笔尖检测阶段3-特征融合

2018-07-11

TensorFlow

修改mobilenet_v2的导入指令，原指定导入的代码位于目录<module ‘nets.mobilenet.mobilenet_v2’ from ‘/home/lq/anaconda3/lib/python3.6/site-packages/slim-0.1-py3.6.egg/nets/mobilenet/mobilenet_v2.py’>下

'from_layer': ['layer_15/expansion_output','layer_19', '', '', '', ''],
            'layer_depth': [-1,-1, 512, 256, 256, 128],

  layer_box_specs = [(0.1, 1.0), (scale, 2.0), (scale, 0.5)]

修改为

models/ssd_mobilenet_v2_feature_extractor.py中

1
2
3

'from_layer': ['layer_10/expansion_output','layer_15/expansion_output',
                        'layer_19'],
    'layer_depth': [-1, -1, -1],

anchor_generators/multiple_grid_anchor_generator.py中

1	layer_box_specs = [(0.05, 1.0)]

训练相关代码

1.1.1 用原图训练

训练

1	python train.py --logtostderr --pipeline_config_path=./training/3/ssd_mobilenet_v2_nib_3.config --train_dir=traininglog/3/ssd_mobilenet_v2_3 --fine_tune_batch_norm=True

config文件中设置参数num_steps: ，控制训练次数限制

tensorboard可视化－查看训练log

1	tensorboard --logdir='traininglog/3/ssd_mobilenet_v2_3'

评估模型

1	python eval.py --logtostderr --pipeline_config_path=./training/3/ssd_mobilenet_v2_nib_3.config --checkpoint_dir=traininglog/3/ssd_mobilenet_v2_3 --eval_dir=eval_log/ssd_mobilenet_v2/3

tensorboard可视化－查看评估log

1	tensorboard --logdir='eval_log/ssd_mobilenet_v2/2'

导出模型（.pb文件，即frozen model）

python export_inference_graph.py --input_type image_tensor --pipeline_config_path training/3/ssd_mobilenet_v2_nib_3.config --trained_checkpoint_prefix traininglog/3/ssd_mobilenet_v2_3/model.ckpt-18329 --output_directory traininglog/3/ssd_mobilenet_v2_3/output_inference_graph.pb