How can I get the number of trainable parameters of a model in Keras?

user_8021672 0
Viewed 18596

I am setting trainable=False in all my layers, implemented through the Model API, but I want to verify whether that is working. model.count_params() returns the total number of parameters, but is there any way in which I can get the total number of trainable parameters, other than looking at the last few lines of model.summary()?

keras python
4 Answers

For TensorFlow 2.0:

import tensorflow.keras.backend as K

trainable_count = np.sum([K.count_params(w) for w in model.trainable_weights])
non_trainable_count = np.sum([K.count_params(w) for w in model.non_trainable_weights])

print('Total params: {:,}'.format(trainable_count + non_trainable_count))
print('Trainable params: {:,}'.format(trainable_count))
print('Non-trainable params: {:,}'.format(non_trainable_count))
user_86093300

For tensorflow.keras this works for me. Its from the tensorflow github code for the function print_layer_summary_with_connections() in layer_utils.py

import numpy as np
from tensorflow.python.util import object_identity

def count_params(weights):
    return int(sum(np.prod(p.shape.as_list())
      for p in object_identity.ObjectIdentitySet(weights)))

if hasattr(model, '_collected_trainable_weights'):
    trainable_count = count_params(model._collected_trainable_weights)
else:
    trainable_count = count_params(model.trainable_weights)

print (trainable_count)
user_85580400

Another way to count trainable parameters is:

model.count_params()
user_38318450
Related