chrishkchris edited a comment on pull request #733: URL: https://github.com/apache/singa/pull/733#issuecomment-643791102
I took a look at your cudnn rnn function: https://github.com/apache/singa/blob/dev/python/singa/layer.py#L1506 you transpose the input so that "inputs has shape of {sequence length, batch size, feature size}" Meanwhile, when I read the cudnn API `cudnnRNNForwardInference()` https://docs.nvidia.com/deeplearning/sdk/cudnn-api/index.html#cudnnRNNForwardInference the description of x is: _xDesc Input. An array of seqLength fully packed tensor descriptors. Each descriptor in the array should have three dimensions that describe the input data format to one recurrent iteration (one descriptor per RNN time-step). **The first dimension (batch size)** of the tensors may decrease from iteration n to iteration n+1 but may not increase. Each tensor descriptor must have the same second dimension (RNN input vector length, inputSize). The third dimension of each tensor should be 1. Input data are expected to be arranged in the column-major order so strides in xDesc should be set as follows:_ See the highlighted text in the above description. The first dimension is batch size? It is a bit confusing for me, so I am not sure what should be the input shape. I suggest checking all those input output format if you don't have further idea for debug. ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected]
