code-Psi0-yth-v21 / src /psi /config /model_psi0.py
Tinghong-Ye22's picture
Add files using upload-large-folder tool
c4f8e4a verified
Raw History Blame Contribute Delete
4 kB
from tyro.conf import subcommand as cmd
from typing import Union, Annotated, Optional, List
from psi.config.config import ModelConfig
from pydantic import Field
class Psi0ModelConfig(ModelConfig):
######################### hfm_action ############################
# pretrained_model_name_or_path: str = "stabilityai/stable-diffusion-3-medium-diffusers"
resnet_store_path: str | None = None # = "cache/checkpoints/resnet18/IN_1M_resnet18.pth"
pretrained_action_header_path: str | None = None
# precomputed_text_encodings: str = "cache/precomputed_text_encodings.pkl"
# ckpt_step: Optional[int] = None
rtc: bool = False
max_delay: int = 8
action_dim: int = 7 #36 # Da
action_chunk_size: int = 6 #30 # Tp
action_exec_horizon: int = 6 #30 # Ta
observation_horizon: int = 1 # To past frames
img_chunk: int = 1 # ?
n_cams: int = 1
use_obs: str = "add_token"
use_future_progress_embedding: bool = False
future_progress_dim: int = 5
dropout: float = 0.1
noise_scheduler: str = "flow"
train_diffusion_steps: int = 1000 # 默认1000步去噪
eval_diffusion_steps: int = 10
share_cam_features: bool = False
early_fusion: bool = False
# observations
odim: int = 15 # 32
# conditions
n_conditions: Optional[int] = 1
token_fusion: Optional[str] = "concat" # "concat", "cross", "perceiver"
# tune_vision_backbone: bool = False
# vision_backbone_lr: float = 1e-5
"""
training loss weight for xyz, rpy, and gripper
should all sumed to 1 in total, eg 0.1x3+0.2x3+0.1x1=1
"""
loss_w: List[float] = Field(default_factory=lambda: [0.1, 0.2, 0.1])
# noise nets
time_dim: int = 256
hidden_dim: int = 1536
num_blocks: int = 6
dim_feedforward: int = 2048
nhead: int = 24
activation: str = "gelu"
view_feature_dim: int = 1920 # views feature dim for views and/ or vlm token dim
use_film: bool = False
combined_temb: bool = False
use_dit: bool = False
################### qwen3vl ####################
# Training Schedule
weight_decay: float = 0.01 # L2 regularization strength
# Core Arguments
model_name_or_path: str = "Qwen/Qwen3-VL-2B-Instruct" # First load, for initalization
# vlm_run_dir: str | None = None
vlm_ckpt_step: str | None = None
tune_vlm: bool = False
tune_mm_llm: bool = False # [TrainingArguments] Train LLM or not
tune_mm_vision: bool = False # [TrainingArguments] Train VIT or not
tune_mm_mlp: bool = False # [TrainingArguments] Train MLP or not
# dataset_use: str = "egodex" # [DataArguments] Dataset specification
# output_dir: str = "./outputs/hfm_qwen3vl" # Output directory for checkpoints
# cache_dir: str = "./cache/models" # [TrainingArguments] Model cache location
gradient_checkpointing: bool = True # [TrainingArguments] Enable gradient checkpointing
# Learning Rate Configuration
lang_backbone_lr: float = 1e-5
mm_projector_lr: float = 1e-5 # [TrainingArguments] Projector-specific LR
vision_tower_lr: float = 1e-6 # [TrainingArguments] Vision encoder LR
optim: str = "adamw_torch" # [TrainingArguments] Optimizer selection
# Sequence Configuration
model_max_length: int = 4096 # [TrainingArguments] Max sequence length
data_flatten: bool = True # [DataArguments] Concatenate batch sequences
data_packing: bool = True # [DataArguments] Using packing data
# Image Processing
max_pixels: int = 576 * 28 * 28 # [DataArguments] Max image pixels (H*W) for image
min_pixels: int = 16 * 28 * 28 # [DataArguments] Min image pixels for image
# lora_r: int = 8 # [TrainingArguments] LoRA r
# lora_alpha: int= 16 # [TrainingArguments] LoRA alpha
# lora_dropout: float = 0.0 # [TrainingArguments] LoRA dropout