Skip to content

P
pytensor
提交 c2b3e4fa authored 作者: Mohammad Pezeshki's avatar Mohammad Pezeshki
浏览文件
操作

gradient.py in pep8

上级 020fc625
隐藏空白字符变更
内嵌 并排
正在显示 包含 70 行增加 和 64 行删除
theano/gradient.py
浏览文件 @ c2b3e4fa
......@@ -86,7 +86,7 @@ def grad_not_implemented(op, x_pos, x, comment=""):
return (NullType((
"This variable is Null because the grad method for "
"input %s (%s) of the %s op is not implemented. %s"
) % (x_pos, x, op, comment)))()
) % (x_pos, x, op, comment)))()
def grad_undefined(op, x_pos, x, comment=""):
......@@ -467,16 +467,17 @@ def grad(cost, wrt, consider_constant=None,
g_cost = known_grads[cost]
else:
g_cost = _float_ones_like(cost)
# g_cost may be Disconnected or NullType. A creative use of the function,
# sure, but nonetheless one we can and should support. So before we try
# to cast it make sure it even has a dtype
# g_cost may be Disconnected or NullType. A creative use of the
# function, sure, but nonetheless one we can and should support.
# So before we try to cast it make sure it even has a dtype
if (hasattr(g_cost.type, 'dtype') and
cost.type.dtype not in tensor.discrete_dtypes):
# Here we enforce the constraint that floating point variables have
# the same dtype as their gradient.
g_cost = g_cost.astype(cost.type.dtype)
cost.type.dtype not in tensor.discrete_dtypes):
# Here we enforce the constraint that floating point variables
# have the same dtype as their gradient.
g_cost = g_cost.astype(cost.type.dtype)
# DO NOT enforce g_cost to be 0 if cost is an integer.
# This is to be enforced by the Op.grad method for the Op that outputs cost.
# This is to be enforced by the Op.grad method for the
# Op that outputs cost.
if hasattr(g_cost.type, 'dtype'):
assert g_cost.type.dtype not in tensor.discrete_dtypes
......@@ -491,10 +492,10 @@ def grad(cost, wrt, consider_constant=None,
' or sparse theano variable' % str(type(g_var)))
if (not isinstance(g_var.type, (NullType, DisconnectedType)) and
'float' not in str(g_var.type.dtype)):
'float' not in str(g_var.type.dtype)):
raise TypeError("Gradients must always be NullType, "
"DisconnectedType, or continuous, but grad was "
"given a known_grad of type "+str(g_var.type))
"given a known_grad of type " + str(g_var.type))
# DO NOT check that these gradients are equal to 0 if var is int
# The gradient is allowed to be non-zero on var in that case
......@@ -846,10 +847,10 @@ def _populate_var_to_app_to_idx(outputs, wrt, consider_constant):
if ipt not in var_to_app_to_idx:
# This object here *must* be an OrderedDict, because
# we iterate over its keys when adding up the terms of
# the gradient on ipt. If it is a regular dict, the grad
# method will return something that is analytically correct,
# but whose order of doing additions depends on the memory
# we iterate over its keys when adding up the terms of the
# gradient on ipt. If it is a regular dict, the grad method
# will return something that is analytically correct, but
# whose order of doing additions depends on the memory
# location of the apply nodes.
var_to_app_to_idx[ipt] = OrderedDict()
app_to_idx = var_to_app_to_idx[ipt]
......@@ -923,8 +924,8 @@ def _populate_grad_dict(var_to_app_to_idx,
grad_dict: A dictionary mapping variables to their gradients.
Should be populated by grad function, which should:
-Set the gradient with respect to the cost to 1
-Load all gradients from known_grads, possibly overriding
the cost
-Load all gradients from known_grads, possibly
overriding the cost
-Set the gradient for disconnected
inputs to a variable with type DisconnectedType()
......@@ -1004,10 +1005,10 @@ def _populate_grad_dict(var_to_app_to_idx,
# call the op's grad method
# Each Op's grad function requires inputs and output_grads
# If the Op destroys any input, but the grad expression uses it,
# then chances are the resulting graph will have a dependency
# cycle. We avoid this cycle by passing (symbolic) copies of
# each destroyed input.
# If the Op destroys any input, but the grad expression uses
# it, then chances are the resulting graph will have a
# dependency cycle. We avoid this cycle by passing (symbolic)
# copies of each destroyed input.
try:
dinputs = [node.inputs[x[0]] for x in
node.op.destroy_map.values()]
......@@ -1030,15 +1031,16 @@ def _populate_grad_dict(var_to_app_to_idx,
# If an output is of an integer dtype, then we just leave it
# alone.
# DO NOT force integer variables to have zero grad. This causes
# bugs where we fail to detect disconnected or undefined gradients.
# DO NOT force integer variables to have integer dtype. This is
# a violation of the op contract.
# bugs where we fail to detect disconnected or undefined
# gradients.
# DO NOT force integer variables to have integer dtype.
# This is a violation of the op contract.
new_output_grads = []
for o, og in zip(node.outputs, output_grads):
o_dt = getattr(o.type, 'dtype', None)
og_dt = getattr(og.type, 'dtype', None)
if (o_dt not in theano.tensor.discrete_dtypes and
og_dt and o_dt != og_dt):
og_dt and o_dt != og_dt):
new_output_grads.append(og.astype(o_dt))
else:
new_output_grads.append(og)
......@@ -1049,7 +1051,7 @@ def _populate_grad_dict(var_to_app_to_idx,
o_dt = getattr(o.type, 'dtype', None)
ng_dt = getattr(ng.type, 'dtype', None)
if (ng_dt is not None and
o_dt not in theano.tensor.discrete_dtypes):
o_dt not in theano.tensor.discrete_dtypes):
assert ng_dt == o_dt
# Someone who had obviously not read the Op contract tried
......@@ -1063,14 +1065,15 @@ def _populate_grad_dict(var_to_app_to_idx,
assert (getattr(ng.type, 'dtype', None)
not in theano.tensor.discrete_dtypes)
# If config.compute_test_value is turned on, check that the gradients
# on the outputs of this node have the right shape.
# We also check the gradient on the inputs later--both checks are needed,
# because some gradients are only ever specified by the user, not computed
# by Op.grad, and some gradients are only computed and returned, but never
# passed as another node's output grads.
# If config.compute_test_value is turned on, check that the
# gradients on the outputs of this node have the right shape.
# We also check the gradient on the inputs later--both checks
# are needed, because some gradients are only ever specified
# by the user, not computed by Op.grad, and some gradients are
# only computed and returned, but never passed as another
# node's output grads.
for idx, packed in enumerate(izip(node.outputs,
new_output_grads)):
new_output_grads)):
orig_output, new_output_grad = packed
if not hasattr(orig_output, 'shape'):
continue
......@@ -1098,14 +1101,14 @@ def _populate_grad_dict(var_to_app_to_idx,
# We can not enforce this, as AdvancedSubtensor1 has an option to
# return the sparse grad for optimization reason.
# for ig, i in zip(input_grads, inputs):
# if (not isinstance(ig.type, (DisconnectedType, NullType)) and
# type(ig.type) != type(i.type)):
# raise ValueError(
# "%s returned the wrong type for gradient terms."
# " Sparse inputs must have sparse grads and dense"
# " inputs must have dense grad. Got %s, expected %s" % (
# str(node.op), ig.type, i.type))
# for ig, i in zip(input_grads, inputs):
# if (not isinstance(ig.type, (DisconnectedType, NullType)) and
# type(ig.type) != type(i.type)):
# raise ValueError(
# "%s returned the wrong type for gradient terms."
# " Sparse inputs must have sparse grads and dense"
# " inputs must have dense grad. Got %s, expected %s" %(
# str(node.op), ig.type, i.type))
# must convert to list in case the op returns a tuple
# we won't be able to post-process out the Nones if it does that
......@@ -1138,7 +1141,8 @@ def _populate_grad_dict(var_to_app_to_idx,
'the grad_undefined or grad_unimplemented helper '
'functions.') % node.op)
# Check that the gradient term for this input has the right shape
# Check that the gradient term for this input
# has the right shape
if hasattr(term, 'shape'):
orig_ipt = inputs[i]
for orig_ipt_v, term_v in get_debug_values(orig_ipt, term):
......@@ -1389,7 +1393,8 @@ class numeric_grad(object):
# create un-initialized memory
x = numpy.ndarray((total_size,), dtype=working_dtype)
if (not out_type is None) and (out_type.startswith('complex')):
# (not out_type is None) --> (out_type is not None) ???
if (out_type is not None) and (out_type.startswith('complex')):
gx = numpy.ndarray((total_size,), dtype=out_type)
else:
gx = numpy.ndarray((total_size,), dtype=working_dtype)
......@@ -1734,10 +1739,10 @@ def jacobian(expression, wrt, consider_constant=None,
from theano.tensor import arange
# Check inputs have the right format
assert isinstance(expression, Variable), \
"tensor.jacobian expects a Variable as `expression`"
"tensor.jacobian expects a Variable as `expression`"
assert expression.ndim < 2, \
("tensor.jacobian expects a 1 dimensional variable as "
"`expression`. If not use flatten to make it a vector")
("tensor.jacobian expects a 1 dimensional variable as "
"`expression`. If not use flatten to make it a vector")
using_list = isinstance(wrt, list)
using_tuple = isinstance(wrt, tuple)
......@@ -1774,9 +1779,9 @@ def jacobian(expression, wrt, consider_constant=None,
sequences=arange(expression.shape[0]),
non_sequences=[expression] + wrt)
assert not updates, \
("Scan has returned a list of updates. This should not "
"happen! Report this to theano-users (also include the "
"script that generated the error)")
("Scan has returned a list of updates. This should not "
"happen! Report this to theano-users (also include the "
"script that generated the error)")
return format_as(using_list, using_tuple, jacobs)
......@@ -1808,9 +1813,9 @@ def hessian(cost, wrt, consider_constant=None,
from theano.tensor import arange
# Check inputs have the right format
assert isinstance(cost, Variable), \
"tensor.hessian expects a Variable as `cost`"
"tensor.hessian expects a Variable as `cost`"
assert cost.ndim == 0, \
"tensor.hessian expects a 0 dimensional variable as `cost`"
"tensor.hessian expects a 0 dimensional variable as `cost`"
using_list = isinstance(wrt, list)
using_tuple = isinstance(wrt, tuple)
......@@ -1823,10 +1828,10 @@ def hessian(cost, wrt, consider_constant=None,
hessians = []
for input in wrt:
assert isinstance(input, Variable), \
"tensor.hessian expects a (list of) Variable as `wrt`"
"tensor.hessian expects a (list of) Variable as `wrt`"
assert input.ndim == 1, \
"tensor.hessian expects a (list of) 1 dimensional variable "\
"as `wrt`"
"tensor.hessian expects a (list of) 1 dimensional variable "\
"as `wrt`"
expr = grad(cost, input, consider_constant=consider_constant,
disconnected_inputs=disconnected_inputs)
......@@ -1834,16 +1839,16 @@ def hessian(cost, wrt, consider_constant=None,
# even if they are connected to cost.
# This should not be an error.
hess, updates = theano.scan(lambda i, y, x: grad(
y[i],
x,
consider_constant=consider_constant,
disconnected_inputs='ignore'),
sequences=arange(expr.shape[0]),
non_sequences=[expr, input])
y[i],
x,
consider_constant=consider_constant,
disconnected_inputs='ignore'),
sequences=arange(expr.shape[0]),
non_sequences=[expr, input])
assert not updates, \
("Scan has returned a list of updates. This should not "
"happen! Report this to theano-users (also include the "
"script that generated the error)")
("Scan has returned a list of updates. This should not "
"happen! Report this to theano-users (also include the "
"script that generated the error)")
hessians.append(hess)
return format_as(using_list, using_tuple, hessians)
......@@ -1974,6 +1979,7 @@ def disconnected_grad(x):
class GradClip(ViewOp):
# See doc in user fct grad_clip
__props__ = ()
def __init__(self, clip_lower_bound, clip_upper_bound):
# We do not put those member in __eq__ or __hash__
# as they do not influence the perform of this op.
......@@ -1996,7 +2002,7 @@ def grad_clip(x, lower_bound, upper_bound):
:param x: the variable we want its gradient inputs clipped
:param lower_bound: The lower bound of the gradient value
:param upper_bound: The upper bound of the gradient value.
:examples:
x = theano.tensor.scalar()
......
Markdown 格式
0% 或
您添加了 0 人 到此讨论。请谨慎行事。
请先完成此评论的编辑！
请 注册 或者 后发表评论