error in step 2 training

[fjremnav]

I am able to train based model. However, it encounters an error in step2 training below when it tries to load PSPNet.pt from step 1.

python train.py --model_name PSPNet --lstm True --use_pre True --noise False --data_dir cityscapes/

Traceback (most recent call last):
File "train.py", line 108, in
train(args)
File "train.py", line 78, in train
init_model = load_model(args)
File "/home/user/4TB/user1/NoisyLSTM/tools/tool.py", line 285, in load_model
pre_model.load_state_dict(new_state_dict, strict=True)
File "/home/user1/anaconda3/envs/remnav/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1045, in load_state_dict
self.class.name, "\n\t".join(error_msgs)))
RuntimeError: Error(s) in loading state_dict for PspNet:
Missing key(s) in state_dict: "layer0.0.weight", "layer0.1.weight", "layer0.1.bias", "layer0.1.running_mean", "layer0.1.running_var", "layer1.0.conv1.
weight", "layer1.0.bn1.weight", "layer1.0.bn1.bias", "layer1.0.bn1.running_mean", "layer1.0.bn1.running_var", "layer1.0.conv2.weight", "layer1.0.bn2.weight",
"layer1.0.bn2.bias", "layer1.0.bn2.running_mean", "layer1.0.bn2.running_var", "layer1.0.conv3.weight", "layer1.0.bn3.weight", "layer1.0.bn3.bias", "layer1.0.b ,,,,,

Any idea why this happens

Thanks,

[wbw520]

Hi,

I think this is a bug by multi-gpu setting. I run the code only with multi-gpu.
If you pre-train the model in one gpu with "--multi_gpu False", one solution is to change the code in
tools/tool.py line 283 "name = k[7:]" to "name = k" . (The former one is used to remove the prefix of saved model parameters caused by nn.DataParallel. It not happens in single gpu.)
Hope it can help you.

good day,

[fjremnav]

It works with your answer.
One more question, how do I test with my own images or videos with test.py?

Thanks,

[wbw520]

The test.py is actually used for evaluate mAP.
If you want to try to visulaize your own image or video, please used the tools/factory.py.
I have updated a new version for it. (But it write for cityscapes (take city of munster as a sample), you may have to change the root setting for you own data)

[fjremnav]

I read factory.py and see groundtruth in the code. Unfortunately, my images/videos do not have ground truths. Could I just comment out groundtruth related codes?

Thanks,

[wbw520]

It is totally ok.

[fjremnav]

My trained model performs quite bad. Do you mind sharing your trained model (PSPNet and PSPNet_Llst_noise) which I can use it as a refence?

Thanks,

[wbw520]

You can download the weight in the following link
https://drive.google.com/file/d/1R2yPUdGUudr3ADTNdsrUgeXye0QGfdwN/view?usp=sharing