Skip to content

Latest commit

 

History

History
133 lines (100 loc) · 4.09 KB

Instructions_FashionIQ.md

File metadata and controls

133 lines (100 loc) · 4.09 KB

FashionIQ

Details on each step are as follows.

Note

To reproduce our reported results, please first download the checkpoints and top-k files and save them to paths noted in DOWNLOAD.md.

We provide all intermediate checkpoints and top-k files, so you could skip training steps and (for instance) develop directly on our Stage 2 model.

Stage 1

Train

# Optional: comet experiment logging --api-key and --workspace
python src/stage1_train.py --dataset FashionIQ \
    --api-key <COMET_API_KEY> --workspace <COMET_WORKSPACE> \
    --blip-learning-rate 2e-5 --blip-max-epoch 20 --num-epochs 20 \
    --batch-size 512 --blip-bs 16 \
    --transform targetpad --target-ratio 1.25 \
    --save-training --save-best --validation-frequency 1 \
    --experiment-name BLIP_stageI_b512_2e-5_cos20 \
    --train
  • by default, the script will create a new folder at ./models/<STAGE1_EXP_FOLDER> for saving the checkpoints.

Validate + extracting the top-k files

python src/validate.py --dataset fashionIQ \
    --stage1-path <STAGE1_EXP_FOLDER> \
    --save-topk --k 200
  • replace <STAGE1_EXP_FOLDER> with the Stage 1 experiment folder;
  • the top-k files will be saved at ./models/<STAGE1_EXP_FOLDER>/.
Alternatively, to reproduce our stage 1 results   
python src/validate.py --dataset fashionIQ \
    --stage1-path stage1/fashionIQ/ \
    --save-topk --k 200
  • which will also save the top-k files at ./models/stage1/fashionIQ/. They should be identical to our provided files.

 

Stage 2

Train

# Optional: comet experiment logging --api-key and --workspace
python src/stage2_train.py --dataset FashionIQ \
    --api-key <COMET_API_KEY> --workspace <COMET_WORKSPACE> \
    --blip-learning-rate 2e-5 --blip-max-epoch 50 \
    --num-epochs 50 --batch-size 16 --blip-bs 16 \
    --transform targetpad --target-ratio 1.25 \
    --save-training --save-best \
    --top-k-path <TOPK_FILE_PATH> --K-value 100 \
    --validation-frequency 1 \
    --experiment-name BLIP_stageII_b16_2e-5_cos50_k100+[BLIP_stageI_b512_2e-5_cos20] \
    --blip-model-path <STAGE1_EXP_FOLDER> \
    --train
  • replace <TOPK_FILE_PATH> with the path of the generated files, you might wish to modify this function;
  • replace <STAGE1_EXP_FOLDER> with the Stage 1 experiment folder;
  • by default, the script will create a new folder at ./models/<STAGE2_EXP_FOLDER> for saving the checkpoints.
Alternatively, to use our extracted top-k file and/or our stage 1 checkpoint   
python src/stage2_train.py --dataset FashionIQ \
    --api-key <COMET_API_KEY> --workspace <COMET_WORKSPACE> \
    --blip-learning-rate 2e-5 --blip-max-epoch 50 \
    --num-epochs 50 --batch-size 16 --blip-bs 16 \
    --transform targetpad --target-ratio 1.25 \
    --save-training --save-best \
    --top-k-path BLIP_stageI_b512_2e-5_cos20 --K-value 100 \
    --validation-frequency 1 \
    --experiment-name BLIP_stageII_b16_2e-5_cos50_k100+[BLIP_stageI_b512_2e-5_cos20] \
    --blip-model-path stage1/fashionIQ \
    --train

 

Validate

python src/validate_stage2.py --dataset fashionIQ \
    --stage1-path <STAGE1_EXP_FOLDER> --stage2-path <STAGE2_EXP_FOLDER> \
    --top-k-path <TOPK_FILE_PATH> \
    --k 100
  • replace <TOPK_FILE_PATH> with the path of the generated files, you might wish to modify this function;
  • replace <STAGE1_or_2_EXP_FOLDER> with the previous experiment folders.
Alternatively, to reproduce our stage 2 results   
python src/validate_stage2.py --dataset fashionIQ \
    --stage1-path stage1/fashionIQ --stage2-path stage2/fashionIQ \
    --top-k-path BLIP_stageI_b512_2e-5_cos20 \
    --k 100