Unknown initializer: GlorotUniform when loading Ke

问题:

I trained my CNN (VGG) through google colab and generated .h5 file. Now problem is, I can predict my output successfully through google colab but when i download that .h5 trained model file and try to predict output on my laptop, I am getting error when loading the model.

Here is the code:

import tensorflow as tf
from tensorflow import keras
import h5py

# Initialization

loaded_model = keras.models.load_model('./train_personCount_model.h5')

And the error:

ValueError: Unknown initializer: GlorotUniform

回答1:

I ran into the same issue. After changing:

from tensorflow import keras

to:

import keras

life is once again worth living.

回答2:

I fixed the problem:

Before:

from keras.models import load_model
classifierLoad = load_model('model/modeltest.h5')

Works for me

import tensorflow as tf 
classifierLoad = tf.keras.models.load_model('model/modeltest.h5')

回答3:

Wow I, just spent 6 Hours of my life trying to figure this out.. Dmitri posted a solution to this here: I trained a keras model on google colab. Now not able to load it locally on my system.

I'm just basically reposting it here because it worked for me.

This looks like some kind of a serialization bug in keras. If you wrap your load_model with the below CustomObjectScope thingy... all should work..

import keras
from keras.models import load_model
from keras.utils import CustomObjectScope
from keras.initializers import glorot_uniform

with CustomObjectScope({'GlorotUniform': glorot_uniform()}):
        model = load_model('imdb_mlp_model.h5')

回答4:

Changing

from keras.models import load_model

to

from tensorflow.keras.models import load_model

solved my problem!

To eliminate errors, import all things directly from Keras or TensorFlow. Mixing both of them in same project may result in problems.

回答5:

I had a same problem and was fixed this way. just don't save the optimizer with the model! just change the save line like this:

the_model.save(file_path,True/False,False)

Second parameter tells Keras to overwrite the model if the file existed or not and the 3rd one tells it not to save the optimizer with the model.

---

Edit: I ran over the problem again on another system today and this did not helped me this time. so i saved the model conf as json and weights as h5 and used them to rebuild the model in another machine. you can do it like this. save like this:

json = model.to_json()
# Save the json on a file
model.save_weights(weights_filepath,save_format="h5")

rebuild the model like this:

# load the json file
# here i use json as loaded content of json file
model = keras.models.model_from_json(json)
model.load_weights(weights_file_path)

回答6:

Something that helped me which wasn't in any of the answers:

custom_objects={'GlorotUniform': glorot_uniform()}

回答7:

if you are loading the architecture and weights separtly, while loading archtiecture of the model change :

models.model_from_json(json)

to :

tf.keras.models.model_from_json(json)

and the problem is solved

回答8:

from tensorflow.keras.initializers import glorot_uniform

loaded_model = tf.keras.models.load_model("pruned.h5",custom_objects={'GlorotUniform': glorot_uniform()})

this worked for me when importing tensorflow keras

回答9:

I had the same problem with a model built with tensorflow 1.11.0 (using tensorflow.python.keras.models.save_model) and loaded with tensoflow 1.11.0 (using tensorflow.python.keras.models.load_model).

I solved it by upgrading everything to tensorflow 1.13.1, after building the model again with the new version, I could load it without this error.