Arch Linux

Description:
Apparently h5py has changed its API in a backward-incompatible manner. I got the following error when I tried to load a Keras model in h5:

Traceback (most recent call last):
File "t.py", line 9, in <module>
model.load_weights(os.path.join(squeezenet_cifar10_path, 'squeeze_net.h5'))
File "/usr/lib/python3.8/site-packages/tensorflow/python/keras/engine/training.py", line 2211, in load_weights
hdf5_format.load_weights_from_hdf5_group(f, self.layers)
File "/usr/lib/python3.8/site-packages/tensorflow/python/keras/saving/hdf5_format.py", line 660, in load_weights_from_hdf5_group
original_keras_version = f.attrs['keras_version'].decode('utf8')
AttributeError: 'str' object has no attribute 'decode'

Additional info:
python-h5py 3.0.0-1
python-tensorflow-opt 2.3.1-2 # I believe all tensorflow split packages are affected

Related upstream ticket: https://github.com/tensorflow/tensorflow/issues/44467

Steps to reproduce:
1. git clone https://github.com/yan12125/SqueezeNet_vs_CIFAR10
2. Run the following script

import os.path

import tensorflow as tf

squeezenet_cifar10_path = './SqueezeNet_vs_CIFAR10/models'
with open(os.path.join(squeezenet_cifar10_path, 'squeeze_net.json')) as f:
model_json = f.read()
model = tf.keras.models.model_from_json(model_json)
model.load_weights(os.path.join(squeezenet_cifar10_path, 'squeeze_net.h5'))

FWIW, by patching two lines in tensorflow, I can load this model

--- /usr/lib/python3.8/site-packages/tensorflow/python/keras/saving/hdf5_format.py 2020-10-05 13:08:53.000000000 +0800
+++ /usr/lib/python3.8/site-packages/tensorflow/python/keras/saving/hdf5_format.py.new 2020-11-01 13:35:07.502299844 +0800
@@ -657,11 +657,11 @@
and weights file.
"""
if 'keras_version' in f.attrs:
- original_keras_version = f.attrs['keras_version'].decode('utf8')
+ original_keras_version = f.attrs['keras_version']
else:
original_keras_version = '1'
if 'backend' in f.attrs:
- original_backend = f.attrs['backend'].decode('utf8')
+ original_backend = f.attrs['backend']
else:
original_backend = None

Not sure if those are all needed changes in tensorflow, though.