Custom dataset

[deeksha1927]

Could you please let us know the folder structure required for custom dataset training?

[selingok033]

If you find the result, could you share with us?

[ANDRESHZ]

for a custom dataset, I use:

"AnyDatasetName" folder:
   -"train" folder:
      -"imgs" folder:
        -"1"..."#last_class" folders:

example image

   -"val" folder:
      -6 validation folder create by the 
        python convert.py --rec_path <datafolder>/<faces_webface_112x112> --make_image_files --make_validation_memfiles

example image

and the main.py comand:

    --data_root /.../datafolder/AnyDatasetName \
    --train_data_path train \
    --val_data_path val \
    --prefix ir50_ms1mv2_adaface \
    --gpus 1 \
    --use_16bit \
    --arch ir_50 \
    --batch_size 6 \
    --num_workers 3 \
    --epochs 13 \
    --lr_milestones 12,20,24 \
    --lr 0.1 \
    --head adaface \
    --m 0.4 \
    --h 0.333 \
    --custom_num_class 62 \
    --low_res_augmentation_prob 0.2 \
    --crop_augmentation_prob 0.25 \
    --photometric_augmentation_prob 0.2 \
    --start_from_model_statedict pretrained/adaface_ir50_ms1mv2.ckpt

[ANDRESHZ]

Any of you have a solution for: this issue #75?

Or find a nice guide to finalize the fine-tuning?

[selingok033]

@ANDRESHZ Thank you for your response but I didn't fully understand what path I should provide in "/<faces_webface_112x112>". Additionally, I would like to mention that I won't be using face data as a custom dataset, instead, I will be using another biometric feature to fine-tune Adaface. Can you please let me know if you can assist me with this?

[ANDRESHZ]

@ANDRESHZ Thank you for your response but I didn't fully understand what path I should provide in "/<faces_webface_112x112>". Additionally, I would like to mention that I won't be using face data as a custom dataset, instead, I will be using another biometric feature to fine-tune Adaface. Can you please let me know if you can assist me with this?

@selingok033 You must use the folder that you want to convert the .rec and the binary files to create the 5 validation datasets.
python convert.py --rec_path <datafolder>/faces_webface_112x112 --make_image_files --make_validation_memfiles
the .rec and .bin files are into the .zip file faces_webface_112x112.zip. (see section one of training .MD file on github).

OK if you will have not used custom dataset like actual "issue thread #81" you will modify the "head" to add the structure EX: into a MLP and get union with the structure after to the load de ".ckpt" file, or would you train from 0 changing the head to admit you new additional metadata.