def rand_crop(img: tf_compat.Tensor):
with tf_compat.name_scope(name):
orig_shape = tf_compat.shape(img)
scale = tf_compat.random_uniform(shape=[1],
minval=scale_range[0],
maxval=scale_range[1])[0]
ratio = tf_compat.random_uniform(shape=[1],
minval=ratio_range[0],
maxval=ratio_range[1])[0]
height = tf_compat.minimum(
tf_compat.cast(
tf_compat.round(
tf_compat.cast(orig_shape[0], dtype=tf_compat.float32)
* scale / ratio),
tf_compat.int32,
),
orig_shape[0],
)
width = tf_compat.minimum(
tf_compat.cast(
tf_compat.round(
tf_compat.cast(orig_shape[1], dtype=tf_compat.float32)
* scale),
tf_compat.int32,
),
orig_shape[1],
)
img = tf_compat.random_crop(img, [height, width, orig_shape[2]])
return img
def _calc_lr():
less = tf_compat.cast(
tf_compat.greater_equal(global_step, milestone_steps),
tf_compat.int64)
updates = tf_compat.reduce_sum(less)
mult_g = tf_compat.pow(gamma,
tf_compat.cast(updates, tf_compat.float32))
return tf_compat.multiply(init_lr, mult_g)
def zero_fraction(inp: tf_compat.Tensor):
nonzero = tf_compat.cast(
tf_compat.reduce_sum(
tf_compat.cast(tf_compat.not_equal(inp, 0),
tf_compat.int64)),
tf_compat.float32,
)
size = tf_compat.size(inp, out_type=tf_compat.float32)
return 1 - tf_compat_div(nonzero, size)
def _calc_lr():
steps = tf_compat.subtract(global_step, start_step)
updates = tf_compat.cond(
after,
lambda: max_updates,
lambda: tf_compat.cast(
tf_compat.floor(tf_compat.divide(steps, step_size)),
tf_compat.int64,
),
)
mult_g = tf_compat.pow(gamma,
tf_compat.cast(updates, tf_compat.float32))
return tf_compat.multiply(init_lr, mult_g)
def _set_constant_mask():
# Assign mask tensor to be 1 for all nonzero values of op_var_tens otherwise 0
# On end step, revert mask to be all 1s
with tf_compat.name_scope(
PruningScope.model(
op,
ks_group,
additional=PruningScope.OPS_UPDATE,
trailing_slash=True,
)):
new_mask = tf_compat.cond(
is_start_step,
lambda: tf_compat.cast(tf_compat.not_equal(op_var_tens, 0.0),
dtype=op_var_tens.dtype),
lambda: tf_compat.ones(op_var_tens.shape,
dtype=op_var_tens.dtype),
)
weight_var = get_tensor_var(op_var_tens)
return tf_compat.group(
tf_compat.assign(mask,
new_mask,
name=PruningScope.OP_MASK_ASSIGN),
tf_compat.assign(weight_var,
masked,
name=PruningScope.OP_WEIGHT_UPDATE),
)
def cent_crop(img: tf_compat.Tensor):
with tf_compat.name_scope(name):
orig_shape = tf_compat.shape(img)
min_size = tf_compat.cond(
tf_compat.greater_equal(orig_shape[0], orig_shape[1]),
lambda: orig_shape[1],
lambda: orig_shape[0],
)
if padding > 0:
orig_shape_list = img.get_shape().as_list()
resize((orig_shape_list[0] + 2 * padding,
orig_shape_list[1] + 2 * padding))
padding_height = tf_compat.add(
tf_compat.cast(
tf_compat.round(
tf_compat.div(
tf_compat.cast(
tf_compat.subtract(orig_shape[0], min_size),
tf_compat.float32,
),
2.0,
)),
tf_compat.int32,
),
padding,
)
padding_width = tf_compat.add(
tf_compat.cast(
tf_compat.round(
tf_compat.div(
tf_compat.cast(
tf_compat.subtract(orig_shape[1], min_size),
tf_compat.float32,
),
2.0,
)),
tf_compat.int32,
),
padding,
)
img = tf_compat.image.crop_to_bounding_box(img, padding_height,
padding_width, min_size,
min_size)
return img
def create_sparsity_mask(
self,
tensor: tf_compat.Tensor,
sparsity: tf_compat.Tensor,
) -> tf_compat.Tensor:
"""
:param tensor: A tensor of a model layer's weights
:param sparsity: the target sparsity to use for assigning the masks
:return: A sparsity mask close to the set sparsity based on the values of
the input tensor
"""
abs_var = tf_compat.abs(tensor) # Magnitudes of weights
sparse_threshold_index = tf_compat.cast(
tf_compat.round(
tf_compat.cast(tf_compat.size(abs_var), tf_compat.float32) *
sparsity),
tf_compat.int32,
)
sparse_threshold_index = tf_compat.minimum(
tf_compat.maximum(sparse_threshold_index, 0),
tf_compat.size(tensor) - 1,
)
try:
argsort = tf_compat.argsort
except Exception:
try:
argsort = tf_compat.contrib.framework.argsort
except Exception:
raise RuntimeError(
"cannot find argsort function in tensorflow_v1, "
"currently unsupported")
# produce tensor where each element is the index in sorted order of abs_var
abs_var_flat = tf_compat.reshape(abs_var, [-1])
element_ranks_flat = tf_compat.scatter_nd(
tf_compat.expand_dims(argsort(abs_var_flat), 1),
tf_compat.range(abs_var_flat.get_shape()[0].value),
abs_var_flat.get_shape(),
)
element_ranks = tf_compat.reshape(element_ranks_flat,
abs_var.get_shape())
return tf_compat.cast(
tf_compat.greater(element_ranks, sparse_threshold_index),
tf_compat.float32,
)
def multi_step_lr_schedule(
global_step: tf_compat.Tensor,
start_step: int,
milestone_steps: List[int],
init_lr: float,
gamma: float,
name: str = "multi_step_lr_schedule",
):
"""
Create a multi step learning rate schedule in the current graph.
Multiplies init_lr by gamma after each milestone has passed.
Ex: lr = init_lr * (gamma ** NUM_UPDATES)
:param global_step: the global step used for training
:param start_step: the step to start the exponential schedule on
:param milestone_steps: a list of steps to decrease the learning rate at,
these are the number of steps that must pass after start_step to decrease lr
:param init_lr: the learning rate to start the schedule with
:param gamma: the decay weight to decrease init_lr by after every step_size interval
:param name: the name scope to create the graph under
:return: the calculated learning rate tensor
"""
with tf_compat.name_scope(name):
global_step = tf_compat.cast(global_step, tf_compat.int64)
milestone_steps = tf_compat.constant(
[mile + start_step for mile in milestone_steps],
dtype=tf_compat.int64,
name="milestone_steps",
)
start_step = tf_compat.constant(start_step,
dtype=tf_compat.int64,
name="start_step")
init_lr = tf_compat.constant(init_lr,
dtype=tf_compat.float32,
name="init_lr")
gamma = tf_compat.constant(gamma,
dtype=tf_compat.float32,
name="gamma")
before = tf_compat.less(global_step, start_step, name="before")
def _calc_lr():
less = tf_compat.cast(
tf_compat.greater_equal(global_step, milestone_steps),
tf_compat.int64)
updates = tf_compat.reduce_sum(less)
mult_g = tf_compat.pow(gamma,
tf_compat.cast(updates, tf_compat.float32))
return tf_compat.multiply(init_lr, mult_g)
learning_rate = tf_compat.cond(before,
lambda: init_lr,
_calc_lr,
name="learning_rate")
return learning_rate
def non_zero_mask_initializer(
shape: tf_compat.TensorShape,
dtype: tf_compat.DType = tf_compat.float32,
partition_info: Any = None, # unsued variable for compatability
) -> tf_compat.Tensor:
dtype = tf_compat.as_dtype(dtype)
if not dtype.is_numpy_compatible or dtype == tf_compat.string:
raise ValueError("Expected numeric or boolean dtype, got %s." %
dtype)
return tf_compat.cast(tf_compat.not_equal(tensor, 0.0),
dtype=dtype)
def preprocess_for_eval(image: tf_compat.Tensor):
"""
The default preprocessing function for test set as defined in Resnet paper
for Cifar datasets
:param image: the image tensor
:return: the preprocessed image
"""
with tf_compat.name_scope("test_preprocess"):
image = tf_compat.cast(image, dtype=tf_compat.float32)
image = tf_compat_div(image, 255.0)
image = tf_compat.image.random_crop(image, [32, 32, 3])
return image
def create_ops(
self,
steps_per_epoch: int,
global_step: tf_compat.Variable,
graph: tf_compat.Graph,
) -> Tuple[List[Union[tf_compat.Tensor, tf_compat.Operation]], Dict[str, Any]]:
"""
Create switch case computing the learning rate at a given global step and
extras created by individual LR modifiers
:param steps_per_epoch: the number of steps per training epoch
:param global_step: the global step used while training
:param graph: the graph to be modified
:return: a tuple (list of empty ops, dict of named ops/tensors for learning
rate and summaries as extras)
"""
mod_ops, mod_extras = super().create_ops(graph, steps_per_epoch, global_step)
name_scope = "{}/{}".format(NM_RECAL, self.__class__.__name__)
with graph.as_default():
with tf_compat.name_scope(name_scope):
pred_fn_pairs = []
global_step = tf_compat.cast(global_step, tf_compat.int64)
for index, child in enumerate(self._lr_modifiers):
with tf_compat.name_scope(str(index)):
_, child_extras = child.create_ops(
steps_per_epoch, global_step, graph
)
child_lr = child_extras[EXTRAS_KEY_LEARNING_RATE]
child_start_step, _ = child.start_end_steps(
steps_per_epoch, after_optim=False
)
child_select = tf_compat.greater_equal(
global_step,
tf_compat.constant(child_start_step, tf_compat.int64),
name="active",
)
pred_fn_pairs.append((child_select, lambda lr=child_lr: lr))
learning_rate = tf_compat.case(pred_fn_pairs)
_add_lr_extras(mod_extras, learning_rate, self.log_types)
return mod_ops, mod_extras
def processor(self, file_path: tf_compat.Tensor, label: tf_compat.Tensor):
"""
:param file_path: the path to the file to load an image from
:param label: the label for the given image
:return: a tuple containing the processed image and label
"""
with tf_compat.name_scope("img_to_tensor"):
img = tf_compat.read_file(file_path)
# Decode and reshape the image to 3 dimensional tensor
# Note: "expand_animations" not available for TF 1.13 and prior,
# hence the reshape trick below
img = tf_compat.image.decode_image(img)
img_shape = tf_compat.shape(img)
img = tf_compat.reshape(img,
[img_shape[0], img_shape[1], img_shape[2]])
img = tf_compat.cast(img, dtype=tf_compat.float32)
if self.pre_resize_transforms:
transforms = (self.pre_resize_transforms.train
if self.train else self.pre_resize_transforms.val)
if transforms:
with tf_compat.name_scope("pre_resize_transforms"):
for trans in transforms:
img = trans(img)
if self._image_size:
res_callable = resize((self.image_size, self.image_size))
img = res_callable(img)
if self.post_resize_transforms:
transforms = (self.post_resize_transforms.train
if self.train else self.post_resize_transforms.val)
if transforms:
with tf_compat.name_scope("post_resize_transforms"):
for trans in transforms:
img = trans(img)
return img, label
def preprocess_for_train(image: tf_compat.Tensor):
"""
The default preprocessing function for train set as defined in Resnet paper
for Cifar datasets
:param image: the image tensor
:return: the preprocessed image
"""
with tf_compat.name_scope("train_preprocess"):
image = tf_compat.cast(image, dtype=tf_compat.float32)
rand_choice = tf_compat.random_uniform(shape=[],
minval=0,
maxval=2,
dtype=tf_compat.int32)
padding = _PADDING
image = tf_compat.cond(
tf_compat.equal(rand_choice, 0),
lambda: tf_compat.pad(image, [[padding, padding],
[padding, padding], [0, 0]]),
lambda: tf_compat.image.random_flip_left_right(image),
)
distorted_image = tf_compat.image.random_crop(image, [32, 32, 3])
return distorted_image
def create_ks_schedule_ops(
global_step: tf_compat.Variable,
begin_step: int,
end_step: int,
update_step_freq: int,
init_sparsity: float,
final_sparsity: float,
exponent: float,
ks_group: str,
) -> Tuple[tf_compat.Tensor, tf_compat.Tensor]:
"""
Create a gradual schedule for model pruning (kernel sparsity).
Creates a sparsity tensor that goes from init_sparsity til final_sparsity
starting at begin_step and ending at end_step.
Uses the global_step to map those.
Additionally creates an update_ready tensor that is True if an update
to the sparsity tensor should be run, False otherwise.
:param global_step: the global optimizer step for the training graph
:param begin_step: the global step to begin pruning at
:param end_step: the global step to end pruning at
:param update_step_freq: the number of global steps between each weight update
:param init_sparsity: the starting value for sparsity of a
weight tensor to be enforce
:param final_sparsity: the end value for sparsity for a weight tensor to be enforce
:param exponent: the exponent to use for interpolating between
init_sparsity and final_sparsity higher values will lead to larger sparsity
steps at the beginning vs the end ie: linear (1) vs cubic (3)
:param ks_group: the group identifier the scope should be created under
:return: a tuple containing the signal for update_ready and the target sparsity
"""
# create the scheduling ops first and the sparsity ops
with tf_compat.name_scope(
PruningScope.general(ks_group,
additional=PruningScope.OPS_SCHEDULE,
trailing_slash=True)):
sched_before = tf_compat.less(global_step, begin_step)
sched_start = tf_compat.equal(global_step, begin_step)
sched_end = tf_compat.equal(global_step, end_step)
sched_active = tf_compat.logical_and(
tf_compat.greater(global_step, begin_step),
tf_compat.less(global_step, end_step),
)
sched_active_inclusive = tf_compat.logical_or(
sched_active, tf_compat.logical_or(sched_start, sched_end))
sched_update = tf_compat.cond(
tf_compat.less_equal(update_step_freq, 0),
lambda: tf_compat.constant(True),
lambda: tf_compat.equal(
tf_compat.mod(
(global_step - begin_step), update_step_freq), 0),
)
sched_update_ready = tf_compat.logical_or(
tf_compat.logical_or(sched_start, sched_end), sched_update)
percentage = tf_compat.minimum(
1.0,
tf_compat.maximum(
0.0,
tf_compat_div(
tf_compat.cast(global_step - begin_step,
tf_compat.float32),
end_step - begin_step,
),
),
)
exp_percentage = 1 - tf_compat.pow(1 - percentage, exponent)
calc_sparsity = (tf_compat.multiply(final_sparsity - init_sparsity,
exp_percentage) + init_sparsity)
# create the update ready tensor and sparsity tensor
with tf_compat.name_scope(
PruningScope.general(ks_group, trailing_slash=True)):
update_ready = tf_compat.logical_and(
sched_active_inclusive,
sched_update_ready,
name=PruningScope.OP_UPDATE_READY,
)
sparsity = tf_compat.case(
[
(sched_before, lambda: tf_compat.constant(0.0)),
(sched_start, lambda: tf_compat.constant(init_sparsity)),
(sched_active, lambda: calc_sparsity),
],
default=lambda: tf_compat.constant(final_sparsity),
name=PruningScope.OP_SPARSITY,
)
# add return state to collections
tf_compat.add_to_collection(
PruningScope.collection_name(ks_group, PruningScope.OP_UPDATE_READY),
update_ready,
)
tf_compat.add_to_collection(
PruningScope.collection_name(ks_group, PruningScope.OP_SPARSITY),
sparsity)
return update_ready, sparsity
def step_lr_schedule(
global_step: tf_compat.Tensor,
start_step: int,
end_step: int,
step_size: int,
init_lr: float,
gamma: float,
name: str = "exponential_lr_schedule",
) -> tf_compat.Tensor:
"""
Create an exponential learning rate schedule in the current graph.
Multiplies init_lr by gamma after each step_size interval has passed.
Ex: lr = init_lr * (gamma ** NUM_UPDATES)
:param global_step: the global step used for training
:param start_step: the step to start the exponential schedule on
:param end_step: the step to end the exponential schedule on,
can be set to -1 and in that event will continually update the LR
:param step_size: the number of steps between each gamma update to the init_lr
:param init_lr: the learning rate to start the schedule with
:param gamma: the decay weight to decrease init_lr by after every step_size interval
:param name: the name scope to create the graph under
:return: the calculated learning rate tensor
"""
with tf_compat.name_scope(name):
global_step = tf_compat.cast(global_step, tf_compat.int64)
max_updates = tf_compat.constant(
(end_step - start_step) // step_size if end_step > 0 else -1,
dtype=tf_compat.int64,
name="max_updates",
)
start_step = tf_compat.constant(start_step,
dtype=tf_compat.int64,
name="start_step")
end_step = tf_compat.constant(end_step,
dtype=tf_compat.int64,
name="end_step")
init_lr = tf_compat.constant(init_lr,
dtype=tf_compat.float32,
name="init_lr")
step_size = tf_compat.constant(step_size,
dtype=tf_compat.int64,
name="step_size")
gamma = tf_compat.constant(gamma,
dtype=tf_compat.float32,
name="gamma")
before = tf_compat.less(global_step, start_step, name="before")
after = tf_compat.logical_and(
tf_compat.greater_equal(global_step, end_step, name="after"),
tf_compat.not_equal(end_step,
tf_compat.constant(-1, tf_compat.int64)),
)
def _calc_lr():
steps = tf_compat.subtract(global_step, start_step)
updates = tf_compat.cond(
after,
lambda: max_updates,
lambda: tf_compat.cast(
tf_compat.floor(tf_compat.divide(steps, step_size)),
tf_compat.int64,
),
)
mult_g = tf_compat.pow(gamma,
tf_compat.cast(updates, tf_compat.float32))
return tf_compat.multiply(init_lr, mult_g)
learning_rate = tf_compat.cond(before,
lambda: init_lr,
_calc_lr,
name="learning_rate")
return learning_rate
