1
2
3
4
5
# Train
python tools/train.py configs/yolox/yolox_tiny_8x8_300e_coco.py
# Test
python tools/test.py configs/yolox/yolox_tiny_8x8_300e_coco.py xxx.pth --eval mAP
Train
Traceback
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
File "./tools/train.py", line 189 in <module>
main()
File "./tools/train.py", line 178 in main
train_detector()
File "mmdet/apis/train.py", line 177 in train_detector
runner.run(data_loaders, cfg.workflow)
File "/home/zuo/code/mmcv/mmcv/runner/epoch_based_runner.py", line 177 in run
epoch_runner(data_loaders[i], **kwargs)
File "/home/zuo/code/mmcv/mmcv/runner/epoch_based_runner.py", line 50 in train
self.run_iter(data_batch, train_mode=True, **kwargs)
File "/home/zuo/code/mmcv/mmcv/runner/epoch_based_runner.py", line 29 in run_iter
self.model.train_step(data_batch, self.optimizer, **kwargs)
File "/home/zuo/code/mmcv/mmcv/parallel/distributed.py", line 177 in train_step
self.module.train_step(*inputs[0], **kwargs[0])
File "mmdet/models/detectors/base.py", line 238 in train_step
losses = self(**data)
File "lib/python3.7/site-packages/torch/nn/modules/module.py", line 727 in _call_impl
result = self.forward(*input, **kwargs)
File "/home/zuo/code/mmcv/mmcv/runner/fp16_utils.py", line 98 in new_func
return old_func(*args, **kwargs)
File "mmdet/models/detectors/base.py", line 172 in forward
return self.forward_train(img, img_metas, **kwargs)
File "mmdet/models/detectors/single_stage.py", line 82~83 in forward_train
x = self.extract_feat(img), line 82 in forward_train
extract_feat() "mmdet/models/detectors/single_stage.py:41", extract_feat(), 提取特征
losses = self.bbox_head.forward_train(x, img_metas, gt_bboxes, gt_labels, gt_bboxes_ignore)
outs = self(x) "mmdet/models/dense_heads/yolox_head.py:196", forward(), 计算输出
losses = self.loss(*loss_inputs) "mmdet/models/dense_heads/yolox_head.py:324", loss(), 计算loss
Test
Traceback
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
_do_evaluate "mmdet/core/evaluation/eval_hooks.py:12"
results = single_gpu_test(runner.model, self.dataloader, show=False)
result = model(return_loss=False, rescale=True, **data) "mmdet/apis/test.py:28"
forward_test(self, imgs, img_metas, **kwargs) "mmdet/models/detectors/base.py:112~154"
simple_test() "mmdet/models/detectors/single_stage.py"
feat = self.extract_feat(img) "mmdet/models/detectors/single_stage.py:41~46"
self.bbox_head.simple_test(feat,..)
simple_test() "mmdet/models/dense_heads/base_dense_head.py:334"
simple_test_bboxes() "mmdet/models/dense_heads/dense_test_mixins.py:17~39"
outs = self.forward(feats)
multi_apply() "mmdet/models/dense_heads/yolox_head.py:207"
results_list = self.get_bboxes()
get_bboxes() "mmdet/models/dense_heads/yolox_head.py:214~295"
bbox2result(det_bboxes,..) "mmdet/core/bbox/transforms.py:100~117"
key_score = self.evaluate(runner, results)
self.dataloader.dataset.evaluate() "/home/zuo/code/mmcv/mmcv/runner/hooks/evaluation.py:359"
evaluate() "mmdet/datasets/voc.py:28~105"
eval_map() "mmdet/core/evaluation/mean_ap.py:297~441"
YOLOX迁移遇到的问题
主要是Dynamic Shape类问题
dataset
由于数据集每张图片中object数量不固定，所以所模型输入tensor gt_bboxes和gt_labels的shape是不固定的。这就涉及到dynamic shape输入。彼时MindSpore对dynamic shape的支持还不完善，因此通过设置一个合理的最大objects数，将gt_bboxes和gt_labels padding到最大shape来规避dynamic shape问题。
另外增加有效object的标识作为输入tensor来表示哪些索引位置的object是真实的，哪些索引处的是padding的，以方便将padding对后续计算loss的影响降到最低。
loss
loss计算流程中get_in_gt_and_in_center_info函数里最后一个步骤：
1
2
3
4
5
# both in boxes and centers. shape:[num_fg, num_gt]
is_in_boxes_and_centers = (
is_in_gts[is_in_gts_or_centers, :] &
is_in_cts[is_in_gts_or_centers, :]
)
is_in_boxes_and_centers的shape会随is_in_gts_or_centers的值不同而不同，也就产生了dynamic shape问题，为了规避这个问题，做了如下修改：
1
2
3
is_in_boxes_and_centers = in_gts * in_cts
zeros = mnp.full_like(is_in_boxes_and_centers, 0)
is_in_boxes_and_centers = P.Select()(is_in_gts_or_centers_repeat, is_in_gts_and_centers, zeros)
通过Select算子，将is_in_gts_or_centers_repeat中True对应的元素赋值给is_in_gts_and_centers，is_in_gts_or_centers_repeat中False对应的位置padding成非法值（这里就是0）。
上面的例子是loss计算流程中dynamic shape的引入点，后续流程还有涉及到dynamic shape的地方，均按照同样的思想进行修改，同样要保证不影响loss的计算结果。