This repo is outdated and will no longer be maintained.

Learning rate multiplier wrapper for optimizers.

pip install git+https://github.com/CyberZHG/keras-lr-multiplier.git

LRMultiplier is a wrapper for optimizers to assign different learning rates to specific layers (or weights). The first argument is the original optimizer which could be either an identifier (e.g. 'Adam') or an initialized object (e.g. Adam(lr=1e-2)). The second argument is a dict that maps prefixes to learning rate multipliers. The multiplier for a weight is the value mapped from the longest matched prefix in the given dict, and the default multiplier 1.0 will be used if there is no prefix matched.

from keras.models import Sequential
from keras.layers import Dense
from keras_lr_multiplier import LRMultiplier

model = Sequential()
model.add(Dense(
    units=5,
    input_shape=(5,),
    activation='tanh',
    name='Dense',
))
model.add(Dense(
    units=2,
    activation='softmax',
    name='Output',
))
model.compile(
    optimizer=LRMultiplier('adam', {'Dense': 0.5, 'Output': 1.5}),
    loss='sparse_categorical_crossentropy',
)

The multiplier can be a callable object. The input argument is the number of steps starting from 0.

from keras import backend as K
from keras_lr_multiplier import LRMultiplier

LRMultiplier('adam', {'Dense': lambda t: 2.0 - K.minimum(1.9, t * 1e-4)})