Fix shape of loaded preprocessed features

bin/run-tc-ldc93s1_new.sh

@@ -14,6 +14,7 @@ fi;
 
 python -u DeepSpeech.py \
   --train_files ${ldc93s1_csv} --train_batch_size 1 \
+  --train_cached_features_path "/tmp/ldc93s1.hdf5" \
   --dev_files ${ldc93s1_csv} --dev_batch_size 1 \
   --test_files ${ldc93s1_csv} --test_batch_size 1 \
   --n_hidden 494 --epoch $epoch_count --random_seed 4567 \

tc-train-tests.sh

@@ -56,6 +56,8 @@ decoder_pkg_url=${DECODER_ARTIFACTS_ROOT}/${decoder_pkg}
 LD_LIBRARY_PATH=${PY37_LDPATH}:$LD_LIBRARY_PATH pip install --verbose --only-binary :all: ${PY37_SOURCE_PACKAGE} --upgrade ${decoder_pkg_url} | cat
 
 pushd ${HOME}/DeepSpeech/ds/
+    # Run twice to test preprocessed features
+    time ./bin/run-tc-ldc93s1_new.sh 104
     time ./bin/run-tc-ldc93s1_new.sh 105
     time ./bin/run-tc-ldc93s1_tflite.sh
 popd

util/preprocess.py

@@ -41,10 +41,9 @@ def preprocess(csv_files, batch_size, numcep, numcontext, alphabet, hdf5_cache_p
             transcript = file.root.transcript[:]
             transcript_len = file.root.transcript_len[:]
 
-            # features are stored flattened, so reshape into
-            # [n_steps, (n_input + 2*n_context*n_input)]
+            # features are stored flattened, so reshape into [n_steps, numcep]
             for i in range(len(features)):
-                features[i] = np.reshape(features[i], [features_len[i], -1])
+                features[i].shape = [features_len[i]+2*numcontext, numcep]
 
             in_data = list(zip(features, features_len,
                                transcript, transcript_len))