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, pool_size, activate_func: str = None) -> None:
self.height, self.width, self.channel = input_shape
self.pool_size = pool_size
assert self.height % pool_size == 0
assert self.width % pool_size == 0
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,
dim_in,
dim_out,
learning_rate,
activate_func: str = None) -> None:
self.dim_in = dim_in
self.dim_out = dim_out
self.learning_rate = learning_rate
self.params = torch.randn((dim_out, dim_in), dtype=floatX) / 500
self.biases = torch.randn((dim_out, ), dtype=floatX) / 500
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)
elif activate_func == 'softmax':
self.activation = other.Softmax()
else:
self.activation = None