Added conda local install, partialy adjusted model to pycandle

install_dependencies.sh

@@ -12,12 +12,12 @@ case ${i} in
     CUDA="${i#*=}"
     shift # past argument=value
     ;;
-    -e=*|--venv=*)
-    VENV="${i#*=}"
+    -a=*|--anaconda=*)
+    ANACONDA="${i#*=}"
     shift # past argument=value
     ;;
-    -p=*|--python_path=*)
-    VENV_PATH="${i#*=}"
+    -e=*|--venv=*)
+    VENV="${i#*=}"
     shift # past argument=value
     ;;
     *)
@@ -26,11 +26,12 @@ case ${i} in
 esac
 done
 
-VENV=${VENV:-true}
+VENV=${VENV:-./venv}
 
-if [ "$VENV" = true ] ; then
-    VENV_PATH=${VENV_PATH:-./venv}
-    source ${VENV_PATH}/bin/activate
+if [ -z ${ANACONDA+x} ] ; then
+    conda activate ${ANACONDA}
+elif [ -z ${VENV+x} ] ; then
+    source ${VENV}/bin/activate
 fi
 
 #TODO: Infer this automatically
@@ -50,4 +51,9 @@ cd ../pytorch_binding && python setup.py install
 cd ../..
 rm -rf warp-ctc
 
+git clone [email protected]:pytorch/audio.git
+cd audio; MACOSX_DEPLOYMENT_TARGET=10.9 CC=clang CXX=clang++ python setup.py install
+cd ..
+rm -rf audio
+
 pip install -r post_requirements.txt

post_requirements.txt

@@ -1,2 +1 @@
--e git://github.com/pytorch/audio.git#egg=torchaudio-0.2
 -e git://github.com/NVIDIA/apex.git#egg=apex

sonosco/datasets/AudioDataLoader.py

@@ -29,6 +29,7 @@ def _collate_fn(self, batch):
             inputs[x][0].narrow(1, 0, batch[x][0].size(1)).copy_(batch[x][0])
             input_percentages[x] = batch[x][0].size(1) / float(max_seqlength)
             target_sizes[x] = len(batch[x][1])
-            targets.extend(batch[x][1])
+            targets.append([batch[x][1]])
 
-        return inputs, torch.IntTensor(targets), input_percentages, torch.from_numpy(target_sizes)
+        # return inputs, torch.IntTensor(targets), input_percentages, torch.from_numpy(target_sizes)
+        return (inputs, input_percentages), torch.IntTensor(targets)

sonosco/models/deepspeech2.py

@@ -82,13 +82,14 @@ def forward(self, input_):
 
 
 class BatchRNN(nn.Module):
-    def __init__(self, input_size, hidden_size, rnn_type=nn.LSTM, batch_norm=True):
+    def __init__(self, input_size, hidden_size, rnn_type=nn.LSTM, batch_norm=True, bidirectional=True):
         super(BatchRNN, self).__init__()
+        self.bidirectional = bidirectional
         self.input_size = input_size
         self.hidden_size = hidden_size
         self.batch_norm = SequenceWise(nn.BatchNorm1d(input_size)) if batch_norm else None
         self.rnn = rnn_type(input_size=input_size, hidden_size=hidden_size,
-                            bidirectional=True, bias=True)
+                            bidirectional=self.bidirectional, bias=True)
 
     def flatten_parameters(self):
         self.rnn.flatten_parameters()
@@ -155,9 +156,12 @@ def __init__(self, rnn_type=nn.LSTM, labels="abc", rnn_hid_size=768, nb_layers=5
 
         self.inference_softmax = InferenceBatchSoftmax()
 
-    def forward(self, x, lengths):
+    def forward(self, xx):
         # if x.is_cuda and self.mixed_precision:
         #     x = x.half()
+        x, input_percentages = xx
+
+        lengths = input_percentages.mul_(int(x.size(3))).int()
         lengths = lengths.cpu().int()
         output_lengths = self.get_seq_lens(lengths)
         x, _ = self.conv(x, output_lengths)
@@ -176,7 +180,7 @@ def forward(self, x, lengths):
         x = x.transpose(0, 1)
         # identity in training mode, softmax in eval mode
         x = self.inference_softmax(x)
-        return x, output_lengths
+        return x
 
     def get_seq_lens(self, input_length):
         """

sonosco/pycandle_train.py

@@ -29,5 +29,5 @@ def load_datasets(manifest_path, batch_size_train, batch_size_test):
 train_loader = load_datasets("./datasets/download_datasets/libri_test_clean_manifest.csv",
                                          batch_size_train=64, batch_size_test=64)
 optimizer = torch.optim.Adam(model.parameters(), lr=0.01)
-model_trainer = ModelTrainer(model, optimizer, F.nll_loss, 20, train_loader, gpu=0)
+model_trainer = ModelTrainer(model, optimizer, F.nll_loss, 20, train_loader, gpu=None)
 model_trainer.start_training()