Merge branch 'inference' of github.com:TimoStoff/event_cnn_minimal in…

…to inference

data_loader/dataset.py

@@ -154,6 +154,10 @@ def __getitem__(self, i, seed=None):
             events_end_idx = np.searchsorted(self.np_ts,
                                              self.voxel_method['t'] + i * self.voxel_method['sliding_window_t'] + self.t0)
 
+            if events_end_idx == events_start_idx:
+                events_end_idx = events_start_idx + 1
+                print('WARNING! Set events_end_idx to events_start_idx + 1 at i={}'.format(i))
+
             timestamp = self.h5_file['events/ts'][events_start_idx]
         else:
             raise Exception("Unsupported voxel_method")

model/unet.py

@@ -62,6 +62,7 @@ def build_prediction_layer(self, num_output_channels, norm=None):
         return ConvLayer(self.base_num_channels if self.skip_type == 'sum' else 2 * self.base_num_channels,
                          num_output_channels, 1, activation=None, norm=norm)
 
+
 class WNet(BaseUNet):
     """
     Recurrent UNet architecture where every encoder is followed by a recurrent convolutional block,
@@ -94,7 +95,6 @@ def __init__(self, unet_kwargs):
     def forward(self, x):
         """
         :param x: N x num_input_channels x H x W
-        :param prev_states: previous LSTM states for every encoder layer
         :return: N x num_output_channels x H x W
         """
 
@@ -159,7 +159,6 @@ def __init__(self, unet_kwargs):
     def forward(self, x):
         """
         :param x: N x num_input_channels x H x W
-        :param prev_states: previous LSTM states for every encoder layer
         :return: N x num_output_channels x H x W
         """
 
@@ -277,7 +276,6 @@ def __init__(self, unet_kwargs):
     def forward(self, x):
         """
         :param x: N x num_input_channels x H x W
-        :param prev_states: previous LSTM states for every encoder layer
         :return: N x num_output_channels x H x W
         """
 
@@ -326,11 +324,9 @@ def __init__(self, unet_kwargs):
         self.decoders = self.build_decoders()
         self.pred = ConvLayer(self.base_num_channels, self.num_output_channels, kernel_size=1, activation=None)
 
-
     def forward(self, x):
         """
         :param x: N x num_input_channels x H x W
-        :param prev_states: previous LSTM states for every encoder layer
         :return: N x num_output_channels x H x W
         """