百度论文复现_3DResNet-引用了Paddle的库
首先感谢百度能够发布论文复现营这种优质的课程,感谢百度提供的算力卡、感谢各位老师精彩的讲解还有班主任芮芮每天雷打不动的鼓励。
这篇程序是根据百度 PaddlePaddle 的论文复现营仿写、复现出来的。主要是学习使用 Paddle 的库。
参加论文复现营的初衷目的是跟随课程想学东西,至于最终的奖励自己现在是不敢想的,
小目标就是希望能够达到 领取结业证书 的条件,算是自己对这一个月努力的一个交代。
已经完成了三次实践作业,和最后跑通论文代码(程序能够跑通,但是还没有进行调优,没有使用预训练模型)。
现在自己已经完成训练和测试的过程,实现了论文代码的跑通,以下是10次迭代和65次迭代结果的展示:
Loss at epoch 9 step 70: [4.3616824], acc: [0.015625]
Loss at epoch 9 step 71: [4.164705], acc: [0.078125]
Loss at epoch 9 step 72: [4.233446], acc: [0.0703125]
Loss at epoch 9 step 73: [4.182964], acc: [0.0859375]
Final loss: [4.182964]
验证集准确率为:0.015331747010350227
Loss at epoch 64 step 71: [1.9090016], acc: [0.453125]
Loss at epoch 64 step 72: [1.7505848], acc: [0.5390625]
Loss at epoch 64 step 73: [1.8691069], acc: [0.53125]
Final loss: [1.8691069]
验证集准确率为:0.023261960595846176
!unzip -q -o /home/aistudio/data/data11460/UCF-101.zip -d data print("数据集解压完成")
!python avi2jpg.py
3. 根据 trainlist01.txt 和 testlist01.txt 划分 训练集和测试集 , 并生成对应的pkl文件(pkl文件很小,只是记录了名称)。pkl 文件中存储的每一条记录,记录的是图片的路径: 视频类名/视频名/图片名
!python jpg2pkl.py
!python data_list_gener.py
!python train.py --use_gpu True --epoch 65
!python test.py --weights 'checkpoints_models/resnet_model' --use_gpu True
eval.py 中的参数 args : Namespace(batch_size=1, config='configs/resnet.txt', filelist=None, infer_topk=1, log_interval=1, model_name='resnet', save_dir='checkpoints_models', use_gpu=True, weights='checkpoints_models/resnet_model')
[INFO: test.py: 133]: Namespace(batch_size=1, config='configs/resnet.txt', filelist=None, infer_topk=1, log_interval=1, model_name='resnet', save_dir='checkpoints_models', use_gpu=True, weights='checkpoints_models/resnet_model')
INFO: config.py: 55: ---------------- Valid Arguments ----------------
INFO: config.py: 59: image_mean:[0.485, 0.456, 0.406]
INFO: config.py: 59: image_std:[0.229, 0.224, 0.225]
W0905 19:17:46.177613 7921 device_context.cc:252] Please NOTE: device: 0, CUDA Capability: 70, Driver API Version: 10.1, Runtime API Version: 9.0
W0905 19:17:46.182003 7921 device_context.cc:260] device: 0, cuDNN Version: 7.3.
验证集准确率为:0.015331747010350227
eval.py 中的参数 args : Namespace(batch_size=1, config='configs/resnet.txt', filelist=None, infer_topk=1, log_interval=1, model_name='resnet',
save_dir='checkpoints_models', use_gpu=True, weights='checkpoints_models/resnet_model')
[INFO: test.py: 133]: Namespace(batch_size=1, config='configs/resnet.txt', filelist=None, infer_topk=1, log_interval=1, model_name='resnet',
save_dir='checkpoints_models', use_gpu=True, weights='checkpoints_models/resnet_model')
[INFO: config.py: 53]: ---------------- Valid Arguments ----------------
INFO: config.py: 57: image_mean:[0.485, 0.456, 0.406]
INFO: config.py: 57: image_std:[0.229, 0.224, 0.225]
W0906 05:12:40.361094 2283 device_context.cc:252] Please NOTE: device: 0, CUDA Capability: 70, Driver API Version: 10.1, Runtime API Version: 9.0
W0906 05:12:40.368430 2283 device_context.cc:260] device: 0, cuDNN Version: 7.3.
验证集准确率为:0.023261960595846176