ubuntu16.04浮現WSHP模型
論文：Weakly and Semi Supervised Human Body Part Parsing via Pose-Guided Knowledge Transfer （CVPR2018）
了解鏈接：https://baijiahao.baidu.com/s?id=1601699961367739276&wfr=spider&for=pc
下面是整體結構：
1，先下載數據集：
https://github.com/MVIG-SJTU/WSHP/tree/master/parsing_network
如下圖的here,裏面有數據集和模型
2，安裝依賴項
pip install tensorflow
pip install -r requirements.txt3，應用
Inference
1,
python inference.py /home/feng/WSHP/parsing_network/dataset/ /home/feng/WSHP/parsing_network/model/model.ckpt-19315 --data_list /home/feng/WSHP/parsing_network/dataset/dance.txt
dance.txt
test2015/COCO_test2015_000000000014.jpg
test2015/COCO_test2015_000000000057.jpg
test2015/COCO_test2015_000000000063.jpg
test2015/COCO_test2015_000000000173.jpg
2,
python real-time-inference.py （我修改了路徑直接讀取final_model+攝像頭哈哈）
Evaluation
python evaluate.py --data-dir ./dataset/ --restore-from ./models/final_model/model.ckpt-19315
（dataset裏面包括pascal_test.txt,Annotations_Pascal_Part和JPEGImages_Pascal_Part文件夾）下面是我用數據集的圖像和我自己的圖片進行推理 的結果
測試結果：
訓練：
## Train our network on the whole dataset, model.ckpt-50000 is the pre-trained weights on COCO dataset
python train.py --data-dir ./dataset/ --data-list dataset/train_all.txt --num-epochs 10 --restore-from models/model.ckpt-50000 --not-restore-last --snapshot-dir snapshots-new-fromcoco --random-scale --random-mirror --save-pred-every 50000
## Finetune the model on the original dataset
python train.py --data-dir ./dataset/ --data-list dataset/pascal_train.txt --num-epochs 90 --restore-from snapshots-new-fromcoco/model.ckpt-213129 --snapshot-dir snapshots-new-fromcoco-finetune --random-scale --random-mirror --save-pred-every 10000訓練問題解決：
輸入訓練命令
python train.py --data-dir ./dataset/ --data-list dataset/train_all.txt --num-epochs 10 --restore-from models/model.ckpt-50000 --not-restore-last --snapshot-dir snapshots-new-fromcoco --random-scale --random-mirror --save-pred-every 50000
進行訓練出現如下問題：
ResourceExhaustedError (see above for traceback): OOM when allocating tensor with shape[2592,2048,5,5] and type float on /job:localhost/replica:0/task:0/device:GPU:0 by allocator GPU_0_bfc
[[Node: gradients/fc1_voc12_c2/convolution_grad/Conv2DBackpropInput = Conv2DBackpropInput[T=DT_FLOAT, data_format="NCHW", dilations=[1, 1, 1, 1], padding="VALID", strides=[1, 1, 1, 1], use_cudnn_on_gpu=true, _device="/job:localhost/replica:0/task:0/device:GPU:0"](gradients/fc1_voc12_c2/convolution_grad/Conv2DBackpropInput-0-VecPermuteNHWCToNCHW-LayoutOptimizer, fc1_voc12_c2/weights/read, gradients/fc1_voc12_c2/convolution_grad/Conv2DBackpropInput-2-TransposeNHWCToNCHW-LayoutOptimizer)]]
Hint: If you want to see a list of allocated tensors when OOM happens, add report_tensor_allocations_upon_oom to RunOptions for current allocation info.
解決方案：
將train.py文件裏面的BATCH_SIZE = 8改成BATCH_SIZE = 4
成功訓練如下：
4，問題解決：
安裝完依賴性進行圖例出現如下問題：
UnknownError (see above for traceback): Failed to get convolution algorithm. ...
原因是目前版本的TensorFlow的ObjectDection中，使用tensorflow-gpu ==1.12 版本会报错，如上，将版本降到tensorflow-gpu ==1.9.0即可正常运行，運行命令如下：
sudo pip install --upgrade --force-reinstall tensorflow-gpu==1.9.0 --user