Skip to content
This repository has been archived by the owner on Jul 1, 2024. It is now read-only.

Commit

Permalink
Immigrate reference operations to a separate module (keras-team#9948)
Browse files Browse the repository at this point in the history
  • Loading branch information
@taehoonlee @fchollet
taehoonlee authored and fchollet committed Apr 17, 2018
1 parent a8ca67f commit d673afd
Show file tree
Hide file tree
Showing 2 changed files with 173 additions and 164 deletions.
173 changes: 9 additions & 164 deletions tests/keras/backend/backend_test.py
View file
Original file line number Diff line number Diff line change
@@ -1,14 +1,15 @@
import pytest
from numpy.testing import assert_allclose
import numpy as np
import scipy.signal as signal
import scipy.sparse as sparse
import warnings
from keras.utils.test_utils import keras_test

from keras import backend as K
from keras.backend import floatx, set_floatx, variable
from keras.utils.conv_utils import convert_kernel
import reference_operations


BACKENDS = [] # Holds a list of all available back-ends

Expand Down Expand Up @@ -192,162 +193,6 @@ def check_composed_tensor_operations(first_function_name, first_function_args,
assert_list_pairwise(z_list)


def normalize_ref_conv(func):
def wrapper(*args):
x = args[0]
w = args[1]
if x.ndim == 3:
w = np.flipud(w)
w = np.transpose(w, (1, 2, 0))
if args[3] == 'channels_last':
x = np.transpose(x, (0, 2, 1))
elif x.ndim == 4:
w = np.fliplr(np.flipud(w))
w = np.transpose(w, (2, 3, 0, 1))
if args[3] == 'channels_last':
x = np.transpose(x, (0, 3, 1, 2))
else:
w = np.flip(np.fliplr(np.flipud(w)), axis=2)
w = np.transpose(w, (3, 4, 0, 1, 2))
if args[3] == 'channels_last':
x = np.transpose(x, (0, 4, 1, 2, 3))

y = func(x, w, args[2], args[3])

if args[3] == 'channels_last':
if y.ndim == 3:
y = np.transpose(y, (0, 2, 1))
elif y.ndim == 4:
y = np.transpose(y, (0, 2, 3, 1))
else:
y = np.transpose(y, (0, 2, 3, 4, 1))

return y

return wrapper


@normalize_ref_conv
def ref_conv(x, w, padding, data_format):
y = []
for i in range(x.shape[0]):
_y = []
for j in range(w.shape[1]):
__y = []
for k in range(w.shape[0]):
__y.append(signal.convolve(x[i, k], w[k, j], mode=padding))
_y.append(np.sum(np.stack(__y, axis=-1), axis=-1))
y.append(_y)
y = np.array(y)
return y


@normalize_ref_conv
def ref_depthwise_conv(x, w, padding, data_format):
y = []
for i in range(x.shape[0]):
_y = []
for j in range(w.shape[0]):
__y = []
for k in range(w.shape[1]):
__y.append(signal.convolve(x[i, j], w[j, k], mode=padding))
_y.append(np.stack(__y, axis=0))
y.append(np.concatenate(_y, axis=0))
y = np.array(y)
return y


def ref_separable_conv(x, w1, w2, padding, data_format):
x2 = ref_depthwise_conv(x, w1, padding, data_format)
return ref_conv(x2, w2, padding, data_format)


def ref_pool(x, pool_size, strides, padding, data_format, pool_mode):
if data_format == 'channels_last':
if x.ndim == 3:
x = np.transpose(x, (0, 2, 1))
elif x.ndim == 4:
x = np.transpose(x, (0, 3, 1, 2))
else:
x = np.transpose(x, (0, 4, 1, 2, 3))

if padding == 'same':
pad = [(0, 0), (0, 0)] + [(s // 2, s // 2) for s in pool_size]
x = np.pad(x, pad, 'constant', constant_values=-np.inf)

# indexing trick
x = np.pad(x, [(0, 0), (0, 0)] + [(0, 1) for _ in pool_size],
'constant', constant_values=0)

if x.ndim == 3:
y = [x[:, :, k:k1:strides[0]]
for (k, k1) in zip(range(pool_size[0]), range(-pool_size[0], 0))]
elif x.ndim == 4:
y = []
for (k, k1) in zip(range(pool_size[0]), range(-pool_size[0], 0)):
for (l, l1) in zip(range(pool_size[1]), range(-pool_size[1], 0)):
y.append(x[:, :, k:k1:strides[0], l:l1:strides[1]])
else:
y = []
for (k, k1) in zip(range(pool_size[0]), range(-pool_size[0], 0)):
for (l, l1) in zip(range(pool_size[1]), range(-pool_size[1], 0)):
for (m, m1) in zip(range(pool_size[2]), range(-pool_size[2], 0)):
y.append(x[:, :, k:k1:strides[0], l:l1:strides[1], m:m1:strides[2]])
y = np.stack(y, axis=-1)
if pool_mode == 'avg':
y = np.mean(np.ma.masked_invalid(y), axis=-1).data
elif pool_mode == 'max':
y = np.max(y, axis=-1)

if data_format == 'channels_last':
if y.ndim == 3:
y = np.transpose(y, (0, 2, 1))
elif y.ndim == 4:
y = np.transpose(y, (0, 2, 3, 1))
else:
y = np.transpose(y, (0, 2, 3, 4, 1))

return y


def ref_rnn(x, w, init, go_backwards=False, mask=None, unroll=False, input_length=None):
w_i, w_h, w_o = w
h = []
o = []

if go_backwards:
t_list = range(x.shape[1] - 1, -1, -1)
else:
t_list = range(x.shape[1])

if mask is not None:
np_mask = K.eval(mask)
else:
np_mask = None

for (i, t) in enumerate(t_list):
h_t = np.dot(x[:, t], w_i)

if w_h is not None:
prev = h[i - 1] if i > 0 else init
h_t1 = np.dot(prev, w_h)
if np_mask is not None:
h_t1[np_mask[:, t] == 0] = prev[np_mask[:, t] == 0]
else:
h_t1 = 0

o_t = h_t + h_t1
if w_o is not None:
o_t = np.dot(o_t, w_o)
o.append(o_t)

if np_mask is not None:
h_t = h_t * np_mask[:, t].reshape(-1, 1)
h.append(h_t + h_t1)

return o[-1], np.stack(o, axis=1), np.stack(h, axis=1)


class TestBackend(object):

def test_is_keras_tensor(self):
Expand Down Expand Up @@ -724,7 +569,7 @@ def rnn_fn(x_k, h_k):
]

for (i, kwargs) in enumerate(kwargs_list):
last_y1, y1, h1 = ref_rnn(x, [wi, wh, None], h0, **kwargs)
last_y1, y1, h1 = reference_operations.rnn(x, [wi, wh, None], h0, **kwargs)
last_y2, y2, h2 = K.rnn(rnn_fn, x_k, h0_k, **kwargs)

assert len(h2) == 1
Expand Down Expand Up @@ -771,8 +616,8 @@ def rnn_fn(x_k, h_k):
y_k = K.dot(x_k, wi_k)
return y_k, []

last_y1, y1, h1 = ref_rnn(x, [wi, None, None], None,
go_backwards=False, mask=None)
last_y1, y1, h1 = reference_operations.rnn(x, [wi, None, None], None,
go_backwards=False, mask=None)
last_y2, y2, h2 = K.rnn(rnn_fn, x_k, [],
go_backwards=False, mask=None)

Expand Down Expand Up @@ -1071,7 +916,7 @@ def test_conv(self, op, input_shape, kernel_shape, padding, data_format):
k = K.backend()
_, x = parse_shape_or_val(input_shape)
_, w = parse_shape_or_val(kernel_shape)
y1 = ref_conv(x, w, padding, data_format)
y1 = reference_operations.conv(x, w, padding, data_format)
y2 = check_two_tensor_operation(
op, x, w, [KTH if k == 'theano' else KC if k == 'cntk' else KTF],
padding=padding, data_format=data_format,
Expand All @@ -1088,7 +933,7 @@ def test_depthwise_conv(self, op, input_shape, kernel_shape, padding, data_forma
k = K.backend()
_, x = parse_shape_or_val(input_shape)
_, w = parse_shape_or_val(kernel_shape)
y1 = ref_depthwise_conv(x, w, padding, data_format)
y1 = reference_operations.depthwise_conv(x, w, padding, data_format)
y2 = check_two_tensor_operation(
op, x, w, [KTH if k == 'theano' else KC if k == 'cntk' else KTF],
padding=padding, data_format=data_format,
Expand All @@ -1108,7 +953,7 @@ def test_depthwise_conv(self, op, input_shape, kernel_shape, padding, data_forma
def test_pool(self, op, input_shape, pool_size, strides, padding, data_format, pool_mode):
k = K.backend()
_, x = parse_shape_or_val(input_shape)
y1 = ref_pool(x, pool_size, strides, padding, data_format, pool_mode)
y1 = reference_operations.pool(x, pool_size, strides, padding, data_format, pool_mode)
y2 = check_single_tensor_operation(
op, x, [KTH if k == 'theano' else KC if k == 'cntk' else KTF],
pool_size=pool_size, strides=strides,
Expand Down Expand Up @@ -1174,7 +1019,7 @@ def test_separable_conv2d(self, op, input_shape, kernel_shape, depth_multiplier,
_, x = parse_shape_or_val(input_shape)
_, depthwise = parse_shape_or_val(kernel_shape + (input_depth, depth_multiplier))
_, pointwise = parse_shape_or_val((1, 1) + (input_depth * depth_multiplier, 7))
y1 = ref_separable_conv(x, depthwise, pointwise, padding, data_format)
y1 = reference_operations.separable_conv(x, depthwise, pointwise, padding, data_format)
if K.backend() == 'cntk':
y2 = cntk_func_three_tensor(
op, input_shape,
Expand Down
Loading

0 comments on commit d673afd

Please sign in to comment.