提交 155c4e01 authored 作者: Chiheb Trabelsi's avatar Chiheb Trabelsi

test_bench_loopfusion.py has been modified in order to respect the flake8 style.

上级 a2503274
...@@ -10,7 +10,7 @@ from __future__ import absolute_import, print_function, division ...@@ -10,7 +10,7 @@ from __future__ import absolute_import, print_function, division
# so state is ignored # so state is ignored
# since this job is not restartable, channel is also ignored # since this job is not restartable, channel is also ignored
import logging, time, sys import logging
import numpy import numpy
from six.moves import xrange from six.moves import xrange
...@@ -18,17 +18,22 @@ from six.moves import xrange ...@@ -18,17 +18,22 @@ from six.moves import xrange
import theano import theano
from theano.compile import shared, pfunc from theano.compile import shared, pfunc
from theano import tensor from theano import tensor
from theano.tensor.nnet import softplus
from theano.tensor.nnet.nnet import softsign from theano.tensor.nnet.nnet import softsign
try:
from PIL import Image
except ImportError:
Image = None
# from PIL import Image
_logger = logging.getLogger('theano.sandbox.cuda.tests.test_bench_loopfusion') _logger = logging.getLogger('theano.sandbox.cuda.tests.test_bench_loopfusion')
def _shared_uniform(rng, low, high, size, dtype, name=None): def _shared_uniform(rng, low, high, size, dtype, name=None):
return shared( return shared(
theano._asarray( theano._asarray(
rng.uniform(low=low, high=high, size=size), rng.uniform(low=low, high=high, size=size),
dtype=dtype), name) dtype=dtype),
name)
class Kouh2008(object): class Kouh2008(object):
...@@ -49,8 +54,10 @@ class Kouh2008(object): ...@@ -49,8 +54,10 @@ class Kouh2008(object):
""" """
if len(w_list) != len(x_list): if len(w_list) != len(x_list):
raise ValueError('w_list must have same len as x_list') raise ValueError('w_list must have same len as x_list')
output = (sum(w * tensor.pow(x, p) for (w, x) in zip(w_list, x_list)))\ output = ((sum(w * tensor.pow(x, p)
/ (theano._asarray(eps, dtype=k.type.dtype) + k + tensor.pow(sum(tensor.pow(x, q) for x in x_list), r)) for (w, x) in zip(w_list, x_list))) /
(theano._asarray(eps, dtype=k.type.dtype) + k +
tensor.pow(sum(tensor.pow(x, q) for x in x_list), r)))
assert output.type.ndim == 2 assert output.type.ndim == 2
self.__dict__.update(locals()) self.__dict__.update(locals())
...@@ -80,10 +87,15 @@ class Kouh2008(object): ...@@ -80,10 +87,15 @@ class Kouh2008(object):
w_sm = theano.tensor.nnet.softmax(w) w_sm = theano.tensor.nnet.softmax(w)
w_list = [w_sm[:, i] for i in xrange(n_terms)] w_list = [w_sm[:, i] for i in xrange(n_terms)]
w_l1 = abs(w).sum() w_l1 = abs(w).sum()
w_l2_sqr = (w**2).sum() w_l2_sqr = (w ** 2).sum()
else: else:
w_list = [shared_uniform(low=-2.0/n_terms, high=2.0/n_terms, size=(n_out,), name='w_%i'%i) w_list = [
for i in xrange(n_terms)] shared_uniform(
low=-2.0 / n_terms,
high=2.0 / n_terms,
size=(n_out,),
name='w_%i' % i)
for i in xrange(n_terms)]
w_l1 = sum(abs(wi).sum() for wi in w_list) w_l1 = sum(abs(wi).sum() for wi in w_list)
w_l2_sqr = sum((wi**2).sum() for wi in w_list) w_l2_sqr = sum((wi**2).sum() for wi in w_list)
...@@ -102,19 +114,27 @@ class Kouh2008(object): ...@@ -102,19 +114,27 @@ class Kouh2008(object):
p = tensor.nnet.sigmoid(p_unbounded) * e_range_mag + e_range_low p = tensor.nnet.sigmoid(p_unbounded) * e_range_mag + e_range_low
q = tensor.nnet.sigmoid(q_unbounded) * e_range_mag + e_range_low q = tensor.nnet.sigmoid(q_unbounded) * e_range_mag + e_range_low
r = tensor.nnet.sigmoid(r_unbounded) * \ r = tensor.nnet.sigmoid(r_unbounded) * \
theano._asarray(1.0/e_range_low - 1.0/e_range_high, dtype=dtype) \ theano._asarray(1.0 / e_range_low - 1.0 / e_range_high,
+ theano._asarray(1.0/e_range_high, dtype=dtype) dtype=dtype) + \
theano._asarray(1.0 / e_range_high, dtype=dtype)
k = softsign(k_unbounded) k = softsign(k_unbounded)
if use_softmax_w: if use_softmax_w:
rval = cls(w_list, x_list, p, q, r, k, rval = cls(w_list, x_list, p, q, r, k,
params=[p_unbounded, q_unbounded, r_unbounded, k_unbounded, w] + params, params=[p_unbounded,
updates=updates) q_unbounded,
r_unbounded,
k_unbounded,
w] + params,
updates=updates)
else: else:
rval = cls(w_list, x_list, p, q, r, k, rval = cls(w_list, x_list, p, q, r, k,
params=[p_unbounded, q_unbounded, r_unbounded, k_unbounded] + w_list + params, params=[p_unbounded,
updates=updates) q_unbounded,
r_unbounded,
k_unbounded] + w_list + params,
updates=updates)
rval.p_unbounded = p_unbounded rval.p_unbounded = p_unbounded
rval.q_unbounded = q_unbounded rval.q_unbounded = q_unbounded
rval.r_unbounded = r_unbounded rval.r_unbounded = r_unbounded
...@@ -126,8 +146,10 @@ class Kouh2008(object): ...@@ -126,8 +146,10 @@ class Kouh2008(object):
return rval return rval
@classmethod @classmethod
def new_filters_expbounds(cls, rng, input, n_in, n_out, n_terms, dtype=None, eps=1e-1, def new_filters_expbounds(cls, rng, input, n_in, n_out, n_terms,
exponent_range=(1.0, 3.0), filter_range=1.0): dtype=None, eps=1e-1,
exponent_range=(1.0, 3.0),
filter_range=1.0):
"""Return a KouhLayer instance with random parameters """Return a KouhLayer instance with random parameters
The parameters are drawn on a range [typically] suitable for fine-tuning by gradient The parameters are drawn on a range [typically] suitable for fine-tuning by gradient
...@@ -161,19 +183,30 @@ class Kouh2008(object): ...@@ -161,19 +183,30 @@ class Kouh2008(object):
def shared_uniform(low, high, size, name): def shared_uniform(low, high, size, name):
return _shared_uniform(rng, low, high, size, dtype, name) return _shared_uniform(rng, low, high, size, dtype, name)
f_list = [shared_uniform(low=-2.0/numpy.sqrt(n_in), high=2.0/numpy.sqrt(n_in), size=(n_in, n_out), name='f_%i'%i) f_list = [shared_uniform(low=-2.0 / numpy.sqrt(n_in),
for i in xrange(n_terms)] high=2.0 / numpy.sqrt(n_in),
size=(n_in, n_out),
b_list = [shared_uniform(low=0, high=.01, size=(n_out,), name='b_%i'%i) name='f_%i' % i)
for i in xrange(n_terms)] for i in xrange(n_terms)]
#x_list = [theano._asarray(eps, dtype=dtype)+softplus(tensor.dot(input, f_list[i])) for i in xrange(n_terms)]
b_list = [shared_uniform(low=0,
high=.01,
size=(n_out,),
name='b_%i' % i)
for i in xrange(n_terms)]
# x_list = [theano._asarray(eps, dtype=dtype) + softplus(tensor.dot(input, f_list[i])) for i in xrange(n_terms)]
filter_range = theano._asarray(filter_range, dtype=dtype) filter_range = theano._asarray(filter_range, dtype=dtype)
half_filter_range = theano._asarray(filter_range/2, dtype=dtype) half_filter_range = theano._asarray(filter_range / 2,
x_list = [theano._asarray(filter_range + eps, dtype=dtype)+half_filter_range * softsign(tensor.dot(input, f_list[i]) + dtype=dtype)
b_list[i]) for i in xrange(n_terms)] x_list = [
theano._asarray(filter_range + eps, dtype=dtype) +
rval = cls.new_expbounds(rng, x_list, n_out, dtype=dtype, params=f_list + b_list, half_filter_range * softsign(
exponent_range=exponent_range) tensor.dot(input, f_list[i]) + b_list[i])
for i in xrange(n_terms)]
rval = cls.new_expbounds(
rng, x_list, n_out, dtype=dtype, params=f_list + b_list,
exponent_range=exponent_range)
rval.f_list = f_list rval.f_list = f_list
rval.input = input # add the input to the returned object rval.input = input # add the input to the returned object
rval.filter_l1 = sum(abs(fi).sum() for fi in f_list) rval.filter_l1 = sum(abs(fi).sum() for fi in f_list)
...@@ -183,6 +216,8 @@ class Kouh2008(object): ...@@ -183,6 +216,8 @@ class Kouh2008(object):
def img_from_weights(self, rows=None, cols=None, row_gap=1, col_gap=1, eps=1e-4): def img_from_weights(self, rows=None, cols=None, row_gap=1, col_gap=1, eps=1e-4):
""" Return an image that visualizes all the weights in the layer. """ Return an image that visualizes all the weights in the layer.
""" """
if Image is None:
raise ImportError("No module named PIL")
n_in, n_out = self.f_list[0].value.shape n_in, n_out = self.f_list[0].value.shape
...@@ -190,10 +225,12 @@ class Kouh2008(object): ...@@ -190,10 +225,12 @@ class Kouh2008(object):
rows = int(numpy.sqrt(n_out)) rows = int(numpy.sqrt(n_out))
if cols is None: if cols is None:
cols = n_out // rows cols = n_out // rows
if n_out % rows: cols += 1 if n_out % rows:
cols += 1
if rows is None: if rows is None:
rows = n_out // cols rows = n_out // cols
if n_out % cols: rows += 1 if n_out % cols:
rows += 1
filter_shape = self.filter_shape filter_shape = self.filter_shape
height = rows * (row_gap + filter_shape[0]) - row_gap height = rows * (row_gap + filter_shape[0]) - row_gap
...@@ -203,34 +240,40 @@ class Kouh2008(object): ...@@ -203,34 +240,40 @@ class Kouh2008(object):
w = self.w.value w = self.w.value
w_col = 0 w_col = 0
def pixel_range(x): def pixel_range(x):
return 255 * (x - x.min()) / (x.max() - x.min() + eps) return 255 * (x - x.min()) / (x.max() - x.min() + eps)
for r in xrange(rows): for r in xrange(rows):
out_r_low = r*(row_gap + filter_shape[0]) out_r_low = r * (row_gap + filter_shape[0])
out_r_high = out_r_low + filter_shape[0] out_r_high = out_r_low + filter_shape[0]
for c in xrange(cols): for c in xrange(cols):
out_c_low = c*(col_gap + filter_shape[1]) out_c_low = c * (col_gap + filter_shape[1])
out_c_high = out_c_low + filter_shape[1] out_c_high = out_c_low + filter_shape[1]
out_tile = out_array[out_r_low:out_r_high, out_c_low:out_c_high, :] out_tile = out_array[out_r_low:out_r_high,
out_c_low:out_c_high,
:]
if c % 3 == 0: # linear filter if c % 3 == 0: # linear filter
if w_col < w.shape[1]: if w_col < w.shape[1]:
out_tile[...] = pixel_range(w[:, w_col]).reshape(filter_shape+(1,)) out_tile[...] = pixel_range(
w[:, w_col]).reshape(filter_shape + (1,))
w_col += 1 w_col += 1
if c % 3 == 1: # E filters if c % 3 == 1: # E filters
if w_col < w.shape[1]: if w_col < w.shape[1]:
# filters after the 3rd do not get rendered, but are skipped over. # filters after the 3rd do not get rendered, but are skipped over.
# there are only 3 colour channels. # there are only 3 colour channels.
for i in xrange(min(self.n_E_quadratic, 3)): for i in xrange(min(self.n_E_quadratic, 3)):
out_tile[:, :, i] = pixel_range(w[:, w_col+i]).reshape(filter_shape) out_tile[:, :, i] = pixel_range(
w[:, w_col + i]).reshape(filter_shape)
w_col += self.n_E_quadratic w_col += self.n_E_quadratic
if c % 3 == 2: # S filters if c % 3 == 2: # S filters
if w_col < w.shape[1]: if w_col < w.shape[1]:
# filters after the 3rd do not get rendered, but are skipped over. # filters after the 3rd do not get rendered, but are skipped over.
# there are only 3 colour channels. # there are only 3 colour channels.
for i in xrange(min(self.n_S_quadratic, 3)): for i in xrange(min(self.n_S_quadratic, 3)):
out_tile[:, :, 2-i] = pixel_range(w[:, w_col+i]).reshape(filter_shape) out_tile[:, :, 2 - i] = pixel_range(
w[:, w_col + i]).reshape(filter_shape)
w_col += self.n_S_quadratic w_col += self.n_S_quadratic
return Image.fromarray(out_array, 'RGB') return Image.fromarray(out_array, 'RGB')
...@@ -264,8 +307,9 @@ class Config(object): ...@@ -264,8 +307,9 @@ class Config(object):
ft_batchsize = 30 ft_batchsize = 30
ft_epoch_len = 50000 ft_epoch_len = 50000
ft_status_interval = 50 # property( lambda s:s.ft_epoch_len/s.ft_batchsize) ft_status_interval = 50 # property(lambda s:s.ft_epoch_len/s.ft_batchsize)
ft_validation_interval = property( lambda s: s.ft_epoch_len/s.ft_batchsize) ft_validation_interval = property(
lambda s: s.ft_epoch_len / s.ft_batchsize)
ft_ntrain_limit = 0 ft_ntrain_limit = 0
ft_test_lag1 = True ft_test_lag1 = True
...@@ -290,14 +334,15 @@ if 0: ...@@ -290,14 +334,15 @@ if 0:
debug = False debug = False
if isinstance(theano.compile.mode.get_default_mode(), if isinstance(theano.compile.mode.get_default_mode(),
theano.compile.debugmode.DebugMode): theano.compile.debugmode.DebugMode):
debug = True debug = True
# get symbolic train set # get symbolic train set
s_lr = theano.tensor.fscalar() s_lr = theano.tensor.fscalar()
if not debug: if not debug:
sshape = (None, 784) sshape = (None, 784)
else: sshape = (None, 3) else:
sshape = (None, 3)
x = theano.tensor.TensorType(dtype=conf.dtype, broadcastable=(0, 0), shape=sshape)() x = theano.tensor.TensorType(dtype=conf.dtype, broadcastable=(0, 0), shape=sshape)()
y = theano.tensor.lvector() y = theano.tensor.lvector()
...@@ -315,7 +360,8 @@ if 0: ...@@ -315,7 +360,8 @@ if 0:
print(layer.params) print(layer.params)
gparams = theano.tensor.grad(cost, layer.params) gparams = theano.tensor.grad(cost, layer.params)
updates = [(p, p - s_lr*gp) for p, gp in zip(layer.params, gparams)] updates = [
(p, p - s_lr * gp) for p, gp in zip(layer.params, gparams)]
train_nll = pfunc([x, y, s_lr], [], updates=updates) train_nll = pfunc([x, y, s_lr], [], updates=updates)
......
Markdown 格式
0%
您添加了 0 到此讨论。请谨慎行事。
请先完成此评论的编辑!
注册 或者 后发表评论