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

docs/inference_model_convertor/demo/tensorflow2paddle.ipynb

@@ -81,7 +81,7 @@
    "source": [
     "## 模型迁移\n",
     "### 1. 获取MobileNetV1的FrozenModel\n",
-    "由于X2Paddle只支持TensorFlow中FrozenModel的转换，如果为纯checkpoint模型，需要参考参考X2Paddle官方[文档](https://github.com/PaddlePaddle/X2Paddle/blob/develop/docs/user_guides/export_tf_model.md)，将其转换为FrozenModel，本示例中提供的模型为FrozenModel，所以无需转换。"
+    "由于X2Paddle只支持TensorFlow中FrozenModel的转换，如果为纯checkpoint模型，需要参考参考X2Paddle官方[文档](https://github.com/PaddlePaddle/X2Paddle/blob/release-1.1/docs/user_guides/export_tf_model.md)，将其转换为FrozenModel，本示例中提供的模型为FrozenModel，所以无需转换。"
    ]
   },
   {
@@ -210,4 +210,4 @@
  },
  "nbformat": 4,
  "nbformat_minor": 4
-}
+}
\ No newline at end of file

x2paddle/core/program.py

@@ -558,9 +558,9 @@ class PaddleGraph(object):
         paddle.save(self.parameters, save_path)
 
     def dygraph2static(self, save_dir, input_shapes=[], input_types=[]):
-        sepc_list = list()
+        spec_list = list()
         for i, name in enumerate(self.inputs):
-            sepc_list.append(
+            spec_list.append(
                 paddle.static.InputSpec(
                     shape=input_shapes[i], name=name, dtype=input_types[i]))
         path = osp.abspath(save_dir)
@@ -574,7 +574,7 @@ class PaddleGraph(object):
         else:
             model.set_dict(restore)
         model.eval()
-        static_model = paddle.jit.to_static(model, input_spec=sepc_list)
+        static_model = paddle.jit.to_static(model, input_spec=spec_list)
         try:
             paddle.jit.save(static_model,
                             osp.join(save_dir, "inference_model/model"))

x2paddle/decoder/onnx_decoder.py

@@ -583,7 +583,8 @@ class ONNXDecoder(object):
             item.name = self.make_variable_name(item.name)
         for node in graph.node:
             node.name = node.output[0]
-            if ":" in node.name and len(node.output) > 1:
+            if ":" in node.name and len(
+                    node.output) > 1 and node.op_type != "LSTM":
                 node.name = node.name.split(':')[0]
             node.name = self.make_variable_name(node.name)
             for i in range(len(node.input)):

x2paddle/op_mapper/caffe2paddle/caffe_op_mapper.py

@@ -966,11 +966,12 @@ class CaffeOpMapper():
             inputs={"x": input.name},
             outputs=[node.layer_name],
             **layer_attrs)
-        self.paddle_graph.add_layer(
-            "paddle.pow",
-            inputs={"x": node.layer_name},
-            outputs=[node.layer_name],
-            exponent=params.power)
+        if params.power != 1:
+            self.paddle_graph.add_layer(
+                "paddle.pow",
+                inputs={"x": node.layer_name,
+                        "y": params.power},
+                outputs=[node.layer_name])
 
     def Reduction(self, node):
         assert len(

x2paddle/op_mapper/onnx2paddle/opset9/opset.py

@@ -2277,14 +2277,35 @@ class OpSet9():
 
         paddings, var_x = self._pad_if_asymmetric(node, pads, val_x)
 
-        output_size = [0, 0]
+        if len(output_size) != 0:
+            paddings = [0] * 4
+            total_paddings = list()
+            total_paddings.append((val_x.out_shapes[0][2] - 1) * strides[
+                0] + dilations[0] * (kernel_shape[0] - 1) + 1 + out_padding[0] -
+                                  output_size[0])
+            total_paddings.append((val_x.out_shapes[0][3] - 1) * strides[
+                1] + dilations[1] * (kernel_shape[1] - 1) + 1 + out_padding[1] -
+                                  output_size[1])
+            if auto_pad == "SAME_UPPER":
+                for i in range(len(total_paddings)):
+                    paddings[2 * i] = total_paddings[0] - total_paddings[0] // 2
+                    paddings[2 * i + 1] = total_paddings[0] // 2
+            else:
+                for i in range(len(total_paddings)):
+                    paddings[2 * i] = total_paddings[0] // 2
+                    paddings[2 * i + 1] = total_paddings[0] - total_paddings[
+                        0] // 2
+        else:
+            output_size = [0, 0]
 
-        output_size[0] = (val_x.out_shapes[0][2] - 1
-                          ) * strides[0] - 2 * paddings[0] + dilations[0] * (
-                              kernel_shape[0] - 1) + 1 + out_padding[0]
-        output_size[1] = (val_x.out_shapes[0][3] - 1
-                          ) * strides[1] - 2 * paddings[1] + dilations[1] * (
-                              kernel_shape[1] - 1) + 1 + out_padding[1]
+            output_size[0] = (
+                val_x.out_shapes[0][2] - 1
+            ) * strides[0] - 2 * paddings[0] + dilations[0] * (
+                kernel_shape[0] - 1) + 1 + out_padding[0]
+            output_size[1] = (
+                val_x.out_shapes[0][3] - 1
+            ) * strides[1] - 2 * paddings[1] + dilations[1] * (
+                kernel_shape[1] - 1) + 1 + out_padding[1]
 
         # Conv2DTranspose缺少output_size，只能在forward里头传进output_size
         inputs_dict = {'x': val_x if isinstance(val_x, str) else val_x.name}
@@ -2327,6 +2348,8 @@ class OpSet9():
         if val_b is not None:
             _rename_or_remove_weight(self.weights, val_b.name,
                                      op_name + '.bias')
+        else:
+            layer_attrs["bias_attr"] = False
         self.paddle_graph.add_layer(
             kernel=paddle_op,
             inputs=inputs_dict,

x2paddle/op_mapper/pytorch2paddle/aten.py

@@ -1315,8 +1315,10 @@ def aten__convolution(mapper, graph, node):
     weights = mapper.pytorch_params[inputs_name[1]]
     if len(weights.shape) == 3:
         op_name = name_generator("conv1d", mapper.nn_name2id)
-    else:
+    elif len(weights.shape) == 4:
         op_name = name_generator("conv2d", mapper.nn_name2id)
+    else:
+        op_name = name_generator("conv3d", mapper.nn_name2id)
     output_name = mapper._get_outputs_name(node)[0]
     layer_outputs = [op_name, output_name]
     layer_inputs = {}
@@ -1364,7 +1366,22 @@ def aten__convolution(mapper, graph, node):
     else:
         layer_attrs['in_channels'] = weights.shape[1] * mapper.attrs[
             inputs_name[8]]
-    if len(weights.shape) == 4:
+    if len(weights.shape) == 3:
+        if mapper.attrs[inputs_name[6]]:
+            graph.add_layer(
+                "paddle.nn.Conv1DTranspose",
+                inputs=layer_inputs,
+                outputs=layer_outputs,
+                scope_name=scope_name,
+                **layer_attrs)
+        else:
+            graph.add_layer(
+                "paddle.nn.Conv1D",
+                inputs=layer_inputs,
+                outputs=layer_outputs,
+                scope_name=scope_name,
+                **layer_attrs)
+    elif len(weights.shape) == 4:
         if mapper.attrs[inputs_name[6]]:
             graph.add_layer(
                 "paddle.nn.Conv2DTranspose",
@@ -1382,14 +1399,14 @@ def aten__convolution(mapper, graph, node):
     else:
         if mapper.attrs[inputs_name[6]]:
             graph.add_layer(
-                "paddle.nn.Conv1DTranspose",
+                "paddle.nn.Conv3DTranspose",
                 inputs=layer_inputs,
                 outputs=layer_outputs,
                 scope_name=scope_name,
                 **layer_attrs)
         else:
             graph.add_layer(
-                "paddle.nn.Conv1D",
+                "paddle.nn.Conv3D",
                 inputs=layer_inputs,
                 outputs=layer_outputs,
                 scope_name=scope_name,