|
Download README.md from jueminghaha/FUSE-RT: direct link, hf CLI and curl.
- Browser
- Download file 5.16 kB
-
https://huggingface.co/jueminghaha/FUSE-RT/resolve/main/README.md
- Command line
-
hf download hf://jueminghaha/FUSE-RT/README.md
-
curl -L -o README.md https://huggingface.co/jueminghaha/FUSE-RT/resolve/main/README.md
5.16 kB
| language: | |
| - en | |
| tags: | |
| - chemistry | |
| - chromatography | |
| - retention-time-prediction | |
| - graph-neural-network | |
| - pytorch | |
| # FUSE-RT checkpoints | |
| This repository distributes the trained FUSE-RT checkpoints for the original **179-method RepoRT cohort**. It contains **180 checkpoints** across 18 experimental configurations, with 10 seeds per configuration. | |
| Model definitions, data preparation, evaluation scripts, and experimental results are available in the [FUSE-RT code repository](https://github.com/jueminghaha/FUSE-RT). | |
| ## Directory overview | |
| | Directory | Purpose | Configurations | Checkpoints | | |
| |---|---|---:|---:| | |
| | [`base/`](https://huggingface.co/jueminghaha/FUSE-RT/tree/main/base) | E1–E9 factorial experiments combining molecular-training strategies and RT architectures | 9 | 90 | | |
| | [`stage/`](https://huggingface.co/jueminghaha/FUSE-RT/tree/main/stage) | E8 auxiliary-data scaling with correctly paired RadonPy property labels | 4 | 40 | | |
| | [`shuffle/`](https://huggingface.co/jueminghaha/FUSE-RT/tree/main/shuffle) | E8 constrained shuffled-label controls at five auxiliary-data fractions | 5 | 50 | | |
| Every experimental directory follows the same layout: | |
| ```text | |
| <group>/<configuration>/seed_<seed>/best_model.pt | |
| ``` | |
| For example, `base/E8/seed_2034/best_model.pt` is the E8 checkpoint for split seed 2034. | |
| ## `base/`: E1–E9 experiments | |
| The molecular-training strategies are **M0**, RT training without an auxiliary objective; **M1**, RadonPy property pretraining followed by RT fine-tuning; and **M2**, joint RT and RadonPy multitask training. | |
| The RT architectures are **R0**, chromatographic-condition encoding with a shared prediction head; **R1**, chromatographic-condition encoding with method-specific prediction heads; and **R2**, method-specific prediction heads without chromatographic-condition encoding. | |
| | Directory | Molecular training | RT architecture | | |
| |---|---|---| | |
| | `base/E1/` | M0: no auxiliary objective | R0: conditions + shared head | | |
| | `base/E2/` | M0: no auxiliary objective | R1: conditions + method-specific heads | | |
| | `base/E3/` | M0: no auxiliary objective | R2: no conditions + method-specific heads | | |
| | `base/E4/` | M1: sequential pretraining | R0: conditions + shared head | | |
| | `base/E5/` | M1: sequential pretraining | R1: conditions + method-specific heads | | |
| | `base/E6/` | M1: sequential pretraining | R2: no conditions + method-specific heads | | |
| | `base/E7/` | M2: joint multitask training | R0: conditions + shared head | | |
| | `base/E8/` | M2: joint multitask training | R1: conditions + method-specific heads | | |
| | `base/E9/` | M2: joint multitask training | R2: no conditions + method-specific heads | | |
| ## `stage/` and `shuffle/`: auxiliary-data experiments | |
| Both groups use the E8 architecture. Percentages denote the requested fraction of the available RadonPy **auxiliary training pool**, not the fraction of RepoRT retention-time observations. | |
| | RadonPy fraction | Correct-label checkpoint directory | Shuffled-label checkpoint directory | | |
| |---:|---|---| | |
| | 0% | `base/E2/` | — | | |
| | 0.05% | `stage/E8_p01/` | `shuffle/E8_p01/` | | |
| | 0.334370% | `stage/E8_p02/` | `shuffle/E8_p02/` | | |
| | 2.236068% | `stage/E8_p03/` | `shuffle/E8_p03/` | | |
| | 14.953488% | `stage/E8_p04/` | `shuffle/E8_p04/` | | |
| | 100% | `base/E8/` | `shuffle/E8_p05/` | | |
| **`stage/`** contains the four intermediate correct-label scaling points. The 0% and 100% endpoints reuse `base/E2/` and `base/E8/`, respectively. Here, “stage” names the scaling-point collection; it does not refer to the M1 sequential-pretraining strategy. | |
| **`shuffle/`** contains the corresponding constrained shuffled-label controls. Within the selected auxiliary training rows, each complete property-target vector and its missing-value mask are reassigned together to another molecule. The permutation prohibits self-assignment, requires different molecular connectivity, and requires Morgan-fingerprint Tanimoto similarity below 0.50. RT labels, molecular inputs, and auxiliary validation labels remain unchanged. These checkpoints support matched comparisons of correct and disrupted molecule–property correspondence. | |
| ## Seed directories and checkpoint files | |
| Every configuration contains the same 10 seed directories: | |
| ```text | |
| seed_2004 seed_2006 seed_2011 seed_2012 seed_2016 | |
| seed_2020 seed_2022 seed_2027 seed_2032 seed_2034 | |
| ``` | |
| Each seed identifies the corresponding molecule-disjoint train/validation/internal-test split and training run. Each directory contains one validation-selected PyTorch checkpoint, `best_model.pt`, including model parameters and the saved configuration, normalization statistics, and vocabularies used by the project evaluation scripts. | |
| To use these files with the GitHub project, retain the configuration and seed subdirectories and map the top-level folders as follows: | |
| | Hugging Face directory | Directory in the GitHub project | | |
| |---|---| | |
| | `base/` | `checkpoints/experiments/` | | |
| | `stage/` | `checkpoints/scaling/` | | |
| | `shuffle/` | `checkpoints/constrained_shuffle/` | | |
| Training and evaluation procedures are documented in the [GitHub README](https://github.com/jueminghaha/FUSE-RT#reproduction). | |