support model Face_cyclegan

x2paddle/op_mapper/onnx_custom_layer/InstanceNormalization.py

+from .register import register
+from x2paddle.core.util import *
+
+
+def InstanceNormalization_shape(input_shape):
+    return input_shape
+
+
+def InstanceNormalization_layer(inputs, name=None):
+    # TODO(lvmengsi@baidu.com): Check the accuracy when using fluid.layers.layer_norm.
+    epsilon = 1e-5
+    mean = fluid.layers.reduce_mean(inputs, dim=[2, 3], keep_dim=True)
+    var = fluid.layers.reduce_mean(fluid.layers.square(inputs - mean),
+                                   dim=[2, 3],
+                                   keep_dim=True)
+    if name is not None:
+        scale_name = name + "_scale"
+        offset_name = name + "_offset"
+    scale_param = fluid.ParamAttr(name=scale_name,
+                                  initializer=fluid.initializer.Constant(1.0),
+                                  trainable=True)
+    offset_param = fluid.ParamAttr(name=offset_name,
+                                   initializer=fluid.initializer.Constant(0.0),
+                                   trainable=True)
+    scale = fluid.layers.create_parameter(attr=scale_param,
+                                          shape=inputs.shape[1:2],
+                                          dtype="float32")
+    offset = fluid.layers.create_parameter(attr=offset_param,
+                                           shape=inputs.shape[1:2],
+                                           dtype="float32")
+
+    tmp = fluid.layers.elementwise_mul(x=(inputs - mean), y=scale, axis=1)
+    tmp = tmp / fluid.layers.sqrt(var + epsilon)
+    tmp = fluid.layers.elementwise_add(tmp, offset, axis=1)
+    return tmp
+
+
+def InstanceNormalization_weights(name, data=None):
+    weights_name = [name + '_scale']
+    return weights_name
+
+
+register(kind='InstanceNormalization',
+         shape=InstanceNormalization_shape,
+         layer=InstanceNormalization_layer,
+         weights=InstanceNormalization_weights)

x2paddle/op_mapper/onnx_custom_layer/__init__.py

+from .register import get_registered_layers
+
+#custom layer import begins
+
+from . import InstanceNormalization
+#custom layer import ends
+
+custom_layers = get_registered_layers()
+
+
+def set_args(f, params):
+    """ set args for function 'f' using the parameters in node.layer.param
+    Args:
+        f (function): a python function object
+        params (object): a object contains attributes needed by f's arguments
+    Returns:
+        arg_names (list): a list of argument names
+        kwargs (dict): a dict contains needed arguments
+    """
+    argc = f.__code__.co_argcount
+    arg_list = f.__code__.co_varnames[0:argc]
+    kwargs = {}
+    for arg_name in arg_list:
+        if hasattr(params, arg_name) and params is not None:
+            kwargs[arg_name] = getattr(params, arg_name)
+    return arg_list, kwargs
+
+
+def has_layer(layer_type):
+    """ test whether this layer exists in custom layer
+    """
+    return layer_type in custom_layers
+
+
+def get_params(layer, layer_type):
+    import re
+    if layer_type.lower() == "deconvolution" or layer_type.lower(
+    ) == "convolutiondepthwise":
+        param_name = '_'.join(('convolution', 'param'))
+    elif layer_type.lower() == "normalize":
+        param_name = '_'.join(('norm', 'param'))
+    elif len(layer_type) - len(re.sub("[A-Z]", "", layer_type)) >= 2:
+        s = ''
+        tmp_name = ''
+        for i, ch in enumerate(layer_type):
+            if i == 0:
+                s += ch.lower()
+                continue
+            elif ch.isupper() and layer_type[i - 1].islower():
+                tmp_name += (s + '_')
+                s = ''
+            s += ch.lower()
+        tmp_name += s
+        param_name = '_'.join((tmp_name, 'param'))
+    else:
+        param_name = '_'.join((layer_type.lower(), 'param'))
+    return getattr(layer, param_name, None)
+
+
+def compute_output_shape(node):
+    """ compute the output shape of custom layer
+    """
+    layer_type = node.layer_type
+    assert layer_type in custom_layers, "layer[%s] not exist in custom layers" % (
+        layer_type)
+    shape_func = custom_layers[layer_type]['shape']
+    layer = node.layer
+    params = get_params(layer, layer_type)
+    arg_names, kwargs = set_args(shape_func, params)
+    input_shape = node.input_shape
+    return shape_func(input_shape, **kwargs)
+
+
+def make_custom_layer(node):
+    """ get the code which implement the custom layer function
+    """
+    layer_type = node.layer_type
+    assert layer_type in custom_layers, "layer[%s] not exist in custom layers" % (
+        layer_type)
+    layer_func = custom_layers[layer_type]['layer']
+    import inspect
+    return inspect.getsource(layer_func), layer_func
+
+
+def deal_weights(node, data=None):
+    """ deal the weights of the custom layer
+    """
+    layer_type = node.layer_type
+    weights_func = custom_layers[layer_type]['weights']
+    name = node.layer_name
+    return weights_func(name, data)

x2paddle/op_mapper/onnx_custom_layer/register.py

+""" this module provides 'register' for registering customized layers
+"""
+
+g_custom_layers = {}
+
+
+def register(kind, shape, layer, weights):
+    """ register a custom layer or a list of custom layers
+
+    Args:
+        @kind (str or list): type name of the layer
+        @shape (function): a function to generate the shape of layer's output
+        @layer (function): a function to generate the paddle code of layer
+        @weights (function): a function to deal with weights data
+
+    Returns:
+        None
+    """
+    assert type(shape).__name__ == 'function', 'shape should be a function'
+    assert type(layer).__name__ == 'function', 'layer should be a function'
+
+    if type(kind) is str:
+        kind = [kind]
+    else:
+        assert type(
+            kind
+        ) is list, 'invalid param "kind" for register, not a list or str'
+
+    for k in kind:
+        assert type(
+            k) is str, 'invalid param "kind" for register, not a list of str'
+        assert k not in g_custom_layers, 'this type[%s] has already been registered' % (
+            k)
+        print('register layer[%s]' % (k))
+        g_custom_layers[k] = {
+            'shape': shape,
+            'layer': layer,
+            'weights': weights
+        }
+
+
+def get_registered_layers():
+    return g_custom_layers