def check_backward(self, x_data, y_grad):
def f(x):
return functions.unpooling_2d(x, self.ksize, outsize=self.outsize,
cover_all=self.cover_all)
gradient_check.check_backward(
f, x_data, y_grad, dtype=numpy.float64,
**self.check_backward_options)
def check_backward(self, x_data, W_data, y_grad):
def f(x, W):
return chainer.functions.embed_id(x, W, self.ignore_label)
gradient_check.check_backward(
f, (x_data, W_data), y_grad, dtype=numpy.float64,
**self.check_backward_options)
def check_backward(self, x_data, W_data, b_data, y_grad,
use_cudnn='never'):
if not self.c_contiguous:
xp = backend.get_array_module(x_data)
x_data = xp.asfortranarray(x_data)
W_data = xp.asfortranarray(W_data)
y_grad = xp.asfortranarray(y_grad)
self.assertFalse(x_data.flags.c_contiguous)
self.assertFalse(W_data.flags.c_contiguous)
self.assertFalse(y_grad.flags.c_contiguous)
if b_data is not None:
b = xp.empty((len(b_data) * 2,), dtype=self.b.dtype)
b[::2] = b_data
b_data = b[::2]
self.assertFalse(b_data.flags.c_contiguous)
args = (x_data, W_data)
if b_data is not None:
args += (b_data,)
def f(*args):
return F.deconvolution_nd(*args, stride=self.stride, pad=self.pad,
outsize=self.outsize, dilate=self.dilate,
groups=self.groups)
with chainer.using_config('use_cudnn', use_cudnn):
with chainer.using_config('autotune', self.autotune):
gradient_check.check_backward(
f, args, y_grad, **self.check_backward_options)
def check_backward(self, link, x_data, y_grad):
params = [link.W]
if not self.nobias:
params.append(link.b)
gradient_check.check_backward(
link, x_data, y_grad, params, **self.check_backward_options)
def check_backward(self, x_data, roi_data, y_grad):
gradient_check.check_backward(
functions.ROIPooling2D(outh=self.outh,
outw=self.outw,
spatial_scale=self.spatial_scale),
(x_data, roi_data), y_grad, no_grads=[False, True],
**self.check_backward_options)
def check_backward(self, c_prev1_data, c_prev2_data, x1_data, x2_data,
c_grad, h_grad):
gradient_check.check_backward(
functions.SLSTM(),
(c_prev1_data, c_prev2_data, x1_data, x2_data),
(c_grad, h_grad), dtype=numpy.float64,
**self.check_backward_options)
def check_backward(self, x_data, t_data, w_data, y_grad, sampler):
def f(x, w):
return functions.negative_sampling(
x, t_data, w, sampler, self.sample_size, reduce=self.reduce)
gradient_check.check_backward(
f, (x_data, w_data), y_grad, **self.check_backward_options)
def check_backward(self, x_data, y_data, gy_data, ggx_data):
def f(y, gy):
return tanh.TanhGrad(x_data).apply((y, gy))[0]
gradient_check.check_backward(
f, (y_data, gy_data), ggx_data, dtype=numpy.float64,
**self.check_backward_options)
def check_backward(self, x_data, g_data):
def func(x):
return functions.cast(x, self.out_type)
gradient_check.check_backward(
func, x_data, g_data, dtype='d',
eps=2.0 ** -2, atol=1e-2, rtol=1e-3)
def check_zero_det(self, x, gy, err):
if self.batched:
x[0, ...] = 0.0
else:
x[...] = 0.0
with self.assertRaises(err):
gradient_check.check_backward(self.det, x, gy)
def check_backward(
self, h_data, xs_data, gh_data, gys_data):
def fun(*args):
if self.hidden_none:
h = None
xs = args
else:
h, = args[:1]
xs = args[1:]
hy, ys = self.rnn(h, xs)
return tuple([hy, ] + list(ys))
params = []
for layer in self.rnn:
for p in layer.params():
params.append(p)
if self.hidden_none:
in_data = xs_data
else:
in_data = [h_data, ] + xs_data
gradient_check.check_backward(
fun, tuple(in_data),
tuple([gh_data, ] + gys_data),
tuple(params), eps=1e-2, rtol=1e-3, atol=1e-3)
def check_backward(self, rx_data, ix_data, rg_data, ig_data):
def f(real, imag):
return getattr(chainer.functions, self.method)((real, imag))
gradient_check.check_backward(
f, (rx_data, ix_data),
(rg_data, ig_data), dtype='d', eps=2.0 ** -2, atol=1e-2, rtol=1e-3)
def check_backward(self, x0_data, x1_data, t_data, gy_data):
def f(x0, x1, t):
return functions.contrastive(x0, x1, t, self.margin, self.reduce)
gradient_check.check_backward(
f, (x0_data, x1_data, t_data), gy_data, dtype='d',
**self.check_backward_options)
def check_backward(self, inputs, grad_outputs, backend_config):
xp = backend_config.xp
if backend_config.use_cuda:
inputs = cuda.to_gpu(inputs)
grad_outputs = cuda.to_gpu(grad_outputs)
x_data, W_data, b_data = inputs
y_grad, = grad_outputs
if not self.c_contiguous:
x_data = xp.asfortranarray(x_data)
W_data = xp.asfortranarray(W_data)
y_grad = xp.asfortranarray(y_grad)
assert not x_data.flags.c_contiguous
assert not W_data.flags.c_contiguous
assert not y_grad.flags.c_contiguous
if b_data is not None:
b = xp.empty((len(b_data) * 2,), dtype=b_data.dtype)
b[::2] = b_data
b_data = b[::2]
assert not b_data.flags.c_contiguous
args = (x_data, W_data)
if b_data is not None:
args = args + (b_data,)
def f(*args):
return F.deconvolution_2d(
*args, stride=self.stride, pad=self.pad, outsize=self.outsize,
dilate=self.dilate, groups=self.groups)
with backend_config:
gradient_check.check_backward(
f, args, y_grad, **self.check_backward_options)
def check_backward(self, x_data, W_data, b_data, y_grad):
xp = backend.get_array_module(x_data)
if not self.c_contiguous:
x_data = xp.asfortranarray(x_data)
W_data = xp.asfortranarray(W_data)
y_grad = xp.asfortranarray(y_grad)
self.assertFalse(x_data.flags.c_contiguous)
self.assertFalse(W_data.flags.c_contiguous)
self.assertFalse(y_grad.flags.c_contiguous)
if b_data is not None:
b = xp.empty((len(b_data) * 2,), dtype=self.b.dtype)
b[::2] = b_data
b_data = b[::2]
self.assertFalse(b_data.flags.c_contiguous)
args = (x_data, W_data)
if b_data is not None:
args = args + (b_data,)
def f(*args):
return F.dilated_convolution_2d(*args, stride=self.stride,
pad=self.pad, dilate=self.dilate,
cover_all=self.cover_all)
with chainer.using_config('use_cudnn', self.use_cudnn):
gradient_check.check_backward(
f, args, y_grad, dtype=numpy.float64,
**self.check_backward_options)
def test_backward_gpu(model, data):
atom_data, adj_data, y_grad = [cuda.to_gpu(d) for d in data]
model.to_gpu()
params = tuple(model.params())
gradient_check.check_backward(model, (atom_data, adj_data), y_grad,
params=params, no_grads=[True, True],
atol=1e3, rtol=1e3)
def check_backward(self, x_data, t_data, w_data, samples_data, gy_data):
def _black_out(x, t, W, samples):
return functions.black_out(x, t, W, samples, self.reduce)
gradient_check.check_backward(
_black_out, (x_data, t_data, w_data, samples_data),
gy_data, atol=1.e-3)
def check_backward(self, x_data, t_data, class_weight, use_cudnn=True):
func = functions.SoftmaxCrossEntropy(
use_cudnn=use_cudnn, cache_score=self.cache_score,
class_weight=class_weight)
gradient_check.check_backward(
func, (x_data, t_data), None, eps=0.02,
**self.check_backward_options)
def check_backward(self, x_data, W_data, b_data, y_grad):
xp = cuda.get_array_module(x_data)
if not self.c_contiguous:
x_data = xp.asfortranarray(x_data)
W_data = xp.asfortranarray(W_data)
y_grad = xp.asfortranarray(y_grad)
self.assertFalse(x_data.flags.c_contiguous)
self.assertFalse(W_data.flags.c_contiguous)
self.assertFalse(y_grad.flags.c_contiguous)
if b_data is not None:
b = xp.empty((len(b_data) * 2,), dtype=self.b.dtype)
b[::2] = b_data
b_data = b[::2]
self.assertFalse(b_data.flags.c_contiguous)
args = (x_data, W_data)
if b_data is not None:
args = args + (b_data,)
with chainer.using_config('use_cudnn', self.use_cudnn):
with chainer.using_config('cudnn_deterministic',
self.cudnn_deterministic):
gradient_check.check_backward(
convolution_2d.Convolution2DFunction(
self.stride, self.pad, self.cover_all),
args, y_grad, **self.check_backward_options)
def test_backward_cpu(readouts, data):
atom_data, y_grad = data
for readout in readouts:
if readout.mode == 'summax':
y_grad = functions.concat((y_grad, y_grad), axis=1).data
gradient_check.check_backward(
readout, atom_data, y_grad, atol=1e-2, rtol=1e-2)
def check_backward(self, x_data, t_data, class_weight, use_cudnn='always'):
with chainer.using_config('use_cudnn', use_cudnn):
func = functions.SoftmaxCrossEntropy(
cache_score=self.cache_score, class_weight=class_weight)
gradient_check.check_backward(
func, (x_data, t_data), None,
**self.check_backward_options)
def test_backward_cpu(model, model_processed, data):
atom_data_processed, atom_data, adj_data, y_grad = data
gradient_check.check_backward(model, (atom_data, adj_data), y_grad,
atol=1e-1, rtol=1e-1)
gradient_check.check_backward(model_processed, (atom_data_processed,
adj_data), y_grad,
atol=1e-1, rtol=1e-1)
def check_backward(self, x_data, y_grad):
if self.dtype == numpy.float16:
tol = 0.1
else:
tol = 1e-4
gradient_check.check_backward(
lambda x: x ** 2, x_data, y_grad, atol=tol, rtol=tol)
def check_backward(self, x_data, W_data, b_data, y_grad,
use_cudnn='never'):
xp = cuda.get_array_module(x_data)
if not self.c_contiguous:
x_data = xp.asfortranarray(x_data)
W_data = xp.asfortranarray(W_data)
y_grad = xp.asfortranarray(y_grad)
self.assertTrue(x_data.flags.f_contiguous)
self.assertTrue(W_data.flags.f_contiguous)
self.assertTrue(y_grad.flags.f_contiguous)
if b_data is not None:
b = xp.empty((len(b_data) * 2,), dtype=b_data.dtype)
b[::2] = b_data
b_data = b[::2]
self.assertFalse(b_data.flags.c_contiguous)
args = (x_data, W_data)
if b_data is not None:
args = args + (b_data,)
ndim = len(self.dims)
with chainer.using_config('use_cudnn', use_cudnn):
gradient_check.check_backward(
convolution_nd.ConvolutionND(
ndim, self.stride, self.pad, self.cover_all),
args, y_grad, **self.check_backward_options)
def check_backward(self, x_data, y_grad):
def f(x):
return functions.elu(x, alpha=self.alpha)
gradient_check.check_backward(
f, x_data, y_grad, dtype=numpy.float64,
**self.check_backward_options)
def check_backward(self, x_data, W_data, b_data, y_grad):
xp = cuda.get_array_module(x_data)
if not self.c_contiguous:
x_data = xp.asfortranarray(x_data)
W_data = xp.asfortranarray(W_data)
y_grad = xp.asfortranarray(y_grad)
self.assertFalse(x_data.flags.c_contiguous)
self.assertFalse(W_data.flags.c_contiguous)
self.assertFalse(y_grad.flags.c_contiguous)
if b_data is not None:
b = xp.empty((len(b_data) * 2,), dtype=self.b.dtype)
b[::2] = b_data
b_data = b[::2]
self.assertFalse(b_data.flags.c_contiguous)
args = (x_data, W_data)
if b_data is not None:
args = args + (b_data,)
def f(*args):
return F.convolution_2d(*args, stride=self.stride, pad=self.pad,
cover_all=self.cover_all,
dilate=self.dilate)
with chainer.using_config('use_cudnn', self.use_cudnn):
with chainer.using_config('cudnn_deterministic',
self.cudnn_deterministic):
with chainer.using_config('autotune', self.autotune):
gradient_check.check_backward(
f, args, y_grad, dtype='d', atol=5e-4, rtol=5e-3)
def check_backward(self, x_data, y_grad):
def f(x):
x_min, x_max = self.x_min_max
return functions.clip(x, x_min, x_max)
gradient_check.check_backward(
f, x_data, y_grad, dtype=numpy.float64)
def check_backward_ignore_nan_with_nonnan_value(inputs):
x0_data, x1_data, _ = inputs
def func(x0, x1):
return chainer_chemistry.functions.mean_squared_error(x0, x1,
ignore_nan=True)
gradient_check.check_backward(func, (x0_data, x1_data), None, eps=1e-2)
def check_backward(self, x_data, y_grad):
params = [self.link.W]
if not self.nobias:
params.append(self.link.b)
gradient_check.check_backward(
self.link, x_data, y_grad, params, eps=1e-2)
def check_backward(self, op, x_data, y_grad):
options = {}
if self.dtype == numpy.float16:
options = {'atol': 5e-3, 'rtol': 5e-2}
gradient_check.check_backward(lambda x: op(x, self.value),
x_data, y_grad,
dtype=numpy.float64, **options)
def check_backward(self, x_data, y_grad):
def f(x):
return functions.get_item(x, self.slices)
gradient_check.check_backward(f, (x_data, ), y_grad, dtype='d')
def check_backward(self, x_data, y_grad):
params = [self.link.linear.W]
if self.initial_bias is not None:
params.append(self.link.linear.b)
gradient_check.check_backward(
self.link, x_data, y_grad, params, atol=1e-2)
def check_backward(self, x_data, y_grad, use_cudnn='always'):
with chainer.using_config('use_cudnn', use_cudnn):
gradient_check.check_backward(
functions.AveragePooling2D(3, 2, 1, False), x_data, y_grad,
**self.check_backward_options)
def check_backward(self, x_data, y_grad):
gradient_check.check_backward(self.link, x_data, y_grad,
(self.link.W, self.link.b),
**self.check_backward_options)
def check_backward(self, x_data, y_grad, **kwargs):
gradient_check.check_backward(functions.batch_inv, x_data, y_grad,
**self.check_backward_options)
def check_backward(self, x_data, g_data):
gradient_check.check_backward(
lambda x: functions.squeeze(x, self.axis),
x_data, g_data, **self.check_backward_options)
def check_backward(self, x_data, y_grad):
gradient_check.check_backward(F.log1p, x_data, y_grad)
def check_backward(self, x_data, y_grad):
gradient_check.check_backward(self.link,
x_data,
y_grad, (self.link.W, self.link.b),
eps=1e-2)
def check_backward(self, x_data, y_grad, use_cudnn=True):
gradient_check.check_backward(functions.AveragePooling2D(
3, 2, 1, False, use_cudnn),
x_data,
y_grad,
eps=1e-2)
def check_backward(self, x1_data, x2_data, axis, y_grad):
x = (x1_data, x2_data)
gradient_check.check_backward(lambda x, y: functions.scale(x, y, axis),
x, y_grad, **self.check_backward_options)
def check_backward(self, args, y_grad, use_cudnn='always'):
with chainer.using_config('use_cudnn', use_cudnn), \
chainer.using_config('train', self.train):
gradient_check.check_backward(self.batch_normalization, args,
y_grad,
**self.check_backward_options)
def check_backward(self, args, y_grad):
with chainer.using_config('train', self.train):
gradient_check.check_backward(
batch_normalization.BatchNormalizationFunction(
mean=None, var=None, decay=self.decay, eps=self.eps), args,
y_grad, **self.check_backward_options)
def check_backward(self, inputs_data, output_grad, atol, rtol):
gradient_check.check_backward(
self.op, inputs_data, output_grad, atol=atol, rtol=rtol,
dtype=numpy.float64)
def check_backward(self, x_data, gy_data, use_cudnn='always'):
with chainer.using_config('use_cudnn', use_cudnn):
gradient_check.check_backward(functions.LogSoftmax(), x_data,
gy_data,
**self.check_backward_options)
def check_backward(self, x_data, y_grad):
def f(x):
return functions.space2depth(x, self.r)
gradient_check.check_backward(f, x_data, y_grad, dtype=numpy.float64,
**self.check_backward_options)
def check_backward(self, x_data, axis, y_grad):
gradient_check.check_backward(
lambda x: functions.cumsum(x, axis), x_data, y_grad,
dtype=numpy.float64, **self.check_backward_options)
def check_backward(self, x_data, t_data, class_weight, use_cudnn='always'):
with chainer.using_config('use_cudnn', use_cudnn):
func = functions.SoftmaxCrossEntropy(cache_score=self.cache_score,
class_weight=class_weight)
gradient_check.check_backward(func, (x_data, t_data), None,
**self.check_backward_options)
# -*- coding: utf-8 -*-
import sys, os
from chainer import cuda, gradient_check, Variable
sys.path.append(os.path.split(os.getcwd())[0])
import model
xp = cuda.cupy
context = xp.random.uniform(-1, 1, (2, 3)).astype(xp.float32)
weight = xp.random.uniform(-1, 1, (2, 1)).astype(xp.float32)
z = xp.full((2, 1), 10.0).astype(xp.float32)
y = model.apply_attention(Variable(context), Variable(weight) / Variable(z))
print y.data
y_grad = xp.random.uniform(-1.0, 1.0, (2, 3)).astype(xp.float32)
gradient_check.check_backward(model.Attention(), (context, weight / z),
y_grad,
eps=1e-2)
def check_backward(self, x_data, axis, y_grad):
gradient_check.check_backward(
lambda x: functions.cumprod(x, axis), x_data, y_grad,
atol=1e-3, dtype=numpy.float64)
def check_backward(self, x_data, t_data, g_data, class_weight):
func = functions.SoftmaxCrossEntropy(cache_score=self.cache_score,
class_weight=class_weight,
reduce='no')
gradient_check.check_backward(func, (x_data, t_data), g_data,
**self.check_backward_options)
def check_backward(self, x_data, gy_data, use_cudnn=True):
gradient_check.check_backward(functions.Softmax(use_cudnn), x_data,
gy_data, **self.check_backward_options)
def check_backward(self, h_data, gloss_data):
gradient_check.check_backward(functions.DeCov(self.reduce), (h_data, ),
gloss_data,
eps=0.02,
atol=1e-3)
def check_backward(self, x_data, grad):
gradient_check.check_backward(functions.HardSigmoid(), x_data, grad,
**self.check_backward_options)
def check_backward(self, x_data, roi_data, y_grad):
gradient_check.check_backward(functions.ROIPooling2D(
outh=self.outh, outw=self.outw, spatial_scale=self.spatial_scale),
(x_data, roi_data),
y_grad,
no_grads=[False, True])
def check_backward(self, x_data, y_grad):
gradient_check.check_backward(chainerrl.functions.SumArrays(),
x_data,
y_grad,
eps=1e-2,
rtol=1e-2)
def check_backward(self, x1_data, x2_data, axis, y_grad):
x = (x1_data, x2_data)
gradient_check.check_backward(lambda x, y: functions.bias(x, y, axis),
x, y_grad)
def check_backward(self, x_data, t_data, y_grad):
def f(x, t):
return functions.huber_loss(x, t, delta=1, reduce=self.reduce)
gradient_check.check_backward(
f, (x_data, t_data), y_grad, **self.backward_options)
def check_backward(self, x_data, y_grad, axis=None):
gradient_check.check_backward(lambda x: functions.logsumexp(x, axis),
x_data, y_grad,
**self.check_backward_option)
def test_backward_gpu(model, data):
# type: (GIN, Tuple[numpy.ndarray, numpy.ndarray, numpy.ndarray]) -> None
atom_data, adj_data, y_grad = map(cuda.to_gpu, data)
model.to_gpu()
gradient_check.check_backward(
model, (atom_data, adj_data), y_grad, atol=1e-5, rtol=1e-5)
def check_backward(self, x_data, g_data):
gradient_check.check_backward(
functions.Tile(self.reps), x_data, g_data,
**self.check_backward_options)