synference 0.1.0__tar.gz → 0.1.2__tar.gz

{synference-0.1.0 → synference-0.1.2}/.github/workflows/docs.yml

@@ -42,17 +42,18 @@ jobs:
           python -m pip install --upgrade pip
           pip install wheel
           sudo apt install pandoc
+          pip install "cosmos-synthesizer@git+https://github.com/synthesizer-project/synthesizer.git"
           pip install .[test,docs]
-          pip install "dense_basis@git+https://github.com/kartheikiyer/dense_basis"
+          pip install "dense_basis@git+https://github.com/kartheikiyer/dense_basis.git"
           pip install "ltu_ili@git+https://github.com/maho3/ltu-ili.git"
+          pip install "cosmos-synthesizer@git+https://github.com/synthesizer-project/synthesizer.git#egg=synthesizer"
           pip install pytest-xdist  # enable parallel pytest execution
-       - name: Download test data
+      - name: Download test data
         run: |
           # Download test grid data 
           mkdir -p data/libraries/
           synthesizer-download --test-grids --dust-grid
-          synference-download --test
-          cd ../
+          python -c "from synference.utils import download_test_data; download_test_data()"
       - name: Sphinx Build
         run: |
           # Test sphinx build (runs all notebooks)

{synference-0.1.0 → synference-0.1.2}/.github/workflows/python-app.yml

@@ -31,9 +31,11 @@ jobs:
     - name: Install dependencies
       run: |
         python -m pip install --upgrade pip
-        WITH_OPENMP=1 pip install .
+        pip install .
         pip install ruff pytest
         pip install "ltu_ili@git+https://github.com/maho3/ltu-ili.git"
+        pip install "cosmos-synthesizer@git+https://github.com/synthesizer-project/synthesizer.git#egg=synthesizer"
+        python -c "from synference.utils import download_test_data; download_test_data()"
     
     - name: Test import
       run: |

{synference-0.1.0 → synference-0.1.2}/.gitignore

@@ -4,7 +4,6 @@ docs/source/_autosummary
 docs/source/auto_examples/
 docs/source/gen_modules/
 docs/source/sg_execution_times.rst
-
 # Python-generated files
 __pycache__/
 *.py[oc]
@@ -50,11 +49,13 @@ uv.lock
 
 __marimo__
 layouts
+docs/source/advanced_topics/config.yaml
 
 tests/test_output/*
 tests/test_grids/*
 !tests/test_libraries/sbi_test_library.hdf5
 !tests/test_libraries/sps_test_library.hdf5
 !docs/build/gfx
-!docs/source/example_models/bpass_db_v4/*
-!docs/source/example_libraries/*
+!docs/source/example_models/BPASS_DB_v4/*
+!docs/source/example_libraries/*
+!docs/source/gfx/*.png

{synference-0.1.0 → synference-0.1.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: synference
-Version: 0.1.0
+Version: 0.1.2
 Summary: Simulation-based inference for SED-fitting of galaxy photometry and spectroscopy
 Project-URL: Homepage, https://github.com/synthesizer-project/synference
 Project-URL: Bug Reports, https://github.com/synthesizer-project/synference/issues
@@ -707,6 +707,7 @@ Requires-Dist: sbi>=0.22.0
 Requires-Dist: scikit-learn
 Requires-Dist: scipy
 Requires-Dist: simple-parsing
+Requires-Dist: tarp
 Requires-Dist: torch
 Requires-Dist: tqdm
 Requires-Dist: unyt
@@ -735,9 +736,7 @@ Requires-Dist: ruff==0.11.7; extra == 'test'
 Description-Content-Type: text/markdown
 
 # Synference
-<center>
-<img src="docs/source/gfx/synference_logo.png"  alt="logo" align="right" width="200px"/>
-</center>
+<img src="https://raw.githubusercontent.com/synthesizer-project/synference/main/docs/source/gfx/synference_logo.png" align="right" width="140px"/>
 
 
 [![workflow](https://github.com/synthesizer-project/synference/actions/workflows/python-app.yml/badge.svg)](https://github.com/synthesizer-project/synference/actions)
@@ -745,23 +744,21 @@ Description-Content-Type: text/markdown
 [![Ruff](https://img.shields.io/endpoint?url=https://raw.githubusercontent.com/astral-sh/ruff/main/assets/badge/v2.json)](https://github.com/astral-sh/ruff)
 [![Contributions welcome](https://img.shields.io/badge/contributions-welcome-brightgreen.svg?style=flat)](https://github.com/synthesizer-project/synference/blob/main/docs/CONTRIBUTING.md)
 [![License: GPLv3](https://img.shields.io/badge/License-GPLv3-blue.svg)](https://www.gnu.org/licenses/gpl-3.0)
-
-<!--[![Documentation Status](https://github.com/synthesizer-project/synference/actions/workflows/static.yml/badge.svg)](https://synthesizer-project.github.io/synference/)-->
+[![Documentation Status](https://github.com/synthesizer-project/synference/actions/workflows/docs.yml/badge.svg)](https://synthesizer-project.github.io/synference/)
 
 ### Overview
 
-Synference is a Python package designed to fit to perform simulation-based inference (SBI, also known as likelihood free inference) SED fitting. It integrates with [Synthesizer](https://synthesizer-project.github.io) for flexible and fast generation of mock spectra and photometry,
-and uses the [LtU-ILI](https://ltu-ili.readthedocs.io/) package for fast, amortised posterior inference.
+Synference is a Python package designed to fit to perform simulation-based inference (SBI, also known as likelihood free inference) SED fitting. It integrates with [Synthesizer](https://synthesizer-project.github.io) for flexible and fast generation of mock spectra and photometry, and uses the [LtU-ILI](https://ltu-ili.readthedocs.io/) package for fast, amortised posterior inference.
 
 ### Key Features
 
 - **Flexible Mock Generation**: Generate mock spectra and photometry using Synthesizer, allowing for full flexibility in almost every aspect of SED generation, including a wide range of post-processed stellar population synthesis grids.
-- **Flexible Training**: Mock photometry creation is seperated from the training of the inference model, allowing for flexible training strategies and the use of different inference models, as well as quickly switching between different feature sets - e.g. different filtersets, different noise models, etc.
+- **Flexible Training**: Mock photometry creation is separated from the training of the inference model, allowing for flexible training strategies and the use of different inference models, as well as quickly switching between different feature sets - e.g. different filtersets, different noise models, etc.
 - **Fast Inference**: Leverage LtU-ILI for fast, amortised posterior inference, enabling efficient fitting of complex models to data.
 
 ### Requirements
 
-synference requires Python 3.10 or higher. It has the following dependencies:
+Synference requires Python 3.10 or higher. It has the following dependencies:
 
 - [synthesizer](https://synthesizer-project.github.io) for mock generation
 - [ltu-ili](https://ltu-ili.readthedocs.io/) for inference
@@ -777,11 +774,17 @@ synference requires Python 3.10 or higher. It has the following dependencies:
 - [tqdm](https://tqdm.github.io/) for progress bars
 - [jax](https://jax.readthedocs.io/) for GPU acceleration (optional, for some inference models)
 
-These dependencies will be automatically installed when you install synference using pip.
+These dependencies will be automatically installed when you install Synference using pip.
 
 ### Installation
 
-The easiest way to currently install synference is to clone the repository and install it in editable mode:
+The easiest way to currently install Synference is from the PyPI repository.
+
+```bash
+pip install synference
+```
+
+If you want the latest development version you can also clone the repository and install it in editable mode:
 
 ```bash
 git clone https://www.github.com/synthesizer-project/synference.git
@@ -789,9 +792,11 @@ cd synference
 pip install -e .
 ```
 
+Note that due to PyPI restrictions you will have to install [LtU-ILI](https://ltu-ili.readthedocs.io/en/latest/install.html) separately, as it can't be included as a dependency.
+
 ### Getting Started
 
-To get started with synference, you can check out the [examples](examples/) directory for usage examples and tutorials. The examples cover various aspects of using synference, including:
+To get started with Synference, you can check out the [examples](examples/) directory for usage examples and tutorials. The examples cover various aspects of using Synference, including:
 
 - Generating mock spectra and photometry with Synthesizer
 - Training inference models with LtU-ILI
@@ -802,7 +807,7 @@ To get started with synference, you can check out the [examples](examples/) dire
 
 The most basic usage, for creating a simple mock catalogue and training a model on it looks like this:
 
-Firstly we setup the Synthesizer based model. More details on how to set up the Synthesizer model can be found in the [Synthesizer documentation](https://synthesizer-project.github.io/). Here we use a BPASS SPS grid, a lognormal star formation history, a single stellar metallicity and a simple emission model including Cloudy nebular emission but no dust reprocessing. The photometric filters used are common JWST/NIRCam wideband filters, but any filters supported by [SVO](https://svo2.cab.inta-csic.es/theory/fps/index.php) or loaded manually can be used. The model parameters are drawn from a Latin hypercube sampling of the parameter space, but this can be done in any way indepedent of synference. All we are providing to the grid generation is a set of *10,000* galaxies with a range of stellar masses, redshifts, metallicities, and star formation histories, and these can be created in any way you like.
+Firstly we setup the Synthesizer based model. More details on how to set up the Synthesizer model can be found in the [Synthesizer documentation](https://synthesizer-project.github.io/). Here we use a BPASS SPS grid, a lognormal star formation history, a single stellar metallicity and a simple emission model including Cloudy nebular emission but no dust reprocessing. The photometric filters used are common JWST/NIRCam wideband filters, but any filters supported by [SVO](https://svo2.cab.inta-csic.es/theory/fps/index.php) or loaded manually can be used. The model parameters are drawn from a Latin hypercube sampling of the parameter space, but this can be done in any way independent of Synference. All we are providing to the grid generation is a set of *10,000* galaxies with a range of stellar masses, redshifts, metallicities, and star formation histories, and these can be created in any way you like.
 
 ```python
 from synthesizer.grid import Grid
@@ -894,17 +899,15 @@ empirical_model_fitter.recover_SED(observed_data_vector)
 
 ```
 
-This is just a basic example to get you started. synference is highly flexible and can be adapted to a wide range of use cases in simulation-based inference for SED fitting.
+This is just a basic example to get you started. Synference is highly flexible and can be adapted to a wide range of use cases in simulation-based inference for SED fitting.
 
 ### Documentation
 
-Work in progress.
+Documentation for Synference is available at [synthesizer-project.github.io/synference](https://synthesizer-project.github.io/synference/). The documentation includes installation instructions, tutorials, API references, and examples to help you get started with using Synference for your own projects.
 
 ### Contributing
 
-We welcome contributions to synference! If you have suggestions, bug reports, or would like to contribute code, please open an issue or submit a pull request on GitHub. Please see our [Code of Conduct](CODE_OF_CONDUCT.md) for more details on how to contribute and interact with the community.
+We welcome contributions to Synference! If you have suggestions, bug reports, or would like to contribute code, please open an issue or submit a pull request on GitHub. Please see our [Code of Conduct](CODE_OF_CONDUCT.md) for more details on how to contribute and interact with the community.
 
 ### License
-This project is licensed under the GNU General Public License v3.0 (GPLv3). See the [LICENSE](LICENSE) file for details. This means you can use, modify, and distribute the code freely, but any derivative works must also be open source and distributed under the same license. We warn users that this software is offered "as is", without any warranty or guarantee of fitness for a particular purpose. synference is under active development, and therefore may change in the future.
-
-
+This project is licensed under the GNU General Public License v3.0 (GPLv3). See the [LICENSE](LICENSE) file for details. This means you can use, modify, and distribute the code freely, but any derivative works must also be open source and distributed under the same license. We warn users that this software is offered "as is", without any warranty or guarantee of fitness for a particular purpose. Synference is under active development, and therefore may change in the future.

{synference-0.1.0 → synference-0.1.2}/README.md

@@ -1,7 +1,5 @@
 # Synference
-<center>
-<img src="docs/source/gfx/synference_logo.png"  alt="logo" align="right" width="200px"/>
-</center>
+<img src="https://raw.githubusercontent.com/synthesizer-project/synference/main/docs/source/gfx/synference_logo.png" align="right" width="140px"/>
 
 
 [![workflow](https://github.com/synthesizer-project/synference/actions/workflows/python-app.yml/badge.svg)](https://github.com/synthesizer-project/synference/actions)
@@ -9,23 +7,21 @@
 [![Ruff](https://img.shields.io/endpoint?url=https://raw.githubusercontent.com/astral-sh/ruff/main/assets/badge/v2.json)](https://github.com/astral-sh/ruff)
 [![Contributions welcome](https://img.shields.io/badge/contributions-welcome-brightgreen.svg?style=flat)](https://github.com/synthesizer-project/synference/blob/main/docs/CONTRIBUTING.md)
 [![License: GPLv3](https://img.shields.io/badge/License-GPLv3-blue.svg)](https://www.gnu.org/licenses/gpl-3.0)
-
-<!--[![Documentation Status](https://github.com/synthesizer-project/synference/actions/workflows/static.yml/badge.svg)](https://synthesizer-project.github.io/synference/)-->
+[![Documentation Status](https://github.com/synthesizer-project/synference/actions/workflows/docs.yml/badge.svg)](https://synthesizer-project.github.io/synference/)
 
 ### Overview
 
-Synference is a Python package designed to fit to perform simulation-based inference (SBI, also known as likelihood free inference) SED fitting. It integrates with [Synthesizer](https://synthesizer-project.github.io) for flexible and fast generation of mock spectra and photometry,
-and uses the [LtU-ILI](https://ltu-ili.readthedocs.io/) package for fast, amortised posterior inference.
+Synference is a Python package designed to fit to perform simulation-based inference (SBI, also known as likelihood free inference) SED fitting. It integrates with [Synthesizer](https://synthesizer-project.github.io) for flexible and fast generation of mock spectra and photometry, and uses the [LtU-ILI](https://ltu-ili.readthedocs.io/) package for fast, amortised posterior inference.
 
 ### Key Features
 
 - **Flexible Mock Generation**: Generate mock spectra and photometry using Synthesizer, allowing for full flexibility in almost every aspect of SED generation, including a wide range of post-processed stellar population synthesis grids.
-- **Flexible Training**: Mock photometry creation is seperated from the training of the inference model, allowing for flexible training strategies and the use of different inference models, as well as quickly switching between different feature sets - e.g. different filtersets, different noise models, etc.
+- **Flexible Training**: Mock photometry creation is separated from the training of the inference model, allowing for flexible training strategies and the use of different inference models, as well as quickly switching between different feature sets - e.g. different filtersets, different noise models, etc.
 - **Fast Inference**: Leverage LtU-ILI for fast, amortised posterior inference, enabling efficient fitting of complex models to data.
 
 ### Requirements
 
-synference requires Python 3.10 or higher. It has the following dependencies:
+Synference requires Python 3.10 or higher. It has the following dependencies:
 
 - [synthesizer](https://synthesizer-project.github.io) for mock generation
 - [ltu-ili](https://ltu-ili.readthedocs.io/) for inference
@@ -41,11 +37,17 @@ synference requires Python 3.10 or higher. It has the following dependencies:
 - [tqdm](https://tqdm.github.io/) for progress bars
 - [jax](https://jax.readthedocs.io/) for GPU acceleration (optional, for some inference models)
 
-These dependencies will be automatically installed when you install synference using pip.
+These dependencies will be automatically installed when you install Synference using pip.
 
 ### Installation
 
-The easiest way to currently install synference is to clone the repository and install it in editable mode:
+The easiest way to currently install Synference is from the PyPI repository.
+
+```bash
+pip install synference
+```
+
+If you want the latest development version you can also clone the repository and install it in editable mode:
 
 ```bash
 git clone https://www.github.com/synthesizer-project/synference.git
@@ -53,9 +55,11 @@ cd synference
 pip install -e .
 ```
 
+Note that due to PyPI restrictions you will have to install [LtU-ILI](https://ltu-ili.readthedocs.io/en/latest/install.html) separately, as it can't be included as a dependency.
+
 ### Getting Started
 
-To get started with synference, you can check out the [examples](examples/) directory for usage examples and tutorials. The examples cover various aspects of using synference, including:
+To get started with Synference, you can check out the [examples](examples/) directory for usage examples and tutorials. The examples cover various aspects of using Synference, including:
 
 - Generating mock spectra and photometry with Synthesizer
 - Training inference models with LtU-ILI
@@ -66,7 +70,7 @@ To get started with synference, you can check out the [examples](examples/) dire
 
 The most basic usage, for creating a simple mock catalogue and training a model on it looks like this:
 
-Firstly we setup the Synthesizer based model. More details on how to set up the Synthesizer model can be found in the [Synthesizer documentation](https://synthesizer-project.github.io/). Here we use a BPASS SPS grid, a lognormal star formation history, a single stellar metallicity and a simple emission model including Cloudy nebular emission but no dust reprocessing. The photometric filters used are common JWST/NIRCam wideband filters, but any filters supported by [SVO](https://svo2.cab.inta-csic.es/theory/fps/index.php) or loaded manually can be used. The model parameters are drawn from a Latin hypercube sampling of the parameter space, but this can be done in any way indepedent of synference. All we are providing to the grid generation is a set of *10,000* galaxies with a range of stellar masses, redshifts, metallicities, and star formation histories, and these can be created in any way you like.
+Firstly we setup the Synthesizer based model. More details on how to set up the Synthesizer model can be found in the [Synthesizer documentation](https://synthesizer-project.github.io/). Here we use a BPASS SPS grid, a lognormal star formation history, a single stellar metallicity and a simple emission model including Cloudy nebular emission but no dust reprocessing. The photometric filters used are common JWST/NIRCam wideband filters, but any filters supported by [SVO](https://svo2.cab.inta-csic.es/theory/fps/index.php) or loaded manually can be used. The model parameters are drawn from a Latin hypercube sampling of the parameter space, but this can be done in any way independent of Synference. All we are providing to the grid generation is a set of *10,000* galaxies with a range of stellar masses, redshifts, metallicities, and star formation histories, and these can be created in any way you like.
 
 ```python
 from synthesizer.grid import Grid
@@ -158,17 +162,15 @@ empirical_model_fitter.recover_SED(observed_data_vector)
 
 ```
 
-This is just a basic example to get you started. synference is highly flexible and can be adapted to a wide range of use cases in simulation-based inference for SED fitting.
+This is just a basic example to get you started. Synference is highly flexible and can be adapted to a wide range of use cases in simulation-based inference for SED fitting.
 
 ### Documentation
 
-Work in progress.
+Documentation for Synference is available at [synthesizer-project.github.io/synference](https://synthesizer-project.github.io/synference/). The documentation includes installation instructions, tutorials, API references, and examples to help you get started with using Synference for your own projects.
 
 ### Contributing
 
-We welcome contributions to synference! If you have suggestions, bug reports, or would like to contribute code, please open an issue or submit a pull request on GitHub. Please see our [Code of Conduct](CODE_OF_CONDUCT.md) for more details on how to contribute and interact with the community.
+We welcome contributions to Synference! If you have suggestions, bug reports, or would like to contribute code, please open an issue or submit a pull request on GitHub. Please see our [Code of Conduct](CODE_OF_CONDUCT.md) for more details on how to contribute and interact with the community.
 
 ### License
-This project is licensed under the GNU General Public License v3.0 (GPLv3). See the [LICENSE](LICENSE) file for details. This means you can use, modify, and distribute the code freely, but any derivative works must also be open source and distributed under the same license. We warn users that this software is offered "as is", without any warranty or guarantee of fitness for a particular purpose. synference is under active development, and therefore may change in the future.
-
-
+This project is licensed under the GNU General Public License v3.0 (GPLv3). See the [LICENSE](LICENSE) file for details. This means you can use, modify, and distribute the code freely, but any derivative works must also be open source and distributed under the same license. We warn users that this software is offered "as is", without any warranty or guarantee of fitness for a particular purpose. Synference is under active development, and therefore may change in the future.

{synference-0.1.0 → synference-0.1.2}/docs/CONTRIBUTING.md

@@ -94,7 +94,7 @@ This will build a local copy of the documentation representative of the currentl
 
 ## Contributing to the Documentation
 
-The synference documentation is written in a combination of reStructuredText, Jupyter notebooks and Python scripts.
+The Synference documentation is written in a combination of reStructuredText, Jupyter notebooks and Python scripts.
 Adding content should be relatively simple if you follow the instructions below.
 
 ### Adding notebooks

{synference-0.1.0 → synference-0.1.2}/docs/source/FAQ/FAQ.rst

@@ -22,9 +22,9 @@ How do I deal with low sampling acceptance?
 -------------------------------------------
 
 A: You will be warned if your sampling acceptance is low during inference. Low sampling acceptance means that the model is predicting posterior samples which are outside the prior proposal range. To deal with this, you can try the following:
-- Increase the prior proposal range to ensure that the true parameters are within the prior support. For some parameters this can be reasonable, for example setting a slightly wider range for stellar mass.
-- Train a better model - low sampling acceptance can indicate that the model is not accurately capturing the posterior distribution. You can try training a more complex model, or using more training data to improve the model's performance.
-- You can check if a specific parameter is causing low sampling acceptance by looking at the acceptance per parameter during inference using the custom torch prior implemented in Synference. You can enable this by running `SBI_Fitter.create_priors(debug_sample_acceptance=True)` before inference. This will log the acceptance rate for each parameter, allowing you to identify any parameters that may be causing issues.
+* Increase the prior proposal range to ensure that the true parameters are within the prior support. For some parameters this can be reasonable, for example setting a slightly wider range for stellar mass.
+* Train a better model - low sampling acceptance can indicate that the model is not accurately capturing the posterior distribution. You can try training a more complex model, or using more training data to improve the model's performance.
+* You can check if a specific parameter is causing low sampling acceptance by looking at the acceptance per parameter during inference using the custom torch prior implemented in Synference. You can enable this by running `SBI_Fitter.create_priors(debug_sample_acceptance=True)` before inference. This will log the acceptance rate for each parameter, allowing you to identify any parameters that may be causing issues.
 
 How do I load a previously trained model for inference?
 -------------------------------------------------------
@@ -40,4 +40,22 @@ A: You can load a previously trained model using the `SBI_Fitter.load_saved_mode
 
 This will load the model from the specified path and move it to the CPU for inference. You can then use this model to perform inference on new data. The model can be trained on GPU and then loaded on CPU for inference as shown above.
 
+What other SBI tools are available for SED fitting and more generally for galaxy inference?
+--------------------------------------------------------------------------------------------
+
+There are a number of great tools out there for specific applications of SBI. This is not an exhaustive list (and feel free to suggest more), but below we've collated some of the useful tools we've found. Please cite the relevant papers if you use these tools. 
+
+1. `SBIPIX <https://github.com/patriglesias/SBIPIX/>`_ - Published
+   `here <https://ui.adsabs.harvard.edu/abs/2025arXiv250604336I/abstract>`_.
+   Specifically for pixel-level inference from galaxy images, this
+   package provides pretrained models as well as the ability to train
+   your own models.
+2. `SBI++ <https://github.com/wangbingjie/sbi_pp>`_ - Published
+   `here <https://ui.adsabs.harvard.edu/abs/2023ApJ...952L..10W/abstract>`_,
+   this tool demonstrates a SBI workflow for SED fitting which can deal
+   with out of distribution noise and missing data.
+3. `SBI-SFHs <https://github.com/patriglesias/SBI_SFHs>`_ - Published
+   `here <https://www.aanda.org/articles/aa/abs/2024/09/aa49909-24/aa49909-24.html>`_
+   this tool recovers galaxy SFHS from spectroscopy.
+
 

{synference-0.1.0 → synference-0.1.2}/docs/source/advanced_topics/advanced_topics.rst

@@ -6,6 +6,5 @@ In this section, we cover some advanced topics and features of Synference that g
 .. toctree::
    :maxdepth: 1
 
-   nle
    custom_loop
    simformer

{synference-0.1.0 → synference-0.1.2}/docs/source/advanced_topics/custom_loop.ipynb

@@ -134,19 +134,138 @@
    "id": "6261165f",
    "metadata": {},
    "source": [
-    "We recommend the Optuna dashboard for monitoring the progress of hyperparameter optimization. This can be launched using the following command:\n",
+    "We recommend [Optuna Dashboard](https://optuna-dashboard.readthedocs.io/en/latest/getting-started.html) for monitoring the progress of hyperparameter optimization. This can be launched using the following command:\n",
     "\n",
     "```bash\n",
     "optuna-dashboard your_sql_link\n",
     "```"
    ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "cc76e813",
+   "metadata": {},
+   "source": [
+    "## Custom Loop with a Fixed Model\n",
+    "\n",
+    "You can also use the custom training loop to train models with fixed parameter values, with a different configuration. This lets you take advantage of the additional features such as the customizable optimizer, model training checkpoints and ongoing training monitoring.\n",
+    "\n",
+    "Here is an example of the yaml file to train a specific fixed model, in this case a Neural Spline Flow, using the Adam optimizer, with 30 hidden features and 14 transforms."
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "cb63e847",
+   "metadata": {},
+   "source": [
+    "```yaml\n",
+    "train_args:\n",
+    "  skip_optimization: True\n",
+    "  validation_fraction: 0.1\n",
+    "  fixed_params:\n",
+    "    model_choice: \"nsf\" # Must be a list\n",
+    "    optimizer_choice: \"Adam\" # Must be a list\n",
+    "    learning_rate: 0.0007460108070908076\n",
+    "    training_batch_size: 79\n",
+    "    stop_after_epochs: 57\n",
+    "    clip_max_norm: 6.656577606872957\n",
+    "    nsf_hidden_features: 30\n",
+    "    nsf_num_transforms: 14\n",
+    "```"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "be69d999",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "config = {\n",
+    "    \"train_args\": {\n",
+    "        \"skip_optimization\": True,\n",
+    "        \"validation_fraction\": 0.1,\n",
+    "        \"fixed_params\": {\n",
+    "            \"model_choice\": \"nsf\",\n",
+    "            \"optimizer_choice\": \"Adam\",\n",
+    "            \"learning_rate\": 0.0007460108070908076,\n",
+    "            \"training_batch_size\": 79,\n",
+    "            \"stop_after_epochs\": 57,\n",
+    "            \"clip_max_norm\": 6.656577606872957,\n",
+    "            \"nsf_hidden_features\": 30,\n",
+    "            \"nsf_num_transforms\": 14,\n",
+    "        },\n",
+    "    }\n",
+    "}\n",
+    "\n",
+    "import yaml\n",
+    "\n",
+    "with open(\"config.yaml\", \"w\") as f:\n",
+    "    yaml.dump(config, f)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "29780005",
+   "metadata": {},
+   "source": [
+    "First we initialize a pre-configured model library."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "0ca21ae1",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from synference import SBI_Fitter, test_data_dir\n",
+    "\n",
+    "fitter = SBI_Fitter.init_from_hdf5(\n",
+    "    model_name=\"test\", hdf5_path=f\"{test_data_dir}/example_model_library.hdf5\"\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "00dae994",
+   "metadata": {},
+   "source": [
+    "Then we create our training arrays."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "89b72d4c",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "fitter.create_feature_array();"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "429f7303",
+   "metadata": {},
+   "source": [
+    "Now we can train a model - note that this is a terrible model with far too small a dataset, purely for demonstration purposes.\n",
+    "\n",
+    "The plotted live training loss curves look slightly odd in a notebook format, but you can see the live progress of your model on your training and validation set, and see where the best performing training epoch is."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "7d8c65ce",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "fitter.run_single_sbi(custom_config_yaml=\"config.yaml\")"
+   ]
   }
  ],
- "metadata": {
-  "language_info": {
-   "name": "python"
-  }
- },
+ "metadata": {},
  "nbformat": 4,
  "nbformat_minor": 5
 }

synference-0.1.2/docs/source/advanced_topics/simformer.ipynb

@@ -0,0 +1,109 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "id": "2291bddd",
+   "metadata": {},
+   "source": [
+    "### The Simformer\n",
+    "\n",
+    "[Gloeckler et al. 2024](https://arxiv.org/abs/2404.09636) introduced the Simformer, for 'all in one simulation based inference'.\n",
+    "\n",
+    "They use a novel probablistic diffusion model with a transformer architecture which learns the full joint distribution of parameters and data, allowing for fast, amortized Bayesian inference, without specifying beforehand which parameters are of interest. This makes the simformer approach particularly well suited to missing data, as the use of an attention mechanism allows the model sample from an arbitrary conditional distribution excluding any missing data."
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "12ddb0a9",
+   "metadata": {},
+   "source": [
+    "The Simformer is currently implemented in two ways. The first way, is a seperate class called ```Simformer_Fitter```, which requires the user to install a [fork of the original simformer repo](https://github.com/tHarvey303/simformer/), which requires quite specific versions of CUDA, PyTorch and jax to work. \n",
+    "\n",
+    "The second way uses the new simformer implementation in the sbi package, which is currently only available in a pull request, but should be merged into the main branch soon. For now we will deal with this implementation, as it is much easier to install and use. There are examples of using the original approach in the examples/simformer folder of the synference repo.\n",
+    "\n",
+    "There are some limitations to the current sbi simformer implementation. Currently it doesn't seem to support serialization (due to the use of lambda functions), so models cannot be saved and loaded like other synference SBI models."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "6284d69d",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from synference import SBI_Fitter, test_data_dir\n",
+    "\n",
+    "fitter = SBI_Fitter.init_from_hdf5(\n",
+    "    model_name=\"test\", hdf5_path=f\"{test_data_dir}/example_model_library.hdf5\"\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "88192510",
+   "metadata": {},
+   "source": [
+    "We will create our training array as normal using the ```SBI_Fitter``` class."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "96c2ff30",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "fitter.create_feature_array();"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "fe118ab9",
+   "metadata": {},
+   "source": [
+    "There are a few parameters to be aware of:\n",
+    "\n",
+    "1. ```sde_type``` : The type of SDE to use. Options are 've` (variance exploding), 'vp' (variance preserving) or 'subvp' (sub variance preserving). This doesn't do anything for the flow based simformer.\n",
+    "2. ```simformer_type```: 'score' or 'flow'- whether to use a score based or flow based simformer. \n",
+    "3  ```learning_rate```: The learning rate to use for training.\n",
+    "4. ```model_kwargs```: A dictionary of additional keyword arguments to pass to the simformer model. These can include:\n",
+    "    - ```num_layers```: The number of transformer layers to use.\n",
+    "    - ```num_heads```: The number of attention heads to use.\n",
+    "    - ```dim_val```: The dimension of the value vectors in the attention mechanism.\n",
+    "    - ```dim_id```: The dimension of the identity vectors in the attention mechanism.\n",
+    "    - ```mlp_ratio``` : The ratio of the hidden dimension to the input dimension in the MLP layers.\n",
+    "    - ```hidden_features```: The number of hidden features to use in the MLP layers.\n",
+    "    - ```time_embedding_dim```: The dimension of the time embedding.\n",
+    "\n",
+    "Like all other synference models we can also set the ```training_batch_size```, ```validation_fraction```, ```stop_after_epochs``` and ```clip_max_norm``` parameters.\n",
+    "\n",
+    "```python\n",
+    "\n",
+    "fitter.run_single_simformer(\n",
+    "    name_append=\"simformer_test\",\n",
+    "    sde_type=\"ve\",\n",
+    "    simformer_type=\"score\",\n",
+    "    learning_rate=1e-5,\n",
+    "    training_batch_size=64,\n",
+    "    model_kwargs={\n",
+    "        \"hidden_features\": 128,\n",
+    "        \"n_layers\": 6,\n",
+    "        \"dim_val\": 64,\n",
+    "        \"dim_id\": 64,\n",
+    "        \"mlp_ratio\": 4,\n",
+    "        \"time_embedding_dim\": 32,\n",
+    "        \"num_heads\": 4,\n",
+    "    },\n",
+    "    load_existing_model=False,\n",
+    "    validation_fraction=0.1,\n",
+    "    stop_after_epochs=30,\n",
+    "    plot=False,  # Currently the LtU-ILI plotting doesn't work with the simformer\n",
+    ")\n",
+    "```"
+   ]
+  }
+ ],
+ "metadata": {},
+ "nbformat": 4,
+ "nbformat_minor": 5
+}