Add more support for mxnet_to_coreml (#14222)

* Add support of mxnet_to_coreml

 * Add PReLU
 * Add group support in Convolution
 * Skip _minus_scalar and _mul_scalar

* Comply with code style

* Add spare line between comment and code
* Add two spare lines between functions
* Change "Convert an leakyrelu" to "Convert a leakyrelu"

* Check if num_group exists in param before assigning value

* Add test cases for test_mxnet_converter

* Add test_tiny_prelu_leakyrelu_random_input
* Add test_really_tiny_conv_random_input_multi_group
* Add test_tiny_conv_random_input_multi_group
* Modify test_conv_random by adding num_group

* Rollback to not setting num_group

* Add "leaky" and "elu" in converter

* Add support of "leaky" and "elu"
* Add unit test for "leaky" and "elu"

tools/coreml/converter/_layers.py

@@ -220,6 +220,49 @@ def convert_activation(net, node, module, builder):
                            output_name = output_name)
 
 
+def convert_leakyrelu(net, node, module, builder):
+    """Convert a leakyrelu layer from mxnet to coreml.
+
+    Parameters
+    ----------
+    network: net
+        A mxnet network object.
+
+    layer: node
+        Node to convert.
+
+    module: module
+        An module for MXNet
+
+    builder: NeuralNetworkBuilder
+        A neural network builder object.
+    """
+
+    input_name, output_name = _get_input_output_name(net, node)
+    name = node['name']
+    inputs = node['inputs']
+    args, _ = module.get_params()
+    mx_non_linearity = _get_attrs(node)['act_type']
+    if mx_non_linearity == 'elu':
+        non_linearity = 'ELU'
+        slope = _get_attrs(node)['slope'] if 'slope' in _get_attrs(node) else 0.25
+        params = slope
+    elif mx_non_linearity == 'leaky':
+        non_linearity = 'LEAKYRELU'
+        slope = _get_attrs(node)['slope'] if 'slope' in _get_attrs(node) else 0.25
+        params = [slope]
+    elif mx_non_linearity == 'prelu':
+        non_linearity = 'PRELU'
+        params = args[_get_node_name(net, inputs[1][0])].asnumpy()
+    else:
+        raise TypeError('Unknown activation type %s' % mx_non_linearity)
+    builder.add_activation(name = name,
+                           non_linearity = non_linearity,
+                           input_name = input_name,
+                           output_name = output_name,
+                           params = params)
+
+
 def convert_elementwise_add(net, node, module, builder):
     """Convert an elementwise add layer from mxnet to coreml.
 
@@ -335,6 +378,7 @@ def convert_convolution(net, node, module, builder):
     border_mode = "valid"
 
     n_filters = int(param['num_filter'])
+    n_groups = int(param['num_group']) if 'num_group' in param else 1
 
     W = args[_get_node_name(net, inputs[1][0])].asnumpy()
     if has_bias:
@@ -361,7 +405,7 @@ def convert_convolution(net, node, module, builder):
         stride_height=stride_height,
         stride_width=stride_width,
         border_mode=border_mode,
-        groups=1,
+        groups=n_groups,
         W=W,
         b=Wb,
         has_bias=has_bias,

tools/coreml/converter/_mxnet_converter.py

@@ -38,11 +38,14 @@
     'elemwise_add'   : _layers.convert_elementwise_add,
     'Reshape'        : _layers.convert_reshape,
     'Deconvolution'  : _layers.convert_deconvolution,
+    'LeakyReLU'      : _layers.convert_leakyrelu,
 }
 
 _MXNET_SKIP_LAYERS = [
     '_MulScalar',
     'Dropout',
+    '_minus_scalar',
+    '_mul_scalar',
 ]
 
 def _mxnet_remove_batch(input_data):

tools/coreml/test/test_mxnet_converter.py

@@ -192,6 +192,33 @@ def test_tiny_tanh_activation_random_input(self):
         net = mx.sym.Activation(net, name='tanh1', act_type="tanh")
         self._test_mxnet_model(net, input_shape=input_shape, mode='random')
 
+    def test_tiny_elu_leakyrelu_random_input(self):
+        np.random.seed(1988)
+        input_shape = (1, 10)
+        net = mx.sym.Variable('data')
+        net = mx.sym.FullyConnected(data=net, name='fc1', num_hidden=5)
+        slope = 0.1
+        net = mx.sym.LeakyReLU(net, name='elu1', act_type="elu", slope=slope)
+        self._test_mxnet_model(net, input_shape=input_shape, mode='random')
+
+    def test_tiny_leaky_leakyrelu_random_input(self):
+        np.random.seed(1988)
+        input_shape = (1, 10)
+        net = mx.sym.Variable('data')
+        net = mx.sym.FullyConnected(data=net, name='fc1', num_hidden=5)
+        slope = 0.1
+        net = mx.sym.LeakyReLU(net, name='leaky1', act_type="leaky", slope=slope)
+        self._test_mxnet_model(net, input_shape=input_shape, mode='random')
+
+    def test_tiny_prelu_leakyrelu_random_input(self):
+        np.random.seed(1988)
+        input_shape = (1, 10)
+        net = mx.sym.Variable('data')
+        net = mx.sym.FullyConnected(data=net, name='fc1', num_hidden=5)
+        gamma = mx.sym.Variable('gamma')
+        net = mx.sym.LeakyReLU(net, gamma=gamma, name='prelu1', act_type="prelu")
+        self._test_mxnet_model(net, input_shape=input_shape, mode='random')
+
     def test_really_tiny_conv_random_input(self):
         np.random.seed(1988)
         input_shape = (1, 1, 10, 10)
@@ -408,6 +435,26 @@ def test_really_tiny_conv_random_input_multi_filter(self):
         )
         self._test_mxnet_model(net, input_shape=input_shape, mode='random')
 
+    def test_really_tiny_conv_random_input_multi_group(self):
+        np.random.seed(1988)
+        input_shape = (1, 16, 10, 10)
+        num_filter = 16
+        num_group = 4
+        kernel = (1, 1)
+        stride = (1, 1)
+        pad = (0, 0)
+        net = mx.sym.Variable('data')
+        net = mx.symbol.Convolution(
+            data=net,
+            num_filter=num_filter,
+            num_group=num_group,
+            kernel=kernel,
+            stride=stride,
+            pad=pad,
+            name='conv_1'
+        )
+        self._test_mxnet_model(net, input_shape=input_shape, mode='random')
+
     def test_tiny_conv_random_3d_input(self):
         np.random.seed(1988)
         input_shape = (1, 3, 10, 10)
@@ -444,6 +491,26 @@ def test_tiny_conv_random_input_multi_filter(self):
         )
         self._test_mxnet_model(net, input_shape=input_shape, mode='random')
 
+    def test_tiny_conv_random_input_multi_group(self):
+        np.random.seed(1988)
+        input_shape = (1, 16, 10, 10)
+        num_filter = 16
+        num_group = 4
+        kernel = (5, 5)
+        stride = (1, 1)
+        pad = (0, 0)
+        net = mx.sym.Variable('data')
+        net = mx.symbol.Convolution(
+            data=net,
+            num_filter=num_filter,
+            num_group=num_group,
+            kernel=kernel,
+            stride=stride,
+            pad=pad,
+            name='conv_1'
+        )
+        self._test_mxnet_model(net, input_shape=input_shape, mode='random')
+
     def test_conv_random(self):
         np.random.seed(1988)
         input_shape = (1, 3, 10, 10)