关于测试问题

[yekanyeeast]

请问作者，我测试的时候会出现分布式训练报错的问题，导致没有结果生成，把关于分布式代码注释掉也没有生成结果，该如何解决这一问题

[AlonzoLeeeooo]

请问作者，我测试的时候会出现分布式训练报错的问题，导致没有结果生成，把关于分布式代码注释掉也没有生成结果，该如何解决这一问题

@yekanyeeast 您好，感谢关注我们的工作。可以提供一下报错日志吗？我试试复现这个错误。

[yekanyeeast]

[AlonzoLeeeooo]

检查一下--images这个参数所对应的路径中存放的文件是否为空，这个应该是由于你的路径中存放的文件有问题，导致124行至224行之间的部分，在迭代image_flist时被跳过了，如下：

SketchRefiner/SIN_test.py

Lines 124 to 224 in bddc77e

	with torch.no_grad():
	for i in range(len(image_flist)):
	if i >= args.num_samples:
	print('\n\nTesting done...\n\n')
	sys.exit(0)
	image = cv2.imread(image_flist[i])
	filename = os.path.basename(image_flist[i]).split('.')[0]
	mask = cv2.imread(os.path.join(args.masks, filename + '.png'))
	edge = cv2.imread(os.path.join(args.edges, filename + '.png'))
	sketch = cv2.imread(os.path.join(args.sketches, filename + '.png'))
	# read refined sketch directly from local path
	refined_sketch = cv2.imread(os.path.join(args.refined_sketches, filename + '.png'))
	refined_sketch = cv2.resize(refined_sketch, [args.size, args.size])
	refined_sketch = refined_sketch / 255.
	_, refined_sketch = cv2.threshold(refined_sketch, thresh=random.uniform(0.65, 0.75), maxval=1.0, type=cv2.THRESH_BINARY)
	refined_sketch = torch.from_numpy(refined_sketch.astype(np.float32)).permute(2, 0, 1).unsqueeze(0).contiguous()
	refined_sketch = refined_sketch.to(device)
	# resize
	image = cv2.resize(image, [args.size, args.size])
	mask = cv2.resize(mask, [args.size, args.size])
	edge = cv2.resize(edge, [args.size, args.size])
	sketch = cv2.resize(sketch, [args.size, args.size])
	# normalize to [0, 1]
	image = image / 255.
	mask = mask / 255.
	edge = edge / 255.
	sketch = sketch / 255.
	# binarize
	thresh = random.uniform(0.65, 0.75)
	_, mask = cv2.threshold(mask, thresh=0.5, maxval=1.0, type=cv2.THRESH_BINARY)
	_, edge = cv2.threshold(edge, thresh=thresh, maxval=1.0, type=cv2.THRESH_BINARY)
	_, sketch = cv2.threshold(sketch, thresh=thresh, maxval=1.0, type=cv2.THRESH_BINARY)
	# to tensor
	image = torch.from_numpy(image.astype(np.float32)).permute(2, 0, 1).unsqueeze(0).contiguous()
	mask = torch.from_numpy(mask.astype(np.float32)).permute(2, 0, 1).unsqueeze(0).contiguous()
	edge = torch.from_numpy(edge.astype(np.float32)).permute(2, 0, 1).unsqueeze(0).contiguous()
	sketch = torch.from_numpy(sketch.astype(np.float32)).permute(2, 0, 1).unsqueeze(0).contiguous()
	# deform sketch
	deformed_sketch = sketch.detach() # [B=1, C=3, H, W]
	# single channel version of elements
	sc_edge = torch.sum(edge / 3, dim=1, keepdim=True)
	sc_sketch = torch.sum(sketch / 3, dim=1, keepdim=True)
	sc_mask = torch.sum(mask / 3, dim=1, keepdim=True)
	# to cuda
	image = image.to(device)
	mask = mask.to(device)
	edge = edge.to(device)
	sc_sketch = sc_sketch.to(device)
	sc_edge = sc_edge.to(device)
	sc_mask = sc_mask.to(device)
	deformed_sketch = deformed_sketch.to(device)
	# forward
	refined_sketch_feature = str_encoder(refined_sketch)
	deformed_sketch_feature = str_encoder(deformed_sketch)
	# forward with refined sketch
	masked_image = image * (1 - mask)
	out_with_refined_sketch = model(torch.cat([masked_image, sc_mask], dim=1), refined_sketch_feature)
	inpainted_with_refined_sketch = mask * out_with_refined_sketch + (1 - mask) * image
	# forward with deformed sketch
	out_with_deformed_sketch = model(torch.cat([masked_image, sc_mask], dim=1), deformed_sketch_feature)
	inpainted_with_deformed_sketch = mask * out_with_deformed_sketch + (1 - mask) * image
	# save sample of sketch
	inpainted_with_deformed_sketch *= 255.0
	inpainted_with_deformed_sketch = inpainted_with_deformed_sketch.permute(0, 2, 3, 1).cpu().detach().numpy().astype(np.uint8)
	cv2.imwrite(args.output + '/' + 'inpainted_with_sketch' + '/' + filename + '.png', inpainted_with_deformed_sketch[0, :, :, :,])
	# save sample of refined sketch
	inpainted_with_refined_sketch *= 255.0
	inpainted_with_refined_sketch = inpainted_with_refined_sketch.permute(0, 2, 3, 1).cpu().detach().numpy().astype(np.uint8)
	cv2.imwrite(args.output + '/' + 'inpainted_with_refined_sketch' + '/' + filename + '.png', inpainted_with_refined_sketch[0, :, :, :,])
	# make data dict
	data = {
	'filename': filename + '.png',
	'image': image,
	'masked_image': masked_image + deformed_sketch * mask,
	'deformed_sketch': deformed_sketch * mask + (1 - mask) * edge,
	'refined_sketch': refined_sketch,
	'edge': edge,
	'inpainted_with_deformed_sketch': mask * out_with_deformed_sketch + (1 - mask) * image,
	'inpainted_with_refined_sketch': mask * out_with_refined_sketch + (1 - mask) * image,
	}
	visualize(data, ['image', 'masked_image', 'deformed_sketch', 'refined_sketch', 'inpainted_with_deformed_sketch', 'inpainted_with_refined_sketch'], os.path.join(args.output, 'samples'))
	print(f"Progress: {i + 1}/{len(image_flist)}")

所以直接运行了227行的cleanup()：

SketchRefiner/SIN_test.py

Line 227 in bddc77e

cleanup()