def test_save_restore(self):
input_size = 20
n_classes = 5
sparse_layers = [
SparseLayer(values=np.array([1, 2, 3]).astype(np.float32),
indices=np.array([[0, 1], [1, 2],
[3, 4]]).astype(np.int16),
dense_shape=np.array([5, 10]).astype(np.int64),
bias=np.array([1, 1, 1, 1, 1]).astype(np.float32)),
SparseLayer(values=np.array([1, 2, 5]).astype(np.float32),
indices=np.array([[0, 2], [1, 2],
[3, 4]]).astype(np.int16),
dense_shape=np.array([10, 5]).astype(np.int64),
bias=np.array([0, 0, 0, 0, 0]).astype(np.float32))
]
network = network_sparse.FullyConnectedClassifierSparse(
input_size=input_size,
n_classes=n_classes,
sparse_layers=sparse_layers,
model_path='temp',
verbose=False)
variables_original = network.sess.run(network.weight_tensors)
network.save_model()
network.load_model()
variables_restored = network.sess.run(network.weight_tensors)
for original, restored in zip(variables_original, variables_restored):
self.assertTrue((original == restored).all())
def test_shapes(self):
input_size = 20
n_classes = 5
sparse_layers = [
SparseLayer(values=np.array([1, 2, 3]).astype(np.float32),
indices=np.array([[0, 1], [1, 2],
[3, 4]]).astype(np.int16),
dense_shape=np.array([5, 10]).astype(np.int64),
bias=np.array([1, 1, 1, 1, 1]).astype(np.float32)),
SparseLayer(values=np.array([1, 2, 5]).astype(np.float32),
indices=np.array([[0, 2], [1, 2],
[3, 4]]).astype(np.int16),
dense_shape=np.array([10, 5]).astype(np.int64),
bias=np.array([0, 0, 0, 0, 0]).astype(np.float32))
]
network = network_sparse.FullyConnectedClassifierSparse(
input_size=input_size,
n_classes=n_classes,
sparse_layers=sparse_layers,
model_path='temp',
verbose=False)
self.assertEqual(network.logits.get_shape().as_list(), [None, 5])
self.assertEqual(network.loss.get_shape().as_list(), [])
values, indices = pruning_utils.get_sparse_values_indices(weights)
shape = np.array(weights.shape).astype(np.int64)
sparse_layers.append(
pruning_utils.SparseLayer(
values=values.astype(np.float32),
#values=values.astype(np.int8),
indices=indices.astype(np.int16),
dense_shape=shape,
bias=bias))
pruning_utils.export_quantized("qweights.txt", 3, 0.25, qlayers, qbiases)
# create sparse classifier
sparse_classifier = network_sparse.FullyConnectedClassifierSparse(
input_size=config_sparse.input_size,
n_classes=config_sparse.n_classes,
sparse_layers=sparse_layers,
model_path=config_sparse.model_path,
activation_fn=config_sparse.activation_fn)
# test sparse classifiergt
runtime = time.time()
accuracy, loss = sparse_classifier.evaluate(
data_provider=test_data_provider, batch_size=config_sparse.batch_size)
runtime = time.time() - runtime
print('Sparse network took', runtime, ' seconds to evaluate.')
print('Accuracy on test with sparse model: {accuracy}, loss on test: {loss}'.
format(accuracy=accuracy, loss=loss))
compRatio = (classifier.params -
sparse_classifier.params) / classifier.params * 100
print('Compression ratio = ', 100 / (100 - compRatio))