add optimizer and aten for shufflenet

x2paddle/core/program.py

@@ -101,11 +101,6 @@ class PaddleGraph(object):
         self.clear_edges()
         outputs_from_nodes = dict()
         for layer_id, layer in self.layers.items():
-            #             if "x5109" in layer.outputs or "x5110" in layer.outputs:
-            #                 print(layer.kernel)
-            #                 print(layer.inputs)
-            #                 print(layer.outputs)
-            #                 print(layer.attrs)
             for input_key, input_var in layer.inputs.items():
                 vs = input_var
                 if not isinstance(vs, list):

x2paddle/decoder/pytorch_decoder.py

@@ -28,6 +28,7 @@ class PyTorchDecoder(object):
         torch._C._jit_pass_lint(graph)
         torch._C._jit_pass_dce(graph)
         torch._C._jit_pass_lint(graph)
-        graph = torch._C._jit_pass_canonicalize(graph)
+        torch._C._jit_pass_canonicalize(graph)
         torch._C._jit_pass_lint(graph)
+        torch._C._jit_pass_constant_propagation(graph)
         return graph

x2paddle/op_mapper/pytorch2paddle/aten.py

@@ -43,7 +43,7 @@ def aten_adaptive_avg_pool2d(mapper, graph, node):
     else:
         mapper._check_input(graph, inputs_node[1], inputs_name[1],
                             current_outputs)
-        layer_attrs["pool_size"] = inputs_name[1]
+        layer_inputs["pool_size"] = inputs_name[1]
         current_inputs.append(inputs_name[1])
     layer_attrs["pool_type"] = string("avg")
 
@@ -93,7 +93,7 @@ def aten_addmm(mapper, graph, node):
     else:
         mapper._check_input(graph, inputs_node[3], inputs_name[3],
                             current_outputs)
-        layer_attrs["beta"] = inputs_name[3]
+        layer_inputs["beta"] = inputs_name[3]
         current_inputs.append(inputs_name[3])
     # 处理输入4，即%151
     if inputs_name[4] in mapper.attrs:
@@ -101,7 +101,7 @@ def aten_addmm(mapper, graph, node):
     else:
         mapper._check_input(graph, inputs_node[4], inputs_name[4],
                             current_outputs)
-        layer_attrs["alpha"] = inputs_name[4]
+        layer_inputs["alpha"] = inputs_name[4]
         current_inputs.append(inputs_name[4])
 
     graph.add_layer(
@@ -175,7 +175,7 @@ def aten_add_(mapper, graph, node):
     else:
         mapper._check_input(graph, inputs_node[2], inputs_name[2],
                             current_outputs)
-        layer_attrs["alpha"] = inputs_name[2]
+        layer_inputs["alpha"] = inputs_name[2]
         current_inputs.append(inputs_name[2])
 
     graph.add_layer(
@@ -203,8 +203,7 @@ def aten___and__(mapper, graph, node):
     mapper._check_input(graph, inputs_node[0], inputs_name[0], current_outputs)
     layer_inputs["x"] = inputs_name[0]
     # 处理输入1，即%288
-    mapper._check_input(
-        graph, inputs_node[1], inputs_name[1], current_outputs, add_dim=True)
+    mapper._check_input(graph, inputs_node[1], inputs_name[1], current_outputs)
     layer_inputs["y"] = inputs_name[1]
     # 获取当前节点输入的list
     current_inputs = list(layer_inputs.values())
@@ -395,7 +394,7 @@ def aten_cat(mapper, graph, node):
     else:
         mapper._check_input(graph, inputs_node[1], inputs_name[1],
                             current_outputs)
-        layer_attrs["axis"] = inputs_name[1]
+        layer_inputs["axis"] = inputs_name[1]
         current_inputs.append(inputs_name[1])
     graph.add_layer(
         "fluid.layers.concat",
@@ -405,6 +404,82 @@ def aten_cat(mapper, graph, node):
     return current_inputs, current_outputs
 
 
+def aten_chunk(mapper, graph, node):
+    """构造分割Tensor的PaddleLayer。
+
+    TorchScript示例:
+        %724 : Tensor[] = aten::chunk(%input.170, %720, %719)
+        参数含义:
+        %724 (Tensor): 输出，分割后的结果。
+        %input.170 (Tensor): 需要进行分割的Tensor。
+        %720 (int): 分割的块数。
+        %719 (int): 分割的维度。
+    """
+    output_name = mapper._get_outputs_name(node)[0]
+    layer_outputs = [output_name]
+    layer_inputs = {}
+    layer_attrs = {}
+    inputs_name, inputs_node = mapper._get_inputs_name(node)
+    # 获取当前节点输出的list
+    current_outputs = [output_name]
+    # 处理输入0，即%input.170
+    mapper._check_input(graph, inputs_node[0], inputs_name[0], current_outputs)
+    layer_inputs["input"] = inputs_name[0]
+    # 获取当前节点输入的list
+    current_inputs = list(layer_inputs.values())
+    # 处理输入1，即%720
+    if inputs_name[1] in mapper.attrs:
+        layer_attrs["num_or_sections"] = mapper.attrs[inputs_name[1]]
+    else:
+        mapper._check_input(graph, inputs_node[1], inputs_name[1],
+                            current_outputs)
+        layer_inputs["num_or_sections"] = inputs_name[1]
+        current_inputs.append(inputs_name[1])
+    # 处理输入2，即%719
+    if inputs_name[2] in mapper.attrs:
+        layer_attrs["dim"] = mapper.attrs[inputs_name[2]]
+    else:
+        mapper._check_input(graph, inputs_node[2], inputs_name[2],
+                            current_outputs)
+        layer_inputs["dim"] = inputs_name[2]
+        current_inputs.append(inputs_name[2])
+    graph.add_layer(
+        "fluid.layers.split",
+        inputs=layer_inputs,
+        outputs=layer_outputs,
+        **layer_attrs)
+    return current_inputs, current_outputs
+
+
+def aten_contiguous(mapper, graph, node):
+    """ 构造在内存中连续存储的PaddleLayer。
+
+    TorchScript示例:
+        %x.7 : Tensor = aten::contiguous(%4058, %4046)
+        参数含义:
+        %x.7 (Tensor): 输出，在内存中连续存储的Tensor。
+        %4058 (Tensor): 原始Tensor。
+        %4046 (int): 存储的形式。
+
+    【注意】Paddle中无此用法，所以此处翻译成赋值。
+    """
+    output_name = mapper._get_outputs_name(node)[0]
+    layer_outputs = [output_name]
+    layer_inputs = {}
+    layer_attrs = {}
+    inputs_name, inputs_node = mapper._get_inputs_name(node)
+    # 获取当前节点输出的list
+    current_outputs = [output_name]
+    # 处理输入0，即%4058
+    mapper._check_input(graph, inputs_node[0], inputs_name[0], current_outputs)
+    layer_inputs["input"] = inputs_name[0]
+    # 获取当前节点输入的list
+    current_inputs = list(layer_inputs.values())
+
+    graph.add_layer("prim.equal", inputs=layer_inputs, outputs=layer_outputs)
+    return current_inputs, current_outputs
+
+
 def aten_conv2d(mapper, graph, node):
     """ 构造conv2d的PaddleLayer。
 
@@ -645,6 +720,35 @@ def aten_flatten(mapper, graph, node):
     return current_inputs, current_outputs
 
 
+def aten_floordiv(mapper, graph, node):
+    """ 构造向上取整除法的PaddleLayer。
+
+    TorchScript示例:
+        %channels_per_group.2 : int = aten::floordiv(%num_channels.2, %3690)
+        参数含义:
+        %channels_per_group.2 (-): 除后的结果。
+        %%num_channels.2 (-): 被除数。
+        %2 (int): 除数。
+    """
+    output_name = mapper._get_outputs_name(node)[0]
+    layer_outputs = [output_name]
+    layer_inputs = {}
+    inputs_name, inputs_node = mapper._get_inputs_name(node)
+    # 获取当前节点输出的list
+    current_outputs = [output_name]
+    # 处理输入0，即%124
+    mapper._check_input(graph, inputs_node[0], inputs_name[0], current_outputs)
+    layer_inputs["x"] = inputs_name[0]
+    # 处理输入1，即%123
+    mapper._check_input(graph, inputs_node[1], inputs_name[1], current_outputs)
+    layer_inputs["y"] = inputs_name[1]
+    # 获取当前节点输入的list
+    current_inputs = list(layer_inputs.values())
+
+    graph.add_layer("prim.floordiv", inputs=layer_inputs, outputs=layer_outputs)
+    return current_inputs, current_outputs
+
+
 def aten___getitem__(mapper, graph, node):
     """ 构造获取list中元素的PaddleLayer。
 
@@ -920,6 +1024,54 @@ def aten_matmul(mapper, graph, node):
     return current_inputs, current_outputs
 
 
+def aten_mean(mapper, graph, node):
+    """ 构造求均值的PaddleLayer。
+
+    TorchScript示例:
+        %x.28 : Tensor = aten::mean(%result.1, %4967, %3, %2)
+        参数含义:
+        %x.28 (Tensor): 输出，求均值后的结果。
+        %result.1 (Tensor): 输入，需要求均值的Tensor。
+        %4967 (int/list): 求平均值运算的维度。
+        %3 (bool): 是否在输出Tensor中保留减小的维度。
+        %2 (Tensor): 结果Tensor。
+    """
+    output_name = mapper._get_outputs_name(node)[0]
+    layer_outputs = [output_name]
+    layer_inputs = {}
+    layer_attrs = {}
+    inputs_name, inputs_node = mapper._get_inputs_name(node)
+    # 获取当前节点输出的list
+    current_outputs = [output_name]
+    # 处理输入0，即%result.1
+    mapper._check_input(graph, inputs_node[0], inputs_name[0], current_outputs)
+    layer_inputs["input"] = inputs_name[0]
+    current_inputs = list(layer_inputs.values())
+    # 处理输入1，即%4967
+    if inputs_name[1] in mapper.attrs:
+        layer_attrs["dim"] = mapper.attrs[inputs_name[1]]
+    else:
+        mapper._check_input(graph, inputs_node[1], inputs_name[1],
+                            current_outputs)
+        layer_inputs["dim"] = inputs_name[1]
+        current_inputs.append(inputs_name[1])
+    # 处理输入2，即%3
+    if inputs_name[1] in mapper.attrs:
+        layer_attrs["keep_dim"] = mapper.attrs[inputs_name[2]]
+    else:
+        mapper._check_input(graph, inputs_node[2], inputs_name[2],
+                            current_outputs)
+        layer_inputs["keep_dim"] = inputs_name[2]
+        current_inputs.append(inputs_name[2])
+
+    graph.add_layer(
+        "fluid.layers.reduce_mean",
+        inputs=layer_inputs,
+        outputs=layer_outputs,
+        **layer_attrs)
+    return current_inputs, current_outputs
+
+
 def aten_mul(mapper, graph, node):
     """ 构造数值相乘的PaddleLayer。
 
@@ -1136,14 +1288,22 @@ def aten_reshape(mapper, graph, node):
     # 处理输入0，即%4700
     mapper._check_input(graph, inputs_node[0], inputs_name[0], current_outputs)
     layer_inputs["x"] = inputs_name[0]
-    # 处理输入1，即%4703
-    mapper._check_input(graph, inputs_node[1], inputs_name[1], current_outputs)
-    layer_inputs["shape"] = inputs_name[1]
     # 获取当前节点输入、输出的list
     current_inputs = list(layer_inputs.values())
+    # 处理输入1，即%4703
+    if inputs_name[1] in mapper.attrs:
+        layer_attrs["shape"] = mapper.attrs[inputs_name[1]]
+    else:
+        mapper._check_input(graph, inputs_node[1], inputs_name[1],
+                            current_outputs)
+        layer_inputs["shape"] = inputs_name[1]
+        current_inputs.append(inputs_name[1])
 
     graph.add_layer(
-        "fluid.layers.reshape", inputs=layer_inputs, outputs=layer_outputs)
+        "fluid.layers.reshape",
+        inputs=layer_inputs,
+        outputs=layer_outputs,
+        **layer_attrs)
     return current_inputs, current_outputs
 
 
@@ -1305,6 +1465,72 @@ def aten_t(mapper, graph, node):
     return current_inputs, current_outputs
 
 
+def aten_transpose(mapper, graph, node):
+    """ 构造矩阵转置的PaddleLayer。
+
+    TorchScript示例:
+        %715 : Tensor = aten::transpose(%x.21, %704, %705)
+        参数含义:
+        %715 (Tensor): 输出，转置后的矩阵。
+        %x.21 (Tensor): 需要转置的Tensor。
+        %704 (int): 转置的维度1。
+        %705 (int): 转置的维度2。
+    """
+    output_name = mapper._get_outputs_name(node)[0]
+    layer_outputs = [output_name]
+    layer_inputs = {}
+    layer_attrs = {}
+    inputs_name, inputs_node = mapper._get_inputs_name(node)
+    # 获取当前节点输出的list
+    current_outputs = [output_name]
+    # 处理输入0，即%x.21
+    mapper._check_input(graph, inputs_node[0], inputs_name[0], current_outputs)
+    layer_inputs["x"] = inputs_name[0]
+    # 处理输入1，即%704
+    mapper._check_input(graph, inputs_node[1], inputs_name[1], current_outputs)
+    dim1 = inputs_name[1]
+    # 处理输入2，即%705
+    mapper._check_input(graph, inputs_node[2], inputs_name[2], current_outputs)
+    dim2 = inputs_name[2]
+    # 获取当前节点输入的list
+    current_inputs = list(layer_inputs.values())
+    graph.add_layer(
+        "prim.shape",
+        inputs={"input": inputs_name[0]},
+        outputs=[output_name + "_shape"])
+    current_outputs.append(output_name + "_shape")
+    graph.add_layer(
+        "prim.len",
+        inputs={"input": output_name + "_shape"},
+        outputs=[output_name + "_len"])
+    current_outputs.append(output_name + "_len")
+    current_inputs.append(output_name + "_shape")
+    graph.add_layer(
+        "prim.len2list",
+        inputs={"len": output_name + "_len"},
+        outputs=[output_name + "_list"])
+    current_outputs.append(output_name + "_list")
+    current_inputs.append(output_name + "_len")
+    graph.add_layer(
+        "prim.replaceitem",
+        inputs={"list": output_name + "_list",
+                "index": dim1,
+                "item": dim2},
+        outputs=[])
+    graph.add_layer(
+        "prim.replaceitem",
+        inputs={"list": output_name + "_list",
+                "index": dim2,
+                "item": dim1},
+        outputs=[])
+    graph.add_layer(
+        "fluid.layers.transpose",
+        inputs=layer_inputs,
+        outputs=layer_outputs,
+        perm=output_name + "_list")
+    return current_inputs, current_outputs
+
+
 def aten_unsqueeze(mapper, graph, node):
     """ 构造插入维度的PaddleLayer。
 
@@ -1333,7 +1559,7 @@ def aten_unsqueeze(mapper, graph, node):
     else:
         mapper._check_input(graph, inputs_node[1], inputs_name[1],
                             current_outputs)
-        layer_attrs["axes"] = inputs_name[1]
+        layer_inputs["axes"] = inputs_name[1]
         current_inputs.append(inputs_name[1])
     graph.add_layer(
         "fluid.layers.unsqueeze",
@@ -1343,6 +1569,51 @@ def aten_unsqueeze(mapper, graph, node):
     return current_inputs, current_outputs
 
 
+def aten_view(mapper, graph, node):
+    """ 构造调整大小的PaddleLayer。
+
+    TorchScript示例:
+        %input.152 : Tensor = aten::view(%x.20, %430)
+        参数含义:
+        %input.152 (Tensor): 输出，view后的Tensor。
+        %x.20 (Tensor): 需要view的Tensor。
+        %430 (list): 形状大小组成的list。
+
+    【注意】view 函数只能用于contiguous后的Tensor上，
+          也就是只能用于内存中连续存储的Tensor。
+          如果对Tensor调用过transpose，permute等操作的话会使该Tensor在内存中变得不再连续，
+          此时就不能再调用view函数。因此，需要先使用contiguous来返回一个contiguous copy。
+          reshape则不需要依赖目标Tensor是否在内存中是连续的。
+    """
+    output_name = mapper._get_outputs_name(node)[0]
+    layer_outputs = [output_name]
+    layer_inputs = {}
+    layer_attrs = {}
+    inputs_name, inputs_node = mapper._get_inputs_name(node)
+    # 获取当前节点输出的list
+    current_outputs = [output_name]
+    # 处理输入0，即%x.20
+    mapper._check_input(graph, inputs_node[0], inputs_name[0], current_outputs)
+    layer_inputs["x"] = inputs_name[0]
+    # 获取当前节点输入、输出的list
+    current_inputs = list(layer_inputs.values())
+    # 处理输入1，即%430
+    if inputs_name[1] in mapper.attrs:
+        layer_attrs["shape"] = mapper.attrs[inputs_name[1]]
+    else:
+        mapper._check_input(graph, inputs_node[1], inputs_name[1],
+                            current_outputs)
+        layer_inputs["shape"] = inputs_name[1]
+        current_inputs.append(inputs_name[1])
+
+    graph.add_layer(
+        "fluid.layers.reshape",
+        inputs=layer_inputs,
+        outputs=layer_outputs,
+        **layer_attrs)
+    return current_inputs, current_outputs
+
+
 def aten_warn(mapper, graph, node):
     """ 构造warning的PaddleLayer。
 
@@ -1370,7 +1641,7 @@ def aten_warn(mapper, graph, node):
     else:
         mapper._check_input(graph, inputs_node[1], inputs_name[1],
                             current_outputs)
-        layer_attrs["stacklevel"] = inputs_name[1]
+        layer_inputs["stacklevel"] = inputs_name[1]
         current_inputs.append(inputs_name[1])
 
     graph.add_layer(

x2paddle/op_mapper/pytorch2paddle/prim.py

@@ -35,6 +35,34 @@ def prim_Constant(mapper, graph, node):
     return [], [output_name]
 
 
+def prim_data(mapper, graph, node):
+    """ 构造Tensor的PaddleLayer。
+
+    TorchScript示例:
+        %4336 : Tensor = prim::data(%out.6)
+        参数含义:
+        %4336 (Tensor): 输出Tensor。
+        %out.6 (Tensor): 原始Tensor。
+
+    【注意】Paddle中无此用法，所以此处翻译成赋值。
+    """
+    output_name = mapper._get_outputs_name(node)[0]
+    layer_outputs = [output_name]
+    layer_inputs = {}
+    layer_attrs = {}
+    inputs_name, inputs_node = mapper._get_inputs_name(node)
+    # 获取当前节点输出的list
+    current_outputs = [output_name]
+    # 处理输入0，即%4336
+    mapper._check_input(graph, inputs_node[0], inputs_name[0], current_outputs)
+    layer_inputs["input"] = inputs_name[0]
+    # 获取当前节点输入的list
+    current_inputs = list(layer_inputs.values())
+
+    graph.add_layer("prim.equal", inputs=layer_inputs, outputs=layer_outputs)
+    return current_inputs, current_outputs
+
+
 def prim_GetAttr(mapper, graph, node):
     """ 获取attribute信息。
 
@@ -66,6 +94,46 @@ def prim_GetAttr(mapper, graph, node):
     return [], [output_name]
 
 
+def prim_If(mapper, graph, node):
+    """ 构造if控制流的PaddleLayer。
+
+    TorchScript示例:
+        %input.5 : Tensor = prim::If(%107)
+          block0():
+            %109 : Tensor = aten::t(%102)
+            %ret.2 : Tensor = aten::addmm(%103, %101, %109, %104, %104)
+            -> (%ret.2)
+          block1():
+            %111 : Tensor = aten::t(%102)
+            ...
+            -> (%output.4)
+        参数含义:
+        %107 (bool): if判断条件。
+        %input.5 (Tensor): if控制流的输出，与%output.4对应。
+    """
+    output_name = mapper._get_outputs_name(node)[0]
+    node_outputs = [output_name]
+    input_node = list(node.inputs())[0].node()
+    script_input_unique_id = list(node.inputs())[0].unique()
+    input_node_name = mapper.outputs_info[script_input_unique_id]
+    mapper._check_input(graph, input_node, input_node_name, node_outputs)
+    graph.add_layer("prim.if", {'input': input_node_name}, [output_name])
+    current_layer = list(graph.layers.values())[-1]
+    block0 = list(node.blocks())[0]
+    block0_graph, graph_inputs0 = mapper.traverse(block0, current_layer)
+    len0 = 0
+    for i, input_name in enumerate(graph_inputs0):
+        current_layer.inputs['input-{}'.format(i)] = input_name
+        len0 = i
+    current_layer.add_block(block0_graph)
+    block1 = list(node.blocks())[1]
+    block1_graph, graph_inputs1 = mapper.traverse(block1, current_layer)
+    for i, input_name in enumerate(graph_inputs1):
+        current_layer.inputs['input-{}'.format(len0 + 1 + i)] = input_name
+    current_layer.add_block(block1_graph)
+    return list(current_layer.inputs.values()), node_outputs
+
+
 def prim_ListConstruct(mapper, graph, node):
     """ 构造list的PaddleLayer。
 
@@ -92,28 +160,30 @@ def prim_ListConstruct(mapper, graph, node):
     return current_inputs, current_outputs
 
 
-def prim_RaiseException(mapper, graph, node):
-    """ 构造抛出异常的PaddleLayer。
+def prim_ListUnpack(mapper, graph, node):
+    """ 构造获取list中元素的PaddleLayer。
 
     TorchScript示例:
-        = prim::RaiseException(%76)
+        %x1.4 : Tensor, %x2.4 : Tensor = prim::ListUnpack(%4354)
         参数含义:
-        %76 (str): 异常信息。
+        %x1.4 (Tensor): 输出，list的第一个元素。
+        %x2.4 (Tensor): 输出，list的第二个元素。
+        %4354 (list): 列表。
     """
-    output_name = mapper._get_outputs_name(node)[0]
-    layer_outputs = [output_name]
+    outputs_name = mapper._get_outputs_name(node)
+    layer_outputs = outputs_name.copy()
     layer_inputs = {}
     inputs_name, inputs_node = mapper._get_inputs_name(node)
     # 获取当前节点输出的list
-    current_outputs = [output_name]
-    # 处理输入0，即%76
+    current_outputs = layer_outputs.copy()
+    # 处理输入0，即%4354
     mapper._check_input(graph, inputs_node[0], inputs_name[0], current_outputs)
     layer_inputs["input"] = inputs_name[0]
     # 获取当前节点输入的list
     current_inputs = list(layer_inputs.values())
 
     graph.add_layer(
-        "prim.exception", inputs=layer_inputs, outputs=layer_outputs)
+        "prim.list_unpack", inputs=layer_inputs, outputs=layer_outputs)
     return current_inputs, current_outputs
 
 
@@ -180,46 +250,6 @@ def prim_Loop(mapper, graph, node):
     return list(current_layer.inputs.values()), node_outputs
 
 
-def prim_If(mapper, graph, node):
-    """ 构造if控制流的PaddleLayer。
-
-    TorchScript示例:
-        %input.5 : Tensor = prim::If(%107)
-          block0():
-            %109 : Tensor = aten::t(%102)
-            %ret.2 : Tensor = aten::addmm(%103, %101, %109, %104, %104)
-            -> (%ret.2)
-          block1():
-            %111 : Tensor = aten::t(%102)
-            ...
-            -> (%output.4)
-        参数含义:
-        %107 (bool): if判断条件。
-        %input.5 (Tensor): if控制流的输出，与%output.4对应。
-    """
-    output_name = mapper._get_outputs_name(node)[0]
-    node_outputs = [output_name]
-    input_node = list(node.inputs())[0].node()
-    script_input_unique_id = list(node.inputs())[0].unique()
-    input_node_name = mapper.outputs_info[script_input_unique_id]
-    mapper._check_input(graph, input_node, input_node_name, node_outputs)
-    graph.add_layer("prim.if", {'input': input_node_name}, [output_name])
-    current_layer = list(graph.layers.values())[-1]
-    block0 = list(node.blocks())[0]
-    block0_graph, graph_inputs0 = mapper.traverse(block0, current_layer)
-    len0 = 0
-    for i, input_name in enumerate(graph_inputs0):
-        current_layer.inputs['input-{}'.format(i)] = input_name
-        len0 = i
-    current_layer.add_block(block0_graph)
-    block1 = list(node.blocks())[1]
-    block1_graph, graph_inputs1 = mapper.traverse(block1, current_layer)
-    for i, input_name in enumerate(graph_inputs1):
-        current_layer.inputs['input-{}'.format(len0 + 1 + i)] = input_name
-    current_layer.add_block(block1_graph)
-    return list(current_layer.inputs.values()), node_outputs
-
-
 def prim_min(mapper, graph, node):
     """ 构造min的PaddleLayer。
 
@@ -245,6 +275,31 @@ def prim_min(mapper, graph, node):
     return current_inputs, current_outputs
 
 
+def prim_RaiseException(mapper, graph, node):
+    """ 构造抛出异常的PaddleLayer。
+
+    TorchScript示例:
+        = prim::RaiseException(%76)
+        参数含义:
+        %76 (str): 异常信息。
+    """
+    output_name = mapper._get_outputs_name(node)[0]
+    layer_outputs = [output_name]
+    layer_inputs = {}
+    inputs_name, inputs_node = mapper._get_inputs_name(node)
+    # 获取当前节点输出的list
+    current_outputs = [output_name]
+    # 处理输入0，即%76
+    mapper._check_input(graph, inputs_node[0], inputs_name[0], current_outputs)
+    layer_inputs["input"] = inputs_name[0]
+    # 获取当前节点输入的list
+    current_inputs = list(layer_inputs.values())
+
+    graph.add_layer(
+        "prim.exception", inputs=layer_inputs, outputs=layer_outputs)
+    return current_inputs, current_outputs
+
+
 def prim_requires_grad(mapper, graph, node):
     """ 构造是否计算梯度的PaddleLayer。
 

x2paddle/op_mapper/pytorch2paddle/prim2code.py

@@ -100,6 +100,12 @@ def prim_exception(layer, indent=1, init_func=[], forward_func=[]):
     forward_func.extend(gen_codes([line], indent=indent))
 
 
+def prim_floordiv(layer, indent=1, init_func=[], forward_func=[]):
+    line = "{} = {} // {}".format(layer.outputs[0],
+                                  get_value(layer, "x"), get_value(layer, "y"))
+    forward_func.extend(gen_codes([line], indent=indent))
+
+
 def prim_if(layer, indent=1, init_func=[], forward_func=[]):
     line = "if {} :".format(get_value(layer, "input"))
     forward_func.extend(gen_codes([line], indent=indent))
@@ -141,6 +147,14 @@ def prim_len(layer, indent=1, init_func=[], forward_func=[]):
     forward_func.extend(gen_codes([line], indent=indent))
 
 
+def prim_len2list(layer, indent=1, init_func=[], forward_func=[]):
+    lines = []
+    lines.append("{} = []".format(layer.outputs[0]))
+    lines.append("for i in range({}):".format(get_value(layer, "len")))
+    lines.append("    {}.append(i)".format(layer.outputs[0]))
+    forward_func.extend(gen_codes(lines, indent=indent))
+
+
 def prim_lt(layer, indent=1, init_func=[], forward_func=[]):
     line = "{} = {} < {}".format(layer.outputs[0],
                                  get_value(layer, "x"), get_value(layer, "y"))
@@ -157,6 +171,11 @@ def prim_list(layer, indent=1, init_func=[], forward_func=[]):
     forward_func.extend(gen_codes([line], indent=indent))
 
 
+def prim_list_unpack(layer, indent=1, init_func=[], forward_func=[]):
+    line = "{} = {}".format(", ".join(layer.outputs), get_value(layer, "input"))
+    forward_func.extend(gen_codes([line], indent=indent))
+
+
 def prim_loop(layer, indent=1, init_func=[], forward_func=[]):
     loop_range = get_value(layer, "input")
     line = "for {} in range({}):".format(layer.outputs[1], loop_range)
@@ -194,6 +213,13 @@ def prim_not(layer, indent=1, init_func=[], forward_func=[]):
     forward_func.extend(gen_codes([line], indent=indent))
 
 
+def prim_replaceitem(layer, indent=1, init_func=[], forward_func=[]):
+    line = "{}[{}] = {}".format(
+        get_value(layer, "list"),
+        get_value(layer, "index"), get_value(layer, "item"))
+    forward_func.extend(gen_codes([line], indent=indent))
+
+
 def prim_requires_grad(layer, indent=1, init_func=[], forward_func=[]):
     line = "{} = not {}.stop_gradient".format(layer.outputs[0],
                                               get_value(layer, "input"))

x2paddle/optimizer/fusion/__init__.py

@@ -14,10 +14,8 @@
 
 from .fc_fuser import FcFuser
 from .fc_fuse_pass import FcFusePass
-from .nn_adaptive_pool2d_fuser import NnAdaptivePool2dFuser
-from .nn_adaptive_pool2d_fuse_pass import NnAdaptivePool2dFusePass
-from .functional_adaptive_pool2d_fuser import FunctionalAdaptivePool2dFuser
-from .functional_adaptive_pool2d_fuse_pass import FunctionalAdaptivePool2dFusePass
+from .adaptive_pool2d_fuser import AdaptivePool2dFuser
+from .adaptive_pool2d_fuse_pass import AdaptivePool2dFusePass
 from .constant_fuser import ConstantFuser
 from .constant_fuse_pass import ConstantFusePass
 from .batchnorm2d_fuser import BatchNorm2dFuser

x2paddle/optimizer/fusion/adaptive_pool2d_fuse_pass.py

+#   Copyright (c) 2020  PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License"
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from x2paddle.optimizer.pass_ import Pass
+from x2paddle.optimizer.fusion import AdaptivePool2dFuser
+from x2paddle.optimizer.pass_manager import pass_register
+
+
+@pass_register
+class AdaptivePool2dFusePass(Pass):
+    name = "adaptive_pool2d_fuse_pass"
+
+    def __init__(self):
+        Pass.__init__(self)
+
+    def apply(self, graph):
+        fuser = AdaptivePool2dFuser()
+        fuser.operate(graph, match_kind="topo")
+
+
+# 用于注册
+adaptive_pool2d_fuse_pass = AdaptivePool2dFusePass()

x2paddle/optimizer/fusion/adaptive_pool2d_fuser.py

+#   Copyright (c) 2020  PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License"
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import numpy as np
+from x2paddle.optimizer.pattern_matcher import FuseBase
+from x2paddle.core.program import PaddleGraph, PaddleLayer
+from x2paddle.core.util import *
+
+
+class AdaptivePool2dFuser(FuseBase):
+    def __init__(self):
+        super(AdaptivePool2dFuser, self).__init__(graph_type="dygraph")
+
+    def build_pattern(self):
+        """ 描述需要替换的adaptive pool2d图结构。
+        adaptive pool2d层模式python实现代码示例:
+            x72 = [6, 6]
+            x73 = x71.shape
+            x75 = 'Exception'
+            x76 = 9223372036854775807
+            x77 = 1
+            x78 = len(x73)
+            x79 = 2
+            x80 = x78 <= x79
+            if x80 :
+                raise RaiseException(x75)
+            x83 = []
+            x84 = -2
+            x85 = x73[x84: x76: x77]
+            x86 = 2
+            x87 = len(x85)
+            x88 = [x86, x87]
+            x89 = min(x88)
+            for _x91 in range(x89):
+                x92 = x72[_x91]
+                x83.append(x92)
+            x93 = fluid.layers.adaptive_pool2d(input=x71, pool_size=x83, pool_type='avg')
+        """
+
+        def gen_name(id):
+            return "x" + str(id)
+
+        self.pattern.add_layer(
+            "prim.constant", inputs={}, outputs=[gen_name(0)], value=[6, 6])
+        self.pattern.add_layer(
+            "prim.shape",
+            inputs={'input': "pool-input-0"},
+            outputs=[gen_name(1)])
+        self.pattern.add_layer(
+            "prim.constant", inputs={}, outputs=[gen_name(2)], value=True)
+        self.pattern.add_layer(
+            "prim.constant",
+            inputs={},
+            outputs=[gen_name(3)],
+            value="Exception")
+        self.pattern.add_layer(
+            "prim.constant",
+            inputs={},
+            outputs=[gen_name(4)],
+            value=9223372036854775807)
+        self.pattern.add_layer(
+            "prim.constant", inputs={}, outputs=[gen_name(5)], value=1)
+        self.pattern.add_layer(
+            "prim.len", inputs={"input": gen_name(1)}, outputs=[gen_name(6)])
+        self.pattern.add_layer(
+            "prim.constant", inputs={}, outputs=[gen_name(7)], value=2)
+        self.pattern.add_layer(
+            "prim.le",
+            inputs={"x": gen_name(6),
+                    "y": gen_name(7)},
+            outputs=[gen_name(8)])
+        self.pattern.add_layer("prim.if", {'input': gen_name(8)}, [gen_name(9)])
+        if_layer = self.pattern.layers[list(self.pattern.layers.keys())[-1]]
+        pattern_block0 = PaddleGraph(if_layer, graph_type="dygraph")
+        pattern_block0.add_layer(
+            "prim.exception",
+            inputs={"input": gen_name(3)},
+            outputs=[gen_name(9)])
+        if_layer.inputs["input-0"] = gen_name(3)
+        if_layer.add_block(pattern_block0)
+        pattern_block1 = PaddleGraph(if_layer, graph_type="dygraph")
+        if_layer.add_block(pattern_block1)
+        self.pattern.add_layer("prim.list", inputs={}, outputs=[gen_name(10)])
+        self.pattern.add_layer(
+            "prim.constant", inputs={}, outputs=[gen_name(11)], value=-2)
+        self.pattern.add_layer(
+            "prim.slice",
+            inputs={
+                "input": gen_name(1),
+                "start": gen_name(11),
+                "end": gen_name(4),
+                "step": gen_name(5)
+            },
+            outputs=[gen_name(12)])
+        self.pattern.add_layer(
+            "prim.constant", inputs={}, outputs=[gen_name(13)], value=2)
+        self.pattern.add_layer(
+            "prim.len", inputs={"input": gen_name(12)}, outputs=[gen_name(14)])
+        self.pattern.add_layer(
+            "prim.list",
+            inputs={"input0": gen_name(13),
+                    "input1": gen_name(14)},
+            outputs=[gen_name(15)])
+        self.pattern.add_layer(
+            "prim.min", inputs={"input": gen_name(15)}, outputs=[gen_name(16)])
+        self.pattern.add_layer("prim.loop", {'input': gen_name(16)},
+                               [gen_name(17), gen_name(18)])
+        loop_layer = self.pattern.layers[list(self.pattern.layers.keys())[-1]]
+        pattern_block = PaddleGraph(loop_layer, graph_type="dygraph")
+        pattern_block.add_layer(
+            "prim.getitem",
+            inputs={"list": gen_name(0),
+                    "index": gen_name(18)},
+            outputs=[gen_name(19)])
+        pattern_block.add_layer(
+            "prim.append",
+            inputs={"list": gen_name(10),
+                    "index": gen_name(19)},
+            outputs=[gen_name(20)])
+        loop_layer.inputs["input-0"] = gen_name(0)
+        loop_layer.inputs["input-2"] = gen_name(10)
+        loop_layer.add_block(pattern_block)
+        pool_attrs = {'pool_type': string("avg")}
+        self.pattern.add_layer(
+            "fluid.layers.adaptive_pool2d",
+            inputs={'input': "pool-input-0",
+                    "pool_size": gen_name(10)},
+            outputs=[gen_name(21)],
+            **pool_attrs)
+        self.pattern.build(inputs={"input-0": "pool-input-0"})
+
+    def insert_new_layer(self, graph, parameters, matches):
+        parameters = graph.parameters
+        new_layer = self.gen_new_layer(parameters, matches)
+        new_layer_id = list(matches.keys())[0]
+        graph.layers[new_layer_id] = new_layer
+        matches.pop(new_layer_id)
+
+    def gen_new_layer(self, parameters, matches):
+        layers_id = list(matches.keys())
+        layer = matches[layers_id[0]]
+        pool_size = layer.attrs["value"]
+        layer = matches[layers_id[1]]
+        input_name = layer.inputs["input"]
+        layer = matches[layers_id[-1]]
+        output_name = layer.outputs[0]
+        pool_type = layer.attrs["pool_type"]
+        attrs = dict()
+        attrs["pool_size"] = pool_size
+        attrs["pool_type"] = pool_type
+        new_layer = PaddleLayer(
+            layers_id[0],
+            "fluid.layers.adaptive_pool2d",
+            inputs={"input": input_name},
+            outputs=[output_name],
+            **attrs)
+        return new_layer

x2paddle/optimizer/fusion/batchnorm2d_fuse_pass.py

+#   Copyright (c) 2020  PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License"
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from x2paddle.optimizer.pass_ import Pass
+from x2paddle.optimizer.fusion import BatchNorm2dFuser
+from x2paddle.optimizer.pass_manager import pass_register
+
+
+@pass_register
+class BatchNorm2dFusePass(Pass):
+    name = "batchnorm2d_fuse_pass"
+
+    def __init__(self):
+        Pass.__init__(self)
+
+    def apply(self, graph):
+        fuser = BatchNorm2dFuser()
+        fuser.operate(graph, match_kind="topo")
+
+
+# 用于注册
+batchnorm2d_fuse_pass = BatchNorm2dFusePass()

x2paddle/optimizer/fusion/batchnorm2d_fuser.py

+#   Copyright (c) 2020  PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License"
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import numpy as np
+from x2paddle.optimizer.pattern_matcher import FuseBase
+from x2paddle.core.program import PaddleGraph, PaddleLayer
+from x2paddle.core.util import *
+
+
+class BatchNorm2dFuser(FuseBase):
+    def __init__(self):
+        super(BatchNorm2dFuser, self).__init__(graph_type="dygraph")
+
+    def build_pattern(self):
+        """ 描述需要替换的batchnorm2d图结构。
+        batchnorm2d层模式python实现代码示例:
+            x2209 = 1
+            x2212 = 'Exception'
+            x2213 = 4
+            x2214 = x2207.shape
+            x2214 = len(x2214)
+            x2215 = x2214 != x2213
+            if x2215 :
+                raise RaiseException(x2212)
+            x2218 = False
+            if x2218 :
+                x2220 = self.x2220
+                x2221 = x2220 + x2209
+                self.x2220 = x2221
+            x2227 = False
+            if x2227 :
+                x2230 = x2207.shape
+                x2231 = 'Exception'
+                x2233 = 0
+                x2234 = 2
+                x2235 = 1
+                x2236 = x2230[x2233]
+                x2237 = len(x2230)
+                x2238 = x2237 - x2234
+                x2241 = x2236
+                for _x2240 in range(x2238):
+                    x2242 = _x2240 + x2234
+                    x2243 = x2230[x2242]
+                    x2244 = x2241 * x2243
+                    x2239 = x2244
+                x2245 = x2239 == x2235
+                if x2245 :
+                    raise RaiseException(x2231)
+            x2248 = self.batchnorm41(x2207)
+        """
+
+        def gen_name(id):
+            return "x" + str(id)
+
+        self.pattern.add_layer(
+            "prim.constant", inputs={}, outputs=[gen_name(0)], value=1)
+        self.pattern.add_layer(
+            "prim.constant", inputs={}, outputs=[gen_name(1)], value=0.1)
+        self.pattern.add_layer(
+            "prim.constant", inputs={}, outputs=[gen_name(2)], value=0.001)
+        self.pattern.add_layer(
+            "prim.constant",
+            inputs={},
+            outputs=[gen_name(3)],
+            value="Exception")
+        self.pattern.add_layer(
+            "prim.constant", inputs={}, outputs=[gen_name(4)], value=4)
+        self.pattern.add_layer(
+            "prim.shape", inputs={'input': "bn-input-0"},
+            outputs=[gen_name(5)])
+        self.pattern.add_layer(
+            "prim.len", inputs={'input': gen_name(5)}, outputs=[gen_name(5)])
+        self.pattern.add_layer(
+            "prim.ne",
+            inputs={"x": gen_name(5),
+                    "y": gen_name(4)},
+            outputs=[gen_name(6)])
+        self.pattern.add_layer("prim.if", {'input': gen_name(6)}, [gen_name(7)])
+        if_layer1 = self.pattern.layers[list(self.pattern.layers.keys())[-1]]
+        pattern_block0 = PaddleGraph(if_layer1, graph_type="dygraph")
+        pattern_block0.add_layer(
+            "prim.exception",
+            inputs={"input": gen_name(3)},
+            outputs=[gen_name(8)])
+        if_layer1.inputs["input-0"] = gen_name(3)
+        if_layer1.add_block(pattern_block0)
+        pattern_block1 = PaddleGraph(if_layer1, graph_type="dygraph")
+        if_layer1.add_block(pattern_block1)
+        self.pattern.add_layer(
+            "prim.constant", inputs={}, outputs=[gen_name(9)], value=False)
+        self.pattern.add_layer("prim.if", {'input': gen_name(9)},
+                               [gen_name(10)])
+        if_layer2 = self.pattern.layers[list(self.pattern.layers.keys())[-1]]
+        pattern_block0 = PaddleGraph(if_layer2, graph_type="dygraph")
+        pattern_block0.add_layer(
+            "fluid.dygraph.base.to_variable",
+            inputs={},
+            outputs=[gen_name(11)],
+            value="params[{}]".format(string(gen_name(11))))
+        pattern_block0.add_layer(
+            "prim.add",
+            inputs={"x": gen_name(11),
+                    "y": gen_name(0)},
+            outputs=[gen_name(12)])
+        pattern_block0.add_layer(
+            "prim.set_attr",
+            inputs={"input": gen_name(12)},
+            outputs=["self." + gen_name(11)])
+        if_layer2.inputs["input-0"] = gen_name(0)
+        if_layer2.add_block(pattern_block0)
+        pattern_block1 = PaddleGraph(if_layer2, graph_type="dygraph")
+        if_layer2.add_block(pattern_block1)
+        self.pattern.add_layer(
+            "prim.constant", inputs={}, outputs=[gen_name(13)], value=True)
+        self.pattern.add_layer(
+            "prim.constant", inputs={}, outputs=[gen_name(14)], value=False)
+        self.pattern.add_layer("prim.if", {'input': gen_name(14)},
+                               [gen_name(15)])
+        if_layer3 = self.pattern.layers[list(self.pattern.layers.keys())[-1]]
+        pattern_block0 = PaddleGraph(if_layer3, graph_type="dygraph")
+        pattern_block0.add_layer(
+            "prim.shape",
+            inputs={'input': "bn-input-0"},
+            outputs=[gen_name(16)])
+        pattern_block0.add_layer(
+            "prim.constant",
+            inputs={},
+            outputs=[gen_name(17)],
+            value="Exception")
+        pattern_block0.add_layer(
+            "prim.constant", inputs={}, outputs=[gen_name(18)], value=True)
+        pattern_block0.add_layer(
+            "prim.constant", inputs={}, outputs=[gen_name(19)], value=0)
+        pattern_block0.add_layer(
+            "prim.constant", inputs={}, outputs=[gen_name(20)], value=2)
+        pattern_block0.add_layer(
+            "prim.constant", inputs={}, outputs=[gen_name(21)], value=1)
+        pattern_block0.add_layer(
+            "prim.getitem",
+            inputs={"list": gen_name(16),
+                    "index": gen_name(19)},
+            outputs=[gen_name(22)])
+        pattern_block0.add_layer(
+            "prim.len", inputs={"input": gen_name(16)}, outputs=[gen_name(23)])
+        pattern_block0.add_layer(
+            "prim.sub",
+            inputs={"x": gen_name(23),
+                    "y": gen_name(20)},
+            outputs=[gen_name(24)])
+        pattern_block0.add_layer(
+            "prim.equal",
+            inputs={"input": gen_name(22)},
+            outputs=[gen_name(25)])
+        pattern_block0.add_layer(
+            "prim.loop",
+            inputs={"input": gen_name(24)},
+            outputs=[gen_name(26), gen_name(27)])
+        loop_layer = pattern_block0.layers[list(pattern_block0.layers.keys())[
+            -1]]
+        pattern_block0_block0 = PaddleGraph(loop_layer, graph_type="dygraph")
+        pattern_block0_block0.add_layer(
+            "prim.add",
+            inputs={"x": gen_name(27),
+                    "y": gen_name(20)},
+            outputs=[gen_name(28)])
+        pattern_block0_block0.add_layer(
+            "prim.getitem",
+            inputs={"list": gen_name(16),
+                    "index": gen_name(28)},
+            outputs=[gen_name(29)])
+        pattern_block0_block0.add_layer(
+            "prim.mul",
+            inputs={"x": gen_name(25),
+                    "y": gen_name(29)},
+            outputs=[gen_name(30)])
+        pattern_block0_block0.add_layer(
+            "prim.equal",
+            inputs={"input": gen_name(30)},
+            outputs=[gen_name(26)])
+        loop_layer.inputs["input-1"] = gen_name(20)
+        loop_layer.inputs["input-2"] = gen_name(16)
+        loop_layer.inputs["input-3"] = gen_name(25)
+        loop_layer.add_block(pattern_block0_block0)
+        pattern_block0.add_layer(
+            "prim.eq",
+            inputs={"x": gen_name(26),
+                    "y": gen_name(21)},
+            outputs=[gen_name(31)])
+        pattern_block0.add_layer(
+            "prim.if", inputs={"input": gen_name(31)}, outputs=[gen_name(32)])
+        if_layer31 = pattern_block0.layers[list(pattern_block0.layers.keys())[
+            -1]]
+        pattern_block0_block0 = PaddleGraph(if_layer31, graph_type="dygraph")
+        pattern_block0_block0.add_layer(
+            "prim.exception",
+            inputs={"input": gen_name(17)},
+            outputs=[gen_name(33)])
+        if_layer31.inputs["input-0"] = gen_name(17)
+        if_layer31.add_block(pattern_block0_block0)
+        pattern_block0_block1 = PaddleGraph(if_layer31, graph_type="dygraph")
+        if_layer31.add_block(pattern_block0_block1)
+        if_layer3.inputs["input-0"] = "bn-input-0"
+        if_layer3.add_block(pattern_block0)
+        pattern_block1 = PaddleGraph(if_layer3, graph_type="dygraph")
+        if_layer3.add_block(pattern_block1)
+        self.pattern.add_layer(
+            "fluid.dygraph.BatchNorm",
+            inputs={"input": "bn-input-0"},
+            outputs=[gen_name(34), gen_name(35)],
+            is_test=True,
+            num_channels=160,
+            momentum=0.1,
+            epsilon=0.001)
+        self.pattern.build(inputs={"input-0": "bn-input-0"})
+
+    def insert_new_layer(self, graph, parameters, matches):
+        new_layer = self.gen_new_layer(parameters, matches)
+        new_layer_id = list(matches.keys())[0]
+        graph.layers[new_layer_id] = new_layer
+        matches.pop(new_layer_id)
+
+    def gen_new_layer(self, parameters, matches):
+        layers_id = list(matches.keys())
+        layer = matches[layers_id[-1]]
+        return layer

x2paddle/optimizer/fusion/constant_fuse_pass.py

+# Licensed under the Apache License, Version 2.0 (the "License"
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from x2paddle.optimizer.pass_ import Pass
+from x2paddle.optimizer.fusion import ConstantFuser
+from x2paddle.optimizer.pass_manager import pass_register
+
+
+@pass_register
+class ConstantFusePass(Pass):
+    name = "constant_fuse_pass"
+
+    def __init__(self):
+        Pass.__init__(self)
+
+    def apply(self, graph):
+        fuser = ConstantFuser()
+        fuser.operate(graph, match_kind="topo")
+
+
+# 用于注册
+constant_fuse_pass = ConstantFuser()

x2paddle/optimizer/fusion/constant_fuser.py

+#   Copyright (c) 2020  PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License"
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import numpy as np
+from x2paddle.optimizer.pattern_matcher import FuseBase
+from x2paddle.core.program import PaddleGraph, PaddleLayer
+from x2paddle.core.util import *
+
+
+class ConstantFuser(FuseBase):
+    def __init__(self):
+        super(ConstantFuser, self).__init__(graph_type="dygraph")
+
+    def build_pattern(self):
+        """ 描述需要替换的constant图结构。
+        constant层模式python实现代码示例:
+            x3 = 10
+            for _x70 in range(x3):
+                ...
+        """
+        self.pattern.add_layer(
+            "prim.constant", inputs={}, outputs=["x1"], value=2)
+        self.pattern.build()
+        self.pattern.outputs = ["x1"]
+
+    def insert_new_layer(self, graph, parameters, matches):
+        def replace_value(layer_connect, match_name, match_value):
+            for k, v in layer_connect.inputs.items():
+                if v == match_name:
+                    layer_connect.inputs.pop(k)
+                    layer_connect.attrs[k] = match_value
+                    break
+            for k, v in layer_connect.attrs.items():
+                if v == match_name:
+                    layer_connect.attrs[k] = match_value
+                    break
+            if layer_connect.kernel == "prim.loop" or \
+            layer_connect.kernel == "prim.if":
+                for block in layer_connect.blocks:
+                    for b_layer_id, b_layer in block.layers.items():
+                        if block.edges_in.get(b_layer_id, 0) != 0 and  \
+                        -1 in block.edges_in[b_layer_id]:
+                            replace_value(b_layer, match_name, match_value)
+
+        layer_id = list(matches.keys())[0]
+        layer = list(matches.values())[0]
+        layer_output_name = layer.outputs[0]
+        layer_value = layer.attrs["value"]
+        if graph.edges_out.get(layer_id, 0) != 0:
+            for layer_id_out in graph.edges_out[layer_id]:
+                layer_connect = graph.layers[layer_id_out]
+                replace_value(layer_connect, layer_output_name, layer_value)

x2paddle/optimizer/fusion/fc_fuser.py

@@ -26,28 +26,23 @@ class FcFuser(FuseBase):
     def build_pattern(self):
         """ 描述需要替换的fc图结构。
         fc层模式python实现代码示例:
-            x149 = 2
-            x151 = x146.shape
-            x151 = len(x151)
-            x152 = x151 == x149
-            if x152 :
-                x147 = self.x147
-                x154 = fluid.layers.transpose(x=x147, perm=[1, 0])
-                x148 = self.x148
-                x155 = fluid.layers.addmm(input=x148, x=x146, y=x154, beta=1, alpha=1)
-                x153 = x155
+            x131 = 2
+            x133 = x128.shape
+            x133 = len(x133)
+            x134 = x133 == x131
+            if x134 :
+                classifier_6_weight = self.classifier_6_weight
+                x136 = fluid.layers.transpose(x=classifier_6_weight, perm=[1, 0])
+                classifier_6_bias = self.classifier_6_bias
+                x137 = fluid.layers.addmm(input=classifier_6_bias, x=x128, y=x136, beta=1, alpha=1)
+                x135 = x137
             else:
-                x147 = self.x147
-                x157 = fluid.layers.transpose(x=x147, perm=[1, 0])
-                x158 = fluid.layers.matmul(x=x146, y=x157)
-                x159 = True
-                if x159 :
-                    x148 = self.x148
-                    x161 = x158 + 1 * x148
-                    x160 = x161
-                else:
-                    x160 = x158
-                x153 = x160
+                classifier_6_weight = self.classifier_6_weight
+                x138 = fluid.layers.transpose(x=classifier_6_weight, perm=[1, 0])
+                x139 = fluid.layers.matmul(x=x128, y=x138)
+                classifier_6_bias = self.classifier_6_bias
+                x140 = x139 + 1 * classifier_6_bias
+                x135 = x140
         """
 
         def gen_name(id):
@@ -117,38 +112,19 @@ class FcFuser(FuseBase):
             outputs=[gen_name(9)])
         if_layer1.inputs["input-1"] = "fc-input-0"
         pattern_block1.add_layer(
-            "prim.constant", inputs={}, outputs=[gen_name(10)], value=True)
-        pattern_block1.add_layer("prim.if", {'input': gen_name(10)},
-                                 [gen_name(11)])
-        if_layer2 = pattern_block1.layers[list(pattern_block1.layers.keys())[
-            -1]]
-        pattern_block1_block0 = PaddleGraph(if_layer2, graph_type="dygraph")
-        pattern_block1_block0.add_layer(
             "fluid.dygraph.base.to_variable",
             inputs={},
             outputs=[gen_name(12)],
             value="params[{}]".format(string(gen_name(12))))
-        pattern_block1_block0.add_layer(
+        pattern_block1.add_layer(
             "prim.add_",
             inputs={"x": gen_name(9),
                     "y": gen_name(12)},
             outputs=[gen_name(13)],
             alpha=1)
-        if_layer2.inputs["input-0"] = gen_name(9)
-        pattern_block1_block0.add_layer(
-            "prim.equal",
-            inputs={'input': gen_name(13)},
-            outputs=[gen_name(11)])
-        if_layer2.add_block(pattern_block1_block0)
-        pattern_block1_block1 = PaddleGraph(if_layer2, graph_type="dygraph")
-        pattern_block1_block1.add_layer(
-            "prim.equal", inputs={'input': gen_name(9)},
-            outputs=[gen_name(11)])
-        if_layer2.inputs["input-1"] = gen_name(9)
         pattern_block1.add_layer(
-            "prim.equal", inputs={'input': gen_name(11)},
+            "prim.equal", inputs={'input': gen_name(13)},
             outputs=[gen_name(4)])
-        if_layer2.add_block(pattern_block1_block1)
         if_layer1.add_block(pattern_block1)
         self.pattern.build(inputs={"input-0": "fc-input-0"})
 

x2paddle/optimizer/optimizer.py

@@ -19,11 +19,8 @@ from x2paddle.optimizer.pass_manager import PassManager
 class GraphOptimizer(object):
     def __init__(self):
         self.passes = [
-            "fc_fuse_pass",
-            #             "nn_adaptive_pool2d_fuse_pass",
-            #             "functional_adaptive_pool2d_fuse_pass",
-            #             "batchnorm2d_fuse_pass",
-            "constant_fuse_pass"
+            "fc_fuse_pass", "adaptive_pool2d_fuse_pass",
+            "batchnorm2d_fuse_pass", "constant_fuse_pass"
         ]
 
     def optimize(self, graph):