checkpoint-5200/config.json

{
  "batch_size": 32,
  "buffer_size": 64,
  "eval_mix": "chris_aubo",
  "frozen_keys": [
    "*hf_model*"
  ],
  "lora": false,
  "lora_config": {
    "bias": "none",
    "lora_alpha": 16,
    "lora_dropout": 0.05,
    "r": 8
  },
  "model": {
    "heads": {
      "action": {
        "args": [],
        "kwargs": {
          "action_dim": 7,
          "action_horizon": 4,
          "dropout_rate": 0.0,
          "n_diffusion_samples": 1,
          "readout_key": "readout_action",
          "token_embedding_size": 768,
          "use_map": false
        },
        "module": "octo.model.components.action_heads",
        "name": "DiffusionActionHead"
      }
    },
    "max_horizon": 10,
    "observation_tokenizers": {
      "primary": {
        "args": [],
        "kwargs": {
          "encoder": {
            "args": [],
            "kwargs": {
              "in_features": 6
            },
            "module": "octo.model.components.vit_encoders",
            "name": "SmallStem16"
          },
          "obs_stack_keys": [
            "image_primary"
          ],
          "task_stack_keys": [
            "image_primary"
          ]
        },
        "module": "octo.model.components.tokenizers",
        "name": "ImageTokenizer"
      }
    },
    "readouts": {
      "action": 1
    },
    "repeat_task_tokens": true,
    "task_tokenizers": {
      "language": {
        "args": [],
        "kwargs": {
          "encoder": "t5-base",
          "finetune_encoder": false
        },
        "module": "octo.model.components.tokenizers",
        "name": "LanguageTokenizer"
      }
    },
    "token_embedding_size": 768,
    "transformer_kwargs": {
      "add_position_embedding": false,
      "attention_dropout_rate": 0.0,
      "dropout_rate": 0.0,
      "mlp_dim": 3072,
      "num_attention_heads": 12,
      "num_layers": 12
    },
    "use_correct_attention": true
  },
  "normalize_method": "sign",
  "obs_token_nums": {
    "primary": 256
  },
  "output_dir": "ljp_aubo_20250610_test",
  "pretrained": {
    "kwargs": {
      "subpath": "oxe-g2-checkpoint-300000"
    },
    "model": "hf://chuanmew/octo_torch"
  },
  "run_name": "aubo",
  "sampler_num_samples": 2560000,
  "seed": 42,
  "subsample_length": 99999,
  "text_processor": {
    "args": [],
    "kwargs": {
      "encode_with_model": false,
      "tokenizer_kwargs": {
        "max_length": 16,
        "padding": "max_length",
        "return_tensors": "np",
        "truncation": true
      },
      "tokenizer_name": "t5-base"
    },
    "module": "octo.components.text_processing",
    "name": "HFTokenizer"
  },
  "train_mix": "ljp_aubo_merged_20250519to20250529_20250604",
  "training_arguments": {
    "bf16": true,
    "dataloader_pin_memory": true,
    "ddp_find_unused_parameters": true,
    "eval_steps": 1000,
    "gradient_accumulation_steps": 1,
    "learning_rate": 3e-05,
    "logging_nan_inf_filter": false,
    "logging_steps": 1000,
    "lr_scheduler_kwargs": {
      "decay_type": "cosine",
      "min_lr_ratio": 0.05,
      "num_decay_steps": 0
    },
    "lr_scheduler_type": "warmup_stable_decay",
    "max_grad_norm": 1.0,
    "max_steps": 5200,
    "optim": "adamw_torch_fused",
    "per_device_eval_batch_size": 1,
    "per_device_train_batch_size": 1,
    "report_to": "wandb",
    "save_safetensors": true,
    "save_steps": 5200,
    "torch_compile": true,
    "warmup_steps": 100,
    "weight_decay": 0.1
  },
  "training_keys": [
    "*action*",
    "*observation_tokenizers*",
    "*transformer.transformer*",
    "*task_projections*",
    "*obs_projections*",
    "*readout_embeddings*",
    "*task_pos_embeddings*",
    "*obs_pos_embeddings*"
  ]
}