Skip to content

Commit d8204d3

Browse files
rwightmanrwightman
committed
Add EfficientNetV2-Small (RW variant) based model weights efficientdetv2_ds w/ 50.1mAP
1 parent cf15a97 commit d8204d3

File tree

3 files changed

+39
-35
lines changed

3 files changed

+39
-35
lines changed

‎README.md

Lines changed: 37 additions & 33 deletions
Original file line numberDiff line numberDiff line change
@@ -16,6 +16,9 @@ Aside from the default model configs, there is a lot of flexibility to facilitat
1616

1717
## Updates
1818

19+
### 2022-01-06
20+
* New `efficientnetv2_ds` weights 50.1 mAP @ 1024x0124, using AGC clipping. Memory use comparable to D3, speed faster than D4. Smaller than optimal training batch size so can probably do better...
21+
1922
### 2021-11-30
2023
* Update `efficientnetv2_dt` weights to a new set, 46.1 mAP @ 768x768, 47.0 mAP @ 896x896 using AGC clipping.
2124
* Add AGC (Adaptive Gradient Clipping support via `timm`). Idea from (`High-Performance Large-Scale Image Recognition Without Normalization` - https://arxiv.org/abs/2102.06171)
@@ -95,39 +98,40 @@ Training sanity checks were done on VOC and OI
9598

9699
The table below contains models with pretrained weights. There are quite a number of other models that I have defined in [model configurations](effdet/config/model_config.py) that use various `timm` backbones.
97100

98-
| Variant | mAP (val2017) | mAP (test-dev2017) | mAP (TF official val2017) | mAP (TF official test-dev2017) | Params (M) | Img Size |
99-
| --- | :---: | :---: | :---: | :---: | :---: | :---: |
100-
| tf_efficientdet_lite0 | 27.1 | TBD | 26.4 | N/A | 3.24 | 320 |
101-
| tf_efficientdet_lite1 | 32.2 | TBD | 31.5 | N/A | 4.25 | 384 |
102-
| efficientdet_d0 | 33.6 | TBD | N/A | N/A | 3.88 | 512 |
103-
| tf_efficientdet_d0 | 34.2 | TBD | 34.3 | 34.6 | 3.88 | 512 |
104-
| tf_efficientdet_d0_ap | 34.8 | TBD | 35.2 | 35.3 | 3.88 | 512 |
105-
| efficientdet_q0 | 35.7 | TBD | N/A | N/A | 4.13 | 512 |
106-
| tf_efficientdet_lite2 | 35.9 | TBD | 35.1 | N/A | 5.25 | 448 |
107-
| efficientdet_d1 | 39.4 | 39.5 | N/A | N/A | 6.62 | 640 |
108-
| tf_efficientdet_lite3 | 39.6 | TBD | 38.8 | N/A | 8.35 | 512 |
109-
| tf_efficientdet_d1 | 40.1 | TBD | 40.2 | 40.5 | 6.63 | 640 |
110-
| tf_efficientdet_d1_ap | 40.8 | TBD | 40.9 | 40.8 | 6.63 | 640 |
111-
| efficientdet_q1 | 40.9 | TBD | N/A | N/A | 6.98 | 640 |
112-
| cspresdext50pan | 41.2 | TBD | N/A | N/A | 22.2 | 640 |
113-
| resdet50 | 41.6 | TBD | N/A | N/A | 27.6 | 640 |
114-
| efficientdet_q2 | 43.1 | TBD | N/A | N/A | 8.81 | 768 |
115-
| cspresdet50 | 43.2 | TBD | N/A | N/A | 24.3 | 768 |
116-
| tf_efficientdet_d2 | 43.4 | TBD | 42.5 | 43 | 8.10 | 768 |
117-
| tf_efficientdet_lite3x | 43.6 | TBD | 42.6 | N/A | 9.28 | 640 |
118-
| tf_efficientdet_lite4 | 44.2 | TBD | 43.2 | N/A | 15.1 | 640 |
119-
| tf_efficientdet_d2_ap | 44.2 | TBD | 44.3 | 44.3 | 8.10 | 768 |
120-
| cspdarkdet53m | 45.2 | TBD | N/A | N/A | 35.6 | 768 |
121-
| efficientdetv2_dt | 45.8 | TBD | N/A | N/A | 13.4 | 768 |
122-
| tf_efficientdet_d3 | 47.1 | TBD | 47.2 | 47.5 | 12.0 | 896 |
123-
| tf_efficientdet_d3_ap | 47.7 | TBD | 48.0 | 47.7 | 12.0 | 896 |
124-
| tf_efficientdet_d4 | 49.2 | TBD | 49.3 | 49.7 | 20.7 | 1024 |
125-
| tf_efficientdet_d4_ap | 50.2 | TBD | 50.4 | 50.4 | 20.7 | 1024 |
126-
| tf_efficientdet_d5 | 51.2 | TBD | 51.2 | 51.5 | 33.7 | 1280 |
127-
| tf_efficientdet_d6 | 52.0 | TBD | 52.1 | 52.6 | 51.9 | 1280 |
128-
| tf_efficientdet_d5_ap | 52.1 | TBD | 52.2 | 52.5 | 33.7 | 1280 |
129-
| tf_efficientdet_d7 | 53.1 | 53.4 | 53.4 | 53.7 | 51.9 | 1536 |
130-
| tf_efficientdet_d7x | 54.3 | TBD | 54.4 | 55.1 | 77.1 | 1536 |
101+
| Variant | mAP (val2017) | mAP (test-dev2017) | mAP (TF official val2017) | mAP (TF official test-dev2017) | Params (M) | Img Size |
102+
|------------------------|:-------------:| :---: | :---: | :---: |:----------:|:--------:|
103+
| tf_efficientdet_lite0 | 27.1 | TBD | 26.4 | N/A | 3.24 | 320 |
104+
| tf_efficientdet_lite1 | 32.2 | TBD | 31.5 | N/A | 4.25 | 384 |
105+
| efficientdet_d0 | 33.6 | TBD | N/A | N/A | 3.88 | 512 |
106+
| tf_efficientdet_d0 | 34.2 | TBD | 34.3 | 34.6 | 3.88 | 512 |
107+
| tf_efficientdet_d0_ap | 34.8 | TBD | 35.2 | 35.3 | 3.88 | 512 |
108+
| efficientdet_q0 | 35.7 | TBD | N/A | N/A | 4.13 | 512 |
109+
| tf_efficientdet_lite2 | 35.9 | TBD | 35.1 | N/A | 5.25 | 448 |
110+
| efficientdet_d1 | 39.4 | 39.5 | N/A | N/A | 6.62 | 640 |
111+
| tf_efficientdet_lite3 | 39.6 | TBD | 38.8 | N/A | 8.35 | 512 |
112+
| tf_efficientdet_d1 | 40.1 | TBD | 40.2 | 40.5 | 6.63 | 640 |
113+
| tf_efficientdet_d1_ap | 40.8 | TBD | 40.9 | 40.8 | 6.63 | 640 |
114+
| efficientdet_q1 | 40.9 | TBD | N/A | N/A | 6.98 | 640 |
115+
| cspresdext50pan | 41.2 | TBD | N/A | N/A | 22.2 | 640 |
116+
| resdet50 | 41.6 | TBD | N/A | N/A | 27.6 | 640 |
117+
| efficientdet_q2 | 43.1 | TBD | N/A | N/A | 8.81 | 768 |
118+
| cspresdet50 | 43.2 | TBD | N/A | N/A | 24.3 | 768 |
119+
| tf_efficientdet_d2 | 43.4 | TBD | 42.5 | 43 | 8.10 | 768 |
120+
| tf_efficientdet_lite3x | 43.6 | TBD | 42.6 | N/A | 9.28 | 640 |
121+
| tf_efficientdet_lite4 | 44.2 | TBD | 43.2 | N/A | 15.1 | 640 |
122+
| tf_efficientdet_d2_ap | 44.2 | TBD | 44.3 | 44.3 | 8.10 | 768 |
123+
| cspdarkdet53m | 45.2 | TBD | N/A | N/A | 35.6 | 768 |
124+
| efficientdetv2_dt | 46.1 | TBD | N/A | N/A | 13.4 | 768 |
125+
| tf_efficientdet_d3 | 47.1 | TBD | 47.2 | 47.5 | 12.0 | 896 |
126+
| tf_efficientdet_d3_ap | 47.7 | TBD | 48.0 | 47.7 | 12.0 | 896 |
127+
| tf_efficientdet_d4 | 49.2 | TBD | 49.3 | 49.7 | 20.7 | 1024 |
128+
| efficientdetv2_ds | 50.1 | TBD | N/A | N/A | 26.6 | 1024 |
129+
| tf_efficientdet_d4_ap | 50.2 | TBD | 50.4 | 50.4 | 20.7 | 1024 |
130+
| tf_efficientdet_d5 | 51.2 | TBD | 51.2 | 51.5 | 33.7 | 1280 |
131+
| tf_efficientdet_d6 | 52.0 | TBD | 52.1 | 52.6 | 51.9 | 1280 |
132+
| tf_efficientdet_d5_ap | 52.1 | TBD | 52.2 | 52.5 | 33.7 | 1280 |
133+
| tf_efficientdet_d7 | 53.1 | 53.4 | 53.4 | 53.7 | 51.9 | 1536 |
134+
| tf_efficientdet_d7x | 54.3 | TBD | 54.4 | 55.1 | 77.1 | 1536 |
131135

132136

133137
See [model configurations](effdet/config/model_config.py) for model checkpoint urls and differences.

‎effdet/config/model_config.py

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -187,7 +187,7 @@ def default_detection_model_configs():
187187
redundant_bias=False,
188188
head_bn_level_first=True,
189189
backbone_args=dict(drop_path_rate=0.2),
190-
url='', # I need to find a machine to train this on
190+
url='https://github.com/rwightman/efficientdet-pytorch/releases/download/v0.1/efficientnetv2_rw_ds_agc-cf589293.pth',
191191
),
192192

193193

‎effdet/version.py

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -1 +1 @@
1-
__version__ = '0.2.6'
1+
__version__ = '0.3.0'

0 commit comments

Comments
 (0)