def testInitWithVariableReuse(self):
with self.cached_session():
p = pruning.Pruning(spec=self.pruning_hparams,
sparsity=self.sparsity)
p_copy = pruning.Pruning(spec=self.pruning_hparams,
sparsity=self.sparsity)
variables.global_variables_initializer().run()
sparsity = p._sparsity.eval()
self.assertAlmostEqual(sparsity, 0.5)
self.assertEqual(p._sparsity.eval(), p_copy._sparsity.eval())
def testInitWithExternalSparsity(self):
with self.cached_session():
p = pruning.Pruning(spec=self.pruning_hparams,
sparsity=self.sparsity)
variables.global_variables_initializer().run()
sparsity = p._sparsity.eval()
self.assertAlmostEqual(sparsity, 0.5)
def testPerLayerBlockSparsity(self):
param_list = [
"block_dims_map=[layer1/weights:1x1,layer2/weights:1x2]",
"block_pooling_function=AVG", "threshold_decay=0.0"
]
test_spec = ",".join(param_list)
pruning_hparams = pruning.get_pruning_hparams().parse(test_spec)
with variable_scope.variable_scope("layer1"):
w1 = constant_op.constant([[-0.1, 0.1], [-0.2, 0.2]],
name="weights")
pruning.apply_mask(w1)
with variable_scope.variable_scope("layer2"):
w2 = constant_op.constant(
[[0.1, 0.1, 0.3, 0.3], [0.2, 0.2, 0.4, 0.4]], name="weights")
pruning.apply_mask(w2)
sparsity = variables.VariableV1(0.5, name="sparsity")
p = pruning.Pruning(pruning_hparams, sparsity=sparsity)
mask_update_op = p.mask_update_op()
with self.cached_session() as session:
variables.global_variables_initializer().run()
session.run(mask_update_op)
mask1_eval = session.run(pruning.get_masks()[0])
mask2_eval = session.run(pruning.get_masks()[1])
self.assertAllEqual(session.run(pruning.get_weight_sparsity()),
[0.5, 0.5])
self.assertAllEqual(mask1_eval, [[0.0, 0.0], [1., 1.]])
self.assertAllEqual(mask2_eval, [[0, 0, 1., 1.], [0, 0, 1., 1.]])
def testInit(self):
p = pruning.Pruning(self.pruning_hparams)
self.assertEqual(p._spec.name, "test")
self.assertAlmostEqual(p._spec.threshold_decay, 0.9)
self.assertEqual(p._spec.pruning_frequency, 10)
self.assertEqual(p._spec.sparsity_function_end_step, 100)
self.assertAlmostEqual(p._spec.target_sparsity, 0.9)
def testWeightSpecificSparsity(self):
param_list = [
"begin_pruning_step=1", "pruning_frequency=1",
"end_pruning_step=100", "target_sparsity=0.5",
"weight_sparsity_map=[layer1:0.6,layer2/weights:0.75,.*kernel:0.6]",
"threshold_decay=0.0"
]
test_spec = ",".join(param_list)
pruning_hparams = pruning.get_pruning_hparams().parse(test_spec)
with variable_scope.variable_scope("layer1"):
w1 = variables.VariableV1(math_ops.linspace(1.0, 100.0, 100),
name="weights")
_ = pruning.apply_mask(w1)
with variable_scope.variable_scope("layer2"):
w2 = variables.VariableV1(math_ops.linspace(1.0, 100.0, 100),
name="weights")
_ = pruning.apply_mask(w2)
with variable_scope.variable_scope("layer3"):
w3 = variables.VariableV1(math_ops.linspace(1.0, 100.0, 100),
name="kernel")
_ = pruning.apply_mask(w3)
p = pruning.Pruning(pruning_hparams)
mask_update_op = p.conditional_mask_update_op()
increment_global_step = state_ops.assign_add(self.global_step, 1)
with self.cached_session() as session:
variables.global_variables_initializer().run()
for _ in range(110):
session.run(mask_update_op)
session.run(increment_global_step)
self.assertAllClose(session.run(pruning.get_weight_sparsity()),
[0.6, 0.75, 0.6])
def testConditionalMaskUpdate(self):
param_list = [
"pruning_frequency=2", "begin_pruning_step=1",
"end_pruning_step=6", "nbins=100"
]
test_spec = ",".join(param_list)
pruning_hparams = pruning.get_pruning_hparams().parse(test_spec)
weights = variables.VariableV1(math_ops.linspace(1.0, 100.0, 100),
name="weights")
masked_weights = pruning.apply_mask(weights)
sparsity = variables.VariableV1(0.00, name="sparsity")
# Set up pruning
p = pruning.Pruning(pruning_hparams, sparsity=sparsity)
p._spec.threshold_decay = 0.0
mask_update_op = p.conditional_mask_update_op()
sparsity_val = math_ops.linspace(0.0, 0.9, 10)
increment_global_step = state_ops.assign_add(self.global_step, 1)
non_zero_count = []
with self.cached_session() as session:
variables.global_variables_initializer().run()
for i in range(10):
session.run(state_ops.assign(sparsity, sparsity_val[i]))
session.run(mask_update_op)
session.run(increment_global_step)
non_zero_count.append(np.count_nonzero(masked_weights.eval()))
# Weights pruned at steps 0,2,4,and,6
expected_non_zero_count = [100, 100, 80, 80, 60, 60, 40, 40, 40, 40]
self.assertAllEqual(expected_non_zero_count, non_zero_count)
def _prune_model(self, session):
pruning_hparams = pruning.get_pruning_hparams().parse(self.pruning_spec)
p = pruning.Pruning(pruning_hparams, sparsity=self.sparsity)
self.mask_update_op = p.conditional_mask_update_op()
variables.global_variables_initializer().run()
for _ in range(20):
session.run(self.mask_update_op)
session.run(self.increment_global_step)
def testUpdateSingleMask(self):
with self.cached_session() as session:
weights = variables.VariableV1(math_ops.linspace(1.0, 100.0, 100),
name="weights")
masked_weights = pruning.apply_mask(weights)
sparsity = variables.VariableV1(0.95, name="sparsity")
p = pruning.Pruning(sparsity=sparsity)
p._spec.threshold_decay = 0.0
mask_update_op = p.mask_update_op()
variables.global_variables_initializer().run()
masked_weights_val = masked_weights.eval()
self.assertAllEqual(np.count_nonzero(masked_weights_val), 100)
session.run(mask_update_op)
masked_weights_val = masked_weights.eval()
self.assertAllEqual(np.count_nonzero(masked_weights_val), 5)
def _blockMasking(self, hparams, weights, expected_mask):
threshold = variables.VariableV1(0.0, name="threshold")
sparsity = variables.VariableV1(0.5, name="sparsity")
test_spec = ",".join(hparams)
pruning_hparams = pruning.get_pruning_hparams().parse(test_spec)
# Set up pruning
p = pruning.Pruning(pruning_hparams, sparsity=sparsity)
with self.cached_session():
variables.global_variables_initializer().run()
_, new_mask = p._maybe_update_block_mask(weights, threshold)
# Check if the mask is the same size as the weights
self.assertAllEqual(new_mask.get_shape(), weights.get_shape())
mask_val = new_mask.eval()
self.assertAllEqual(mask_val, expected_mask)
def testPartitionedVariableMasking(self):
partitioner = partitioned_variables.variable_axis_size_partitioner(40)
with self.cached_session() as session:
with variable_scope.variable_scope("", partitioner=partitioner):
sparsity = variables.VariableV1(0.5, name="Sparsity")
weights = variable_scope.get_variable(
"weights", initializer=math_ops.linspace(1.0, 100.0, 100))
masked_weights = pruning.apply_mask(
weights, scope=variable_scope.get_variable_scope())
p = pruning.Pruning(sparsity=sparsity)
p._spec.threshold_decay = 0.0
mask_update_op = p.mask_update_op()
variables.global_variables_initializer().run()
masked_weights_val = masked_weights.eval()
session.run(mask_update_op)
masked_weights_val = masked_weights.eval()
self.assertAllEqual(np.count_nonzero(masked_weights_val), 50)
def train():
"""Train CIFAR-10 for a number of steps."""
with tf.Graph().as_default():
global_step = tf.contrib.framework.get_or_create_global_step()
# Get images and labels for CIFAR-10.
images, labels = cifar10.distorted_inputs()
# Build a Graph that computes the logits predictions from the
# inference model.
logits = cifar10.inference(images)
# Calculate loss.
loss = cifar10.loss(logits, labels)
# Build a Graph that trains the model with one batch of examples and
# updates the model parameters.
train_op = cifar10.train(loss, global_step)
# Parse pruning hyperparameters
pruning_hparams = pruning.get_pruning_hparams().parse(
FLAGS.pruning_hparams)
# Create a pruning object using the pruning hyperparameters
pruning_obj = pruning.Pruning(pruning_hparams, global_step=global_step)
# Use the pruning_obj to add ops to the training graph to update the masks
# The conditional_mask_update_op will update the masks only when the
# training step is in [begin_pruning_step, end_pruning_step] specified in
# the pruning spec proto
mask_update_op = pruning_obj.conditional_mask_update_op()
# Use the pruning_obj to add summaries to the graph to track the sparsity
# of each of the layers
pruning_obj.add_pruning_summaries()
class _LoggerHook(tf.train.SessionRunHook):
"""Logs loss and runtime."""
def begin(self):
self._step = -1
def before_run(self, run_context):
self._step += 1
self._start_time = time.time()
return tf.train.SessionRunArgs(loss) # Asks for loss value.
def after_run(self, run_context, run_values):
duration = time.time() - self._start_time
loss_value = run_values.results
if self._step % 10 == 0:
num_examples_per_step = 128
examples_per_sec = num_examples_per_step / duration
sec_per_batch = float(duration)
format_str = (
'%s: step %d, loss = %.2f (%.1f examples/sec; %.3f '
'sec/batch)')
print(format_str %
(datetime.datetime.now(), self._step, loss_value,
examples_per_sec, sec_per_batch))
with tf.train.MonitoredTrainingSession(
checkpoint_dir=FLAGS.train_dir,
hooks=[
tf.train.StopAtStepHook(last_step=FLAGS.max_steps),
tf.train.NanTensorHook(loss),
_LoggerHook()
],
config=tf.ConfigProto(log_device_placement=FLAGS.
log_device_placement)) as mon_sess:
while not mon_sess.should_stop():
mon_sess.run(train_op)
# Update the masks
mon_sess.run(mask_update_op)