train_manager.yaml

---
mode: 'train'
device: [0]                 # e.g. [0,1,2,3]
batch_size: 1
num_workers: 10             # num_workers == batchsize_per_gpu works best for my model
rank_metric: 'f1_score'     # precision

## training config
fix_randomness: true
epochs: 1000
lr: 0.001
wd: 0                       # weight_decay

## evaluation
nn_thresh: 0.7
mutual_nn: true

## debug
debug_times:              false
dmsg_freq: 10
ignore_warnings:          true

## data
dataset_type: homography
backup_path: 'assets/dataset/train_results' 

homography: 
    data_path: ['assets/dataset/dataset_h5']
    resume: false
    checkpoint_path: 'assets/dataset/train_results/mode=train.input=linetr.lr=0.001.bs=30.wd=0.time=2022-03-26@13-24/model_best.pth.tar' 

linetr:
    mode: 'train'
    image_shape: !!python/tuple [480,640]   # height, width
    min_length: 16
    token_distance: 8
    max_tokens: 21
    remove_borders: 1
    max_keylines: -1

    descriptor_dim: 256
    keyline_encoder: [32, 64, 128, 256]
    n_heads: 4
    n_line_descriptive_layers: 12
    d_inner: 1024    # d_inner at the Feed_Forward Layer
...