This python package provides a library that accelerates the training of arbitrary neural networks created with Keras using importance sampling.
# Keras imports
from importance_sampling.training import ImportanceTraining, \
ApproximateImportanceTraining
x_train, y_train, x_val, y_val = load_data()
model = create_keras_model()
model.compile(
optimizer="adam",
loss="categorical_crossentropy",
metrics=["accuracy"]
)
ImportanceTraining(model).fit(
x_train, y_train,
batch_size=32,
epochs=10,
verbose=1,
validation_data=(x_val, y_val)
)
model.evaluate(x_val, y_val)
Importance sampling for Deep Learning is an active research field and this library is undergoing development so your mileage may vary.
Ours
- Biased Importance Sampling for Deep Neural Network Training [preprint]
By others
- Stochastic optimization with importance sampling for regularized loss minimization [pdf]
- Variance reduction in SGD by distributed importance sampling [pdf]
Normally if you already have a functional Keras installation you just need to
pip install keras-importance-sampling
.
Keras
> 2- A Keras backend among Tensorflow, Theano and CNTK
transparent-keras
blinker
numpy
matplotlib
,seaborn
,scikit-learn
are optional (used by the plot scripts)
The module has a dedicated documentation site but you can also read the source code and the examples to get an idea of how the library should be used and extended.
In the examples
folder you can find some Keras examples that have been edited
(minimally) to use importance sampling.
In this section we will showcase part of the API that can be used to train neural networks with importance sampling.
# Import what is needed to build the Keras model
from keras import backend as K
from keras.layers import Dense
from keras.models import Sequential
# Import a toy dataset and the importance training
from importance_sampling.datasets import CanevetICML2016
from importance_sampling.training import ImportanceTraining
def create_nn():
"""Build a simple fully connected NN"""
model = Sequential([
Dense(40, activation="tanh", input_shape=(2,)),
Dense(40, activation="tanh"),
Dense(1, activation="sigmoid")
])
model.compile(
optimizer="adam",
loss="binary_crossentropy",
metrics=["accuracy"]
)
return model
if __name__ == "__main__":
# Load the data
dataset = CanevetICML2016(N=1024)
x_train, y_train = dataset.train_data[:]
x_test, y_test = dataset.test_data[:]
y_train, y_test = y_train.argmax(axis=1), y_test.argmax(axis=1)
# Create the NN and keep the initial weights
model = create_nn()
weights = model.get_weights()
# Train with uniform sampling
K.set_value(model.optimizer.lr, 0.01)
model.fit(
x_train, y_train,
batch_size=64, epochs=10,
validation_data=(x_test, y_test)
)
# Train with biased importance sampling
model.set_weights(weights)
K.set_value(model.optimizer.lr, 0.01)
ImportanceTraining(model, forward_batch_size=1024).fit(
x_train, y_train,
batch_size=64, epochs=3,
validation_data=(x_test, y_test)
)
The following terminal commands train a small VGG-like network to ~0.55% error on MNIST (the numbers are from a CPU). It is not optimized, it just showcases that with importance sampling 6 times less iterations are required in this case.
$ # Train a small cnn with mnist for 500 mini-batches using importance $ # sampling with bias to achieve ~ 0.55% error (on the CPU) $ time ./importance_sampling.py \ > small_cnn \ > oracle-loss \ > model \ > predicted \ > mnist \ > /tmp/is \ > --hyperparams 'batch_size=i128;lr=f0.003;lr_reductions=I10000;k=f0.5' \ > --train_for 500 --validate_every 500 real 6m16.476s user 24m46.800s sys 5m36.592s $ $ # And with uniform sampling to achieve the same accuracy (learning rate is $ # smaller because with uniform sampling the variance is too big) $ time ./importance_sampling.py \ > small_cnn \ > oracle-loss \ > uniform \ > unweighted \ > mnist \ > /tmp/uniform \ > --hyperparams 'batch_size=i128;lr=f0.001;lr_reductions=I1000' \ > --train_for 3000 --validate_every 3000 real 10m36.836s user 47m36.316s sys 7m14.412s