TypeError: Unexpected keyword argument passed to optimizer: learning_rate

问题

I am trying to load a Keras model which was trained on an Azure VM (NC promo). But I am getting the following error.

TypeError: Unexpected keyword argument passed to optimizer:learning_rate

EDIT:

Here is the code snippet that I am using to load my model:

from keras.models import load_model
model = load_model('my_model_name.h5')

回答1:

Did you use a custom optimizer?

If so, you can load like this:

model = load_model('my_model_name.h5', custom_objects={
    'Adam': lambda **kwargs: hvd.DistributedOptimizer(keras.optimizers.Adam(**kwargs))
})

Alternatively you can load your model with model = load_model('my_model_name.h5', compile=False) and then add an optimizer and recompile, but that will lose your saved weights.

回答2:

This happened to me too. Most likely because the learning_rate was renamed from version 2.2.* to 2.3.0 in September 2018. (see release notes: https://github.com/keras-team/keras/releases : Rename lr to learning_rate for all optimizers. )

This worked for me: sudo pip install keras --upgrade

回答3:

i got the same error while i was working in two different PC. in some versions of tensorflow is tf.keras.optimizers.SGD(lr = x) while in other verions istf.keras.optimizers.SGD(learning rate = x).

回答4:

In my case I found the best solution is to use h5py to change name of the variable from "learning_rate" -> "lr" as suggested in the previous posts.

import h5py
data_p = f.attrs['training_config']
data_p = data_p.decode().replace("learning_rate","lr").encode()
f.attrs['training_config'] = data_p
f.close()

回答5:

I had the same problem. Using Keras version 2.3.1 and TensorFlow-GPU version 1.13, I had to upgrade Tensorflow-GPU to version 1.15

pip uninstall tensorflow-gpu
pip install tensorflow-gpu==1.15

回答6:

I've had a similar problem.

You you have this issue, try to use lr instead of learning_rate when defining the learning rate in your optimizer.

回答7:

I am also experiencing this when I try to load my model on another machine. Also trained the initial modal on an azure VM. I have tried the suggestions above and can't figure out what is causing it. Any other thoughts? Below is my code to train the model.

Models were trained and are being used in my api projects using the following versions: keras 2.3.0 tensorflow 1.14.0

history = model.fit(X, y,validation_split=0.1, \
                epochs=20, \
                batch_size=1000, \
                class_weight = cw)

回答8:

I was running into the same thing. You will have to upgrade to Tensorlfow 2.0 and Keras, or match the two systems together.

回答9:

That issue usual on dependencies difference between the kernel where that model has been trained and the dependencies versions where the model is being loaded.

If you have installed the latest version of Tensorflow now (2.1) try to load the model like this:

import tensorflow as tf
print(tf.__version__)
print("Num GPUs Available: ", 
       len(tf.config.experimental.list_physical_devices('GPU')))
# Checking the version for incompatibilities and GPU list devices 
# for a fast check on GPU drivers installation. 

model_filepath = './your_model_path.h5'

model = tf.keras.models.load_model(
    model_filepath,
    custom_objects=None,
    compile=False
)

Compile=False only if the model has already compiled.