WARNING Models created here work with the BoatController app in android/BoatController
Simply copy the compiled model into 
    BoatController/app/src/pytorch/assets/banks.ptl
=======================================================================================

20241229 - Usage summary

Here is an example of how to augment the dataset:
    # ./mkframes.py -f 30 -t /data/river/tags -d /data/river/newtags /data/river/movies/gelise-1.mp4
    # ./client.py --datadir /data/river/newtags 
    # ./tagger.py /data/river/newtags/*jpg
    Merge the content newtag/  into the main tag/ directory.

Training & predicting
    # ./main.py -c catalog/ train --help
    # ./main.py -c catalog/ predict BasicModel-k9x9-f32:16:8:4-conv /data/river/movies/seine4.mp4
    

TODO
- IoU for loss function
- Integrate tqdm
- Generate final model with --seed 0 to optimize it

Varying the positional layer position in the netowrk:
$ ./client.py -c positional-at-beginning rank /data/river/test-tags
BasicModel-k9x9-f32:16:8:4-conv	93.95
BasicModel-k9x9-f32:16:8:4-linear	93.44

$ ./client.py -c positional-at-end rank /data/river/test-tags
BasicModel-k9x9-f32:16:8:4-conv	93.91
BasicModel-k9x9-f32:16:8:4-linear	93.37

Varying the positional layers:

$ ./client.py -c positional-parameters/ rank /data/river/test-tags/
BasicModel-k9x9-f32:16:8:4-conv	94.41
BasicModel-k9x9-f32:16:8:4-linear	94.36
$ ./client.py --catalog height-only-positional rank /data/river/test-tags
BasicModel-k9x9-f32:16:8:4	92.62
BanksModel-k5x5-codec16:16	91.39
$ ./client.py --catalog height-and-width-positional rank /data/river/test-tags
BasicModel-k9x9-f32:16:8:4	93.03
BanksModel-k5x5-codec16:16	91.16


20241221 - Tools
Two main tools here:

# ./train.py --catalog <directory> train --epochs N --checkpoint C
runs training of the models defined  in train.py::train() for N epochs, 
saving state every C epochs.

# ./client.py --catalog <direcory> 
runs various tasks and metrics through the model.

Varying the amount of random crops in the dataset:


$ ./client.py --catalog nocrop rank /data/river/test-tags
BasicModel-k5x5-f32:16:8	92.38
BanksModel-k3x3-codec16:16	92.15
BanksModel-k5x5-codec16:16	91.85
BasicModel-k3x3-f32:16:8	91.51
BanksModel-k7x7-codec16:16	91.10
BasicModel-k7x7-f32:16:8	91.01

$ ./client.py --catalog half-crop rank /data/river/test-tags
BasicModel-k8x8-f32:16:8	93.19
BasicModel-k8x8-f32:16:16	93.11
BasicModel-k9x9-f16:32:64	93.03
BasicModel-k8x8-f32:32:32	92.99
BasicModel-k9x9-f32:16:8	92.98
BasicModel-k8x8-f16:32:64	92.97
BasicModel-k9x9-f32:32:32	92.33
BasicModel-k5x5-f32:16:8	92.29
BasicModel-k9x9-f32:16:16	92.29
BasicModel-k7x7-f32:16:8	91.94
BasicModel-k7x7-f16:32:64	91.92
BanksModel-k7x7-codec16:16	91.82
BanksModel-k5x5-codec16:32	91.81
BanksModel-k5x5-codec16:16	91.78
BanksModel-k9x9-codec16:16	91.56
BanksModel-k3x3-codec16:16	91.37
BanksModel-k8x8-codec16:16	91.02
BasicModel-k3x3-f32:16:8	90.74


$ ./client.py --catalog full-crop rank /data/river/test-tags
Basic4Model-k9x9-f32:16:8:4	93.21
Basic4Model-k8x8-f32:16:8:4	93.04
BanksModel-k5x5-codec16:16	92.09
BasicModel-k7x7-f32:16:8	92.02
BanksModel-k5x5-codec16:32	91.64
BasicModel-k8x8-f32:16:8	91.63
BanksModel-k3x3-codec16:16	91.49
BanksModel-k7x7-codec16:16	91.31
BasicModel-k5x5-f32:16:8	91.20
BasicModel-k3x3-f32:16:8	90.79