[深度学习][原创]mmaction2预测结果top5得分改为0-1

Posted 2022-01-04 FL1623863129

测试版本：mmaction0.20.0

测试环境:ubuntu18.04

当我们测试tsn模型：

python demo/demo.py configs/recognition/tsn/tsn_r50_video_inference_1x1x3_100e_kinetics400_rgb.py checkpoints/tsn_r50_1x1x3_100e_kinetics400_rgb_20200614-e508be42.pth demo/demo.mp4 tools/data/kinetics/label_map_k400.txt

结果显示：

load checkpoint from local path: checkpoints/tsn_r50_1x1x3_100e_kinetics400_rgb_20200614-e508be42.pth
The top-5 labels with corresponding scores are:
arm wrestling:  29.61644
rock scissors paper:  10.754842
shaking hands:  9.908401
clapping:  9.189913
massaging feet:  8.305306
显然上面得分>1不符合0-1概率，因此如何改为0-1概率显示呢。

方法如下：

经过测试发现tsn_r50_video_inference_1x1x3_100e_kinetics400_rgb.py加入test_cfg已经没有效果，正确做法就是：

去mmaction2-0.20.0/configs/_base_/models/tsn_r50.py将tesct_cfg改为

test_cfg = dict(average_clips='prob', test_crops=1))

最终文件内容结果：

# model settings
model = dict(
    type='Recognizer2D',
    backbone=dict(
        type='ResNet',
        pretrained='torchvision://resnet50',
        depth=50,
        norm_eval=False),
    cls_head=dict(
        type='TSNHead',
        num_classes=400,
        in_channels=2048,
        spatial_type='avg',
        consensus=dict(type='AvgConsensus', dim=1),
        dropout_ratio=0.4,
        init_std=0.01),
    # model training and testing settings
    train_cfg=None,
    test_cfg = dict(average_clips='prob', test_crops=1))