def test_nonzero_slope(self):
bottom = Array.rand(256, 256).astype(np.float32) * 255
actual = Array.zeros(bottom.shape, np.float32)
relu(bottom, bottom, actual, 2.4)
expected = np.clip(bottom, 0.0, float('inf')) + \
2.4 * np.clip(bottom, float('-inf'), 0.0)
self._check(actual, expected)
def setup(self, bottom, top):
conv_param = self.layer_param.convolution_param
channels, height, width = bottom[0].shape
num_output = conv_param.num_output
assert channels % self.group == 0, \
"Number of channels should be a multiple of group."
assert num_output % self.group == 0, \
"Number of outputs should be a multiple of group."
self.bias_term = conv_param.bias_term
if self.weights is not None:
logging.debug("Skipping parameter initialization")
else:
weights_shape = (num_output, channels // self.group, self.kernel_h,
self.kernel_w)
weight_filler = conv_param.weight_filler
if weight_filler.type == 'gaussian':
self.weights = weight_filler.mean + weight_filler.std * \
Array.standard_normal(
weights_shape).astype(np.float32)
self.weight_diff = Array.empty_like(self.weights)
else:
raise Exception("Filler not implemented for weight filler \
type {}".format(weight_filler.type))
if self.bias_term:
self.bias = Array((num_output, ), np.float32)
self.bias_diff = Array.empty_like(self.bias)
filler = conv_param.bias_filler
if filler.type == 'constant':
self.bias.fill(filler.value)
else:
raise Exception("Filler not implemented for bias filler \
type {}".format(filler.type))
def test_simple_mul(self):
a = Array.rand(256, 256).astype(np.float32)
b = Array.rand(256, 256).astype(np.float32)
actual = a * b
expected = np.multiply(a, b)
self._check(actual, expected)
def setup(self, bottom, top):
conv_param = self.layer_param.convolution_param
channels, height, width = bottom[0].shape
num_output = conv_param.num_output
assert channels % self.group == 0, \
"Number of channels should be a multiple of group."
assert num_output % self.group == 0, \
"Number of outputs should be a multiple of group."
self.bias_term = conv_param.bias_term
if self.weights is not None:
logging.debug("Skipping parameter initialization")
else:
weights_shape = (num_output, channels // self.group,
self.kernel_h, self.kernel_w)
weight_filler = conv_param.weight_filler
if weight_filler.type == 'gaussian':
self.weights = weight_filler.mean + weight_filler.std * \
Array.standard_normal(
weights_shape).astype(np.float32)
self.weight_diff = Array.empty_like(self.weights)
else:
raise Exception("Filler not implemented for weight filler \
type {}".format(weight_filler.type))
if self.bias_term:
self.bias = Array((num_output, ), np.float32)
self.bias_diff = Array.empty_like(self.bias)
filler = conv_param.bias_filler
if filler.type == 'constant':
self.bias.fill(filler.value)
else:
raise Exception("Filler not implemented for bias filler \
type {}".format(filler.type))
def test_gradient(self):
for kernel_h in range(3, 5):
for kernel_w in range(3, 5):
channels = 12
height = 3
width = 5
bottom = Array.zeros((5, channels, height, width), np.int32)
bottom_diff = Array.zeros_like(bottom)
for n in range(5):
for c in range(channels):
bottom[n, c] = Array.array(
[[1, 2, 5, 2, 3],
[9, 4, 1, 4, 8],
[1, 2, 5, 2, 3]]).astype(np.int32)
param = self.layer[5]
param.pooling_param.kernel_h = kernel_h
param.pooling_param.kernel_w = kernel_w
param.pooling_param.stride = 2
param.pooling_param.pad = 1
layer = PoolingLayer(param)
top_shape = layer.get_top_shape(bottom)
top = Array.zeros(top_shape, np.int32)
top_diff = Array.zeros_like(top)
checker = GradientChecker(1e-4, 1e-2)
checker.check_gradient_exhaustive(layer, bottom, bottom_diff,
top, top_diff)
def test_nonzero_slope(self):
bottom = Array.rand(256, 256).astype(np.float32) * 255
actual = Array.zeros(bottom.shape, np.float32)
relu(bottom, bottom, actual, 2.4)
expected = np.clip(bottom, 0.0, float('inf')) + \
2.4 * np.clip(bottom, float('-inf'), 0.0)
self._check(actual, expected)
def test_simple_add(self):
a = Array.rand(256, 256).astype(np.float32)
b = Array.rand(256, 256).astype(np.float32)
actual = a + b
expected = np.add(a, b)
self._check(actual, expected)
def test_gradient(self):
for kernel_h in range(3, 5):
for kernel_w in range(3, 5):
channels = 12
height = 3
width = 5
bottom = Array.zeros((5, channels, height, width), np.int32)
bottom_diff = Array.zeros_like(bottom)
for n in range(5):
for c in range(channels):
bottom[n, c] = Array.array([[1, 2, 5, 2, 3],
[9, 4, 1, 4, 8],
[1, 2, 5, 2,
3]]).astype(np.int32)
param = self.layer[5]
param.pooling_param.kernel_h = kernel_h
param.pooling_param.kernel_w = kernel_w
param.pooling_param.stride = 2
param.pooling_param.pad = 1
layer = PoolingLayer(param)
top_shape = layer.get_top_shape(bottom)
top = Array.zeros(top_shape, np.int32)
top_diff = Array.zeros_like(top)
checker = GradientChecker(1e-4, 1e-2)
checker.check_gradient_exhaustive(layer, bottom, bottom_diff,
top, top_diff)
def backward(self, bottom_data, bottom_diff, top_data, top_diff):
padded_ratio = Array.zeros(1, bottom_data.shape[1] + self.size - 1,
bottom_data.shape[2], bottom_data.shape[3])
accum_ratio = Array.zeros(1, 1, bottom_data.shape[2],
bottom_data.shape[3])
accum_ratio_times_bottom = Array.zeros(1, 1, bottom_data.shape[2],
bottom_data.shape[3])
cache_ratio_value = 2.0 * self.apha * self.beta / self.size
bottom_diff = np.pow(self.scale, -self.beta)
bottom_diff *= top_diff
inverse_pre_pad = self.size - (self.size + 1) / 2
for n in range(bottom_data.shape[0]):
padded_ratio[0, inverse_pre_pad] = top_diff[n] * top_data[n]
padded_ratio[0, inverse_pre_pad] /= self.scale[n]
accum_ratio.fill(0)
for c in range(self.size - 1):
accum_ratio += padded_ratio[0, c]
for c in range(bottom_data.shape[1]):
accum_ratio += padded_ratio[0, c + self.size - 1]
accum_ratio_times_bottom += bottom_data[n, c] * accum_ratio
bottom_data[n, c] += -cache_ratio_value * \
accum_ratio_times_bottom
accum_ratio += -1 * padded_ratio[0, c]
def test_blase(self):
A = Array.rand(256, 256).astype(np.float32)
x = Array.rand(256, 256).astype(np.float32)
actual = Array.rand(256, 256).astype(np.float32)
gemm(A, x, actual, 1.0, 0.0, 256, 256, 256)
self._check(actual, np.dot(A, x))
def test_simple(self):
channels = 12
height = 3
width = 5
bottom = Array.zeros((5, channels, height, width), np.int32)
for n in range(5):
for c in range(channels):
bottom[n, c] = Array.array([[1, 2, 5, 2, 3], [9, 4, 1, 4, 8],
[1, 2, 5, 2, 3]]).astype(np.int32)
param = self.layer[5]
param.pooling_param.kernel_size = 2
param.pooling_param.stride = 1
layer = PoolingLayer(param)
actual_shape = layer.get_top_shape(bottom)
actual = Array.zeros(actual_shape, np.int32)
layer.setup(bottom, actual)
layer.forward(bottom, actual)
for n in range(5):
for c in range(channels):
np.testing.assert_array_equal(
actual[n, c],
np.array([[9, 5, 5, 8], [9, 5, 5, 8]]).astype(np.int32))
bottom = Array.zeros_like(bottom)
for n in range(5):
for c in range(channels):
actual[n, c] = Array.array([[1, 1, 1, 1],
[1, 1, 1, 1]]).astype(np.int32)
layer.backward(bottom, actual)
for n in range(5):
for c in range(channels):
np.testing.assert_array_equal(
bottom[n, c],
np.array([[0, 0, 2, 0, 0], [2, 0, 0, 0, 2],
[0, 0, 2, 0, 0]]).astype(np.int32))
def matrix_mult_sample():
A = Array.array([[1, 2], [3, 4]])
B = Array.array([[10, 3], [7, 4]])
C = dot(A.transpose(), B.transpose())
return C
def matrix_mult_sample():
B = Array.array([[10, 3], [7, 4]])
# transposition
A = Array.array([[1, 2], [3, 4]])
A = Array.transpose(A)
C = dot(B, A)
return C
def test_no_padding(self):
a = (Array.rand(256, 256) * 255).astype(np.float32)
weights = (Array.rand(5, 5) * 2).astype(np.float32)
actual = Array.zeros((254, 254), np.float32)
convolve(a, weights, actual, (0, 0), (1, 1))
expected = signal.convolve(a, np.fliplr(np.flipud(weights)),
mode='same')[2:, 2:]
self._check(actual, expected)
def test_simple(self):
a = (Array.rand(256, 256) * 255).astype(np.float32)
weights = (Array.rand(3, 3) * 2).astype(np.float32)
actual = Array.zeros(a.shape, np.float32)
convolve(a, weights, actual, (1, 1), (1, 1))
expected = signal.convolve(a, np.fliplr(np.flipud(weights)),
mode='same')
self._check(actual, expected)
def test_no_padding(self):
a = (Array.rand(256, 256) * 255).astype(np.float32)
weights = (Array.rand(5, 5) * 2).astype(np.float32)
actual = Array.zeros((254, 254), np.float32)
convolve(a, weights, actual, (0, 0), (1, 1))
expected = signal.convolve(a,
np.fliplr(np.flipud(weights)),
mode='same')[2:, 2:]
self._check(actual, expected)
def test_simple(self):
a = (Array.rand(256, 256) * 255).astype(np.float32)
weights = (Array.rand(3, 3) * 2).astype(np.float32)
actual = Array.zeros(a.shape, np.float32)
convolve(a, weights, actual, (1, 1), (1, 1))
expected = signal.convolve(a,
np.fliplr(np.flipud(weights)),
mode='same')
self._check(actual, expected)
def matrix_mult_sample():
A = Array.array([[1, 2], [3, 4]])
B = Array.array([[10, 3], [7, 4]])
M = Array.transpose(A)
M = Array.array([[1, 1], [3, 0]])
C = dot(M, B)
return C
def matrix_mult_sample():
A = Array.array([[1, 2], [3, 4]])
B = Array.array([[10, 3], [7, 4]])
# double transposition
A = Array.transpose(A)
B = Array.transpose(B)
C = dot(A, B)
return C
def test_forward_simple(self):
channels = 12
height = 3
width = 5
bottom = Array.rand(
5, channels, height, width).astype(np.float32)
bottom = bottom * 256 - 128
layer = ReluLayer(self.layer[3])
actual = Array.zeros(layer.get_top_shape(bottom), np.float32)
layer.forward(bottom, actual)
expected = np.clip(bottom, 0.0, float('inf')).astype(np.float32)
np.testing.assert_allclose(actual, expected)
def test_multiple_nested(self):
def matrix_mult_complex(A, B, C):
return dot(dot(A, B), C)
A = Array.rand(3, 3)
B = Array.rand(3, 3)
C = Array.rand(3, 3)
expected = matrix_mult_complex(A, B, C)
actual = dgemmify(matrix_mult_complex)(A, B, C)
self._check(actual, expected)
def _forward_test(self, param, in_shape):
conv_param = param.convolution_param
in_batch = Array.rand(*in_shape).astype(np.float32) * 255
conv = ConvLayer(param)
top_shape = conv.get_top_shape(in_batch)
expected_conv = NaiveConv(conv_param)
actual = Array.zeros(top_shape, np.float32)
expected = Array.zeros(top_shape, np.float32)
conv.setup(in_batch, actual)
conv.forward(in_batch, actual)
expected_conv(in_batch, conv.weights, conv.bias, expected)
self._check(actual, expected)
def test_simple(self):
A = Array.rand(256, 256).astype(np.float32)
x = Array.rand(256, 256).astype(np.float32)
b = Array.rand(256, 256).astype(np.float32)
@blasc
def fn(A, x, b):
v1 = T(A)
v2 = dot(v1, x)
v3 = v2 - b
return v3
self._check(fn(A, x, b), np.transpose(A).dot(x) - b)
def test_simple(self):
a = Array.rand(256, 256).astype(np.float32) * 255
actual_mask = Array.zeros((254, 254), np.float32)
actual = Array.zeros((254, 254), np.float32)
actual.fill(float('-inf'))
expected_mask = Array.zeros((254, 254), np.float32)
expected = Array.zeros((254, 254), np.float32)
expected.fill(float('-inf'))
max_pool(a, actual, actual_mask, (2, 2))
py_max_pool(a, expected, expected_mask, (2, 2), (1, 1), (0, 0))
self._check(actual, expected)
self._check(actual_mask, expected_mask)
def _forward_test(self, param, in_shape):
conv_param = param.convolution_param
in_batch = Array.rand(*in_shape).astype(np.float32) * 255
conv = ConvLayer(param)
top_shape = conv.get_top_shape(in_batch)
expected_conv = NaiveConv(conv_param)
actual = Array.zeros(top_shape, np.float32)
expected = Array.zeros(top_shape, np.float32)
conv.setup(in_batch, actual)
conv.forward(in_batch, actual)
expected_conv(in_batch, conv.weights, conv.bias, expected)
self._check(actual, expected)
def test_simple(self):
a = Array.rand(256, 256).astype(np.float32) * 255
actual_mask = Array.zeros((254, 254), np.float32)
actual = Array.zeros((254, 254), np.float32)
actual.fill(float('-inf'))
expected_mask = Array.zeros((254, 254), np.float32)
expected = Array.zeros((254, 254), np.float32)
expected.fill(float('-inf'))
max_pool(a, actual, actual_mask, (2, 2))
py_max_pool(a, expected, expected_mask,
(2, 2), (1, 1), (0, 0))
self._check(actual, expected)
self._check(actual_mask, expected_mask)
def forward(self, bottom, top):
# initialize scale to constant value
self.scale.fill(self.k)
padded_square = Array.zeros((bottom.shape[1] + self.size - 1,
bottom.shape[2], bottom.shape[3]),
bottom.dtype)
alpha_over_size = self.alpha / self.size
for n in range(bottom.shape[0]):
padded_square[self.pre_pad:bottom.shape[1] + self.pre_pad] = \
np.square(bottom[n])
for c in range(self.size):
self.scale[n] += alpha_over_size * padded_square[c]
# for c in range(1, bottom.shape[1]):
# self.scale[n, c] = self.scale[n, c - 1] + \
# alpha_over_size * padded_square[c + self.size - 1] - \
# alpha_over_size * padded_square[c - 1]
print('asdfasdf', c)
self.scale[n, 1:] = self.scale[n, :-1] + \
alpha_over_size * padded_square[self.size:self.size + bottom.shape[1] - 1] - \
alpha_over_size * padded_square[:bottom.shape[1] - 1]
top[:] = np.power(self.scale, -self.beta) * bottom
def main():
# Smaller Dataset
h = 1200 # height (number of rows, or column length)
w = 1000 # width (number of columns, or row length)
TOTAL_SIZE = h * w
length = 100
# Larger Dataset
# TOTAL_SIZE = 500000000
# h = 500000 # height (number of rows, or column length)
# w = 1000 # width (number of columns, or row length)
# length = 5000
block_set = Array.array(list(range(TOTAL_SIZE))) # sample dataset
block_set = block_set.reshape(h, w)
block_set = block_set.astype(np.float32)
start_time = time.time()
result = np.array(dcRemoval(block_set.flatten(), length, h).reshape(h, w))
time_total = time.time() - start_time
print "SEJITS dcRemoval Time: ", time_total, " seconds"
print "RESULT: ", result
return result
def test_range_of_fifty(self):
TOTAL_SIZE = 50
height = 25 # height: (number of rows, or column length)
width = 2 # width: (number of columns, or row length)
pfov_length = 5
# Creating a sample dataset for the SEJITS tests
sejits_block_set = Array.array(list(range(TOTAL_SIZE)))
sejits_block_set = sejits_block_set.reshape(height, width)
sejits_block_set = sejits_block_set.astype(np.float32)
# Creating a sample dataset for the Python tests
pyop_block_set = np.array(list(range(TOTAL_SIZE)))
pyop_block_set = pyop_block_set.reshape(height, width)
pyop_block_set = pyop_block_set.astype(np.float32)
python_result = dcRemPython(pyop_block_set.flatten(1), height, pfov_length).astype(
np.float32).reshape((height, width), order='F').astype(np.int32)
sejits_result = np.array(
dcRemSejits(sejits_block_set.flatten(), pfov_length, height)).astype(
np.float32).reshape((height, width)).astype(np.int32)
print sejits_result
self._check(python_result, sejits_result)
def test_two_inputs(self):
a = Array.rand(256, 256).astype(np.float32) * 255.0 - 128.0
b = Array.rand(256, 256).astype(np.float32) * 255.0 - 128.0
negative_slope = 0.0
@smap2
def fn(x, y):
if x > 0:
return y
else:
return negative_slope * y
actual = fn(a, b)
expected = b
expected[a <= 0] *= negative_slope
self._check(actual, expected)
def test_simple(self):
a = Array.rand(256, 256).astype(np.float32)
actual = transpose(a)
expected = np.transpose(a)
self._check(actual, expected)
def test_simple_mul_scalar(self):
a = Array.rand(256, 256).astype(np.float32)
actual = a * 3.0
expected = np.multiply(a, 3.0)
self._check(actual, expected)
actual = 3.0 * a
self._check(actual, expected)
def test_simple_add_scalar(self):
a = Array.rand(256, 256).astype(np.float32)
actual = a + 3.0
expected = np.add(a, 3.0)
self._check(actual, expected)
actual = 3.0 + a
self._check(actual, expected)
def setup(self, bottom, top):
weights_shape = (self.num_output, bottom.shape[0])
weight_filler = self.layer_param.inner_product_param.weight_filler
if weight_filler.type == 'gaussian':
self.weights = weight_filler.mean + weight_filler.std * \
Array.standard_normal(
weights_shape).astype(np.float32)
else:
raise Exception("Filler not implemented for weight filler"
"type {}".format(weight_filler.type))
def test_simple(self):
channels = 12
height = 3
width = 5
bottom = Array.zeros((5, channels, height, width), np.int32)
for n in range(5):
for c in range(channels):
bottom[n, c] = Array.array(
[[1, 2, 5, 2, 3],
[9, 4, 1, 4, 8],
[1, 2, 5, 2, 3]]).astype(np.int32)
param = self.layer[5]
param.pooling_param.kernel_size = 2
param.pooling_param.stride = 1
layer = PoolingLayer(param)
actual_shape = layer.get_top_shape(bottom)
actual = Array.zeros(actual_shape, np.int32)
layer.setup(bottom, actual)
layer.forward(bottom, actual)
for n in range(5):
for c in range(channels):
np.testing.assert_array_equal(
actual[n, c],
np.array([
[9, 5, 5, 8],
[9, 5, 5, 8]
]).astype(np.int32))
bottom = Array.zeros_like(bottom)
for n in range(5):
for c in range(channels):
actual[n, c] = Array.array(
[[1, 1, 1, 1],
[1, 1, 1, 1]]).astype(np.int32)
layer.backward(bottom, actual)
for n in range(5):
for c in range(channels):
np.testing.assert_array_equal(
bottom[n, c],
np.array([[0, 0, 2, 0, 0],
[2, 0, 0, 0, 2],
[0, 0, 2, 0, 0]]).astype(np.int32))
def __init__(self, layer_param):
super(InnerProductLayer, self).__init__(layer_param)
param = self.layer_param.inner_product_param
self.num_output = param.num_output
self.bias_term = param.bias_term
if self.bias_term:
self.bias = Array.zeros(self.num_output)
filler = param.bias_filler
if filler.type == 'constant':
self.bias.fill(filler.value)
else:
raise Exception("Filler not implemented for bias filler \
type {}".format(filler.type))
def test_simple(self):
a = Array.rand(256, 256).astype(np.float32)
@smap
def fn(x):
if x > 0:
return x
else:
return 0
actual = fn(a)
expected = np.copy(a)
expected[expected < 0] = 0
self._check(actual, expected)
def test_backward_simple(self):
channels = 12
height = 3
width = 5
bottom = Array.rand(
5, channels, height, width).astype(np.float32)
bottom = bottom * 256 - 128
top_diff = Array.rand(
5, channels, height, width).astype(np.float32)
top_diff = top_diff * 256 - 128
top = np.zeros(top_diff.shape, np.float32)
actual = Array.zeros(bottom.shape, np.float32)
layer = ReluLayer(self.layer[3])
layer.backward(bottom, actual, top, top_diff)
expected = np.multiply(top_diff,
np.greater(bottom, Array.zeros(bottom.shape,
np.float32)))
np.testing.assert_allclose(actual, expected)
def test_simple(self):
bottom = Array.rand(3, 8, 32, 32).astype(np.float32)
actual = Array.zeros_like(bottom)
layer = LRNLayer(self.layer[4])
param = layer.layer_param.lrn_param
alpha = param.alpha
size = param.local_size
beta = param.beta
layer.setup(bottom, actual)
layer.forward(bottom, actual)
expected = Array.zeros_like(bottom)
for n in range(bottom.shape[0]):
for c in range(bottom.shape[1]):
for h in range(bottom.shape[2]):
for w in range(bottom.shape[3]):
c_start = c - (size - 1) // 2
c_end = min(c_start + size, bottom.shape[1])
scale = 1
for i in range(c_start, c_end):
value = bottom[n, i, h, w]
scale += value * value * alpha / size
expected = bottom[n, c, h, w] / pow(scale, beta)
self.assertTrue(
abs(actual[n, c, h, w] - expected) < 1e-4)
def forward(self, bottom_data, bottom_label, top):
accuracy = 0
data = bottom_data
label = bottom_label
dim = np.prod(data.shape) / data.shape[0]
# Perform a partial sort to find top_k
for i in range(data.shape[0]):
vec = Array.array([[data[i * dim + j], j] for j in range(dim)])
vec.partition((0, self.top_k))
# If label is in top_k increase accuracy
for k in range(self.top_k):
if vec[k][1] == label[i]:
accuracy += 1
top[0] = accuracy / data.shape[0]
def forward(self, bottom_data, bottom_label, top):
accuracy = 0
data = bottom_data
label = bottom_label
dim = np.prod(data.shape) / data.shape[0]
# Perform a partial sort to find top_k
for i in range(data.shape[0]):
vec = Array.array(
[[data[i * dim + j], j] for j in range(dim)])
vec.partition((0, self.top_k))
# If label is in top_k increase accuracy
for k in range(self.top_k):
if vec[k][1] == label[i]:
accuracy += 1
top[0] = accuracy / data.shape[0]
def matrix_mult_sample():
A = Array.array([[1, 2], [3, 4]])
B = Array.array([[10, 3], [7, 4]])
M = Array.transpose(A)
N = Array.transpose(B)
C = dot(M, N)
# the double assignment after the dot call
M = Array.array([[1, 1], [3, 0]])
N = Array.array([[1, 0], [7, 0]])
return C
def matrix_mult_sample():
A = Array.array([[1, 2], [3, 4]])
B = Array.array([[10, 3], [7, 4]])
A = Array.transpose(A)
B = Array.transpose(B)
C = dot(A, B)
# the double transpose in place after the dot call
A = Array.transpose(A)
B = Array.transpose(B)
return C
def test_range_of_ten(self):
TOTAL_SIZE = 10
height = 10 # height: (number of rows, or column length)
width = 1 # width: (number of columns, or row length)
pfov_length = 5
# Creating a sample dataset for the SEJITS tests
sejits_block_set = Array.array(list(range(TOTAL_SIZE)))
sejits_block_set = sejits_block_set.reshape(height, width)
sejits_block_set = sejits_block_set.astype(np.float32)
# Creating a sample dataset for the Python tests
pyop_block_set = np.array(list(range(TOTAL_SIZE)))
pyop_block_set = pyop_block_set.reshape(height, width)
pyop_block_set = pyop_block_set.astype(np.float32)
python_result = dcRemPython(pyop_block_set, height, pfov_length)
sejits_result = np.array(dcRemSejits(sejits_block_set, pfov_length, height))
self._check(python_result, sejits_result)
def add_blob(self, blob, shape):
self.blobs[blob] = Array.zeros(shape, np.float32)
def setup(self, bottom, top):
self.sum_multiplier = Array.ones((bottom.shape[1]), np.float32)
def setup(self, bottom_data, bottom_label, top):
self.prob = Array.zeros_like(bottom_data)
self.softmax_layer.setup(bottom_data, self.prob)
class ConvLayer(BaseLayer):
def __init__(self, param):
super(ConvLayer, self).__init__(param)
conv_param = param.convolution_param
if conv_param.kernel_size:
self.kernel_h = conv_param.kernel_size
self.kernel_w = conv_param.kernel_size
else:
self.kernel_h = conv_param.kernel_h
self.kernel_w = conv_param.kernel_w
assert (self.kernel_h, self.kernel_w) > (0, 0), \
"Filter dimensions cannot be zero."
self.padding = (conv_param.pad, conv_param.pad)
self.stride = (conv_param.stride, conv_param.stride)
assert conv_param.num_output > 0, "Layer must have at least one output"
self.group = conv_param.group
self.weights = None
self.bias_term = None
self.bias = None
self.kernel_size = self.kernel_h, self.kernel_w
self.im2col = Im2Col(self.kernel_size, self.stride, self.padding)
def get_top_shape(self, bottom):
conv_param = self.layer_param.convolution_param
height_out = (bottom.shape[2] + 2 * self.padding[0] - self.kernel_h) // \
self.stride[0] + 1
width_out = (bottom.shape[3] + 2 * self.padding[1] - self.kernel_w) // \
self.stride[1] + 1
return bottom.shape[0], conv_param.num_output, height_out, width_out
def setup(self, bottom, top):
conv_param = self.layer_param.convolution_param
channels, height, width = bottom[0].shape
num_output = conv_param.num_output
assert channels % self.group == 0, \
"Number of channels should be a multiple of group."
assert num_output % self.group == 0, \
"Number of outputs should be a multiple of group."
self.bias_term = conv_param.bias_term
if self.weights is not None:
logging.debug("Skipping parameter initialization")
else:
weights_shape = (num_output, channels // self.group,
self.kernel_h, self.kernel_w)
weight_filler = conv_param.weight_filler
if weight_filler.type == 'gaussian':
self.weights = weight_filler.mean + weight_filler.std * \
Array.standard_normal(
weights_shape).astype(np.float32)
self.weight_diff = Array.empty_like(self.weights)
else:
raise Exception("Filler not implemented for weight filler \
type {}".format(weight_filler.type))
if self.bias_term:
self.bias = Array((num_output, ), np.float32)
self.bias_diff = Array.empty_like(self.bias)
filler = conv_param.bias_filler
if filler.type == 'constant':
self.bias.fill(filler.value)
else:
raise Exception("Filler not implemented for bias filler \
type {}".format(filler.type))
# @meta
def forward(self, bottom, top):
weights = self.weights.reshape(self.weights.shape[0],
np.prod(self.weights.shape[1:]))
for bottom_data, top_data in zip(bottom, top):
self.col_data = self.im2col(
bottom_data, self.kernel_size, self.padding, self.stride)
col_offset = self.col_data.shape[0] // self.group
weight_offset = top_data.shape[0] // self.group
top_offset = top_data.shape[0] // self.group
for g in range(self.group):
top_data[g * top_offset:(g + 1) * top_offset] = (
weights[g * weight_offset:(g + 1) * weight_offset].dot(
self.col_data[g * col_offset:(g + 1) * col_offset])
).reshape(top_data[g * top_offset:(g + 1) * top_offset].shape)
if self.bias_term:
for output_data, bias in zip(top_data, self.bias):
output_data += bias
# out_groups = top.shape[1] // self.group
# in_groups = bottom.shape[1] // self.group
# for i in range(len(top)):
# for group in range(self.group):
# for out_group in range(out_groups):
# for in_group in range(in_groups):
# convolve(
# bottom[i, in_group + group * in_groups],
# self.weights[out_group + group * out_groups,
# in_group],
# top[i, out_group + group * out_groups],
# self.padding, self.stride)
# if self.bias_term:
# for j in range(len(self.bias)):
# # TODO: Add support for sugar
# # top[i, j] += self.bias[j]
# top[i, j] += self.bias[j]
def backward(self, top_data, top_diff, bottom_data, bottom_diff):
if self.propagate_down:
self.weight_diff.fill(0)
if self.bias is not None:
self.bias_diff.fill(0)
for i in range(top_data.shape[0]):
if self.bias is not None and self.propagate_down:
curr_bias_diff = self.bias_diff[i]
for n in range(top_data.shape[1]):
for data, bias_diff in zip(top_data[i, n], curr_bias_diff):
top_data += bias_diff
if self.propagate_down:
for n in range(top_data.shape[1]):
self.col_data = self.im2col(bottom_data, self.kernel_size,
self.padding, self.stride)
col_offset = self.col_data.shape[0] // self.group
weight_offset = top_data.shape[0] // self.group
top_offset = top_data.shape[0] // self.group
# TODO: Clean this up with helper functions
for g in range(self.group):
self.weight_diff[i, g * weight_offset:(g + 1) * weight_offset] += \
top_diff[i, g * top_offset:
(g + 1) * top_offset].dot(
self.col_data[
i, g * col_offset:(g + 1) * col_offset])
for g in range(self.group):
self.col_data[g * col_offset:(g + 1) * col_offset] = \
self.weights[
g * weight_offset:(g + 1) * weight_offset].dot(
self.top_diff[n, g * top_offset:
(g + 1) * top_offset])
bottom_diff[i] = self.col2im(self.col_data,
self.kernel_size,
self.padding, self.stride)
def test_simple(self):
bottom = Array.rand(256, 256).astype(np.float32) * 255
actual = Array.zeros(bottom.shape, np.float32)
relu(bottom, bottom, actual, 0.0)
expected = np.clip(bottom, 0.0, float('inf'))
self._check(actual, expected)
def setup(self, bottom, top):
self.scale = Array.zeros_like(bottom)
def __call__(self, *args):
output = Array.zeros(self.out_shape, args[0].dtype)
self._c_function(args[0], output)
return output
def __call__(self, *args):
output = Array.empty(self.out_shape, np.float32)
self._c_function(args[0], output)
return output