def conv_efficient(self, x, w, b, output_size, vec_idx_key, strides=1):
batches = x.shape[0]
depth_i = x.shape[1]
filter_size = w.shape[2]
depth_o = w.shape[0]
if 4 == x.ndim: # 原始规格:
input_size = x.shape[2] #
p = int(((output_size - 1) * strides + filter_size - input_size) /
2) # padding尺寸
# logger.debug("padding begin..")
if p > 0: # 需要padding处理
x_pad = Tools.padding(x, p, self.dataType)
else:
x_pad = x
st = time.time()
logger.debug("vecting begin..")
# 可以根据自己的硬件环境，在三种优化方式中选择较快的一种
x_col = self.vectorize4conv_batches(x_pad, filter_size,
output_size, strides)
#x_col = spd.vectorize4conv_batches(x_pad, filter_size, output_size, strides)
#x_col = vec_by_idx(x_pad, filter_size, filter_size,vec_idx_key,0, strides)
logger.debug("vecting end.. %f s" % (time.time() - st))
else: # x_col规格
x_col = x
w_row = w.reshape(depth_o, x_col.shape[1])
conv = np.zeros((batches, depth_o, (output_size * output_size)),
dtype=self.dataType)
st1 = time.time()
logger.debug("matmul begin..")
#不广播，提高处理效率
for batch in range(batches):
conv[batch] = Tools.matmul(w_row, x_col[batch]) + b
logger.debug("matmul end.. %f s" % (time.time() - st1))
conv_return = conv.reshape(batches, depth_o, output_size, output_size)
return conv_return
def conv4dw(self, x, w, output_size, b=0, strides=1, x_v=False):
batches = x.shape[0]
depth_i = x.shape[1]
filter_size = w.shape[2] # 过滤器尺寸,对应卷积层误差矩阵尺寸
x_per_filter = filter_size * filter_size
depth_o = w.shape[1]
if False == x_v: # 原始规格:
input_size = x.shape[2] #
p = int(((output_size - 1) * strides + filter_size - input_size) /
2) # padding尺寸
if p > 0: # 需要padding处理
x_pad = Tools.padding(x, p, self.dataType)
else:
x_pad = x
logger.debug("vec4dw begin..")
x_col = self.vectorize4convdw_batches(x_pad, filter_size,
output_size, strides)
logger.debug("vec4dw end..")
else: # x_col规格
x_col = x
w_row = w.reshape(batches, depth_o, x_per_filter)
conv = np.zeros(
(batches, depth_i, depth_o, (output_size * output_size)),
dtype=self.dataType)
logger.debug("conv4dw matmul begin..")
for batch in range(batches):
for col in range(depth_i):
conv[batch, col] = Tools.matmul(w_row[batch], x_col[batch,
col])
conv_sum = np.sum(conv, axis=0)
# transpose而不是直接reshape避免错位
conv = conv_sum.transpose(1, 0, 2).reshape(depth_o, depth_i,
output_size, output_size)
logger.debug("conv4dw matmul end..")
return conv, x_col