) as f:
ret = []
f.readline()
f.readline()
for line in f:
line = line.split(' ')
while line[-1].strip().isdigit() is False:
line = line[:-1]
ret.append([' '.join(line[0:-1]).strip(), int(line[-1])])
attr_type = pd.DataFrame(ret, columns=['attr_name', 'type'])
attr_type['attr_index'] = ['attr_' + str(i) for i in range(1000)]
attr_type.set_index('attr_index', inplace=True)
with torch.no_grad():
net.eval()
evaluator = const.EVALUATOR()
cat_pred_lines = []
for sample_idx, sample in enumerate(inf_dataloader):
for key in sample:
sample[key] = sample[key].to(const.device)
output = net(sample)
# all_lines = evaluator.add(output, sample)
# gt = all_lines[0][0]
# pred = all_lines[0][1]
# correct = all_lines[0][-1]
category_type = sample['category_type'][0].cpu().numpy()
# lm_size = int(output['lm_pos_map'].shape[2])
writer.add_figure(
'heatmaps/{}'.format(const.lm2name[i]), fig, step)
# save and evaluate model
if (sample_idx + 1) == total_step:
# if (sample_idx + 1)%5 == 0:
print('\nSaving Model....')
net.set_buffer('step', step)
torch.save(net.state_dict(), 'models/' + const.MODEL_NAME)
print('OK.')
if const.VAL_WHILE_TRAIN:
print('Now Evaluate..')
with torch.no_grad():
net.eval()
evaluator = const.EVALUATOR(category_topk=(1, 2,
3),
attr_topk=(1, 2, 3))
for j, sample in enumerate(val_dataloader):
# move samples to GPU/CPU
for key in sample:
sample[key] = sample[key].to(const.device)
# perform inference
output = net(sample)
# add result to evaluator
evaluator.add(output, sample)
if (j + 1) % 100 == 0:
print('Val Step [{}/{}]'.format(
j + 1, val_step))
# get result from evaluator