Add cuda detection and fix problem due to PyTorch Bug #43227

config.json

@@ -1,8 +1,9 @@
 {
     "name"          :   "Urban Testing",
+    "csvPath"       : "metdata/UrbanSound8K.csv",
     "data"          :   {
                             "type"      :   "CSVDataManager",
-                            "path"      :   "/home/kiran/Documents/DATA/UrbanSound8K",
+                            "path"      :   "/data/UrbanSound8K",
                             "format"    :   "audio",
                             "loader"    :   {
                                                 "shuffle"       : true,

data/data_manager.py

@@ -126,7 +126,10 @@ def __init__(self, config):
 
         self.load_func = load_formats[config['format']]
 
-        mfile = os.path.join(self.dir_path, 'metadata/UrbanSound8K.csv') 
+        csvPath = 'metadata/UrbanSound8K.csv' if not config['csvPath'] else config['csvPath']
+        print(f'Loading csv configuration from {csvPath}..')
+
+        mfile = os.path.join(self.dir_path, csvPath)
         metadata_df = pd.read_csv(mfile).sample(frac=1)
         self.metadata_df = self._remove_too_small(metadata_df, 1)
 

net/model.py

@@ -61,7 +61,7 @@ def forward(self, batch):
         # xt -> (batch, time, channel*freq)
         batch, time = x.size()[:2]
         x = x.reshape(batch, time, -1)
-        x_pack = torch.nn.utils.rnn.pack_padded_sequence(x, lengths, batch_first=True)
+        x_pack = torch.nn.utils.rnn.pack_padded_sequence(x, lengths.cpu(), batch_first=True)
 
         # x -> (batch, time, lstm_out)
         x_pack, hidden = self.net['recur'](x_pack)
@@ -123,7 +123,7 @@ def forward(self, batch):
         # x -> (batch, time, channel*freq)
         batch, time = x.size()[:2]
         x = x.reshape(batch, time, -1)
-        x_pack = torch.nn.utils.rnn.pack_padded_sequence(x, lengths, batch_first=True)
+        x_pack = torch.nn.utils.rnn.pack_padded_sequence(x, lengths.cpu(), batch_first=True)
 
         # x -> (batch, time, lstm_out)
         x_pack, hidden = self.net['recur'](x_pack)

run.py

@@ -11,6 +11,15 @@
 
 from eval import ClassificationEvaluator, AudioInference
 
+def p(msg): print(msg)
+
+def isCuda():
+  import torch
+  device = torch.device("cuda")
+  print(f'typeof(device)={type(device)} device={device}')
+  return str(device)=='cuda'
+
+print(f"IsCuda? {isCuda()}")
 
 def _get_transform(config, name):
     tsf_name = config['transforms']['type']
@@ -52,6 +61,7 @@ def eval_main(checkpoint):
 
 def infer_main(file_path, config, checkpoint):
     # Fix bugs
+    p('infer_main')
     if checkpoint is None:
         model = getattr(net_module, config['model']['type'])()
     else:
@@ -84,6 +94,7 @@ def train_main(config, resume):
     m_name = config['model']['type']
     model = getattr(net_module, m_name)(classes, config=config)
     num_classes = len(classes)
+    print(f'num_classes = {num_classes}')
 
 
     loss = getattr(net_module, config['train']['loss'])
@@ -112,10 +123,11 @@ def train_main(config, resume):
                       lr_scheduler=lr_scheduler,
                       train_logger=train_logger)
 
+    p('calling train()')
     trainer.train()
     return trainer
     #duration = 1; freq = 440
-    #os.system('play --no-show-progress --null --channels 1 synth %s sine %f'%(duration, freq))
+      #os.system('play --no-show-progress --null --channels 1 synth %s sine %f'%(duration, freq))
 
 def _test_loader(config):
 

train/base_trainer.py

@@ -18,6 +18,7 @@ def __init__(self, model, loss, metrics, optimizer, resume, config, train_logger
         self.logger = logging.getLogger(self.__class__.__name__)
 
         self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+        print(f'BaseTrainer.device={self.device}')
 
         self.model = model.to(self.device)
 

train/trainer.py

@@ -57,10 +57,12 @@ def _train_epoch(self, epoch):
         self.writer.set_step(epoch) 
 
         _trange = tqdm(self.data_loader, leave=True, desc='')
+        print(f'tqdm _trange={_trange}')
 
         for batch_idx, batch in enumerate(_trange):
             batch = [b.to(self.device) for b in batch]
             data, target = batch[:-1], batch[-1]
+            print(f'len(data) = {len(data)}')
             data = data if len(data) > 1 else data[0] 
             #data, target = data.to(self.device), target.to(self.device)