One more issue about the model

[MariaWang96]

One more issue about the model

[Ray-tju]

Hi @MariaWang96
I uploaded the two files you need, I hope it can solve your problem. Note that the file "attention_1d" should add to the dir named "training".

[Ray-tju]

The file https://github.com/leilimaster/MFIRRN/blob/main/model/Mfirrn.py in ./model has one module named 'attention'. I tried to add another module like:

import torch.nn as nn class SELayer(nn.Module): def init(self, channel, reduction=16): super(SELayer, self).init() self.avg_pool = nn.AdaptiveAvgPool2d(1) self.fc = nn.Sequential( nn.Linear(channel, channel // reduction, bias=False), nn.ReLU(inplace=True), nn.Linear(channel // reduction, channel, bias=False), nn.Sigmoid() )

def forward(self, x):
    b, c, _, _ = x.size()
    y = self.avg_pool(x).view(b, c)
    y = self.fc(y).view(b, c, 1, 1)
    return x * y.expand_as(x)

as a .py file and then 'import attention' in Mfirrn.py.

But, when I run the benchmark.py, it can't work.

error: RuntimeError: Error(s) in loading state_dict for LLNet: Unexpected key(s) in state_dict: "attention.fc.0.weight", "attention.fc.2.weight".

Could you give some guidance?
Hi @MariaWang96
I re-uploaded the file named "Mfirrn", please download it again.

[MariaWang96]

The file https://github.com/leilimaster/MFIRRN/blob/main/model/Mfirrn.py in ./model has one module named 'attention'. I tried to add another module like:
import torch.nn as nn class SELayer(nn.Module): def init(self, channel, reduction=16): super(SELayer, self).init() self.avg_pool = nn.AdaptiveAvgPool2d(1) self.fc = nn.Sequential( nn.Linear(channel, channel // reduction, bias=False), nn.ReLU(inplace=True), nn.Linear(channel // reduction, channel, bias=False), nn.Sigmoid() )

def forward(self, x):
    b, c, _, _ = x.size()
    y = self.avg_pool(x).view(b, c)
    y = self.fc(y).view(b, c, 1, 1)
    return x * y.expand_as(x)

as a .py file and then 'import attention' in Mfirrn.py.
But, when I run the benchmark.py, it can't work.
error: RuntimeError: Error(s) in loading state_dict for LLNet: Unexpected key(s) in state_dict: "attention.fc.0.weight", "attention.fc.2.weight".
Could you give some guidance?
Hi @MariaWang96
I re-uploaded the file named "Mfirrn", please download it again.

I put these two attention files in ./training and then 'from training import attention,attention_1d' in Mfirrn.py, it works!
Thanks for your time and help.
I will close this issue.

[Ray-tju]

The file https://github.com/leilimaster/MFIRRN/blob/main/model/Mfirrn.py in ./model has one module named 'attention'. I tried to add another module like:
import torch.nn as nn class SELayer(nn.Module): def init(self, channel, reduction=16): super(SELayer, self).init() self.avg_pool = nn.AdaptiveAvgPool2d(1) self.fc = nn.Sequential( nn.Linear(channel, channel // reduction, bias=False), nn.ReLU(inplace=True), nn.Linear(channel // reduction, channel, bias=False), nn.Sigmoid() )

def forward(self, x):
    b, c, _, _ = x.size()
    y = self.avg_pool(x).view(b, c)
    y = self.fc(y).view(b, c, 1, 1)
    return x * y.expand_as(x)

as a .py file and then 'import attention' in Mfirrn.py.
But, when I run the benchmark.py, it can't work.
error: RuntimeError: Error(s) in loading state_dict for LLNet: Unexpected key(s) in state_dict: "attention.fc.0.weight", "attention.fc.2.weight".
Could you give some guidance?
Hi @MariaWang96
I re-uploaded the file named "Mfirrn", please download it again.

I put these two attention files in ./training and then 'from training import attention,attention_1d' in Mfirrn.py, it works! Thanks for your time and help. I will close this issue.

The file https://github.com/leilimaster/MFIRRN/blob/main/model/Mfirrn.py in ./model has one module named 'attention'. I tried to add another module like:
import torch.nn as nn class SELayer(nn.Module): def init(self, channel, reduction=16): super(SELayer, self).init() self.avg_pool = nn.AdaptiveAvgPool2d(1) self.fc = nn.Sequential( nn.Linear(channel, channel // reduction, bias=False), nn.ReLU(inplace=True), nn.Linear(channel // reduction, channel, bias=False), nn.Sigmoid() )

def forward(self, x):
    b, c, _, _ = x.size()
    y = self.avg_pool(x).view(b, c)
    y = self.fc(y).view(b, c, 1, 1)
    return x * y.expand_as(x)

as a .py file and then 'import attention' in Mfirrn.py.
But, when I run the benchmark.py, it can't work.
error: RuntimeError: Error(s) in loading state_dict for LLNet: Unexpected key(s) in state_dict: "attention.fc.0.weight", "attention.fc.2.weight".
Could you give some guidance?
Hi @MariaWang96
I re-uploaded the file named "Mfirrn", please download it again.

I put these two attention files in ./training and then 'from training import attention,attention_1d' in Mfirrn.py, it works! Thanks for your time and help. I will close this issue.

Hi @MariaWang96

Thanks for your attention to our work!

The file https://github.com/leilimaster/MFIRRN/blob/main/model/Mfirrn.py in ./model has one module named 'attention'. I tried to add another module like:
import torch.nn as nn class SELayer(nn.Module): def init(self, channel, reduction=16): super(SELayer, self).init() self.avg_pool = nn.AdaptiveAvgPool2d(1) self.fc = nn.Sequential( nn.Linear(channel, channel // reduction, bias=False), nn.ReLU(inplace=True), nn.Linear(channel // reduction, channel, bias=False), nn.Sigmoid() )

def forward(self, x):
    b, c, _, _ = x.size()
    y = self.avg_pool(x).view(b, c)
    y = self.fc(y).view(b, c, 1, 1)
    return x * y.expand_as(x)

as a .py file and then 'import attention' in Mfirrn.py.
But, when I run the benchmark.py, it can't work.
error: RuntimeError: Error(s) in loading state_dict for LLNet: Unexpected key(s) in state_dict: "attention.fc.0.weight", "attention.fc.2.weight".
Could you give some guidance?
Hi @MariaWang96
I re-uploaded the file named "Mfirrn", please download it again.

I put these two attention files in ./training and then 'from training import attention,attention_1d' in Mfirrn.py, it works! Thanks for your time and help. I will close this issue.

Hi @MariaWang96
I just re-tested the performance of the model, and the results are shown in the table below. Note that our GPU is Nvidia RTX 3090, and the test environment is cuda V11.1, Pytorch 1.7.

Thank you for your attention to this work！

Extracting params take 2.542s
[ 0, 30] Mean: 2.839, Std: 1.550
[30, 60] Mean: 3.557, Std: 1.669
[60, 90] Mean: 4.572, Std: 2.175
[ 0, 90] Mean: 3.656, Std: 0.711
Extracting params take 13.159s
[ 0, 30] Mean: 4.323, Std: 3.748
[30, 60] Mean: 5.070, Std: 4.933
[60, 90] Mean: 5.962, Std: 6.983
[ 0, 90] Mean: 5.119, Std: 0.670

[MariaWang96]

Actually, I suspected your result on AFLW2000-3D before. In my experiment, when I get 3.678 on AFLW2000-3D, I could get 4.785 on AFLW at the same time.

Before you published the weight file, I retrian your model on my machine following your paper, and evaluate on AFLW2000-3D and AFLW.

When set 'def calc_nme(pts68_fit_all, option='ori'):' in bechmark_aflw2000.py:
the evaluation result is
AFLW2000-3D: 2.974 3.983 5.220 4.059
AFLW: 4.307 5.049 6.059 5.138

When set 'def calc_nme(pts68_fit_all, option='re'):' in bechmark_aflw2000.py:
the evaluation result is
AFLW2000-3D: 2.765 3.226 4.704 3.565
AFLW: 4.307 5.049 6.059 5.138

But now, it seems I'm wrong.

[Ray-tju]

Actually, I suspected your result on AFLW2000-3D before. In my experiment, when I get 3.678 on AFLW2000-3D, I could get 4.785 on AFLW at the same time.

Before you published the weight file, I retrian your model on my machine following your paper, and evaluate on AFLW2000-3D and AFLW.

When set 'def calc_nme(pts68_fit_all, option='ori'):' in bechmark_aflw2000.py: the evaluation result is AFLW2000-3D: 2.974 3.983 5.220 4.059 AFLW: 4.307 5.049 6.059 5.138

When set 'def calc_nme(pts68_fit_all, option='re'):' in bechmark_aflw2000.py: the evaluation result is AFLW2000-3D: 2.765 3.226 4.704 3.565 AFLW: 4.307 5.049 6.059 5.138

But now, it seems I'm wrong.

Replace the "benchmark.py" in the baseline with the "benchmark.py" we just released to get the correct result, but due to the randomness of multi-granular segmentation, the evaluation result will fluctuate in the range of 3.650-3.690.
If you have other questions, you can start the issue again
Thank you for your attention to this work！