[PaddlePaddle Hackathon] Task 66 #4317

gbstack · 2021-10-15T07:06:39Z

PR types

New features

PR changes

APIs

Describe

你好，
该PR根据 #4225 添加了多尺度测试

使用方法

在配置文件中添加MultiscaleTestResize配置，具体使用方法可以参考新添加的配置文件：configs/faster_rcnn/faster_rcnn_r34_fpn_multiscaletest_1x_coco.yml

测试方法

评估

python tools/eval.py --config configs/faster_rcnn/faster_rcnn_r34_fpn_multiscaletest_1x_coco.yml -o weights=/home/gbstack/Downloads/faster_rcnn_r34_fpn_1x_coco.pdparams

（该测试使用了backbone为resnet34的faster rcnn配置，需要修改上面的权重文件为正确位置）

预测

tools/infer.py --config configs/faster_rcnn/faster_rcnn_r34_fpn_multiscaletest_1x_coco.yml --infer_img=/dataset/coco2017/val2017/000000000139.jpg -o weights=/home/gbstack/Downloads/faster_rcnn_r34_fpn_1x_coco.pdparams

评估效果对比

根据上面的配置文件的配置（resnet34-faster-rcnn-fpn）在coco2017验证集上进行评估

多尺度

单尺度

…t method)

jerrywgz

需要增加multiscale-test的使用文档

jerrywgz · 2021-10-18T06:21:36Z

ppdet/metrics/metrics.py

@@ -97,7 +97,11 @@ def update(self, inputs, outputs):
        for k, v in outputs.items():
            outs[k] = v.numpy() if isinstance(v, paddle.Tensor) else v

-        im_id = inputs['im_id']
+        # multi-scale inputs: all inputs have same im_id


这里是否只支持batch size=1？

是的，只支持batch size 1。因为我看到静态图的实现好像也是batch size 1，就和他一致了

static/ppdet/utils/post_process.py:207

def mstest_box_post_process(result, config, num_classes): """ Multi-scale Test Only available for batch_size=1 now. """ post_bbox = {} use_flip = False

jerrywgz · 2021-10-18T06:34:26Z

ppdet/modeling/architectures/meta_arch.py

+
+    def merge_multi_scale_predictions(self, outs):
+        # default values for architectures not included in following list
+        num_classes = 80


这里RCNN是否可以统一使用bbox_post_process中nms的参数
另外num_classe能否可以调整成输入参数

好的，新提交的代码已修改为统一使用bbox_post_process中nms的参数。

您说的num_classe调整成输入参数，是指从配置文件里面配置吗？如果是配置文件的话，现在的num_classes会使用配置文件中的BBoxHead.num_classes值

现在是否仅支持rcnn系列模型做ms-test，是的话可以在这里给出提示信息

是的。最新的提交版本已添加了错误信息

…ascadeRCNN

jerrywgz · 2021-10-18T07:55:54Z

需要增加multiscale-test的使用文档和限定条件

… scale test mode

gbstack · 2021-10-18T08:36:46Z

最新的版本添加了使用文档（位置在docs/tutorials/config_annotation/multi_scale_test_config.md），您看是否合适呢

heavengate

加一下单测吧，加到ppdet/modeling/tests目录下，可以参考下https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.2/ppdet/modeling/tests/test_architectures.py
或者
https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.2/ppdet/modeling/tests/test_ops.py

gbstack · 2021-10-18T09:36:14Z

加一下单测吧，加到ppdet/modeling/tests目录下，可以参考下https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.2/ppdet/modeling/tests/test_architectures.py 或者 https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.2/ppdet/modeling/tests/test_ops.py

好的，测试文件已添加（位于ppdet/modeling/tests/test_mstest.py）

jerrywgz · 2021-10-18T11:17:59Z

ppdet/modeling/tests/test_mstest.py

+        cfg = load_config(self.mstest_cfg_file)
+        trainer = Trainer(cfg, mode='test')
+
+        if not os.path.exists(cfg.weights):


权重使用https://paddledet.bj.bcebos.com/models/faster_rcnn_r34_fpn_1x_coco.pdparams
测试图片放在test文件夹中并使用相对路径

好的，最新的提交版本已修改

jerrywgz · 2021-10-18T11:18:48Z

docs/tutorials/config_annotation/multi_scale_test_config.md

+
+`target_size: [700 , 900]` property is used to specify different scales. 
+
+It can be plugged into evaluation process or test (inference) process, by adding `MultiscaleTestResize` entry to `EvalReader.sample_transforms` or `TestReader.sample_transforms`


说明目前支持rcnn系列模型，并且batch_size要求为1

好的，最新的提交版本已修改

gbstack · 2021-10-19T01:18:31Z

加一下单测吧，加到ppdet/modeling/tests目录下，可以参考下https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.2/ppdet/modeling/tests/test_architectures.py 或者 https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.2/ppdet/modeling/tests/test_ops.py

好的，测试文件已添加（位于ppdet/modeling/tests/test_mstest.py）

lyuwenyu · 2021-10-19T07:52:40Z

ppdet/data/transform/batch_operators.py

@@ -58,15 +58,21 @@ def __call__(self, samples, context=None):
        """
        coarsest_stride = self.pad_to_stride

-        max_shape = np.array([data['image'].shape for data in samples]).max(
+        # multi scale input is nested list
+        if type(samples) in (list, tuple) and type(samples[0]) in (list, tuple):


isinstance(samples, typing.Sequence) and len(samples) > 0 and isinstance(samples[0], typing.Sequence)

好的，最新的提交版本已修改

lyuwenyu · 2021-10-19T07:53:05Z

ppdet/engine/trainer.py

@@ -428,7 +428,11 @@ def _eval_with_loader(self, loader):
            for metric in self._metrics:
                metric.update(data, outs)

-            sample_num += data['im_id'].numpy().shape[0]
+            # multi-scale inputs: all inputs have same im_id
+            if type(data) in (list, tuple):


好的，最新的提交版本已修改

lyuwenyu · 2021-10-19T07:53:16Z

ppdet/engine/trainer.py

@@ -471,7 +475,10 @@ def predict(self,
            outs = self.model(data)

            for key in ['im_shape', 'scale_factor', 'im_id']:
-                outs[key] = data[key]
+                if type(data) in (tuple, list):


好的，最新的提交版本已修改

…ples sequence

jerrywgz

LGTM，辛苦向develop分支也提交个pr吧

gbstack · 2021-10-20T09:10:52Z

LGTM，辛苦向develop分支也提交个pr吧

好的，已提交PR。提交PR时发生的冲突需要我去合并吗？如果需要的话，是只合并我的修改造成的冲突，还是都修改呢？（包括release/2.2分支和develop分支之间的冲突）

jerrywgz · 2021-10-20T09:12:48Z

LGTM，辛苦向develop分支也提交个pr吧

好的，已提交PR。提交PR时发生的冲突需要我去合并吗？如果需要的话，是只合并我的修改造成的冲突，还是都修改呢？（包括release/2.2和develop的冲突）

建议把改动先手动加到develop分支上再提交pr，这个分支直接向develop分支提pr的话冲突会比较多
另：这个pr已经可以合入了

gbstack · 2021-10-20T09:17:07Z

LGTM，辛苦向develop分支也提交个pr吧

好的，已提交PR。提交PR时发生的冲突需要我去合并吗？如果需要的话，是只合并我的修改造成的冲突，还是都修改呢？（包括release/2.2和develop的冲突）

建议把改动先手动加到develop分支上再提交pr，这个分支直接向develop分支提pr的话冲突会比较多另：这个pr已经可以合入了

好的

qingqing01 · 2021-10-27T04:48:57Z

ppdet/modeling/architectures/meta_arch.py

 from ppdet.core.workspace import register
+from static.ppdet.utils.post_process import nms


不能从static文件夹中导入 nms

好的，新提交的PR已修复 #4375

This reverts commit 47fc188.

gbstack added 9 commits October 14, 2021 17:20

allow forwarding for multi scale inputs

3726603

fixed PadBatch not working with multi scale input

5a0ddf7

add faster_rcnn_r34_fpn_multiscaletest_1x_coco.yml

90f3e4b

fixed multi scale input checking for batch transforms (PadBatch)

9251968

less scales and faster inference for multi scale test configuration file

edaf3fa

fixed retrieving im_id from multi-scale inputs

837f47a

add multi scale prediction boxes merging

3236df6

allow multi scale testing for tools/infer.py (actually trainer.predic…

8889f8f

…t method)

add MultiscaleTestResize setting for TestReader

c69350e

gbstack mentioned this pull request Oct 15, 2021

【PaddlePaddle Hackathon】任务总览 PaddlePaddle/Paddle#35940

Closed

jerrywgz reviewed Oct 18, 2021

View reviewed changes

use BBoxPostProcess.nms parameter for multi scale prediction NMS of C…

9516048

…ascadeRCNN

gbstack added 3 commits October 18, 2021 15:58

add documentation for multi scale testing

1c88cc4

raise exception if the network architecture is not supported in multi…

5708c7a

… scale test mode

move documentation file

6478c1e

heavengate reviewed Oct 18, 2021

View reviewed changes

add test case for multi scale test

9a4f8c5

jerrywgz reviewed Oct 18, 2021

View reviewed changes

gbstack added 3 commits October 19, 2021 09:47

load pretrained weights from URL

d229c30

add image files for testing. use relative path for input images

a9d99ca

update documentation

4e31d8c

lyuwenyu reviewed Oct 19, 2021

View reviewed changes

use isinstance for sequence type checking. add empty checking for sam…

7ecbb8c

…ples sequence

jerrywgz approved these changes Oct 20, 2021

View reviewed changes

heavengate approved these changes Oct 20, 2021

View reviewed changes

jerrywgz merged commit 47fc188 into PaddlePaddle:release/2.2 Oct 20, 2021

qingqing01 reviewed Oct 27, 2021

View reviewed changes

qingqing01 added a commit that referenced this pull request Oct 27, 2021

Revert " [PaddlePaddle Hackathon] Task 66 (#4317)"

da3cf8d

This reverts commit 47fc188.

This was referenced Oct 27, 2021

Revert " [PaddlePaddle Hackathon] Task 66" #4374

Merged

[BUG] 10月22更新后出现安装异常 No module named 'static' #4368

Closed

qingqing01 added a commit that referenced this pull request Oct 27, 2021

Revert " [PaddlePaddle Hackathon] Task 66 (#4317)" (#4374)

6220fd3

This reverts commit 47fc188.

This was referenced Oct 27, 2021

多尺度测试代码中，移除了对static模块的依赖 #4375

Closed

add multi-scale test #4377

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[PaddlePaddle Hackathon] Task 66 #4317

[PaddlePaddle Hackathon] Task 66 #4317

gbstack commented Oct 15, 2021

jerrywgz left a comment

jerrywgz Oct 18, 2021

gbstack Oct 18, 2021

jerrywgz Oct 18, 2021

gbstack Oct 18, 2021 •

edited

Loading

jerrywgz Oct 18, 2021

gbstack Oct 18, 2021

jerrywgz commented Oct 18, 2021

gbstack commented Oct 18, 2021

heavengate left a comment

gbstack commented Oct 18, 2021

jerrywgz Oct 18, 2021

gbstack Oct 19, 2021

jerrywgz Oct 18, 2021

gbstack Oct 19, 2021

gbstack commented Oct 19, 2021

lyuwenyu Oct 19, 2021

gbstack Oct 19, 2021

lyuwenyu Oct 19, 2021

gbstack Oct 19, 2021

lyuwenyu Oct 19, 2021

gbstack Oct 19, 2021

jerrywgz left a comment

gbstack commented Oct 20, 2021 •

edited

Loading

jerrywgz commented Oct 20, 2021 •

edited

Loading

gbstack commented Oct 20, 2021

qingqing01 Oct 27, 2021

gbstack Oct 27, 2021


		`target_size: [700 , 900]` property is used to specify different scales.

		It can be plugged into evaluation process or test (inference) process, by adding `MultiscaleTestResize` entry to `EvalReader.sample_transforms` or `TestReader.sample_transforms`

		from ppdet.core.workspace import register
		from static.ppdet.utils.post_process import nms

[PaddlePaddle Hackathon] Task 66 #4317

[PaddlePaddle Hackathon] Task 66 #4317

Conversation

gbstack commented Oct 15, 2021

jerrywgz left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

gbstack Oct 18, 2021 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

jerrywgz commented Oct 18, 2021

gbstack commented Oct 18, 2021

heavengate left a comment

Choose a reason for hiding this comment

gbstack commented Oct 18, 2021

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

gbstack commented Oct 19, 2021

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

jerrywgz left a comment

Choose a reason for hiding this comment

gbstack commented Oct 20, 2021 • edited Loading

jerrywgz commented Oct 20, 2021 • edited Loading

gbstack commented Oct 20, 2021

Choose a reason for hiding this comment

Choose a reason for hiding this comment

gbstack Oct 18, 2021 •

edited

Loading

gbstack commented Oct 20, 2021 •

edited

Loading

jerrywgz commented Oct 20, 2021 •

edited

Loading