Pre-trained model in Pytorch Version about dexined

[sooperset]

@ZhangYuef Thanks for sharing!
Not sure why this error is popped up like

model = DexiNet()
model.load_state_dict(torch.load('24_model.pth'))

---

RuntimeError Traceback (most recent call last)
in
1 model = DexiNet()
----> 2 model.load_state_dict(checkpoint)
3 print('done')

/opt/conda/lib/python3.7/site-packages/torch/nn/modules/module.py in load_state_dict(self, state_dict, strict)
1043 if len(error_msgs) > 0:
1044 raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format(
-> 1045 self.class.name, "\n\t".join(error_msgs)))
1046 return _IncompatibleKeys(missing_keys, unexpected_keys)
1047

RuntimeError: Error(s) in loading state_dict for DexiNet:
size mismatch for dblock_3.denselayer1.conv1.weight: copying a param with shape torch.Size([256, 128, 3, 3]) from checkpoint, the shape in current model is torch.Size([256, 128, 1, 1]).
size mismatch for dblock_3.denselayer2.conv1.weight: copying a param with shape torch.Size([256, 256, 3, 3]) from checkpoint, the shape in current model is torch.Size([256, 256, 1, 1]).
size mismatch for dblock_4.denselayer1.conv1.weight: copying a param with shape torch.Size([512, 256, 3, 3]) from checkpoint, the shape in current model is torch.Size([512, 256, 1, 1]).
size mismatch for dblock_4.denselayer2.conv1.weight: copying a param with shape torch.Size([512, 512, 3, 3]) from checkpoint, the shape in current model is torch.Size([512, 512, 1, 1]).
size mismatch for dblock_4.denselayer3.conv1.weight: copying a param with shape torch.Size([512, 512, 3, 3]) from checkpoint, the shape in current model is torch.Size([512, 512, 1, 1]).
size mismatch for dblock_5.denselayer1.conv1.weight: copying a param with shape torch.Size([512, 512, 3, 3]) from checkpoint, the shape in current model is torch.Size([512, 512, 1, 1]).
size mismatch for dblock_5.denselayer2.conv1.weight: copying a param with shape torch.Size([512, 512, 3, 3]) from checkpoint, the shape in current model is torch.Size([512, 512, 1, 1]).
size mismatch for dblock_5.denselayer3.conv1.weight: copying a param with shape torch.Size([512, 512, 3, 3]) from checkpoint, the shape in current model is torch.Size([512, 512, 1, 1]).
size mismatch for dblock_6.denselayer1.conv1.weight: copying a param with shape torch.Size([256, 512, 3, 3]) from checkpoint, the shape in current model is torch.Size([256, 512, 1, 1]).
size mismatch for dblock_6.denselayer2.conv1.weight: copying a param with shape torch.Size([256, 256, 3, 3]) from checkpoint, the shape in current model is torch.Size([256, 256, 1, 1]).
size mismatch for dblock_6.denselayer3.conv1.weight: copying a param with shape torch.Size([256, 256, 3, 3]) from checkpoint, the shape in current model is torch.Size([256, 256, 1, 1]).

But, I changed the first convolution layer of _DenseLayer with kernel=3 and padding=1, then I could successfully load the checkpoint.

class _DenseLayer(nn.Sequential):
    def __init__(self, input_features, out_features):
        super(_DenseLayer, self).__init__()

        # self.add_module('relu2', nn.ReLU(inplace=True)),
        self.add_module('conv1', nn.Conv2d(input_features, out_features,
                                           kernel_size=3, stride=1, padding=1, bias=True)),
        self.add_module('norm1', nn.BatchNorm2d(out_features)),
        self.add_module('relu1', nn.ReLU(inplace=True)),
        self.add_module('conv2', nn.Conv2d(out_features, out_features,
                                           kernel_size=3, stride=1, padding=1, bias=True)),
        self.add_module('norm2', nn.BatchNorm2d(out_features))
        # double check the norm1 comment if necessary and put norm after conv2

    def forward(self, x):
        x1, x2 = x
        # maybe I should put here a RELU
        new_features = super(_DenseLayer, self).forward(F.relu(x1))  # F.relu()
        return 0.5 * (new_features + x2), x2

[ZhangYuef]

This is a checkpoint from epoch 24 trained on BIPED dataset. Download link is here.
Training result is as the following:

Hope this unofficial one can help : )

i have placed this file in checkpoint folder and in 24/ i am getting this error

Error(s) in loading state_dict for DexiNed:
Missing key(s) in state_dict: "block_cat.conv.weight", "block_cat.conv.bias", "block_cat.bn.weight", "block_cat.bn.bias", "block_cat.bn.running_mean", "block_cat.bn.running_var".
Unexpected key(s) in state_dict: "block_cat.weight", "block_cat.bias".

Sorry to forget mention, this checkpoint is trained based on the model of commit a9f6ade. Link from here.

[xavysp]

Hi @LionXFQ sorry I will update the DexiNed-Pytorch weights soon, I am preparing the DexiNed extension in pytorch version

[ZhangYuef]

Also looking forward to the pretrained weights for Pytorch model.

[Palzer]

+1 here too.

[xavysp]

Sorry guys it will take a little longer, I cannot turn on my PC in the LAB, working remotely. But hope solve this issue in the coming days and update the DexiNed-Pytorch checkpoint

Cheers.

Xavier

[ZhangYuef]

Sorry guys it will take a little longer, I cannot turn on my PC in the LAB, working remotely. But hope solve this issue in the coming days and update the DexiNed-Pytorch checkpoint

Cheers.

Xavier

I have trained one on BIPED dataset myself. I can share if it may help anyone.

[anandhu1436]

Sorry guys it will take a little longer, I cannot turn on my PC in the LAB, working remotely. But hope solve this issue in the coming days and update the DexiNed-Pytorch checkpoint
Cheers.
Xavier

I have trained one on BIPED dataset myself. I can share if it may help anyone.

please share

[ZhangYuef]

This is a checkpoint from epoch 24 trained on BIPED dataset. Download link is here.

BTW, this checkpoint is trained based on the model of commit. Link here.

Training result is as the following:

Hope this unofficial one can help : )

[anandhu1436]

This is a checkpoint from epoch 24 trained on BIPED dataset. Download link is here.

Training result is as the following:

Hope this unofficial one can help : )

i have placed this file in checkpoint folder and in 24/ i am getting this error

Error(s) in loading state_dict for DexiNed:
Missing key(s) in state_dict: "block_cat.conv.weight", "block_cat.conv.bias", "block_cat.bn.weight", "block_cat.bn.bias", "block_cat.bn.running_mean", "block_cat.bn.running_var".
Unexpected key(s) in state_dict: "block_cat.weight", "block_cat.bias".

[ZhangYuef]

Thanks @soomiles for pointing out ! I did change the model layer as you point out as the official code had dimension error at that time. : )

Let's wait for the official one!

[priyamdey]

Hi. Is the official PyTorch pretrained model available yet?

[xavysp]

Hi. Is the official PyTorch pretrained model available yet?

Hi sorry, now you can find the official one.

[ManuBN786]

Sorry guys it will take a little longer, I cannot turn on my PC in the LAB, working remotely. But hope solve this issue in the coming days and update the DexiNed-Pytorch checkpoint
Cheers.
Xavier

I have trained one on BIPED dataset myself. I can share if it may help anyone.

Please do. and send the link. Thanks