Reproducibility and over-removal problem #5
Description
I ran the following training script:
training scripts
python -m train \
--gin_configs=configs/360_dino.gin \
--gin_bindings="Config.data_dir='Datasets/on-the-go/mountain'" \
--gin_bindings="Config.checkpoint_dir='output/mountain/run_1/checkpoints'" \
--gin_bindings="Config.patch_size=32" \
--gin_bindings="Config.dilate=4" \
--gin_bindings="Config.data_loss_type='on-the-go'"
python -m train \
--gin_configs=configs/360_dino.gin \
--gin_bindings="Config.data_dir='Datasets/on-the-go/fountain'" \
--gin_bindings="Config.checkpoint_dir='output/fountain/run_1/checkpoints'" \
--gin_bindings="Config.patch_size=32" \
--gin_bindings="Config.dilate=4" \
--gin_bindings="Config.data_loss_type='on-the-go'"
python -m train \
--gin_configs=configs/360_dino.gin \
--gin_bindings="Config.data_dir='Datasets/on-the-go/corner'" \
--gin_bindings="Config.checkpoint_dir='output/corner/run_1/checkpoints'" \
--gin_bindings="Config.patch_size=32" \
--gin_bindings="Config.dilate=4" \
--gin_bindings="Config.data_loss_type='on-the-go'"
python -m train \
--gin_configs=configs/360_dino.gin \
--gin_bindings="Config.data_dir='Datasets/on-the-go/patio'" \
--gin_bindings="Config.checkpoint_dir='output/patio/run_1/checkpoints'" \
--gin_bindings="Config.patch_size=32" \
--gin_bindings="Config.dilate=4" \
--gin_bindings="Config.data_loss_type='on-the-go'" \
--gin_bindings="Config.train_render_every = 5000" \
--gin_bindings="Config.H = 1080" \
--gin_bindings="Config.W = 1920" \
--gin_bindings="Config.factor = 4" \
--gin_bindings="Config.feat_rate = 2"
python -m train \
--gin_configs=configs/360_dino.gin \
--gin_bindings="Config.data_dir='Datasets/on-the-go/spot'" \
--gin_bindings="Config.checkpoint_dir='output/spot/run_1/checkpoints'" \
--gin_bindings="Config.patch_size=32" \
--gin_bindings="Config.dilate=4" \
--gin_bindings="Config.data_loss_type='on-the-go'"
python -m train \
--gin_configs=configs/360_dino.gin \
--gin_bindings="Config.data_dir='Datasets/on-the-go/patio_high'" \
--gin_bindings="Config.checkpoint_dir='output/patio_high/run_1/checkpoints'" \
--gin_bindings="Config.patch_size=32" \
--gin_bindings="Config.dilate=4" \
--gin_bindings="Config.data_loss_type='on-the-go'"
and found that our eval_metrics/psnr was much worse than the paper's metrics:
| Mountain | Fountain | Corner | Patio | Spot | Patio-High | |
|---|---|---|---|---|---|---|
| onthego (paper) | 20.15 | 20.11 | 24.22 | 20.78 | 23.33 | 21.41 |
| onthego (our env) | 18.34 | 15.59 | 20.82 | 16.71 | 16.32 | 14.59 |
In addition, we found that the fountain was missing in the rendered fountain model.
Do you have any suggestion to fix it and reproduce the paper's PSNR?
