Some weights and baises are missing in the graphunet pre-trained model

[Almadani92]

Hi bardiadoosti,

Thanks for updating the trained model and sharing it again. However, there are some missing weights in the shared model. I run the updated testing code and get the following error:

RuntimeError: Error(s) in loading state_dict for DataParallel: Missing key(s) in state_dict: "module.pool1.fc.weight", "module.pool1.fc.bias", "module.pool2.fc.weight", "module.pool2.fc.bias", "module.pool3.fc.weight", "module.pool3.fc.bias", "module.pool4.fc.weight", "module.pool4.fc.bias", "module.pool5.fc.weight", "module.pool5.fc.bias".

It seems to me that the pooling 1, 2, 3, 4, and 5 layers weights and biases are missing in the updated trained model. Please, let me know if I am wrong.

Thanks.

[hedjm]

@Almadani92 Did you try this code, I replied to you in your comment previously.

if args.pretrained_model != '':
    loaded_model = torch.load(args.pretrained_model)
    resnet = {}
    graphnet = {}
    graphunet = {}
    for key, val in loaded_model.items():
        if 'resnet' in key:
            resnet[key.replace('module.resnet.', '')] = val
        if 'graphnet' in key:
            graphnet[key.replace('module.graphnet.', '')] = val
        if 'graphunet' in key:
            graphunet[key.replace('module.graphunet.', '')] = val
    model.resnet.load_state_dict(resnet)
    model.graphnet.load_state_dict(graphnet)
    model.graphunet.load_state_dict(graphunet)
    losses = np.load(args.pretrained_model[:-4] + '-losses.npy').tolist()
    start = len(losses)
else:
    losses = []
    start = 0

[Almadani92]

@Almadani92 Did you try this code, I replied to you in your comment previously.

if args.pretrained_model != '':
    loaded_model = torch.load(args.pretrained_model)
    resnet = {}
    graphnet = {}
    graphunet = {}
    for key, val in loaded_model.items():
        if 'resnet' in key:
            resnet[key.replace('module.resnet.', '')] = val
        if 'graphnet' in key:
            graphnet[key.replace('module.graphnet.', '')] = val
        if 'graphunet' in key:
            graphunet[key.replace('module.graphunet.', '')] = val
    model.resnet.load_state_dict(resnet)
    model.graphnet.load_state_dict(graphnet)
    model.graphunet.load_state_dict(graphunet)
    losses = np.load(args.pretrained_model[:-4] + '-losses.npy').tolist()
    start = len(losses)
else:
    losses = []
    start = 0

Many thanks @hedjm,

Actually, it seems to me that really there are missing weights and biases for all pooling layers in the new pre-trained graphunet model. I run the code, unfortunately it didn't work.

[bardiadoosti]

Oh it seems that I mistakenly trained a wrong model. Please give me a couple of days and I will publish the correct weights.

[Almadani92]

Oh it seems that I mistakenly trained a wrong model. Please give me a couple of days and I will publish the correct weights.

Many thanks @bardiadoosti for your efforts. Actually, it would be better if you publish the whole model weight parameters, I mean for the whole HopeNet model, not just for the GraphuNet, if it is possible.

Thanks in advance.

[Almadani92]

Dear @bardiadoosti ,

Thanks for posting the GraphUnet pre-trained model. However, could you please let me know how did you calculate the average error noted in the paper. Since I have calculated and found it equal to 20 mm for the test dataset.

Thanks.

[bardiadoosti]

Yes it is correct. To quickly share the weights I just trained for 2K iterations. The full training cycle is 10K.

[Almadani92]

Thanks @bardiadoosti ,

I think the training code contains 5K training iterations with learning rate equal to 0.001. did you reduce the learning rate while you were training the new network, such that you have to train for 10K iterations? Also, did you change any other parameters while training the new network?

Thanks in advance.

[payne53]

Yes it is correct. To quickly share the weights I just trained for 2K iterations. The full training cycle is 10K. Dear @bardiadoosti , Thanks for posting the GraphUnet pre-trained model. However, do the rest model weight parameters remain the same as previous version? I am a little confused about the training strategy. Do you train the model end-to-end or just train the graphunet? Thanks in advance.

[Almadani92]

Yes it is correct. To quickly share the weights I just trained for 2K iterations. The full training cycle is 10K. Dear @bardiadoosti , Thanks for posting the GraphUnet pre-trained model. However, do the rest model weight parameters remain the same as previous version? I am a little confused about the training strategy. Do you train the model end-to-end or just train the graphunet? Thanks in advance.

Dear @payne53 the author posted the model corresponding to the last part of the whole Hopenet model, such that it contains the graphunet network (as it is shown in figure 3 in the paper). The input he used in the training is the predicted 2d points which can be extracted via the camera projection matrix, so he did not use the rgb images in this training.

However @bardiadoosti , I managed to train the whole Hopenet network using the code provided. I trained it for 10000 epochs. Unfortunately, I got an Average Error of 21.24 mm. Is the provided training code is the one you are using or there are some modifications.

Thanks and regards,

Murad.

[zc-alexfan]

Just wondering if there is any updates on this thread. Thanks