diff --git a/configs/razor/subnets/README.md b/configs/razor/subnets/README.md
index ad7a716b..093dedd7 100644
--- a/configs/razor/subnets/README.md
+++ b/configs/razor/subnets/README.md
@@ -62,18 +62,18 @@ CUDA_VISIBLE_DEVICES=0 PORT=29500 ./tools/dist_test.sh configs/razor/subnets/yol
Here we provide the baseline version of YOLO Series with NAS backbone.
-| Model | size | box AP | Params(M) | FLOPs(G) | Config | Download |
-| :------------------------: | :--: | :----: | :----------: | :------: | :----------------------------------------------------------------------------------------------------------------------------------------: | :---------------------------------------------------------------------------------------------------------------------------------------------------------------------: |
-| yolov5-s | 640 | 37.7 | 7.235 | 8.265 | [config](https://github.com/open-mmlab/mmyolo/blob/main/configs/yolov5/yolov5_s-v61_syncbn_fast_8xb16-300e_coco.py) | [model](https://download.openmmlab.com/mmyolo/v0/yolov5/yolov5_s-v61_syncbn_fast_8xb16-300e_coco/yolov5_s-v61_syncbn_fast_8xb16-300e_coco_20220918_084700-86e02187.pth) |
-| yolov5_s_spos_shufflenetv2 | 640 | 37.9 | 7.04(-2.7%) | 7.03 | [config](https://github.com/open-mmlab/mmyolo/tree/dev/configs/razor/subnets/yolov5_s_spos_shufflenetv2_syncbn_8xb16-300e_coco.py) | [model](https://download.openmmlab.com/mmrazor/v1/spos/yolov5/yolov5_s_spos_shufflenetv2_syncbn_8xb16-300e_coco_20230109_155302-777fd6f1.pth) |
-| yolov6-s | 640 | 44.0 | 18.869 | 24.253 | [config](https://github.com/open-mmlab/mmyolo/blob/main/configs/yolov6/yolov6_s_syncbn_fast_8xb32-400e_coco.py) | [model](https://download.openmmlab.com/mmyolo/v0/yolov6/yolov6_s_syncbn_fast_8xb32-400e_coco/yolov6_s_syncbn_fast_8xb32-400e_coco_20221102_203035-932e1d91.pth) |
-| yolov6_l_attentivenas_a6 | 640 | 44.5 | 18.38(-2.6%) | 8.49 | [config](https://github.com/open-mmlab/mmyolo/tree/dev/configs/razor/subnets/yolov6_l_attentivenas_a6_d12_syncbn_fast_16xb16-300e_coco.py) | [model](https://download.openmmlab.com/mmrazor/v1/attentivenas/yolov6/yolov6_l_attentivenas_a6_d12_syncbn_fast_16xb16-300e_coco_20230108_174944-4970f0b7.pth) |
-| RTMDet-tiny | 640 | 41.0 | 4.8 | 8.1 | [config](./rtmdet_l_syncbn_fast_8xb32-300e_coco.py) | [model](https://download.openmmlab.com/mmyolo/v0/rtmdet/rtmdet_tiny_syncbn_fast_8xb32-300e_coco/rtmdet_tiny_syncbn_fast_8xb32-300e_coco_20230102_140117-dbb1dc83.pth) |
-| rtmdet_tiny_ofa_lat31 | 960 | 41.1 | 3.91(-18.5%) | 6.09 | [config](https://github.com/open-mmlab/mmyolo/tree/dev/configs/razor/subnets/rtmdet_tiny_ofa_lat31_syncbn_16xb16-300e_coco.py) | [model](https://download.openmmlab.com/mmrazor/v1/ofa/rtmdet/rtmdet_tiny_ofa_lat31_syncbn_16xb16-300e_coco_20230108_222141-24ff87dex.pth) |
+| Model | size | box AP | Params(M) | FLOPs(G) | Config | Download |
+| :------------------------: | :--: | :----: | :----------: | :------: | :---------------------------------------------------------------------------------------------------------------------------------------: | :------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------: |
+| yolov5-s | 640 | 37.7 | 7.235 | 8.265 | [config](https://github.com/open-mmlab/mmyolo/blob/main/configs/yolov5/yolov5_s-v61_syncbn_fast_8xb16-300e_coco.py) | [model](https://download.openmmlab.com/mmyolo/v0/yolov5/yolov5_s-v61_syncbn_fast_8xb16-300e_coco/yolov5_s-v61_syncbn_fast_8xb16-300e_coco_20220918_084700-86e02187.pth) \| [log](https://download.openmmlab.com/mmyolo/v0/yolov5/yolov5_s-v61_syncbn_fast_8xb16-300e_coco/yolov5_s-v61_syncbn_fast_8xb16-300e_coco_20220918_084700.log.json) |
+| yolov5_s_spos_shufflenetv2 | 640 | 38.0 | 7.04(-2.7%) | 7.03 | [config](https://github.com/open-mmlab/mmyolo/tree/dev/configs/razor/subnets/yolov5_s_spos_shufflenetv2_syncbn_8xb16-300e_coco.py) | [model](https://download.openmmlab.com/mmrazor/v1/yolo_nas_backbone/yolov5_s_spos_shufflenetv2_syncbn_8xb16-300e_coco_20230211_220635-578be9a9.pth) \| log |
+| yolov6-s | 640 | 44.0 | 18.869 | 24.253 | [config](https://github.com/open-mmlab/mmyolo/blob/main/configs/yolov6/yolov6_s_syncbn_fast_8xb32-400e_coco.py) | [model](https://download.openmmlab.com/mmyolo/v0/yolov6/yolov6_s_syncbn_fast_8xb32-400e_coco/yolov6_s_syncbn_fast_8xb32-400e_coco_20221102_203035-932e1d91.pth) \| [log](https://download.openmmlab.com/mmyolo/v0/yolov6/yolov6_s_syncbn_fast_8xb32-400e_coco/yolov6_s_syncbn_fast_8xb32-400e_coco_20221102_203035.log.json) |
+| yolov6_l_attentivenas_a6 | 640 | 45.3 | 18.38(-2.6%) | 8.49 | [config](https://github.com/open-mmlab/mmyolo/tree/dev/configs/razor/subnets/yolov6_l_attentivenas_a6_d12_syncbn_fast_8xb32-300e_coco.py) | [model](https://download.openmmlab.com/mmrazor/v1/yolo_nas_backbone/yolov6_l_attentivenas_a6_d12_syncbn_fast_8xb32-300e_coco_20230211_222409-dcc72668.pth) \| log |
+| RTMDet-tiny | 640 | 41.0 | 4.8 | 8.1 | [config](./rtmdet_l_syncbn_fast_8xb32-300e_coco.py) | [model](https://download.openmmlab.com/mmyolo/v0/rtmdet/rtmdet_tiny_syncbn_fast_8xb32-300e_coco/rtmdet_tiny_syncbn_fast_8xb32-300e_coco_20230102_140117-dbb1dc83.pth) \| [log](https://download.openmmlab.com/mmyolo/v0/rtmdet/rtmdet_tiny_syncbn_fast_8xb32-300e_coco/rtmdet_tiny_syncbn_fast_8xb32-300e_coco_20230102_140117.log.json) |
+| rtmdet_tiny_ofa_lat31 | 960 | 41.3 | 3.91(-18.5%) | 6.09 | [config](https://github.com/open-mmlab/mmyolo/tree/dev/configs/razor/subnets/rtmdet_tiny_ofa_lat31_syncbn_16xb16-300e_coco.py) | [model](https://download.openmmlab.com/mmrazor/v1/yolo_nas_backbone/rtmdet_tiny_ofa_lat31_syncbn_16xb16-300e_coco_20230214_210623-449bb2a0.pth) \| log |
**Note**:
1. For fair comparison, the training configuration is consistent with the original configuration and results in an improvement of about 0.2-0.5% AP.
-2. `yolov5_s_spos_shufflenetv2` achieves 37.9% AP with only 7.042M parameters, directly instead of the backbone, and outperforms `yolov5_s` with a similar size by more than 0.2% AP.
+2. `yolov5_s_spos_shufflenetv2` achieves 38.0% AP with only 7.042M parameters, directly instead of the backbone, and outperforms `yolov5_s` with a similar size by more than 0.3% AP.
3. With the efficient backbone of `yolov6_l_attentivenas_a6`, the input channels of `YOLOv6RepPAFPN` are reduced. Meanwhile, modify the **deepen_factor** and the neck is made deeper to restore the AP.
-4. with the `rtmdet_tiny_ofa_lat31` backbone with only 3.315M parameters and 3.634G flops, we can modify the input resolution to 960, with a similar model size compared to `rtmdet_tiny` and exceeds `rtmdet_tiny` by 0.1% AP, reducing the size of the whole model to 3.91 MB.
+4. with the `rtmdet_tiny_ofa_lat31` backbone with only 3.315M parameters and 3.634G flops, we can modify the input resolution to 960, with a similar model size compared to `rtmdet_tiny` and exceeds `rtmdet_tiny` by 0.4% AP, reducing the size of the whole model to 3.91 MB.
diff --git a/configs/razor/subnets/rtmdet_tiny_ofa_lat31_syncbn_16xb16-300e_coco.py b/configs/razor/subnets/rtmdet_tiny_ofa_lat31_syncbn_16xb16-300e_coco.py
index 82d696be..04d8c2d8 100644
--- a/configs/razor/subnets/rtmdet_tiny_ofa_lat31_syncbn_16xb16-300e_coco.py
+++ b/configs/razor/subnets/rtmdet_tiny_ofa_lat31_syncbn_16xb16-300e_coco.py
@@ -4,17 +4,13 @@ _base_ = [
]
checkpoint_file = 'https://download.openmmlab.com/mmrazor/v1/ofa/ofa_mobilenet_subnet_8xb256_in1k_note8_lat%4031ms_top1%4072.8_finetune%4025.py_20221214_0939-981a8b2a.pth' # noqa
-fix_subnet = 'https://download.openmmlab.com/mmrazor/v1/ofa/rtmdet/OFA_SUBNET_NOTE8_LAT31.yaml' # noqa
+fix_subnet = 'https://download.openmmlab.com/mmrazor/v1/yolo_nas_backbone/OFA_SUBNET_NOTE8_LAT31.yaml' # noqa
deepen_factor = 0.167
widen_factor = 1.0
channels = [40, 112, 160]
train_batch_size_per_gpu = 16
img_scale = (960, 960)
-_base_.base_lr = 0.002
-_base_.optim_wrapper.optimizer.lr = 0.002
-_base_.param_scheduler[1].eta_min = 0.002 * 0.05
-
_base_.nas_backbone.out_indices = (2, 4, 5)
_base_.nas_backbone.conv_cfg = dict(type='mmrazor.OFAConv2d')
_base_.nas_backbone.init_cfg = dict(
@@ -36,6 +32,14 @@ _base_.model.bbox_head.head_module.in_channels = channels[0]
_base_.model.bbox_head.head_module.feat_channels = channels[0]
_base_.model.bbox_head.head_module.widen_factor = widen_factor
+_base_.model.test_cfg = dict(
+ multi_label=True,
+ nms_pre=1000,
+ min_bbox_size=0,
+ score_thr=0.05,
+ nms=dict(type='nms', iou_threshold=0.6),
+ max_per_img=100)
+
train_pipeline = [
dict(type='LoadImageFromFile', file_client_args=_base_.file_client_args),
dict(type='LoadAnnotations', with_bbox=True),
@@ -43,12 +47,12 @@ train_pipeline = [
type='Mosaic',
img_scale=img_scale,
use_cached=True,
- max_cached_images=40,
+ max_cached_images=20,
+ random_pop=False,
pad_val=114.0),
dict(
type='mmdet.RandomResize',
- # img_scale is (width, height)
- scale=(img_scale[0] * 2, img_scale[1] * 2),
+ scale=(1280, 1280),
ratio_range=(0.5, 2.0), # note
resize_type='mmdet.Resize',
keep_ratio=True),
@@ -56,7 +60,15 @@ train_pipeline = [
dict(type='mmdet.YOLOXHSVRandomAug'),
dict(type='mmdet.RandomFlip', prob=0.5),
dict(type='mmdet.Pad', size=img_scale, pad_val=dict(img=(114, 114, 114))),
- dict(type='YOLOv5MixUp', use_cached=True, max_cached_images=20),
+ dict(
+ type='YOLOXMixUp',
+ img_scale=(960, 960),
+ ratio_range=(1.0, 1.0),
+ max_cached_images=10,
+ use_cached=True,
+ random_pop=False,
+ pad_val=(114, 114, 114),
+ prob=0.5),
dict(type='mmdet.PackDetInputs')
]
@@ -81,23 +93,17 @@ train_dataloader = dict(
test_pipeline = [
dict(type='LoadImageFromFile', file_client_args=_base_.file_client_args),
- dict(type='YOLOv5KeepRatioResize', scale=img_scale),
- dict(
- type='LetterResize',
- scale=img_scale,
- allow_scale_up=False,
- pad_val=dict(img=114)),
+ dict(type='mmdet.Resize', scale=(960, 960), keep_ratio=True),
+ dict(type='mmdet.Pad', size=(960, 960), pad_val=dict(img=(114, 114, 114))),
dict(type='LoadAnnotations', with_bbox=True, _scope_='mmdet'),
dict(
type='mmdet.PackDetInputs',
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
- 'scale_factor', 'pad_param'))
+ 'scale_factor'))
]
-batch_shapes_cfg = dict(img_size=img_scale[0])
-
val_dataloader = dict(
- dataset=dict(pipeline=test_pipeline, batch_shapes_cfg=batch_shapes_cfg))
+ dataset=dict(pipeline=test_pipeline, batch_shapes_cfg=None))
test_dataloader = val_dataloader
diff --git a/configs/razor/subnets/yolov5_s_spos_shufflenetv2_syncbn_8xb16-300e_coco.py b/configs/razor/subnets/yolov5_s_spos_shufflenetv2_syncbn_8xb16-300e_coco.py
index 39884047..beb4941c 100644
--- a/configs/razor/subnets/yolov5_s_spos_shufflenetv2_syncbn_8xb16-300e_coco.py
+++ b/configs/razor/subnets/yolov5_s_spos_shufflenetv2_syncbn_8xb16-300e_coco.py
@@ -5,7 +5,6 @@ _base_ = [
checkpoint_file = 'https://download.openmmlab.com/mmrazor/v1/spos/spos_shufflenetv2_subnet_8xb128_in1k_flops_0.33M_acc_73.87_20211222-1f0a0b4d_v3.pth' # noqa
fix_subnet = 'https://download.openmmlab.com/mmrazor/v1/spos/spos_shufflenetv2_subnet_8xb128_in1k_flops_0.33M_acc_73.87_20211222-1f0a0b4d_subnet_cfg_v3.yaml' # noqa
-
widen_factor = 1.0
channels = [160, 320, 640]
diff --git a/configs/razor/subnets/yolov6_l_attentivenas_a6_d12_syncbn_fast_16xb16-300e_coco.py b/configs/razor/subnets/yolov6_l_attentivenas_a6_d12_syncbn_fast_8xb32-300e_coco.py
similarity index 100%
rename from configs/razor/subnets/yolov6_l_attentivenas_a6_d12_syncbn_fast_16xb16-300e_coco.py
rename to configs/razor/subnets/yolov6_l_attentivenas_a6_d12_syncbn_fast_8xb32-300e_coco.py
diff --git a/tests/test_downstream/test_mmrazor.py b/tests/test_downstream/test_mmrazor.py
index ebf6806e..dc3090d2 100644
--- a/tests/test_downstream/test_mmrazor.py
+++ b/tests/test_downstream/test_mmrazor.py
@@ -11,9 +11,9 @@ from mmyolo.testing import get_detector_cfg
'razor/subnets/'
'yolov5_s_spos_shufflenetv2_syncbn_8xb16-300e_coco.py', 'razor/subnets/'
'rtmdet_tiny_ofa_lat31_syncbn_16xb16-300e_coco.py', 'razor/subnets/'
- 'yolov6_l_attentivenas_a6_d12_syncbn_fast_16xb16-300e_coco.py'
+ 'yolov6_l_attentivenas_a6_d12_syncbn_fast_8xb32-300e_coco.py'
])
-def test_razor_backbone_forward(cfg_file):
+def test_razor_backbone_init(cfg_file):
model = get_detector_cfg(cfg_file)
model_cfg = copy.deepcopy(model.backbone)
from mmrazor.registry import MODELS