Merge pull request #1 from PaddlePaddle/develop

develop

FAQ.md

@@ -2,3 +2,32 @@
 
 **Q1. TensorFlow模型转换过程中，提示『Unknown shape for input tensor[tensor name: "input"]， Please define shape of input here』？**  
 A：该提示信息表示无法从TensorFlow的pb模型中获取到输入tensor(tensor名为"input:)的shape信息，所以需要用户手动在提示后输入详细的shape信息，如None,224,224,3 其中None表示Batch 
+
+
+**Q2. TensorFlow模型转换失败怎么解决?**  
+A: 目前TensorFlow模型转换失败存在几个问题。1) 存在暂未支持的OP，此信息会在转换时输出; 2) NHWC优化导致部分参数出错；3）Batch维度带来的出错 4）其它
+
+对于（1）问题，建议自行添加或发起Issue；
+
+其中（2）、(3)、（4）问题目前没有明确的报错信息，当您遇到模型转换失败时，请尝试如下的步骤后，再进行转换测试
+
+```
+x2paddle -f tensorflow -m tf.pb -s pd-model --without_data_format_optimization --define_input_shape
+```
+
+#### without_data_format_optimization : 关闭NHWC优化
+TensorFlow的CV模型，大多的输入格式为`NHWC`，而Paddle目前仅支持`NCHW`，如若直接转换，需要在conv2d、pool2d等操作前后添加transpose解决，这样会带来性能的损耗。X2Paddle在模型转换过程中，对此问题进行了优化，避免transpose操作带来的性能问题，但目前仅在部分模型上进行了测试，不一定适用于其它模型，因此，如若模型转换存在问题时，我们建议你关闭NHWC的优化。
+
+在模型转换时添加参数 --without_data_format_optimization
+```
+x2paddle -f tensorflow -m tf.pb -s pd-model --without_data_format_optimization
+```
+
+### define_input_shape : 固定Batch大小
+受限于不同框架的运行机制，在转换过程中，Batch维度也有一定可能会带来模型转换失败的问题。可以尝试固定Batch维度后再转换
+
+在模型转换时添加参数 --define_input_shape
+```
+x2paddle -f tensorflow -m tf.pb -s pd-model --define_input_shape
+```
+如原tensorflow模型的输入shape为`[None, 224, 224, 3]`，可添加参数后，根据提示，把输入的shape修改为`[2, 224, 224, 3]`

x2paddle/op_mapper/tf_op_mapper.py

@@ -17,6 +17,7 @@ from x2paddle.core.op_mapper import OpMapper
 from x2paddle.core.util import *
 import inspect
 import numpy
+import sys
 
 
 # compute padding size for SAME mode
@@ -83,18 +84,31 @@ class TFOpMapper(OpMapper):
             del self.graph.input_nodes[idx]
 
         print("Total nodes: {}".format(len(self.graph.topo_sort)))
+        unsupported_ops = set()
         for node_name in self.graph.topo_sort:
             node = self.graph.get_node(node_name)
             op = node.layer_type
             if op in self.directly_map_ops:
+                if len(unsupported_ops) > 0:
+                    continue
                 self.directly_map(node)
             elif op in self.elementwise_ops:
+                if len(unsupported_ops) > 0:
+                    continue
                 self.elementwise_map(node)
             elif hasattr(self, op):
+                if len(unsupported_ops) > 0:
+                    continue
                 func = getattr(self, op)
                 func(node)
             else:
-                raise Exception("OP: [{}] not support yet".format(op))
+                unsupported_ops.add(op)
+        if len(unsupported_ops) > 0:
+            print("=========={} Ops are not supported yet======".format(
+                len(unsupported_ops)))
+            for op in unsupported_ops:
+                print("========== {} ==========".format(op))
+            sys.exit(-1)
 
     def directly_map(self, node):
         assert node.layer_type in self.directly_map_ops
@@ -773,7 +787,15 @@ class TFOpMapper(OpMapper):
             begin = [begin[i] for i in [0, 3, 1, 2]]
             end = [end[i] for i in [0, 3, 1, 2]]
 
-        attr = {"axes": range(len(strides)), "starts": begin, "ends": end}
+        for i in range(len(end)):
+            if end[i] == 0:
+                end[i] = 999999
+
+        attr = {
+            "axes": [i for i in range(len(strides))],
+            "starts": begin,
+            "ends": end
+        }
         node.fluid_code.add_layer("slice",
                                   inputs=input,
                                   output=node,
@@ -955,3 +977,37 @@ class TFOpMapper(OpMapper):
                                   inputs=input,
                                   output=node,
                                   param_attr=attr)
+
+    def ResizeNearestNeighbor(self, node):
+        input = self.graph.get_node(node.layer.input[0], copy=True)
+        resize_shape = self.graph.get_node(node.layer.input[1], copy=True)
+        self.omit_nodes.append(resize_shape.layer_name)
+        if resize_shape.layer_type == "Const":
+            resize_shape = resize_shape.value.tolist()
+        else:
+            resize_shape = self.decoder.infer_shape_tensor(resize_shape)
+        align_corners = node.get_attr("align_corners")
+        attr = {"align_corners": align_corners, "out_shape": resize_shape}
+        node.fluid_code.add_layer("resize_nearest",
+                                  inputs=input,
+                                  output=node,
+                                  param_attr=attr)
+
+    def ResizeBilinear(self, node):
+        input = self.graph.get_node(node.layer.input[0], copy=True)
+        resize_shape = self.graph.get_node(node.layer.input[1], copy=True)
+        self.omit_nodes.append(resize_shape.layer_name)
+        if resize_shape.layer_type == "Const":
+            resize_shape = resize_shape.value.tolist()
+        else:
+            resize_shape = self.decoder.infer_shape_tensor(resize_shape)
+        align_corners = node.get_attr("align_corners")
+        attr = {
+            "align_corners": align_corners,
+            "out_shape": resize_shape,
+            "align_mode": 1
+        }
+        node.fluid_code.add_layer("resize_bilinear",
+                                  inputs=input,
+                                  output=node,
+                                  param_attr=attr)