Upload RFM model
Browse files- config.yaml +109 -0
config.yaml
ADDED
|
@@ -0,0 +1,109 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
data:
|
| 2 |
+
dataloader_num_workers: 0
|
| 3 |
+
dataloader_pin_memory: false
|
| 4 |
+
dataset_preference_ratio: 0.7
|
| 5 |
+
dataset_type: default
|
| 6 |
+
eval_datasets:
|
| 7 |
+
- abraranwar/libero_rfm
|
| 8 |
+
- ykorkmaz/libero_failure_rfm
|
| 9 |
+
- HenryZhang/metaworld_rewind_rfm_eval
|
| 10 |
+
eval_subset_size: 500
|
| 11 |
+
eval_subsets:
|
| 12 |
+
- - libero256_10
|
| 13 |
+
- - libero_10_failure
|
| 14 |
+
- - metaworld_rewind_eval
|
| 15 |
+
force_reprocess: false
|
| 16 |
+
fps: 10
|
| 17 |
+
max_frames: 16
|
| 18 |
+
max_frames_for_preprocessing: 64
|
| 19 |
+
max_trajectories: -1
|
| 20 |
+
model_type: default
|
| 21 |
+
n_wrong_tasks: 5
|
| 22 |
+
num_bins: 10
|
| 23 |
+
num_proc: 1
|
| 24 |
+
preference_ratio: 1.0
|
| 25 |
+
preference_strategy_ratio:
|
| 26 |
+
- 0.4
|
| 27 |
+
- 0.3
|
| 28 |
+
- 0.3
|
| 29 |
+
- 0.0
|
| 30 |
+
progress_ratio: 0.5
|
| 31 |
+
resized_height: 128
|
| 32 |
+
resized_width: 128
|
| 33 |
+
rewind_lengths: null
|
| 34 |
+
samples_per_trajectory: 1
|
| 35 |
+
seed: 42
|
| 36 |
+
shuffle: true
|
| 37 |
+
train_datasets:
|
| 38 |
+
- abraranwar/libero_rfm
|
| 39 |
+
- ykorkmaz/libero_failure_rfm
|
| 40 |
+
- HenryZhang/metaworld_rewind_rfm_train
|
| 41 |
+
train_subsets:
|
| 42 |
+
- - libero256_90
|
| 43 |
+
- - libero_90_failure
|
| 44 |
+
- - metaworld_rewind_train
|
| 45 |
+
video_frame_sampling: uniform
|
| 46 |
+
debug: false
|
| 47 |
+
logging:
|
| 48 |
+
print_trainable_parameters: true
|
| 49 |
+
save_model: true
|
| 50 |
+
save_processor: true
|
| 51 |
+
use_wandb: true
|
| 52 |
+
wandb_entity: clvr
|
| 53 |
+
wandb_project: rfm
|
| 54 |
+
wandb_run_name: rfm
|
| 55 |
+
mode: train
|
| 56 |
+
model:
|
| 57 |
+
base_model_id: Qwen/Qwen2.5-VL-3B-Instruct
|
| 58 |
+
torch_dtype: bfloat16
|
| 59 |
+
train_language_model: false
|
| 60 |
+
train_preference_head: true
|
| 61 |
+
train_progress_head: true
|
| 62 |
+
train_similarity_head: false
|
| 63 |
+
train_value_head: true
|
| 64 |
+
train_vision_encoder: true
|
| 65 |
+
trust_remote_code: true
|
| 66 |
+
peft:
|
| 67 |
+
bias: none
|
| 68 |
+
lora_alpha: 64
|
| 69 |
+
lora_dropout: 0.05
|
| 70 |
+
peft_vision_encoder: true
|
| 71 |
+
r: 32
|
| 72 |
+
target_modules:
|
| 73 |
+
- q_proj
|
| 74 |
+
- k_proj
|
| 75 |
+
- v_proj
|
| 76 |
+
- o_proj
|
| 77 |
+
- gate_proj
|
| 78 |
+
- up_proj
|
| 79 |
+
- down_proj
|
| 80 |
+
use_peft: false
|
| 81 |
+
training:
|
| 82 |
+
beta: 0.1
|
| 83 |
+
bf16: true
|
| 84 |
+
ddp_bucket_cap_mb: 25
|
| 85 |
+
ddp_find_unused_parameters: true
|
| 86 |
+
do_eval: true
|
| 87 |
+
eval_steps: 100
|
| 88 |
+
evaluation_strategy: steps
|
| 89 |
+
fp16: false
|
| 90 |
+
gradient_accumulation_steps: 1
|
| 91 |
+
gradient_checkpointing: true
|
| 92 |
+
learning_rate: 2.0e-05
|
| 93 |
+
logging_steps: 1
|
| 94 |
+
lr_scheduler_type: cosine
|
| 95 |
+
max_grad_norm: 10.0
|
| 96 |
+
max_seq_length: 1024
|
| 97 |
+
max_steps: 5000
|
| 98 |
+
num_gpus: 2
|
| 99 |
+
num_train_epochs: -1
|
| 100 |
+
output_dir: ./logs/rfm_progpref_peft_vision
|
| 101 |
+
per_device_eval_batch_size: 8
|
| 102 |
+
per_device_train_batch_size: 8
|
| 103 |
+
prediction_loss_only: true
|
| 104 |
+
remove_unused_columns: false
|
| 105 |
+
resume_from_checkpoint: null
|
| 106 |
+
save_steps: 200
|
| 107 |
+
save_strategy: steps
|
| 108 |
+
warmup_ratio: 0.1
|
| 109 |
+
warmup_steps: 0
|