# Forward + Backward Prop
optimizer.zero_grad()
torch.set_grad_enabled(True)
outputs = model.forward(inputs)
predictions = torch.max(outputs, 1)[1]
loss = criterion(outputs, labels.long().squeeze(1))
loss.backward()
optimizer.step()
# statistics
running_loss += loss.item()
writer.add_scalar('data/Train BatchWise Loss', loss.item(),
total_iter_num)
_total_iou, per_class_iou, num_images_per_class = utils.get_iou(
predictions, labels, n_classes=config.train.numClasses)
total_iou += _total_iou
# Print loss every 20 Batches
# if (iter_num % 20) == 0:
# print('Epoch{} Batch{} BatchLoss: {:.4f} '.format(epoch, iter_num, loss.item()))
# Log Epoch Loss
epoch_loss = running_loss / (len(trainLoader))
writer.add_scalar('data/Train Epoch Loss', epoch_loss, total_iter_num)
print('\nTrain Epoch Loss: {:.4f}'.format(epoch_loss))
# Log mIoU
miou = total_iou / ((len(trainLoader)) * config.train.batchSize)
writer.add_scalar('data/Train mIoU', miou, total_iter_num)
print('Train mIoU: {:.4f}'.format(miou))
total_iou = 0.0
for ii, sample_batched in enumerate(tqdm(testLoader)):
inputs, labels = sample_batched
# Forward pass of the mini-batch
inputs = inputs.to(device)
labels = labels.to(device)
with torch.no_grad():
outputs = model(inputs)
predictions = torch.max(outputs, 1)[1]
_total_iou, per_class_iou, num_images_per_class = utils.get_iou(
predictions,
labels.long().squeeze(1),
n_classes=config.train.numClasses)
total_iou += _total_iou
# Save output images, one at a time, to results
img_tensor = inputs.detach().cpu()
output_tensor = outputs.detach().cpu()
label_tensor = labels.detach().cpu()
# Extract each tensor within batch and save results
for iii, sample_batched in enumerate(
zip(img_tensor, output_tensor, label_tensor)):
image, output, label = sample_batched
result_path = os.path.join(
results_dir, SUBDIR_RESULT,
def _ocr_extend_dataset(self, data, path):
paths = glob.glob(f'{path}/**/ocr_output.json', recursive=True)
for path_ in paths:
with open(path_, "r") as f:
_data = json.load(f)
name = Path(path_).parent.stem
# load ground truth
gt_name = f"{name}.json"
if gt_name not in data:
continue
gt_dict = data[gt_name]
data_key = f"ocr_extend_{name}.json"
texts = []
bboxes_for_train = []
input_bboxes = []
keys_type = []
formal_keys = []
match = 0
not_match = 0
# _data is list
for label in _data:
text = str(label["text"])
text = text.replace('\n', '').replace(' ', '')
match_index = -1
match_value = 0
now_box_list = [*label["location"][0], *label["location"][2]]
now_box = {
key: value
for key, value in zip(["x1", "y1", "x2", "y2"],
now_box_list)
}
for i, bbox_for_train in enumerate(gt_dict["bboxes"]):
test_box = {
key: value
for key, value in zip(["x1", "y1", "x2", "y2"],
bbox_for_train)
}
# LOGGER.info(now_box, test_box)
iou = get_iou(now_box, test_box)
if iou > 0.5 and iou > match_value:
match_index = i
match_value = iou
if match_index != -1:
keys_type.append(gt_dict["keys_type"][match_index])
formal_keys.append(gt_dict["formal_keys"][match_index])
else:
# LOGGER.info("didn't find match")
keys_type.append("")
formal_keys.append("")
# construct the graph
input_bbox_format, bbox_for_train = [
*now_box_list, 1, 1, text, None
], now_box_list
texts.append(text)
bboxes_for_train.append(bbox_for_train)
input_bboxes.append(input_bbox_format)
data[data_key]["image_file"] = gt_dict["image_file"]
data[data_key]["keys_type"] = keys_type
data[data_key]["formal_keys"] = formal_keys
data[data_key]["category"] = gt_dict["category"]
# data[key]["text"] = "".join(texts)
data[data_key]["texts"] = texts
data[data_key]["bboxes"] = bboxes_for_train
data[data_key]["input_bboxes"] = input_bboxes