FasterRCNN Anchor Generator Op #11218

xingyuanbu · 2018-06-06T03:37:39Z

Anchor Generator Op
Match with detectron

qingqing01 · 2018-06-11T07:19:48Z

paddle/fluid/operators/detection/anchor_generator_op.cc

+  void Make() override {
+    AddInput("Input",
+             "(Tensor, default Tensor<float>), "
+             "the input feature data of AnchorGeneratorOp. "


the input feature data of AnchorGeneratorOp ->

the input feature is tensor with a rank of 4.

qingqing01 · 2018-06-11T07:20:54Z

paddle/fluid/operators/detection/anchor_generator_op.cc

+             "The layout is NCHW.");
+    AddOutput("Anchors",
+              "(Tensor, default Tensor<float>), the output anchors of "
+              "AnchorGeneratorOp. The layout is [H, W, num_anchors, 4]. "


the output anchors of AnchorGeneratorOp ->

the output is a tensor with a rank of 4.

qingqing01 · 2018-06-11T07:24:22Z

paddle/fluid/operators/detection/anchor_generator_op.cc

+              "AnchorGeneratorOp. The layout is [H, W, num_anchors, 4]. "
+              "H is the height of input, W is the width of input, num_anchors "
+              "is the box count of each position. "
+              "Each anchor is in (x1, y1, x2, y2) format");


x1, y1, x2, y2 -> xmin, ymin, xmax, ymax ?

qingqing01 · 2018-06-11T07:25:09Z

paddle/fluid/operators/detection/anchor_generator_op.cc

+              "Each anchor is in (x1, y1, x2, y2) format");
+    AddOutput("Variances",
+              "(Tensor, default Tensor<float>), the expanded variances of "
+              "AnchorGeneratorOp. The layout is [H, W, num_anchors, 4]. "


of AnchorGeneratorOp 这种话没有意义。

qingqing01 · 2018-06-11T07:26:29Z

paddle/fluid/operators/detection/anchor_generator_op.cc

+              "H is the height of input, W is the width of input, num_anchors "
+              "is the box count of each position. "
+              "Each variance is in (x, y, w, h) format "
+              "which is unrelated with Anchor's");


Each variance is in (x, y, w, h) format which is unrelated with Anchor's

这句话不太懂。

qingqing01 · 2018-06-11T07:54:43Z

python/paddle/fluid/layers/detection.py

+
+def anchor_generator(input,
+                     anchor_sizes=[64, 128, 256, 512],
+                     aspect_ratios=[0.5, 1.0, 2.0],


anchor_sizes=None aspect_ratios=None stride=None

用户会依据图片大小调整的这些变量，不要设置缺省值较好些？

qingqing01 · 2018-06-11T07:55:51Z

python/paddle/fluid/tests/unittests/test_anchor_generator_op.py

+
+        self.anchor_sizes = [64, 128, 256, 512]
+        self.anchor_sizes = np.array(self.anchor_sizes).astype(
+            'float32').tolist()


self.anchor_sizes = [64., 128., 256., 512.]

即可，去掉line 57.

qingqing01 · 2018-06-11T07:56:23Z

python/paddle/fluid/tests/unittests/test_anchor_generator_op.py

+            'float32').tolist()
+        self.aspect_ratios = [0.5, 1, 2]
+        self.aspect_ratios = np.array(self.aspect_ratios).astype(
+            'float32').tolist()


同上，没有必要转成np.array再转回来吧。

qingqing01 · 2018-06-11T07:58:01Z

python/paddle/fluid/tests/unittests/test_anchor_generator_op.py

+        out_var = np.tile(self.variances, (self.layer_h, self.layer_w,
+                                           self.num_anchors, 1))
+        self.out_anchors = out_anchors.astype('float32')
+        self.out_var = out_var.astype('float32')


test_prior_box_op.py 本身就写的有点乱，写法不要按照这个来。

这里把anchor生成的代码写到class外，写成一个函数吧。

qingqing01 · 2018-06-11T08:00:12Z

python/paddle/fluid/tests/unittests/test_anchor_generator_op.py

+    def set_data(self):
+        self.init_test_params()
+        self.init_test_input()
+        self.init_test_output()


这里的单测只有一个，输入、输出的初始化不是为了不同的单测复用，就没必要写成两个函数，里面的内容可以直接展开到这里吧。

感觉不展开继续使用init_test_output这样的方式更清晰，更模块化一些。我已经将anchor生成代码写到class外了，你可以看一看现在的风格是否恰当~

qingqing01 · 2018-06-29T06:21:06Z

python/paddle/fluid/layers/detection.py

+    """
+    **Anchor generator operator**
+
+    Generate prior boxes for SSD(Single Shot MultiBox Detector) algorithm.


Not for SSD.

qingqing01 · 2018-06-29T06:21:43Z

python/paddle/fluid/layers/detection.py

+    Each position of the input produce N prior boxes, N is determined by
+    the count of min_sizes, max_sizes and aspect_ratios, The size of the
+    box is in range(min_size, max_size) interval, which is generated in
+    sequence according to the aspect_ratios.


There is no min_sizes, max_sizes here, please update the doc.

qingqing01 · 2018-06-29T06:24:35Z

python/paddle/fluid/layers/detection.py

+    Args:
+       input(Variable): The Input feature map, the format is NCHW.
+       anchor_sizes(list|tuple|float value): Anchor sizes of generated anchors.
+       Defalut: [64, 128, 256, 512]


list|tuple|float value -> list|tuple|float

There is no default value in the interface. And need to explain what is anchor size?

qingqing01 · 2018-06-29T06:25:35Z

python/paddle/fluid/layers/detection.py

+def anchor_generator(input,
+                     anchor_sizes=None,
+                     aspect_ratios=None,
+                     variance=[0.1, 0.1, 0.2, 0.2],


这里的默认值是常用值吗？否则，不要设置。暴露给用户的接口要谨慎。

variance=[0.1, 0.1, 0.2, 0.2] is a common value, and is rarely to be set to another value.

qingqing01 · 2018-06-29T06:26:15Z

python/paddle/fluid/layers/detection.py

+       anchor_sizes(list|tuple|float value): Anchor sizes of generated anchors.
+       Defalut: [64, 128, 256, 512]
+       aspect_ratios(list|tuple|float value): The aspect ratios of generated
+            anchors. Default: [0.5, 1.0, 2.0].


same as above.

qingqing01 · 2018-06-29T06:27:44Z

python/paddle/fluid/layers/detection.py

+             The layout is [H, W, num_anchors, 4].
+             H is the height of input, W is the width of input,
+             num_anchors is the box count of each position.
+             Each anchor is in (x1, y1, x2, y2) format.


x1, y1, x2, y2 -> xmin, ymin, xmax, ymax? 以及说明下，是否normalize？

qingqing01 · 2018-06-29T06:30:41Z

python/paddle/fluid/layers/detection.py

+            aspect_ratios=[0.5, 1.0, 2.0],
+            variance=[0.1, 0.1, 0.2, 0.2],
+            stride=[16.0, 16.0],
+            offset=0.5)


Examples .. code-block:: python 前后注意留空格，否则文档格式不正确。

Examples: .. code-block:: python anchor, variance = anchor_generator( # ... 下面代码注意缩进

qingqing01 · 2018-06-29T06:33:24Z

paddle/fluid/operators/detection/anchor_generator_op.cc

+
+    AddAttr<std::vector<float>>(
+        "anchor_sizes",
+        "(vector<float>) List of RPN anchor sizes "


List of RPN anchor sizes

need more details, RPN用全拼。

qingqing01 · 2018-06-29T06:34:48Z

paddle/fluid/operators/detection/anchor_generator_op.cc

+                   "(float) "
+                   "Anchor center offset, with a default of 0.5")
+        .SetDefault(0.5);
+    AddComment(R"DOC(


上面是否有默认值和Python端保持一致？

0.5 is the default value both in python and C++

qingqing01 · 2018-06-29T06:35:56Z

paddle/fluid/operators/detection/anchor_generator_op.cc

+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("Input"),
+                   "Input(Input) of AnchorGeneratorOp should not be null.");


Also need to check outputs.

PADDLE_ENFORCE(ctx->HasInput("Anchors"), "xxxx"); PADDLE_ENFORCE(ctx->HasInput("Variances"), "xxxx");

qingqing01 · 2018-06-29T06:46:30Z

python/paddle/fluid/layers/detection.py

@@ -909,3 +910,95 @@ def _is_list_or_tuple_and_equal(data, length, err_info):
    box.stop_gradient = True
    var.stop_gradient = True
    return mbox_locs_concat, mbox_confs_concat, box, var
+
+
+def anchor_generator(input,


Need unit testing for the interface in https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/tests/test_detection.py

* Add anchor generator operator for Faster-RCNN. * Add unittest testing. * Add Python API.

reyoung · 2018-09-27T01:49:57Z

paddle/fluid/operators/detection/anchor_generator_op.cc

+    auto input_dims = ctx->GetInputDim("Input");
+    PADDLE_ENFORCE(input_dims.size() == 4, "The layout of input is NCHW.");
+
+    auto anchor_sizes = ctx->Attrs().Get<std::vector<float>>("anchor_sizes");


auto --> auto&

reyoung · 2018-09-27T01:50:04Z

paddle/fluid/operators/detection/anchor_generator_op.cc

+    PADDLE_ENFORCE(input_dims.size() == 4, "The layout of input is NCHW.");
+
+    auto anchor_sizes = ctx->Attrs().Get<std::vector<float>>("anchor_sizes");
+    auto aspect_ratios = ctx->Attrs().Get<std::vector<float>>("aspect_ratios");


auto --> auto&

reyoung · 2018-09-27T01:50:13Z

paddle/fluid/operators/detection/anchor_generator_op.cc

+
+    auto anchor_sizes = ctx->Attrs().Get<std::vector<float>>("anchor_sizes");
+    auto aspect_ratios = ctx->Attrs().Get<std::vector<float>>("aspect_ratios");
+    auto stride = ctx->Attrs().Get<std::vector<float>>("stride");


auto --> auto&

reyoung · 2018-09-27T01:50:17Z

paddle/fluid/operators/detection/anchor_generator_op.cc

+    auto anchor_sizes = ctx->Attrs().Get<std::vector<float>>("anchor_sizes");
+    auto aspect_ratios = ctx->Attrs().Get<std::vector<float>>("aspect_ratios");
+    auto stride = ctx->Attrs().Get<std::vector<float>>("stride");
+    auto variances = ctx->Attrs().Get<std::vector<float>>("variances");


auto --> auto&

reyoung · 2018-09-27T01:50:40Z

paddle/fluid/operators/detection/anchor_generator_op.cc

+    size_t num_anchors = aspect_ratios.size() * anchor_sizes.size();
+
+    std::vector<int64_t> dim_vec(4);
+    dim_vec[0] = input_dims[2];


Check input_dims.size() >= 4

reyoung · 2018-09-27T01:50:45Z

paddle/fluid/operators/detection/anchor_generator_op.cc

+
+    std::vector<int64_t> dim_vec(4);
+    dim_vec[0] = input_dims[2];
+    dim_vec[1] = input_dims[3];


Check input_dims.size() >= 4

xingyuanbu added 5 commits June 4, 2018 22:50

add anchor generator unittest

19181e1

anchor generator op .h

79d286d

add anchor generator op .cu .cc

456bb2c

add anchro generator op into CMakelist.txt

3ae6208

fix and remove image from inputs

c1b831d

qingqing01 requested review from kuke and qingqing01 June 6, 2018 04:27

refine and follow format

c831787

xingyuanbu force-pushed the fasterrcnn branch from 6fbc6bb to c831787 Compare June 6, 2018 06:05

xingyuanbu added 2 commits June 6, 2018 19:09

Merge branch 'develop' into fasterrcnn

1ddb1d7

add python api

8dedcec

qingqing01 reviewed Jun 11, 2018

View reviewed changes

follow comments

be9a82e

qingqing01 reviewed Jun 29, 2018

View reviewed changes

xingyuanbu added 2 commits June 29, 2018 16:28

follow comments

6394330

fix

4b04afd

xingyuanbu force-pushed the fasterrcnn branch from 59d5278 to 716253e Compare June 29, 2018 16:17

fix

8ac7b0a

xingyuanbu force-pushed the fasterrcnn branch from 716253e to 8ac7b0a Compare June 29, 2018 16:43

qingqing01 approved these changes Jul 2, 2018

View reviewed changes

qingqing01 merged commit 5056d3e into PaddlePaddle:develop Jul 2, 2018

kuke pushed a commit to kuke/Paddle that referenced this pull request Aug 25, 2018

FasterRCNN Anchor Generator Op (PaddlePaddle#11218)

b7d31df

* Add anchor generator operator for Faster-RCNN. * Add unittest testing. * Add Python API.

reyoung reviewed Sep 27, 2018

View reviewed changes

FasterRCNN Anchor Generator Op #11218

FasterRCNN Anchor Generator Op #11218

Uh oh!

Conversation

xingyuanbu commented Jun 6, 2018

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

qingqing01 Jun 29, 2018 •

edited

Loading