Adapt warpctc grad op for gradient checking #7414
Conversation
wanghaoshuang
commented
Jan 10, 2018
wanghaoshuang
commented
- Fix warpctc grad op
- Add check grad test
- Add sequence scale functor
2. Add check grad test
wanghaoshuang
changed the title
| public: | ||
| void operator()(const platform::CPUDeviceContext& context, | ||
| framework::LoDTensor& seq, const T* scales, | ||
| const size_t num_seq) { |
There was a problem hiding this comment.
It seems there is no need to pass num_seq, it can be got from framework::LoDTensor& seq.
| T* seq_data = seq.mutable_data<T>(context.GetPlace()); | ||
|
|
||
| int threads = 1024; | ||
| int grid = (seq.numel() * seq_width + threads - 1) / threads; |
There was a problem hiding this comment.
For the CUDA kernel, we can use one block to process one sample.
int threads = 1024;
int grid = abs_offset_lod[level].size();| } | ||
| }; | ||
|
|
||
| template class ScaleLoDTensorFunctor<platform::CUDADeviceContext, float>; |
There was a problem hiding this comment.
double type is also needed.
There was a problem hiding this comment.
Also, need to register double type for paddle/operators/warpctc_op.cc, please help to enhance it.
There was a problem hiding this comment.
It seems that API in warp-CTC doesn't support for double.
|
|
||
| def test_check_grad(self): | ||
| self.outputs['WarpCTCGrad'] = self.gradient | ||
| self.check_grad(["Logits"], "Loss", max_relative_error=0.01) |
There was a problem hiding this comment.
Can the gradient check pass when max_relative_error is lower?
There was a problem hiding this comment.
It will fail while max_relative_error = 0.005.
There was a problem hiding this comment.
Changed max_relative_error from 0.01 to 0.007.
2. Remove num_seq arguments. 3. Refine CUDA kernel of ScaleLoDTensorFunctor. 4. Change max_relative_error of gradient unitest to 0.007
1. Fix kernel 2. Add more test case
2 participants
