Reorganize parameters in training configuration

There is a whole lot of parameters in the training configuration. Some parameters are also not in the right section. As we prepare for the configuration file, all sections should make sense.

changed milestone to %DAN-P5: Improve training

added P2 Quick Win labels

assigned to @yschneider

marked this issue as related to #170

Here is a suggested organization of parameters:

• mlflow (not changed)
• dataset
  • datasets
  • train
  • val
  • test
  • tokens
  • max_char_prediction
• model
  • models (can be removed)
  • additional_tokens
  • encoder_params
    • encoder (= models/encoder param)
    • dropout
    • enc_dim (might also be needed in the decoder, to be checked)
    • nb_layers
    • h_max
    • w_max
  • decoder_params
    • decoder (= models/decoder param)
    • l_max
    • dec_num_layers
    • dec_num_heads
    • dec_res_dropout
    • dec_pred_dropout
    • dec_att_dropout
    • dec_dim_feedforward
    • attention_win
• training
  • data
    • batch_size
    • load_in_memory
    • worker_per_gpu
    • preprocessings
    • augmentation
  • device
    • use_ddp
    • ddp_port
    • nb_gpu
    • force_cpu
    • use_amp
  • output_folder
  • max_nb_epochs
  • load_epoch
  • optimizers
  • lr_schedulers
  • label_noise_scheduler
  • metrics
    • train_metrics
    • valid_metrics (= eval_metrics param)
  • validation
    • eval_on_valid
    • eval_on_valid_interval
    • set_name_focus_metric
  • transfer_learning
    • encoder (= model_params/transfer_learning/encoder param)
    • decoder (= model_params/transfer_learning/decoder param)
    • charset (= model_params/transfered_charset param)

mentioned in merge request !242 (merged)

closed with merge request !242 (merged)