def testSVD_BreaksDownLayerWeights(self):
model = _build_model()
first_conv_layer = model.layers[2]
self.assertLen(first_conv_layer.weights, 2)
model_for_inference = svd.SVD(rank=16).compress_model(model)
first_conv_layer = model_for_inference.layers[2]
self.assertLen(first_conv_layer.weights, 3)
def testSVD_HasReasonableAccuracy_TFLite(self):
model = _build_model()
_train_model(model)
model_for_inference = svd.SVD(rank=16).compress_model(model)
saved_model_dir = _save_as_saved_model(model_for_inference)
compressed_tflite_file = _convert_to_tflite(saved_model_dir)
accuracy = _test_tflite(compressed_tflite_file)
self.assertGreater(accuracy, 0.60)
def testSVD_ReducesSavedModelSize(self):
model = _build_model()
original_saved_model_dir = _save_as_saved_model(model)
model_for_inference = svd.SVD(rank=16).compress_model(model)
saved_model_dir = _save_as_saved_model(model_for_inference)
original_size = _get_directory_size_in_bytes(original_saved_model_dir)
compressed_size = _get_directory_size_in_bytes(saved_model_dir)
self.assertLess(compressed_size, original_size / 3)
def testSVD_ReducesTFLiteModelSize(self):
model = _build_model()
original_saved_model_dir = _save_as_saved_model(model)
original_tflite_file = _convert_to_tflite(original_saved_model_dir)
model_for_inference = svd.SVD(rank=16).compress_model(model)
saved_model_dir = _save_as_saved_model(model_for_inference)
compressed_tflite_file = _convert_to_tflite(saved_model_dir)
original_size = os.path.getsize(original_tflite_file)
compressed_size = os.path.getsize(compressed_tflite_file)
self.assertLess(compressed_size, original_size / 6)
def testSVD_HasReasonableAccuracy_TF(self):
model = _build_model()
_train_model(model)
model_for_inference = svd.SVD(rank=16).compress_model(model)
_, (x_test, y_test) = _get_dataset()
loss_fn = tf.keras.losses.SparseCategoricalCrossentropy(
from_logits=True)
model_for_inference.compile(optimizer='adam',
loss=loss_fn,
metrics=['accuracy'])
results = model_for_inference.evaluate(x_test, y_test)
self.assertGreater(results[1], 0.60)
def testSVD_PreservesPretrainedWeights(self):
i = tf.keras.layers.Input(shape=(2), name='input')
output = tf.keras.layers.Dense(3, name='fc1')(i)
model = tf.keras.Model(inputs=[i], outputs=[output])
dense_layer_weights = model.layers[1].get_weights()
algorithm = svd.SVD(rank=1)
model_for_inference = algorithm.compress_model(model)
dense_layer_compressed_weights = model_for_inference.layers[
1].get_weights()
# kernel
w1, w2 = algorithm.compress_training_weights(
tf.constant(dense_layer_weights[0]))
assert (w1 == dense_layer_compressed_weights[0]).numpy().all()
assert (w2 == dense_layer_compressed_weights[1]).numpy().all()
# bias
assert (
dense_layer_weights[1] == dense_layer_compressed_weights[2]).all()
