Transfer weights from monai's hovernet to original author's hovernet

[rajeshtims]

If I use monailabel to train Hovernet, how can I copy these weights to the original author's model (state dict mapping)?
The state dictionary mapping is quite strange here; names are completely different and not sure which key in monai's hovernet corresponds to which key in the original author's model.

[KumoLiu]

Hi @rajeshtims, I think you could refer to these mapping rules.

MONAI/monai/networks/nets/hovernet.py

Lines 623 to 650 in 0ffe55a

	def _remap_preact_resnet_model(model_url: str):
	pattern_conv0 = re.compile(r"^(conv0\.\/)(.+)$")
	pattern_block = re.compile(r"^(d\d+)\.(.+)$")
	pattern_layer = re.compile(r"^(.+\.d\d+)\.units\.(\d+)(.+)$")
	pattern_bna = re.compile(r"^(.+\.d\d+)\.blk_bna\.(.+)")
	# download the pretrained weights into torch hub's default dir
	weights_dir = os.path.join(torch.hub.get_dir(), "preact-resnet50.pth")
	download_url(model_url, fuzzy=True, filepath=weights_dir, progress=False)
	state_dict = torch.load(weights_dir, map_location=None if torch.cuda.is_available() else torch.device("cpu"))[
	"desc"
	]
	for key in list(state_dict.keys()):
	new_key = None
	if pattern_conv0.match(key):
	new_key = re.sub(pattern_conv0, r"conv0.conv\2", key)
	elif pattern_block.match(key):
	new_key = re.sub(pattern_block, r"res_blocks.\1.\2", key)
	if pattern_layer.match(new_key):
	new_key = re.sub(pattern_layer, r"\1.layers.denselayer_\2.layers\3", new_key)
	elif pattern_bna.match(new_key):
	new_key = re.sub(pattern_bna, r"\1.bna_block.\2", new_key)
	if new_key:
	state_dict[new_key] = state_dict[key]
	del state_dict[key]
	if "upsample2x" in key:
	del state_dict[key]
	return state_dict

Hope it can help you, thanks!

[rajeshtims]

Thanks for the quick reply. I am a bit confused about the mapping of the decoder. Could you please provide me some information about how to get the decoder mapping?

[KumoLiu]

Hi @rajeshtims, we don't have the decoder mapping, maybe you need to manually export the structure of the network and compare them.
Thanks!