class Trainer(object):
def __init__(self, args):
super(Trainer, self).__init__()
self.args = args
if cfg.SEED > 0:
random.seed(cfg.SEED)
torch.manual_seed(cfg.SEED)
torch.cuda.manual_seed_all(cfg.SEED)
self.num_gpus = torch.cuda.device_count()
self.distributed = self.num_gpus > 1
if self.distributed:
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(
backend="nccl", init_method="env://"
)
self.device = torch.device("cuda")
self.rl_stage = False
self.setup_logging()
self.setup_dataset()
self.setup_network()
self.val_evaler = Evaler(
eval_ids = cfg.DATA_LOADER.VAL_ID,
gv_feat = cfg.DATA_LOADER.VAL_GV_FEAT,
att_feats = cfg.DATA_LOADER.VAL_ATT_FEATS,
eval_annfile = cfg.INFERENCE.VAL_ANNFILE
)
self.test_evaler = Evaler(
eval_ids = cfg.DATA_LOADER.TEST_ID,
gv_feat = cfg.DATA_LOADER.TEST_GV_FEAT,
att_feats = cfg.DATA_LOADER.TEST_ATT_FEATS,
eval_annfile = cfg.INFERENCE.TEST_ANNFILE
)
self.scorer = Scorer()
def setup_logging(self):
self.logger = logging.getLogger(cfg.LOGGER_NAME)
self.logger.setLevel(logging.INFO)
if self.distributed and dist.get_rank() > 0:
return
ch = logging.StreamHandler(stream=sys.stdout)
ch.setLevel(logging.INFO)
formatter = logging.Formatter("[%(levelname)s: %(asctime)s] %(message)s")
ch.setFormatter(formatter)
self.logger.addHandler(ch)
if not os.path.exists(cfg.ROOT_DIR):
os.makedirs(cfg.ROOT_DIR)
fh = logging.FileHandler(os.path.join(cfg.ROOT_DIR, cfg.LOGGER_NAME + '.txt'))
fh.setLevel(logging.INFO)
fh.setFormatter(formatter)
self.logger.addHandler(fh)
self.logger.info('Training with config:')
self.logger.info(pprint.pformat(cfg))
def setup_network(self):
model = models.create(cfg.MODEL.TYPE)
if self.distributed:
# this should be removed if we update BatchNorm stats
self.model = torch.nn.parallel.DistributedDataParallel(
model.to(self.device),
device_ids = [self.args.local_rank],
output_device = self.args.local_rank,
broadcast_buffers = False
)
else:
self.model = torch.nn.DataParallel(model).cuda()
if self.args.resume > 0:
self.model.load_state_dict(
torch.load(self.snapshot_path("caption_model", self.args.resume),
map_location=lambda storage, loc: storage)
)
self.optim = Optimizer(self.model)
self.xe_criterion = losses.create(cfg.LOSSES.XE_TYPE).cuda()
self.rl_criterion = losses.create(cfg.LOSSES.RL_TYPE).cuda()
def setup_dataset(self):
self.coco_set = datasets.coco_dataset.CocoDataset(
image_ids_path = cfg.DATA_LOADER.TRAIN_ID,
input_seq = cfg.DATA_LOADER.INPUT_SEQ_PATH,
target_seq = cfg.DATA_LOADER.TARGET_SEQ_PATH,
gv_feat_path = cfg.DATA_LOADER.TRAIN_GV_FEAT,
att_feats_folder = cfg.DATA_LOADER.TRAIN_ATT_FEATS,
seq_per_img = cfg.DATA_LOADER.SEQ_PER_IMG,
max_feat_num = cfg.DATA_LOADER.MAX_FEAT
)
def setup_loader(self, epoch):
self.training_loader = datasets.data_loader.load_train(
self.distributed, epoch, self.coco_set)
def eval(self, epoch):
if (epoch + 1) % cfg.SOLVER.TEST_INTERVAL != 0:
return None
if self.distributed and dist.get_rank() > 0:
return None
val_res = self.val_evaler(self.model, 'val_' + str(epoch + 1))
self.logger.info('######## Epoch (VAL)' + str(epoch + 1) + ' ########')
self.logger.info(str(val_res))
test_res = self.test_evaler(self.model,'test_' + str(epoch + 1))
self.logger.info('######## Epoch (TEST)' + str(epoch + 1) + ' ########')
self.logger.info(str(test_res))
val = 0
for score_type, weight in zip(cfg.SCORER.TYPES, cfg.SCORER.WEIGHTS):
val -= val_res[score_type] * weight
return val
def snapshot_path(self, name, epoch):
snapshot_folder = os.path.join(cfg.ROOT_DIR, 'snapshot')
return os.path.join(snapshot_folder, name + "_" + str(epoch) + ".pth")
def save_model(self, epoch):
if (epoch + 1) % cfg.SOLVER.SNAPSHOT_ITERS != 0:
return
if self.distributed and dist.get_rank() > 0:
return
snapshot_folder = os.path.join(cfg.ROOT_DIR, 'snapshot')
if not os.path.exists(snapshot_folder):
os.mkdir(snapshot_folder)
torch.save(self.model.state_dict(), self.snapshot_path("caption_model", epoch+1))
def make_kwargs(self, indices, input_seq, target_seq, gv_feat, att_feats, att_mask):
seq_mask = (input_seq > 0).type(torch.cuda.LongTensor)
seq_mask[:,0] += 1
seq_mask_sum = seq_mask.sum(-1)
max_len = int(seq_mask_sum.max())
input_seq = input_seq[:, 0:max_len].contiguous()
target_seq = target_seq[:, 0:max_len].contiguous()
kwargs = {
cfg.PARAM.INDICES: indices,
cfg.PARAM.INPUT_SENT: input_seq,
cfg.PARAM.TARGET_SENT: target_seq,
cfg.PARAM.GLOBAL_FEAT: gv_feat,
cfg.PARAM.ATT_FEATS: att_feats,
cfg.PARAM.ATT_FEATS_MASK: att_mask
}
return kwargs
def scheduled_sampling(self, epoch):
if epoch > cfg.TRAIN.SCHEDULED_SAMPLING.START:
frac = (epoch - cfg.TRAIN.SCHEDULED_SAMPLING.START) // cfg.TRAIN.SCHEDULED_SAMPLING.INC_EVERY
ss_prob = min(cfg.TRAIN.SCHEDULED_SAMPLING.INC_PROB * frac, cfg.TRAIN.SCHEDULED_SAMPLING.MAX_PROB)
self.model.module.ss_prob = ss_prob
def display(self, iteration, data_time, batch_time, losses, loss_info):
if iteration % cfg.SOLVER.DISPLAY != 0:
return
if self.distributed and dist.get_rank() > 0:
return
info_str = ' (DataTime/BatchTime: {:.3}/{:.3}) losses = {:.5}'.format(data_time.avg, batch_time.avg, losses.avg)
self.logger.info('Iteration ' + str(iteration) + info_str +', lr = ' + str(self.optim.get_lr()))
for name in sorted(loss_info):
self.logger.info(' ' + name + ' = ' + str(loss_info[name]))
data_time.reset()
batch_time.reset()
losses.reset()
def forward(self, kwargs):
if self.rl_stage == False:
logit = self.model(**kwargs)
loss, loss_info = self.xe_criterion(logit, kwargs[cfg.PARAM.TARGET_SENT])
else:
ids = kwargs[cfg.PARAM.INDICES]
gv_feat = kwargs[cfg.PARAM.GLOBAL_FEAT]
att_feats = kwargs[cfg.PARAM.ATT_FEATS]
att_mask = kwargs[cfg.PARAM.ATT_FEATS_MASK]
# max
kwargs['BEAM_SIZE'] = 1
kwargs['GREEDY_DECODE'] = True
kwargs[cfg.PARAM.GLOBAL_FEAT] = gv_feat
kwargs[cfg.PARAM.ATT_FEATS] = att_feats
kwargs[cfg.PARAM.ATT_FEATS_MASK] = att_mask
self.model.eval()
with torch.no_grad():
seq_max, logP_max = self.model.module.decode(**kwargs)
self.model.train()
rewards_max, rewards_info_max = self.scorer(ids, seq_max.data.cpu().numpy().tolist())
rewards_max = utils.expand_numpy(rewards_max)
ids = utils.expand_numpy(ids)
gv_feat = utils.expand_tensor(gv_feat, cfg.DATA_LOADER.SEQ_PER_IMG)
att_feats = utils.expand_tensor(att_feats, cfg.DATA_LOADER.SEQ_PER_IMG)
att_mask = utils.expand_tensor(att_mask, cfg.DATA_LOADER.SEQ_PER_IMG)
# sample
kwargs['BEAM_SIZE'] = 1
kwargs['GREEDY_DECODE'] = False
kwargs[cfg.PARAM.GLOBAL_FEAT] = gv_feat
kwargs[cfg.PARAM.ATT_FEATS] = att_feats
kwargs[cfg.PARAM.ATT_FEATS_MASK] = att_mask
seq_sample, logP_sample = self.model.module.decode(**kwargs)
rewards_sample, rewards_info_sample = self.scorer(ids, seq_sample.data.cpu().numpy().tolist())
rewards = rewards_sample - rewards_max
rewards = torch.from_numpy(rewards).float().cuda()
loss = self.rl_criterion(seq_sample, logP_sample, rewards)
loss_info = {}
for key in rewards_info_sample:
loss_info[key + '_sample'] = rewards_info_sample[key]
for key in rewards_info_max:
loss_info[key + '_max'] = rewards_info_max[key]
return loss, loss_info
def train(self):
self.model.train()
self.optim.zero_grad()
iteration = 0
for epoch in range(cfg.SOLVER.MAX_EPOCH):
if epoch == cfg.TRAIN.REINFORCEMENT.START:
self.rl_stage = True
self.setup_loader(epoch)
start = time.time()
data_time = AverageMeter()
batch_time = AverageMeter()
losses = AverageMeter()
for _, (indices, input_seq, target_seq, gv_feat, att_feats, att_mask) in enumerate(self.training_loader):
data_time.update(time.time() - start)
input_seq = input_seq.cuda()
target_seq = target_seq.cuda()
gv_feat = gv_feat.cuda()
att_feats = att_feats.cuda()
att_mask = att_mask.cuda()
kwargs = self.make_kwargs(indices, input_seq, target_seq, gv_feat, att_feats, att_mask)
loss, loss_info = self.forward(kwargs)
loss.backward()
utils.clip_gradient(self.optim.optimizer, self.model,
cfg.SOLVER.GRAD_CLIP_TYPE, cfg.SOLVER.GRAD_CLIP)
self.optim.step()
self.optim.zero_grad()
self.optim.scheduler_step('Iter')
batch_time.update(time.time() - start)
start = time.time()
losses.update(loss.item())
self.display(iteration, data_time, batch_time, losses, loss_info)
iteration += 1
if self.distributed:
dist.barrier()
self.save_model(epoch)
val = self.eval(epoch)
self.optim.scheduler_step('Epoch', val)
self.scheduled_sampling(epoch)
if self.distributed:
dist.barrier()
class Trainer(object):
def __init__(self, args):
super(Trainer, self).__init__()
self.args = args
if cfg.SEED > 0:
np.random.seed(int(cfg.SEED))
random.seed(cfg.SEED)
torch.manual_seed(cfg.SEED)
torch.cuda.manual_seed_all(cfg.SEED)
self.setup_logging()
self.setup_dataset()
self.setup_network()
self.val_evaler = Evaler(eval_ids=cfg.DATA_LOADER.VAL_ID,
gv_feat=cfg.DATA_LOADER.VAL_GV_FEAT,
att_feats=cfg.DATA_LOADER.VAL_ATT_FEATS,
eval_annfile=cfg.INFERENCE.VAL_ANNFILE)
self.test_evaler = Evaler(eval_ids=cfg.DATA_LOADER.TEST_ID,
gv_feat=cfg.DATA_LOADER.TEST_GV_FEAT,
att_feats=cfg.DATA_LOADER.TEST_ATT_FEATS,
eval_annfile=cfg.INFERENCE.TEST_ANNFILE)
self.scorer = Scorer()
self._init_ppo()
def _init_ppo(self):
self.nenvs = 4 * cfg.TRAIN.BATCH_SIZE
self.noptepochs = 1
self.envsperbatch = cfg.TRAIN.BATCH_SIZE
self.clip_range = 0.1
assert self.nenvs % cfg.TRAIN.BATCH_SIZE == 0
self.batch_next = None
self.mv_approxkl = 0
self.mv_violate = 0
self.mv_entropy = 0
self.mv_total = 0
def setup_logging(self):
self.logger = logging.getLogger(cfg.LOGGER_NAME)
self.logger.setLevel(logging.INFO)
ch = logging.StreamHandler(stream=sys.stdout)
ch.setLevel(logging.INFO)
formatter = logging.Formatter(
"[%(levelname)s: %(asctime)s] %(message)s")
ch.setFormatter(formatter)
self.logger.addHandler(ch)
if not os.path.exists(cfg.ROOT_DIR):
os.makedirs(cfg.ROOT_DIR)
fh = logging.FileHandler(
os.path.join(cfg.ROOT_DIR, cfg.LOGGER_NAME + '.txt'))
fh.setLevel(logging.INFO)
fh.setFormatter(formatter)
self.logger.addHandler(fh)
self.logger.info('Training with config:')
self.logger.info(pprint.pformat(cfg))
def setup_network(self):
model = models.create(cfg.MODEL.TYPE)
self.trainer = model.cuda()
self.checkpointer = CaptionCheckpointer(
self.trainer, os.path.join(cfg.ROOT_DIR, "snapshot"))
if self.args.resume > 0:
self.checkpointer.load(
self.snapshot_path("caption_model", self.args.resume))
self.predictor = models.create(cfg.MODEL.TYPE).cuda()
self.predictor.load_state_dict(self.trainer.state_dict())
self.optim = Optimizer(self.trainer)
def setup_dataset(self):
self.coco_set = datasets.coco_dataset.CocoDataset(
image_ids_path=cfg.DATA_LOADER.TRAIN_ID,
input_seq=cfg.DATA_LOADER.INPUT_SEQ_PATH,
target_seq=cfg.DATA_LOADER.TARGET_SEQ_PATH,
gv_feat_path=cfg.DATA_LOADER.TRAIN_GV_FEAT,
att_feats_folder=cfg.DATA_LOADER.TRAIN_ATT_FEATS,
seq_per_img=1,
max_feat_num=cfg.DATA_LOADER.MAX_FEAT)
def setup_loader(self, epoch):
self.training_loader = datasets.data_loader.load_train(
False, epoch, self.coco_set)
def snapshot_path(self, name, epoch):
snapshot_folder = os.path.join(cfg.ROOT_DIR, 'snapshot')
return os.path.join(snapshot_folder, name + "_" + str(epoch) + ".pth")
def save_model(self, iteration):
snapshot_folder = os.path.join(cfg.ROOT_DIR, 'snapshot')
if not os.path.exists(snapshot_folder):
os.mkdir(snapshot_folder)
torch.save(self.trainer.state_dict(),
self.snapshot_path("caption_model", iteration))
def get_batch(self):
epoch = 0
while True:
self.setup_loader(epoch)
for x in self.training_loader:
yield epoch, x
epoch += 1
def _set_kwargs(self, kwargs, repeat_factor):
# sample
kwargs['GREEDY_DECODE'] = False
kwargs["NEED_PD"] = True
kwargs["REPEAT_FACTOR"] = repeat_factor
return kwargs
def _prefix_rewards(self, kwargs, seq_prefix):
kwargs[cfg.PARAM.MAX_GEN_LEN] = seq_prefix.shape[-1]
kwargs[cfg.PARAM.GEN_RESULT] = utils.expand_tensor(seq_prefix, 5)
with torch.no_grad():
seq_sample = self.predictor.extend_trajectory(
**kwargs).detach().cpu().numpy().tolist()
return seq_sample
def _sample_trajectory(self, kwargs):
self._set_kwargs(kwargs, 1)
with torch.no_grad():
seq_sample, log_prob_sample = self.predictor.decode(**kwargs)
seq_sample_list = seq_sample.detach().cpu().numpy().tolist()
log_prob_sample, seq_sample = [
_.detach().cpu().numpy()
for _ in [log_prob_sample, seq_sample]
]
indices = kwargs[cfg.PARAM.INDICES]
rewards_sample, _ = self.scorer(indices, seq_sample_list)
repeat_factor = 5
kwargs = self._set_kwargs(kwargs, repeat_factor)
kwargs['gx'], kwargs['encoder_out'], kwargs['p_att_feats'] , kwargs['att_mask']= \
self.predictor.init_gx_encoder_out_p_att_feats_att_mask(**kwargs)
indices = utils.expand_numpy(kwargs[cfg.PARAM.INDICES], repeat_factor)
advs = np.zeros_like(seq_sample, dtype=np.float32)
for k in range(cfg.MODEL.SEQ_LEN):
baseline, _ = self.scorer(
indices,
self._prefix_rewards(
kwargs,
torch.from_numpy(seq_sample[:, :k]).cuda()))
baseline = baseline.reshape(-1, repeat_factor)
advs[:, k] = rewards_sample - baseline.mean(-1)
if seq_sample[:, k].sum() == 0:
break
seq_sample = seq_sample[:, None, :]
advs = np.clip(advs[:, None, :], -1, 1)
#advs = advs[:, None, :]
log_prob_sample = log_prob_sample[:, None, ...]
return seq_sample, log_prob_sample, advs
def runner_run(self, iteration):
mb_indices = []
mb_gv_feat = []
mb_att_feats = []
mb_att_mask = []
mb_sample_logprobs = []
mb_gen_result = []
mb_advs = []
for _ in range(self.nenvs // cfg.TRAIN.BATCH_SIZE):
# data - indices, input_seq, target_seq, gv_feat, att_feats, att_mask
epoch, data = next(self.batch_next)
iteration += 1
indices = data[0]
mb_indices.append(indices.reshape(-1, 1))
for x, y in zip(data[-3:],
[mb_gv_feat, mb_att_feats, mb_att_mask]):
y.append(x.numpy())
gv_feat, att_feats, att_mask = [_.cuda() for _ in data[-3:]]
kwargs = {
cfg.PARAM.INDICES: indices,
cfg.PARAM.GLOBAL_FEAT: gv_feat,
cfg.PARAM.ATT_FEATS: att_feats,
cfg.PARAM.ATT_FEATS_MASK: att_mask
}
seq_sample, log_prob_sample, rewards = self._sample_trajectory(
kwargs)
trajectory = [log_prob_sample, seq_sample, rewards]
for x, y in zip(trajectory,
[mb_sample_logprobs, mb_gen_result, mb_advs]):
y.append(x)
max_att_num = np.max([_.shape[1] for _ in mb_att_feats])
for k, x in enumerate(mb_att_feats):
after = max_att_num - x.shape[1]
mb_att_feats[k] = np.pad(x, ((0, 0), (0, after), (0, 0)),
mode="constant")
mb_att_mask[k] = np.pad(mb_att_mask[k], ((0, 0), (0, after)),
mode="constant")
mb_indices, mb_gv_feat, mb_att_feats, mb_att_mask, \
mb_sample_logprobs, mb_gen_result, mb_advs = [np.vstack(_) for _ in [
mb_indices, mb_gv_feat, mb_att_feats, mb_att_mask,
mb_sample_logprobs, mb_gen_result, mb_advs
]]
return iteration, epoch, [
mb_indices, mb_gv_feat, mb_att_feats, mb_att_mask,
mb_sample_logprobs, mb_gen_result, mb_advs
]
def mb_train(self, kwargs):
kwargs = self._set_kwargs(kwargs, 1)
_, neglogpac = self.trainer.decode(**kwargs)
sample_logprobs, gen_result, advs = [
kwargs[_] for _ in
[cfg.PARAM.SAMPLE_LOGPROBS, cfg.PARAM.GEN_RESULT, cfg.PARAM.ADVS]
]
trajectory = [sample_logprobs, gen_result, advs]
for k, _ in enumerate(trajectory):
trajectory[k] = _.view(-1, *_.shape[2:])
sample_logprobs, gen_result, advs = trajectory
mask = gen_result > 0
mask = torch.cat(
[mask.new_full((mask.shape[0], 1), True), mask[:, :-1]], 1)
kl_div = torch.exp(sample_logprobs) * (sample_logprobs - neglogpac)
kl_div = kl_div.sum(-1)
kl_div = torch.masked_select(kl_div, mask)
entropy = torch.sum(torch.exp(neglogpac) * (-neglogpac), dim=-1)
entropy = entropy[mask].mean()
neglogpac = torch.gather(neglogpac, 2,
gen_result.unsqueeze(-1)).squeeze(-1)
sample_logprobs = torch.gather(sample_logprobs, 2,
gen_result.unsqueeze(-1)).squeeze(-1)
advs_close_zero = (-1e-5 < advs) & (advs < 1e-5)
mask &= ~advs_close_zero
neglogpac = -torch.masked_select(neglogpac, mask)
oldneglogpac = -torch.masked_select(sample_logprobs, mask)
advs = torch.masked_select(advs, mask)
ratio = torch.exp(oldneglogpac - neglogpac)
pg_losses = -advs * ratio
pg_losses2 = -advs * torch.clamp(ratio, 1.0 - self.clip_range,
1.0 + self.clip_range)
self.mv_total = 0.9 * self.mv_total + 0.1 * pg_losses.shape[0]
pg_loss = torch.max(pg_losses,
pg_losses2).sum() / (cfg.TRAIN.BATCH_SIZE * 16)
mask_positive = (advs > 0) & (ratio > 1 + self.clip_range)
mask_negative = (advs < 0) & (ratio < 1 - self.clip_range)
mask_total = mask_positive | mask_negative
kl_div = kl_div.mean()
loss = pg_loss
self.mv_approxkl = 0.9 * self.mv_approxkl + 0.1 * kl_div.item()
self.mv_entropy = 0.9 * self.mv_entropy + 0.1 * entropy.item()
self.mv_violate = 0.9 * self.mv_violate + 0.1 * mask_total.sum().item()
return loss
def train(self):
self.batch_next = self.get_batch()
# eval - crucial to disable dropout
self.trainer.eval()
self.predictor.eval()
epoch, iteration = 0, 0
val_current = None
val_best = self._compute_val(iteration)
self.logger.info("val @iteration 0: {}".format(val_best))
while True:
iteration, epoch, data = self.runner_run(iteration)
envinds = np.arange(self.nenvs)
for _ in range(self.noptepochs):
np.random.shuffle(envinds)
for start in range(0, self.nenvs, self.envsperbatch):
end = start + self.envsperbatch
mbenvinds = envinds[start:end]
indices = data[0][mbenvinds].reshape(-1)
gv_feat, att_feats, att_mask, sample_logprobs, gen_result, advs = \
[torch.from_numpy(x).cuda()
for x in [_[mbenvinds] for _ in data[1:]]]
kwargs = {
cfg.PARAM.INDICES: indices,
cfg.PARAM.GLOBAL_FEAT: gv_feat,
cfg.PARAM.ATT_FEATS: att_feats,
cfg.PARAM.ATT_FEATS_MASK: att_mask,
cfg.PARAM.SAMPLE_LOGPROBS: sample_logprobs,
cfg.PARAM.GEN_RESULT: gen_result,
cfg.PARAM.ADVS: advs
}
loss = self.mb_train(kwargs)
self.optim.zero_grad()
loss.backward()
#torch.nn.utils.clip_grad_norm_(self.trainer.parameters(), 0.5, 2)
self.optim.step()
time.sleep(1)
if iteration % 64 == 0:
self.predictor.load_state_dict(self.trainer.state_dict())
val_current = self._compute_val(iteration)
if val_best is None:
val_best = val_current
self.logger.info(
"val_current @iteration {}: {}, val_predictor: {}".format(
iteration, val_current, val_best))
self.logger.info("mv_approxkl: {}".format(self.mv_approxkl))
self.logger.info("mv_entropy: {}".format(self.mv_entropy))
self.logger.info("mv_violate: {}".format(self.mv_violate))
self.logger.info("mv_total: {}".format(self.mv_total))
if val_best <= val_current:
self.save_model(23)
val_best = val_current
if iteration == 3584:
break
def _compute_val(self, iteration):
val_res = self.val_evaler(self.trainer, 'val_' + str(iteration), 1)
val = 0
for score_type, weight in zip(cfg.SCORER.TYPES, cfg.SCORER.WEIGHTS):
# plus!!!
val += val_res[score_type] * weight
test_res = self.test_evaler(self.trainer, 'test_' + str(iteration), 2)
self.logger.info('######## Iter (TEST) ' + str(iteration) +
' ########')
self.logger.info(str(test_res))
# crucial!
self.trainer.eval()
return val