Skip to content

P
pytensor
提交 70b5f2c1 authored 作者: Iban Harlouchet's avatar Iban Harlouchet 提交者： Frederic
浏览文件
操作

flake8 for tensor/nnet/nnet.py

上级 f4edcc59
显示空白字符变更
内嵌 并排
正在显示 包含 39 行增加 和 38 行删除
theano/tensor/nnet/nnet.py
浏览文件 @ 70b5f2c1
......@@ -15,6 +15,7 @@ from six.moves import xrange
import theano
from theano import gof
from theano import scalar
from theano.tensor import basic as tensor
from theano.tensor import subtensor
from theano.tensor import elemwise
......@@ -27,12 +28,12 @@ from theano.gradient import DisconnectedType
from theano.gradient import grad_not_implemented
from theano.tensor.type import values_eq_approx_remove_nan
############
#
# TENSOR OPS
#
class SoftmaxWithBias(gof.Op):
"""
An L{Op} for the output of neural-net multiclass classifiers.
......@@ -300,11 +301,11 @@ class SoftmaxGrad(gof.Op):
dy, sm = inp
g, = grads
tmp = g + tensor.neg(tensor.sum(g*sm, axis=1).dimshuffle((0, 'x')))
tmp = g + tensor.neg(tensor.sum(g * sm, axis=1).dimshuffle((0, 'x')))
g_dy = tmp * sm
tmp2 = tensor.sum(dy*sm, axis=1).dimshuffle((0, 'x'))
g_sm = tmp*dy - g *tmp2
tmp2 = tensor.sum(dy * sm, axis=1).dimshuffle((0, 'x'))
g_sm = tmp * dy - g * tmp2
return g_dy, g_sm
......@@ -571,12 +572,15 @@ class Softmax(gof.Op):
softmax_op = Softmax()
def softmax_graph(c):
return tensor.exp(c) / tensor.exp(c).sum(axis=-1, keepdims=True)
def softmax(c):
return softmax_op(c)
@opt.register_specialize('fast_compile_gpu')
@gof.local_optimizer([softmax_op])
def local_softmax_with_bias(node):
......@@ -593,9 +597,9 @@ def local_softmax_with_bias(node):
# tensor.DimShuffle) since specialization comes
# relatively late in optimization, we don't want to
# put in extra DimShuffles un-necessarily.
if (x_in.owner and isinstance(x_in.owner.op,
tensor.DimShuffle)
and list(x_in.owner.inputs[0].type.broadcastable) == [False]):
if (x_in.owner and
isinstance(x_in.owner.op, tensor.DimShuffle) and
list(x_in.owner.inputs[0].type.broadcastable) == [False]):
# cut out the DimShuffle that was broadcasting a vector
vectors.append(x_in.owner.inputs[0])
else:
......@@ -673,8 +677,7 @@ def softmax_simplifier(numerators, denominators):
numerators.append(softmax_op(x))
return numerators, denominators
opt.local_mul_canonizer.add_simplifier(softmax_simplifier,
'softmax_simplifier')
opt.local_mul_canonizer.add_simplifier(softmax_simplifier, 'softmax_simplifier')
if 0:
@opt.register_specialize
......@@ -836,7 +839,7 @@ class CrossentropySoftmaxArgmax1HotWithBias(gof.Op):
# TODO: Is this correct? It used to be y, not y_idx
nll = tensor.TensorType(x.type.dtype,
y_idx.type.broadcastable)()
y_idx.type.broadcastable).make_variable()
# nll = TensorType(x.dtype, y.broadcastable)
sm = x.type()
am = y_idx.type()
......@@ -866,15 +869,14 @@ class CrossentropySoftmaxArgmax1HotWithBias(gof.Op):
if any(y_idx < 0):
raise ValueError("y_i value out of bounds")
sm = numpy.zeros_like(x) # softmax
nll = numpy.zeros(x.shape[0], dtype=node.outputs[0].type.
dtype) # nll(y | softmax(x))
nll = numpy.zeros(x.shape[0], dtype=node.outputs[0].type.dtype) # nll(y | softmax(x))
am = numpy.zeros_like(y_idx)
for i in xrange(sm.shape[0]):
# add the bias vector to the i'th row of x
row = x[i] + b
# get the maximum value of i'th row for numerically safe
#softmax / nll
# softmax / nll
am[i] = numpy.argmax(row)
m = row[am[i]]
......@@ -1083,8 +1085,7 @@ class CrossentropySoftmax1HotWithBiasDx(gof.Op):
y_idx_range = tensor.arange(y_idx.shape[0])
g_dy = tensor.sum(
g_dx * subtensor.AdvancedIncSubtensor()(
sm, tensor.fill(dy, -1), y_idx_range, y_idx),
axis=1)
sm, tensor.fill(dy, -1), y_idx_range, y_idx), axis=1)
g_sm = dy.dimshuffle(0, 'x') * g_dx
g_y_idx = grad_not_implemented(self, 2, y_idx)
return [g_dy, g_sm, g_y_idx]
......@@ -1226,8 +1227,7 @@ def crossentropy_softmax_max_and_argmax_1hot_with_bias(x, b, y_idx, **kwargs):
unnecessary? e.g. CrossentropySoftmaxArgmax1HotWithBias should return
the appropriate information (i.e. the max probability)?
"""
(xent, softmax) = crossentropy_softmax_1hot_with_bias(x, b, y_idx,
**kwargs)
(xent, softmax) = crossentropy_softmax_1hot_with_bias(x, b, y_idx, **kwargs)
(max_pr, argmax) = tensor.max_and_argmax(softmax, axis=-1)
return (xent, softmax, max_pr, argmax)
......@@ -1251,8 +1251,8 @@ class CrossentropyCategorical1HotGrad(gof.Op):
g_coding_strg, = out
g_coding = numpy.zeros_like(coding_dist)
for i in xrange(len(g_y)):
g_coding[i, true_one_of_n[i]] = -g_y[i] / coding_dist[i,
true_one_of_n[i]]
g_coding[i, true_one_of_n[i]] = (-g_y[i] /
coding_dist[i, true_one_of_n[i]])
g_coding_strg[0] = g_coding
def infer_shape(self, node, in_shapes):
......@@ -1346,9 +1346,10 @@ def crossentropy_to_crossentropy_with_softmax_with_bias(fgraph):
sm, one_of_n = node.inputs
if sm.owner and sm.owner.op == softmax_with_bias:
x, b = sm.owner.inputs
new_nll, new_sm, new_am = crossentropy_softmax_argmax_1hot_with_bias(x, b,
one_of_n)
fgraph.replace_all_validate([(nll, new_nll), (sm, new_sm)],
new_nll, new_sm, new_am = crossentropy_softmax_argmax_1hot_with_bias(
x, b, one_of_n)
fgraph.replace_all_validate(
[(nll, new_nll), (sm, new_sm)],
reason="crossentropy_to_crossentropy_with_softmax_with_bias")
return True
......@@ -1381,16 +1382,18 @@ def crossentropy_to_crossentropy_with_softmax(fgraph):
sm, one_of_n = node.inputs
if sm.owner and sm.owner.op == softmax_op:
x, = sm.owner.inputs
new_nll, new_sm, new_am = crossentropy_softmax_argmax_1hot_with_bias(x,
tensor.zeros_like(x[0]), one_of_n)
fgraph.replace_all_validate([(nll, new_nll), (sm, new_sm)],
new_nll, new_sm, new_am = crossentropy_softmax_argmax_1hot_with_bias(
x, tensor.zeros_like(x[0]), one_of_n)
fgraph.replace_all_validate(
[(nll, new_nll), (sm, new_sm)],
reason="crossentropy_to_crossentropy_with_softmax")
return True
if sm.owner and sm.owner.op == softmax_with_bias:
x, b = sm.owner.inputs
new_nll, new_sm, new_am = crossentropy_softmax_argmax_1hot_with_bias(x, b,
one_of_n)
fgraph.replace_all_validate([(nll, new_nll), (sm, new_sm)],
fgraph.replace_all_validate(
[(nll, new_nll), (sm, new_sm)],
reason="crossentropy_to_crossentropy_with_softmax")
return True
......@@ -1415,8 +1418,8 @@ def local_softmax_grad_to_crossentropy_with_softmax_grad(node):
if (g_coding_dist.owner and
g_coding_dist.owner.op == crossentropy_categorical_1hot_grad):
g_nll, coding_dist, true_one_of_n = g_coding_dist.owner.inputs
dx = crossentropy_softmax_1hot_with_bias_dx(g_nll,
coding_dist, true_one_of_n)
dx = crossentropy_softmax_1hot_with_bias_dx(g_nll, coding_dist,
true_one_of_n)
return [dx]
......@@ -1428,7 +1431,8 @@ def local_argmax_pushdown(node):
(softmax_op, softplus, tensor.exp, tensor.log, tensor.tanh, sigmoid,
softmax_with_bias):
if theano.config.warn.argmax_pushdown_bug:
logging.getLogger('theano.tensor.nnet.nnet').warn("WARNING: there "
logging.getLogger('theano.tensor.nnet.nnet').warn(
"WARNING: there "
"was a bug in Theano fixed on May 27th, 2010 in this case."
" I.E. when we take the max of a softplus, softmax, exp, "
"log, tanh, sigmoid, softmax_with_bias op, we were doing "
......@@ -1657,15 +1661,15 @@ def local_advanced_indexing_crossentropy_onehot_grad(node):
if isinstance(denom.owner.op, subtensor.AdvancedSubtensor):
# Base case
adv_subtensor = denom
#out_grad /= 1.
# out_grad /= 1.
elif denom.owner.op == tensor.mul:
# Try to find the AdvancedSubtensor node mentionned above,
# and the output gradient
for i, input in enumerate(denom.owner.inputs):
if input.owner and isinstance(input.owner.op,
subtensor.AdvancedSubtensor):
other_inputs = [in_ for (j,
in_) in enumerate(denom.owner.inputs) if j != i]
other_inputs = [in_ for (j, in_) in
enumerate(denom.owner.inputs) if j != i]
if len(other_inputs) == 1:
rest = other_inputs[0]
else:
......@@ -1894,16 +1898,14 @@ def categorical_crossentropy(coding_dist, true_dist):
"""
if true_dist.ndim == coding_dist.ndim:
return -tensor.sum(true_dist * tensor.log(coding_dist), axis=coding_dist.ndim-1)
return -tensor.sum(true_dist * tensor.log(coding_dist),
axis=coding_dist.ndim - 1)
elif true_dist.ndim == coding_dist.ndim - 1:
return crossentropy_categorical_1hot(coding_dist, true_dist)
else:
raise TypeError('rank mismatch between coding and true distributions')
from theano import scalar
class Prepend_scalar_constant_to_each_row(gof.Op):
__props__ = ()
......@@ -2026,7 +2028,7 @@ local_log_softmax = gof.PatternSub(in_pattern=(tensor.log, (softmax_op, 'x')),
# don't do register_stabilize, this is to make local_log_softmax run
# only after another more specific optimization that stabilizes cross entropy
#opt.register_stabilize(local_log_softmax, name = 'local_log_softmax')
# opt.register_stabilize(local_log_softmax, name = 'local_log_softmax')
opt.register_specialize(local_log_softmax, 'fast_compile_gpu', name='local_log_softmax')
......
theano/tests/test_flake8.py
浏览文件 @ 70b5f2c1
......@@ -88,7 +88,6 @@ whitelist_flake8 = [
"tensor/signal/conv.py",
"tensor/signal/tests/test_conv.py",
"tensor/signal/tests/test_downsample.py",
"tensor/nnet/nnet.py",
"tensor/nnet/Conv3D.py",
"tensor/nnet/__init__.py",
"tensor/nnet/ConvTransp3D.py",
......
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