seeker
57 lines
1.2 KiB
YAML
57 lines
1.2 KiB
YAML
seed: 11
|
|
|
|
paths:
|
|
model_dir: models/qwen2.5_0.5B_vl
|
|
pai_raw_dir: data/raw/pai
|
|
manifest: data/processed/overfit_manifest.jsonl
|
|
stage1_dir: outputs/stage1_vlm
|
|
stage2_dir: outputs/stage2_expert
|
|
|
|
vlm:
|
|
repo_id: thkim0305/qwen2.5_0.5B_vl
|
|
dtype: float16
|
|
attn_implementation: eager
|
|
max_reasoning_tokens: 48
|
|
num_additional_image_tokens: 1
|
|
|
|
data:
|
|
repo_id: nvidia/PhysicalAI-Autonomous-Vehicles
|
|
num_clips: 1
|
|
camera: camera_front_wide_120fov
|
|
num_frames: 1
|
|
history_steps: 8
|
|
source_future_steps: 64
|
|
expert_waypoints: 16
|
|
time_step: 0.1
|
|
|
|
stage1:
|
|
steps: 80
|
|
learning_rate: 0.0001
|
|
weight_decay: 0.0
|
|
adam_epsilon: 0.0001
|
|
freeze_vision_tower: true
|
|
freeze_multimodal_projector: true
|
|
train_last_n_language_layers: 8
|
|
train_final_norm: true
|
|
train_lm_head: false
|
|
gradient_checkpointing: true
|
|
log_every: 5
|
|
|
|
stage2:
|
|
steps: 1000
|
|
learning_rate: 0.0001
|
|
weight_decay: 0.0
|
|
adam_epsilon: 0.0001
|
|
expert_hidden_size: 512
|
|
expert_intermediate_size: 2432
|
|
expert_num_layers: 24
|
|
expert_num_attention_heads: 8
|
|
expert_num_key_value_heads: 2
|
|
expert_head_dim: 64
|
|
expert_non_causal_attention: true
|
|
action_projection_hidden_size: 256
|
|
action_fourier_features: 20
|
|
inference_steps: 10
|
|
log_every: 50
|
|
checkpoint_every: 50
|