Checkpoint step 400 (epoch 0.20)

Checkpoint step 300 (epoch 0.15)

Checkpoint step 200 (epoch 0.10)

Checkpoint step 100 (epoch 0.05)

Delete results.json

Delete REPORT.md

Delete images

Upload 32 files

fix: also update inference.py to support loading from HF repo ID

fix: use snapshot_download instead of hf_hub_download for LoRA loading

Checkpoint step 100 (epoch 0.05)

Fix safetensors shared tensor crash: override _save in OmniTrainer to save components separately

Reduce save/push interval to every 100 steps

Push intermediate results to Hub on every checkpoint + increase T4 batch_size to 2

Delete train_kaggle.py

Add train_kaggle.py — multi-GPU (2xT4) version using accelerate DDP instead of DataParallel

Update README: SlimSAM references + infer_dataset.py usage + updated file table

Add infer_dataset.py — samples from each task, runs inference, generates REPORT.md with GT vs predicted comparisons

Fix: add gradient_checkpointing_enable/disable + config proxy to SatelliteOmniLISA for Trainer compatibility

Fix AttributeError: total_mem → total_memory"

Switch SAM to SlimSAM (9.7M params, 10x smaller) — drop-in replacement, same mask_decoder API

Fix OOM on Colab/T4: lazy dataset, auto bf16/fp16, batch_size=1 for 16GB, smaller SAM placeholders, zero workers

Add setup.sh for automated dependency installation

Add requirements.txt with transformers version pinning for Qwen3.5

Update README with full documentation

Add inference script

Add training script

Add model README

initial commit