Skip to content

TUNING_LOG.md

Latest commit

聽

History

History
311 lines (272 loc) 路 25.5 KB

TUNING_LOG.md

File metadata and controls

311 lines (272 loc) 路 25.5 KB
Raw

Complete Tuning Log

Penn Action

3DA (best) with Pr-VIPE, UNIK, HDM-BG, 3D Deep, PoseMap, MultitaskCNN, STAR: alt text

The complete experiment tuning logs:

masked pretraining decoder d_model n_head num_layers freeze T1? T1-lr #epochs T2-lr (ft-lr) #epochs clf-acc
Complete Experiments. 5% held-out validation (n = 2326)
no linear 256 8 4 yes 1e-4 1000 1e-5, wd=1e-4 500 85.11%
no linear 256 8 4 no 1e-4 1000 1e-5, wd=1e-4 200 89.70%
no linear 256 8 4 no 1e-4 1000 1e-5, wd=1e-4 400 91.10%
no linear 256 8 4 no 1e-4 1000 1e-5, wd=1e-4 500 91.01%
no linear 256 8 4 no 1e-4 1000 1e-5, wd=1e-4 700 89.42%
no linear 256 8 4 finetune layer #4 1e-4 1000 1e-5, wd=1e-4 400 86.89%
no linear 256 8 4 finetune layer #4 1e-4 1000 1e-5, wd=1e-4 490 88.39%
no MLP 256 8 4 no 1e-5 1000 1e-5, wd=1e-4 600 89.51%
no MLP 256 8 4 no 1e-5 1000 1e-5, wd=1e-4 700 90.45%
no MLP 256 8 4 no 1e-5 1000 1e-5, wd=1e-4 1000 89.89%
30% masked pretraining - random frames
30% linear 256 8 4 no 1e-4 500 1e-5, wd=1e-4 500 87.17%
30% linear 256 8 4 no 1e-4 500 1e-5, wd=1e-4 700 87.27%
30% linear 256 8 4 no 1e-4 500 1e-5, wd=1e-4 1000 89.23%
30% linear 256 8 4 no 1e-4 500 1e-5, wd=1e-4 1500 87.73%
30% MLP 256 8 4 no 1e-4 500 1e-5, wd=1e-4 500 86.89%
30% MLP 256 8 4 no 1e-4 500 1e-5, wd=1e-4 1000 87.45%
30% linear 256 8 4 no 1e-4 1K 1e-5, wd=1e-4 500 89.70%
30% linear 256 8 4 no 1e-4 1K 1e-5, wd=1e-4 1000 89.98%
30% linear 256 8 4 no 1e-4 2K 1e-5, wd=1e-4 1000 87.36%
cosine scheduler didn't improve the performance...
30% linear 256 8 4 no 1e-4, batch-cosine 300 1e-5, wd=1e-4 500 86.99%
30% linear 256 8 4 no 1e-4, batch-cosine 300 1e-5, wd=1e-4 1000 88.20%
30% linear 256 8 4 no 1e-4, batch-cosine 300 1e-5, wd=1e-4 1200 87.92%
30% linear 256 8 4 no 1e-4, batch-cosine 300 1e-5, wd=1e-4, epoch-cosine 800 86.52%
30% linear 256 8 4 no 1e-4, batch-cosine 300 1e-5, wd=1e-4, epoch-cosine 1000 86.80%
30% linear 256 8 4 no 1e-4, batch-cosine 300 1e-5, wd=1e-4, epoch-cosine 1500 86.80%
40% masked pretraining - random frames
40% linear 256 8 4 no 1e-4 500 1e-5, wd=1e-4 400 87.55%
40% linear 256 8 4 no 1e-4 500 1e-5, wd=1e-4 500 88.20%
40% linear 256 8 4 no 1e-4 500 1e-5, wd=1e-4 600 87.36%
40% linear 256 8 4 no 1e-4 500 1e-5, wd=1e-4 700 87.83%
20% masked pretraining - random frames
40% linear 256 8 4 no 1e-4 500 1e-5, wd=1e-4 400 87.73%
40% linear 256 8 4 no 1e-4 500 1e-5, wd=1e-4 500 88.67%
40% linear 256 8 4 no 1e-4 500 1e-5, wd=1e-4 700 89.04%
40% linear 256 8 4 no 1e-4 500 1e-5, wd=1e-4 1000 88.30%
40% linear 256 8 4 no 1e-4 500 1e-5, wd=1e-4 1000 88.39%
30% masked pretraining - random global joints
30% linear 256 8 4 no 1e-4 200 1e-5, wd=1e-4 200 91.48%
30% linear 256 8 4 no 1e-4 200 1e-5, wd=1e-4 500 91.01%
30% linear 256 8 4 no 1e-4 300 1e-5, wd=1e-4 200 91.67%
30% linear 256 8 4 no 1e-4 300 1e-5, wd=1e-4 300 92.42%
30% linear 256 8 4 no 1e-4 300 1e-5, wd=1e-4 350 91.76%
30% linear 256 8 4 no 1e-4 300 1e-5, wd=1e-4 400 91.95%
30% linear 256 8 4 no 1e-4 300 1e-5, wd=1e-4 500 91.39%
30% linear 256 8 4 no 1e-4 500 1e-5, wd=1e-4 200 91.48%
30% linear 256 8 4 no 1e-4 500 1e-5, wd=1e-4 300 91.57%
30% linear 256 8 4 no 1e-4 500 1e-5, wd=1e-4 500 91.57%
30% linear 256 8 4 no 1e-4 600 1e-5, wd=1e-4 200 91.20%
30% linear 256 8 4 no 1e-4 600 1e-5, wd=1e-4 300 92.88%
30% linear 256 8 4 no 1e-4 600 1e-5, wd=1e-4 400 92.42%
30% linear 256 8 4 no 1e-4 600 1e-5, wd=1e-4 500 91.57%
30% linear 256 8 4 no 1e-4 600 1e-5, wd=1e-4 600 92.13%
30% linear 256 8 4 no 1e-4 600 1e-5, wd=1e-4 900 92.42%
30% linear 256 8 4 no 1e-4 600 1e-5, wd=1e-4 1000 91.95%
30% linear 256 8 4 no 1e-4 1000 1e-5, wd=1e-4 1000 94.66%
ablation study: too many layers can cause overfitting...
30% linear 512 8 8 no 1e-5 300 1e-5, wd=1e-4 100 89.89%
30% linear 512 8 8 no 1e-5 300 1e-5, wd=1e-4 300 86.52%
ablation study: subtraction-based bones
30%, subtraction-bone linear 256 8 4 no 1e-4 200 1e-5, wd=1e-4 200 91.76%
30%, subtraction-bone linear 256 8 4 no 1e-4 1000 1e-5, wd=1e-4 300 91.67%
30%, subtraction-bone linear 256 8 4 no 1e-4 1000 1e-5, wd=1e-4 500 92.32%
30%, subtraction-bone linear 256 8 4 no 1e-4 1000 1e-5, wd=1e-4 700 92.32%
30%, subtraction-bone linear 256 8 4 no 1e-4 1000 1e-5, wd=1e-4 1000 90.92%
ablation study: concatenation-based bones
30%, subtraction-bone linear 256 8 4 no 1e-4 200 1e-5, wd=1e-4 200 90.26%
30%, subtraction-bone linear 256 8 4 no 1e-4 1000 1e-5, wd=1e-4 500 92.51%
30%, subtraction-bone linear 256 8 4 no 1e-4 1000 1e-5, wd=1e-4 1000 93.16%
ablation study: parameterization-based bones
30%, subtraction-bone linear 256 8 4 no 1e-4 1000 1e-5, wd=1e-4 500 93.91%
30%, subtraction-bone linear 256 8 4 no 1e-4 1000 1e-5, wd=1e-4 1000 93.45%
CascadeFormer 1.1 (convolution enhanced)
30% linear 256 8 4 no 1e-4 100 1e-5, wd=1e-4 100 89.70%
30% linear 256 8 4 no 1e-4 200 1e-5, wd=1e-4 200 93.35%
30% linear 256 8 4 no 1e-4 300 1e-5, wd=1e-4 100 92.51%
30% linear 256 8 4 no 1e-4 300 1e-5, wd=1e-4 200 92.98%
30% linear 256 8 4 no 1e-4 300 1e-5, wd=1e-4 300 92.79%
30% linear 256 8 4 no 1e-4 500 1e-5, wd=1e-4 500 94.10%
CascadeFormer 1.2 (spatial transformer)
30% linear 256 8 4 no 1e-4 200 1e-5, wd=1e-4 100 94.10%
30% linear 256 8 4 no 1e-4 200 1e-5, wd=1e-4 120 94.10%
30% linear 256 8 4 no 1e-4 200 1e-5, wd=1e-4 200 94.01%
masked pretraining decoder d_model n_head num_layers freeze T1? T1-lr #epochs T2-lr (ft-lr) #epochs clf-acc
CascadeFormer 2.0 (interleaving spatial-temporal attention)
30% linear 832 8 2 no 1e-4 50 1e-4, wd=1e-2 50 83.15% - overfitting!
30% linear 832 8 2 no 1e-4 100 1e-4, wd=1e-2 70 91.29%
30% linear 832 8 2 no 1e-4 100 1e-4, wd=1e-2 80 92.32%
30% linear 832 8 2 no 1e-4 100 1e-4, wd=1e-2 90 90.73%
30% linear 832 8 2 no 1e-4 100 1e-4, wd=1e-2 100 91.67%
30% linear 832 8 2 no 1e-4 200 1e-4, wd=1e-2 100 86.05% - overfitting!
30% linear 832 8 2 no 1e-4 200 1e-4, wd=1e-2 200 86.05% - overfitting!

NTU RGB+D 60

batch-level padding OR fixed-length sequence

alt text

Many folks have done data augmentation (still using only skeleton data):

  1. E1: joint modality only
  2. E2: joint + bone modalities
  3. E4: joint + bone + joint motion + bone motion modalities

People usually:

train separate networks for each modality and ensemble their outputs

alt text

The current best training setup (95%-5% train-val split):

cross-subject evaluation

The complete experiment tuning logs

cross-subject evaluation:

masked pretraining decoder d_model n_head num_layers freeze T1? T1-lr #epochs T2-lr (ft-lr) #epochs accuracy
regular pretraining
no linear 256 8 4 no 1e-4 100 1e-5, wd=1e-4 50 70.46%
no linear 256 8 4 no 1e-4 300 1e-5, wd=1e-4 50 71.33%
no linear 256 8 4 no 1e-4 300 1e-5, wd=1e-4 100 71.91%
30% masked pretraining - random frames
30% linear 256 8 4 no 1e-4 100 1e-5, wd=1e-4 100 70.45%
30% linear 256 8 4 no 1e-4 100 1e-5, wd=1e-4 300 71.67%
30% masked pretraining - random global joints
30% linear 256 8 4 no 1e-4 100 1e-5, wd=1e-4 100 71.27%
30% linear 256 8 4 no 1e-4 300 1e-5, wd=1e-4 300 72.65%
30% linear 256 8 4 no 1e-4 300 1e-5, wd=1e-4 500 72.31%
30% linear 256 8 4 no 1e-4 500 1e-5, wd=1e-4 100 72.32%
30% linear 256 8 4 no 1e-4 500 3e-5, wd=1e-4, cosine + warmup 100 72.33%
30% linear 256 8 4 no 1e-4 500 3e-5, wd=1e-4, cosine + warmup 300 73.21%
30% linear 256 8 4 no 1e-4 500 3e-5, wd=1e-4, cosine + warmup 500 72.94%
strong backbone
30% linear 256 8 4 no 1e-4 1000 3e-5, wd=1e-4, cosine + warmup 300 73.56%
30% linear 256 8 4 no 1e-4 1000 3e-5, wd=1e-4, cosine + warmup 500 73.17%
30% linear 256 8 4 no 1e-4 1000 3e-5, wd=1e-4, cosine + warmup 1000 72.51%
larger model
30% linear 512 8 8 no 1e-4 200 1e-5, wd=1e-4 100 74.79%
30% linear 512 8 8 no 1e-4 200 1e-5, wd=1e-4 200 74.75%
30% linear 512 8 12 no 1e-4 100 1e-5, wd=1e-4 100 75.12%
30% linear 512 8 12 no 1e-4 200 1e-5, wd=1e-4 200 75.22%
ablation study: subtraction-based bones
30% linear 512 8 8 no 1e-4 100 1e-5, wd=1e-4 100 74.23%
ablation study: concatenation-based bones
30% linear 512 8 8 no 1e-4 100 1e-5, wd=1e-4 100 73.81%
CascadeFormer 1.1 with data augmentation
30% linear 512 8 12 no 1e-4 100 1e-5, wd=1e-4 100 70.68%
CascadeFormer 1.1 with trimmed-uniform random sampling
30%, 64+1.0 linear 512 8 12 no 1e-4 50 3e-5, wd=1e-2, CosineAnnealingWarmRestarts, epoch-level 50 72.22%
30%, 64+1.0 linear 512 8 12 no 1e-4 50 3e-5, wd=1e-2, CosineAnnealingWarmRestarts, batch-level 10 71.84%
30%, 64+1.0 linear 512 8 12 no 1e-4 100 3e-5, wd=1e-2, CosineAnnealingWarmRestarts, batch-level 10 71.18%
30%, 64+1.0 linear 512 8 12 no 1e-4 100 3e-5, wd=1e-2, CosineAnnealingWarmRestarts, batch-level 20 70.33%
30%, 64+1.0 linear 512 8 12 no 1e-4 100 3e-5, wd=1e-2, CosineAnnealingLR, epoch-level 20 72.08%
30%, 64+1.0 linear 512 8 12 no 1e-4 100 3e-5, wd=1e-2, CosineAnnealingLR, epoch-level 30 73.01%
30%, 64+1.0 linear 512 8 12 no 1e-4 100 3e-5, wd=1e-2, CosineAnnealingLR, epoch-level 50 73.28%
30%, 64+1.0 linear 512 8 12 no 1e-4 100 3e-5, wd=1e-2, CosineAnnealingLR, epoch-level 100 73.97%
30%, 64+1.0 linear 512 8 12 no 1e-4 200 3e-5, wd=1e-2, CosineAnnealingLR, epoch-level 20 71.60%
30%, 64+1.0 linear 512 8 12 no 1e-4 200 3e-5, wd=1e-2, CosineAnnealingLR, epoch-level 50 72.67%
30%, 64+1.0 linear 512 8 12 no 1e-4 200 3e-5, wd=1e-2, CosineAnnealingLR, epoch-level 100 73.22%
30%, 64+1.0 linear 512 8 8 no 1e-4 100 3e-5, wd=1e-2, CosineAnnealingLR, epoch-level 50 72.78%
30%, 64+1.0 linear 512 8 8 no 1e-4 100 3e-5, wd=1e-2, CosineAnnealingLR, epoch-level 100 73.07%
a larger hidden size/latent space?!
30%, 64+1.0 linear 768 8 8 no 1e-4 300 3e-5, wd=1e-2, CosineAnnealingLR, epoch-level 100 73.48%
30%, 64+1.0 linear 768 8 8 no 1e-4 300 3e-5, wd=1e-2, CosineAnnealingLR, epoch-level 300 74.10%
add dropout: 0.1 for T1, 0.1 for T2, 0.5 for head
30%, 64+1.0 linear 768 8 8 no 1e-4 100 3e-5, wd=1e-2, CosineAnnealingLR, epoch-level 100 73.85%
hidden size 1024
30%, 64+1.0 linear 1024 8 8 no 1e-4 100 3e-5, wd=1e-2, CosineAnnealingLR, epoch-level 100 73.61%
100% augmentation
30%, 64+1.0 linear 1024 8 8 no 1e-4 100 3e-5, wd=1e-2, CosineAnnealingLR, epoch-level 100 73.33%
CascadeFormer 1.2
30%, 64+1.0 linear 512 8 8 no 1e-4 50 3e-5, wd=1e-2, CosineAnnealingLR, epoch-level 50 71.79%
30%, 64+1.0 linear 512 8 8 freeze-25-unfreeze 1e-4 100 2.6e-4, wd=5e-3, CosineAnnealingLR, batch-level 100 72.10%
30%, 64+1.0 linear 400, revised 8 8 no 1e-4 100 2.6e-4, wd=5e-3, CosineAnnealingLR, batch-level 100 73.56%
CascadeFormer 1.3 (GCN)
30%, 64+1.0 linear 400, revised 8 8 no 1e-4 100 2.6e-4, wd=5e-3, CosineAnnealingLR, batch-level 100 68.98%
CascadeFormer 1.1 but biomechanics-aware frame embeddings
30%, 64+1.0, bioaware linear 256 8 8 no 1e-4 100 3e-5, wd=1e-2, CosineAnnealingLR, batch-level 100 73.36%
30%, 64+1.0, bioaware linear 768 16 16 no 1e-4, CosineAnnealingLR 400 3e-5, wd=1e-2, CosineAnnealingLR, batch-level 100 TBD
30%, 64+1.0, bioaware linear 768 16 16 no 1e-4, CosineAnnealingLR 400 3e-5, wd=1e-2, CosineAnnealingLR, batch-level 200 TBD
30%, 64+1.0, bioaware linear 768 16 16 no 1e-4, CosineAnnealingLR 400 3e-5, wd=1e-2, CosineAnnealingLR, batch-level 400 running - CUDA_VISIBLE_DEVICES=1 nohup taskset -c 21-30 python3 baseline/action_recognition/cascadeformer_1_0/joint/ntu_60_own/NTU_main.py --pretrain --num_epochs 400 > large_bioaware_1_0.txt 2>&1 &
CascadeFormer 2.0
30%, 64+1.0 linear 800 8 2 no 1e-4 100 3e-5, wd=1e-2, CosineAnnealingLR, epoch-level 100 54.48%
30%, 64+1.0 linear 800 8 2 no 1e-4 100 3e-5, wd=1e-2, CosineAnnealingLR, epoch-level 100 58.29%, 20% validation
30%, 64+1.0 linear 800 8 4 no 1e-4 25 3e-5, wd=1e-2, CosineAnnealingLR, epoch-level 25 58.29%, 20% validation

Baseline - Experiment (NW-UCLA, cross-view)

alt text

The current best training setup (95%-5% train-val split):

masked pretraining decoder d_model n_head num_layers freeze T1? T1-lr #epochs T2-lr (ft-lr) #epochs accuracy
30%, SF data loader linear 256 8 4 no 1e-4 1000 3e-5, wd=1e-4, cosine + warmup 300 88.79%

The complete experiment tuning logs:

Use SkateFormer data loader instead of my own data loader because:

  1. normalization: first-joint centering; min-max normalization
  2. regularization: random rotation, random scaling, random sampling, random dropout (joint/axis)
  3. data augmentation: dataset duplication (repeat N=10 times)
masked pretraining decoder d_model n_head num_layers freeze T1? T1-lr #epochs T2-lr (ft-lr) #epochs accuracy
small backbone
30%, SF data loader linear 256 8 4 no 1e-4 100 3e-5, wd=1e-4, cosine + warmup 20 84.48%
30%, SF data loader linear 256 8 4 no 1e-4 100 3e-5, wd=1e-4, cosine + warmup 50 86.64%
30%, SF data loader linear 256 8 4 no 1e-4 100 3e-5, wd=1e-4, cosine + warmup 100 85.13%
30%, SF data loader linear 256 8 4 no 1e-4 100 3e-5, wd=1e-4, cosine + warmup 200 84.27%
30%, SF data loader linear 256 8 4 no 1e-4 200 3e-5, wd=1e-4, cosine + warmup 100 85.56%
30%, SF data loader linear 256 8 4 no 1e-4 200 3e-5, wd=1e-4, cosine + warmup 200 87.28%
30%, SF data loader linear 256 8 4 no 1e-4 200 3e-5, wd=1e-4, cosine + warmup 300 85.13%
medium backbone
30%, SF data loader linear 256 8 4 no 1e-4 500 3e-5, wd=1e-4, cosine + warmup 100 84.05%
30%, SF data loader linear 256 8 4 no 1e-4 500 3e-5, wd=1e-4, cosine + warmup 200 84.05%
30%, SF data loader linear 256 8 4 no 1e-4 500 3e-5, wd=1e-4, cosine + warmup 500 87.93%
30%, SF data loader linear 256 8 4 no 1e-4 500 3e-5, wd=1e-4, cosine + warmup 1000 85.99%
strong backbone
30%, SF data loader linear 256 8 4 no 1e-4 1000 3e-5, wd=1e-4, cosine + warmup 200 86.21%
30%, SF data loader linear 256 8 4 no 1e-4 1000 3e-5, wd=1e-4, cosine + warmup 300 88.79%
30%, SF data loader linear 256 8 4 no 1e-4 1000 3e-5, wd=1e-4, cosine + warmup 500 88.15%
30%, SF data loader linear 256 8 4 no 1e-4 1000 3e-5, wd=1e-4, cosine + warmup 1000 87.72%
ablation study: subtraction-based bones
30%, SF data loader linear 256 8 4 no 1e-4 100 3e-5, wd=1e-4, cosine + warmup 100 85.56%
30%, SF data loader linear 256 8 4 no 1e-4 300 3e-5, wd=1e-4, cosine + warmup 300 85.13%
30%, SF data loader linear 256 8 4 no 1e-4 1000 3e-5, wd=1e-4, cosine + warmup 100 83.19%
30%, SF data loader linear 256 8 4 no 1e-4 1000 3e-5, wd=1e-4, cosine + warmup 300 82.54%
30%, SF data loader linear 256 8 4 no 1e-4 1000 3e-5, wd=1e-4, cosine + warmup 1000 79.31%
ablation study: concatenation-based bones
30%, SF data loader linear 256 8 4 no 1e-4 100 3e-5, wd=1e-4, cosine + warmup 100 84.27%
30%, SF data loader linear 256 8 4 no 1e-4 200 3e-5, wd=1e-4, cosine + warmup 200 84.70%
30%, SF data loader linear 256 8 4 no 1e-4 300 3e-5, wd=1e-4, cosine + warmup 300 88.15%
masked pretraining decoder d_model n_head num_layers freeze T1? T1-lr #epochs T2-lr (ft-lr) #epochs accuracy
CascadeFormer 1.1 (convolution enhanced)
p: probability of dropping axis/joint regularization
30%, p=0.5, repeat=10 linear 256 8 4 no 1e-4 3 3e-5, wd=1e-4, cosine + warmup 3 82.54%
30%, p=0.5, repeat=10 linear 256 8 4 no 1e-4 200 3e-5, wd=1e-4, cosine + warmup 100 85.99%
30%, p=0.5, repeat=10 linear 256 8 4 no 1e-4 200 3e-5, wd=1e-4, cosine + warmup 200 83.84%
30%, p=0.5, repeat=10 linear 256 8 4 no 1e-4 300 3e-5, wd=1e-4, cosine + warmup 100 84.70%
30%, p=0.5, repeat=10 linear 256 8 4 no 1e-4 300 3e-5, wd=1e-4, cosine + warmup 300 87.07%
30%, p=0.5, repeat=10 linear 256 8 4 no 1e-4 300 3e-5, wd=1e-4, cosine + warmup 400 86.85%
try LESS regularization!
30%, p=0.1, repeat=10 linear 256 8 4 no 1e-4 100 3e-5, wd=1e-4, cosine + warmup 80 84.27%
30%, p=0.1, repeat=10 linear 256 8 4 no 1e-4 100 3e-5, wd=1e-4, cosine + warmup 90 87.07%
30%, p=0.1, repeat=10 linear 256 8 4 no 1e-4 100 3e-5, wd=1e-4, cosine + warmup 95 86.85%
30%, p=0.1, repeat=10 linear 256 8 4 no 1e-4 100 3e-5, wd=1e-4, cosine + warmup 99 86.42%
30%, p=0.1, repeat=10 linear 256 8 4 no 1e-4 100 3e-5, wd=1e-4, cosine + warmup 100 91.16%
30%, p=0.1, repeat=10 linear 256 8 4 no 1e-4 100 1e-5, wd=1e-4, cosine + warmup 100 87.72%
30%, p=0.1, repeat=10, 64 batch size linear 256 8 4 no 1e-4 100 3e-5, wd=1e-4, cosine + warmup 100 84.70%
30%, p=0.0, repeat=10, 64 batch size linear 256 8 4 no 1e-4 100 3e-5, wd=1e-4, cosine + warmup 100 86.64%
30%, p=0.1, repeat=10 linear 256 8 4 no 1e-4 100 3e-5, wd=1e-4, cosine + warmup, but p=0.0 during finetuning 100 85.56%
30%, p=0.1, repeat=10 linear 256 8 4 no 1e-4 100 3e-5, wd=1e-4, cosine + warmup, but repeat=15 during finetuning 100 87.07%
30%, p=0.1, repeat=10 linear 256 8 4 no 1e-4 100 3e-5, wd=1e-4, cosine + warmup 110 86.64%
30%, p=0.1, repeat=10 linear 256 8 4 no 1e-4 200 3e-5, wd=1e-4, cosine + warmup 50 86.21%
30%, p=0.1, repeat=10 linear 256 8 4 no 1e-4 200 3e-5, wd=1e-4, cosine + warmup 100 87.50%
30%, p=0.1, repeat=10 linear 256 8 4 no 1e-4 200 3e-5, wd=1e-4, cosine + warmup 200 85.34%
try MORE repeat
30%, p=0.1, repeat=15 linear 256 8 4 no 1e-4 100 3e-5, wd=1e-4, cosine + warmup 50 85.13%
30%, p=0.1, repeat=15 linear 256 8 4 no 1e-4 100 3e-5, wd=1e-4, cosine + warmup 100 85.99%

save the best checkpoint

This is my current backbone: | 30%, p=0.1, repeat=10 | linear | 256 | 8 | 4 | no | 1e-4 |

running 1207438!
CUDA_VISIBLE_DEVICES=0 taskset -c 21-30 nohup python3 baseline/action_recognition/cascadeformer_1_0/joint/ntu_60_own/NTU_main.py --num_epochs 150 > 150_epoch.txt 2>&1 &

Ablation Study: bone representation (Penn Action and NTU/CS)

dataset #videos #actions dimension #joints outperform SoTA?
Penn Action, subtraction-bone 2,326 15 2D 13 92.32% ~ 93.4% (HDM-BG)
Penn Action, concatenation-bone 2,326 15 2D 13 93.16% ~ 93.4% (HDM-BG)
Penn Action, parameterization-bone 2,326 15 2D 13 93.91% > 93.4% (HDM-BG)
N-UCLA, subtraction-bone 1,494 12 3D 20 85.56% < 98.3% (SkateFormer)
N-UCLA, concatenation-bone 1,494 12 3D 20 88.15% < 98.3% (SkateFormer)
NTU/CS, subtraction-bone 56,880 60 3D 25 74.23% << 92.6% (SkateFormer) - cross subject
NTU/CS, concatenation-bone 56,880 60 3D 25 73.81% << 92.6% (SkateFormer) - cross subject