Can not load pertrained weights in VGG19

[se7enXF]

New Issue Checklist

• I have read the Contribution Guidelines
• I searched for existing GitHub issues

Issue Description

When I use VGG19 model with pretrained weights, some errors arose. Maybe it is mismatching of
numpy and tensorlayer2 that cause the problem. The weight download process works well.

Reproducible Code

• OS:Windows10
• Python3.6
• numpy1.16.4
• tensorlayer2.1.0

import tensorlayer as tl
cnn = tl.models.vgg19(pretrained=True, end_with='conv5_4')

Here is the error:

Traceback (most recent call last):
  File "D:/TAM_Net/model.py", line 197, in <module>
    test = THZoom([None, image_size, image_size, 3])
  File "D:/TAM_Net/model.py", line 181, in __init__
    self.encoder = tl.models.vgg19(pretrained=True, end_with='conv5_4', mode='static')
  File "C:\Users\dell\Miniconda3\envs\tf2\lib\site-packages\tensorlayer\models\vgg.py", line 317, in vgg19
    restore_model(model, layer_type='vgg19')
  File "C:\Users\dell\Miniconda3\envs\tf2\lib\site-packages\tensorlayer\models\vgg.py", line 169, in restore_model
    npz = np.load(os.path.join('models', model_saved_name[layer_type]), encoding='latin1').item()
  File "C:\Users\dell\Miniconda3\envs\tf2\lib\site-packages\numpy\lib\npyio.py", line 447, in load
    pickle_kwargs=pickle_kwargs)
  File "C:\Users\dell\Miniconda3\envs\tf2\lib\site-packages\numpy\lib\format.py", line 696, in read_array
    raise ValueError("Object arrays cannot be loaded when "
ValueError: Object arrays cannot be loaded when allow_pickle=False

[JingqingZ]

This issue is mentioned by #1019 and has been fixed in PR#1021. Please pull the latest TensorLayer code and modify your code accordingly. Thanks!

[se7enXF]

I use pip installing the TL2 and the version is the latest 2.1.0.
I edit the TF2 source code in python with 'allow_pickle=True' and it works well.
Hope you can add some more useful model in TL, such as ResNet, YOLO. TL is very useful and thank you.

[rinabell]

same problem. and my version is the latest 2.1.1.
but what puzzles me is when I ran codes in tensorlayer/examples/pretrained_cnn/, tutorial_models_vgg16.py worked prefect and tutorial_models_vgg19.py did't.
I guess it is the problem of tensorlayer/tensorlayer/models/vgg.py, cauz http://www.cs.toronto.edu/~frossard/post/vgg16/ can be acceessed and link to vgg19 cannot.
maybe change the link to a more stable one will help.
also hope codes in tensorlayer/examples/ could show how to load local vgg*.npz.
thanks a lot.

[se7enXF]

@rinabell

About vgg16 and vgg19

The pretrained weights of VGG16 and VGG19 are in different encodings which causes the different loading code in TL source code.

    if layer_type == 'vgg16':
        npz = np.load(os.path.join('models', model_saved_name[layer_type]))
        ...
    elif layer_type == 'vgg19':
        # npz = np.load(os.path.join('models', model_saved_name[layer_type]), encoding='latin1').item()
        npz = np.load(os.path.join('models', model_saved_name[layer_type]), encoding='latin1', allow_pickle=True).item()
        ...

Now I think you know why VGG16 works well. If you have downloaded the pretrained weights of VGG19 but failed to loading, please change the TL code tensorlayer/models/vgg.py in python site-packages as shown upper.

About downloading error

If you are in China might cause some TIMEOUT error. I suggest you to try more because I did it and I am in China.
To load the local vgg*.npz, you can find out by reading the source code and here is my submission:

• The pertrained *.npz will be downloaded into $PROJECT_DIR/models.
• Make a directory in your project root directory and download the pretrained *.npz into it. (Make sure that your project root is python execute root)
• The TL code will load *.npz from $PROJECT_DIR/models