def forward(self, w_value, x_value, b_value, is_numba):
self.w_value = w_value
self.x_value = x_value
self.b_value = b_value
if is_numba:
x_value, col, col_w, out = self._forward(w_value, x_value, b_value,
self.pad, self.stride)
self.x_value = x_value
self.col = col
self.col_w = col_w
return out
else:
FN, C, FH, FW = self.w_value.shape
N, C, H, W = x_value.shape
out_h = 1 + int((H + 2 * self.pad - FH) / self.stride)
out_w = 1 + int((W + 2 * self.pad - FW) / self.stride)
col = tff.im2col(x_value, FH, FW, self.stride, self.pad)
col_w = self.w_value.reshape(FN, -1).T
out = np.dot(col, col_w) + self.b_value
out = out.reshape(N, out_h, out_w, -1).transpose(0, 3, 1, 2)
self.x_value = x_value
self.col = col
self.col_w = col_w
return out
def _forward(w_value, x_value, b_value, pad, stride):
FN, C, FH, FW = w_value.shape
N, C, H, W = x_value.shape
out_h = 1 + int((H + 2 * pad - FH) / stride)
out_w = 1 + int((W + 2 * pad - FW) / stride)
col = tff.im2col(x_value, FH, FW, stride, pad)
col_w = w_value.reshape(FN, -1).T
out = np.dot(col, col_w) + b_value
out = out.reshape(N, out_h, out_w, -1).transpose(0, 3, 1, 2)
return x_value, col, col_w, out
def _forward(x_value, filter_h, filter_w, pad, stride):
N, C, H, W = x_value.shape
out_h = int(1 + (H - filter_h) / stride)
out_w = int(1 + (W - filter_w) / stride)
col = tff.im2col(x_value, filter_h, filter_w, stride, pad)
col = col.reshape(-1, filter_h * filter_w)
arg_max = np.argmax(col, axis=1)
out = np.max(col, axis=1)
out = out.reshape(N, out_h, out_w, C).transpose(0, 3, 1, 2)
return arg_max, out
def _forward(w_value, x_value, pad, stride):
FN, C, FH, FW = w_value.shape
N, C, H, W = x_value.shape
out_h = int(1 + (H - FH) / stride)
out_w = int(1 + (W - FW) / stride)
col = tff.im2col(x_value, FH, FW, stride, pad)
col = col.reshape(-1, FH * FW)
arg_max = np.argmax(col, axis=1)
out = np.max(col, axis=1)
out = out.reshape(N, out_h, out_w, C).transpose(0, 3, 1, 2)
return arg_max, out
def forward(self, x_value, is_train=True, is_numba=False):
self.x_value = x_value
if is_numba:
arg_max, out = self._forward(self.x_value, self.filter_h,
self.filter_w, self.pad, self.stride)
self.arg_max = arg_max
return out
else:
N, C, H, W = self.x_value.shape
out_h = int(1 + (H - self.filter_h) / self.stride)
out_w = int(1 + (W - self.filter_w) / self.stride)
col = tff.im2col(self.x_value, self.filter_h, self.filter_w,
self.stride, self.pad)
col = col.reshape(-1, self.filter_h * self.filter_w)
arg_max = np.argmax(col, axis=1)
out = np.max(col, axis=1)
out = out.reshape(N, out_h, out_w, C).transpose(0, 3, 1, 2)
self.arg_max = arg_max
return out
def forward(self, w_value, x_value, is_numba):
self.w_value = w_value
self.x_value = x_value
if is_numba:
arg_max, out = self._forward(self.w_value, self.x_value, self.pad,
self.stride)
self.arg_max = arg_max
return out
else:
FN, C, FH, FW = self.w_value.shape
N, C, H, W = self.x_value.shape
out_h = int(1 + (H - FH) / self.stride)
out_w = int(1 + (W - FW) / self.stride)
col = tff.im2col(self.x_value, FH, FW, self.stride, self.pad)
col = col.reshape(-1, FH * FW)
arg_max = np.argmax(col, axis=1)
out = np.max(col, axis=1)
out = out.reshape(N, out_h, out_w, C).transpose(0, 3, 1, 2)
self.arg_max = arg_max
return out