Merge remote-tracking branch 'upstream/develop' into fixed_Misspell

docs/inference_model_convertor/op_list.md

@@ -117,7 +117,7 @@ Aten:
 | 125 | aten::complex | 126 | aten::real | 127 | aten::imag | 128 | aten::fft\_rfftn |
 | 129 | aten::fft\_irfftn | 130 | aten::hardsigmoid | 131 | aten::hardswish | 132 | aten::linear |
 | 133 | aten::rsqrt | 134 | aten::replication\_pad1d | 135 | aten::full | 136 | aten::group\_norm |
-| 137 | aten::argmax | 138 | aten::copy |  |  |  |  |
+| 137 | aten::argmax | 138 | aten::copy | 139 | aten::upsample\_trilinear3d |  |  |
 
 Prim:
 | 序号 | OP | 序号 | OP | 序号 | OP | 序号 | OP |

x2paddle/__init__.py

-__version__ = "1.3.6"
+__version__ = "1.3.7"
 
 from .core.program import PaddleGraph
 

x2paddle/convert.py

@@ -192,6 +192,14 @@ def tf2paddle(model_path,
             ConverterCheck(
                 task="TensorFlow", time_info=time_info,
                 lite_state="Success").start()
+    # for convert survey
+    logging.info("================================================")
+    logging.info("")
+    logging.info(
+        "Model Convertd! Fill this survey to help X2Paddle better, https://iwenjuan.baidu.com/?code=npyd51 "
+    )
+    logging.info("")
+    logging.info("================================================")
 
 
 def caffe2paddle(proto_file,
@@ -240,6 +248,14 @@ def caffe2paddle(proto_file,
         if not disable_feedback:
             ConverterCheck(
                 task="Caffe", time_info=time_info, lite_state="Success").start()
+    # for convert survey
+    logging.info("================================================")
+    logging.info("")
+    logging.info(
+        "Model Convertd! Fill this survey to help X2Paddle better, https://iwenjuan.baidu.com/?code=npyd51 "
+    )
+    logging.info("")
+    logging.info("================================================")
 
 
 def onnx2paddle(model_path,
@@ -293,6 +309,14 @@ def onnx2paddle(model_path,
         if not disable_feedback:
             ConverterCheck(
                 task="ONNX", time_info=time_info, lite_state="Success").start()
+    # for convert survey
+    logging.info("================================================")
+    logging.info("")
+    logging.info(
+        "Model Convertd! Fill this survey to help X2Paddle better, https://iwenjuan.baidu.com/?code=npyd51 "
+    )
+    logging.info("")
+    logging.info("================================================")
 
 
 def pytorch2paddle(module,
@@ -364,6 +388,14 @@ def pytorch2paddle(module,
             ConverterCheck(
                 task="PyTorch", time_info=time_info,
                 lite_state="Success").start()
+    # for convert survey
+    logging.info("================================================")
+    logging.info("")
+    logging.info(
+        "Model Convertd! Fill this survey to help X2Paddle better, https://iwenjuan.baidu.com/?code=npyd51 "
+    )
+    logging.info("")
+    logging.info("================================================")
 
 
 def main():

x2paddle/op_mapper/caffe2paddle/caffe_custom_layer/detectionoutput.py

@@ -34,7 +34,7 @@ class DetectionOutput(object):
         pbv = priorbox_list[1]
         pb = paddle.reshape(x=pb, shape=[-1, 4])
         pbv = paddle.reshape(x=pbv, shape=[-1, 4])
-        pb_dim = fluid.layers.shape(pb)[0]
+        pb_dim = paddle.shape(pb)[0]
         loc = paddle.reshape(x0, shape=[-1, pb_dim, 4])
         conf_flatten = paddle.reshape(x1, shape=[0, pb_dim, -1])
         out = fluid.layers.detection_output(

x2paddle/op_mapper/caffe2paddle/caffe_custom_layer/normalize.py

@@ -13,7 +13,6 @@
 # limitations under the License.
 
 import paddle
-import paddle.fluid as fluid
 
 
 class Normalize(object):
@@ -21,7 +20,7 @@ class Normalize(object):
         self.axis = axis
 
     def __call__(self, x, param):
-        l2_norm = fluid.layers.l2_normalize(x=x, axis=1)
+        l2_norm = paddle.norm(x=x, p=2, axis=1, keepdim=True)
         param = paddle.reshape(param, [param.shape[-1]])
         perm = list(range(len(l2_norm.shape)))
         perm.pop(self.axis)

x2paddle/op_mapper/caffe2paddle/caffe_custom_layer/priorbox.py

@@ -13,7 +13,87 @@
 # limitations under the License.
 
 import paddle
-import paddle.fluid as fluid
+from paddle import _C_ops
+from paddle import in_dynamic_mode
+from paddle.common_ops_import import Variable, LayerHelper, check_variable_and_dtype, check_type, check_dtype
+
+
+@paddle.jit.not_to_static
+def prior_box(input,
+              image,
+              min_sizes,
+              max_sizes=None,
+              aspect_ratios=[1.],
+              variance=[0.1, 0.1, 0.2, 0.2],
+              flip=False,
+              clip=False,
+              steps=[0.0, 0.0],
+              offset=0.5,
+              min_max_aspect_ratios_order=False,
+              name=None):
+    helper = LayerHelper("prior_box", **locals())
+    dtype = helper.input_dtype()
+    check_variable_and_dtype(
+        input, 'input', ['uint8', 'int8', 'float32', 'float64'], 'prior_box')
+
+    def _is_list_or_tuple_(data):
+        return (isinstance(data, list) or isinstance(data, tuple))
+
+    if not _is_list_or_tuple_(min_sizes):
+        min_sizes = [min_sizes]
+    if not _is_list_or_tuple_(aspect_ratios):
+        aspect_ratios = [aspect_ratios]
+    if not (_is_list_or_tuple_(steps) and len(steps) == 2):
+        raise ValueError('steps should be a list or tuple ',
+                         'with length 2, (step_width, step_height).')
+
+    min_sizes = list(map(float, min_sizes))
+    aspect_ratios = list(map(float, aspect_ratios))
+    steps = list(map(float, steps))
+
+    cur_max_sizes = None
+    if max_sizes is not None and len(max_sizes) > 0 and max_sizes[0] > 0:
+        if not _is_list_or_tuple_(max_sizes):
+            max_sizes = [max_sizes]
+        cur_max_sizes = max_sizes
+
+    if in_dynamic_mode():
+        attrs = ('min_sizes', min_sizes, 'aspect_ratios', aspect_ratios,
+                 'variances', variance, 'flip', flip, 'clip', clip, 'step_w',
+                 steps[0], 'step_h', steps[1], 'offset', offset,
+                 'min_max_aspect_ratios_order', min_max_aspect_ratios_order)
+        if cur_max_sizes is not None:
+            attrs += ('max_sizes', cur_max_sizes)
+        box, var = _C_ops.prior_box(input, image, *attrs)
+        return box, var
+    else:
+        attrs = {
+            'min_sizes': min_sizes,
+            'aspect_ratios': aspect_ratios,
+            'variances': variance,
+            'flip': flip,
+            'clip': clip,
+            'step_w': steps[0],
+            'step_h': steps[1],
+            'offset': offset,
+            'min_max_aspect_ratios_order': min_max_aspect_ratios_order
+        }
+
+        if cur_max_sizes is not None:
+            attrs['max_sizes'] = cur_max_sizes
+
+        box = helper.create_variable_for_type_inference(dtype)
+        var = helper.create_variable_for_type_inference(dtype)
+        helper.append_op(
+            type="prior_box",
+            inputs={"Input": input,
+                    "Image": image},
+            outputs={"Boxes": box,
+                     "Variances": var},
+            attrs=attrs, )
+        box.stop_gradient = True
+        var.stop_gradient = True
+        return box, var
 
 
 class PriorBox(object):
@@ -32,8 +112,7 @@ class PriorBox(object):
         }
 
     def __call__(self, x0, x1):
-        box, var = fluid.layers.prior_box(
-            input=x0, image=x1, **self.priorbox_layer_attrs)
+        box, var = prior_box(input=x0, image=x1, **self.priorbox_layer_attrs)
         box = paddle.reshape(x=box, shape=[1, 1, -1])
         var = paddle.reshape(x=var, shape=[1, 1, -1])
         out = paddle.concat(x=[box, var], axis=1)

x2paddle/op_mapper/caffe2paddle/caffe_custom_layer/roipooling.py

@@ -13,7 +13,51 @@
 # limitations under the License.
 
 import paddle
-import paddle.fluid as fluid
+from paddle import _C_ops
+from paddle import in_dynamic_mode
+from paddle.common_ops_import import Variable, LayerHelper, check_variable_and_dtype, check_type, check_dtype
+
+
+@paddle.jit.not_to_static
+def roi_pool(input,
+             rois,
+             pooled_height,
+             pooled_width,
+             spatial_scale=1.0,
+             rois_num=None,
+             name=None):
+    if in_dynamic_mode():
+        assert rois_num is not None, "rois_num should not be None in dygraph mode."
+        pool_out, argmaxes = _C_ops.roi_pool(
+            input, rois, rois_num, "pooled_height", pooled_height,
+            "pooled_width", pooled_width, "spatial_scale", spatial_scale)
+        return pool_out, argmaxes
+
+    else:
+        check_variable_and_dtype(input, 'input', ['float32'], 'roi_pool')
+        check_variable_and_dtype(rois, 'rois', ['float32'], 'roi_pool')
+        helper = LayerHelper('roi_pool', **locals())
+        dtype = helper.input_dtype()
+        pool_out = helper.create_variable_for_type_inference(dtype)
+        argmaxes = helper.create_variable_for_type_inference(dtype='int32')
+
+        inputs = {
+            "X": input,
+            "ROIs": rois,
+        }
+        if rois_num is not None:
+            inputs['RoisNum'] = rois_num
+        helper.append_op(
+            type="roi_pool",
+            inputs=inputs,
+            outputs={"Out": pool_out,
+                     "Argmax": argmaxes},
+            attrs={
+                "pooled_height": pooled_height,
+                "pooled_width": pooled_width,
+                "spatial_scale": spatial_scale
+            })
+        return pool_out, argmaxes
 
 
 class ROIPooling(object):
@@ -26,6 +70,5 @@ class ROIPooling(object):
 
     def __call__(self, x0, x1):
         slice_x1 = paddle.slice(input=x1, axes=[1], starts=[1], ends=[5])
-        out = fluid.layers.roi_pool(
-            input=x0, rois=slice_x1, **self.roipooling_layer_attrs)
+        out = roi_pool(input=x0, rois=slice_x1, **self.roipooling_layer_attrs)
         return out

x2paddle/op_mapper/caffe2paddle/caffe_custom_layer/select.py

@@ -13,7 +13,6 @@
 # limitations under the License.
 
 import paddle
-import paddle.fluid as fluid
 
 
 class Select(object):

x2paddle/op_mapper/caffe2paddle/caffe_op_mapper.py

@@ -429,13 +429,13 @@ class CaffeOpMapper():
         assert params.local_size % 2 == 1
         alpha = params.alpha / float(params.local_size)
         layer_attrs = {
-            "n": params.local_size,
-            "k": params.k,
+            "size": params.local_size,
             "alpha": alpha,
             "beta": params.beta,
+            "k": params.k,
         }
         self.paddle_graph.add_layer(
-            "paddle.fluid.layers.lrn",
+            "paddle.nn.LocalResponseNorm",
             inputs={"input": input.name},
             outputs=[node.layer_name],
             **layer_attrs)
@@ -1209,10 +1209,10 @@ class CaffeOpMapper():
         input = self.graph.get_input_node(node, idx=0, copy=True)
         params = node.layer.shuffle_channel_param
         self.paddle_graph.add_layer(
-            "paddle.fluid.layers.shuffle_channel",
+            "paddle.nn.functional.channel_shuffle",
             inputs={"x": input.name},
             outputs=[node.layer_name],
-            group=params.group)
+            groups=params.group)
 
     def Upsample(self, node):
         assert len(

x2paddle/op_mapper/onnx2paddle/onnx_custom_layer/__init__.py

@@ -18,3 +18,5 @@ from .pad_all_dim2 import PadAllDim2
 from .pad_all_dim4 import PadAllDim4
 from .pad_all_dim4_one_input import PadAllDim4WithOneInput
 from .nms import NMS
+from .roi_align import ROIAlign
+from .roi_pooling import ROIPooling

x2paddle/op_mapper/onnx2paddle/onnx_custom_layer/nms.py

@@ -13,9 +13,9 @@
 # limitations under the License.
 
 import paddle
-from paddle.fluid import core
-from paddle.fluid.framework import Variable, in_dygraph_mode
-from paddle.fluid.layer_helper import LayerHelper
+from paddle import _C_ops
+from paddle import in_dynamic_mode
+from paddle.common_ops_import import Variable, LayerHelper
 
 
 def multiclass_nms(bboxes,
@@ -33,12 +33,12 @@ def multiclass_nms(bboxes,
                    name=None):
     helper = LayerHelper('multiclass_nms3', **locals())
 
-    if in_dygraph_mode():
+    if in_dynamic_mode():
         attrs = ('background_label', background_label, 'score_threshold',
                  score_threshold, 'nms_top_k', nms_top_k, 'nms_threshold',
                  nms_threshold, 'keep_top_k', keep_top_k, 'nms_eta', nms_eta,
                  'normalized', normalized)
-        output, index, nms_rois_num = core.ops.multiclass_nms3(bboxes, scores,
+        output, index, nms_rois_num = _C_ops.multiclass_nms3(bboxes, scores,
                                                              rois_num, *attrs)
         if not return_index:
             index = None

x2paddle/op_mapper/onnx2paddle/onnx_custom_layer/roi_align.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 paddle
+from paddle import _C_ops
+from paddle import in_dynamic_mode
+from paddle.common_ops_import import Variable, LayerHelper, check_variable_and_dtype, check_type, check_dtype
+
+
+@paddle.jit.not_to_static
+def roi_align(input,
+              rois,
+              pooled_height,
+              pooled_width,
+              spatial_scale=1.0,
+              sampling_ratio=-1,
+              rois_num=None,
+              aligned=False,
+              name=None):
+    if in_dynamic_mode():
+        assert rois_num is not None, "rois_num should not be None in dygraph mode."
+        align_out = _C_ops.roi_align(
+            input, rois, rois_num, "pooled_height", pooled_height,
+            "pooled_width", pooled_width, "spatial_scale", spatial_scale,
+            "sampling_ratio", sampling_ratio, "aligned", aligned)
+        return align_out
+
+    else:
+        check_variable_and_dtype(input, 'input', ['float32', 'float64'],
+                                 'roi_align')
+        check_variable_and_dtype(rois, 'rois', ['float32', 'float64'],
+                                 'roi_align')
+        helper = LayerHelper('roi_align', **locals())
+        dtype = helper.input_dtype()
+        align_out = helper.create_variable_for_type_inference(dtype)
+        inputs = {
+            "X": input,
+            "ROIs": rois,
+        }
+        if rois_num is not None:
+            inputs['RoisNum'] = rois_num
+        helper.append_op(
+            type="roi_align",
+            inputs=inputs,
+            outputs={"Out": align_out},
+            attrs={
+                "pooled_height": pooled_height,
+                "pooled_width": pooled_width,
+                "spatial_scale": spatial_scale,
+                "sampling_ratio": sampling_ratio,
+                "aligned": aligned,
+            })
+        return align_out
+
+
+class ROIAlign(object):
+    def __init__(self, pooled_height, pooled_width, spatial_scale,
+                 sampling_ratio):
+        self.roialign_layer_attrs = {
+            "pooled_height": pooled_height,
+            "pooled_width": pooled_width,
+            "spatial_scale": spatial_scale,
+            'sampling_ratio': sampling_ratio,
+        }
+
+    def __call__(self, x0, x1, x2):
+        out = roi_align(
+            input=x0, rois=x1, rois_num=x2, **self.roialign_layer_attrs)
+        return out

x2paddle/op_mapper/onnx2paddle/onnx_custom_layer/roi_pooling.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 paddle
+from paddle import _C_ops
+from paddle import in_dynamic_mode
+from paddle.common_ops_import import Variable, LayerHelper, check_variable_and_dtype, check_type, check_dtype
+
+
+@paddle.jit.not_to_static
+def roi_pool(input,
+             rois,
+             pooled_height,
+             pooled_width,
+             spatial_scale=1.0,
+             rois_num=None,
+             name=None):
+    if in_dynamic_mode():
+        assert rois_num is not None, "rois_num should not be None in dygraph mode."
+        pool_out, argmaxes = _C_ops.roi_pool(
+            input, rois, rois_num, "pooled_height", pooled_height,
+            "pooled_width", pooled_width, "spatial_scale", spatial_scale)
+        return pool_out, argmaxes
+
+    else:
+        check_variable_and_dtype(input, 'input', ['float32'], 'roi_pool')
+        check_variable_and_dtype(rois, 'rois', ['float32'], 'roi_pool')
+        helper = LayerHelper('roi_pool', **locals())
+        dtype = helper.input_dtype()
+        pool_out = helper.create_variable_for_type_inference(dtype)
+        argmaxes = helper.create_variable_for_type_inference(dtype='int32')
+
+        inputs = {
+            "X": input,
+            "ROIs": rois,
+        }
+        if rois_num is not None:
+            inputs['RoisNum'] = rois_num
+        helper.append_op(
+            type="roi_pool",
+            inputs=inputs,
+            outputs={"Out": pool_out,
+                     "Argmax": argmaxes},
+            attrs={
+                "pooled_height": pooled_height,
+                "pooled_width": pooled_width,
+                "spatial_scale": spatial_scale
+            })
+        return pool_out, argmaxes
+
+
+class ROIPooling(object):
+    def __init__(self, pooled_height, pooled_width, spatial_scale):
+        self.roipooling_layer_attrs = {
+            "pooled_height": pooled_height,
+            "pooled_width": pooled_width,
+            "spatial_scale": spatial_scale
+        }
+
+    def __call__(self, x0, x1):
+        out = roi_pool(input=x0, rois=x1, **self.roipooling_layer_attrs)
+        return out

x2paddle/op_mapper/onnx2paddle/opset9/opset.py

@@ -262,6 +262,8 @@ class OpSet9():
         shape = node.out_shapes[0]
 
         if hasattr(node.weight, "shape") and len(node.weight.shape) == 0:
+            if node.weight == float('inf') or node.weight == float('-inf'):
+                node.weight = string(node.weight)
             self.paddle_graph.add_layer(
                 "paddle.full",
                 inputs={},
@@ -536,12 +538,14 @@ class OpSet9():
             'pooled_width': pooled_width,
             'spatial_scale': spatial_scale,
             'sampling_ratio': sampling_ratio,
-            'rois_num': val_rois_num,
         }
         self.paddle_graph.add_layer(
-            'paddle.fluid.layers.roi_align',
-            inputs={'input': val_x.name,
-                    'rois': val_rois.name},
+            'custom_layer:ROIAlign',
+            inputs={
+                'input': val_x.name,
+                'rois': val_rois.name,
+                'rois_num': val_rois_num
+            },
             outputs=[node.name],
             **layer_attrs)
 
@@ -558,7 +562,7 @@ class OpSet9():
             'spatial_scale': spatial_scale,
         }
         self.paddle_graph.add_layer(
-            'paddle.fluid.layers.roi_pool',
+            'custom_layer:ROIPooling',
             inputs={'input': val_x.name,
                     'rois': val_rois.name},
             outputs=[node.name],
@@ -792,6 +796,8 @@ class OpSet9():
         if len(value) == 1:
             value = value.tolist()
             value = value[0]
+            if value == float('inf') or value == float('-inf'):
+                value = string(value)
             self.paddle_graph.add_layer(
                 "paddle.full",
                 inputs={},
@@ -1093,6 +1099,12 @@ class OpSet9():
         val_x = self.graph.get_input_node(node, idx=0, copy=True)
         starts, ends, axes, steps = None, None, None, None
         layer_attrs = {}
+        if val_x.dtype == 'uint8':
+            self.paddle_graph.add_layer(
+                'paddle.cast',
+                inputs={"x": val_x.name},
+                outputs=[val_x.name],
+                dtype=string('int32'))
         if len(node.inputs) > 1:
             starts = self.graph.get_input_node(node, idx=1, copy=True)
             ends = self.graph.get_input_node(node, idx=2, copy=True)
@@ -1121,7 +1133,8 @@ class OpSet9():
                 starts_value = starts_value.copy()
                 ends_value = ends_value.copy()
                 for idx in range(len(ends_value)):
-                    if starts_value[idx] >= val_x.out_shapes[0][axes[
+                    if len(val_x.out_shapes[0]) != 0 and starts_value[
+                            idx] >= val_x.out_shapes[0][axes[
                                 idx]] and val_x.out_shapes[0][axes[idx]] > 0:
                         starts_value[idx] = val_x.out_shapes[0][axes[idx]] - 1
                         ends_value[idx] = val_x.out_shapes[0][axes[idx]]
@@ -1178,11 +1191,16 @@ class OpSet9():
                 inputs={"input": val_x.name},
                 outputs=[node.name],
                 **layer_attrs)
+        if val_x.dtype == 'uint8':
+            self.paddle_graph.add_layer(
+                'paddle.cast',
+                inputs={"x": node.name},
+                outputs=[node.name],
+                dtype=string('uint8'))
 
     @print_mapping_info
     def ConstantOfShape(self, node):
         val_shape = self.graph.get_input_node(node, idx=0, copy=True)
-        val_y = self.graph.get_node(node.layer.output[0], copy=True)
 
         value = node.get_attr('value')
         dtype = value.dtype
@@ -1191,6 +1209,8 @@ class OpSet9():
                                  'this is not supported')
         if len(value) == 1:
             value = value[0]
+            if value == float('inf') or value == float('-inf'):
+                value = string(value)
             layer_attrs = {'dtype': string(dtype), 'fill_value': value}
             self.paddle_graph.add_layer(
                 "paddle.full",
@@ -1550,11 +1570,15 @@ class OpSet9():
         val_x = self.graph.get_input_node(node, idx=0, copy=True)
         output_shape = val_x.out_shapes[0]
         axis = node.get_attr('axis', 1)
-        shape_list = [1, 1]
         if axis == 0:
-            for s in output_shape:
-                shape_list[1] *= s
+            self.paddle_graph.add_layer(
+                'paddle.reshape',
+                inputs={"x": val_x.name},
+                outputs=[node.name],
+                shape=[1, -1])
         else:
+            if len(output_shape) != 0:
+                shape_list = [1, 1]
                 for s in output_shape[:axis]:
                     shape_list[0] *= s
                 for s in output_shape[axis:]:
@@ -1564,6 +1588,19 @@ class OpSet9():
                     inputs={"x": val_x.name},
                     outputs=[node.name],
                     shape=shape_list)
+            else:
+                # flatten + reshape
+                self.paddle_graph.add_layer(
+                    "paddle.flatten",
+                    inputs={"input": val_x.name},
+                    outputs=[val_x.name + "_flatten"],
+                    start_axis=[0],
+                    stop_axis=[axis])
+                self.paddle_graph.add_layer(
+                    'paddle.reshape',
+                    inputs={'x': val_x.name + "_flatten"},
+                    outputs=[node.name],
+                    shape=[0, -1])
 
     @print_mapping_info
     def Gemm(self, node):
@@ -1790,7 +1827,11 @@ class OpSet9():
     def Squeeze(self, node):
         val_x = self.graph.get_input_node(node, idx=0, copy=True)
         axes = node.get_attr('axes')
-        if len(val_x.out_shapes[0]) == 1:
+        if axes is None:
+            axes_node = self.graph.get_input_node(node, idx=1, copy=True)
+            axes = _const_weight_or_none(axes_node, necessary=True)
+        # deal with scalar(0D) tensor
+        if len(val_x.out_shapes[0]) <= 1 and len(axes) == 1 and axes[0] == 0:
             self.paddle_graph.add_layer(
                 "paddle.cast",
                 inputs={"x": val_x.name},
@@ -1829,67 +1870,23 @@ class OpSet9():
         val_x = self.graph.get_input_node(node, idx=1, copy=True)
         val_y = self.graph.get_input_node(node, idx=2, copy=True)
 
-        not_condition = condition.name + '_not'
-        self.paddle_graph.add_layer(
-            "paddle.logical_not",
-            inputs={"x": condition.name},
-            outputs=[not_condition])
-        cast_not_condition = not_condition + '_cast'
-        self.paddle_graph.add_layer(
-            "paddle.cast",
-            inputs={"x": not_condition},
-            outputs=[cast_not_condition],
-            dtype=string(val_x.dtype))
-        cast_condition = condition.name + '_cast'
         self.paddle_graph.add_layer(
-            "paddle.cast",
-            inputs={"x": condition.name},
-            outputs=[cast_condition],
-            dtype=string(val_x.dtype))
-        mul_val_x = val_x.name + '_mul'
-        self.paddle_graph.add_layer(
-            "paddle.multiply",
-            inputs={'x': val_x.name,
-                    'y': cast_condition},
-            outputs=[mul_val_x])
-        mul_val_y = val_y.name + '_mul'
-        self.paddle_graph.add_layer(
-            "paddle.multiply",
-            inputs={'x': val_y.name,
-                    'y': cast_not_condition},
-            outputs=[mul_val_y])
-
-        self.paddle_graph.add_layer(
-            "paddle.add",
-            inputs={'x': mul_val_x,
-                    'y': mul_val_y},
+            "paddle.where",
+            inputs={
+                'condition': condition.name,
+                'x': val_x.name,
+                'y': val_y.name
+            },
             outputs=[node.name])
 
     @print_mapping_info
     def NonZero(self, node):
         val_x = self.graph.get_input_node(node, idx=0, copy=True)
-        val_x_dim = len(val_x.out_shapes[0])
-        if val_x_dim == 1:
         self.paddle_graph.add_layer(
             "paddle.nonzero",
             inputs={"x": val_x.name},
-                outputs=[val_x.name])
-            self.paddle_graph.add_layer(
-                "paddle.transpose",
-                inputs={"x": val_x.name},
-                outputs=[node.layer_name],
-                perm=[1, 0])
-        if val_x_dim > 1:
-            self.paddle_graph.add_layer(
-                "paddle.nonzero",
-                inputs={"x": val_x.name},
-                outputs=[val_x.name])
-            self.paddle_graph.add_layer(
-                "paddle.split",
-                inputs={"x": val_x.name},
             outputs=[val_x.name],
-                num_or_sections=1,
-                axis=val_x_dim)
+            as_tuple=True)
         self.paddle_graph.add_layer(
             "paddle.concat", inputs={"x": val_x.name}, outputs=[node.name])
 
@@ -2565,18 +2562,33 @@ class OpSet9():
     def TopK(self, node):
         val_x = self.graph.get_input_node(node, idx=0, copy=True)
         val_k = self.graph.get_input_node(node, idx=1, copy=True)
-        if val_k.dtype != "int32":
-            self.paddle_graph.add_layer(
-                "paddle.cast",
-                inputs={"x": val_k.name},
-                outputs=[val_k.name],
-                dtype=string('int32'))
         layer_attrs = dict()
         layer_attrs["axis"] = node.get_attr('axis', -1)
         layer_attrs["largest"] = True if node.get_attr('largest',
                                                        1) == 1 else False
         layer_attrs["sorted"] = True if node.get_attr('sorted',
                                                       1) == 1 else False
+        k = _const_weight_or_none(val_k)
+        if isinstance(k, (list, tuple, np.ndarray)):
+            k = k[0]
+        # If k can get the value directly, it is used as an attribute; otherwise it is used as an input tensor
+        if k is not None:
+            layer_attrs["k"] = k
+            self.paddle_graph.add_layer(
+                "paddle.topk",
+                inputs={"x": val_x.name},
+                outputs=[
+                    "{}_p{}".format(node.layer_name, 0),
+                    "{}_p{}".format(node.layer_name, 1)
+                ],
+                **layer_attrs)
+        else:
+            if val_k.dtype != "int32":
+                self.paddle_graph.add_layer(
+                    "paddle.cast",
+                    inputs={"x": val_k.name},
+                    outputs=[val_k.name],
+                    dtype=string('int32'))
             self.paddle_graph.add_layer(
                 "paddle.topk",
                 inputs={"x": val_x.name,

x2paddle/op_mapper/prim2code.py

@@ -612,7 +612,7 @@ def prim_shape_dim(layer,
                    forward_func=[],
                    layer_id=None,
                    different_attrs=None):
-    line = "{} = fluid.layers.shape({})[{}]".format(
+    line = "{} = paddle.shape({})[{}]".format(
         layer.outputs[0],
         get_value(layer, "input", different_attrs),
         get_value(layer, "dim", different_attrs))

x2paddle/op_mapper/pytorch2paddle/aten.py

@@ -1189,129 +1189,66 @@ def aten___contains__(mapper, graph, node):
 
 
 def aten_constant_pad_nd(mapper, graph, node):
-    """ 构造填充固定值的PaddleLayer。
-    TorchScript示例:
+    """
+    TorchScript Code:
         %58 : Tensor = aten::constant_pad_nd(%input1.24, %4876, %42)
-        参数含义:
-        %58 (Tensor): 输出，填充后的Tensor。
-        %input1.24 (Tensor): 需要填充的Tensor。
-        %4876 (list): 填充大小。
-        %42 (-): 填充值。
+        Parameter meaning:
+        %58 (Tensor): Output Tensor
+        %input1.24 (Tensor): Input Tensor
+        %4876 (list): pad
+        %42 (-): value
     """
     scope_name = mapper.normalize_scope_name(node)
     op_name = name_generator("pad", mapper.nn_name2id)
     output_name = mapper._get_outputs_name(node)[0]
-    layer_outputs = [op_name, 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
+    # Output list
     current_outputs = [output_name]
-    # 处理输入0，即%input1.24
+    # process Input Tensor
     mapper._check_input(graph, inputs_node[0], inputs_name[0], current_outputs,
                         scope_name)
-    layer_inputs["input"] = inputs_name[0]
-    # 处理输入1，即%4876
-    is_padding_tensor = False
+    # process pad
+    padding_attr = None
     if inputs_name[1] in mapper.attrs:
-        layer_attrs["padding"] = mapper.attrs[inputs_name[1]]
+        padding_attr = mapper.attrs[inputs_name[1]]
     else:
         mapper._check_input(graph, inputs_node[1], inputs_name[1],
                             current_outputs, scope_name)
         layer_inputs["pad"] = inputs_name[1]
-        is_padding_tensor = True
-    # 获取当前节点输入的list
-    current_inputs = list(layer_inputs.values())
-    # 处理输入2，即%42
+    # process value
     layer_attrs["value"] = mapper.attrs[inputs_name[2]]
 
-    if not is_padding_tensor:
-        graph.add_layer(
-            "prim.shape",
-            inputs={"input": inputs_name[0]},
-            outputs=[inputs_name[0] + "_shape"],
-            scope_name=scope_name)
-        graph.add_layer(
-            "prim.len",
-            inputs={"input": inputs_name[0] + "_shape"},
-            outputs=[inputs_name[0] + "_len"],
-            scope_name=scope_name)
-
-    def add_pad_layers(kernel, dim):
-        graph.add_layer(
-            "prim.ne",
-            inputs={"x": inputs_name[0] + "_len"},
-            outputs=[inputs_name[0] + "_cond"],
-            scope_name=scope_name,
-            y=dim)
+    if padding_attr is not None:
+        layer_inputs["x"] = inputs_name[0]
+        kernel_name = "paddle.nn.functional.pad"
+        if len(padding_attr) == 2:
+            layer_attrs["pad"] = [0, 0, 0, 0, 0, 0] + padding_attr
+        elif len(padding_attr) == 4:
+            layer_inputs["x"] = inputs_name[0]
+            layer_attrs["pad"] = [0, 0, 0, 0] + padding_attr
+        elif len(padding_attr) == 6:
+            layer_inputs["x"] = inputs_name[0]
+            layer_attrs["pad"] = [0, 0] + padding_attr
+        else:
+            layer_inputs["x"] = inputs_name[0]
+            layer_attrs["pad"] = padding_attr
         graph.add_layer(
-            "prim.if", {'input': inputs_name[0] + "_cond"},
-            outputs=[inputs_name[0] + "_if", output_name],
-            scope_name=scope_name)
-        if_layer = graph.layers[list(graph.layers.keys())[-1]]
-        block = PaddleGraph(source_type="pytorch", parent_layer=if_layer)
-        block.add_layer(
-            "prim.sub",
-            inputs={"y": inputs_name[0] + "_len"},
-            outputs=[inputs_name[0] + "_len0"],
-            scope_name=scope_name,
-            alpha=1.0,
-            x=dim)
-        block.add_layer(
-            "prim.len2list",
-            inputs={"len": inputs_name[0] + "_len0"},
-            outputs=[inputs_name[0] + "_list"],
-            scope_name=scope_name)
-        block.add_layer(
-            "paddle.unsqueeze",
-            inputs={"x": inputs_name[0],
-                    "axis": inputs_name[0] + "_list"},
-            outputs=[inputs_name[0] + "_var"],
-            scope_name=scope_name)
-        block.add_layer(
-            kernel,
-            inputs={"input": inputs_name[0] + "_var"},
-            outputs=copy.deepcopy(layer_outputs),
-            scope_name=scope_name,
-            **layer_attrs)
-        block.add_layer(
-            "paddle.squeeze",
-            inputs={"x": output_name,
-                    "axis": inputs_name[0] + "_list"},
-            outputs=[output_name],
-            scope_name=scope_name)
-        if_layer.add_block(block)
-        block = PaddleGraph(source_type="pytorch", parent_layer=if_layer)
-        layer_inputs["input"] = inputs_name[0]
-        block.add_layer(
-            kernel,
+            kernel_name,
             inputs=layer_inputs,
-            outputs=layer_outputs,
+            outputs=[output_name],
             scope_name=scope_name,
             **layer_attrs)
-        if_layer.add_block(block)
-        if_layer.inputs["input-0"] = inputs_name[0]
-        if_layer.inputs["input-1"] = inputs_name[0] + "_len"
-
-    if not is_padding_tensor:
-        if len(layer_attrs["padding"]) == 2:
-            layer_outputs[0] = layer_outputs[0].replace("pad", "pad1d")
-            add_pad_layers("paddle.nn.Pad1D", 3)
-        elif len(layer_attrs["padding"]) == 4:
-            layer_outputs[0] = layer_outputs[0].replace("pad", "pad2d")
-            add_pad_layers("paddle.nn.Pad2D", 4)
-        elif len(layer_attrs["padding"]) == 6:
-            layer_outputs[0] = layer_outputs[0].replace("pad", "pad3d")
-            add_pad_layers("paddle.nn.Pad3D", 5)
-        else:
-            raise Exception("The lenght of padding list must be 2, 4 or 6!")
     else:
+        layer_inputs["input"] = inputs_name[0]
         graph.add_layer(
             "custom_layer:Pad",
             inputs=layer_inputs,
             outputs=[output_name],
             scope_name=scope_name,
             **layer_attrs)
+    current_inputs = list(layer_inputs.values())
     return current_inputs, current_outputs
 
 
@@ -6025,7 +5962,7 @@ def aten_upsample_bilinear2d(mapper, graph, node):
             inputs={"input": inputs_name[1]},
             outputs=[inputs_name[1] + "_isinstance"],
             scope_name=scope_name,
-            cls="paddle.fluid.Variable")
+            cls="paddle.static.Variable")
         # TODO(syf): paddle.Variable
         graph.add_layer(
             "prim.if", {"input": inputs_name[1] + "_isinstance"},
@@ -6065,6 +6002,66 @@ def aten_upsample_bilinear2d(mapper, graph, node):
     return current_inputs, current_outputs
 
 
+def aten_upsample_trilinear3d(mapper, graph, node):
+    """
+    TorchScript Code:
+        %4997 : Tensor = aten::upsample_trilinear3d(%x.13, %4963, %5421, %4995)
+        Parameter meaning:
+        %4997 (Tensor): Output Tensor
+        %x.13 (Tensor): Input Tensor
+        %4963 (list): output_size
+        %5421 (bool): align_corners
+        %4995 (float): scale_factors
+    """
+    scope_name = mapper.normalize_scope_name(node)
+    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)
+    # Output list
+    current_outputs = [output_name]
+    # process Input Tensor
+    mapper._check_input(graph, inputs_node[0], inputs_name[0], current_outputs,
+                        scope_name)
+    layer_inputs["x"] = inputs_name[0]
+    current_inputs = list(layer_inputs.values())
+    # process output_size
+    if inputs_name[1] in mapper.attrs:
+        layer_attrs["size"] = mapper.attrs[inputs_name[1]]
+    else:
+        mapper._check_input(graph, inputs_node[1], inputs_name[1],
+                            current_outputs, scope_name)
+        layer_inputs["size"] = inputs_name[1]
+        current_inputs.append(inputs_name[1])
+    # process align_corners
+    if inputs_name[2] in mapper.attrs:
+        layer_attrs["align_corners"] = mapper.attrs[inputs_name[2]]
+    else:
+        mapper._check_input(graph, inputs_node[2], inputs_name[2],
+                            current_outputs, scope_name)
+        layer_inputs["align_corners"] = inputs_name[2]
+        current_inputs.append(inputs_name[2])
+    # process scale_factor
+    if inputs_name[3] in mapper.attrs:
+        layer_attrs["scale_factor"] = mapper.attrs[inputs_name[3]]
+    else:
+        mapper._check_input(graph, inputs_node[3], inputs_name[3],
+                            current_outputs, scope_name)
+        layer_inputs["scale_factor"] = inputs_name[3]
+        current_inputs.append(inputs_name[3])
+    layer_attrs["align_mode"] = 0
+    layer_attrs["mode"] = string("trilinear")
+    layer_attrs["data_format"] = string("NCDHW")
+    graph.add_layer(
+        "paddle.nn.functional.interpolate",
+        inputs=layer_inputs,
+        outputs=layer_outputs,
+        scope_name=scope_name,
+        **layer_attrs)
+    return current_inputs, current_outputs
+
+
 def aten_upsample_nearest2d(mapper, graph, node):
     """ 构造使用nearest上采样的PaddleLayer。
     TorchScript示例:
@@ -6103,7 +6100,7 @@ def aten_upsample_nearest2d(mapper, graph, node):
             inputs={"input": inputs_name[1]},
             outputs=[inputs_name[1] + "_isinstance"],
             scope_name=scope_name,
-            cls="paddle.fluid.Variable")
+            cls="paddle.static.Variable")
         # TODO(syf): paddle.Variable
         graph.add_layer(
             "prim.if", {"input": inputs_name[1] + "_isinstance"},

x2paddle/op_mapper/pytorch2paddle/pytorch_custom_layer/instance_norm.py

@@ -14,7 +14,7 @@
 
 import paddle
 from paddle.nn.functional import instance_norm
-from paddle.fluid.initializer import Constant
+from paddle.nn.initializer import Constant
 
 
 class InstanceNorm(paddle.nn.Layer):

x2paddle/optimizer/fusion/__init__.py

@@ -40,3 +40,7 @@ from .trace_fc_fuser import TraceFcFuser
 from .trace_fc_fuse_pass import TraceFcFusePass
 from .onnx_layernorm_fuser import LayerNormFuser
 from .onnx_layernorm_fuse_pass import LayerNormFusePass
+from .onnx_gelu_fuser import GeluFuser
+from .onnx_gelu_fuse_pass import GeluFusePass
+from .replace_div_to_scale import Div2Scale
+from .replace_div_to_scale_pass import Div2ScalePass

x2paddle/optimizer/fusion/interpolate_bilinear_fuser.py

@@ -46,7 +46,7 @@ class InterpolateBilinearFuser(FuseBase):
                 if x2271 :
                     x2274 = x2197[0]
                     x2275 = x2197[1]
-                    x2233_isinstance = isinstance(x2233, paddle.fluid.Variable)
+                    x2233_isinstance = isinstance(x2233, paddle.static.Variable)
                     if x2233_isinstance :
                         x2233 = x2233.numpy().tolist()
                     x2276 = paddle.nn.functional.interpolate(x=x2181, size=x2233, scale_factor=x2274, align_corners=False, align_mode=0, mode='bilinear')
@@ -146,7 +146,7 @@ class InterpolateBilinearFuser(FuseBase):
             "prim.isinstance",
             inputs={"input": "interpolate-input-3"},
             outputs=["interpolate-input-0_isinstance"],
-            cls="paddle.fluid.Variable")
+            cls="paddle.static.Variable")
         pattern_block_block.add_layer(
             "prim.if", {"input": "interpolate-input-0_isinstance"},
             outputs=["interpolate-input-0_if1"])

x2paddle/optimizer/fusion/onnx_gelu_fuse_pass.py

+#   Copyright (c) 2022  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 GeluFuser
+from x2paddle.optimizer.pass_manager import pass_register
+
+
+@pass_register
+class GeluFusePass(Pass):
+    name = "onnx_gelu_fuse_pass"
+
+    def __init__(self):
+        Pass.__init__(self)
+
+    def apply(self, graph):
+        fuser = GeluFuser()
+        fuser.operate(graph, match_kind="edge")
+
+
+# register gelu pass
+onnx_gelu_fuse_pass = GeluFusePass()

x2paddle/optimizer/fusion/onnx_gelu_fuser.py

+#   Copyright (c) 2022  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 copy
+import numpy as np
+from collections import OrderedDict
+from x2paddle.optimizer.pattern_matcher import FuseBase
+from x2paddle.core.program import PaddleGraph, PaddleLayer
+from x2paddle.core.util import *
+
+
+class GeluFuser(FuseBase):
+    def __init__(self):
+        super(GeluFuser, self).__init__()
+
+    def build_pattern(self):
+        """
+        code describe:
+            x2paddle_332 = paddle.full(dtype='float32', shape=[1], fill_value=1.4142135381698608)
+            x2paddle_335 = paddle.full(dtype='float32', shape=[1], fill_value=1.0)
+            x2paddle_338 = paddle.full(dtype='float32', shape=[1], fill_value=0.5)
+            x2paddle_333 = paddle.divide(x=x2paddle_331, y=x2paddle_332)
+            x2paddle_334 = paddle.erf(x=x2paddle_333)
+            x2paddle_336 = paddle.add(x=x2paddle_334, y=x2paddle_335)
+            x2paddle_337 = paddle.multiply(x=x2paddle_331, y=x2paddle_336)
+            x2paddle_339 = paddle.multiply(x=x2paddle_337, y=x2paddle_338)
+        """
+
+        def gen_name(id):
+            return "x" + str(id)
+
+        self.pattern.add_layer(
+            "paddle.full",
+            inputs={},
+            outputs=[gen_name(0)],
+            shape=[1],
+            fill_value=1.4142135381698608)
+        self.pattern.add_layer(
+            "paddle.full",
+            inputs={},
+            outputs=[gen_name(1)],
+            shape=[1],
+            fill_value=1.0)
+        self.pattern.add_layer(
+            "paddle.full",
+            inputs={},
+            outputs=[gen_name(2)],
+            shape=[1],
+            fill_value=0.5)
+        self.pattern.add_layer(
+            "paddle.divide",
+            inputs={"x": "gelu-input-0",
+                    "y": gen_name(0)},
+            outputs=[gen_name(3)])
+        self.pattern.add_layer(
+            "paddle.erf", inputs={"x": gen_name(3)}, outputs=[gen_name(4)])
+        self.pattern.add_layer(
+            "paddle.add",
+            inputs={"x": gen_name(4),
+                    "y": gen_name(1)},
+            outputs=[gen_name(5)])
+        self.pattern.add_layer(
+            "paddle.multiply",
+            inputs={"x": "gelu-input-0",
+                    "y": gen_name(5)},
+            outputs=[gen_name(6)])
+        self.pattern.add_layer(
+            "paddle.multiply",
+            inputs={"x": gen_name(6),
+                    "y": gen_name(2)},
+            outputs=[gen_name(7)])
+        self.pattern.build(inputs={"input-0": "gelu-input-0", })
+
+    def insert_new_layer(self, graph, parameters, matches):
+        new_layer, new_layer_id = self.gen_new_layer(parameters, matches)
+        graph.layers[new_layer_id] = new_layer
+        matches.pop(new_layer_id)
+
+    def gen_new_layer(self, parameters, matches):
+        layer_id_list = list(matches.keys())
+        layer_id_list.sort(key=int)
+        layer_inputs = list()
+        layer_inputs_ids = list()
+        fill_value_list = list()
+        for layer_id, layer in matches.items():
+            if layer.kernel == "paddle.divide":
+                layer_inputs.append(layer.inputs["x"])
+                layer_inputs_ids.append(layer_id)
+            if layer.kernel == "paddle.multiply":
+                output_name = layer.outputs[0]
+        new_layer = PaddleLayer(
+            layer_id_list[0],
+            "paddle.nn.GELU",
+            inputs={"x": layer_inputs[0]},
+            outputs=[output_name],
+            approximate=False)
+        return new_layer, layer_inputs_ids[0]

x2paddle/optimizer/fusion/replace_div_to_scale.py

+#   Copyright (c) 2022  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 copy
+import numpy as np
+from collections import OrderedDict
+from x2paddle.optimizer.pattern_matcher import FuseBase
+from x2paddle.core.program import PaddleGraph, PaddleLayer
+from x2paddle.core.util import *
+
+
+class Div2Scale(FuseBase):
+    def __init__(self):
+        super(Div2Scale, self).__init__()
+
+    def build_pattern(self):
+        """
+        code describe:
+            x2paddle_296 = paddle.full(dtype='float32', shape=[1], fill_value=8.0)
+            x2paddle_293 = paddle.transpose(x=x2paddle_292, perm=[0, 2, 1, 3])
+            x2paddle_294 = paddle.transpose(x=x2paddle_260, perm=[0, 2, 3, 1])
+            x2paddle_295 = paddle.matmul(x=x2paddle_293, y=x2paddle_294)
+            x2paddle_297 = paddle.divide(x=x2paddle_295, y=x2paddle_296)
+        """
+
+        def gen_name(id):
+            return "x" + str(id)
+
+        self.pattern.add_layer(
+            "paddle.full",
+            inputs={},
+            outputs=[gen_name(0)],
+            shape=[1],
+            fill_value=8)
+        self.pattern.add_layer(
+            "paddle.transpose",
+            inputs={"x": "div2scale-input-0"},
+            outputs=[gen_name(1)],
+            perm=[0, 2, 1, 3])
+        self.pattern.add_layer(
+            "paddle.transpose",
+            inputs={"x": "div2scale-input-1"},
+            outputs=[gen_name(2)],
+            perm=[0, 2, 1, 3])
+        self.pattern.add_layer(
+            "paddle.matmul",
+            inputs={"x": gen_name(1),
+                    "y": gen_name(2)},
+            outputs=[gen_name(3)])
+        self.pattern.add_layer(
+            "paddle.divide",
+            inputs={"x": gen_name(3),
+                    "y": gen_name(0)},
+            outputs=[gen_name(4)])
+        self.pattern.build(inputs={
+            "input-0": "div2scale-input-0",
+            "input-1": "div2scale-input-1",
+        })
+
+    def insert_new_layer(self, graph, parameters, matches):
+        new_layer, new_layer_id = self.gen_new_layer(parameters, matches)
+        graph.layers[new_layer_id] = new_layer
+        matches_copy = copy.deepcopy(matches)
+        for layer_id, layer in matches_copy.items():
+            if layer.kernel in ["paddle.transpose", "paddle.matmul"]:
+                matches.pop(layer_id)
+        matches.pop(new_layer_id)
+
+    def gen_new_layer(self, parameters, matches):
+        layer_id_list = list(matches.keys())
+        layer_id_list.sort(key=int)
+        layer_inputs = list()
+        layer_inputs_ids = list()
+        fill_value = 0
+        for layer_id, layer in matches.items():
+            if layer.kernel == "paddle.full":
+                fill_value = layer.attrs["fill_value"]
+            if layer.kernel == "paddle.divide":
+                layer_inputs.append(layer.inputs["x"])
+                layer_inputs_ids.append(layer_id)
+                output_name = layer.outputs[0]
+        new_layer = PaddleLayer(
+            layer_id_list[0],
+            "paddle.scale",
+            inputs={"x": layer_inputs[0]},
+            outputs=[output_name],
+            scale=1 / fill_value)
+        return new_layer, layer_inputs_ids[0]

x2paddle/optimizer/fusion/replace_div_to_scale_pass.py

+#   Copyright (c) 2022  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 Div2Scale
+from x2paddle.optimizer.pass_manager import pass_register
+
+
+@pass_register
+class Div2ScalePass(Pass):
+    name = "replace_div_to_scale_pass"
+
+    def __init__(self):
+        Pass.__init__(self)
+
+    def apply(self, graph):
+        fuser = Div2Scale()
+        fuser.operate(graph, match_kind="edge")
+
+
+# register huggingface div2scale pass
+replace_div_to_scale_pass = Div2ScalePass()

x2paddle/optimizer/fusion/trace_fc_fuser.py

@@ -118,7 +118,9 @@ class TraceFcFuser(FuseBase):
             (1, 0))
         self.rm_params.add(weight_name)
         bias_numpy = parameters[bias_name]
-        parameters["{}.bias".format(linear_name)] = np.squeeze(bias_numpy)
+        if len(bias_numpy.shape) == 2:
+            bias_numpy = np.squeeze(bias_numpy)
+        parameters["{}.bias".format(linear_name)] = bias_numpy
         self.rm_params.add(bias_name)
         new_layer = PaddleLayer(
             layers_id[0],

x2paddle/optimizer/optimizer.py

@@ -37,7 +37,11 @@ class GraphOptimizer(object):
                 "prelu_fuse_pass", "transpose_eliminate_pass"
             ]
         elif source_frame == "onnx":
-            self.passes = ["onnx_layernorm_fuse_pass"]
+            self.passes = [
+                "onnx_layernorm_fuse_pass",
+                "onnx_gelu_fuse_pass",
+                "replace_div_to_scale_pass",
+            ]
         else:
             self.passes = []
 
@@ -54,7 +58,7 @@ class GraphOptimizer(object):
                     before_len = len(graph.layers)
                     pass_.apply(graph)
                     after_len = len(graph.layers)
-                    if after_len < before_len:
+                    if after_len <= before_len:
                         show_pass_log = True
                     if before_len == after_len:
                         break

x2paddle/utils.py

@@ -103,15 +103,7 @@ class PaddleDtypes():
             self.t_int64 = paddle.int64
             self.t_bool = paddle.bool
         else:
-            self.t_float16 = "paddle.fluid.core.VarDesc.VarType.FP16"
-            self.t_float32 = "paddle.fluid.core.VarDesc.VarType.FP32"
-            self.t_float64 = "paddle.fluid.core.VarDesc.VarType.FP64"
-            self.t_uint8 = "paddle.fluid.core.VarDesc.VarType.UINT8"
-            self.t_int8 = "paddle.fluid.core.VarDesc.VarType.INT8"
-            self.t_int16 = "paddle.fluid.core.VarDesc.VarType.INT16"
-            self.t_int32 = "paddle.fluid.core.VarDesc.VarType.INT32"
-            self.t_int64 = "paddle.fluid.core.VarDesc.VarType.INT64"
-            self.t_bool = "paddle.fluid.core.VarDesc.VarType.BOOL"
+            raise Exception("Paddle>=2.0.0 is required, Please update version!")
 
 
 is_new_version = check_version()