未验证 提交 d98fb5d4 编写于 作者: S sunxl1988 提交者: GitHub

update HTC model (#1276)

* test=master update htc

* test=master update htc cfg
上级 3bc034aa
......@@ -23,4 +23,4 @@ The results on COCO 2017val are shown in the below table. (results on test-dev a
| Backbone | Lr schd | Inf time (fps) | box AP | mask AP | Download |
|:---------:|:-------:|:--------------:|:------:|:-------:|:--------:|
| R-50-FPN | 1x | 11 | 42.2 | 36.5 | [model](https://paddlemodels.bj.bcebos.com/object_detection/htc_r50_fpn_1x.pdparams ) |
| R-50-FPN | 1x | 11 | 42.7 | 36.8 | [model](https://paddlemodels.bj.bcebos.com/object_detection/htc_r50_fpn_1x.pdparams ) |
architecture: HybridTaskCascade
use_gpu: true
max_iters: 100000
max_iters: 180000
snapshot_iter: 10000
log_smooth_window: 50
save_dir: output
......@@ -110,12 +110,13 @@ LearningRate:
schedulers:
- !PiecewiseDecay
gamma: 0.1
milestones: [60000, 80000]
milestones: [120000, 160000]
- !LinearWarmup
start_factor: 0.1
steps: 1000
OptimizerBuilder:
clip_grad_by_norm: 35.0
optimizer:
momentum: 0.9
type: Momentum
......@@ -124,8 +125,9 @@ OptimizerBuilder:
type: L2
TrainReader:
batch_size: 2
batch_size: 1
worker_num: 2
shuffle: true
dataset:
!COCODataSet
dataset_dir: dataset/coco
......
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