def write_prediction(self, config, test_set):
"""Performs an epoch of evaluation
Args:
config: (Config) with batch_size and dir_answers
test_set:(Dataset) instance
Returns:
files: (list) of path to files
perp: (float) perplexity on test set
"""
# 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)]
# iterate over the dataset
n_words, ce_words = 0, 0 # sum of ce for all words + nb of words
for img, formula in minibatches(test_set, config.batch_size):
fd = self._get_feed_dict(img,
training=False,
formula=formula,
dropout=1)
ce_words_eval, n_words_eval, ids_eval = self.sess.run(
[self.ce_words, self.n_words, self.pred_test.ids],
feed_dict=fd)
# TODO(guillaume): move this logic into tf graph
if self._config.decoding == "greedy":
ids_eval = np.expand_dims(ids_eval, axis=1)
elif self._config.decoding == "beam_search":
ids_eval = np.transpose(ids_eval, [0, 2, 1])
# print("---------------------------------------------------------------after decoding :")
# print(ids_eval)
n_words += n_words_eval
ce_words += ce_words_eval
# print("---------------------------------------------------------------formula and prediction :")
for form, preds in zip(formula, ids_eval):
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)
perp = -np.exp(ce_words / float(n_words))
return files, perp
def write_prediction(self, config, test_set):
"""Performs an epoch of evaluation
Args:
config: (Config) with batch_size and dir_answers
test_set:(Dataset) instance
Returns:
files: (list) of path to files
perp: (float) perplexity on test set
"""
# 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)]
nbatches = (len(test_set) + config.batch_size - 1) // config.batch_size
prog = Progbar(nbatches)
n_words, ce_words = 0, 0 # sum of ce for all words + nb of words
for i, (img,
formula) in enumerate(minibatches(test_set,
config.batch_size)):
fd = self._get_feed_dict(img, formula=formula, dropout=1)
ce_words_eval, n_words_eval, ids_eval = self.sess.run(
[self.ce_words, self.n_words, self.pred_test.ids],
feed_dict=fd)
if self._config.decoding == "greedy":
ids_eval = np.expand_dims(ids_eval, axis=1)
elif self._config.decoding == "beam_search":
ids_eval = np.transpose(ids_eval, [0, 2, 1])
n_words += n_words_eval
ce_words += ce_words_eval
for form, preds in zip(formula, ids_eval):
refs.append(form)
for j, pred in enumerate(preds):
hyps[j].append(pred)
prog.update(i + 1,
[("perplexity", -np.exp(ce_words / float(n_words)))])
files = write_answers(refs, hyps, self._vocab.id_to_tok,
config.dir_answers, self._vocab.id_end)
perp = -np.exp(ce_words / float(n_words))
return files, perp
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()
}