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 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 testCreateMask2D(self):
width = 10
height = 20
with self.cached_session():
weights = variables.VariableV1(random_ops.random_normal(
[width, height], stddev=1),
name="weights")
masked_weights = pruning.apply_mask(
weights, variable_scope.get_variable_scope())
variables.global_variables_initializer().run()
weights_val = weights.eval()
masked_weights_val = masked_weights.eval()
self.assertAllEqual(weights_val, masked_weights_val)
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 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 inference(images):
"""Build the CIFAR-10 model.
Args:
images: Images returned from distorted_inputs() or inputs().
Returns:
Logits.
"""
# We instantiate all variables using tf.compat.v1.get_variable() instead of
# tf.Variable() in order to share variables across multiple GPU training runs.
# If we only ran this model on a single GPU, we could simplify this function
# by replacing all instances of tf.compat.v1.get_variable() with tf.Variable().
#
# While instantiating conv and local layers, we add mask and threshold
# variables to the layer by calling the pruning.apply_mask() function.
# Note that the masks are applied only to the weight tensors
# conv1
with tf.variable_scope('conv1') as scope:
kernel = _variable_with_weight_decay('weights',
shape=[5, 5, 3, 64],
stddev=5e-2,
wd=0.0)
conv = tf.nn.conv2d(images,
pruning.apply_mask(kernel, scope), [1, 1, 1, 1],
padding='SAME')
biases = _variable_on_cpu('biases', [64], tf.constant_initializer(0.0))
pre_activation = tf.nn.bias_add(conv, biases)
conv1 = tf.nn.relu(pre_activation, name=scope.name)
_activation_summary(conv1)
# pool1
pool1 = tf.nn.max_pool(conv1,
ksize=[1, 3, 3, 1],
strides=[1, 2, 2, 1],
padding='SAME',
name='pool1')
# norm1
norm1 = tf.nn.lrn(pool1,
4,
bias=1.0,
alpha=0.001 / 9.0,
beta=0.75,
name='norm1')
# conv2
with tf.variable_scope('conv2') as scope:
kernel = _variable_with_weight_decay('weights',
shape=[5, 5, 64, 64],
stddev=5e-2,
wd=0.0)
conv = tf.nn.conv2d(norm1,
pruning.apply_mask(kernel, scope), [1, 1, 1, 1],
padding='SAME')
biases = _variable_on_cpu('biases', [64], tf.constant_initializer(0.1))
pre_activation = tf.nn.bias_add(conv, biases)
conv2 = tf.nn.relu(pre_activation, name=scope.name)
_activation_summary(conv2)
# norm2
norm2 = tf.nn.lrn(conv2,
4,
bias=1.0,
alpha=0.001 / 9.0,
beta=0.75,
name='norm2')
# pool2
pool2 = tf.nn.max_pool(norm2,
ksize=[1, 3, 3, 1],
strides=[1, 2, 2, 1],
padding='SAME',
name='pool2')
# local3
with tf.variable_scope('local3') as scope:
# Move everything into depth so we can perform a single matrix multiply.
reshape = tf.reshape(pool2, [BATCH_SIZE, -1])
dim = reshape.get_shape()[1].value
weights = _variable_with_weight_decay('weights',
shape=[dim, 384],
stddev=0.04,
wd=0.004)
biases = _variable_on_cpu('biases', [384],
tf.constant_initializer(0.1))
local3 = tf.nn.relu(
tf.matmul(reshape, pruning.apply_mask(weights, scope)) + biases,
name=scope.name)
_activation_summary(local3)
# local4
with tf.variable_scope('local4') as scope:
weights = _variable_with_weight_decay('weights',
shape=[384, 192],
stddev=0.04,
wd=0.004)
biases = _variable_on_cpu('biases', [192],
tf.constant_initializer(0.1))
local4 = tf.nn.relu(
tf.matmul(local3, pruning.apply_mask(weights, scope)) + biases,
name=scope.name)
_activation_summary(local4)
# linear layer(WX + b),
# We don't apply softmax here because
# tf.nn.sparse_softmax_cross_entropy_with_logits accepts the unscaled logits
# and performs the softmax internally for efficiency.
with tf.variable_scope('softmax_linear') as scope:
weights = _variable_with_weight_decay('weights', [192, NUM_CLASSES],
stddev=1 / 192.0,
wd=0.0)
biases = _variable_on_cpu('biases', [NUM_CLASSES],
tf.constant_initializer(0.0))
softmax_linear = tf.add(tf.matmul(local4,
pruning.apply_mask(weights, scope)),
biases,
name=scope.name)
_activation_summary(softmax_linear)
return softmax_linear