def c_d_w_b_yn1(self, c_d_yn, y_n, y_n_1, w, is_c_d_yn1=1):
"""
stride, padding are both the integer value in the feed forward case.
get derivative of cost w.r.t. weight, bias, prev_layer.
ARGUMENT:
c_d_yn (batch) x (channel_n) x (height) x (width)
y_n (batch) x (channel_n) x (height) x (width)
y_n_1 (batch) x (channel_n_1) x (height') x (width')
w (channel_n_1) x (channel_n) x (kern) x (kern)
"""
c_d_xn = self._c_d_xn(c_d_yn, y_n)
c_d_b = np.sum(c_d_xn, axis=(0,2,3))
#### c_d_w ####
## y_n_1 (*) flipped(c_d_xn)
# patch_stride = stride
# padding = padding
# slide_stride = 1
c_d_w = slid.slid_win_4d_flip(np.swapaxes(y_n_1,0,1), np.swapaxes(c_d_xn,0,1),
1, self.stride, self.padding, slid.convolution())
assert c_d_w.shape == w.shape
#### c_d_yn1 ####
## c_d_xn (*) w ##
# patch_stride = 1/stride
# padding = (kern-padding-1)/stride
# slide_stride = 1/stride
pad2 = Fraction(w.shape[-1] - self.padding - 1, self.stride)
c_d_yn1 = slid.slid_win_4d_flip(c_d_xn, w[:,:,::-1,::-1], Fraction(1, self.stride),
Fraction(1, self.stride), pad2, slid.convolution())
assert c_d_yn1.shape == y_n_1.shape
return c_d_w, c_d_b, c_d_yn1
def act_forward(self, prev_layer, w, b):
"""
NOTE:
w is actually the flipped kernel:
y = x (*) kernel = x (*) flipped(w)
ARGUMENTS:
prev_layer: (batch) x (channel_in) x (height) x (width)
w: (channel_in) x (channel_out) x (kernel) x (kernel)
b: (channel_out)
OUTPUT:
(batch) x (channel_out) x (height') x (width')
please refer to slid_win_4d_flip for height' and width'
"""
ret = slid.slid_win_4d_flip(prev_layer, np.swapaxes(w, 0, 1),
self.stride, 1, self.padding, slid.convolution())
b_exp = b[np.newaxis, :, np.newaxis, np.newaxis]
return np.clip(ret+b_exp, 0, np.finfo(np.float64).max) # ReLU
