.has_key(iteration)) or is_eval_start:
predictions, best_val_score, best_flag, current_score = eval_model(
model_cnn, model, params, loader, board, iteration, notifier,
val_result_history, best_val_score)
infos['best_val_score'] = best_val_score
infos['val_result_history'] = val_result_history
train_utils.save_infos(infos, opt)
if best_flag:
train_utils.save_best_result(predictions, opt)
train_utils.save_model_best(model, model_cnn, infos, opt)
early_stop_cnt = 0
else:
early_stop_cnt += 1
is_eval_start = False
if epoch >= opt.max_epochs and opt.max_epochs != -1:
msg = "max epoch"
logger.info(msg)
break
# auto update model
if opt.use_auto_learning_rate == 1 and current_score is not None:
if early_stop_cnt > opt.auto_early_stop_cnt or current_score < opt.auto_early_stop_score:
early_stop_cnt = 0
train_process_index += 1
msg = opt.id + " early stop " + str(train_process_index)
logger.info(msg)
infos['train_process_index'] = train_process_index
train_utils.save_infos(infos, opt)
if train_process_index >= len(train_process):
notifier.send(opt.id + " early stop", msg)
logger.info("break")
break
train_step = train_process[train_process_index]
optimizer_cnn = None
optimizer = None
opt.learning_rate = train_step.learning_rate
opt.cnn_learning_rate = train_step.cnn_learning_rate
opt.finetune_cnn_after = train_step.finetune_cnn_after
opt.start_from_best = opt.auto_start_from_best
# model_cnn_path = os.path.join(opt.auto_start_from_best, opt.id + '_model_cnn_best.pth')
# model_cnn.load_state_dict(torch.load(model_cnn_path))
# model_cnn = model_cnn.cuda()
# model_cnn = nn.DataParallel(model_cnn)
#
# model_path = os.path.join(opt.auto_start_from_best, opt.id + '_model_best.pth')
# model.load_state_dict(torch.load(model_path))
# model = model.cuda()
del model_cnn
del model
torch.cuda.empty_cache()
model_cnn = models.setup_cnn(opt)
model_cnn = model_cnn.cuda()
model_cnn = nn.DataParallel(model_cnn)
model = models.setup(opt)
model = model.cuda()
model_cnn.train()
model.train()
update_lr_flag = True
# start train
# Update the iteration and epoch
iteration += 1
if update_lr_flag:
if opt.finetune_cnn_after >= 0 and epoch >= opt.finetune_cnn_after:
finetune_cnn_start = True
else:
finetune_cnn_start = False
optimizer_cnn = train_utils.get_cnn_optimizer(
model_cnn, optimizer_cnn, finetune_cnn_start, opt)
train_utils.update_lr(epoch, optimizer, optimizer_cnn,
finetune_cnn_start, opt)
update_lr_flag = False
if opt.reinforce_start >= 0 and epoch >= opt.reinforce_start:
use_reinforce = True
else:
use_reinforce = False
optimizer = get_optimizer(optimizer, epoch, model, model_cnn)
start_total = time.time()
start = time.time()
optimizer.zero_grad()
if finetune_cnn_start:
optimizer_cnn.zero_grad()
# batch data
data = loader.get_batch('train', batch_size)
images = data['images']
bus = None
if models.has_bu(opt.caption_model):
bus = data['bus']
if opt.verbose:
print('data {:.3f}'.format(time.time() - start))
start = time.time()
fc_feats, att_feats, bu_feats = train_cnn(model_cnn, images, bus,
fc_expander, att_expander,
bu_expander, use_reinforce)
if opt.verbose:
print('model_cnn {:.3f}'.format(time.time() - start))
# get input data
params['fc_feats'] = fc_feats
params['att_feats'] = att_feats
params['bu_feats'] = bu_feats
# get target data
params['labels'] = data['labels']
params['masks'] = data['masks']
params['tokens'] = data['tokens']
params['gts'] = data['gts']
params['targets'] = data['targets']
# crit_pg, crit_rl, crit_ctc, crit_c, crit_ac, crit,
train_loss, reward_mean, use_reinforce = train_model(
params, iteration, epoch, board)
# update the gradient
update_gradient(optimizer, optimizer_cnn, finetune_cnn_start)
time_batch = time.time() - start_total
left_time = (opt.save_checkpoint_every -
iteration % opt.save_checkpoint_every) * time_batch
s_left_time = utils.format_time(left_time)
msg = "id {} iter {} (epoch {}), train_loss = {:.3f}, lr = {} lr_cnn = {} f_cnn = {} rf = {} r = {:.3f} early_stop_cnt = {} time/batch = {:.3f}s time/eval = {}" \
.format(opt.id, iteration, epoch, train_loss, opt.current_lr, opt.current_cnn_lr, finetune_cnn_start,
use_reinforce, reward_mean, early_stop_cnt, time_batch, s_left_time)
logger.info(msg)
if opt.use_tensorboard:
if iteration % opt.tensorboard_for_train_every == 0:
board.loss_train(train_loss, iteration)
if data['bounds']['wrapped']:
epoch += 1
update_lr_flag = True
# Write the training loss summary
if iteration % opt.losses_log_every == 0:
loss_history[iteration] = train_loss
lr_history[iteration] = opt.current_lr
finetune_cnn_history[iteration] = finetune_cnn_start
# update infos
infos['iter'] = iteration
infos['epoch'] = epoch
infos['iterators'] = loader.iterators
infos['best_val_score'] = best_val_score
infos['opt'] = opt
infos['val_result_history'] = val_result_history
infos['loss_history'] = loss_history
infos['lr_history'] = lr_history
infos['finetune_cnn_history'] = finetune_cnn_history
if opt.use_auto_learning_rate == 1:
infos['train_process_index'] = train_process_index
if opt.save_snapshot_every > 0 and iteration % opt.save_snapshot_every == 0:
train_utils.save_model(model, model_cnn, infos, opt)
loader.terminate()
def train(opt):
notifier = notify()
notifier.login()
# init path
if not os.path.exists(opt.eval_result_path):
os.makedirs(opt.eval_result_path)
config_file = os.path.join(opt.eval_result_path, opt.id + '_config.txt')
with open(config_file, 'w') as f:
f.write("{}\n".format(json.dumps(vars(opt), sort_keys=True, indent=2)))
torch.backends.cudnn.benchmark = True
if opt.use_tensorboard:
board = tensorboard.TensorBoard()
board.start(opt.id, opt.tensorboard_ip)
# board = trans_client.TransClient()
# board.start(opt.id)
print(opt.cnn_model)
loader = DataLoaderThreadBu(opt)
opt.vocab_size = loader.vocab_size
opt.seq_length = loader.seq_length
vocab = loader.get_vocab()
batch_size = loader.batch_size
try:
if opt.is_load_infos == 1:
infos = train_utils.load_infos(opt)
else:
infos = {}
except:
infos = {}
print('load infos error')
iteration = infos.get('iter', 0)
epoch = infos.get('epoch', 0)
val_result_history = infos.get('val_result_history', {})
loss_history = infos.get('loss_history', {})
lr_history = infos.get('lr_history', {})
loader.iterators = infos.get('iterators', loader.iterators)
if opt.load_best_score == 1:
best_val_score = infos.get('best_val_score', None)
else:
best_val_score = None
model_cnn = models.setup_cnn(opt)
model_cnn = nn.DataParallel(model_cnn.cuda())
model = models.setup(opt)
model.cuda()
train_utils.save_model_conf(model_cnn, model, opt)
update_lr_flag = True
model_cnn.train()
model.train()
if opt.seq_per_img > 1:
fc_expander = utils.FeatExpander(opt.seq_per_img)
att_expander = utils.FeatExpander(opt.seq_per_img)
bu_expander = utils.FeatExpander(opt.seq_per_img)
# crit = Criterion.LanguageModelWeightNewCriterion()
crit = Criterion.LanguageModelWithProbWeightCriterion(
opt.prob_weight_alpha)
crit_rl = None
# print(model_cnn)
optimizer = optim.Adam(model.parameters(),
lr=opt.learning_rate,
betas=(opt.optim_alpha, opt.optim_beta),
eps=opt.optim_epsilon)
optimizer_cnn = None
finetune_cnn_start = False
early_stop_cnt = 0
params = {}
params['model'] = model
params['crit'] = crit
params['vocab'] = vocab
while True:
# try:
if update_lr_flag:
if opt.finetune_cnn_after >= 0 and epoch >= opt.finetune_cnn_after:
finetune_cnn_start = True
else:
finetune_cnn_start = False
optimizer_cnn = train_utils.finetune_cnn(model_cnn, optimizer_cnn,
finetune_cnn_start, opt)
train_utils.update_lr(epoch, optimizer, optimizer_cnn,
finetune_cnn_start, opt)
update_lr_flag = False
start_total = time.time()
start = time.time()
optimizer.zero_grad()
if finetune_cnn_start:
optimizer_cnn.zero_grad()
# batch data
data = loader.get_batch('train', batch_size)
images = data['images']
labels = data['labels']
masks = data['masks']
tokens = data['tokens']
gts = data['gts']
if opt.verbose:
print('data {:.3f}'.format(time.time() - start))
# train cnn
fc_feats, att_feats, bu_feats = model_cnn(images)
if opt.seq_per_img > 1:
fc_feats = fc_expander(fc_feats)
att_feats = att_expander(att_feats)
bu_feats = bu_expander(bu_feats)
params['fc_feats'] = fc_feats
params['att_feats'] = att_feats
params['bu_feats'] = bu_feats
params['labels'] = labels
params['masks'] = masks
params['tokens'] = tokens
params['gts'] = gts
if opt.reinforce_start >= 0 and epoch >= opt.reinforce_start:
use_reinforce = True
if crit_rl is None:
if opt.is_aic_data:
crit_rl = Criterion.RewardCriterionAIC(opt, vocab)
else:
crit_rl = Criterion.RewardCriterion(opt)
params['crit_rl'] = crit_rl
train_loss, reward_mean, sample_mean, greedy_mean = train_utils.train_reinforce(
params, opt)
if opt.use_tensorboard:
board.val("sample_mean", sample_mean, iteration)
board.val("greedy_mean", greedy_mean, iteration)
else:
use_reinforce = False
params['crit'] = crit
train_loss, reward_mean = train_utils.train_with_prob_weight(
params, opt)
# update the gradient
utils.clip_gradient(optimizer, opt.grad_clip)
optimizer.step()
if finetune_cnn_start:
utils.clip_gradient(optimizer_cnn, opt.grad_clip)
optimizer_cnn.step()
msg = "iter {} (epoch {}), train_loss = {:.3f}, lr = {} lr_cnn = {} f_cnn = {} rf = {} r = {:.3f} time/batch = {:.3f}" \
.format(iteration, epoch, train_loss, opt.current_lr, opt.current_cnn_lr, finetune_cnn_start,
use_reinforce, reward_mean, time.time() - start_total)
logger.info(msg)
if opt.use_tensorboard:
board.loss_train(train_loss, iteration)
# Update the iteration and epoch
if not opt.is_eval_start:
iteration += 1
if data['bounds']['wrapped']:
epoch += 1
update_lr_flag = True
# Write the training loss summary
if (iteration % opt.losses_log_every == 0):
loss_history[iteration] = train_loss
lr_history[iteration] = opt.current_lr
# make evaluation on validation set, and save model
if (iteration % opt.save_checkpoint_every == 0):
if opt.is_every_eval:
# eval model
eval_kwargs = {
'split': opt.val_split,
'dataset': opt.input_json
}
eval_kwargs.update(vars(opt))
print("start eval ...")
val_loss, predictions, lang_stats, str_stats = eval_utils.eval_split_with_region_bu(
model_cnn, model, crit, loader, eval_kwargs)
if opt.use_tensorboard:
board.accuracy(lang_stats, iteration)
board.loss_val(val_loss, iteration)
print("end eval ...")
msg = "iteration = {} val_loss = {} str_stats = {}".format(
iteration, val_loss, str_stats)
notifier.send(opt.id + " val result", opt.id + " :\n" + msg)
logger.info(msg)
train_utils.save_result(str_stats + ',' + str(val_loss),
predictions, opt)
val_result_history[iteration] = {
'loss': val_loss,
'lang_stats': lang_stats,
'predictions': predictions
}
# Save model if is improving on validation result
if opt.language_eval == 1:
eval_metric = opt.eval_metric
current_score = lang_stats[eval_metric]
else:
current_score = -val_loss
best_flag = False
if best_val_score is None or current_score > best_val_score:
best_val_score = current_score
best_flag = True
else:
best_flag = True
infos['iter'] = iteration
infos['epoch'] = epoch
infos['iterators'] = loader.iterators
infos['best_val_score'] = best_val_score
infos['opt'] = opt
infos['val_result_history'] = val_result_history
infos['loss_history'] = loss_history
infos['lr_history'] = lr_history
infos['vocab'] = loader.get_vocab()
train_utils.save_model(model, model_cnn, infos, opt)
if best_flag:
train_utils.save_model_best(model, model_cnn, infos, opt)
early_stop_cnt = 0
else:
early_stop_cnt += 1
if epoch >= opt.max_epochs and opt.max_epochs != -1:
break
if opt.is_eval_start:
iteration += 1
loader.terminate()