config.json

{
  "model_type": "kiri-ocr",
  "architectures": ["KiriOCR"],
  "library_name": "pytorch",
  "torch_dtype": "float32",

  "hidden_size": 256,
  "num_hidden_layers": 7,
  "num_attention_heads": 8,

  "image_height": 48,
  "image_width": 640,
  "max_decoder_length": 260,

  "vocab_size": 966,
  "unk_token": "<unk>",

  "collapse_whitespace": true,
  "unicode_nfc": true,

  "encoder": {
    "dim": 256,
    "num_layers": 4,
    "num_heads": 8,
    "feedforward_dim": 1024,
    "dropout": 0.15
  },

  "decoder": {
    "enabled": true,
    "dim": 256,
    "num_layers": 3,
    "num_heads": 8,
    "feedforward_dim": 1024
  },

  "ctc": {
    "enabled": true,
    "fusion_alpha": 0.5
  },

  "language_model": {
    "enabled": true,
    "fusion_enabled_eval": true,
    "fusion_alpha": 0.35
  },

  "inference": {
    "use_fp16": false,
    "use_autocast": false,
    "beam_width": 4,
    "beam_length_penalty": 0.6,
    "eos_logp_bias": 5.0,
    "eos_logp_boost": 5.0,
    "eos_bias_until_length": 3,
    "repeat_last_penalty": 3.0,
    "unk_logp_penalty": 2.0,
    "max_length_ratio": 1.5,
    "max_length_pad": 10,
    "memory_max_length_ratio": 0.75
  },

  "preprocessing": {
    "mean": 0.5,
    "std": 0.5,
    "pad_value": 128
  },

  "transformers_version": "4.40.0"
}