未验证 提交 653604c0 编写于 作者: F Feng Ni 提交者: GitHub

Fix swin and add swin ppyoloe (#7857)

* refine swin configs and codes

* fix swin ppyoloe

* fix swin for ema and distill training

* fix configs for CI

* fix docs, test=document_fix
上级 237b19d9
......@@ -23,7 +23,7 @@
| ResNet50-vd-SSLDv2-FPN | Faster | 1 | 2x | ---- | 42.3 | [下载链接](https://paddledet.bj.bcebos.com/models/faster_rcnn_r50_vd_fpn_ssld_2x_coco.pdparams) | [配置文件](./faster_rcnn_r50_vd_fpn_ssld_2x_coco.yml) |
| Swin-Tiny-FPN | Faster | 2 | 1x | ---- | 42.6 | [下载链接](https://paddledet.bj.bcebos.com/models/faster_rcnn_swin_tiny_fpn_1x_coco.pdparams) | [配置文件](./faster_rcnn_swin_tiny_fpn_1x_coco.yml) |
| Swin-Tiny-FPN | Faster | 2 | 2x | ---- | 44.8 | [下载链接](https://paddledet.bj.bcebos.com/models/faster_rcnn_swin_tiny_fpn_2x_coco.pdparams) | [配置文件](./faster_rcnn_swin_tiny_fpn_2x_coco.yml) |
| Swin-Tiny-FPN | Faster | 2 | 3x | ---- | 45.3 | [下载链接](https://paddledet.bj.bcebos.com/models/faster_rcnn_swin_tiny_fpn_3x_coco.pdparams) | [配置文件](./faster_rcnn_swin_tiny_fpn_3x_coco.yml) |
| Swin-Tiny-FPN | Faster | 2 | 3x | ---- | 45.3 | [下载链接](https://paddledet.bj.bcebos.com/models/faster_rcnn_swin_tiny_fpn_3x_coco.pdparams) | [配置文件](../swin/faster_rcnn_swin_tiny_fpn_3x_coco.yml) |
## Citations
```
......
......@@ -15,8 +15,6 @@ OptimizerBuilder:
optimizer:
type: AdamW
weight_decay: 0.05
param_groups:
-
params: ['absolute_pos_embed', 'relative_position_bias_table', 'norm']
weight_decay: 0.
- params: ['absolute_pos_embed', 'relative_position_bias_table', 'norm']
weight_decay: 0.0
......@@ -14,9 +14,3 @@ LearningRate:
- !LinearWarmup
start_factor: 0.1
steps: 1000
OptimizerBuilder:
clip_grad_by_norm: 1.0
optimizer:
type: AdamW
weight_decay: 0.05
_BASE_: [
'faster_rcnn_swin_tiny_fpn_1x_coco.yml',
]
weights: output/faster_rcnn_swin_tiny_fpn_3x_coco/model_final
epoch: 36
LearningRate:
base_lr: 0.0001
schedulers:
- !PiecewiseDecay
gamma: 0.1
milestones: [24, 33]
- !LinearWarmup
start_factor: 0.1
steps: 1000
OptimizerBuilder:
clip_grad_by_norm: 1.0
optimizer:
type: AdamW
weight_decay: 0.05
# Swin Transformer
## COCO Model Zoo
| 骨架网络 | 网络类型 | 每张GPU图片个数 | 学习率策略 |推理时间(fps) | Box AP | 下载 | 配置文件 |
| :------------------- | :------------- | :-----: | :-----: | :------------: | :-----: | :-----------------------------------------------------: | :-----: |
| swin_T_224 | Faster R-CNN | 2 | 36e | ---- | 45.3 | [下载链接](https://paddledet.bj.bcebos.com/models/faster_rcnn_swin_tiny_fpn_3x_coco.pdparams) | [配置文件](./faster_rcnn_swin_tiny_fpn_3x_coco.yml) |
| swin_T_224 | PP-YOLOE+ | 8 | 36e | ---- | 43.6 | [下载链接](https://paddledet.bj.bcebos.com/models/ppyoloe_plus_swin_tiny_36e_coco.pdparams) | [配置文件](./ppyoloe_plus_swin_tiny_36e_coco.yml) |
## Citations
```
@article{liu2021Swin,
title={Swin Transformer: Hierarchical Vision Transformer using Shifted Windows},
author={Liu, Ze and Lin, Yutong and Cao, Yue and Hu, Han and Wei, Yixuan and Zhang, Zheng and Lin, Stephen and Guo, Baining},
journal={arXiv preprint arXiv:2103.14030},
year={2021}
}
@inproceedings{liu2021swinv2,
title={Swin Transformer V2: Scaling Up Capacity and Resolution},
author={Ze Liu and Han Hu and Yutong Lin and Zhuliang Yao and Zhenda Xie and Yixuan Wei and Jia Ning and Yue Cao and Zheng Zhang and Li Dong and Furu Wei and Baining Guo},
booktitle={International Conference on Computer Vision and Pattern Recognition (CVPR)},
year={2022}
}
```
_BASE_: [
'../datasets/coco_detection.yml',
'../runtime.yml',
'../faster_rcnn/_base_/faster_rcnn_r50_fpn.yml',
'../faster_rcnn/_base_/faster_fpn_reader.yml',
]
weights: output/faster_rcnn_swin_tiny_fpn_3x_coco/model_final
pretrain_weights: https://paddledet.bj.bcebos.com/models/pretrained/swin_tiny_patch4_window7_224_22kto1k_pretrained.pdparams
FasterRCNN:
backbone: SwinTransformer
neck: FPN
rpn_head: RPNHead
bbox_head: BBoxHead
bbox_post_process: BBoxPostProcess
SwinTransformer:
arch: 'swin_T_224' # ['swin_T_224', 'swin_S_224', 'swin_B_224', 'swin_L_224', 'swin_B_384', 'swin_L_384']
ape: false
drop_path_rate: 0.1
patch_norm: true
out_indices: [0, 1, 2, 3]
worker_num: 2
TrainReader:
sample_transforms:
- Decode: {}
- RandomResizeCrop: {resizes: [400, 500, 600], cropsizes: [[384, 600], ], prob: 0.5}
- RandomResize: {target_size: [[480, 1333], [512, 1333], [544, 1333], [576, 1333], [608, 1333], [640, 1333], [672, 1333], [704, 1333], [736, 1333], [768, 1333], [800, 1333]], keep_ratio: True, interp: 2}
- RandomFlip: {prob: 0.5}
- NormalizeImage: {is_scale: true, mean: [0.485, 0.456, 0.406], std: [0.229, 0.224, 0.225]}
- Permute: {}
batch_transforms:
- PadBatch: {pad_to_stride: 32}
batch_size: 2
shuffle: true
drop_last: true
collate_batch: false
EvalReader:
sample_transforms:
- Decode: {}
- Resize: {interp: 2, target_size: [800, 1333], keep_ratio: True}
- NormalizeImage: {is_scale: true, mean: [0.485, 0.456, 0.406], std: [0.229, 0.224, 0.225]}
- Permute: {}
batch_transforms:
- PadBatch: {pad_to_stride: 32}
batch_size: 1
TestReader:
inputs_def:
image_shape: [-1, 3, 640, 640] # TODO deploy: set fixes shape currently
sample_transforms:
- Decode: {}
- Resize: {interp: 2, target_size: 640, keep_ratio: True}
- Pad: {size: 640}
- NormalizeImage: {is_scale: true, mean: [0.485, 0.456, 0.406], std: [0.229, 0.224, 0.225]}
- Permute: {}
batch_size: 1
epoch: 36
LearningRate:
base_lr: 0.0001
schedulers:
- !PiecewiseDecay
gamma: 0.1
milestones: [24, 33]
- !LinearWarmup
start_factor: 0.1
steps: 1000
OptimizerBuilder:
clip_grad_by_norm: 1.0
optimizer:
type: AdamW
weight_decay: 0.05
param_groups:
- params: ['absolute_pos_embed', 'relative_position_bias_table', 'norm']
weight_decay: 0.0
_BASE_: [
'../datasets/coco_detection.yml',
'../runtime.yml',
'../ppyoloe/_base_/ppyoloe_plus_crn.yml',
'../ppyoloe/_base_/ppyoloe_plus_reader.yml',
]
depth_mult: 0.33 # s version
width_mult: 0.50
log_iter: 50
snapshot_epoch: 4
weights: output/ppyoloe_plus_swin_tiny_36e_coco/model_final
pretrain_weights: https://paddledet.bj.bcebos.com/models/pretrained/swin_tiny_patch4_window7_224_22kto1k_pretrained.pdparams
architecture: PPYOLOE
norm_type: sync_bn
use_ema: true
ema_decay: 0.9998
ema_black_list: ['proj_conv.weight']
custom_black_list: ['reduce_mean']
PPYOLOE:
backbone: SwinTransformer
neck: CustomCSPPAN
yolo_head: PPYOLOEHead
post_process: ~
SwinTransformer:
arch: 'swin_T_224' # ['swin_T_224', 'swin_S_224', 'swin_B_224', 'swin_L_224', 'swin_B_384', 'swin_L_384']
ape: false
drop_path_rate: 0.1
patch_norm: true
out_indices: [1, 2, 3]
PPYOLOEHead:
static_assigner_epoch: 12
nms:
nms_top_k: 10000
keep_top_k: 300
score_threshold: 0.01
nms_threshold: 0.7
TrainReader:
batch_size: 8
epoch: 36
LearningRate:
base_lr: 0.0001
schedulers:
- !PiecewiseDecay
gamma: 0.1
milestones: [24, 33]
- !LinearWarmup
start_factor: 0.1
steps: 1000
OptimizerBuilder:
clip_grad_by_norm: 1.0
optimizer:
type: AdamW
weight_decay: 0.05
param_groups:
- params: ['absolute_pos_embed', 'relative_position_bias_table', 'norm']
weight_decay: 0.0
......@@ -191,8 +191,6 @@ class WindowAttention(nn.Layer):
relative_coords[:, :, 1] += self.window_size[1] - 1
relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
self.relative_position_index = relative_coords.sum(-1) # Wh*Ww, Wh*Ww
self.register_buffer("relative_position_index",
self.relative_position_index)
self.qkv = nn.Linear(dim, dim * 3, bias_attr=qkv_bias)
self.attn_drop = nn.Dropout(attn_drop)
......@@ -425,7 +423,6 @@ class BasicLayer(nn.Layer):
""" A basic Swin Transformer layer for one stage.
Args:
dim (int): Number of input channels.
input_resolution (tuple[int]): Input resolution.
depth (int): Number of blocks.
num_heads (int): Number of attention heads.
window_size (int): Local window size.
......@@ -500,10 +497,7 @@ class BasicLayer(nn.Layer):
cnt = 0
for h in h_slices:
for w in w_slices:
try:
img_mask[:, h, w, :] = cnt
except:
pass
img_mask[:, h, w, :] = cnt
cnt += 1
......@@ -572,15 +566,12 @@ class PatchEmbed(nn.Layer):
@register
@serializable
class SwinTransformer(nn.Layer):
""" Swin Transformer
A PaddlePaddle impl of : `Swin Transformer: Hierarchical Vision Transformer using Shifted Windows` -
https://arxiv.org/pdf/2103.14030
""" Swin Transformer backbone
Args:
img_size (int | tuple(int)): Input image size. Default 224
arch (str): Architecture of FocalNet
pretrain_img_size (int | tuple(int)): Input image size. Default 224
patch_size (int | tuple(int)): Patch size. Default: 4
in_chans (int): Number of input image channels. Default: 3
num_classes (int): Number of classes for classification head. Default: 1000
embed_dim (int): Patch embedding dimension. Default: 96
depths (tuple(int)): Depth of each Swin Transformer layer.
num_heads (tuple(int)): Number of attention heads in different layers.
......@@ -619,6 +610,7 @@ class SwinTransformer(nn.Layer):
pretrained=None):
super(SwinTransformer, self).__init__()
assert arch in MODEL_cfg.keys(), "Unsupported arch: {}".format(arch)
pretrain_img_size = MODEL_cfg[arch]['pretrain_img_size']
embed_dim = MODEL_cfg[arch]['embed_dim']
depths = MODEL_cfg[arch]['depths']
......@@ -748,7 +740,7 @@ class SwinTransformer(nn.Layer):
(0, 3, 1, 2))
outs.append(out)
return tuple(outs)
return outs
@property
def out_shape(self):
......
......@@ -236,7 +236,7 @@ def get_sine_pos_embed(pos_tensor,
"""generate sine position embedding from a position tensor
Args:
pos_tensor (torch.Tensor): Shape as `(None, n)`.
pos_tensor (Tensor): Shape as `(None, n)`.
num_pos_feats (int): projected shape for each float in the tensor. Default: 128
temperature (int): The temperature used for scaling
the position embedding. Default: 10000.
......@@ -245,7 +245,7 @@ def get_sine_pos_embed(pos_tensor,
be `[pos(y), pos(x)]`. Defaults: True.
Returns:
torch.Tensor: Returned position embedding # noqa
Tensor: Returned position embedding # noqa
with shape `(None, n * num_pos_feats)`.
"""
scale = 2. * math.pi
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册