Key mismatch while loading the model?

[Jumabek]

I am having issue loading the trained checkpoint to FPNSSD512 model. How can I fix that?

RuntimeError: Error(s) in loading state_dict for FPNSSD512:
	Missing key(s) in state_dict: "fpn.conv1.weight", "fpn.bn1.running_var", "fpn.bn1.bias", "fpn.bn1.running_mean", "fpn.bn1.weight", "fpn.layer1.0.conv1.weight", "fpn.layer1.0.bn1.running_var", "fpn.layer1.0.bn1.bias", "fpn.layer1.0.bn1.running_mean", "fpn.layer1.0.bn1.weight", "fpn.layer1.0.conv2.weight", "fpn.layer1.0.bn2.running_var", "fpn.layer1.0.bn2.bias", 

        Unexpected key(s) in state_dict: "module.fpn.conv1.weight", "module.fpn.bn1.weight", "module.fpn.bn1.bias", "module.fpn.bn1.running_mean", "module.fpn.bn1.running_var", "module.fpn.layer1.0.conv1.weight"

[Jumabek]

following code before loading the checkpoint solved the issue

if device == 'cuda':
    net = torch.nn.DataParallel(net)
    cudnn.benchmark = True

[ahkarami]

Dear @Jumabek, I have also your reported issue. my script is something like this:

import torch
import torch.backends.cudnn as cudnn
from models.fpnssd.net import FPNSSD512


# Print the PyTorch Version:
print(torch.__version__)  # 0.4.0


# *************** Parameters **************** #
# Check use GPU or not
use_gpu = torch.cuda.is_available()  # use GPU
if use_gpu:
    device = torch.device("cuda:0")  
else:
    device = torch.device("cpu")


# ** Loading Pre-Trained Weights:
net = FPNSSD512(num_classes=20).to(device)
net = torch.nn.DataParallel(net)
cudnn.benchmark = True
# download pre-trained weights from:
# https://drive.google.com/open?id=1yy_kUnm_hZR3uk9yLcaQSMwxVn7wApTU
net.load_state_dict(torch.load('./fpnssd512_20_trained.pth'))
net.eval()

However, I got your reported error. Would you please help me to address this issue?

[ahkarami]

Dear @kuangliu, Would you please answer my above question?

[Jumabek]

@ahkarami sorry for late reply. While I do not fully understand the issue. Can you run the code below: I added net = torch.nn.DataParallel(net) after loading the model

import torch
import torch.backends.cudnn as cudnn
from models.fpnssd.net import FPNSSD512


# Print the PyTorch Version:
print(torch.__version__)  # 0.4.0


# *************** Parameters **************** #
# Check use GPU or not
use_gpu = torch.cuda.is_available()  # use GPU
if use_gpu:
    device = torch.device("cuda:0")  
else:
    device = torch.device("cpu")


# ** Loading Pre-Trained Weights:
net = FPNSSD512(num_classes=20).to(device)
net = torch.nn.DataParallel(net)
cudnn.benchmark = True
# download pre-trained weights from:
# https://drive.google.com/open?id=1yy_kUnm_hZR3uk9yLcaQSMwxVn7wApTU
net.load_state_dict(torch.load('./fpnssd512_20_trained.pth'))
net = torch.nn.DataParallel(net)
net.eval()

[ahkarami]

Dear @Jumabek, Thank you for your reply. Sorry for my inconvenience. I have tested your recommended script, but unfortunately the error is remain. The error is:

Traceback (most recent call last):
  File "/home/user/TorchCV/Attempt1.py", line 54, in <module>
    net.load_state_dict(torch.load('./fpnssd512_20_trained.pth'))
  File "/opt/pytorch4/torch/nn/modules/module.py", line 721, in load_state_dict
    self.__class__.__name__, "\n\t".join(error_msgs)))
RuntimeError: Error(s) in loading state_dict for FPNSSD512:
	Missing key(s) in state_dict: "fpn.conv1.weight", "fpn.bn1.running_mean", "fpn.bn1.running_var", ...
	Unexpected key(s) in state_dict: "extractor.conv1.weight", "extractor.bn1.weight", "extractor.bn1.bias", ....

Process finished with exit code 1

It is worth nothing that I have tested the above code on a system which has just one GTX 1080ti GPU (with CUDA 9.0 & cuDNN 7).

[dearleiii]

Hi I followed your code and seems helped me solve the issue of unexpected key, But I'm wondering what's the reason for it to occur? Why is DataParallel help to solve it?

[zacario-li]

@ahkarami I meet the same issue with you. Have you fixed it now?

[ahkarami]

Dear @zacario-li, Unfortunately I couldn't address the issue. I can train & test model by my own GPU (i.e., my trained models are correct) but the released pre-trained model has the above issue. I think the problem related to this fact that the pre-trained model has been trained on a machine with multi GPU but now we want to use it in a machine with just one GPU. However, In this case using the torch.nn.DataParallel(net) command must address the problem, but we saw that this command can't solve the problem!!!

[root-master]

If you want to load the weights after DataParallel use: net.module.load_state_dict(pertained_weights) If you want to load the weights before DataParallel use: net.load_state_dict(pertained_weights)

[silkylove]

Dear @ahkarami , I think the pretrained fpnssd model provided by @kuangliu is not the same as /models/fpnssd/net.py. Actually, he said that he just replaced vgg16 by fpn50 in ssd512 which is /models/ssd/net.py. So you could not use the model created by /models/fpnssd/net.py to load the wights in /models/ssd/net.py as the keys are not matched. The solution to use his provided pretrained model is to train his ssd512 model with fpn50 not fpnssd512 model in /models/fpnssd/net.py. Also, it seems that he did not put all of his examples on this github or he delete something before pushing.

[ahkarami]

Dear @silkylove, Thank you very much for your useful information. Could you load & use his pre-trained network? If yes, would you please release its loading code?

[silkylove]

Dear @ahkarami , Ok, I will release the code after I get similar performence compared to his pretrained fpnssd512 model.

[ahkarami]

Thank you very much @silkylove.

[silkylove]

@ahkarami Please check my code. https://github.com/silkylove/ObjectDetection/tree/master/example/fpnssd I also uploaded the training log with adam with 100 epochs which could get 73.95mAP until now. I am now training SGD with 200 epochs on that which I think would get higher mAP, I will release the training log later. Also, you can uncommen this line in eval.py https://github.com/silkylove/ObjectDetection/blob/master/example/fpnssd/eval.py#L25 to got his pertrained model's performence (about 56mAP). And make sure not to use dataparallel.

[ahkarami]

Dear @silkylove, Thank you very much for your time. Your implemented and modified code is really valuable. It would be also great If you upload your pre-trained model (e.g., in Google Drive).

[silkylove]

@ahkarami I uploaded the sgd training and eval log. And with sgd, I can only got aound 76% mAP now. The pretrained model was in here.

[ahkarami]

@silkylove, Thank you very much.