def predict(self, sample=None):
pred_sents = []
actual_sents = []
img_files = []
for batch in self.valid_gen:
batch = self.batch_to_device(batch)
if self.beamsearch:
translated_sentence = batch_translate_beam_search(
batch['img'], self.model)
else:
translated_sentence = translate(batch['img'], self.model)
pred_sent = self.vocab.batch_decode(translated_sentence.tolist())
actual_sent = self.vocab.batch_decode(batch['tgt_output'].tolist())
img_files.extend(batch['filenames'])
pred_sents.extend(pred_sent)
actual_sents.extend(actual_sent)
if sample != None and len(pred_sents) > sample:
break
return pred_sents, actual_sents, img_files
def gen_pseudo_labels(self, outfile=None):
pred_sents = []
img_files = []
probs_sents = []
for idx, batch in enumerate(tqdm.tqdm(self.valid_gen)):
batch = self.batch_to_device(batch)
if self.model.seq_modeling != 'crnn':
if self.beamsearch:
translated_sentence = batch_translate_beam_search(
batch['img'], self.model)
prob = None
else:
translated_sentence, prob = translate(
batch['img'], self.model)
pred_sent = self.vocab.batch_decode(
translated_sentence.tolist())
else:
translated_sentence, prob = translate_crnn(
batch['img'], self.model)
pred_sent = self.vocab.batch_decode(
translated_sentence.tolist(), crnn=True)
pred_sents.extend(pred_sent)
img_files.extend(batch['filenames'])
probs_sents.extend(prob)
assert len(pred_sents) == len(img_files) and len(img_files) == len(
probs_sents)
with open(outfile, 'w', encoding='utf-8') as f:
for anno in zip(img_files, pred_sents, probs_sents):
f.write('||||'.join([anno[0], anno[1],
str(float(anno[2]))]) + '\n')
def batch_predict(self, images):
"""
param: images : list of ndarray
"""
batch_dict, indices = self.batch_process(images)
list_keys = [
i for i in batch_dict
if batch_dict[i] != batch_dict.default_factory()
]
result = list([])
for width in list_keys:
batch = batch_dict[width]
batch = np.asarray(batch)
batch = torch.FloatTensor(batch)
batch = batch.to(self.config['device'])
if self.config['predictor']['beamsearch']:
sent = batch_translate_beam_search(batch, model=self.model)
else:
sent = translate(batch, self.model).tolist()
batch_text = self.vocab.batch_decode(sent)
result.extend(batch_text)
# sort text result to original coordinate
def get_index(element):
return element[1]
z = zip(result, indices)
sorted_result = sorted(z, key=get_index)
result, _ = zip(*sorted_result)
return result
def predict(self, sample=None):
pred_sents = []
actual_sents = []
img_files = []
probs_sents = []
imgs_sents = []
for idx, batch in enumerate(tqdm.tqdm(self.valid_gen)):
batch = self.batch_to_device(batch)
if self.model.seq_modeling != 'crnn':
if self.beamsearch:
translated_sentence = batch_translate_beam_search(
batch['img'], self.model)
prob = None
else:
translated_sentence, prob = translate(
batch['img'], self.model)
pred_sent = self.vocab.batch_decode(
translated_sentence.tolist())
else:
translated_sentence, prob = translate_crnn(
batch['img'], self.model)
pred_sent = self.vocab.batch_decode(
translated_sentence.tolist(), crnn=True)
actual_sent = self.vocab.batch_decode(batch['tgt_output'].tolist())
pred_sents.extend(pred_sent)
actual_sents.extend(actual_sent)
imgs_sents.extend(batch['img'])
img_files.extend(batch['filenames'])
probs_sents.extend(prob)
# Visualize in tensorboard
if idx == 0:
try:
num_samples = self.config['monitor']['num_samples']
fig = plt.figure(figsize=(12, 15))
imgs_samples = imgs_sents[:num_samples]
preds_samples = pred_sents[:num_samples]
actuals_samples = actual_sents[:num_samples]
probs_samples = probs_sents[:num_samples]
for id_img in range(len(imgs_samples)):
img = imgs_samples[id_img]
img = img.permute(1, 2, 0)
img = img.cpu().detach().numpy()
ax = fig.add_subplot(num_samples,
1,
id_img + 1,
xticks=[],
yticks=[])
plt.imshow(img)
ax.set_title(
"LB: {} \n Pred: {:.4f}-{}".format(
actuals_samples[id_img], probs_samples[id_img],
preds_samples[id_img]),
color=('green' if actuals_samples[id_img]
== preds_samples[id_img] else 'red'),
fontdict={
'fontsize': 18,
'fontweight': 'medium'
})
self.writer.add_figure('predictions vs. actuals',
fig,
global_step=self.iter)
except Exception as error:
print(error)
continue
if sample != None and len(pred_sents) > sample:
break
return pred_sents, actual_sents, img_files, probs_sents, imgs_sents