[FRONTEND][TENSORFLOW] Some bug fixes for tensorflow NCHW data_format (#3514)

nnvm/python/nnvm/frontend/tensorflow.py

@@ -205,8 +205,12 @@ def _conv(opname):
         # NCHW Layout require weights transpose
         if attr['data_format'] == 'NCHW':
             tmp_shape = attr['_input_shapes'][inputs[1]]
+            if opname == 'conv':
                 tmp_shape = [tmp_shape[ii] for ii in (3, 2, 0, 1)]
                 inputs[1] = _sym.transpose(inputs[1], axes=(3, 2, 0, 1))
+            else:
+                tmp_shape = [tmp_shape[ii] for ii in (2, 3, 0, 1)]
+                inputs[1] = _sym.transpose(inputs[1], axes=(2, 3, 0, 1))
             attr['_input_shapes'][inputs[1]] = tmp_shape
 
         input_shape = attr['_input_shapes'][inputs[0]]
@@ -238,12 +242,12 @@ def _conv(opname):
                 attr['dilations'] = (attr['dilations'][1], attr['dilations'][2])
             attr['strides'] = (attr['strides'][1], attr['strides'][2])
         elif attr['data_format'] == 'NCHW':
-            depth_mult, _, kernel_h, kernel_w = weights_shape
+            _, depth_mult, kernel_h, kernel_w = weights_shape
             attr['kernel_shape'] = (weights_shape[2], weights_shape[3])
             if opname == 'conv':
                 attr['channels'] = weights_shape[0]
             else:
-                attr['channels'] = input_shape[0] * depth_mult
+                attr['channels'] = input_shape[1] * depth_mult
                 if attr['channels'] < 0:
                     attr['channels'] *= -1
 
@@ -256,6 +260,9 @@ def _conv(opname):
 
 
         if opname == 'depthwise':
+            if depth_mult > 1:
+                raise tvm.error.OpNotImplemented('depth_mult > 1 of operator DepthwiseConv2dNative'
+                                                 ' is not supported.')
             attr['groups'] = attr['channels']
 
         # Fix padding
@@ -459,7 +466,11 @@ def _reshape():
 
 def _bias_add():
     def _impl(inputs, attr, params):
-        return _sym.broadcast_add(inputs[0], inputs[1])
+        if attr['data_format'].decode("utf-8") == 'NCHW':
+            bias = _sym.reshape(inputs[1], newshape=(1, -1, 1, 1))
+        else:
+            bias = inputs[1]
+        return _sym.broadcast_add(inputs[0], bias)
     return _impl
 
 def _squeeze():

python/tvm/relay/frontend/tensorflow.py

@@ -361,8 +361,12 @@ def _conv(opname):
         # NCHW Layout require weights transpose
         if attr['data_format'] == 'NCHW':
             tmp_shape = attr['_input_shapes'][inputs[1]]
+            if opname == 'conv':
                 tmp_shape = [tmp_shape[ii] for ii in (3, 2, 0, 1)]
                 inputs[1] = _op.transpose(inputs[1], axes=(3, 2, 0, 1))
+            else:
+                tmp_shape = [tmp_shape[ii] for ii in (2, 3, 0, 1)]
+                inputs[1] = _op.transpose(inputs[1], axes=(2, 3, 0, 1))
             attr['_input_shapes'][inputs[1]] = tmp_shape
 
         input_shape = attr['_input_shapes'][inputs[0]]
@@ -394,12 +398,12 @@ def _conv(opname):
                 attr['dilations'] = (attr['dilations'][1], attr['dilations'][2])
             attr['strides'] = (attr['strides'][1], attr['strides'][2])
         elif attr['data_format'] == 'NCHW':
-            depth_mult, _, kernel_h, kernel_w = weights_shape
+            _, depth_mult, kernel_h, kernel_w = weights_shape
             attr['kernel_shape'] = (weights_shape[2], weights_shape[3])
             if opname == 'conv':
                 attr['channels'] = weights_shape[0]
             else:
-                attr['channels'] = input_shape[0] * depth_mult
+                attr['channels'] = input_shape[1] * depth_mult
                 if attr['channels'] < 0:
                     attr['channels'] *= -1
 
@@ -411,8 +415,10 @@ def _conv(opname):
                   'not valid.'
             raise tvm.error.OpAttributeInvalid(msg.format(attr['data_format']))
 
-
         if opname == 'depthwise':
+            if depth_mult > 1:
+                raise tvm.error.OpNotImplemented('depth_mult > 1 of operator DepthwiseConv2dNative'
+                                                 ' is not supported.')
             attr['groups'] = attr['channels']
 
         # Fix padding

tests/python/frontend/tensorflow/test_forward.py

@@ -223,7 +223,7 @@ def test_forward_pooling():
 # Convolution
 # -----------
 
-def _test_convolution(tensor_in_sizes, filter_in_sizes,
+def _test_convolution(opname, tensor_in_sizes, filter_in_sizes,
                       dilations, strides, padding, data_format):
     """ One iteration of convolution with given shapes and attributes """
 
@@ -244,6 +244,7 @@ def _test_convolution(tensor_in_sizes, filter_in_sizes,
             strides = [1, 1] + strides
             dilations = [1, 1] + dilations
 
+        if opname == 'conv':
             nn_ops.conv2d(in_data,
                           in_filter,
                           strides=strides,
@@ -253,18 +254,74 @@ def _test_convolution(tensor_in_sizes, filter_in_sizes,
 
             compare_tf_with_tvm(np.reshape(data_array, tensor_in_sizes).astype('float32'),
                                 'Placeholder:0', 'Conv2D:0')
+        else:
+            nn_ops.depthwise_conv2d_native(in_data,
+                          in_filter,
+                          strides=strides,
+                          dilations=dilations,
+                          padding=padding,
+                          data_format=data_format)
+
+            compare_tf_with_tvm(np.reshape(data_array, tensor_in_sizes).astype('float32'),
+                                'Placeholder:0', 'DepthwiseConv2dNative:0')
 
 def test_forward_convolution():
     if is_gpu_available():
-        _test_convolution([4, 176, 8, 8], [1, 1, 176, 32], [1, 1], [1, 1], 'SAME', 'NCHW')
-        _test_convolution([4, 19, 17, 17], [3, 3, 19, 19], [1, 1], [2, 2], 'VALID', 'NCHW')
-        _test_convolution([4, 124, 17, 17], [1, 1, 124, 19], [1, 1], [1, 1], 'SAME', 'NCHW')
-        _test_convolution([4, 12, 17, 17], [3, 3, 12, 32], [1, 1], [2, 2], 'VALID', 'NCHW')
-
-    _test_convolution([4, 8, 8, 176], [1, 1, 176, 32], [1, 1], [1, 1], 'SAME', 'NHWC')
-    _test_convolution([4, 17, 17, 19], [3, 3, 19, 19], [1, 1], [2, 2], 'VALID', 'NHWC')
-    _test_convolution([4, 17, 17, 124], [1, 1, 124, 19], [1, 1], [1, 1], 'SAME', 'NHWC')
-    _test_convolution([4, 17, 17, 12], [3, 3, 12, 32], [1, 1], [2, 2], 'VALID', 'NHWC')
+        _test_convolution('conv', [4, 176, 8, 8], [1, 1, 176, 32], [1, 1], [1, 1], 'SAME', 'NCHW')
+        _test_convolution('conv', [4, 19, 17, 17], [3, 3, 19, 19], [1, 1], [2, 2], 'VALID', 'NCHW')
+        _test_convolution('conv', [4, 124, 17, 17], [1, 1, 124, 19], [1, 1], [1, 1], 'SAME', 'NCHW')
+        _test_convolution('conv', [4, 12, 17, 17], [3, 3, 12, 32], [1, 1], [2, 2], 'VALID', 'NCHW')
+        _test_convolution('depthwise', [4, 176, 8, 8], [1, 1, 176, 1], [1, 1], [1, 1], 'SAME', 'NCHW')
+        _test_convolution('depthwise', [4, 19, 17, 17], [3, 3, 19, 1], [1, 1], [2, 2], 'VALID', 'NCHW')
+        _test_convolution('depthwise', [4, 124, 17, 17], [1, 1, 124, 1], [1, 1], [1, 1], 'SAME', 'NCHW')
+        _test_convolution('depthwise', [4, 12, 17, 17], [3, 3, 12, 1], [1, 1], [2, 2], 'VALID', 'NCHW')
+
+    _test_convolution('conv', [4, 8, 8, 176], [1, 1, 176, 32], [1, 1], [1, 1], 'SAME', 'NHWC')
+    _test_convolution('conv', [4, 17, 17, 19], [3, 3, 19, 19], [1, 1], [2, 2], 'VALID', 'NHWC')
+    _test_convolution('conv', [4, 17, 17, 124], [1, 1, 124, 19], [1, 1], [1, 1], 'SAME', 'NHWC')
+    _test_convolution('conv', [4, 17, 17, 12], [3, 3, 12, 32], [1, 1], [2, 2], 'VALID', 'NHWC')
+    _test_convolution('depthwise', [4, 8, 8, 176], [1, 1, 176, 1], [1, 1], [1, 1], 'SAME', 'NHWC')
+    _test_convolution('depthwise', [4, 17, 17, 19], [3, 3, 19, 1], [1, 1], [2, 2], 'VALID', 'NHWC')
+    _test_convolution('depthwise', [4, 17, 17, 124], [1, 1, 124, 1], [1, 1], [1, 1], 'SAME', 'NHWC')
+    _test_convolution('depthwise', [4, 17, 17, 12], [3, 3, 12, 1], [1, 1], [2, 2], 'VALID', 'NHWC')
+
+#######################################################################
+# BiasAdd
+# -----------
+def _test_biasadd(tensor_in_sizes, data_format):
+    """ One iteration of biasadd with given shapes and attributes """
+
+    total_size_1 = 1
+    for s in tensor_in_sizes:
+        total_size_1 *= s
+    tensor_bias_sizes = [tensor_in_sizes[1]] if data_format == 'NCHW' else [tensor_in_sizes[3]]
+    total_size_2 = tensor_bias_sizes[0]
+    # Initializes the input tensor with array containing incrementing
+    # numbers from 1.
+    data_array = [f * 1.0 for f in range(1, total_size_1 + 1)]
+    bias_array = [f * 1.0 for f in range(1, total_size_2 + 1)]
+
+    with tf.Graph().as_default():
+        in_data = array_ops.placeholder(shape=tensor_in_sizes, dtype='float32')
+        in_bias = constant_op.constant(bias_array, shape=tensor_bias_sizes, dtype='float32')
+        nn_ops.bias_add(in_data,
+                        in_bias,
+                        data_format=data_format)
+
+        compare_tf_with_tvm(np.reshape(data_array, tensor_in_sizes).astype('float32'),
+                            'Placeholder:0', 'BiasAdd:0')
+
+def test_forward_biasadd():
+    if is_gpu_available():
+        _test_biasadd([4, 176, 8, 8], 'NCHW')
+        _test_biasadd([1, 100, 1, 1], 'NCHW')
+        _test_biasadd([4, 19, 17, 17], 'NCHW')
+        _test_biasadd([4, 124, 3, 3], 'NCHW')
+
+    _test_biasadd([4, 8, 8, 176], 'NHWC')
+    _test_biasadd([1, 1, 1, 100], 'NHWC')
+    _test_biasadd([4, 17, 17, 19], 'NHWC')
+    _test_biasadd([4, 3, 3, 124], 'NHWC')
 
 #######################################################################
 # SpaceToBatchND