fix u2 nnet out frames num

speechx/.gitignore

@@ -1 +1,2 @@
 tools/valgrind*
+*log

speechx/speechx/nnet/nnet_itf.h

@@ -21,15 +21,15 @@
 
 namespace ppspeech {
 
-struct NnetOut{
-  // nnet out, maybe logprob or prob
-  kaldi::Vector<kaldi::BaseFloat> logprobs;
-  int32 vocab_dim;
+struct NnetOut {
+    // nnet out, maybe logprob or prob
+    kaldi::Vector<kaldi::BaseFloat> logprobs;
+    int32 vocab_dim;
 
-  // nnet state. Only using in Attention model.
-  std::vector<std::vector<kaldi::BaseFloat>> encoder_outs;
+    // nnet state. Only using in Attention model.
+    std::vector<std::vector<kaldi::BaseFloat>> encoder_outs;
 
-  NnetOut() : logprobs({}), vocab_dim(-1), encoder_outs({}) {} 
+    NnetOut() : logprobs({}), vocab_dim(-1), encoder_outs({}) {}
 };
 
 

speechx/speechx/nnet/u2_nnet.cc

@@ -313,10 +313,8 @@ void U2Nnet::ForwardEncoderChunkImpl(
     // call.
     std::vector<paddle::Tensor> inputs = {
         feats, offset, /*required_cache_size, */ att_cache_, cnn_cache_};
-    VLOG(3) << "inputs size: " << inputs.size();
     CHECK(inputs.size() == 4);
     std::vector<paddle::Tensor> outputs = forward_encoder_chunk_(inputs);
-    VLOG(3) << "outputs size: " << outputs.size();
     CHECK(outputs.size() == 3);
 
 #ifdef USE_GPU
@@ -351,10 +349,12 @@ void U2Nnet::ForwardEncoderChunkImpl(
     // current offset in decoder frame
     // not used in nnet
     offset_ += chunk_out.shape()[1];
+    VLOG(2) << "encoder out chunk size: " << chunk_out.shape()[1] << " total: " << offset_ ;
+    
 
     // collects encoder outs.
-    VLOG(2) << "encoder_outs_ size: " << encoder_outs_.size();
     encoder_outs_.push_back(chunk_out);
+    VLOG(2) << "encoder_outs_ size: " << encoder_outs_.size();
 
 #ifdef TEST_DEBUG
     {

speechx/speechx/nnet/u2_nnet.h

@@ -137,7 +137,9 @@ class U2Nnet : public U2NnetBase {
     // debug
     void FeedEncoderOuts(paddle::Tensor& encoder_out);
 
-    const std::vector<paddle::Tensor>& EncoderOuts() const {return encoder_outs_; }
+    const std::vector<paddle::Tensor>& EncoderOuts() const {
+        return encoder_outs_;
+    }
 
   private:
     U2ModelOptions opts_;

speechx/speechx/nnet/u2_nnet_main.cc

@@ -95,29 +95,29 @@ int main(int argc, char* argv[]) {
         //                    kaldi::kCopyData);
         // }
 
-        int32 num_chunks = (feature.NumRows() - chunk_size) / chunk_stride + 1;
         int32 frame_idx = 0;
         std::vector<kaldi::Vector<kaldi::BaseFloat>> prob_vec;
         int32 ori_feature_len = feature.NumRows();
+        int32 num_chunks = feature.NumRows() / chunk_stride + 1;
+        LOG(INFO) << "num_chunks: " << num_chunks;
 
         for (int chunk_idx = 0; chunk_idx < num_chunks; ++chunk_idx) {
-            kaldi::Vector<kaldi::BaseFloat> feature_chunk(chunk_size *
-                                                          feat_dim);
-
-            int32 feature_chunk_size = 0;
+            int32 this_chunk_size = 0;
             if (ori_feature_len > chunk_idx * chunk_stride) {
-                feature_chunk_size = std::min(
+                this_chunk_size = std::min(
                     ori_feature_len - chunk_idx * chunk_stride, chunk_size);
             }
-            if (feature_chunk_size < receptive_field_length) {
+            if (this_chunk_size < receptive_field_length) {
                 LOG(WARNING) << "utt: " << utt << " skip last "
-                             << feature_chunk_size << " frames, expect is "
+                             << this_chunk_size << " frames, expect is "
                              << receptive_field_length;
                 break;
             }
 
+            kaldi::Vector<kaldi::BaseFloat> feature_chunk(this_chunk_size *
+                                                          feat_dim);
             int32 start = chunk_idx * chunk_stride;
-            for (int row_id = 0; row_id < chunk_size; ++row_id) {
+            for (int row_id = 0; row_id < this_chunk_size; ++row_id) {
                 kaldi::SubVector<kaldi::BaseFloat> feat_row(feature, start);
                 kaldi::SubVector<kaldi::BaseFloat> feature_chunk_row(
                     feature_chunk.Data() + row_id * feat_dim, feat_dim);

speechx/tools/clang-format.sh

@@ -0,0 +1,3 @@
+#!/bin/bash
+
+find speechx -name '*.c' -o -name '*.h' -not -path "*kaldi*" | xargs -I{} clang-format -i  {}