Skip to content

Latest commit

 

History

History
40 lines (35 loc) · 3.2 KB

note.md

File metadata and controls

40 lines (35 loc) · 3.2 KB

10.06.2019

Problem:

  • VGG isn't good enough for my problem. Final feature is too small.

Solution:

  • 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.

10.07.2019

Idea

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

Result: No

10.09.2019

Idea

  1. Collect inferred part-action pairs and inspect for meaningful statistical patterns in part-action pair frequencies
  2. Reject hard (soft) negatives based on part-action pairs in inference/training, hopefully it improves instance-level performance
  3. 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

Result

pending

Idea

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

Idea

Suppress part-part edges between different humans

10.18.2019

Idea: Change lifting func from .sum to .mean

  • 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