def __init__(self, tensor1, tensor2):
super(Add, self).__init__(tensor1.value + tensor2.value)
self.requires_grad = tensor1.requires_grad or tensor2.requires_grad
if tensor1.requires_grad:
back_channel = Node(tensor1, self.der_pos1)
self.tensor1 = tensor1
self.channels.append(back_channel)
if tensor2.requires_grad:
back_channel = Node(tensor2, self.der_pos2)
self.tensor2 = tensor2
self.channels.append(back_channel)
def __init__(self, tensor1):
super(Negate, self).__init__(-tensor1.value)
self.requires_grad = tensor1.requires_grad
if tensor1.requires_grad:
back_channel = Node(tensor1, self.der_pos)
self.channels.append(back_channel)
def __init__(self, y_pred, y_target):
super(LogLikelihood,
self).__init__(value=-(y_target.value * torch.log(y_pred.value)))
self.y_pred = y_pred
self.y_target = y_target
back_channel = Node(autoVariable=y_pred, vjp=self.der)
self.channels.append(back_channel)
def __init__(self, y_pred, y_target):
super(MeanAbsoluteError,
self).__init__(value=torch.abs(y_pred.value - y_target.value))
self.y_pred = y_pred
self.y_target = y_target
back_channel = Node(autoVariable=y_pred, vjp=self.der)
self.channels.append(back_channel)
def __init__(self, y_pred, y_target):
super(SquareError,
self).__init__(value=0.5 * ((y_pred.value - y_target.value)**2))
self.y_pred = y_pred
self.y_target = y_target
back_channel = Node(autoVariable=y_pred, vjp=self.der)
self.channels.append(back_channel)
def __init__(self, tensor: autoTensor, *idx):
super(Transpose, self).__init__(self._transpose(tensor, *idx))
self.requires_grad = tensor.requires_grad
self.idx = list(idx)
if tensor.requires_grad:
back_channel = Node(tensor, self.der)
self.channels.append(back_channel)
def __init__(self, inputs, mask):
super(Dpout, self).__init__(value=inputs.value * mask)
self.requires_grad = inputs.requires_grad
if self.requires_grad:
self.mask = mask
back_channel = Node(inputs, self.der)
self.channels.append(back_channel)
def __init__(self, tensor1):
super(Exp, self).__init__(torch.exp(tensor1.value))
self.requires_grad = tensor1.requires_grad
if tensor1.requires_grad:
back_channel = Node(tensor1, self.der_pos1)
self.tensor1 = tensor1
self.channels.append(back_channel)
def __init__(self, y_pred, y_target):
super(BinaryCrossEntropy, self).__init__(
value=-(y_target.value * torch.log(y_pred.value) +
(1 - y_target.value) * torch.log(1 - y_pred.value)))
self.y_pred = y_pred
self.y_target = y_target
back_channel = Node(autoVariable=y_pred, vjp=self.der)
self.channels.append(back_channel)
def __init__(self, inputs):
super(Flatten2d,
self).__init__(value=inputs.value.view(inputs.size()[0], -1))
self.requires_grad = inputs.requires_grad
if self.requires_grad:
self.inputs_size = inputs.size()
back_channel = Node(inputs, self.der)
self.channels.append(back_channel)
def __init__(self, tensor1, axis=0):
super(Sum, self).__init__(tensor1.value.sum(dim=axis, keepdim=True))
self.requires_grad = tensor1.requires_grad
if tensor1.requires_grad:
back_channel = Node(tensor1, self.der_pos1)
self.shape = tensor1.value.size()
self.channels.append(back_channel)
def __init__(self, tensor1, pow_val):
super(Power, self).__init__(tensor1.value**pow_val)
self.requires_grad = tensor1.requires_grad
if tensor1.requires_grad:
back_channel = Node(tensor1, self.der_pos1)
self.tensor1 = tensor1
self.pow_val = pow_val
self.channels.append(back_channel)
def __init__(self, image_block, filters, bias, padding=0, stride=1):
conv = F.conv2d(input=image_block.value,
weight=filters.value,
bias=bias.value,
stride=(stride, stride),
padding=(padding, padding))
super(Conv2d, self).__init__(value=conv)
self.requires_grad = image_block.requires_grad or filters.requires_grad or bias.requires_grad
self.padding = padding
self.stride = stride
self.image_block = image_block
self.filters = filters
self.bias = bias
if image_block.requires_grad:
back_channel = Node(image_block, self.der_image)
self.channels.append(back_channel)
if filters.requires_grad:
back_channel = Node(filters, self.der_filters)
self.channels.append(back_channel)
if bias.requires_grad:
back_channel = Node(bias, self.der_bias)
self.channels.append(back_channel)
