cp311 →
cp312

The cp311 notebook uses pycolmap 3.11.1 installed from the /kaggle/input/pycolmap3-11/ dataset. cp312 upgrades to pycolmap 4.0.3 installed from a user notebook dataset, with a cp312-compiled wheel. All MASt3R source paths are also migrated from /kaggle/input/mast3r-fix/ to /kaggle/input/notebooks/stpeteishii/mast3r-cp312/. print(sys.version) is also added early on to explicitly verify the active Python version.

cp311 installs faiss-gpu-cu12 in a single command from the mast3r wheels folder. cp312 installs three separate wheels explicitly — faiss_gpu_cu12, nvidia_cublas_cu12, and nvidia_cuda_runtime_cu12 — from a dedicated dataset, ensuring correct CUDA runtime linkage for cp312.

ASMK's compiled C-extensions (hamming, etc.) are built per Python version. cp312 adds explicit manual loading of the cp312-compiled .so files using importlib.util.spec_from_file_location, since the cp311 binaries cannot be used.

Two monkey-patches are added in cp312 to fix FAISS compatibility issues.

① FaissGpuL2Index patch — create_index is overridden to return a CPU IndexFlatL2, bypassing GPU initialization that fails in the cp312 environment.

② Codebook.quantize patch — Rebuilds the FAISS index on demand and unifies the return signature: 3-tuple (des, word_ids, image_ids) for build_ivf and 2-tuple (des, word_ids) for query_ivf, fixing an argument mismatch that would crash ASMK retrieval.

Newer PyTorch versions require explicit whitelisting of classes used during torch.load(). cp312 adds torch.serialization.add_safe_globals([argparse.Namespace]) before loading the MASt3R checkpoint, preventing a WeightsOnlyError that would occur silently in cp311's older torch version.

This is the most significant change. The pycolmap 4.0.3 Database high-level API changed incompatibly, so cp311's approach of importing and calling the function from the external h5_to_db utility no longer works.

cp312 reimplements import_into_colmap from scratch using raw sqlite3:

The resulting database schema is identical to what pycolmap's own writer would produce — the difference is only in how it is created.

In pycolmap 3.x, cam_from_world was a plain attribute holding the pose object. In pycolmap 4.x it became a method that returns the pose. The call site in reconstruct_from_db must be updated accordingly.

cp312 adds an early-exit check after shortlisting: if no valid image pairs were found, the dataset is skipped cleanly with an informative message rather than passing an empty list into match_with_mast3r_and_save and potentially crashing downstream.

The except block in run_one_dataset now calls traceback.print_exc() in addition to printing the error message string, making failures significantly easier to diagnose since the full stack trace is shown.

cp312 introduces a log() helper with sys.stdout.flush() and wraps every major step — import_into_colmap, run_verify_matches_safe, incremental_mapping — with entry/exit log lines. This makes it easy to pinpoint where a hang or crash occurs during reconstruction.

run_mast3r_pipeline reduces the thread pool from 2 workers to 1. With the heavier memory footprint of the cp312 stack (larger pycolmap, explicit CUDA wheels), running two datasets concurrently risks OOM errors on Kaggle's GPU instances.