def _run_train_epoch(self, config, train_set, val_set, epoch, lr_schedule):
"""Performs an epoch of training
Args:
config: Config instance
train_set: Dataset instance
val_set: Dataset instance
epoch: (int) id of the epoch, starting at 0
lr_schedule: LRSchedule instance that takes care of learning proc
Returns:
score: (float) model will select weights that achieve the highest score
"""
# logging
batch_size = config.batch_size
nbatches = (len(train_set) + batch_size - 1) // batch_size
prog = Progbar(nbatches)
self.model.train()
self.encoder.train()
self.decoder.train()
train_loader = torch.utils.data.DataLoader(
ImgFormulaDataset(train_set),
batch_size=batch_size,
shuffle=True,
num_workers=3,
pin_memory=True)
# for i, (img, formula) in enumerate(train_loader):
for i, (img, formula) in enumerate(minibatches(train_set, batch_size)):
img = pad_batch_images_2(img)
img = torch.FloatTensor(img) # (N, W, H, C)
formula, formula_length = pad_batch_formulas(
formula, self._vocab.id_pad, self._vocab.id_end)
img = img.permute(0, 3, 1, 2) # (N, C, W, H)
formula = torch.LongTensor(formula) # (N,)
loss_eval = self.getLoss(img,
formula=formula,
lr=lr_schedule.lr,
dropout=config.dropout,
training=True)
prog.update(i + 1, [("loss", loss_eval), ("lr", lr_schedule.lr)])
# update learning rate
lr_schedule.update(batch_no=epoch * nbatches + i)
self.logger.info("- Training: {}".format(prog.info))
self.logger.info("- Config: (before evaluate, we need to see config)")
config.show(fun=self.logger.info)
# evaluation
config_eval = Config({
"dir_answers": self._dir_output + "formulas_val/",
"batch_size": config.batch_size
})
scores = self.evaluate(config_eval, val_set)
score = scores["perplexity"]
lr_schedule.update(score=score)
return score
def _get_feed_dict(self, img, formula=None, lr=None, dropout=1):
"""Returns a dict 网络的输入"""
img = pad_batch_images(img)
fd = {
self.img: img,
self.dropout: dropout,
}
if formula is not None:
formula, formula_length = pad_batch_formulas(formula, self._vocab.id_pad, self._vocab.id_end)
# print img.shape, formula.shape
fd[self.formula] = formula
fd[self.formula_length] = formula_length
if lr is not None:
fd[self.lr] = lr
return fd
def _run_evaluate_epoch(self, config, test_set):
"""Performs an epoch of evaluation
Args:
test_set: Dataset instance
params: (dict) with extra params in it
- "dir_name": (string)
Returns:
scores: (dict) scores["acc"] = 0.85 for instance
"""
self.model.eval()
self.encoder.eval()
self.decoder.eval()
# initialize containers of references and predictions
if self._config.decoding == "greedy":
refs, hyps = [], [[]]
elif self._config.decoding == "beam_search":
refs, hyps = [], [[] for i in range(self._config.beam_size)]
references = list() # references (true captions) for calculating BLEU-4 score
hypotheses = list() # hypotheses (predictions)
with torch.no_grad():
nbatches = len(test_set)
prog = Progbar(nbatches)
test_loader = torch.utils.data.DataLoader(ImgFormulaDataset(test_set),
batch_size=nbatches,
shuffle=True, num_workers=3, pin_memory=True)
for i, (img, formula) in enumerate(minibatches(test_set, nbatches)):
# print(type(img), len(img), img[0].shape)
# print(type(formula), formula)
# Move to GPU, if available
img = pad_batch_images_2(img)
img = torch.FloatTensor(img) # (N, W, H, C)
formula, formula_length = pad_batch_formulas(formula, self._vocab.id_pad, self._vocab.id_end)
img = img.permute(0, 3, 1, 2) # (N, C, W, H)
formula = torch.LongTensor(formula) # (N,)
img = img.to(self.device)
formula = formula.to(self.device)
# Forward prop.
imgs = self.encoder(img)
scores, caps_sorted, decode_lengths, alphas, sort_ind = self.decoder(imgs, formula, torch.LongTensor([[len(i)] for i in formula]))
# Since we decoded starting with <start>, the targets are all words after <start>, up to <end>
targets = caps_sorted[:, 1:]
# Remove timesteps that we didn't decode at, or are pads
# pack_padded_sequence is an easy trick to do this
scores, _ = pack_padded_sequence(scores, decode_lengths, batch_first=True)
targets, _ = pack_padded_sequence(targets, decode_lengths, batch_first=True)
# Calculate loss
loss = self.criterion(scores, targets)
print(scores.shape, targets.shape)
print(loss)
alpha_c = 1.
# Add doubly stochastic attention regularization
loss += alpha_c * ((1. - alphas.sum(dim=1)) ** 2).mean()
loss_eval = loss.item()
prog.update(i + 1, [("loss", loss_eval), ("perplexity", np.exp(loss_eval))])
# Store references (true captions), and hypothesis (prediction) for each image
# If for n images, we have n hypotheses, and references a, b, c... for each image, we need -
# references = [[ref1a, ref1b, ref1c], [ref2a, ref2b], ...], hypotheses = [hyp1, hyp2, ...]
# print("---------------------------------------------------------------formula and prediction :")
for form, preds in zip(formula, scores):
refs.append(form)
# print(form, " ---------- ", preds[0])
for i, pred in enumerate(preds):
hyps[i].append(pred)
files = write_answers(refs, hyps, self._vocab.id_to_tok, config.dir_answers, self._vocab.id_end)
scores = score_files(files[0], files[1])
# perp = - np.exp(ce_words / float(n_words))
# scores["perplexity"] = perp
self.logger.info("- Evaluating: {}".format(prog.info))
return {
"perplexity": loss.item()
}
sess = tf.Session()
sess.run(tf.global_variables_initializer())
feed_dicts = []
batch_size = config.batch_size
for i, (_img, _formula) in enumerate(minibatches(train_set, batch_size)):
fd = {
inputs: _img,
dropout: 0.2,
training: True,
learning_rate: 0.0001,
}
if _formula is not None:
_formula, _formula_length = pad_batch_formulas(
_formula,
vocab.id_pad,
vocab.id_end
)
fd[formula] = _formula
fd[formula_length] = _formula_length
feed_dicts.append(fd)
# all_img = []
# all_formula = []
# for i, (_img, _formula) in enumerate(minibatches(train_set, batch_size)):
# all_img.append(_img)
# if _formula is not None:
# _formula, _formula_length = pad_batch_formulas(
# _formula,
# vocab.id_pad,
# vocab.id_end
