vipsania 1.0.0__tar.gz

vipsania-1.0.0/.gitignore

@@ -0,0 +1,178 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# poetry
+poetry.lock
+
+# checkpoints
+wandb/
+checkpoints/
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+cover/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+.pybuilder/
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+#   For a library or package, you might want to ignore these files since the code is
+#   intended to run in multiple environments; otherwise, check them in:
+# .python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# UV
+#   Similar to Pipfile.lock, it is generally recommended to include uv.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+#uv.lock
+
+# poetry
+#   Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+#   https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
+#poetry.lock
+
+# pdm
+#   Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
+#pdm.lock
+#   pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it
+#   in version control.
+#   https://pdm.fming.dev/latest/usage/project/#working-with-version-control
+.pdm.toml
+.pdm-python
+.pdm-build/
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/
+
+# PyCharm
+#  JetBrains specific template is maintained in a separate JetBrains.gitignore that can
+#  be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
+#  and can be added to the global gitignore or merged into this file.  For a more nuclear
+#  option (not recommended) you can uncomment the following to ignore the entire idea folder.
+#.idea/
+
+# PyPI configuration file
+.pypirc

vipsania-1.0.0/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Richard Krieg
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

vipsania-1.0.0/PKG-INFO

@@ -0,0 +1,206 @@
+Metadata-Version: 2.5
+Name: vipsania
+Version: 1.0.0
+Summary: An unsupervised deep learning ab-initio gene finder.
+Project-URL: Homepage, https://github.com/gaius-augustus/vipsania
+Project-URL: Issues, https://github.com/gaius-augustus/vipsania/issues
+Author-email: Richard Krieg <irkri@irkri.net>, Mario Stanke <mario.stanke@uni-greifswald.de>
+License-Expression: MIT
+License-File: LICENSE
+Keywords: bioinformatics,deep learning,gene finding,genome annotation,genome foundation model,hidden Markov model,masked language modelling,unsupervised learning
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Science/Research
+Classifier: Operating System :: OS Independent
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Classifier: Topic :: Scientific/Engineering :: Bio-Informatics
+Requires-Python: >=3.12
+Requires-Dist: bricks2marble[tf]
+Requires-Dist: tensorflow<2.20
+Requires-Dist: wandb
+Description-Content-Type: text/markdown
+
+# Vipsania
+
+**Vipsania is an unsupervised deep learning *ab-initio* gene finder.**
+
+Unlike other deep learning gene finders, Vipsania is never shown a reference annotation. It is
+trained purely on raw genomic sequences with a masked language modelling objective: nucleotides are
+hidden and the model learns to predict them from their context. Gene structure emerges as the
+representation a hidden Markov model inside the network needs in order to explain the sequence, so
+no curated gene set, no related-species annotation and no protein evidence are required at any
+point.
+
+For you, this means:
+
+- **Annotate any eukaryotic genome from a FASTA file alone.** No extrinsic evidence, no reference
+  annotation, no gene-structure training data of any kind.
+- **Adapt the model to your species on the fly.** Because training needs nothing but sequence, the
+  genome you want to annotate is itself valid training data. A short finetuning run on your target
+  genome directly before annotating gives the best results — see
+  [Finetuning](#finetuning-recommended).
+- **Get standard output.** Annotations are written as GenePred, GFF3 or GTF.
+
+## Installation
+
+Vipsania requires `python>=3.12`. It is recommended to install into a clean virtual environment
+(e.g. using `conda`, `pyenv` or `uv`).
+
+    $ python -m pip install vipsania
+
+This installs the `vipsania` command, which is everything you need to annotate a genome. For
+training, use the [config files](/configs) provided and follow the instructions [here](#training).
+
+Alternatively, you can clone the repository, install Vipsania from source and use the
+[scripts](/scripts) for annotation and training.
+
+    $ git clone https://github.com/gaius-augustus/vipsania
+    $ python -m pip install -e vipsania/
+
+Installing Vipsania pulls in [bricks2marble](https://github.com/gaius-augustus/bricks2marble) and
+[hidten](https://github.com/gaius-augustus/hidten), which in turn install TensorFlow with CUDA
+support. A GPU is strongly recommended; annotating a large genome on CPU only is possible, but
+slow.
+
+## Annotating a genome
+
+Annotation is done with `vipsania annotate`. It needs a trained model and a FASTA file of the
+genome you want to annotate.
+
+    $ vipsania annotate <model_id> genome.fa -o annotation.gff3 --finetune
+
+The same command is available as `python -m vipsania annotate ...`, and from a cloned repository as
+`python scripts/annotate.py ...`.
+
+The first argument names the model to use, either by the **clade** it was trained for or by its
+**model ID**, both listed below. The first time a model is used, Vipsania downloads it (about 100
+MB) and keeps it, so every later annotation with it starts immediately.
+
+    $ vipsania annotate Insecta genome.fa -o annotation.gff3 --finetune
+
+A clade name gives you the model we recommend for that clade at the moment. Should a clade get a
+better model later, `vipsania download <clade>` picks it up; annotating on its own keeps using the
+model it already has. A model ID always refers to that one model.
+
+A model you trained yourself is given in the same way: its ID is the name of its run folder, which
+Vipsania looks for in the directories around the working directory. Set `--model_dir` to use a
+different directory that holds the model folder directly.
+
+Progress is printed to the terminal while the genome is annotated, together with a log file next to
+the output.
+
+### Choosing a model
+
+Vipsania is trained per clade. Use the model of the clade your species belongs to, choosing the
+most specific one available.
+
+| Clade | Model ID | Training species | Test species | Average locus F1 | Average locus F1 (pretrained) |
+| --- | --- | ---: | ---: | ---: | ---: |
+| Alveolata | sd2zcj7u | 30 | 6 | 0.613 | 0.582 |
+| Amoebozoa | ezhpj2qm | 15 | 7 | 0.437 | 0.416 |
+| Arthropoda | ihe1jk30 | 200 | 6 | 0.207 | 0.217 |
+| Chlorophyta | faeijtmk | 26 | 6 | 0.579 | 0.558 |
+| Cnidaria | b5vtieo0 | 82 | 10 | 0.453 | 0.406 |
+| Discoba | gcra9d9y | 20 | 4 | 0.696 | 0.696 |
+| Echinodermata | sx9zjl7p | 52 | 8 | 0.444 | 0.440 |
+| Fungi | fh1kg88z | 200 | 7 | 0.660 | 0.645 |
+| Insecta | 58hsuobw | 200 | 9 | 0.537 | 0.474 |
+| Nematoda | zspca2rb | 73 | 8 | 0.539 | 0.517 |
+| Porifera | 7bjiexcu | 75 | 6 | 0.480 | 0.416 |
+| Rhodophyta | 6kmw3wme | 20 | 5 | 0.405 | 0.301 |
+| Spiralia | v5ej8oyt | 200 | 6 | 0.374 | 0.262 |
+| Stramenopiles | r6p9z9jw | 32 | 6 | 0.426 | 0.413 |
+| Streptophyta | j9m0cdmk | 200 | 13 | 0.602 | 0.579 |
+| Tunicata | hcehc7ff | 34 | 6 | 0.523 | 0.494 |
+| Vertebrata | etb1go6q | 200 | 13 | 0.406 | 0.327 |
+| | | | | | |
+| *none of the above* | cg6grhms | 33 | 7 | 0.436 | 0.416 |
+
+Use the last entry for eukaryotes that belong to none of the clades above — it was trained on
+exactly such species. It has no clade name of its own; pass `Other` or its model ID.
+
+The two F1 columns are the average locus F1 over the test species of that clade, measured against
+their reference annotations: the first with `--finetune`, the second with the pretrained model
+alone. Finetuning helps in almost every clade, which is why we recommend it.
+
+Whichever model you pick, finetuning it on your genome adapts it to your species — see
+[Finetuning](#finetuning-recommended).
+
+The species every model was trained on are listed in
+[docs/training_species.tsv](/docs/training_species.tsv), one row per species with its NCBI
+assembly accession. Since Vipsania never learns from annotations, the quality of a species'
+reference annotation, or whether it has one at all, leaves no trace in the model, and unlike for a
+supervised gene finder, finding your species or a close relative in the list is not an argument
+against that model.
+
+The test species behind the two F1 columns are in
+[docs/test_species.tsv](/docs/test_species.tsv), with precision, sensitivity and F1 at base,
+exon and locus level for each of them, once with and once without finetuning.
+
+### Where models are stored
+
+The first time a model ID is used it is downloaded into `~/.cache/vipsania/models` and read from
+there ever after. See [docs/download.md](/docs/download.md) for changing that location, and for
+fetching models in advance on machines that have no internet access when they annotate.
+
+### Output formats
+
+The output format is chosen by the file suffix of `-o/--output`:
+
+| suffix           | format                                      |
+| ---------------- | ------------------------------------------- |
+| `.gp`            | GenePred                                    |
+| `.gff3` / `.gff` | GFF3                                        |
+| `.gtf`           | GTF                                         |
+
+If `-o` is omitted, the annotation is written next to the input FASTA as `vipsania_[model_id].gff`,
+named after the model that produced it even when you asked for it by clade. The protein and coding
+sequences of the predicted genes can be written out at the same time with `--protein proteins.fa`
+and `--coding coding.fa`.
+
+### Finetuning (recommended)
+
+    $ vipsania annotate <model_id> genome.fa -o annotation.gff3 --finetune
+
+`--finetune` trains the model on the genome you are about to annotate before predicting anything.
+Since Vipsania is unsupervised, this needs nothing except the FASTA file that you already have. The
+model adapts to the codon usage, repeat landscape and intron statistics of your species, which
+consistently improves the annotation.
+
+The finetuned checkpoint is saved next to the output file, so you can reuse it for further
+annotations of the same genome or closely related species without finetuning again. Use the
+`--model_dir` argument in these cases.
+
+All options of the annotation are listed in [docs/annotate.md](/docs/annotate.md), or with
+`vipsania annotate --help`.
+
+## Training
+
+Vipsania models are trained with `vipsania train` on a set of local FASTA files. Ready-to-use
+configurations for the three model sizes are in [configs](/configs); rather than editing one, you
+list the FASTA files you want to train on in [configs/train.json](/configs/train.json) and pass it
+as an override:
+
+    $ vipsania train configs/base_10M.json -oc configs/train.json
+
+Training runs can be logged to [Weights & Biases](https://wandb.ai) with `--online entity/project`.
+Everything a run produces — configuration, checkpoints and metrics — is written to a folder below
+`./checkpoints`, named after the run. That name is exactly the model ID that `vipsania annotate`
+expects.
+
+A published model can also be trained further on species of your choice, which extends the
+`--finetune` option of `vipsania annotate` to a set of genomes. See
+[docs/training.md](/docs/training.md) for that, for the configuration format, and for what to
+change when a run does not fit into GPU memory.
+
+## Built on
+
+Vipsania is built on two libraries developed in our group:
+
+- [bricks2marble](https://github.com/gaius-augustus/bricks2marble) — efficient handling of
+  nucleotide sequences and genome annotations, plus the pre- and postprocessing around genome
+  annotation.
+- [hidten](https://github.com/gaius-augustus/hidten-docs) — hidden Markov models as differentiable,
+  highly parallel layers inside deep learning models.

vipsania-1.0.0/README.md

@@ -0,0 +1,182 @@
+# Vipsania
+
+**Vipsania is an unsupervised deep learning *ab-initio* gene finder.**
+
+Unlike other deep learning gene finders, Vipsania is never shown a reference annotation. It is
+trained purely on raw genomic sequences with a masked language modelling objective: nucleotides are
+hidden and the model learns to predict them from their context. Gene structure emerges as the
+representation a hidden Markov model inside the network needs in order to explain the sequence, so
+no curated gene set, no related-species annotation and no protein evidence are required at any
+point.
+
+For you, this means:
+
+- **Annotate any eukaryotic genome from a FASTA file alone.** No extrinsic evidence, no reference
+  annotation, no gene-structure training data of any kind.
+- **Adapt the model to your species on the fly.** Because training needs nothing but sequence, the
+  genome you want to annotate is itself valid training data. A short finetuning run on your target
+  genome directly before annotating gives the best results — see
+  [Finetuning](#finetuning-recommended).
+- **Get standard output.** Annotations are written as GenePred, GFF3 or GTF.
+
+## Installation
+
+Vipsania requires `python>=3.12`. It is recommended to install into a clean virtual environment
+(e.g. using `conda`, `pyenv` or `uv`).
+
+    $ python -m pip install vipsania
+
+This installs the `vipsania` command, which is everything you need to annotate a genome. For
+training, use the [config files](/configs) provided and follow the instructions [here](#training).
+
+Alternatively, you can clone the repository, install Vipsania from source and use the
+[scripts](/scripts) for annotation and training.
+
+    $ git clone https://github.com/gaius-augustus/vipsania
+    $ python -m pip install -e vipsania/
+
+Installing Vipsania pulls in [bricks2marble](https://github.com/gaius-augustus/bricks2marble) and
+[hidten](https://github.com/gaius-augustus/hidten), which in turn install TensorFlow with CUDA
+support. A GPU is strongly recommended; annotating a large genome on CPU only is possible, but
+slow.
+
+## Annotating a genome
+
+Annotation is done with `vipsania annotate`. It needs a trained model and a FASTA file of the
+genome you want to annotate.
+
+    $ vipsania annotate <model_id> genome.fa -o annotation.gff3 --finetune
+
+The same command is available as `python -m vipsania annotate ...`, and from a cloned repository as
+`python scripts/annotate.py ...`.
+
+The first argument names the model to use, either by the **clade** it was trained for or by its
+**model ID**, both listed below. The first time a model is used, Vipsania downloads it (about 100
+MB) and keeps it, so every later annotation with it starts immediately.
+
+    $ vipsania annotate Insecta genome.fa -o annotation.gff3 --finetune
+
+A clade name gives you the model we recommend for that clade at the moment. Should a clade get a
+better model later, `vipsania download <clade>` picks it up; annotating on its own keeps using the
+model it already has. A model ID always refers to that one model.
+
+A model you trained yourself is given in the same way: its ID is the name of its run folder, which
+Vipsania looks for in the directories around the working directory. Set `--model_dir` to use a
+different directory that holds the model folder directly.
+
+Progress is printed to the terminal while the genome is annotated, together with a log file next to
+the output.
+
+### Choosing a model
+
+Vipsania is trained per clade. Use the model of the clade your species belongs to, choosing the
+most specific one available.
+
+| Clade | Model ID | Training species | Test species | Average locus F1 | Average locus F1 (pretrained) |
+| --- | --- | ---: | ---: | ---: | ---: |
+| Alveolata | sd2zcj7u | 30 | 6 | 0.613 | 0.582 |
+| Amoebozoa | ezhpj2qm | 15 | 7 | 0.437 | 0.416 |
+| Arthropoda | ihe1jk30 | 200 | 6 | 0.207 | 0.217 |
+| Chlorophyta | faeijtmk | 26 | 6 | 0.579 | 0.558 |
+| Cnidaria | b5vtieo0 | 82 | 10 | 0.453 | 0.406 |
+| Discoba | gcra9d9y | 20 | 4 | 0.696 | 0.696 |
+| Echinodermata | sx9zjl7p | 52 | 8 | 0.444 | 0.440 |
+| Fungi | fh1kg88z | 200 | 7 | 0.660 | 0.645 |
+| Insecta | 58hsuobw | 200 | 9 | 0.537 | 0.474 |
+| Nematoda | zspca2rb | 73 | 8 | 0.539 | 0.517 |
+| Porifera | 7bjiexcu | 75 | 6 | 0.480 | 0.416 |
+| Rhodophyta | 6kmw3wme | 20 | 5 | 0.405 | 0.301 |
+| Spiralia | v5ej8oyt | 200 | 6 | 0.374 | 0.262 |
+| Stramenopiles | r6p9z9jw | 32 | 6 | 0.426 | 0.413 |
+| Streptophyta | j9m0cdmk | 200 | 13 | 0.602 | 0.579 |
+| Tunicata | hcehc7ff | 34 | 6 | 0.523 | 0.494 |
+| Vertebrata | etb1go6q | 200 | 13 | 0.406 | 0.327 |
+| | | | | | |
+| *none of the above* | cg6grhms | 33 | 7 | 0.436 | 0.416 |
+
+Use the last entry for eukaryotes that belong to none of the clades above — it was trained on
+exactly such species. It has no clade name of its own; pass `Other` or its model ID.
+
+The two F1 columns are the average locus F1 over the test species of that clade, measured against
+their reference annotations: the first with `--finetune`, the second with the pretrained model
+alone. Finetuning helps in almost every clade, which is why we recommend it.
+
+Whichever model you pick, finetuning it on your genome adapts it to your species — see
+[Finetuning](#finetuning-recommended).
+
+The species every model was trained on are listed in
+[docs/training_species.tsv](/docs/training_species.tsv), one row per species with its NCBI
+assembly accession. Since Vipsania never learns from annotations, the quality of a species'
+reference annotation, or whether it has one at all, leaves no trace in the model, and unlike for a
+supervised gene finder, finding your species or a close relative in the list is not an argument
+against that model.
+
+The test species behind the two F1 columns are in
+[docs/test_species.tsv](/docs/test_species.tsv), with precision, sensitivity and F1 at base,
+exon and locus level for each of them, once with and once without finetuning.
+
+### Where models are stored
+
+The first time a model ID is used it is downloaded into `~/.cache/vipsania/models` and read from
+there ever after. See [docs/download.md](/docs/download.md) for changing that location, and for
+fetching models in advance on machines that have no internet access when they annotate.
+
+### Output formats
+
+The output format is chosen by the file suffix of `-o/--output`:
+
+| suffix           | format                                      |
+| ---------------- | ------------------------------------------- |
+| `.gp`            | GenePred                                    |
+| `.gff3` / `.gff` | GFF3                                        |
+| `.gtf`           | GTF                                         |
+
+If `-o` is omitted, the annotation is written next to the input FASTA as `vipsania_[model_id].gff`,
+named after the model that produced it even when you asked for it by clade. The protein and coding
+sequences of the predicted genes can be written out at the same time with `--protein proteins.fa`
+and `--coding coding.fa`.
+
+### Finetuning (recommended)
+
+    $ vipsania annotate <model_id> genome.fa -o annotation.gff3 --finetune
+
+`--finetune` trains the model on the genome you are about to annotate before predicting anything.
+Since Vipsania is unsupervised, this needs nothing except the FASTA file that you already have. The
+model adapts to the codon usage, repeat landscape and intron statistics of your species, which
+consistently improves the annotation.
+
+The finetuned checkpoint is saved next to the output file, so you can reuse it for further
+annotations of the same genome or closely related species without finetuning again. Use the
+`--model_dir` argument in these cases.
+
+All options of the annotation are listed in [docs/annotate.md](/docs/annotate.md), or with
+`vipsania annotate --help`.
+
+## Training
+
+Vipsania models are trained with `vipsania train` on a set of local FASTA files. Ready-to-use
+configurations for the three model sizes are in [configs](/configs); rather than editing one, you
+list the FASTA files you want to train on in [configs/train.json](/configs/train.json) and pass it
+as an override:
+
+    $ vipsania train configs/base_10M.json -oc configs/train.json
+
+Training runs can be logged to [Weights & Biases](https://wandb.ai) with `--online entity/project`.
+Everything a run produces — configuration, checkpoints and metrics — is written to a folder below
+`./checkpoints`, named after the run. That name is exactly the model ID that `vipsania annotate`
+expects.
+
+A published model can also be trained further on species of your choice, which extends the
+`--finetune` option of `vipsania annotate` to a set of genomes. See
+[docs/training.md](/docs/training.md) for that, for the configuration format, and for what to
+change when a run does not fit into GPU memory.
+
+## Built on
+
+Vipsania is built on two libraries developed in our group:
+
+- [bricks2marble](https://github.com/gaius-augustus/bricks2marble) — efficient handling of
+  nucleotide sequences and genome annotations, plus the pre- and postprocessing around genome
+  annotation.
+- [hidten](https://github.com/gaius-augustus/hidten-docs) — hidden Markov models as differentiable,
+  highly parallel layers inside deep learning models.

vipsania-1.0.0/configs/base_10M.json

@@ -0,0 +1,107 @@
+{
+    "model": {
+        "n_layers": 10,
+        "d_hidden": 256,
+        "stripes": [["lru"], ["ffn"], ["hmm"]],
+        "restrict_stripes": {"2": [4]},
+
+        "spliced_loss": {
+            "source": ["site", "site", "outer", "inner"],
+            "weight": [0.05, 0.0, 0.0, 0.0],
+            "adaptk": [[0.01, 100], null, null, null],
+            "target": ["position", "position", "position", "position"]
+        },
+
+        "input_repeat_masked": true,
+
+        "lru": {
+            "embed_norm": "layer",
+            "readout": true,
+            "readout_norm": "layer",
+            "pre_readout_activation": "swish",
+
+            "d_latent": 256,
+            "init_radius": [[0.0, 1.0], [0.0, 1.0], [0.0, 1.0]],
+            "init_phase": [[1.047, 2.094], [3.142, 4.189], [5.236, 6.283]],
+            "init_mixing_ratio": [0.34, 0.33, 0.33],
+            "initial_state_memory": 0.99,
+            "max_tree_depth": 18
+        },
+        "hmm": {
+            "embed": 128,
+            "embed_norm": "layer",
+            "embed_activation": "softmax",
+            "readout": true,
+            "readout_type": "conv",
+            "readout_conv_kernel": 9,
+
+            "intron_state_chain": 2,
+            "intron_chain_starts": true,
+            "intron_chain_skips": true,
+            "intron_chain_loop": false,
+            "intron_chain_transitions": false,
+
+            "parallel_factor": 125,
+            "use_reverse_strand": true,
+            "emitter_share_noncoding": true,
+            "emitter_share_frames": true,
+            "train_emitter": true,
+            "initial_exon_len": 1000,
+            "initial_intron_len": [100, 10000],
+            "initial_ir_len": 10000,
+            "transitioner_share_noncoding": false,
+            "transitioner_share_frames": true,
+            "train_transitions": true,
+            "train_start_dist": true
+        },
+        "ffn": {
+            "embed_norm": "layer",
+
+            "type": "glu",
+            "units": 512,
+            "activation_hidden": "swish"
+        }
+    },
+
+    "dataset": {
+        "T": 20000,
+        "B": 8,
+
+        "masking": 0.05,
+        "token": 1.0,
+        "same": 0,
+        "false": 0,
+
+        "repeats_loss_weight": 1e-4,
+        "drop_N_threshold": 0.001,
+        "drop_repeats_threshold": 0,
+
+        "train_paths": []
+    },
+
+    "trainer": {
+        "epochs": 1000,
+        "train_steps": 800,
+
+        "start_lr": 1e-7,
+        "warmup_final_lr": 1e-3,
+        "lr": 5e-4,
+        "warmup_steps": 10,
+        "decay_steps": 90,
+
+        "no_weight_decay": [
+            "bias", "beta", "gamma", "annotation", "nu_log", "theta_log", "B_re", "B_im"
+        ],
+        "gradient_clip_norm": 1.0,
+        "gradient_accumulation_steps": 8,
+
+        "hyperparameter_schedule": [{
+            "parameter": "spliced_loss_adapter.loss_weights",
+            "list_index": 0,
+            "start": 0.0,
+            "idle": 100,
+            "target": 0.05,
+            "warmup": 5
+        }]
+    }
+}