Skip to content

Commit c4b6947

Browse files
lxg39lxg39
authored
feat: add training configs and training weights of Volo(d2,d3,d4) (#731)
1 parent b8a6782 commit c4b6947

File tree

5 files changed

+209
-5
lines changed

5 files changed

+209
-5
lines changed

‎configs/volo/README.md

Lines changed: 3 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -22,7 +22,9 @@ Our reproduced model performance on ImageNet-1K is reported as follows.
2222
| Model | Context | Top-1 (%) | Top-5 (%) | Params (M) | Recipe | Weight |
2323
|-----------------|-----------|-------|------------|------------------------------------------------------------------------------------------|----------------------------------------------------------------------------------|----------------------------------------------------------------------------------|
2424
| volo_d1 | D910x8-G | 82.59 | 95.99 | 27 | [yaml](https://github.com/mindspore-lab/mindcv/blob/main/configs/volo/volo_d1_ascend.yaml) | [weights](https://download.mindspore.cn/toolkits/mindcv/volo/volo_d1-c7efada9.ckpt) |
25-
25+
| volo_d2 | D910x8-G | 82.95 | 96.13 | 59 | [yaml](https://github.com/mindspore-lab/mindcv/blob/main/configs/volo/volo_d2_ascend.yaml) | [weights](https://download.mindspore.cn/toolkits/mindcv/volo/volo_d2-0910a460.ckpt) |
26+
| volo_d3 | D910x8-G | 83.38 | 96.28 | 87 | [yaml](https://github.com/mindspore-lab/mindcv/blob/main/configs/volo/volo_d3_ascend.yaml) | [weights](https://download.mindspore.cn/toolkits/mindcv/volo/volo_d3-25916c36.ckpt) |
27+
| volo_d4 | D910x8-G | 82.5 | 95.86 | 193 | [yaml](https://github.com/mindspore-lab/mindcv/blob/main/configs/volo/volo_d4_ascend.yaml) | [weights](https://download.mindspore.cn/toolkits/mindcv/volo/volo_d4-6c88cd33.ckpt) |
2628
</div>
2729

2830
#### Notes

‎configs/volo/volo_d2_ascend.yaml

Lines changed: 67 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,67 @@
1+
# system
2+
mode: 0
3+
distribute: True
4+
num_parallel_workers: 8
5+
val_while_train: True
6+
val_interval: 1
7+
8+
# dataset
9+
dataset: 'imagenet'
10+
data_dir: '/path/to/imagenet'
11+
shuffle: True
12+
dataset_download: False
13+
batch_size: 128
14+
drop_remainder: True
15+
16+
# augmentation
17+
image_resize: 224
18+
scale: [0.08, 1.0]
19+
ratio: [0.75, 1.33]
20+
hflip: 0.5
21+
vflip: 0.
22+
interpolation: 'bicubic'
23+
auto_augment: 'randaug-m9-mstd0.5-inc1'
24+
re_prob: 0.25
25+
mixup: 0.2
26+
cutmix: 1.0
27+
cutmix_prob: 1.0
28+
crop_pct: 0.96
29+
color_jitter: [0.4, 0.4, 0.4]
30+
31+
# model config
32+
model: 'volo_d2'
33+
num_classes: 1000
34+
pretrained: False
35+
ckpt_path: ''
36+
keep_checkpoint_max: 10
37+
ckpt_save_dir: './ckpt/'
38+
ckpt_save_policy: 'top_k'
39+
drop_path_rate: 0.2
40+
dataset_sink_mode: True
41+
amp_level: 'O3'
42+
val_amp_level: 'O3'
43+
ema: True
44+
ema_decay: 0.9995
45+
46+
# loss
47+
loss: 'CE'
48+
label_smoothing: 0.1
49+
50+
# lr scheduler
51+
scheduler: 'warmup_cosine_decay'
52+
lr: 0.0007
53+
min_lr: 0.00001
54+
epoch_size: 300
55+
warmup_epochs: 10
56+
decay_epochs: 290
57+
decay_rate: 0.1
58+
59+
# optimizer
60+
opt: 'adamw'
61+
weight_decay: 0.05
62+
momentum: 0.9
63+
filter_bias_and_bn: True
64+
loss_scale_type: 'dynamic'
65+
loss_scale: 2048
66+
use_nesterov: False
67+
drop_overflow_update: True

‎configs/volo/volo_d3_ascend.yaml

Lines changed: 68 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,68 @@
1+
# system
2+
mode: 0
3+
distribute: True
4+
num_parallel_workers: 8
5+
val_while_train: True
6+
val_interval: 1
7+
8+
# dataset
9+
dataset: 'imagenet'
10+
data_dir: '/path/to/imagenet'
11+
shuffle: True
12+
dataset_download: False
13+
batch_size: 64
14+
drop_remainder: True
15+
16+
# augmentation
17+
image_resize: 224
18+
scale: [0.08, 1.0]
19+
ratio: [0.75, 1.33]
20+
hflip: 0.5
21+
vflip: 0.
22+
interpolation: 'bicubic'
23+
auto_augment: 'randaug-m9-mstd0.5-inc1'
24+
re_prob: 0.25
25+
mixup: 0.2
26+
cutmix: 1.0
27+
cutmix_prob: 1.0
28+
crop_pct: 0.96
29+
color_jitter: [0.4, 0.4, 0.4]
30+
31+
# model
32+
model: 'volo_d3'
33+
num_classes: 1000
34+
pretrained: False
35+
ckpt_path: ''
36+
keep_checkpoint_max: 10
37+
ckpt_save_dir: './ckpt'
38+
ckpt_save_policy: 'top_k'
39+
drop_path_rate: 0.5
40+
dataset_sink_mode: True
41+
amp_level: 'O3'
42+
val_amp_level: 'O3'
43+
ema: True
44+
ema_decay: 0.9995
45+
gradient_accumulation_steps: 2
46+
47+
# loss
48+
loss: 'CE'
49+
label_smoothing: 0.1
50+
51+
# lr scheduler
52+
scheduler: 'warmup_cosine_decay'
53+
lr: 0.0005
54+
min_lr: 0.00001
55+
epoch_size: 3
56+
warmup_epochs: 10
57+
decay_epochs: 290
58+
decay_rate: 0.1
59+
60+
# optimizer
61+
opt: 'adamw'
62+
weight_decay: 0.05
63+
momentum: 0.9
64+
filter_bias_and_bn: True
65+
loss_scale_type: 'dynamic'
66+
loss_scale: 1024
67+
use_nesterov: False
68+
drop_overflow_update: True

‎configs/volo/volo_d4_ascend.yaml

Lines changed: 67 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,67 @@
1+
# system
2+
mode: 0
3+
distribute: True
4+
num_parallel_workers: 8
5+
val_while_train: True
6+
val_interval: 1
7+
8+
# dataset
9+
dataset: 'imagenet'
10+
data_dir: '/path/to/imagenet'
11+
shuffle: True
12+
dataset_download: False
13+
batch_size: 32
14+
drop_remainder: True
15+
16+
# augmentation
17+
image_resize: 224
18+
scale: [0.08, 1.0]
19+
ratio: [0.75, 1.33]
20+
hflip: 0.5
21+
vflip: 0.
22+
interpolation: 'bicubic'
23+
auto_augment: 'randaug-m9-mstd0.5-inc1'
24+
re_prob: 0.25
25+
mixup: 0.2
26+
cutmix: 1.0
27+
cutmix_prob: 1.0
28+
crop_pct: 1.15
29+
color_jitter: [0.4, 0.4, 0.4]
30+
31+
# model
32+
model: 'volo_d4'
33+
num_classes: 1000
34+
pretrained: False
35+
ckpt_path: ''
36+
keep_checkpoint_max: 10
37+
ckpt_save_dir: './ckpt'
38+
ckpt_save_policy: 'top_k'
39+
drop_path_rate: 0.5
40+
dataset_sink_mode: True
41+
amp_level: 'O3'
42+
val_amp_level: 'O3'
43+
ema: True
44+
ema_decay: 0.9995
45+
46+
# loss
47+
loss: 'CE'
48+
label_smoothing: 0.1
49+
50+
# lr scheduler
51+
scheduler: 'warmup_cosine_decay'
52+
lr: 0.00025
53+
min_lr: 0.00001
54+
epoch_size: 300
55+
warmup_epochs: 10
56+
decay_epochs: 290
57+
decay_rate: 0.1
58+
59+
# optimizer
60+
opt: 'adamw'
61+
weight_decay: 0.05
62+
momentum: 0.9
63+
filter_bias_and_bn: True
64+
loss_scale_type: 'dynamic'
65+
loss_scale: 1024
66+
use_nesterov: False
67+
drop_overflow_update: True

‎mindcv/models/volo.py

Lines changed: 4 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -43,10 +43,10 @@ def _cfg(url='', **kwargs):
4343

4444

4545
default_cfgs = {
46-
'volo_d1': _cfg(url=''),
47-
'volo_d2': _cfg(url=''),
48-
'volo_d3': _cfg(url=''),
49-
'volo_d4': _cfg(url='', crop_pct=1.15),
46+
'volo_d1': _cfg(url='https://download.mindspore.cn/toolkits/mindcv/volo/volo_d1-c7efada9.ckpt'),
47+
'volo_d2': _cfg(url='https://download.mindspore.cn/toolkits/mindcv/volo/volo_d2-0910a460.ckpt'),
48+
'volo_d3': _cfg(url='https://download.mindspore.cn/toolkits/mindcv/volo/volo_d3-25916c36.ckpt'),
49+
'volo_d4': _cfg(url='https://download.mindspore.cn/toolkits/mindcv/volo/volo_d4-6c88cd33.ckpt', crop_pct=1.15),
5050
'volo_d5': _cfg(url='', crop_pct=1.15)
5151
}
5252

0 commit comments

Comments
 (0)