Download src/data/__init__.py from joelinator/dflow-novo-model: direct link, hf CLI and curl.
- Browser
- Download file 1.43 kB
-
https://huggingface.co/joelinator/dflow-novo-model/resolve/main/src/data/__init__.py
- Command line
-
hf download hf://joelinator/dflow-novo-model/src/data/__init__.py
-
curl -L -o __init__.py https://huggingface.co/joelinator/dflow-novo-model/resolve/main/src/data/__init__.py
1.43 kB
| from .constants import AA_MASSES_DICT, M_H, STANDARD_AMINO_ACIDS | |
| __all__ = [ | |
| "AA_MASSES_DICT", | |
| "M_H", | |
| "STANDARD_AMINO_ACIDS", | |
| "MAX_PEPTIDE_LENGTH", | |
| "MIN_PEPTIDE_LENGTH", | |
| "NUM_LENGTH_CLASSES", | |
| "SpectrumDataSet", | |
| "build_dataloader", | |
| "build_vocabulary", | |
| "apply_length_padding", | |
| "class_to_length", | |
| "clamp_length", | |
| "decoder_output_size", | |
| "decoder_output_token_ids", | |
| "get_aa_masses", | |
| "get_dataset", | |
| "get_output_aa_masses", | |
| "length_to_active_mask", | |
| "length_to_class", | |
| "spectrum_collate", | |
| "validate_peptide_length", | |
| ] | |
| _LENGTH_NAMES = { | |
| "MAX_PEPTIDE_LENGTH", | |
| "MIN_PEPTIDE_LENGTH", | |
| "NUM_LENGTH_CLASSES", | |
| "apply_length_padding", | |
| "class_to_length", | |
| "clamp_length", | |
| "length_to_active_mask", | |
| "length_to_class", | |
| "validate_peptide_length", | |
| } | |
| _DATA_NAMES = { | |
| "SpectrumDataSet", | |
| "build_dataloader", | |
| "build_vocabulary", | |
| "decoder_output_size", | |
| "decoder_output_token_ids", | |
| "get_aa_masses", | |
| "get_dataset", | |
| "get_output_aa_masses", | |
| "spectrum_collate", | |
| } | |
| def __getattr__(name: str): | |
| if name in _LENGTH_NAMES: | |
| from . import lengths as lengths_module | |
| return getattr(lengths_module, name) | |
| if name in _DATA_NAMES: | |
| from . import data as data_module | |
| return getattr(data_module, name) | |
| raise AttributeError(f"module {__name__!r} has no attribute {name!r}") | |