spatial_training

Minimal utilities for VLM SFT + Pose supervision.

• training_scripts/train_sft.py: SFT training
• training_scripts/train_pose.py: Pose training (adds pose loss)
• training_scripts/build_pose_dataset.py: build a HuggingFace dataset folder via save_to_disk()
• utils/: collators, resize transform, pose loss, feature extractor

1) Setup (Conda + pip)

cd /Projects/SG_VLN_HumanData/spatial_training
conda create -y -p /root/conda_envs/spatial_training python=3.11
conda activate /root/conda_envs/spatial_training
python3 -m pip install -U pip
python3 -m pip install -r requirements.txt

2) Train

Single GPU:

export CUDA_VISIBLE_DEVICES=0

SFT:

python3 training_scripts/train_sft.py \
  --model_id /Projects/SG_VLN_HumanData/SG-VLN/sft_pipeline/text_adapted_model \
  --train_dataset_dir /Projects/SG_VLN_HumanData/spatial_training/data/habitat_web_pose_v1/train \
  --eval_dataset_dir /Projects/SG_VLN_HumanData/spatial_training/data/habitat_web_pose_v1/validation \
  --output_dir ./dump/sft_training_continue

Pose:

python3 training_scripts/train_pose.py \
  --model_id /Projects/SG_VLN_HumanData/SG-VLN/sft_pipeline/text_adapted_model \
  --train_dataset_dir /Projects/SG_VLN_HumanData/spatial_training/data/habitat_web_pose_v1/train \
  --eval_dataset_dir /Projects/SG_VLN_HumanData/spatial_training/data/habitat_web_pose_v1/validation \
  --output_dir ./dump/pose_training_continue \
  --resume_from_checkpoint /Projects/SG_VLN_HumanData/spatial_training/dump/pose_training_test/checkpoint-9500