提交 a015ea8f 编写于 作者: C chengduoZH

refine conv2d naive function

上级 b504a234
...@@ -3,30 +3,50 @@ import numpy as np ...@@ -3,30 +3,50 @@ import numpy as np
from op_test import OpTest from op_test import OpTest
def conv2d_forward_naive(input, filter, group, conv_param):
in_n, in_c, in_h, in_w = input.shape
out_c, f_c, f_h, f_w = filter.shape
assert f_c * group == in_c
assert np.mod(out_c, group) == 0
sub_out_c = out_c / group
stride, pad = conv_param['stride'], conv_param['pad']
out_h = 1 + (in_h + 2 * pad - f_h) / stride
out_w = 1 + (in_w + 2 * pad - f_w) / stride
out = np.zeros((in_n, out_c, out_h, out_w))
input_pad = np.pad(input, ((0, ), (0, ), (pad, ), (pad, )),
mode='constant',
constant_values=0)
for i in range(out_h):
for j in range(out_w):
for g in range(group):
input_pad_masked = input_pad[:, g * f_c:(
g + 1) * f_c, i * stride:i * stride + f_h, j * stride:j *
stride + f_w]
f_sub = filter[g * sub_out_c:(g + 1) * sub_out_c, :, :, :]
for k in range(sub_out_c):
out[:, g * sub_out_c + k, i, j] = np.sum(input_pad_masked *
f_sub[k, :, :, :],
axis=(1, 2, 3))
return out
class TestConv2dOp(OpTest): class TestConv2dOp(OpTest):
def setUp(self): def setUp(self):
self.init_groups() self.init_groups()
self.op_type = "conv2d" self.op_type = "conv2d"
batch_size = 2 input_size = [2, 3, 5, 5] # NCHW
input_channels = 3 assert np.mod(input_size[1], self.groups) == 0
input_height = 5 f_c = input_size[1] / self.groups
input_width = 5 filter_size = [6, f_c, 3, 3]
output_channels = 6 conv2d_param = {'stride': 1, 'pad': 0}
filter_height = 3
filter_width = 3 input = np.random.random(input_size).astype("float32")
stride = 1 filter = np.random.random(filter_size).astype("float32")
padding = 0
output_height = (input_height - filter_height + 2 * padding output = conv2d_forward_naive(input, filter, self.groups, conv2d_param)
) / stride + 1
output_width = (input_width - filter_width + 2 * padding) / stride + 1
input = np.random.random((batch_size, input_channels, input_height,
input_width)).astype("float32")
filter = np.random.random(
(output_channels, input_channels / self.groups, filter_height,
filter_width)).astype("float32")
output = np.ndarray(
(batch_size, output_channels, output_height, output_width))
self.inputs = {'Input': input, 'Filter': filter} self.inputs = {'Input': input, 'Filter': filter}
self.attrs = { self.attrs = {
...@@ -34,39 +54,6 @@ class TestConv2dOp(OpTest): ...@@ -34,39 +54,6 @@ class TestConv2dOp(OpTest):
'paddings': [0, 0], 'paddings': [0, 0],
'groups': self.groups 'groups': self.groups
} }
output_group_channels = output_channels / self.groups
input_group_channels = input_channels / self.groups
for batchid in xrange(batch_size):
for group in xrange(self.groups):
for outchannelid in range(group * output_group_channels,
(group + 1) * output_group_channels):
for rowid in xrange(output_height):
for colid in xrange(output_width):
start_h = (rowid * stride) - padding
start_w = (colid * stride) - padding
output_value = 0.0
for inchannelid in range(
group * input_group_channels,
(group + 1) * input_group_channels):
for frowid in xrange(filter_height):
for fcolid in xrange(filter_width):
input_value = 0.0
inrowid = start_h + frowid
incolid = start_w + fcolid
if ((inrowid >= 0 and
inrowid < input_height) and
(incolid >= 0 and
incolid < input_width)):
input_value = input[batchid][
inchannelid][inrowid][incolid]
filter_value = filter[outchannelid][
inchannelid % input_group_channels][
frowid][fcolid]
output_value += input_value * filter_value
output[batchid][outchannelid][rowid][
colid] = output_value
self.outputs = {'Output': output} self.outputs = {'Output': output}
def test_check_output(self): def test_check_output(self):
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册