def __init__(self,
input_shape,
out_channel,
kernel_size,
stride,
learning_rate,
activate_func: str = None):
self.in_h, self.in_w, self.in_channel = input_shape
self.learning_rate = learning_rate
self.out_channel = out_channel
self.kernel_h, self.kernel_w = kernel_size
self.stride_h, self.stride_w = stride
# ignore padding
assert (self.in_h - self.kernel_h) % self.stride_h == 0
assert (self.in_w - self.kernel_w) % self.stride_w == 0
self.out_h = (self.in_h - self.kernel_h) // self.stride_h + 1
self.out_w = (self.in_w - self.kernel_w) // self.stride_w + 1
self.filters = torch.randn(
(self.kernel_h, self.kernel_w, self.in_channel, out_channel),
dtype=floatX)
self.biases = torch.randn((self.out_channel, ), dtype=floatX)
self.filters_gradient = torch.empty(
(self.kernel_h, self.kernel_w, self.in_channel, out_channel),
dtype=floatX)
if activate_func == 'relu':
self.activation = other.Relu()
elif activate_func == 'sigmoid':
self.activation = other.Sigmoid(100)
elif activate_func == 'tanh':
self.activation = other.Tanh(100)
else:
self.activation = None
def __init__(self,
input_shape,
out_channel,
kernel_size,
stride,
learning_rate,
padding=(0, 0),
activate_func: str = None):
self.in_h, self.in_w, self.in_channel = input_shape
self.learning_rate = learning_rate
self.out_channel = out_channel
self.kernel_h, self.kernel_w = kernel_size
self.stride_h, self.stride_w = stride
self.in_pad_h, self.in_pad_w = padding
# padding
self.real_in_h = self.in_h
self.real_in_w = self.in_w
self.in_h += 2 * self.in_pad_h
self.in_w += 2 * self.in_pad_w
assert (self.in_h - self.kernel_h) % self.stride_h == 0
assert (self.in_w - self.kernel_w) % self.stride_w == 0
self.out_h = (self.in_h - self.kernel_h) // self.stride_h + 1
self.out_w = (self.in_w - self.kernel_w) // self.stride_w + 1
# xaiver 初始化
n_in = self.in_h * self.in_w * self.in_channel
n_out = self.out_h * self.out_w * self.out_channel
coe = math.sqrt(6) / math.sqrt(n_in + n_out)
print('conv coe', coe)
self.filters = torch.tensor(np.random.uniform(
-coe,
coe,
size=(self.kernel_h, self.kernel_w, self.in_channel, out_channel)),
dtype=floatX,
device=device)
self.biases = torch.tensor(np.random.uniform(
-coe, coe, size=(self.out_channel, )),
dtype=floatX,
device=device)
# self.filters = torch.randn(
# (self.kernel_h, self.kernel_w, self.in_channel, out_channel),
# dtype=floatX, device=device)
# self.biases = torch.randn((self.out_channel,), dtype=floatX, device=device)
# self.filters_gradient = torch.empty(
# (self.kernel_h, self.kernel_w, self.in_channel, out_channel),
# dtype=floatX, device=device)
if activate_func == 'relu':
self.activation = other.Relu()
elif activate_func == 'mfm':
self.activation = other.MFM()
else:
self.activation = None