lebeg commented on a change in pull request #12468: [MXNET-807] Support integer
label type in ctc_loss operator
URL: https://github.com/apache/incubator-mxnet/pull/12468#discussion_r215702787
##########
File path: src/operator/contrib/ctc_loss-inl.h
##########
@@ -256,66 +255,69 @@ class CTCLossOp : public Operator {
exceed_cudnn_limit = false;
Stream<xpu> *s = ctx.get_stream<xpu>();
- Tensor<xpu, 3, real_t> data =
+ MSHADOW_TYPE_SWITCH(in_data[ctc_loss::kLabel].type_flag_, DType, {
+ Tensor<xpu, 3, real_t> data =
in_data[ctc_loss::kData].get<xpu, 3, real_t>(s);
- Tensor<xpu, 2, real_t> labels =
- in_data[ctc_loss::kLabel].get<xpu, 2, real_t>(s);
+ Tensor<xpu, 2, DType> labels =
+ in_data[ctc_loss::kLabel].get<xpu, 2, DType>(s);
- Tensor<xpu, 1, real_t> costs =
+ Tensor<xpu, 1, real_t> costs =
out_data[ctc_loss::kOut].get<xpu, 1, real_t>(s);
- Tensor<xpu, 3, real_t> grad =
+ Tensor<xpu, 3, real_t> grad =
out_data[ctc_loss::kGrad].get<xpu, 3, real_t>(s);
- int max_seq_len = data.size(0);
- int batch_size = data.size(1);
- int alphabet_size = data.size(2);
-
- // data_lengths
- std::vector<int> data_lengths(batch_size, max_seq_len);
- if (param_.use_data_lengths) {
- int kInputLength = 2;
- IndexTensorToVector(in_data[kInputLength].get<xpu, 1, real_t>(s),
&data_lengths);
- }
-
- // label_lengths
- std::vector<int> packed_labels;
- std::vector<int> label_lengths(batch_size);
-
- if (param_.use_label_lengths) {
- int kLabelLength = 2+param_.use_data_lengths;
- exceed_cudnn_limit = PackLabelByLength(labels,
in_data[kLabelLength].get<xpu, 1, real_t>(s),
- &packed_labels, &label_lengths);
- } else {
- exceed_cudnn_limit = LabelTensorToPackedVector(labels,
param_.blank_label == 0?0:-1,
- &packed_labels,
&label_lengths);
- }
-
-// CUDNN is disabled due to lack of support for input lengths
-/* #if defined(__CUDACC__) && MXNET_USE_CUDNN == 1 && CUDNN_MAJOR >= 7 */
-/* if (!exceed_cudnn_limit) { */
-/* cudnn_forward(ctx, s, data, costs, grad, */
-/* &data_lengths, &label_lengths, &packed_labels, */
-/* max_seq_len, batch_size, alphabet_size, */
-/* req[ctc_loss::kGrad] != mxnet::kNullOp); */
-/* } else { */
-/* baidu_forward(ctx, s, data, costs, grad, */
-/* &data_lengths, &label_lengths, &packed_labels, */
-/* batch_size, alphabet_size, req[ctc_loss::kGrad] !=
mxnet::kNullOp); */
-/* } */
-/* #else */
-
- baidu_forward(ctx, s, data, costs, grad,
- &data_lengths, &label_lengths, &packed_labels,
- batch_size, alphabet_size, req[ctc_loss::kGrad] !=
mxnet::kNullOp);
-
- if (param_.use_data_lengths) {
- // baidu warp CTC implementation sometimes includes undefined gradients
- // for data outside of length mask. Setting to 0 to make it consistent
- // with CPU implementation.
- int kInputLength = 2;
- mxnet_op::SequenceMask(grad, in_data[kInputLength].get<xpu, 1,
real_t>(s),
- static_cast<real_t>(0));
- }
+ int max_seq_len = data.size(0);
+ int batch_size = data.size(1);
+ int alphabet_size = data.size(2);
+
+ // data_lengths
+ std::vector<int> data_lengths(batch_size, max_seq_len);
+ if (param_.use_data_lengths) {
+ int kInputLength = 2;
+ IndexTensorToVector(in_data[kInputLength].get<xpu, 1, real_t>(s),
&data_lengths);
+ }
+
+ // label_lengths
+ std::vector<int> packed_labels;
+ std::vector<int> label_lengths(batch_size);
+
+ if (param_.use_label_lengths) {
+ int kLabelLength = 2 + param_.use_data_lengths;
+ exceed_cudnn_limit =
+ PackLabelByLength(labels, in_data[kLabelLength].get<xpu, 1,
DType>(s),
+ &packed_labels, &label_lengths);
+ } else {
+ exceed_cudnn_limit = LabelTensorToPackedVector(labels,
param_.blank_label == 0?0:-1,
Review comment:
Sorry, should be `0 ? 0 : -1`
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
With regards,
Apache Git Services
