VIBE
VIBE copied to clipboard
About Training Accuracy Question
- operating system and the version: ubuntu18
- python version: 3.6
- pytorch version: 1.6
Did anyone else have the same problem? The accuracy is decreasing, and the error is increasing.
Here is my train config:
GPU name -> GeForce RTX 2080 Ti GPU feat -> _CudaDeviceProperties(name='GeForce RTX 2080 Ti', major=7, minor=5, total_memory=11019MB, multi_processor_count=68) {'CUDNN': CfgNode({'BENCHMARK': True, 'DETERMINISTIC': False, 'ENABLED': True}), 'DATASET': CfgNode({'SEQLEN': 16, 'OVERLAP': 0.5}), 'DEBUG': False, 'DEBUG_FREQ': 5, 'DEVICE': 'cuda', 'EXP_NAME': 'vibe', 'LOGDIR': 'results/vibe_tests/02-12-2020_16-03-31_vibe', 'LOSS': {'D_MOTION_LOSS_W': 0.5, 'KP_2D_W': 300.0, 'KP_3D_W': 300.0, 'POSE_W': 60.0, 'SHAPE_W': 0.06}, 'MODEL': {'TEMPORAL_TYPE': 'gru', 'TGRU': {'ADD_LINEAR': True, 'BIDIRECTIONAL': False, 'HIDDEN_SIZE': 1024, 'NUM_LAYERS': 2, 'RESIDUAL': True}}, 'NUM_WORKERS': 8, 'OUTPUT_DIR': 'results/vibe_tests', 'SEED_VALUE': -1, 'TRAIN': {'BATCH_SIZE': 32, 'DATASETS_2D': ['Insta'], 'DATASETS_3D': ['MPII3D'], 'DATASET_EVAL': 'ThreeDPW', 'DATA_2D_RATIO': 0.6, 'END_EPOCH': 30, 'GEN_LR': 5e-05, 'GEN_MOMENTUM': 0.9, 'GEN_OPTIM': 'Adam', 'GEN_WD': 0.0, 'LR_PATIENCE': 5, 'MOT_DISCR': {'ATT': {'DROPOUT': 0.2, 'LAYERS': 3, 'SIZE': 1024}, 'FEATURE_POOL': 'attention', 'HIDDEN_SIZE': 1024, 'LR': 0.0001, 'MOMENTUM': 0.9, 'NUM_LAYERS': 2, 'OPTIM': 'Adam', 'UPDATE_STEPS': 1, 'WD': 0.0001}, 'NUM_ITERS_PER_EPOCH': 500, 'PRETRAINED': '', 'PRETRAINED_REGRESSOR': 'data/vibe_data/spin_model_checkpoint.pth.tar', 'RESUME': '', 'START_EPOCH': 0}} InstaVariety number of dataset objects 130431 MPII3D Dataset overlap ratio: 0 Loaded mpii3d dataset from data/vibe_db/mpii3d_train_db.pt mpii3d - number of dataset objects 59934 AMASS dataset number of videos: 1012067 3DPW Dataset overlap ratio: 0.0 Loaded 3dpw dataset from data/vibe_db/3dpw_val_db.pt 3dpw - number of dataset objects 635 => loaded pretrained model from 'data/vibe_data/spin_model_checkpoint.pth.tar' is not a pretrained model!!!! => no checkpoint found at ''
@jiandandian2 Did you ever end up resolving this? I see the exact same behavior.
@jkamalu solved by adding the 3dpw training dataset https://github.com/mkocabas/VIBE/issues/140
How did you add the 3dpw? @jiandandian2
@jkamalu solved by adding the 3dpw training dataset #140
I think the author dont want us use 3pdw for training. #99