lingyi0110/mai2bert

How lingyi0110/mai2bert works

mai2bert is a self-supervised pretraining and supervised chart-difficulty prediction framework for structured maimai event sequences. Its main lifecycle is to discover and parse chart files, build augmented event representations and splits, store processed data, train or fine-tune sequence models, then load a checkpoint to predict difficulty and uncertainty for chart inputs. The graph preserves both training and inference branches. The README is empty, and several modules were not sampled; their paths identify subsystems, but uncertain internal call links are omitted.

Main entry points: prepare.py Preparation pipeline · regression.py Difficulty regressor · batch.py Batch prediction · train.py Training command

Components 27

Chart preparation

Prepare command prepare.py
Chart discovery src/preprocessing/discovery.py
Simai parser src/lib/parser/simai.py
MA2 parser src/lib/parser/ma2.py
Event representation src/preprocessing/representation.py
Chart augmentation src/lib/augmentation
Dataset splits src/preprocessing/split.py
Preparation pipeline src/preprocessing/prepare.py

Dataset pipeline

Processed HDF5 src/preprocessing/store.py
Event dataset src/data/dataset.py
Batch collation src/data/batching.py
Pretraining masking src/data/masking.py

Sequence models

Event embedding src/models/embedding.py
Sequence encoder src/models/encoder.py
Pretraining model src/models/pretraining.py
Difficulty regressor src/models/regression.py

Training

Training command train.py
Pretraining loop src/training/pretrain.py
Regression training src/training/regression.py
Transfer training src/training/transfer.py

Prediction

Inference command infer.py
Batch prediction src/inference/batch.py
Checkpoint loader src/inference/loader.py
Prediction runtime src/inference/predict.py
Prediction files

Other components

Chart files
Model checkpoint

Connections 33

Chart files → Prepare command: supplies charts
Prepare command → Preparation pipeline: runs preparation
Preparation pipeline → Chart discovery: discovers charts
Preparation pipeline → Simai parser: parses charts
Preparation pipeline → Event representation: builds events
Preparation pipeline → Chart augmentation: creates variants
Preparation pipeline → Dataset splits: assigns splits
Preparation pipeline → Processed HDF5: writes arrays
Chart files → Inference command: supplies charts
Inference command → Batch prediction: dispatches batch
Batch prediction → Simai parser: parses Simai
Batch prediction → MA2 parser: parses MA2
Batch prediction → Prediction runtime: submits items
Inference command → Checkpoint loader: loads model
Model checkpoint → Checkpoint loader: provides weights
Checkpoint loader → Difficulty regressor: restores model
Prediction runtime → Difficulty regressor: predicts difficulty
Batch prediction → Prediction files: writes CSV and JSON
Training command → Pretraining loop: starts pretraining
Training command → Regression training: starts regression
Training command → Transfer training: starts transfer
Processed HDF5 → Event dataset: supplies events
Event dataset → Batch collation: provides samples
Batch collation → Pretraining masking: collates for masking
Pretraining masking → Pretraining model: provides masked batches
Batch collation → Difficulty regressor: provides batches
Pretraining model → Event embedding: encodes events
Difficulty regressor → Event embedding: encodes events
Event embedding → Sequence encoder: builds representations
Pretraining loop → Pretraining model: trains model
Regression training → Difficulty regressor: trains regressor
Transfer training → Difficulty regressor: adapts model
Regression training → Model checkpoint: saves weights

More Python diagrams: serengil/deepface · verl-project/verl · marimo-team/marimo · sinaptik-ai/pandas-ai · scrapegraphai/scrapegraph-ai

Drawn by GitDiagram on September 29, 2026 from the repository's file tree, README and source files. Read as Markdown ·