Implement training in the script

Depends #5 (closed)

We now need to update worker_pylaia/train.py to support training from a local dataset.

New arguments:

• --img-path, Path, required, path to the training dataset's images
• --data-path, Path, required, path to the training dataset

Define the following functions:

• get_model_training_config(): this will return a JSON version of the following YAML file (inspired by the original training YAML config file)

img_dirs: <args.img_path>
syms: <args.syms_path>
tr_txt_table: <args.data_path> / "train.txt"
va_txt_table: <args.data_path> / "val.txt"
common:
  train_path: <created_tmpdir>
data:
  batch_size: 8
  color_mode: L
optimizer:
  learning_rate: 0.0005
  name: RMSProp
scheduler:
  active: true
  monitor: va_loss
train:
  augment_training: false
  early_stopping_patience: 50
  delimiters: 
   - <space>
trainer:
  auto_select_gpus: true
  max_epochs: 1

• train_model(): this will make a call to laia.scripts.htr.train_ctc.run (renamed as train_model) defined here with the following arguments

run(
    syms=config["syms"],
    img_dirs=config["img_dirs"],
    tr_txt_table=config["tr_txt_table"],
    va_txt_table=config["va_txt_table"],
    common=CommonArgs(**config["common"]),
    train=TrainArgs(**config["train"]),
    optimizer=OptimizerArgs(**config["optimizer"]),
    scheduler=SchedulerArgs(**config["scheduler"]),
    data=DataArgs(**config["data"]),
    trainer=TrainerArgs(**config["trainer"]),
)

Running the script will call first get_model_creation_config then create_model to create the model. Then it will call get_model_training_config() and train_model() to train the model.

To run the script, the following directory architecture is now expected (line images may be in subfolders):

<args.data_path>
│   train.txt
│   val.txt    
│
└───images
│   │   line001.jpg
│   │   line002.jpg
|   |   ...