未验证 提交 582d3c25 编写于 作者: C ceci3 提交者: GitHub

add blazeface nas1 result based on latency (#235)


* add blazeface_nas1
上级 ce624b3a
architecture: BlazeFace
max_iters: 320000
pretrain_weights:
use_gpu: true
snapshot_iter: 10000
log_smooth_window: 20
log_iter: 20
metric: WIDERFACE
save_dir: output
weights: output/blazeface_nas/model_final/
# 1(label_class) + 1(background)
num_classes: 2
BlazeFace:
backbone: BlazeNet
output_decoder:
keep_top_k: 750
nms_threshold: 0.3
nms_top_k: 5000
score_threshold: 0.01
min_sizes: [[16.,24.], [32., 48., 64., 80., 96., 128.]]
use_density_prior_box: false
BlazeNet:
blaze_filters: [[12, 12], [12, 12, 2], [12, 12]]
double_blaze_filters: [[12, 16, 32, 2], [32, 32, 32], [32, 16, 72, 2], [72, 24, 72]]
with_extra_blocks: true
lite_edition: false
LearningRate:
base_lr: 0.001
schedulers:
- !PiecewiseDecay
gamma: 0.1
milestones: [240000, 300000]
OptimizerBuilder:
optimizer:
momentum: 0.0
type: RMSPropOptimizer
regularizer:
factor: 0.0005
type: L2
TrainReader:
inputs_def:
image_shape: [3, 640, 640]
fields: ['image', 'gt_bbox', 'gt_class']
dataset:
!WIDERFaceDataSet
dataset_dir: dataset/wider_face
anno_path: wider_face_split/wider_face_train_bbx_gt.txt
image_dir: WIDER_train/images
sample_transforms:
- !DecodeImage
to_rgb: true
- !NormalizeBox {}
- !RandomDistort
brightness_lower: 0.875
brightness_upper: 1.125
is_order: true
- !ExpandImage
max_ratio: 4
prob: 0.5
- !CropImageWithDataAchorSampling
anchor_sampler:
- [1, 10, 1.0, 1.0, 1.0, 1.0, 0.0, 0.0, 0.2, 0.0]
batch_sampler:
- [1, 50, 1.0, 1.0, 1.0, 1.0, 0.0, 0.0, 1.0, 0.0]
- [1, 50, 0.3, 1.0, 1.0, 1.0, 0.0, 0.0, 1.0, 0.0]
- [1, 50, 0.3, 1.0, 1.0, 1.0, 0.0, 0.0, 1.0, 0.0]
- [1, 50, 0.3, 1.0, 1.0, 1.0, 0.0, 0.0, 1.0, 0.0]
- [1, 50, 0.3, 1.0, 1.0, 1.0, 0.0, 0.0, 1.0, 0.0]
target_size: 640
- !RandomInterpImage
target_size: 640
- !RandomFlipImage
is_normalized: true
- !Permute {}
- !NormalizeImage
is_scale: false
mean: [104, 117, 123]
std: [127.502231, 127.502231, 127.502231]
batch_size: 8
use_process: true
worker_num: 8
shuffle: true
EvalReader:
inputs_def:
fields: ['image', 'im_id']
dataset:
!WIDERFaceDataSet
dataset_dir: dataset/wider_face
anno_path: wider_face_split/wider_face_val_bbx_gt.txt
image_dir: WIDER_val/images
sample_transforms:
- !DecodeImage
to_rgb: true
- !NormalizeBox {}
- !ResizeImage
interp: 1
target_size: 640
use_cv2: false
- !Permute {}
- !NormalizeImage
is_scale: false
mean: [104, 117, 123]
std: [127.502231, 127.502231, 127.502231]
batch_size: 1
TestReader:
inputs_def:
image_shape: [3,640,640]
fields: ['image', 'im_id', 'im_shape']
dataset:
!ImageFolder
use_default_label: true
sample_transforms:
- !DecodeImage
to_rgb: true
- !ResizeImage
interp: 1
target_size: 640
use_cv2: false
- !Permute {}
- !NormalizeImage
is_scale: false
mean: [104, 117, 123]
std: [127.502231, 127.502231, 127.502231]
batch_size: 1
......@@ -36,6 +36,7 @@ FaceDetection的目标是提供高效、高速的人脸检测解决方案,包
| BlazeFace | 原始版本 | 640 | 8 | 32w | **0.915** | **0.892** | **0.797** | [模型](https://paddlemodels.bj.bcebos.com/object_detection/blazeface_original.tar) |
| BlazeFace | Lite版本 | 640 | 8 | 32w | 0.909 | 0.885 | 0.781 | [模型](https://paddlemodels.bj.bcebos.com/object_detection/blazeface_lite.tar) |
| BlazeFace | NAS版本 | 640 | 8 | 32w | 0.837 | 0.807 | 0.658 | [模型](https://paddlemodels.bj.bcebos.com/object_detection/blazeface_nas.tar) |
| BlazeFace | NAS_V2版本 | 640 | 8 | 32W | 0.870 | 0.837 | 0.685 | [模型](https://paddlemodels.bj.bcebos.com/object_detection/blazeface_nas2.tar)
| FaceBoxes | 原始版本 | 640 | 8 | 32w | 0.878 | 0.851 | 0.576 | [模型](https://paddlemodels.bj.bcebos.com/object_detection/faceboxes_original.tar) |
| FaceBoxes | Lite版本 | 640 | 8 | 32w | 0.901 | 0.875 | 0.760 | [模型](https://paddlemodels.bj.bcebos.com/object_detection/faceboxes_lite.tar) |
......@@ -254,6 +255,7 @@ cd dataset/fddb/evaluation
- 原始版本: 参考原始论文复现;
- Lite版本: 使用3x3卷积替换5x5卷积,更少的网络层数和通道数;
- NAS版本: 使用神经网络搜索算法构建网络结构,相比于`Lite`版本,NAS版本需要更少的网络层数和通道数。
- NAS_V2版本1: 基于PaddleSlim中SANAS算法在blazeface-NAS的基础上搜索出来的结构,相比`NAS`版本,NAS_V2版本的精度平均高出3个点,在855芯片上的硬件延时相对`NAS`版本仅增加5%。
### FaceBoxes
**简介:**
......
......@@ -40,6 +40,7 @@ optimized network structure.
| BlazeFace | Original | 640 | 8 | 32w | **0.915** | **0.892** | **0.797** | [model](https://paddlemodels.bj.bcebos.com/object_detection/blazeface_original.tar) |
| BlazeFace | Lite | 640 | 8 | 32w | 0.909 | 0.885 | 0.781 | [model](https://paddlemodels.bj.bcebos.com/object_detection/blazeface_lite.tar) |
| BlazeFace | NAS | 640 | 8 | 32w | 0.837 | 0.807 | 0.658 | [model](https://paddlemodels.bj.bcebos.com/object_detection/blazeface_nas.tar) |
| BlazeFace | NAS_V2 | 640 | 8 | 32W | 0.870 | 0.837 | 0.685 | [model](https://paddlemodels.bj.bcebos.com/object_detection/blazeface_nas2.tar)
| FaceBoxes | Original | 640 | 8 | 32w | 0.878 | 0.851 | 0.576 | [model](https://paddlemodels.bj.bcebos.com/object_detection/faceboxes_original.tar) |
| FaceBoxes | Lite | 640 | 8 | 32w | 0.901 | 0.875 | 0.760 | [model](https://paddlemodels.bj.bcebos.com/object_detection/faceboxes_lite.tar) |
......@@ -275,6 +276,7 @@ regression parameters of a bounding box as a weighted mean between the overlappi
- Lite: Replace 5x5 conv with 3x3 conv, fewer network layers and conv channels.
- NAS: use `Neural Architecture Search` algorithm to optimized network structure,
less network layer and conv channel number than `Lite`.
- NAS_V2: this version of model architecture searched based on blazeface-NAS by the SANAS in PaddleSlim, the average precision is 3% higher than blazeface-NAS, the latency is only 5% higher than blazeface-NAS on chip 855.
### FaceBoxes
**Introduction:**
......
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