Crosstalk Caffe Bug Fixes

Resolved bugs due to:
1. Scaling in batch normalization was done twice
2. Splicing did not take into account a default axis
3. Did not support newer protobuf versions.
4. Stdout was not being ignored during validation

Source/CNTKv2LibraryDll/tensorboard/TensorBoardFileWriter.cpp

@@ -228,7 +228,7 @@ namespace CNTK
                     break;
 
                 default:
-                    fprintf(stderr, "TensorBoardFileWriter: Unsupported data type: %d ", dtype);
+                    fprintf(stderr, "TensorBoardFileWriter: Unsupported data type: %d ", static_cast<int>(dtype));
                     break;
                 }
 

bindings/python/cntk/contrib/crosstalkcaffe/adapter/bvlccaffe/caffeadapter.py

@@ -214,13 +214,16 @@ class SetupCaffeParameters(object):
         '''
         cntk_layer_def.parameters = cntkmodel.CntkSpliceParameters()
 
-        cntk_layer_def.parameters.axis = caffe_parameters.axis
+        if caffe_parameters is not None:
+            cntk_layer_def.parameters.axis = caffe_parameters.axis
+        else:
+            cntk_layer_def.parameters.axis = 1
         output_tensor = inputs_info[0].tensor[:]
         output_tensor[0] = 0
         for input_info in inputs_info:
             if not output_tensor[1:] == input_info.tensor[1:]:
                 raise IndexError('Non-align tensor information\n')
-            output_tensor[0] += input_info.tensor[caffe_parameters.axis - 1]
+            output_tensor[0] += input_info.tensor[cntk_layer_def.parameters.axis - 1]
         cntk_layer_def.tensor = output_tensor
 
     @staticmethod
@@ -541,9 +544,9 @@ class CaffeAdapter(baseadapter.Adapter):
     @staticmethod
     def _get_layer_parameters(raw_layer):
         convert_name = raw_layer.type.lower() + 'param'
-        for term in dir(raw_layer):
-            if term.lower().replace('_', '') == convert_name:
-                return getattr(raw_layer, term)
+        for (descriptor, attr) in raw_layer.ListFields():
+            if descriptor.name.lower().replace('_','') == convert_name:
+                return attr
         return None
 
     def _try_special_case_wrapper(self, raw_layer):

bindings/python/cntk/contrib/crosstalkcaffe/unimodel/cntkinstance.py

@@ -203,9 +203,9 @@ class ApiSetup(object):
             mean_tensor = cntk_layer.parameter_tensor[0]
             variance_tensor = cntk_layer.parameter_tensor[1]
             global_scale = cntk_layer.parameter_tensor[2].data[0]
-            scale_init = 1 / global_scale if global_scale != 0 else 0
-            mean_init = np.asarray(mean_tensor.data, dtype=np.float32) * scale_init
-            var_init = np.asarray(variance_tensor.data, dtype=np.float32) * scale_init
+            moving_average_factor = 1 / global_scale if global_scale != 0 else 0
+            mean_init = np.asarray(mean_tensor.data, dtype=np.float32) * moving_average_factor
+            var_init = np.asarray(variance_tensor.data, dtype=np.float32) * moving_average_factor
             if len(cntk_layer.parameter_tensor) == 5:
                 scale_tensor = cntk_layer.parameter_tensor[3]
                 bias_tensor = cntk_layer.parameter_tensor[4]

bindings/python/cntk/contrib/crosstalkcaffe/validation/validcaffe.py

@@ -47,9 +47,9 @@ class CaffeValidCore(ValidCore):
         caffe_solver = CaffeResolver()
         caffe = caffe_solver.caffe
         if not caffe_solver.runtime():
-            sys.stdout.write('No caffe runtime support, ignore validation...\n')
+            sys.__stdout__.write('No caffe runtime support, ignore validation...\n')
             return
-        sys.stdout.write('Start valid feature map...\n')
+        sys.__stdout__.write('Start valid feature map...\n')
         caffe.set_mode_gpu()
         caffe.set_device(0)
         net = caffe.Net(source_solver.model_path, source_solver.weights_path, caffe.TEST)
@@ -66,7 +66,7 @@ class CaffeValidCore(ValidCore):
             test_result = net.blobs[target].data
             power_error = np.power(gt_result.flatten() - test_result.flatten(), 2).sum()
             rsme_diff = np.sqrt(power_error / gt_result.size)
-            sys.stdout.write('Validating %s with RMSE = %s, MAX = %s, MIN = %s\n' %
-                             (target, str(rsme_diff), str(gt_result.max()), str(gt_result.min())))
-        sys.stdout.write('Validation finished...\n')
-        sys.stdout.flush()
+            sys.__stdout__.write(('Validating %s with RMSE = %s, MAX = %s, MIN = %s\n' %
+                             (target, str(rsme_diff), str(gt_result.max()), str(gt_result.min()))))
+        sys.__stdout__.write('Validation finished...\n')
+        sys.__stdout__.flush()