search

jackyjsy / CVPR21Chal-SLR

This repo contains the official code of our work SAM-SLR which won the CVPR 2021 Challenge on Large Scale Signer Independent Isolated Sign Language Recognition.
Creative Commons Zero v1.0 Universal
205 stars 50 forks source link

the SL-GCN motion-related data prediction acc very low ? #18

Closed LiangSiyv closed 2 years ago

LiangSiyv commented 2 years ago

I download your code and barely not change anything. But the SL-GCN can't recognize the joint-motion and bone-motion very well. below are my result use the train_joint_motion.yaml and train_bone_motion.yaml: clipboard

below yaml file I just change the data_path and label_path, and not change other things! If possible, do you know why this result is?

Experiment_name: sign_joint_motion_final_NEU_data

# feeder
feeder: feeders.feeder.Feeder
train_feeder_args:
  data_path: ../../../Datas/TAUSL/AUTSL/skeleton/val_data_joint_motion.npy
  label_path: ../../../Datas/TAUSL/AUTSL/skeleton/val_gt.pkl
  debug: False
  random_choose: True
  random_shift: True
  window_size: 100
  random_mirror: True
  random_mirror_p: 0.5
  normalization: True
  is_vector: True

test_feeder_args:
  data_path: ./data/sign/27_2/train_data_joint_motion.npy
  label_path: ../../../Datas/TAUSL/AUTSL/skeleton/train_label.pkl
  random_mirror: False
  normalization: True
  is_vector: True

# model
model: model.decouple_gcn_attn.Model
model_args:
  num_class: 226
  num_point: 27
  num_person: 1
  graph: graph.sign_27.Graph
  groups: 16
  block_size: 41
  graph_args:
    labeling_mode: 'spatial'

#optim
weight_decay: 0.0001
base_lr: 0.1
step: [150, 200]

# training
device: [0,1]
keep_rate: 0.9
only_train_epoch: 1
batch_size: 64
test_batch_size: 64
num_epoch: 250
nesterov: True
warm_up_epoch: 20
LiangSiyv commented 2 years ago

I'm so sorry. After I use your provided traindata in SAMv2 link, the acc of joint_motino and bone_motion could reach to 90+%. I still don't know why, but sorry again to bother you!