如何从保存的预训练模型的 h5 文件中找到层数？

Question

I have VGG-16 weights saved in h5 format. I want to see the number of layers in the network. How can I do that?

I tried using:

file = h5py.File(vgg16.h5)

after that, I checked for file.attrs but after this point I don't know which command to use to find the number of layers in the network.

Answer 1

If you type this code in spyder provided, you have installed Keras library. You will get all the layers in h5.file

from keras.models import load_model

classifier=load_model('my_model.h5')

classifier.summary()

Answer 2

In case you have only weights and not the model structure, you can use keys() method to get all the layers name in that weight file.

For example: I have one weight file: saved-weight.h5

If I want to know what are the layers present in this weight file, you can do the following:

import h5py
file = h5py.File('saved-weight.h5')

layer_names = file.keys()

# output 
layer_names =  <KeysViewHDF5 ['add', 'bn_3', 'bn_5', 'bn_7', 
    'concatenate', 'conv_1', 'conv_2', 'conv_3',
    'dropout', 'fc_8', 'fc_9', 
    'gru_10', 'gru_10_back', 'gru_11', 'gru_11_back', 
    'input_1', 'input_3', 'input_4', 'labels', 
    'lambda', 'lambda_1', 'lambda_2', 'lambda_3', 
    'maxpool_3', 'maxpool_5', 'model', 'permute', 'reshape']>

These are the layers present in the saved-weights file

Answer 3

Without your data or code, it is hard to provide more details. To demonstrate h5py methods to access h5 data, here is a simple example that creates a h5 file with 1 group with 3 datasets. After the group and datasets are created, there is a loop to print the dataset name, shape and dtype.

import h5py, numpy as np
h5f=h5py.File('SO_54511719.h5','w')

ds_data = np.random.random(100).reshape(10,10)
group1 = h5f.create_group('group1')
group1.create_dataset('ds_1', data=ds_data)
group1.create_dataset('ds_2', data=ds_data)
group1.create_dataset('ds_3', data=ds_data)

print ('number of datasets in group:', len(group1))
for (dsname, dsvalue) in group1.items() :
    print ('for',dsname,':')
    print ('shape:',dsvalue.shape)
    print ('dtype:',dsvalue.dtype)

h5f.close()

Output looks like this:

number of datasets in group: 3
for ds_1 :
shape: (10, 10)
dtype: float64
for ds_2 :
shape: (10, 10)
dtype: float64
for ds_3 :
shape: (10, 10)
dtype: float64