Closed jcreinhold closed 5 years ago
At least when 1-cycle policy is enabled, dropout > 0 causes backprop to fail. It states that an inplace operation has changed a value needed for the gradient.
Set inplace=False in dropout blocks solved the issue
At least when 1-cycle policy is enabled, dropout > 0 causes backprop to fail. It states that an inplace operation has changed a value needed for the gradient.