def test_loss_sensitivity(net_const: Callable, inp_arr: numpy.ndarray,
labs_arr: numpy.ndarray):
with tf_compat.Graph().as_default():
out, inp = net_const()
labels = tf_compat.placeholder(tf_compat.float32,
[None, *labs_arr.shape[1:]],
name="logits")
loss = batch_cross_entropy_loss(out, labels)
op_vars = pruning_loss_sens_op_vars()
with tf_compat.Session() as sess:
sess.run(tf_compat.global_variables_initializer())
def add_ops_creator(step: int):
return []
def feed_dict_creator(step: int):
return {inp: inp_arr, labels: labs_arr}
analysis = pruning_loss_sens_one_shot(op_vars, loss, 5,
add_ops_creator,
feed_dict_creator)
for res in analysis.results:
assert res.name
assert isinstance(res.index, int)
assert len(res.sparse_measurements) > 0
assert len(res.averages) > 0
assert res.sparse_average > 0
assert res.sparse_integral > 0
def test_mnist_registry(key: str, pretrained: Union[bool, str],
test_input: bool):
with tf_compat.Graph().as_default():
inputs = tf_compat.placeholder(tf_compat.float32, [None, 28, 28, 1],
name="inputs")
logits = ModelRegistry.create(key, inputs)
with tf_compat.Session() as sess:
if test_input:
sess.run(tf_compat.global_variables_initializer())
out = sess.run(
logits,
feed_dict={inputs: numpy.random.random((1, 28, 28, 1))})
assert out.sum() != 0
if pretrained:
ModelRegistry.load_pretrained(key, pretrained)
if test_input:
out = sess.run(logits,
feed_dict={
inputs: numpy.random.random(
(1, 28, 28, 1))
})
assert out.sum() != 0
def test_trainable_params_modifier_with_training():
modifier = TrainableParamsModifier(
params=["mlp_net/fc1/weight"],
trainable=False,
params_strict=False,
)
manager = ScheduledModifierManager([modifier])
steps_per_epoch = 5
batch_size = 2
with tf_compat.Graph().as_default() as graph:
logits, inputs = mlp_net()
labels = tf_compat.placeholder(tf_compat.float32,
[None, *logits.shape[1:]])
loss = batch_cross_entropy_loss(logits, labels)
global_step = tf_compat.train.get_or_create_global_step()
num_trainable_variabls_init = len(tf_compat.trainable_variables())
mod_ops, mod_extras = manager.create_ops(steps_per_epoch)
assert len(
tf_compat.trainable_variables()) < num_trainable_variabls_init
# Get the variables returned by the trainable_params modifier
non_trainable_vars = mod_extras[EXTRAS_KEY_VAR_LIST]
trainable_vars = tf_compat.trainable_variables()
train_op = tf_compat.train.AdamOptimizer(learning_rate=1e-4).minimize(
loss, global_step=global_step)
with tf_compat.Session(graph=graph) as sess:
sess.run(tf_compat.global_variables_initializer())
manager.initialize_session(sess)
init_non_trainable_vars = [
var.eval(session=sess) for var in non_trainable_vars
]
init_trainable_vars = [
var.eval(session=sess) for var in trainable_vars
]
batch_lab = numpy.random.random((batch_size, *logits.shape[1:]))
batch_inp = numpy.random.random((batch_size, *inputs.shape[1:]))
for epoch in range(10):
for step in range(steps_per_epoch):
sess.run(train_op,
feed_dict={
inputs: batch_inp,
labels: batch_lab
})
sess.run(global_step)
# Compare initial and final variable values
for idx, init_non_trainable_var in enumerate(
init_non_trainable_vars):
final_non_trainable_var = non_trainable_vars[idx].eval(
session=sess)
assert numpy.array_equal(init_non_trainable_var,
final_non_trainable_var)
for idx, init_trainable_var in enumerate(init_trainable_vars):
final_trainable_var = trainable_vars[idx].eval(session=sess)
assert not numpy.array_equal(init_trainable_var,
final_trainable_var)
manager.complete_graph()
def test_lifecycle(
self,
modifier_lambda: Callable[[], SetLearningRateModifier],
graph_lambda: Callable[[], tf_compat.Graph],
steps_per_epoch: int,
optim_lambda,
):
modifier = modifier_lambda()
graph = graph_lambda()
with graph.as_default():
global_step = tf_compat.train.get_or_create_global_step()
# Further set up for loss, optimizer and training op
x_batch = graph.get_tensor_by_name("inp:0")
y_pred = graph.get_tensor_by_name("out:0")
n_inputs = x_batch.shape[1]
n_outputs = y_pred.shape[1]
y_lab = tf_compat.placeholder(tf_compat.float32,
shape=(None, n_outputs),
name="y")
mod_ops, mod_extras = modifier.create_ops(steps_per_epoch,
global_step=global_step,
graph=graph)
assert len(mod_ops) == 0
assert len(mod_extras) == 2
assert EXTRAS_KEY_LEARNING_RATE in mod_extras
assert EXTRAS_KEY_SUMMARIES in mod_extras
learning_rate = mod_extras[EXTRAS_KEY_LEARNING_RATE]
with tf_compat.name_scope("train"):
optimizer = optim_lambda(learning_rate=learning_rate)
loss = tf_compat.losses.mean_squared_error(y_lab, y_pred)
training_op = optimizer.minimize(loss, global_step=global_step)
np.random.seed(12)
batch_size = 8
batch_x = np.random.randn(batch_size, n_inputs)
batch_lab = np.random.randn(batch_size, n_outputs)
with tf_compat.Session(graph=graph) as sess:
sess.run(tf_compat.global_variables_initializer())
for epoch in range(
int(max(modifier.start_epoch, modifier.end_epoch)) + 5):
for step in range(steps_per_epoch):
gs = sess.run(global_step)
expected = modifier.learning_rate
optim_lr = sess.run(_get_lr(optimizer))
assert (
abs(optim_lr - expected) <= EPSILON
), "Failed at epoch:{} step:{} global_step:{}".format(
epoch, step, gs)
sess.run(
training_op,
feed_dict={
x_batch: batch_x,
y_lab: batch_lab
},
)
def test_lifecycle(
self,
modifier_lambda: Callable[[], GMPruningModifier],
graph_lambda: Callable[[], tf_compat.Graph],
steps_per_epoch: int,
):
modifier = modifier_lambda()
graph = graph_lambda()
with graph.as_default():
global_step = tf_compat.train.get_or_create_global_step()
step_placeholder = tf_compat.placeholder(dtype=tf_compat.int64,
name="step")
global_assign = global_step.assign(step_placeholder)
inp = graph.get_tensor_by_name("inp:0")
out = graph.get_tensor_by_name("out:0")
mod_ops, mod_extras = modifier.create_ops(steps_per_epoch,
global_step, graph)
assert len(mod_ops) == 1
assert mod_ops[0] is not None
assert len(mod_extras) == 1
assert EXTRAS_KEY_SUMMARIES in mod_extras
assert modifier.prune_op_vars
assert len(modifier.prune_op_vars) > 0
last_sparsities = [0.0 for _ in range(len(modifier.prune_op_vars))]
with tf_compat.Session(graph=graph) as sess:
sess.run(tf_compat.global_variables_initializer())
modifier.initialize_session(sess)
step_counter = 0
inp_arr = numpy.random.random((1, *inp.shape[1:]))
for epoch in range(int(modifier.end_epoch + 5.0)):
for step in range(steps_per_epoch):
res = sess.run(out, feed_dict={inp: inp_arr})
assert res.sum() > 0
step_counter += 1
sess.run(global_assign,
feed_dict={step_placeholder: step_counter})
sess.run(mod_ops)
for index, op_vars in enumerate(
modifier.prune_op_vars):
mask_sparsity = eval_tensor_sparsity(op_vars.mask)
masked_sparsity = eval_tensor_sparsity(
op_vars.masked)
assert abs(mask_sparsity - masked_sparsity) < 1e-5
if epoch < modifier.start_epoch:
assert masked_sparsity < 1e-2
else:
assert masked_sparsity == last_sparsities[
index]
last_sparsities[index] = masked_sparsity
modifier.complete_graph(graph, sess)
def test_resnets(key: str, pretrained: Union[bool, str], test_input: bool,
const: Callable):
input_shape = ModelRegistry.input_shape(key)
# test out the stand alone constructor
with tf_compat.Graph().as_default():
inputs = tf_compat.placeholder(tf_compat.float32, [None, *input_shape],
name="inputs")
logits = const(inputs, training=False)
if test_input:
with tf_compat.Session() as sess:
sess.run(tf_compat.global_variables_initializer())
out = sess.run(
logits,
feed_dict={inputs: numpy.random.random((1, *input_shape))})
assert out.sum() != 0
# test out the registry
with tf_compat.Graph().as_default():
inputs = tf_compat.placeholder(tf_compat.float32, [None, *input_shape],
name="inputs")
logits = ModelRegistry.create(key, inputs, training=False)
with tf_compat.Session() as sess:
if test_input:
sess.run(tf_compat.global_variables_initializer())
out = sess.run(
logits,
feed_dict={inputs: numpy.random.random((1, *input_shape))})
assert out.sum() != 0
if pretrained:
ModelRegistry.load_pretrained(key, pretrained)
if test_input:
out = sess.run(
logits,
feed_dict={
inputs: numpy.random.random((1, *input_shape))
},
)
assert out.sum() != 0
def test_module_analyzer(
model, init_weights, layer_name, params, zeroed_params, total_flops
):
g = model(init_weights)
with tf_compat.Session(graph=g) as sess:
sess.run(tf_compat.global_variables_initializer())
desc_dict = analyze_module(sess, g, op_names=[layer_name])
desc = desc_dict[layer_name]
assert desc.params == params
assert desc.zeroed_params == zeroed_params
assert desc.total_flops == total_flops
def test_mnist():
with tf_compat.Graph().as_default():
inputs = tf_compat.placeholder(tf_compat.float32, [None, 28, 28, 1],
name="inputs")
logits = mnist_net(inputs)
with tf_compat.Session() as sess:
sess.run(tf_compat.global_variables_initializer())
out = sess.run(
logits,
feed_dict={inputs: numpy.random.random((1, 28, 28, 1))})
assert out.sum() != 0
def test_approx_ks_loss_sensitivity(net_const: Callable):
with tf_compat.Graph().as_default() as graph:
out, inp = net_const()
with tf_compat.Session() as sess:
sess.run(tf_compat.global_variables_initializer())
analysis = pruning_loss_sens_magnitude(graph)
for res in analysis.results:
assert res.name
assert isinstance(res.index, int)
assert len(res.sparse_measurements) > 0
assert len(res.averages) > 0
assert res.sparse_average > 0
assert res.sparse_integral > 0
def _load_model(args, sess, checkpoint_path=None):
sess.run([
tf_compat.global_variables_initializer(),
tf_compat.local_variables_initializer(),
])
checkpoint_path = checkpoint_path or args.checkpoint_path
ModelRegistry.load_pretrained(
args.arch_key,
pretrained=args.pretrained,
pretrained_dataset=args.pretrained_dataset,
pretrained_path=checkpoint_path,
sess=sess,
)
if checkpoint_path:
LOGGER.info(
"Loaded model weights from checkpoint: {}".format(checkpoint_path))
def test_apply_op_vars_masks(
sparsity_val: float,
net_const: Callable,
inp_arr: numpy.ndarray,
var_names: List[str],
):
group = "test-group"
with tf_compat.Graph().as_default() as graph:
out, inp = net_const()
sparsity = tf_compat.placeholder(dtype=tf_compat.float32,
name="sparsity_placeholder")
update_ready = tf_compat.placeholder(dtype=tf_compat.bool,
name="update_ready")
pruning_op_vars = get_or_create_graph_ops_pruning(
graph,
var_names,
sparsity,
update_ready,
True,
None,
group,
UnstructuredPruningMaskCreator(),
)
with tf_compat.Session() as sess:
sess.run(tf_compat.global_variables_initializer())
for op_vars in pruning_op_vars:
sess.run(
op_vars.update,
feed_dict={
sparsity: sparsity_val,
update_ready: True
},
)
apply_op_vars_masks(pruning_op_vars, group, sess)
for op_vars in pruning_op_vars:
var_sparsity = eval_tensor_sparsity(op_vars.op_input)
assert abs(var_sparsity - sparsity_val) < 1e-2
def test_complete_graph(
self,
modifier_lambda: Callable[[], Modifier],
graph_lambda: Callable[[], tf_compat.Graph],
steps_per_epoch: int,
):
modifier = modifier_lambda()
graph = graph_lambda()
with graph.as_default():
global_step = tf_compat.train.get_or_create_global_step()
with pytest.raises(RuntimeError):
modifier.complete_graph(graph, None)
mod_ops, mod_extras = modifier.create_ops(steps_per_epoch,
global_step, graph)
with tf_compat.Session() as sess:
sess.run(tf_compat.global_variables_initializer())
modifier.initialize_session(sess)
modifier.complete_graph(graph, sess)
def _validate(dataset: ImageFolderDataset, size: int):
with tf_compat.Graph().as_default():
batch_size = 16
with tf_compat.device("/cpu:0"):
print("loading datasets")
dataset_len = len(dataset)
assert dataset_len > 0
tf_dataset = dataset.build(
batch_size,
repeat_count=2,
shuffle_buffer_size=10,
prefetch_buffer_size=batch_size,
num_parallel_calls=4,
)
handle, iterator, (tf_iter, ) = create_split_iterators_handle(
[tf_dataset])
images, labels = iterator.get_next()
with tf_compat.Session() as sess:
sess.run([
tf_compat.global_variables_initializer(),
tf_compat.local_variables_initializer(),
])
iter_handle = sess.run(tf_iter.string_handle())
sess.run(tf_iter.initializer)
for _ in range(5):
batch_x, batch_lab = sess.run([images, labels],
feed_dict={handle: iter_handle})
assert batch_x.shape[0] == 16
assert batch_x.shape[1] == size
assert batch_x.shape[2] == size
assert batch_x.shape[3] == 3
assert batch_lab.shape[0] == 16
assert batch_lab.shape[1] == 10
def test_create_op_pruning_conv(sparsity_val: float,
mask_creator: PruningMaskCreator):
group = "test-group"
is_grouped_mask = isinstance(mask_creator, GroupedPruningMaskCreator)
with tf_compat.Graph().as_default() as graph:
inp = tf_compat.placeholder(tf_compat.float32, [None, 8, 8, 64])
with tf_compat.name_scope("conv"):
weights = tf_compat.Variable(tf_compat.random_normal(
[3, 3, 64, 64]),
name="weights")
bias = tf_compat.Variable(tf_compat.random_normal([64]),
name="bias")
conv = tf_compat.nn.conv2d(inp,
weights,
strides=[1, 1, 1, 1],
padding="SAME",
name="conv")
add = tf_compat.add(conv, bias, name="bias_add")
relu = tf_compat.nn.relu(add, name="relu")
sparsity = tf_compat.placeholder(dtype=tf_compat.float32,
name="sparsity_placeholder")
update_ready = tf_compat.placeholder(dtype=tf_compat.bool,
name="update_ready")
conv_op = graph.get_operation_by_name("conv/conv")
conv_op_input = get_op_input_var(conv_op, VAR_INDEX_FROM_TRAINABLE)
pruning_op_vars = create_op_pruning(
conv_op,
conv_op_input,
sparsity,
update_ready,
True,
None,
group,
mask_creator=mask_creator,
)
with tf_compat.Session() as sess:
sess.run(tf_compat.global_variables_initializer())
sess.run(
pruning_op_vars.update,
feed_dict={
sparsity: sparsity_val,
update_ready: False
},
)
err_threshold = 1e-3 if not is_grouped_mask else 0.05
mask_sparsity = eval_tensor_sparsity(pruning_op_vars.mask)
weight_sparsity = eval_tensor_sparsity(pruning_op_vars.op_input)
assert mask_sparsity < err_threshold
assert abs(mask_sparsity - weight_sparsity) <= 1e-4
masked_sparsity = eval_tensor_sparsity(pruning_op_vars.masked)
assert masked_sparsity < err_threshold
sess.run(
pruning_op_vars.update,
feed_dict={
sparsity: sparsity_val,
update_ready: True
},
)
mask_sparsity = eval_tensor_sparsity(pruning_op_vars.mask)
assert abs(mask_sparsity - sparsity_val) < err_threshold
masked_sparsity = eval_tensor_sparsity(pruning_op_vars.masked)
assert abs(masked_sparsity - sparsity_val) < err_threshold
res = sess.run(relu,
feed_dict={inp: numpy.random.random((4, 8, 8, 64))})
assert res.sum() > 0.0
if is_grouped_mask:
# Check that every value in the mask_creator grouping
# is the same within the mask. Assumes grouping applies
# an absolte mean to each grouping
grouped_mask = mask_creator.group_tensor(pruning_op_vars.mask)
mask_vals_are_grouped = tf_compat.reduce_all(
tf_compat.logical_or(
tf_compat.equal(grouped_mask, 0.0),
tf_compat.equal(grouped_mask, 1.0),
))
assert sess.run(mask_vals_are_grouped)
# limitations under the License.
# flake8: noqa
import math
from sparseml.tensorflow_v1.optim import (
EXTRAS_KEY_LEARNING_RATE,
EXTRAS_KEY_SUMMARIES,
ScheduledModifierManager,
)
from sparseml.tensorflow_v1.utils import tf_compat
with tf_compat.Graph().as_default() as graph:
CREATE_MODEL_GRAPH = None
global_step = tf_compat.train.get_or_create_global_step()
manager = ScheduledModifierManager.from_yaml("/PATH/TO/config.yaml")
mod_ops, mod_extras = manager.create_ops(
steps_per_epoch=math.ceil(len(TRAIN_DATASET) / TRAIN_BATCH_SIZE))
summary_ops = mod_extras[EXTRAS_KEY_SUMMARIES]
learning_rate = mod_extras[EXTRAS_KEY_LEARNING_RATE]
with tf_compat.Session() as sess:
sess.run(tf_compat.global_variables_initializer())
for epoch in range(manager.max_epochs):
for batch in range(TRAIN_BATCH_SIZE):
sess.run(TRAIN_OP)
sess.run(mod_ops)
def test_get_or_create_graph_ops_pruning(
sparsity_val: float,
net_const: Callable,
inp_arr: numpy.ndarray,
var_names: List[str],
mask_creator: PruningMaskCreator,
):
group = "test-group"
is_grouped_mask = isinstance(mask_creator, GroupedPruningMaskCreator)
with tf_compat.Graph().as_default() as graph:
out, inp = net_const()
sparsity = tf_compat.placeholder(dtype=tf_compat.float32,
name="sparsity_placeholder")
update_ready = tf_compat.placeholder(dtype=tf_compat.bool,
name="update_ready")
pruning_op_vars = get_or_create_graph_ops_pruning(
graph,
var_names,
sparsity,
update_ready,
True,
None,
group,
mask_creator,
)
pruning_op_vars_sec = get_or_create_graph_ops_pruning(
graph,
var_names,
sparsity,
update_ready,
True,
None,
group,
mask_creator,
)
assert len(pruning_op_vars) >= len(
var_names) # get at least 1 match per regex
assert len(pruning_op_vars) == len(pruning_op_vars_sec)
for op_vars, op_vars_sec in zip(pruning_op_vars, pruning_op_vars_sec):
assert op_vars.op == op_vars_sec.op
# import pdb
# pdb.set_trace()
assert op_vars.update == op_vars_sec.update
assert op_vars.mask == op_vars_sec.mask
assert op_vars.masked == op_vars_sec.masked
with tf_compat.Session() as sess:
sess.run(tf_compat.global_variables_initializer())
for op_vars in pruning_op_vars:
sess.run(
op_vars.update,
feed_dict={
sparsity: sparsity_val,
update_ready: False
},
)
print(op_vars.mask.shape)
# When we reduce the values a mask can take, there can be higher error
err_threshold = 1e-2 if not is_grouped_mask else 1e-1
mask_sparsity = eval_tensor_sparsity(op_vars.mask)
weight_sparsity = eval_tensor_sparsity(op_vars.op_input)
assert mask_sparsity < err_threshold
assert weight_sparsity == mask_sparsity
masked_sparsity = eval_tensor_sparsity(op_vars.masked)
assert masked_sparsity < err_threshold
sess.run(
op_vars.update,
feed_dict={
sparsity: sparsity_val,
update_ready: True
},
)
mask_sparsity = eval_tensor_sparsity(op_vars.mask)
assert abs(mask_sparsity - sparsity_val) < err_threshold
masked_sparsity = eval_tensor_sparsity(op_vars.masked)
assert abs(masked_sparsity - sparsity_val) < err_threshold
res = sess.run(out, feed_dict={inp: inp_arr})
assert res.sum() > 0.0
if is_grouped_mask:
# Check that every value in the mask_creator grouping
# is the same within the mask. Assumes grouping applies
# an absolte mean to each grouping
grouped_mask = mask_creator.group_tensor(op_vars.mask)
mask_vals_are_grouped = tf_compat.reduce_all(
tf_compat.logical_or(
tf_compat.equal(grouped_mask, 0.0),
tf_compat.equal(grouped_mask, 1.0),
))
assert sess.run(mask_vals_are_grouped)
def test_get_or_create_ks_schedule_ops(
begin_step: int,
end_step: int,
update_step_freq: int,
init_sparsity: float,
final_sparsity: float,
exponent: float,
):
group = "test-group"
with tf_compat.Graph().as_default():
global_step = tf_compat.train.get_or_create_global_step()
step_placeholder = tf_compat.placeholder(dtype=tf_compat.int64,
name="step")
global_assign = global_step.assign(step_placeholder)
update_ready, sparsity = get_or_create_ks_schedule_ops(
global_step,
begin_step,
end_step,
update_step_freq,
init_sparsity,
final_sparsity,
exponent,
group,
)
update_ready_sec, sparsity_sec = get_or_create_ks_schedule_ops(
global_step,
begin_step,
end_step,
update_step_freq,
init_sparsity,
final_sparsity,
exponent,
group,
)
assert update_ready == update_ready_sec
assert sparsity == sparsity_sec
with tf_compat.Session() as sess:
sess.run(tf_compat.global_variables_initializer())
last_update_step = None
last_update_sparsity = None
for step in range(end_step + 10):
sess.run(global_assign, feed_dict={step_placeholder: step})
update_ready_val = sess.run(update_ready)
sparsity_val = sess.run(sparsity)
if step < begin_step:
assert not update_ready_val
assert abs(sparsity_val) < 1e-5
elif step <= begin_step:
assert update_ready_val
assert abs(sparsity_val - init_sparsity) < 1e-5
last_update_step = step
last_update_sparsity = sparsity_val
elif step == end_step:
assert update_ready_val
assert abs(sparsity_val - final_sparsity) < 1e-5
last_update_step = step
last_update_sparsity = sparsity_val
elif step > end_step:
assert not update_ready_val
assert abs(sparsity_val - final_sparsity) < 1e-5
else:
# check if update should be ready
check_ready = (last_update_step is None or
step >= last_update_step + update_step_freq)
assert sparsity_val > last_update_sparsity
if check_ready:
assert update_ready_val
last_update_step = step
last_update_sparsity = sparsity_val
else:
assert not update_ready_val
def test_lrs_with_manager(optim_lambda):
manager = ScheduledModifierManager(modifiers=[
SetLearningRateModifier(learning_rate=0.1, start_epoch=0),
LearningRateModifier(
lr_class="ExponentialLR",
lr_kwargs={"gamma": 0.9},
start_epoch=5,
end_epoch=10,
init_lr=0.01,
),
LearningRateModifier(
lr_class="MultiStepLR",
lr_kwargs={
"gamma": 0.95,
"milestones": [15, 18]
},
start_epoch=12,
end_epoch=20,
init_lr=0.05,
),
])
assert manager.max_epochs == 20
assert manager.min_epochs == 0
graph = mlp_graph_lambda()
steps_per_epoch = 100
with graph.as_default():
global_step = tf_compat.train.get_or_create_global_step()
# Further set up for loss, optimizer and training op
x_batch = graph.get_tensor_by_name("inp:0")
y_pred = graph.get_tensor_by_name("out:0")
n_inputs = x_batch.shape[1]
n_outputs = y_pred.shape[1]
y_lab = tf_compat.placeholder(tf_compat.float32,
shape=(None, n_outputs),
name="y")
mod_ops, mod_extras = manager.create_ops(steps_per_epoch,
global_step=global_step,
graph=graph)
assert len(mod_ops) == 1
assert len(mod_extras) == 2
assert EXTRAS_KEY_LEARNING_RATE in mod_extras
assert EXTRAS_KEY_SUMMARIES in mod_extras
assert isinstance(mod_extras[EXTRAS_KEY_SUMMARIES], list)
assert len(mod_extras[EXTRAS_KEY_SUMMARIES]) == 1
learning_rate = mod_extras[EXTRAS_KEY_LEARNING_RATE]
with tf_compat.name_scope("train"):
optimizer = optim_lambda(learning_rate=learning_rate)
loss = tf_compat.losses.mean_squared_error(y_lab, y_pred)
training_op = optimizer.minimize(loss, global_step=global_step)
np.random.seed(12)
batch_size = 8
batch_x = np.random.randn(batch_size, n_inputs)
batch_lab = np.random.randn(batch_size, n_outputs)
with tf_compat.Session(graph=graph) as sess:
sess.run(tf_compat.global_variables_initializer())
for epoch in range(manager.max_epochs + 5):
# for now hardcoding the tests to get out the door
if epoch < 5:
expected = 0.1
elif epoch < 10:
expected = 0.01 * 0.9**(epoch - 5)
elif epoch < 12:
expected = 0.01 * 0.9**4
elif epoch < 15:
expected = 0.05
elif epoch < 18:
expected = 0.05 * 0.95
else:
expected = 0.05 * 0.95**2
for step in range(steps_per_epoch):
gs = sess.run(global_step)
optim_lr = sess.run(_get_lr(optimizer))
assert (abs(optim_lr - expected) <= EPSILON
), "Failed at epoch:{} step:{} global_step:{}".format(
epoch, step, gs)
sess.run(
training_op,
feed_dict={
x_batch: batch_x,
y_lab: batch_lab
},
)
def test_create_op_pruning_fc(sparsity_val):
group = "test-group"
with tf_compat.Graph().as_default() as graph:
inp = tf_compat.placeholder(tf_compat.float32, [None, 64])
with tf_compat.name_scope("fc"):
weights = tf_compat.Variable(tf_compat.random_normal([64, 64]),
name="weights")
bias = tf_compat.Variable(tf_compat.random_normal([64]),
name="bias")
matmul = tf_compat.matmul(inp, weights, name="matmul")
add = tf_compat.add(matmul, bias, name="bias_add")
relu = tf_compat.nn.relu(add, name="relu")
sparsity = tf_compat.placeholder(dtype=tf_compat.float32,
name="sparsity_placeholder")
update_ready = tf_compat.placeholder(dtype=tf_compat.bool,
name="update_ready")
matmul_op = graph.get_operation_by_name("fc/matmul")
matmul_op_input = get_op_input_var(matmul_op, VAR_INDEX_FROM_TRAINABLE)
pruning_op_vars = create_op_pruning(
matmul_op,
matmul_op_input,
sparsity,
update_ready,
True,
None,
group,
UnstructuredPruningMaskCreator(),
)
with tf_compat.Session() as sess:
sess.run(tf_compat.global_variables_initializer())
sess.run(
pruning_op_vars.update,
feed_dict={
sparsity: sparsity_val,
update_ready: False
},
)
mask_sparsity = eval_tensor_sparsity(pruning_op_vars.mask)
weight_sparsity = eval_tensor_sparsity(pruning_op_vars.op_input)
assert mask_sparsity < 1e-3
assert mask_sparsity == weight_sparsity
masked_sparsity = eval_tensor_sparsity(pruning_op_vars.masked)
assert masked_sparsity < 1e-3
sess.run(
pruning_op_vars.update,
feed_dict={
sparsity: sparsity_val,
update_ready: True
},
)
mask_sparsity = eval_tensor_sparsity(pruning_op_vars.mask)
assert abs(mask_sparsity - sparsity_val) < 1e-3
masked_sparsity = eval_tensor_sparsity(pruning_op_vars.masked)
assert abs(masked_sparsity - sparsity_val) < 1e-3
res = sess.run(relu, feed_dict={inp: numpy.random.random((4, 64))})
assert res.sum() > 0.0
def test_gm_pruning_training_with_manager():
modifier = GMPruningModifier(
params=["mlp_net/fc1/weight", "mlp_net/fc3/weight"],
init_sparsity=0.05,
final_sparsity=0.8,
start_epoch=2.0,
end_epoch=7.0,
update_frequency=1.0,
)
sec_modifier = GMPruningModifier(
params=["mlp_net/fc2/weight"],
init_sparsity=0.05,
final_sparsity=0.8,
start_epoch=2.0,
end_epoch=7.0,
update_frequency=1.0,
)
manager = ScheduledModifierManager([modifier, sec_modifier])
steps_per_epoch = 5
batch_size = 2
with tf_compat.Graph().as_default() as graph:
logits, inputs = mlp_net()
labels = tf_compat.placeholder(tf_compat.float32,
[None, *logits.shape[1:]])
loss = batch_cross_entropy_loss(logits, labels)
global_step = tf_compat.train.get_or_create_global_step()
train_op = tf_compat.train.AdamOptimizer(learning_rate=1e-4).minimize(
loss, global_step=global_step)
mod_ops, mod_extras = manager.create_ops(steps_per_epoch)
last_sparsities = [0.0 for _ in range(len(modifier.prune_op_vars))]
with tf_compat.Session(graph=graph) as sess:
sess.run(tf_compat.global_variables_initializer())
manager.initialize_session(sess)
batch_lab = numpy.random.random((batch_size, *logits.shape[1:]))
batch_inp = numpy.random.random((batch_size, *inputs.shape[1:]))
for epoch in range(int(modifier.end_epoch + 2.0)):
for step in range(steps_per_epoch):
sess.run(train_op,
feed_dict={
inputs: batch_inp,
labels: batch_lab
})
sess.run(global_step)
sess.run(mod_ops)
update_ready_val = sess.run(modifier.update_ready)
sess.run(modifier.sparsity)
for index, op_vars in enumerate(modifier.prune_op_vars):
mask_sparsity = eval_tensor_sparsity(op_vars.mask)
masked_sparsity = eval_tensor_sparsity(op_vars.masked)
assert abs(mask_sparsity - masked_sparsity) < 1e-5
if epoch < modifier.start_epoch:
assert masked_sparsity < 1e-2
assert not update_ready_val
elif epoch >= modifier.end_epoch:
assert abs(masked_sparsity -
modifier.final_sparsity) < 1e-2
assert not update_ready_val
else:
assert masked_sparsity >= last_sparsities[
index] - 1e-2
last_sparsities[index] = masked_sparsity
manager.complete_graph()
for op_vars in modifier.prune_op_vars:
assert (abs(modifier.final_sparsity -
eval_tensor_sparsity(op_vars.op_input)) < 1e-2)
def test_get_or_create_ks_scheduled_graph_ops(
begin_step: int,
end_step: int,
update_step_freq: int,
init_sparsity: float,
final_sparsity: float,
exponent: float,
net_const: Callable,
inp_arr: numpy.ndarray,
var_names: List[str],
):
group = "test-group"
with tf_compat.Graph().as_default() as graph:
global_step = tf_compat.train.get_or_create_global_step()
step_placeholder = tf_compat.placeholder(dtype=tf_compat.int64,
name="step")
global_assign = global_step.assign(step_placeholder)
out, inp = net_const()
(
update_op,
pruning_op_vars,
update_ready,
sparsity,
) = get_or_create_ks_scheduled_graph_ops(
graph,
global_step,
var_names,
begin_step,
end_step,
update_step_freq,
init_sparsity,
final_sparsity,
exponent,
True,
group,
UnstructuredPruningMaskCreator(),
)
(
update_op_sec,
pruning_op_vars_sec,
update_ready,
sparsity,
) = get_or_create_ks_scheduled_graph_ops(
graph,
global_step,
var_names,
begin_step,
end_step,
update_step_freq,
init_sparsity,
final_sparsity,
exponent,
True,
group,
UnstructuredPruningMaskCreator(),
)
assert update_op == update_op_sec
assert update_ready == update_ready
assert sparsity == sparsity
assert len(pruning_op_vars) == 3
assert len(pruning_op_vars) >= len(
var_names) # at least 1 regex match per name
for op_vars, op_vars_sec in zip(pruning_op_vars, pruning_op_vars_sec):
assert op_vars.op == op_vars_sec.op
assert op_vars.update == op_vars_sec.update
assert op_vars.mask == op_vars_sec.mask
assert op_vars.masked == op_vars_sec.masked
with tf_compat.Session() as sess:
sess.run(tf_compat.global_variables_initializer())
last_update_sparsity = None
for step in range(end_step + 10):
sess.run(global_assign, feed_dict={step_placeholder: step})
update_ready_val = sess.run(update_ready)
sparsity_val = sess.run(sparsity)
sess.run(update_op)
for op_var in pruning_op_vars:
mask_sparsity = eval_tensor_sparsity(op_var.mask)
masked_sparsity = eval_tensor_sparsity(op_var.masked)
weight_sparsity = eval_tensor_sparsity(op_vars.op_input)
assert abs(mask_sparsity - masked_sparsity) < 1e-5
if step < begin_step:
assert abs(masked_sparsity) < 1e-2
assert not update_ready_val
elif step == begin_step:
assert abs(masked_sparsity - init_sparsity) < 1e-2
assert abs(sparsity_val - init_sparsity) < 1e-5
assert update_ready_val
last_update_sparsity = masked_sparsity
elif step == end_step:
assert update_ready_val
assert abs(masked_sparsity - final_sparsity) < 1e-2
assert abs(sparsity_val - final_sparsity) < 1e-5
last_update_sparsity = masked_sparsity
elif step > end_step:
assert not update_ready_val
assert abs(masked_sparsity - final_sparsity) < 1e-2
else:
assert masked_sparsity >= last_update_sparsity - 1e-2
assert sparsity_val >= last_update_sparsity - 1e-2
assert abs(weight_sparsity - masked_sparsity) <= 1e-2
last_update_sparsity = masked_sparsity
if step < end_step and update_ready_val:
steps_count = sess.run(global_step) - begin_step
steps_range = end_step - begin_step
expected = _expected_sparsity(
steps_count,
steps_range,
init_sparsity,
final_sparsity,
exponent,
)
assert abs(sparsity_val - expected) < 1e-5
res = sess.run(out, feed_dict={inp: inp_arr})
assert res.sum() >= 0.0
