- VGG isn't good enough for my problem. Final feature is too small.
- Switch back to ResNet-152, compute each segment separately. Maybe get Feng Shi on solving the repeated computation part. Should speed up quite a bit.
Train .sum and evaluate with .max, see if we get better mAP
- bear: CUDA_VISIBLE_DEVICES=0 python vcoco.py --batch-size 1 --prefetch 4 --epochs 100 --extra-feature --resume tmp/checkpoints/vcoco/exp0
- bear: CUDA_VISIBLE_DEVICES=1 python vcoco.py --batch-size 1 --prefetch 4 --epochs 100 --extra-feature --NRT --resume tmp/checkpoints/vcoco/exp1
- bear: CUDA_VISIBLE_DEVICES=2 python vcoco.py --batch-size 1 --prefetch 4 --epochs 100 --extra-feature --NRE --resume tmp/checkpoints/vcoco/exp2
- bear: CUDA_VISIBLE_DEVICES=3 python vcoco.py --batch-size 1 --prefetch 4 --epochs 100 --extra-feature --NRT --NRE --resume tmp/checkpoints/vcoco/exp3
- Collect inferred part-action pairs and inspect for meaningful statistical patterns in part-action pair frequencies
- Reject hard (soft) negatives based on part-action pairs in inference/training, hopefully it improves instance-level performance
- Similar to the tangram model: EM on compositional grammar## Experiment
- camel: CUDA_VISIBLE_DEVICES=0 python vcoco.py --batch-size 1 --prefetch 4 --epochs 100 --extra-feature --resume tmp/checkpoints/vcoco/exp0.pg --model-type PG
- camel: CUDA_VISIBLE_DEVICES=1 python vcoco.py --batch-size 1 --prefetch 4 --epochs 100 --extra-feature --NRT --resume tmp/checkpoints/vcoco/exp1.pg --model-type PG
- camel: CUDA_VISIBLE_DEVICES=2 python vcoco.py --batch-size 1 --prefetch 4 --epochs 100 --extra-feature --NRE --resume tmp/checkpoints/vcoco/exp2.pg --model-type PG
- camel: CUDA_VISIBLE_DEVICES=3 python vcoco.py --batch-size 1 --prefetch 4 --epochs 100 --extra-feature --NRT --NRE --resume tmp/checkpoints/vcoco/exp3.pg --model-type PG
pending
Adjusting prop-layer
parameter
- bear: CUDA_VISIBLE_DEVICES=0 python vcoco.py --batch-size 1 --prefetch 4 --epochs 100 --extra-feature --resume tmp/checkpoints/vcoco/exp4 --data-root /mnt/hdd-12t/share/v-coco/ --log-root ../../log/vcoco/exp4 --prop-layer 1
- bear: CUDA_VISIBLE_DEVICES=1 python vcoco.py --batch-size 1 --prefetch 4 --epochs 100 --extra-feature --resume tmp/checkpoints/vcoco/exp5 --data-root /mnt/hdd-12t/share/v-coco/ --log-root ../../log/vcoco/exp5 --prop-layer 2
- bear: CUDA_VISIBLE_DEVICES=2 python vcoco.py --batch-size 1 --prefetch 4 --epochs 100 --extra-feature --resume tmp/checkpoints/vcoco/exp6 --data-root /mnt/hdd-12t/share/v-coco/ --log-root ../../log/vcoco/exp6 --prop-layer 3
- bear: CUDA_VISIBLE_DEVICES=3 python vcoco.py --batch-size 1 --prefetch 4 --epochs 100 --extra-feature --resume tmp/checkpoints/vcoco/exp7 --data-root /mnt/hdd-12t/share/v-coco/ --log-root ../../log/vcoco/exp7 --prop-layer 4
Suppress part-part edges between different humans
- bear: CUDA_VISIBLE_DEVICES=1 python vcoco.py --batch-size 1 --prefetch 4 --epochs 100 --extra-feature --resume tmp/checkpoints/vcoco/exp8 --data-root /mnt/hdd-12t/share/v-coco/ --log-root ../../log/vcoco/exp8 --prop-layer 2