Some mods in docstrings and arguments

bob/learn/pytorch/config/mccnn/wmca_mccnn.py

@@ -58,7 +58,7 @@ bob_hldi_instance_train = BatlPadDatabase(
 #==============================================================================
 # Initialize the torch dataset, subselect channels from the pretrained files if needed.
 
-SELECTED_CHANNELS = [0,1,2] # selects only color, depth and infrared
+SELECTED_CHANNELS = [0,1,2,3] # selects  color, depth, infrared and thermal
 
 img_transform_train = transforms.Compose([ChannelSelect(selected_channels = SELECTED_CHANNELS),transforms.ToPILImage(),transforms.RandomHorizontalFlip(),transforms.ToTensor()])# Add p=0.5 later
 
@@ -74,7 +74,9 @@ dataset = DataFolder(data_folder=data_folder_train,
 #==============================================================================
 # Load the architecture
 
-NUM_CHANNELS = 3
+NUM_CHANNELS = 4
+
+assert(len(SELECTED_CHANNELS)==NUM_CHANNELS)
 
 network=MCCNN(num_channels = NUM_CHANNELS)
 
@@ -87,7 +89,7 @@ learning_rate=0.0001
 seed = 3
 output_dir = 'training_mccn'
 use_gpu = False
-adapted_layers = 'conv1-ffc'
+adapted_layers = 'conv1-group1-block1-ffc'
 adapt_reference_channel = False
 verbose = 2
 

bob/learn/pytorch/extractor/image/MCCNN.py

@@ -23,13 +23,16 @@ class MCCNNExtractor(Extractor):
 
   """
   
-  def __init__(self, num_channels=4, transforms = transforms.Compose([transforms.ToTensor()]), model_file=None):
+  def __init__(self, num_channels_used=4, transforms = transforms.Compose([transforms.ToTensor()]), model_file=None):
     """ Init method
 
     Parameters
     ----------
-    num_channels: int
-      The number of channels present in the input
+    num_channels_used: int
+      The number of channels to be used by the network. This could be 
+      different from the number of channels present in the input image. For instance, 
+      when used together with 'ChannelSelect' transform. The value of `num_channels_used`
+      should be the number of channels eventually used by the network (i.e., output of transform).
     model_file: str
       The path of the trained PAD network to load
     transforms: :py:mod:`torchvision.transforms` 
@@ -41,22 +44,26 @@ class MCCNNExtractor(Extractor):
     
     # model
     self.transforms = transforms 
-    self.network = MCCNN(num_channels=num_channels)
+    self.network = MCCNN(num_channels=num_channels_used)
+
+    logger.debug('Initiliazed model with lightCNN weights')
     
     #self.network=self.network.to(device)
 
     if model_file is None:
       # do nothing (used mainly for unit testing) 
-      logger.info("No pretrained file provided")
+      logger.debug("No pretrained file provided")
       pass
     else:
 
+
       # With the new training
+      logger.debug('Starting to load the pretrained PAD model')
       cp = torch.load(model_file)
       if 'state_dict' in cp:
         self.network.load_state_dict(cp['state_dict'])
 
-      logger.info('Loaded the pretrained PAD model')    
+      logger.debug('Loaded the pretrained PAD model')    
  
     self.network.eval()
 
@@ -69,6 +76,9 @@ class MCCNNExtractor(Extractor):
     ----------
     image : 3D :py:class:`numpy.ndarray` (floats)
       The multi-channel image to extract the score from. Its size must be num_channelsx128x128;
+      Note: the value of `num_channels` is the number of channels as obtained from the preprocessed
+      data. The actual number of channels used may vary, for instance
+      if ChannelSelect transform is used, the number of channels used would change.
       
     Returns
     -------

bob/learn/pytorch/test/test.py

@@ -323,7 +323,7 @@ def test_extractors():
 
   # MCCNN 
   from ..extractor.image import MCCNNExtractor
-  extractor = MCCNNExtractor(num_channels=4)
+  extractor = MCCNNExtractor(num_channels_used=4)
   # this architecture expects num_channelsx128x128 Multi channel images
   data = numpy.random.rand(4, 128, 128).astype("float32")
   output = extractor(data)