add roberta aten

x2paddle/convert.py

@@ -92,7 +92,7 @@ def arg_parser():
         "--input_shapes",
         "-is",
         action='append',
-        default=[],
+        default=None,
         help="define the inputs' shape")
 
     return parser
@@ -207,12 +207,15 @@ def pytorch2paddle(model_path, save_dir, input_shapes):
     graph_opt = GraphOptimizer()
     graph_opt.optimize(mapper.graph)
     print("Model optimized.")
-    real_input_shapes = list()
-    for shape in input_shapes:
-        sp = shape[1:-1].split(",")
-        for i, s in enumerate(sp):
-            sp[i] = int(s)
-        real_input_shapes.append(sp)
+    if input_shapes is not None:
+        real_input_shapes = list()
+        for shape in input_shapes:
+            sp = shape[1:-1].split(",")
+            for i, s in enumerate(sp):
+                sp[i] = int(s)
+            real_input_shapes.append(sp)
+    else:
+        real_input_shapes = None
     mapper.graph.gen_model(save_dir, real_input_shapes)
 
 

x2paddle/core/program.py

@@ -226,7 +226,7 @@ class PaddleGraph(object):
             indent=1)
         f.close()
 
-    def gen_model(self, save_dir, input_shapes):
+    def gen_model(self, save_dir, input_shapes=None):
         if not os.path.exists(save_dir):
             os.makedirs(save_dir)
         if self.graph_type == "static":
@@ -264,8 +264,9 @@ class PaddleGraph(object):
         else:
             self.gen_dygraph_code(save_dir)
             self.dump_dygraph_parameter(save_dir)
-
-#             self.dygraph2static(save_dir, input_shapes)  #[[None, 3, 224, 224]]
+            if input_shapes is not None:
+                # 如果input_shapes非空，则导出推理模型；其值类似[[None, 3, 224, 224]]
+                self.dygraph2static(save_dir, input_shapes)
 
     def dump_parameter(self, param_name, param, save_dir):
         if not os.path.exists(save_dir):
@@ -318,6 +319,8 @@ class PaddleGraph(object):
 
         update(self.layers)
         self.inputs = list(set(self.inputs))
+        if self.inputs is not None:
+            self.inputs.sort()
 
     def get_dygraph_outputs(self):
         for layer_id, layer in self.layers.items():
@@ -389,7 +392,8 @@ class PaddleGraph(object):
 
         for layer_id, layer in self.layers.items():
             if ("paddle.nn" in layer.kernel and "functional" not in layer.kernel
-                ) or layer.kernel == "fluid.dygraph.base.to_variable":
+                ) or layer.kernel == "fluid.dygraph.base.to_variable" or \
+               "paddle.fluid.dygraph" in layer.kernel:
                 line = "{}".format(
                     layer.outputs[0]
                 ) if layer.kernel == "fluid.dygraph.base.to_variable" and not layer.attrs[

x2paddle/op_mapper/pytorch2paddle/aten.py

@@ -765,6 +765,51 @@ def aten_cos(mapper, graph, node):
     return current_inputs, current_outputs
 
 
+def aten_cumsum(mapper, graph, node):
+    """ 构造与前一个元素累加的PaddleLayer。
+
+    TorchScript示例:
+        %56 : Tensor = aten::cumsum(%mask.1, %46, %48)
+        参数含义:
+        %56 (Tensor): 输出，累加后的结果。
+        %mask.1 (Tensor): 输入，需要累加的Tensor。
+        %46 (int): 累加的维度。
+        %48 (int/None): Tensor的类型。
+    """
+    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%mask.1
+    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%46
+    if inputs_name[1] in mapper.attrs:
+        layer_attrs["axis"] = mapper.attrs[inputs_name[1]]
+    else:
+        mapper._check_input(graph, inputs_node[1], inputs_name[1],
+                            current_outputs)
+        layer_inputs["axis"] = inputs_name[1]
+        current_inputs.append(inputs_name[1])
+    # 处理输入1，即%48，代表dtype
+    if mapper.attrs[inputs_name[2]] is None:
+        layer_attrs["dtype"] = None
+    else:
+        layer_attrs["dtype"] = dtype_dict[mapper.attrs[inputs_name[2]]]
+
+    graph.add_layer(
+        "paddle.cumsum",
+        inputs=layer_inputs,
+        outputs=layer_outputs,
+        **layer_attrs)
+    return current_inputs, current_outputs
+
+
 def aten_detach(mapper, graph, node):
     """ 构造返回一个新的Tensor，从当前计算图中分离下来的，但是仍指向原变量的存放位置的PaddleLayer。
 
@@ -990,6 +1035,8 @@ def aten_embedding(mapper, graph, node):
     # 处理输入0，即%57
     weights = mapper.pytorch_params[inputs_name[0]]
     mapper.paddle_params[embedding_name + ".weight"] = weights
+    #     layer_attrs["num_embeddings"] = weights.shape[0]
+    #     layer_attrs["embedding_dim"] = weights.shape[1]
     layer_attrs["size"] = weights.shape
     # 处理输入1，即%input_ids.1
     mapper._check_input(graph, inputs_node[1], inputs_name[1], current_outputs)
@@ -999,10 +1046,11 @@ def aten_embedding(mapper, graph, node):
     # 处理输入2，即%45
     layer_attrs["padding_idx"] = mapper.attrs[inputs_name[2]]
     # 处理输入4，即%46
+    #     layer_attrs["sparse"] = mapper.attrs[inputs_name[4]]
     layer_attrs["is_sparse"] = mapper.attrs[inputs_name[4]]
 
     graph.add_layer(
-        "paddle.nn.Embedding",
+        "paddle.fluid.dygraph.Embedding",
         inputs=layer_inputs,
         outputs=layer_outputs,
         **layer_attrs)
@@ -2885,6 +2933,44 @@ def aten_softplus(mapper, graph, node):
     return current_inputs, current_outputs
 
 
+def aten_stack(mapper, graph, node):
+    """ 构造堆叠Tensor的PaddleLayer。
+
+    TorchScript示例:
+        %x.222 : Tensor = aten::stack(%32, %7)
+        参数含义:
+        %x.222 (Tensor): 输出，堆叠后的结果。
+        %i.12 (Tensor): 需要堆叠的Tensor组成的Tensor。
+        %7 (int): 堆叠的轴。
+    """
+    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%13
+    mapper._check_input(graph, inputs_node[0], inputs_name[0], current_outputs)
+    layer_inputs["x"] = inputs_name[0]
+    # 获取当前节点输入的list
+    current_inputs = list(layer_inputs.values())
+    # 处理输入1，即%12
+    if inputs_name[1] in mapper.attrs:
+        layer_attrs["axis"] = mapper.attrs[inputs_name[1]]
+    else:
+        mapper._check_input(graph, inputs_node[1], inputs_name[1],
+                            current_outputs)
+        layer_inputs["axis"] = inputs_name[1]
+        current_inputs.append(inputs_name[1])
+    graph.add_layer(
+        "paddle.stack",
+        inputs=layer_inputs,
+        outputs=layer_outputs,
+        **layer_attrs)
+    return current_inputs, current_outputs
+
+
 def aten_sub(mapper, graph, node):
     """ 构造数值相减的PaddleLayer。
 
@@ -3017,44 +3103,6 @@ def aten_split(mapper, graph, node):
     return current_inputs, current_outputs
 
 
-def aten_stack(mapper, graph, node):
-    """ 构造堆叠Tensor的PaddleLayer。
-
-    TorchScript示例:
-        %x.222 : Tensor = aten::stack(%32, %7)
-        参数含义:
-        %x.222 (Tensor): 输出，堆叠后的结果。
-        %i.12 (Tensor): 需要堆叠的Tensor组成的Tensor。
-        %7 (int): 堆叠的轴。
-    """
-    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%13
-    mapper._check_input(graph, inputs_node[0], inputs_name[0], current_outputs)
-    layer_inputs["x"] = inputs_name[0]
-    # 获取当前节点输入的list
-    current_inputs = list(layer_inputs.values())
-    # 处理输入1，即%12
-    if inputs_name[1] in mapper.attrs:
-        layer_attrs["axis"] = mapper.attrs[inputs_name[1]]
-    else:
-        mapper._check_input(graph, inputs_node[1], inputs_name[1],
-                            current_outputs)
-        layer_inputs["axis"] = inputs_name[1]
-        current_inputs.append(inputs_name[1])
-    graph.add_layer(
-        "paddle.stack",
-        inputs=layer_inputs,
-        outputs=layer_outputs,
-        **layer_attrs)
-    return current_inputs, current_outputs
-
-
 def aten_transpose(mapper, graph, node):
     """ 构造矩阵转置的PaddleLayer。
 
@@ -3152,14 +3200,16 @@ def aten_to(mapper, graph, node):
     inputs_name, inputs_node = mapper._get_inputs_name(node)
     # 获取当前节点输出的list
     current_outputs = [output_name]
-    assert len(inputs_name) == 5, "Paddle only support converting the dtype!"
     # 处理输入0，即%13
     mapper._check_input(graph, inputs_node[0], inputs_name[0], current_outputs)
     layer_inputs["x"] = inputs_name[0]
     # 获取当前节点输入的list
     current_inputs = list(layer_inputs.values())
     # 处理输入1，即%12
-    layer_attrs["dtype"] = dtype_dict[mapper.attrs[inputs_name[1]]]
+    if len(inputs_name) == 6:
+        layer_attrs["dtype"] = dtype_dict[mapper.attrs[inputs_name[2]]]
+    else:
+        layer_attrs["dtype"] = dtype_dict[mapper.attrs[inputs_name[1]]]
 
     graph.add_layer(
         "fluid.layers.cast",
@@ -3169,6 +3219,41 @@ def aten_to(mapper, graph, node):
     return current_inputs, current_outputs
 
 
+def aten_type_as(mapper, graph, node):
+    """ 构造转换Tensor类型的PaddleLayer。
+
+    TorchScript示例:
+        %57 : Tensor = aten::type_as(%56, %mask.1)
+        参数含义:
+        %57 (Tensor): 输出，改变类型后的Tensor。
+        %56 (Tensor): 需要改变类型的Tensor。
+        %mask.1 (Tensor): 转换成与该Tensor相一致的类型。
+    """
+    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%56
+    mapper._check_input(graph, inputs_node[0], inputs_name[0], current_outputs)
+    layer_inputs["x"] = inputs_name[0]
+    # 获取当前节点输入的list
+    current_inputs = list(layer_inputs.values())
+    # 处理输入0，即%mask.1
+    mapper._check_input(graph, inputs_node[1], inputs_name[1], current_outputs)
+    graph.add_layer(
+        "prim.type",
+        inputs={"input": inputs_name[1]},
+        outputs=[inputs_name[1] + "_type"])
+    layer_inputs["dtype"] = inputs_name[1] + "_type"
+    current_inputs.append(inputs_name[1])
+
+    graph.add_layer(
+        "fluid.layers.cast", inputs=layer_inputs, outputs=layer_outputs)
+    return current_inputs, current_outputs
+
+
 def aten_unsqueeze(mapper, graph, node):
     """ 构造插入维度的PaddleLayer。
 

x2paddle/op_mapper/pytorch2paddle/prim.py

@@ -29,7 +29,6 @@ def prim_Constant(mapper, graph, node):
     output = list(node.outputs())[0]
     value = output.toIValue()
     output_type = output.type()
-    mapper.attrs[output_name] = value
     if isinstance(value, str):
         value = string(value)
     if str(output_type) == "Tensor":
@@ -45,6 +44,7 @@ def prim_Constant(mapper, graph, node):
     if "9223372036854775807" in str(value):
         import math
         value = int(math.pow(2, 31) - 1)
+    mapper.attrs[output_name] = value
     graph.add_layer(
         "prim.constant", inputs={}, outputs=[output_name], value=value)
     return [], [output_name]

x2paddle/optimizer/fusion/adaptive_pool2d_fuser.py

@@ -25,28 +25,25 @@ class AdaptivePool2dFuser(FuseBase):
     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 = fluid.layers.shape(x71)
-            x78 = len(x73)
-            x80 = x78 <= x79
-            if x80 :
-                raise RaiseException(x75)
-            x83 = []
-            x85 = x73[x84: x76: x77]
-            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')
+            x68 = fluid.layers.shape(input=x60)
+            x69 = len(x68)
+            x70 = x69 <= 2
+            if x70 :
+                raise RaiseException('Exception')
+            x73 = []
+            x74 = x68[-2: 2147483647: 1]
+            x75 = len(x74)
+            x76 = [2, x75]
+            x77 = min(x76)
+            for _x79 in range(x77):
+                x80 = [6, 6][_x79]
+                x73.append(x80)
+            x81 = fluid.layers.adaptive_pool2d(input=x60, pool_size=x73, 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(
             "fluid.layers.shape",
             inputs={'input': "pool-input-0"},
@@ -54,38 +51,33 @@ class AdaptivePool2dFuser(FuseBase):
         self.pattern.add_layer(
             "prim.len", inputs={"input": gen_name(1)}, outputs=[gen_name(6)])
         self.pattern.add_layer(
-            "prim.le",
-            inputs={"x": gen_name(6),
-                    "y": "pool-input-1"},
-            outputs=[gen_name(8)])
+            "prim.le", inputs={"x": gen_name(6)}, outputs=[gen_name(8)], y=2)
         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": "pool-input-6"},
-            outputs=[gen_name(9)])
-        if_layer.inputs["input-0"] = "pool-input-6"
+            inputs={},
+            outputs=[gen_name(9)],
+            input="Exception")
         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.slice",
-            inputs={
-                "input": gen_name(1),
-                "start": "pool-input-2",
-                "end": "pool-input-3",
-                "step": "pool-input-4"
-            },
-            outputs=[gen_name(12)])
+            inputs={"input": gen_name(1), },
+            outputs=[gen_name(12)],
+            start=-1,
+            end=100,
+            step=1)
         self.pattern.add_layer(
             "prim.len", inputs={"input": gen_name(12)}, outputs=[gen_name(14)])
         self.pattern.add_layer(
             "prim.list",
-            inputs={"input0": "pool-input-4",
-                    "input1": gen_name(14)},
-            outputs=[gen_name(15)])
+            inputs={"input1": gen_name(14)},
+            outputs=[gen_name(15)],
+            input0=2)
         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)},
@@ -94,16 +86,15 @@ class AdaptivePool2dFuser(FuseBase):
         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)])
+            inputs={"index": gen_name(18)},
+            outputs=[gen_name(19)],
+            list=[6, 6])
         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.inputs["input-0"] = gen_name(10)
         loop_layer.add_block(pattern_block)
         pool_attrs = {'pool_type': string("avg")}
         self.pattern.add_layer(
@@ -112,15 +103,7 @@ class AdaptivePool2dFuser(FuseBase):
                     "pool_size": gen_name(10)},
             outputs=[gen_name(21)],
             **pool_attrs)
-        self.pattern.build(inputs={
-            "input-0": "pool-input-0",
-            "input-1": "pool-input-1",
-            "input-2": "pool-input-2",
-            "input-3": "pool-input-3",
-            "input-4": "pool-input-4",
-            "input-5": "pool-input-5",
-            "input-6": "pool-input-6"
-        })
+        self.pattern.build(inputs={"input-0": "pool-input-0", })
 
     def insert_new_layer(self, graph, parameters, matches):
         parameters = graph.parameters
@@ -131,9 +114,9 @@ class AdaptivePool2dFuser(FuseBase):
 
     def gen_new_layer(self, parameters, matches):
         layers_id = list(matches.keys())
+        layer = matches[layers_id[11]]
+        pool_size = layer.attrs["list"]
         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]

x2paddle/optimizer/optimizer.py

@@ -19,14 +19,10 @@ from x2paddle.optimizer.pass_manager import PassManager
 class GraphOptimizer(object):
     def __init__(self):
         self.passes = [
-            "constant_fuse_pass",
-            "batchnorm2d_fuse_pass",
-            "interpolate_bilinear_fuse_pass",
-            "fc_fuse_pass",
-            #             "interpolate_bilinear_fuse_pass",
-            #             "fc_fuse_pass",
-            #             "adaptive_pool2d_fuse_pass", "batchnorm2d_fuse_pass",
-            #             "constant_fuse_pass", "reshape_fuse_pass", "dropout_fuse_pass"
+            "constant_fuse_pass", "batchnorm2d_fuse_pass",
+            "interpolate_bilinear_fuse_pass", "fc_fuse_pass",
+            "adaptive_pool2d_fuse_pass", "reshape_fuse_pass",
+            "dropout_fuse_pass"
         ]
 
     def optimize(self, graph):