def optimizer_step(self, epoch, batch_idx, optimizer, optimizer_idx, *args,
**kwargs):
# warm up lr
opt = self.opt
iteration = self.trainer.global_step
if opt.use_warmup and (iteration < opt.noamopt_warmup):
opt.current_lr = opt.learning_rate * \
(iteration+1) / opt.noamopt_warmup
utils.set_lr(optimizer, opt.current_lr)
super().optimizer_step(epoch, batch_idx, optimizer, optimizer_idx,
*args, **kwargs)
def on_epoch_start(self, trainer, pl_module):
# Update lr/training stage/scheduled sampling prob etc.
opt = pl_module.opt
model = pl_module.model
epoch = trainer.current_epoch
optimizer = trainer.optimizers[0]
if not opt.noamopt and not opt.reduce_on_plateau:
# Assign the learning rate
if epoch > opt.learning_rate_decay_start and opt.learning_rate_decay_start >= 0:
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
else:
opt.current_lr = opt.learning_rate
utils.set_lr(optimizer, opt.current_lr) # set the decayed rate
# Assign the scheduled sampling prob
if epoch > opt.scheduled_sampling_start and opt.scheduled_sampling_start >= 0:
frac = (epoch - opt.scheduled_sampling_start
) // opt.scheduled_sampling_increase_every
opt.ss_prob = min(opt.scheduled_sampling_increase_prob * frac,
opt.scheduled_sampling_max_prob)
model.ss_prob = opt.ss_prob
# If start self critical training
if opt.self_critical_after != -1 and epoch >= opt.self_critical_after:
sc_flag = True
init_scorer(opt.cached_tokens)
else:
sc_flag = False
# If start structure loss training
if opt.structure_after != -1 and epoch >= opt.structure_after:
struc_flag = True
init_scorer(opt.cached_tokens)
else:
struc_flag = False
# drop_worst flag
if opt.drop_worst_after != -1 and epoch >= opt.drop_worst_after:
drop_worst_flag = True
else:
drop_worst_flag = False
pl_module.struc_flag = struc_flag
pl_module.sc_flag = sc_flag
pl_module.drop_worst_flag = drop_worst_flag
def train(opt):
################################
# Build dataloader
################################
loader = DataLoader(opt)
opt.vocab_size = loader.vocab_size
opt.seq_length = loader.seq_length
##########################
# Initialize infos
##########################
infos = {
'iter': 0,
'epoch': 0,
'loader_state_dict': None,
'vocab': loader.get_vocab(),
}
# Load old infos(if there is) and check if models are compatible
if opt.start_from is not None and os.path.isfile(
os.path.join(opt.start_from, 'infos_' + opt.id + '.pkl')):
with open(os.path.join(opt.start_from, 'infos_' + opt.id + '.pkl'),
'rb') as f:
infos = utils.pickle_load(f)
saved_model_opt = infos['opt']
need_be_same = [
"caption_model", "rnn_type", "rnn_size", "num_layers"
]
for checkme in need_be_same:
assert getattr(saved_model_opt, checkme) == getattr(
opt, checkme
), "Command line argument and saved model disagree on '%s' " % checkme
infos['opt'] = opt
#########################
# Build logger
#########################
# naive dict logger
histories = defaultdict(dict)
if opt.start_from is not None and os.path.isfile(
os.path.join(opt.start_from, 'histories_' + opt.id + '.pkl')):
with open(os.path.join(opt.start_from, 'histories_' + opt.id + '.pkl'),
'rb') as f:
histories.update(utils.pickle_load(f))
# tensorboard logger
tb_summary_writer = SummaryWriter(opt.checkpoint_path)
##########################
# Build model
##########################
opt.vocab = loader.get_vocab()
model = models.setup(opt).cuda()
del opt.vocab
# Load pretrained weights:
if opt.start_from is not None and os.path.isfile(
os.path.join(opt.start_from, 'model.pth')):
model.load_state_dict(
torch.load(os.path.join(opt.start_from, 'model.pth')))
# Wrap generation model with loss function(used for training)
# This allows loss function computed separately on each machine
lw_model = LossWrapper(model, opt)
# Wrap with dataparallel
dp_model = torch.nn.DataParallel(model)
dp_model.vocab = getattr(model, 'vocab', None) # nasty
dp_lw_model = torch.nn.DataParallel(lw_model)
##########################
# Build optimizer
##########################
if opt.noamopt:
assert opt.caption_model in [
'transformer', 'bert', 'm2transformer'
], 'noamopt can only work with transformer'
optimizer = utils.get_std_opt(model,
optim_func=opt.optim,
factor=opt.noamopt_factor,
warmup=opt.noamopt_warmup)
elif opt.reduce_on_plateau:
optimizer = utils.build_optimizer(model.parameters(), opt)
optimizer = utils.ReduceLROnPlateau(
optimizer,
factor=opt.reduce_on_plateau_factor,
patience=opt.reduce_on_plateau_patience)
else:
optimizer = utils.build_optimizer(model.parameters(), opt)
# Load the optimizer
if opt.start_from is not None and os.path.isfile(
os.path.join(opt.start_from, "optimizer.pth")):
optimizer.load_state_dict(
torch.load(os.path.join(opt.start_from, 'optimizer.pth')))
#########################
# Get ready to start
#########################
iteration = infos['iter']
epoch = infos['epoch']
# For back compatibility
if 'iterators' in infos:
infos['loader_state_dict'] = {
split: {
'index_list': infos['split_ix'][split],
'iter_counter': infos['iterators'][split]
}
for split in ['train', 'val', 'test']
}
loader.load_state_dict(infos['loader_state_dict'])
if opt.load_best_score == 1:
best_val_score = infos.get('best_val_score', None)
if opt.noamopt:
optimizer._step = iteration
# flag indicating finish of an epoch
# Always set to True at the beginning to initialize the lr or etc.
epoch_done = True
# Assure in training mode
dp_lw_model.train()
# Start training
try:
while True:
# Stop if reaching max epochs
if epoch >= opt.max_epochs and opt.max_epochs != -1:
break
if epoch_done:
if not opt.noamopt and not opt.reduce_on_plateau:
# Assign the learning rate
if epoch > opt.learning_rate_decay_start and opt.learning_rate_decay_start >= 0:
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
else:
opt.current_lr = opt.learning_rate
utils.set_lr(optimizer,
opt.current_lr) # set the decayed rate
# Assign the scheduled sampling prob
if epoch > opt.scheduled_sampling_start and opt.scheduled_sampling_start >= 0:
frac = (epoch - opt.scheduled_sampling_start
) // opt.scheduled_sampling_increase_every
opt.ss_prob = min(
opt.scheduled_sampling_increase_prob * frac,
opt.scheduled_sampling_max_prob)
model.ss_prob = opt.ss_prob
# If start self critical training
if opt.self_critical_after != -1 and epoch >= opt.self_critical_after:
sc_flag = True
init_scorer(opt.cached_tokens)
else:
sc_flag = False
# If start structure loss training
if opt.structure_after != -1 and epoch >= opt.structure_after:
struc_flag = True
init_scorer(opt.cached_tokens)
else:
struc_flag = False
epoch_done = False
start = time.time()
if opt.use_warmup and (iteration < opt.noamopt_warmup):
opt.current_lr = opt.learning_rate * (iteration +
1) / opt.noamopt_warmup
utils.set_lr(optimizer, opt.current_lr)
# Load data from train split (0)
data = loader.get_batch('train')
print('Read data:', time.time() - start)
torch.cuda.synchronize()
start = time.time()
tmp = [
data['fc_feats'], data['att_feats'], data['trace_feats'],
data['box_feats'], data['labels'], data['masks'],
data['att_masks'], data['trace_masks']
]
tmp = [_ if _ is None else _.cuda() for _ in tmp]
fc_feats, att_feats, trace_feats, box_feats, labels, masks, att_masks, trace_masks = tmp
optimizer.zero_grad()
model_out = dp_lw_model(fc_feats, att_feats, trace_feats,
box_feats, labels, masks, att_masks,
trace_masks, data['gts'],
torch.arange(0, len(data['gts'])), sc_flag,
struc_flag)
loss = model_out['loss'].mean()
loss.backward()
if opt.grad_clip_value != 0:
getattr(torch.nn.utils, 'clip_grad_%s_' %
(opt.grad_clip_mode))(model.parameters(),
opt.grad_clip_value)
if not torch.isnan(loss):
if opt.language_eval == 1:
print('Doing final model evaluation, not updating model.')
else:
optimizer.step()
else:
print('Meet nan loss', data['gts'], model_out)
train_loss = loss.item()
torch.cuda.synchronize()
end = time.time()
if struc_flag:
print("iter {} (epoch {}), train_loss = {:.3f}, lm_loss = {:.3f}, struc_loss = {:.3f}, time/batch = {:.3f}" \
.format(iteration, epoch, train_loss, model_out['lm_loss'].mean().item(), model_out['struc_loss'].mean().item(), end - start))
elif not sc_flag:
print("iter {} (epoch {}), train_loss = {:.3f}, time/batch = {:.3f}" \
.format(iteration, epoch, train_loss, end - start))
else:
print("iter {} (epoch {}), avg_reward = {:.3f}, time/batch = {:.3f}" \
.format(iteration, epoch, model_out['reward'].mean(), end - start))
# Update the iteration and epoch
iteration += 1
if data['bounds']['wrapped']:
epoch += 1
epoch_done = True
# Write the training loss summary
if (iteration % opt.losses_log_every == 0):
tb_summary_writer.add_scalar('train_loss', train_loss,
iteration)
if opt.noamopt:
opt.current_lr = optimizer.rate()
elif opt.reduce_on_plateau:
opt.current_lr = optimizer.current_lr
tb_summary_writer.add_scalar('learning_rate', opt.current_lr,
iteration)
tb_summary_writer.add_scalar('scheduled_sampling_prob',
model.ss_prob, iteration)
if sc_flag:
tb_summary_writer.add_scalar('avg_reward',
model_out['reward'].mean(),
iteration)
elif struc_flag:
tb_summary_writer.add_scalar(
'lm_loss', model_out['lm_loss'].mean().item(),
iteration)
tb_summary_writer.add_scalar(
'struc_loss', model_out['struc_loss'].mean().item(),
iteration)
tb_summary_writer.add_scalar(
'reward', model_out['reward'].mean().item(), iteration)
tb_summary_writer.add_scalar(
'reward_var', model_out['reward'].var(1).mean(),
iteration)
histories['loss_history'][
iteration] = train_loss if not sc_flag else model_out[
'reward'].mean()
histories['lr_history'][iteration] = opt.current_lr
histories['ss_prob_history'][iteration] = model.ss_prob
# update infos
infos['iter'] = iteration
infos['epoch'] = epoch
infos['loader_state_dict'] = loader.state_dict()
# make evaluation on validation set, and save model
if opt.language_eval == 1 or (iteration % opt.save_checkpoint_every == 0 and not opt.save_every_epoch) or \
(epoch_done and opt.save_every_epoch):
# eval model
eval_kwargs = {'split': 'val', 'dataset': opt.input_json}
eval_kwargs.update(vars(opt))
assert (opt.task in ['caption', 'c_joint_t'] and opt.eval_task == 'caption') or \
(opt.task in ['trace', 'c_joint_t'] and opt.eval_task == 'trace') or \
(opt.task == 'pred_both' and opt.eval_task == 'pred_both')
if opt.eval_task == 'caption':
val_loss, predictions, lang_stats = eval_utils.eval_split(
dp_model, lw_model.crit_caption, loader, 'caption',
eval_kwargs)
elif opt.eval_task == 'trace':
val_loss = None
# This is a little time consuming due to the linear programming solve.
val_loss = eval_utils.eval_trace_generation(
dp_model,
lw_model.crit_trace,
loader,
window_size=0,
eval_kwargs=eval_kwargs
) # Adjust the window_size as needed
lang_stats = None
predictions = None
elif opt.eval_task == 'pred_both':
val_loss, predictions, lang_stats = eval_utils.eval_split(
dp_model, lw_model.crit_caption, loader, 'both',
eval_kwargs) # caption generation
val_loss_trace = eval_utils.eval_trace_generation(
dp_model,
lw_model.crit_trace,
loader,
window_size=0,
eval_kwargs=eval_kwargs
) # Adjust the window_size as needed
if opt.language_eval == 1:
break # The language eval is done during testing, after the training finishes.
if opt.reduce_on_plateau:
if 'CIDEr' in lang_stats:
optimizer.scheduler_step(-lang_stats['CIDEr'])
else:
optimizer.scheduler_step(val_loss)
# Write validation result into summary
tb_summary_writer.add_scalar('validation loss', val_loss,
iteration)
if lang_stats is not None:
for k, v in lang_stats.items():
tb_summary_writer.add_scalar(k, v, iteration)
histories['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 best_val_score is None or current_score > best_val_score:
best_val_score = current_score
best_flag = True
# Dump miscalleous informations
infos['best_val_score'] = best_val_score
# '''
utils.save_checkpoint(opt, model, infos, optimizer, histories)
if opt.save_history_ckpt:
utils.save_checkpoint(
opt,
model,
infos,
optimizer,
append=str(epoch)
if opt.save_every_epoch else str(iteration))
if best_flag:
utils.save_checkpoint(opt,
model,
infos,
optimizer,
append='best')
# '''
except (RuntimeError, KeyboardInterrupt):
print('Save ckpt on exception ...')
# '''
utils.save_checkpoint(opt, model, infos, optimizer)
print('Save ckpt done.')
stack_trace = traceback.format_exc()
print(stack_trace)
def train(opt):
################################
# 创建dataloader
################################
loader = DataLoader(opt)
opt.vocab_size = loader.vocab_size
opt.seq_length = loader.seq_length
##########################
# 初始化训练信息
##########################
infos = {
'iter': 0,
'epoch': 0,
'loader_state_dict': None,
'vocab': loader.get_vocab(),
'stage': 1,
'stage_saved': 1 # 用于中断处理，记录了中断时的状态，用于判定是否重新加载最佳模型
}
if opt.start_from is not None and os.path.isfile(
os.path.join(opt.start_from, 'infos_' + opt.id + '.pkl')):
with open(os.path.join(opt.start_from, 'infos_' + opt.id + '.pkl'),
'rb') as f:
infos = utils.pickle_load(f)
saved_model_opt = infos['opt']
need_be_same = [
"caption_model", "rnn_type", "rnn_size", "num_layers"
]
for checkme in need_be_same:
assert getattr(saved_model_opt, checkme) == getattr(
opt, checkme
), "Command line argument and saved model disagree on '%s' " % checkme
infos['opt'] = opt
#########################
# 创建logger
#########################
# 文件logger
histories = defaultdict(dict)
if opt.start_from is not None and os.path.isfile(
os.path.join(opt.start_from, 'histories_' + opt.id + '.pkl')):
with open(os.path.join(opt.start_from, 'histories_' + opt.id + '.pkl'),
'rb') as f:
histories.update(utils.pickle_load(f))
# tensorboard logger
tb_summary_writer = SummaryWriter(opt.checkpoint_path)
##########################
# 创建模型
##########################
opt.vocab = loader.get_vocab()
model = models.setup(opt).cuda()
del opt.vocab
if opt.finetune_only == 1:
if os.path.isfile(os.path.join(opt.start_from, 'model_best.pth')):
model.load_state_dict(
torch.load(os.path.join(opt.start_from, 'model_best.pth')))
else:
if opt.start_from is not None and os.path.isfile(
os.path.join(opt.start_from, 'model.pth')):
model.load_state_dict(
torch.load(os.path.join(opt.start_from, 'model.pth')))
# 作者注：面向模型的loss封装，便于将loss计算独立，便于多卡时减小No.0 GPU的负载
lw_model = LossWrapper(model, opt)
# 多GPU封装
dp_model = torch.nn.DataParallel(model)
dp_model.vocab = getattr(model, 'vocab', None)
dp_lw_model = torch.nn.DataParallel(lw_model)
model.set_stage(infos['stage'])
##########################
# 创建优化器
##########################
if opt.noamopt:
assert opt.caption_model in [
'transformer', 'bert', 'm2transformer'
], 'noamopt can only work with transformer'
optimizer = utils.get_std_opt(model,
optim_func=opt.optim,
factor=opt.noamopt_factor,
warmup=opt.noamopt_warmup)
elif opt.reduce_on_plateau:
optimizer = utils.build_optimizer(model.parameters(), opt)
optimizer = utils.ReduceLROnPlateau(
optimizer,
factor=opt.reduce_on_plateau_factor,
patience=opt.reduce_on_plateau_patience)
else:
optimizer = utils.build_optimizer(model.parameters(), opt)
if opt.finetune_only == 1:
if os.path.isfile(os.path.join(opt.start_from, "optimizer_best.pth")):
optimizer.load_state_dict(
torch.load(os.path.join(opt.start_from, 'optimizer_best.pth')))
else:
if opt.start_from is not None and os.path.isfile(
os.path.join(opt.start_from, "optimizer.pth")):
optimizer.load_state_dict(
torch.load(os.path.join(opt.start_from, 'optimizer.pth')))
#########################
# 训练
#########################
# 准备阶段
iteration = infos['iter']
epoch = infos['epoch']
loader.load_state_dict(infos['loader_state_dict'])
if opt.load_best_score == 1:
best_val_score = infos.get('best_val_score', None)
if opt.noamopt:
optimizer._step = iteration
# 作者注：轮次完成标志量，用于新轮次可能的训练参数调整
epoch_done = True
eval_done = False
dp_lw_model.train()
# 开始训练啦！经典训练
if infos['stage'] == 1 and opt.finetune_only != 1:
try:
while True:
# 达到最大epoch限制，跳出经典训练
if epoch >= opt.max_epochs_base != -1:
if eval_done:
break
else:
# 末尾再评估一次
eval_kwargs = {
'split': 'base_val',
'dataset': opt.input_json
}
eval_kwargs.update(vars(opt))
val_loss, predictions, lang_stats, _ = eval_utils.eval_split(
dp_model, lw_model.crit, loader, eval_kwargs)
if opt.reduce_on_plateau:
if 'CIDEr' in lang_stats:
optimizer.scheduler_step(-lang_stats['CIDEr'])
else:
optimizer.scheduler_step(val_loss)
# 将评估结果写入日志
tb_summary_writer.add_scalar('validation loss',
val_loss, iteration)
if lang_stats is not None:
for k, v in lang_stats.items():
tb_summary_writer.add_scalar(k, v, iteration)
histories['val_result_history'][iteration] = {
'loss': val_loss,
'lang_stats': lang_stats,
'predictions': predictions
}
# 根据CIDEr指标选择最佳模型
if opt.language_eval == 1:
current_score = lang_stats['CIDEr']
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
infos['best_val_score'] = best_val_score
utils.save_checkpoint(opt, model, infos, optimizer,
histories)
if opt.save_history_ckpt:
utils.save_checkpoint(
opt,
model,
infos,
optimizer,
append=str(epoch)
if opt.save_every_epoch else str(iteration))
if best_flag:
utils.save_checkpoint(opt,
model,
infos,
optimizer,
append='best')
break
eval_done = False
# 设置学习参数
if epoch_done:
# Transformer相关
if not opt.noamopt and not opt.reduce_on_plateau:
if epoch > opt.learning_rate_decay_start >= 0:
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_base * decay_factor
else:
opt.current_lr = opt.learning_rate_base
utils.set_lr(optimizer, opt.current_lr)
# scheduled sampling
if epoch > opt.scheduled_sampling_start >= 0:
frac = (epoch - opt.scheduled_sampling_start
) // opt.scheduled_sampling_increase_every
opt.ss_prob = min(
opt.scheduled_sampling_increase_prob * frac,
opt.scheduled_sampling_max_prob)
model.ss_prob = opt.ss_prob
# SCST
if opt.self_critical_after != -1 and epoch >= opt.self_critical_after:
sc_flag = True
init_scorer(opt.cached_tokens)
else:
sc_flag = False
# 结构损失
if opt.structure_after != -1 and epoch >= opt.structure_after:
struc_flag = True
init_scorer(opt.cached_tokens)
else:
struc_flag = False
epoch_done = False
# start = time.time()
# Transformer Warmup
if opt.use_warmup and (iteration < opt.noamopt_warmup):
opt.current_lr = opt.learning_rate_base * (
iteration + 1) / opt.noamopt_warmup
utils.set_lr(optimizer, opt.current_lr)
data = loader.get_batch('base_train')
# print('\r Read data:', time.time() - start, end="")
torch.cuda.synchronize()
start = time.time()
tmp = [
data['fc_feats'], data['att_feats'], data['labels'],
data['masks'], data['att_masks']
]
tmp = [_ if _ is None else _.cuda() for _ in tmp]
fc_feats, att_feats, labels, masks, att_masks = tmp
optimizer.zero_grad()
model_out = dp_lw_model(fc_feats, att_feats, labels, masks,
att_masks, data['gts'],
torch.arange(0, len(data['gts'])),
sc_flag, struc_flag)
loss = model_out['loss'].mean()
loss.backward()
# 梯度截断
if opt.grad_clip_value != 0:
getattr(torch.nn.utils, 'clip_grad_{}_'.format(
opt.grad_clip_mode))(model.parameters(),
opt.grad_clip_value)
optimizer.step()
train_loss = loss.item()
torch.cuda.synchronize()
end = time.time()
# 输出
if struc_flag:
print('Base Training:', "iter {} (epoch {}), train_loss = {:.3f}, lm_loss = {:.3f}, struc_loss = {:.3f}, time/batch = {:.3f}" \
.format(iteration, epoch, train_loss, model_out['lm_loss'].mean().item(), model_out['struc_loss'].mean().item(), end - start))
elif not sc_flag:
print('Base Training:', "iter {} (epoch {}), train_loss = {:.3f}, time/batch = {:.3f}" \
.format(iteration, epoch, train_loss, end - start))
else:
print('Base Training:', "iter {} (epoch {}), avg_reward = {:.3f}, time/batch = {:.3f}" \
.format(iteration, epoch, model_out['reward'].mean(), end - start))
# 更新迭代计数器，如果到达epoch边界，需要调整一些参数
iteration += 1
if data['bounds']['wrapped']:
epoch += 1
epoch_done = True
# 将训练结构写入到日志中
if iteration % opt.losses_log_every == 0:
tb_summary_writer.add_scalar('train_loss', train_loss,
iteration)
if opt.noamopt:
opt.current_lr = optimizer.rate()
elif opt.reduce_on_plateau:
opt.current_lr = optimizer.current_lr
tb_summary_writer.add_scalar('learning_rate',
opt.current_lr, iteration)
tb_summary_writer.add_scalar('scheduled_sampling_prob',
model.ss_prob, iteration)
if sc_flag:
tb_summary_writer.add_scalar(
'avg_reward', model_out['reward'].mean(),
iteration)
elif struc_flag:
tb_summary_writer.add_scalar(
'lm_loss', model_out['lm_loss'].mean().item(),
iteration)
tb_summary_writer.add_scalar(
'struc_loss',
model_out['struc_loss'].mean().item(), iteration)
tb_summary_writer.add_scalar(
'reward', model_out['reward'].mean().item(),
iteration)
tb_summary_writer.add_scalar(
'reward_var', model_out['reward'].var(1).mean(),
iteration)
histories['loss_history'][
iteration] = train_loss if not sc_flag else model_out[
'reward'].mean()
histories['lr_history'][iteration] = opt.current_lr
histories['ss_prob_history'][iteration] = model.ss_prob
# 信息更新
infos['iter'] = iteration
infos['epoch'] = epoch
infos['loader_state_dict'] = loader.state_dict()
# 根据需要，在两个模式下评估模型
if (iteration % opt.save_checkpoint_every == 0
and not opt.save_every_epoch) or (
epoch_done and opt.save_every_epoch):
eval_kwargs = {
'split': 'base_val',
'dataset': opt.input_json
}
eval_kwargs.update(vars(opt))
val_loss, predictions, lang_stats, _ = eval_utils.eval_split(
dp_model, lw_model.crit, loader, eval_kwargs)
if opt.reduce_on_plateau:
if 'CIDEr' in lang_stats:
optimizer.scheduler_step(-lang_stats['CIDEr'])
else:
optimizer.scheduler_step(val_loss)
# 将评估结果写入日志
tb_summary_writer.add_scalar('validation loss', val_loss,
iteration)
if lang_stats is not None:
for k, v in lang_stats.items():
tb_summary_writer.add_scalar(k, v, iteration)
histories['val_result_history'][iteration] = {
'loss': val_loss,
'lang_stats': lang_stats,
'predictions': predictions
}
# 根据CIDEr指标选择最佳模型
if opt.language_eval == 1:
current_score = lang_stats['CIDEr']
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
infos['best_val_score'] = best_val_score
utils.save_checkpoint(opt, model, infos, optimizer,
histories)
if opt.save_history_ckpt:
utils.save_checkpoint(
opt,
model,
infos,
optimizer,
append=str(epoch)
if opt.save_every_epoch else str(iteration))
if best_flag:
utils.save_checkpoint(opt,
model,
infos,
optimizer,
append='best')
eval_done = True
except (RuntimeError, KeyboardInterrupt):
print('Save ckpt on exception ...')
utils.save_checkpoint(opt, model, infos, optimizer)
print('Save ckpt done.')
stack_trace = traceback.format_exc()
print(stack_trace)
os._exit(0)
infos['stage'] = 2
# dummy配置下，不进行微调
if opt.train_only == 1:
# 微调训练
infos['stage'] = 2
epoch_done = True
loader.reset_iterator('support')
# 加载最佳模型，如果中断位置在第二阶段，则不进行模型加载
if opt.start_from and infos['stage_saved'] == 2:
pass
else:
# 否则加载stage 1的最佳模型进行微调
print('Finetuning:', "loading best model from stage 1")
model.load_state_dict(
torch.load(os.path.join(opt.start_from,
'model_best' + '.pth')))
optimizer.load_state_dict(
torch.load(
os.path.join(opt.start_from, 'optimizer_best' + '.pth')))
lw_model = LossWrapper(model, opt)
# 多GPU封装
dp_model = torch.nn.DataParallel(model)
dp_model.vocab = getattr(model, 'vocab', None)
dp_lw_model = torch.nn.DataParallel(lw_model)
model.set_stage(infos['stage'])
infos['stage_saved'] = 2
# 冻结除了最后一个logit层之外的所有参数
for name, parameter in dp_lw_model.module.named_parameters():
if 'logit' not in name:
parameter.requires_grad = False
else:
parameter.requires_grad = True
# 因为计数器没有清零，所以这里是直接加上去
max_epochs_all = opt.max_epochs_base + opt.max_epochs_finetune
# 提前准备：相关学习参数是否跟随
if opt.learning_rate_decay_start_finetune < 0:
opt.learning_rate_decay_start_finetune = opt.learning_rate_decay_start - opt.max_epochs_base
if opt.learning_rate_finetune < 0:
opt.learning_rate_finetune = opt.learning_rate_base
if opt.scheduled_sampling_start_finetune < 0:
opt.scheduled_sampling_start_finetune = opt.scheduled_sampling_start - opt.max_epochs_base
try:
while True:
# 达到最大epoch限制，跳出
if epoch >= max_epochs_all != -2:
utils.save_checkpoint(opt,
model,
infos,
optimizer,
histories,
append='finetune')
break
# 设置学习参数
if epoch_done:
# Transformer相关
if not opt.noamopt and not opt.reduce_on_plateau:
if epoch > opt.learning_rate_decay_start_finetune + opt.max_epochs_base >= 0:
frac = (epoch -
opt.learning_rate_decay_start_finetune -
opt.max_epochs_base
) // opt.learning_rate_decay_every_finetune
decay_factor = opt.learning_rate_decay_rate_finetune**frac
opt.current_lr = opt.learning_rate_finetune * decay_factor
else:
opt.current_lr = opt.learning_rate_finetune
utils.set_lr(optimizer, opt.current_lr)
# scheduled sampling
if epoch > opt.scheduled_sampling_start_finetune + opt.max_epochs_base >= 0:
frac = (
epoch - opt.scheduled_sampling_start_finetune -
opt.max_epochs_base
) // opt.scheduled_sampling_increase_every_finetune
opt.ss_prob = min(
opt.scheduled_sampling_increase_prob_finetune *
frac, opt.scheduled_sampling_max_prob_finetune)
model.ss_prob = opt.ss_prob
# SCST
if opt.self_critical_after != -1 and epoch >= opt.self_critical_after:
sc_flag = True
init_scorer(opt.cached_tokens)
else:
sc_flag = False
# 结构损失
if opt.structure_after != -1 and epoch >= opt.structure_after:
struc_flag = True
init_scorer(opt.cached_tokens)
else:
struc_flag = False
epoch_done = False
# start = time.time()
# Transformer Warmup
# if opt.use_warmup and (iteration < opt.noamopt_warmup):
# opt.current_lr = opt.learning_rate * (iteration + 1) / opt.noamopt_warmup
# utils.set_lr(optimizer, opt.current_lr)
data = loader.get_batch('support')
torch.cuda.synchronize()
start = time.time()
tmp = [
data['fc_feats'], data['att_feats'], data['labels'],
data['masks'], data['att_masks']
]
tmp = [_ if _ is None else _.cuda() for _ in tmp]
fc_feats, att_feats, labels, masks, att_masks = tmp
optimizer.zero_grad()
model_out = dp_lw_model(fc_feats, att_feats, labels, masks,
att_masks, data['gts'],
torch.arange(0, len(data['gts'])),
sc_flag, struc_flag)
loss = model_out['loss'].mean()
loss.backward()
# 梯度截断
if opt.grad_clip_value != 0:
getattr(torch.nn.utils, 'clip_grad_{}_'.format(
opt.grad_clip_mode))(model.parameters(),
opt.grad_clip_value)
optimizer.step()
train_loss = loss.item()
torch.cuda.synchronize()
end = time.time()
# 输出
if struc_flag:
print('Finetuning:', "iter {} (epoch {}), train_loss = {:.3f}, lm_loss = {:.3f}, struc_loss = {:.3f}, time/batch = {:.3f}" \
.format(iteration, epoch, train_loss, model_out['lm_loss'].mean().item(), model_out['struc_loss'].mean().item(), end - start))
elif not sc_flag:
print('Finetuning:', "iter {} (epoch {}), train_loss = {:.3f}, time/batch = {:.3f}" \
.format(iteration, epoch, train_loss, end - start))
else:
print('Finetuning:', "iter {} (epoch {}), avg_reward = {:.3f}, time/batch = {:.3f}" \
.format(iteration, epoch, model_out['reward'].mean(), end - start))
# 更新迭代计数器，如果到达epoch边界，需要调整一些参数
iteration += 1
if data['bounds']['wrapped']:
epoch += 1
epoch_done = True
# 将训练结构写入到日志中
if iteration % opt.losses_log_every == 0:
tb_summary_writer.add_scalar('train_loss', train_loss,
iteration)
if opt.noamopt:
opt.current_lr = optimizer.rate()
elif opt.reduce_on_plateau:
opt.current_lr = optimizer.current_lr
tb_summary_writer.add_scalar('learning_rate',
opt.current_lr, iteration)
tb_summary_writer.add_scalar('scheduled_sampling_prob',
model.ss_prob, iteration)
if sc_flag:
tb_summary_writer.add_scalar(
'avg_reward', model_out['reward'].mean(),
iteration)
elif struc_flag:
tb_summary_writer.add_scalar(
'lm_loss', model_out['lm_loss'].mean().item(),
iteration)
tb_summary_writer.add_scalar(
'struc_loss',
model_out['struc_loss'].mean().item(), iteration)
tb_summary_writer.add_scalar(
'reward', model_out['reward'].mean().item(),
iteration)
tb_summary_writer.add_scalar(
'reward_var', model_out['reward'].var(1).mean(),
iteration)
histories['loss_history'][
iteration] = train_loss if not sc_flag else model_out[
'reward'].mean()
histories['lr_history'][iteration] = opt.current_lr
histories['ss_prob_history'][iteration] = model.ss_prob
# 信息更新
infos['iter'] = iteration
infos['epoch'] = epoch
infos['loader_state_dict'] = loader.state_dict()
if (iteration % opt.save_checkpoint_every == 0
and not opt.save_every_epoch) or (
epoch_done and opt.save_every_epoch):
utils.save_checkpoint(opt,
model,
infos,
optimizer,
histories,
append='finetune')
if opt.save_history_ckpt:
utils.save_checkpoint(
opt,
model,
infos,
optimizer,
append=str(epoch)
if opt.save_every_epoch else str(iteration))
except (RuntimeError, KeyboardInterrupt):
print('Save ckpt on exception ...')
utils.save_checkpoint(opt, model, infos, optimizer)
print('Save ckpt done.')
stack_trace = traceback.format_exc()
print(stack_trace)
os._exit(0)
