Keras segmentation fault on load_model on Linux and not on Windows - linux

Question

I prototyped a Python deep learning piece of code working on Windows and I can't make it work on Linux. I identified that the problem comes from load_model.
Here is the piece of Python code that behaves differently on Windows and in Linux.
Both Keras installations were made from the github source repository from Keras Team because the model format is not recognized by the standard Keras package, a patch was done very recently for the characters format in the Github source code.
Do you have an idea of what's going on?
The code:
from keras.models import load_model, Model
import sys
import keras
import tensorflow as tf
import os
import platform
print("----------------------------------------------")
print("Operating system:")
print (os.name)
print(platform.system())
print(platform.release())
print("----------------------------------------------")
print("Python version:")
print(sys.version)
print("----------------------------------------------")
print("Tensorflow version: ", tf.__version__)
print("----------------------------------------------")
print("Keras version : ", keras.__version__)
print("----------------------------------------------")
yolo_model = load_model("model.h5")
Windows output:
Using TensorFlow backend.
----------------------------------------------
Operating system:
nt
Windows
7
----------------------------------------------
Python version:
3.6.0 (v3.6.0:41df79263a11, Dec 23 2016, 08:06:12) [MSC v.1900 64 bit (AMD64)]
----------------------------------------------
Tensorflow version: 1.4.0
----------------------------------------------
Keras version : 2.1.2
----------------------------------------------
2018-01-06 21:54:37.700794: I C:\tf_jenkins\home\workspace\rel-win\M\windows\PY\36\tensorflow\core\platform\cpu_feature_guard.cc:137] Your CPU supports instruc
ions that this TensorFlow binary was not compiled to use: AVX AVX2
C:\Users\David\AppData\Local\Programs\Python\Python36\lib\site-packages\keras-2.1.2-py3.6.egg\keras\models.py:252: UserWarning: No training configuration found
in save file: the model was *not* compiled. Compile it manually.
Linux output:
Using TensorFlow backend.
----------------------------------------------
Operating system:
posix
Linux
4.9.0-5-amd64
----------------------------------------------
Python version:
3.5.3 (default, Jan 19 2017, 14:11:04)
[GCC 6.3.0 20170118]
----------------------------------------------
Tensorflow version: 1.4.1
----------------------------------------------
Keras version : 2.1.2
----------------------------------------------
----------------------------------------------
2018-01-06 21:47:58.099715: I tensorflow/core/platform/cpu_feature_guard.cc:137] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.1 SSE4.2 AVX
Erreur de segmentation
The french Erreur de segmentation means Segmentation fault
Thank you for your help!
Glassfrog

Answer 1

I only found a workaround.
As the model file was a data conversion from another weights file in another format, I went and regenerated the Keras model for the latest version of Keras.
Now It works.
But I still don't know what caused the segmentation fault.

Answer 2

From what I can tell, the segfault happens at the model creation, but I have no idea why.
I could debug this by saving the model and weights independently:
from keras.models import load_model
x = load_model('combined_model.h5') # runs only on the source machine
with open('model.json', 'w') as fp:
fp.write(x.to_json())
x.save_weights('weights.h5')
on the other machine I tried to load the model from the JSON file, but got the segmentation fault as well:
from keras.models import model_from_json
with open('model.json', 'r') as fp:
model = model_from_json(fp.read()) # segfaults here
If it is possible to simply re-create the model on the target machine by creating the Sequential model again, you can simply load the weights in the model:
from keras import Sequential
# ...
new_model = Sequential()
# [...] run your model creation here...
new_model.load_weights('weights.h5')
new_model.predict(...) # this should work now