I am writing a piece of code to better understand how autograd and pytorch pipeline in general works. When I define a new layer using torch.autograd.Function class with 4 inputs I get RuntimeError: leaf variable was used in an inplace operation error. When I define the same class but with only 3 inputs I get the correct output.
This is the layer I am defining in its simplest form:
class ExpRecR_(torch.autograd.Function):
@staticmethod
def forward(ctx, input, p1, p2, p3):
ctx.input = input
ctx.p1 = p1
ctx.p2 = p2
ctx.p3 = p3
ctx.mark_dirty(input)
ctx.mark_dirty(p1)
ctx.mark_dirty(p2)
ctx.mark_dirty(p3)
loss = 0
return torch.Tensor([loss])
@staticmethod
def backward(ctx, grad_output):
# saved tensors - tuple of tensors, so we need get first
input = ctx.input
p1 = ctx.p1
p2 = ctx.p2
p3 = ctx.p3
out = torch.zeros(1,1)
out_p2 = torch.zeros(1,1)
out_p3 = torch.zeros(1,1)
return None, Variable(out)*grad_output, Variable(out_p2)*grad_output, Variable(out_p3)*grad_output
and I check this layer using torch.autograd.gradcheck
test = gradcheck(ExpRecR_.apply, (Variable(timestamp_pt,),
nn.Parameter(-torch.ones(1, 1)*0.2),
nn.Parameter(torch.ones(1, 1)*0.2),
nn.Parameter(torch.ones(1, 1)*0.1)) , eps=1e-3, atol=1e-4)
this produces following error:
RuntimeError Traceback (most recent call last)
<ipython-input-144-b5c59cb4deb3> in <module>()
2 nn.Parameter(-torch.ones(1, 1)*0.2),
3 nn.Parameter(torch.ones(1, 1)*0.2),
----> 4 nn.Parameter(torch.ones(1, 1)*0.1)) , eps=1e-3, atol=1e-4)
~/anaconda/envs/py36/lib/python3.6/site-packages/torch/autograd/gradcheck.py in gradcheck(func, inputs, eps, atol, rtol, raise_exception)
174 return _as_tuple(func(*input))[i].data
175
--> 176 analytical, reentrant, correct_grad_sizes = get_analytical_jacobian(_as_tuple(inputs), o)
177 numerical = get_numerical_jacobian(fn, inputs, inputs, eps)
178
~/anaconda/envs/py36/lib/python3.6/site-packages/torch/autograd/gradcheck.py in get_analytical_jacobian(input, output)
108 for jacobian_c in (jacobian, jacobian_reentrant):
109 zero_gradients(input)
--> 110 output.backward(grad_output, create_graph=True)
111 for jacobian_x, (d_x, x) in zip(jacobian_c, iter_variables(input)):
112 if d_x is None:
~/anaconda/envs/py36/lib/python3.6/site-packages/torch/autograd/variable.py in backward(self, gradient, retain_graph, create_graph, retain_variables)
165 Variable.
166 """
--> 167 torch.autograd.backward(self, gradient, retain_graph, create_graph, retain_variables)
168
169 def register_hook(self, hook):
~/anaconda/envs/py36/lib/python3.6/site-packages/torch/autograd/__init__.py in backward(variables, grad_variables, retain_graph, create_graph, retain_variables)
97
98 Variable._execution_engine.run_backward(
---> 99 variables, grad_variables, retain_graph)
100
101
RuntimeError: leaf variable was used in an inplace operation
When I use the layer like this:
gradcheck(ExpRecR_.apply, (Variable(timestamp_pt,),
nn.Parameter(-torch.ones(1, 1)*0.2),
nn.Parameter(torch.ones(1, 1)*0.2),
Variable(torch.ones(1, 1)*0.1)) , eps=1e-3, atol=1e-4)
I get True, which is the correct output.
Can someone help me clarify why I get this error, is there a limit on how many inputs the layer can have?