how can I get box or TSR and TSSR when predict not eval？

[TheHonestBob]

No description provided.

[hsl323]

I apologize for the late response.

The box is the content that should be given for inference. You can use the OCR information that exists in the dataset or use your own box.

As you can see in Instructions for proposed method in the readme, the inference code has the argument attention_vis. You can get TSR, and TSSR using this argument.

If you want to get something that is not synthesized with the image for visualization, you can use the model's seventh output, att_set

garnet/CODE/inference.py

Lines 106 to 121 in c5a7adb

	_, _, _, _, result, _, att_set, att = net(x)
	# save result
	result = (1 - box_mask) * im + box_mask * result.cpu()
	img = (torch.clamp(result[0] + 1, 0, 2) * 127.5).cpu().detach().numpy().transpose(1, 2, 0).astype(np.uint8)
	img = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)
	img = cv2.resize(img, (H, W))
	cv2.imwrite(os.path.join(args.result_path, name), img)
	if args.attention_vis:
	stroke_att = []
	surround_att = []
	gate_att = []
	for i in range(0, 4):
	stroke_att_map = attention_visualize(result.cpu(), att_set[i][0], args.input_size)
	surround_att_map = attention_visualize(result.cpu(), att_set[i][1], args.input_size)

att_set[i][0]: TSR on the i-th layer
att_set[i][1]: TSSR on the i-th layer

[TheHonestBob]

thanks for your reply, I will try it.