def test_a_epoch(name, data, model, result_file, label2id, lr, batch_size, en):
model.training = False
full_img_paths = []
evaluator = Evaluator(name, label2id)
for images, labels, img_paths in tqdm(data,
desc=name,
total=np.math.ceil(len(data))):
inp = Variable(images.cuda())
seq_out = model(inp)
pred = argmax(seq_out)
evaluator.append_data(0, pred, labels)
# print(predicted)
full_img_paths.extend(img_paths)
evaluator.gen_results()
evaluator.print_results()
evaluator.write_results(
result_file, "epoch = {2}; GOOGLE NET; lr={0}; batch_size={1}".format(
lr, batch_size, en))
return evaluator, full_img_paths
def train_a_epoch(name, data, model, optimizer, criterion, res_file, lr,
batch_size, label2id, en):
model = model.train(True)
evaluator = Evaluator(name, label2id)
print("evaluator loaded")
i = 0
pbar = tqdm(data, desc=name, total=np.math.ceil(len(data)))
for images, labels, _ in pbar:
# zero the parameter gradients
sys.stdout.flush()
optimizer.zero_grad()
if i == 0:
pbar.total = np.math.ceil(len(data))
inp = Variable(images.cuda())
seq_out = model(inp)
pred = argmax(seq_out)
loss = criterion(seq_out, Variable(torch.cuda.LongTensor(labels)))
pbar.set_description("{0} loss: {1}".format(name, to_scalar(loss)))
evaluator.append_data(to_scalar(loss), pred, labels)
loss.backward()
optimizer.step()
i += 1
print("Training Done")
evaluator.gen_results()
evaluator.print_results()
evaluator.write_results(
res_file, "epoch = {2}; GOOGLE NET; lr={0}; batch_size={1}".format(
lr, batch_size, en))
return model
def test_a_epoch(name, data, model, result_file, cfg, en):
pred_list = []
true_list = []
evaluator = Evaluator(name,
label2id=cfg['LABEL_2_ID'],
pure_labels=cfg['PURE_LABELS'])
for images, labels in tqdm(data, desc=name, total=np.math.ceil(len(data))):
seq_out = model(Variable(torch.from_numpy(images).float().cuda()))
pred = argmax(seq_out)
evaluator.append_data(0, pred, labels)
# print(predicted)
pred_list.extend(pred)
true_list.extend(labels)
evaluator.gen_results()
evaluator.print_results()
evaluator.write_results(
result_file, "epoch = {2}; GOOGLE NET; lr={0}; batch_size={1}".format(
cfg['LEARNING_RATE'], cfg['BATCH_SIZE'], en))
return evaluator, pred_list, true_list
def test_a_epoch(name, data, model, result_file, label2id):
model.training = False
pred_list = []
true_list = []
evaluator = Evaluator(name, label2id)
for images, labels in tqdm(data, desc=name, total=np.math.ceil(len(data))):
inp = Variable(images.cuda())
seq_out = model(inp)
pred = argmax(seq_out)
evaluator.append_data(0, pred, labels)
# print(predicted)
pred_list.extend(pred)
true_list.extend(labels)
evaluator.gen_results()
evaluator.print_results()
evaluator.write_results(result_file, '')
return evaluator, pred_list, true_list
def train_a_epoch(name, data, model, optimizer, criterion, res_file, lr,
batch_size, label2id, en):
model.training = True
evaluator = Evaluator(name, label2id)
print("evaluator loaded")
i = 0
for images, labels, _ in tqdm(data,
desc=name,
total=np.math.ceil(len(data))):
# zero the parameter gradients
sys.stdout.flush()
optimizer.zero_grad()
model.zero_grad()
# image = crop(sample.image, cfg.MAX_IMG_SIZE)
inp = Variable(images.cuda())
# inp = torch.transpose(inp, 1, 3)
seq_out = model(inp)
pred = argmax(seq_out)
loss = criterion(seq_out, Variable(torch.cuda.LongTensor(labels)))
evaluator.append_data(to_scalar(loss), pred, labels)
loss.backward()
optimizer.step()
i += 1
print("Training Done")
evaluator.gen_results()
evaluator.print_results()
evaluator.write_results(
res_file, "epoch = {2}; GOOGLE NET; lr={0}; batch_size={1}".format(
lr, batch_size, en))
return model