Automatically download model weights

.github/workflows/tests.yml

@@ -14,7 +14,7 @@ jobs:
     runs-on: ${{ matrix.os }}
     strategy:
       matrix:
-        os: [ubuntu-latest, windows-latest, macos-latest]
+        os: [ubuntu-latest, windows-latest]
 
     steps:
     - uses: actions/checkout@v2

README.md

@@ -8,8 +8,6 @@ If you use Casanovo in your work, please cite the following publication:
 
 - Yilmaz, M., Fondrie, W. E., Bittremieux, W., Oh, S. & Noble, W. S. *De novo* mass spectrometry peptide sequencing with a transformer model. in *Proceedings of the 39th International Conference on Machine Learning - ICML '22* vol. 162 25514–25522 (PMLR, 2022). [https://proceedings.mlr.press/v162/yilmaz22a.html](https://proceedings.mlr.press/v162/yilmaz22a.html)
 
-Data and pre-trained model weights are available [on Zenodo](https://zenodo.org/record/6791263).
-
 ## Documentation
 
 #### https://casanovo.readthedocs.io/en/latest/
@@ -18,7 +16,7 @@ Data and pre-trained model weights are available [on Zenodo](https://zenodo.org/
 
 We recommend to run Casanovo in a dedicated **Anaconda** environment.
 This helps keep your environment for Casanovo and its dependencies separate from your other Python environments.
-**This is especially helpful because Casanovo works within a specific range of Python versions (3.8 ≥ Python version > 3.10).**
+**This is especially helpful because Casanovo works within a specific range of Python versions (3.8 ≥ Python version ≥ 3.10).**
 
 - Check out the [Windows](https://docs.anaconda.com/anaconda/install/windows/#), [MacOS](https://docs.anaconda.com/anaconda/install/mac-os/), and [Linux](https://docs.anaconda.com/anaconda/install/linux/) installation instructions.
 
@@ -49,9 +47,9 @@ The base environment most likely will not work.
 
 ### Installation
 
-Install Casanovo as a Python package from this repository (requires 3.8 ≥ [Python version] > 3.10 , dependencies will be installed automatically as needed):
+Install Casanovo as a Python package from this repository (requires 3.8 ≥ [Python version] ≥ 3.10 , dependencies will be installed automatically as needed):
 
-```
+``` sh
 pip install casanovo
 ```
 
@@ -60,12 +58,22 @@ Once installed, Casanovo can be used with a simple command line interface.
 All auxiliary data, model, and training-related parameters can be specified in a user created `.yaml` configuration file.
 See [`casanovo/config.yaml`](https://github.com/Noble-Lab/casanovo/blob/main/casanovo/config.yaml) for the default configuration that was used to obtain the reported results.
 
+
+### Model weights
+
+When running Casanovo in `denovo` or `eval` mode, Casanovo needs compatible pretrained model weights to make predictions.
+Model weights can be found on the [Releases page](https://github.com/Noble-Lab/casanovo/releases) under the "Assets" for each release (file extension: .ckpt).
+The model file can then be specified using the `--model` command-line parameter when executing Casanovo.
+To assist users, if no model file is specified Casanovo will try to download and use a compatible model file automatically.
+
+Not all releases might have a model file included on the [Releases page](https://github.com/Noble-Lab/casanovo/releases), in which case model weights for alternative releases with the same major version number can be used.
+
 ### Example commands
 
 - To run _de novo_ sequencing:
 
 ```
-casanovo --mode=denovo --model=path/to/pretrained.ckpt --peak_path=path/to/predict/spectra.mgf --config=path/to/config.yaml --output=path/to/output
+casanovo --mode=denovo --peak_path=path/to/predict/spectra.mgf --output=path/to/output
 ```
 
 Casanovo can predict peptide sequences for MS/MS data in mzML, mzXML, and MGF files.
@@ -74,15 +82,15 @@ This will write peptide predictions for the given MS/MS spectra to the specified
 - To evaluate _de novo_ sequencing performance based on known spectrum annotations:
 
 ```
-casanovo --mode=eval --model=path/to/pretrained.ckpt --peak_path=path/to/test/annotated_spectra.mgf --config=path/to/config.yaml
+casanovo --mode=eval --peak_path=path/to/test/annotated_spectra.mgf
 ```
 
 To evaluate the peptide predictions, ground truth peptide labels need to be provided as an annotated MGF file.
 
 - To train a model from scratch:
 
 ```
-casanovo --mode=train --peak_path=path/to/train/annotated_spectra.mgf --peak_path_val=path/to/validation/annotated_spectra.mgf --config=path/to/config.yaml
+casanovo --mode=train --peak_path=path/to/train/annotated_spectra.mgf --peak_path_val=path/to/validation/annotated_spectra.mgf
 ```
 
 Training and validation MS/MS data need to be provided as annotated MGF files.
@@ -95,16 +103,13 @@ We will demonstrate how to use Casanovo using a small walkthrough example on a s
 The example MGF file is available at [`sample_data/sample_preprocessed_spectra.mgf`](https://github.com/Noble-Lab/casanovo/blob/main/sample_data/sample_preprocessed_spectra.mgf`).
 
 1. Install Casanovo (see above for details).
-2. Download the `casanovo_pretrained_model_weights.zip` from [Zenodo](https://zenodo.org/record/6791263). Place these models in a location that you can easily access and know the path of.
-    - We will be `using pretrained_excl_mouse.ckpt` for this job.
-3. Copy the example `config.yaml` file into a location you can easily access. 
-4. Ensure you are in the proper anaconda environment by typing `conda activate casanovo_env`. (If you named your environment differently, type in that name instead.)
-5. Run this command:
+2. Ensure you are in the proper anaconda environment by typing `conda activate casanovo_env`. (If you named your environment differently, type in that name instead.)
+3. Run this command:
 ```
-casanovo --mode=denovo --model=[PATH_TO]/pretrained_excl_mouse.ckpt --peak_path=[PATH_TO]/sample_preprocessed_spectra.mgf --config=[PATH_TO]/config.yaml
+casanovo --mode=denovo --peak_path=[PATH_TO]/sample_preprocessed_spectra.mgf
 ```
-Make sure you use the proper filepath to the `pretrained_excl_mouse.ckpt` file.
-    - Note: If you want to get the output mzTab file in different location than the working directory, specify an alternative output location using the `--output` parameter.
+
+Note: If you want to store the output mzTab file in a different location than the current working directory, specify an alternative output location using the `--output` parameter.
 
 This job will take very little time to run (< 1 minute).
 
@@ -127,8 +132,29 @@ Run the following command in your command prompt to see all possible command-lin
 casanovo --help
 ```
 
+Additionally, you can use a configuration file to fully customize Casanovo.
+You can find the `config.yaml` configuration file that is used by default [here](https://github.com/Noble-Lab/casanovo/blob/main/casanovo/config.yaml).
+
 **I get a "CUDA out of memory" error when trying to run Casanovo. Help!**
 
 This means that there was not enough (free) memory available on your GPU to run Casanovo, which is especially likely to happen when you are using a smaller, consumer-grade GPU.
 We recommend trying to decrease the `train_batch_size` or `predict_batch_size` options in the [config file](https://github.com/Noble-Lab/casanovo/blob/main/casanovo/config.yaml) (depending on whether the error occurred during `train` or `denovo` mode) to reduce the number of spectra that are processed simultaneously.
 Additionally, we recommend shutting down any other processes that may be running on the GPU, so that Casanovo can exclusively use the GPU.
+
+**How do I solve a "PermissionError: GitHub API rate limit exceeded" error when trying to run Casanovo?**
+
+When running Casanovo in `denovo` or `eval` mode, Casanovo needs compatible pretrained model weights to make predictions.
+If no model weights file is specified using the `--model` command-line parameter, Casanovo will automatically try to download the latest compatible model file from GitHub and save it to its cache for subsequent use.
+However, the GitHub API is limited to maximum 60 requests per hour per IP address.
+Consequently, if Casanovo has been executed multiple times already, it might temporarily not be able to communicate with GitHub.
+You can avoid this error by explicitly specifying the model file using the `--model` parameter.
+
+**I see "NotImplementedError: The operator 'aten::index.Tensor'..." when using a Mac with an Apple Silicon chip.**
+
+Casanovo can leverage Apple's Metal Performance Shaders (MPS) on newer Mac computers, which requires that the `PYTORCH_ENABLE_MPS_FALLBACK` is set to `1`:
+
+``` sh
+export PYTORCH_ENABLE_MPS_FALLBACK=1
+```
+
+This will need to be set with each new shell session, or you can add it to your `.bashrc` / `.zshrc` to set this environment variable by default.

casanovo/casanovo.py

@@ -1,16 +1,27 @@
 """The command line entry point for Casanovo."""
 import datetime
+import functools
 import logging
 import os
+import re
+import shutil
 import sys
+import warnings
+from typing import Optional, Tuple
 
+warnings.filterwarnings("ignore", category=DeprecationWarning)
+
+import appdirs
 import click
-import psutil
+import github
 import pytorch_lightning as pl
+import requests
 import torch
+import tqdm
 import yaml
 
 from . import __version__
+from . import utils
 from .data import ms_io
 from .denovo import model_runner
 
@@ -61,11 +72,11 @@
 )
 def main(
     mode: str,
-    model: str,
+    model: Optional[str],
     peak_path: str,
-    peak_path_val: str,
-    config: str,
-    output: str,
+    peak_path_val: Optional[str],
+    config: Optional[str],
+    output: Optional[str],
 ):
     """
     \b
@@ -105,10 +116,12 @@ def main(
     root.addHandler(file_handler)
     # Disable dependency non-critical log messages.
     logging.getLogger("depthcharge").setLevel(logging.INFO)
+    logging.getLogger("github").setLevel(logging.WARNING)
     logging.getLogger("h5py").setLevel(logging.WARNING)
     logging.getLogger("numba").setLevel(logging.WARNING)
     logging.getLogger("pytorch_lightning").setLevel(logging.WARNING)
     logging.getLogger("torch").setLevel(logging.WARNING)
+    logging.getLogger("urllib3").setLevel(logging.WARNING)
 
     # Read parameters from the config file.
     if config is None:
@@ -163,13 +176,30 @@ def main(
     }
     # Add extra configuration options and scale by the number of GPUs.
     n_gpus = torch.cuda.device_count()
-    config["n_workers"] = len(psutil.Process().cpu_affinity())
+    config["n_workers"] = utils.n_workers()
     if n_gpus > 1:
-        config["n_workers"] = config["n_workers"] // n_gpus
         config["train_batch_size"] = config["train_batch_size"] // n_gpus
 
     pl.utilities.seed.seed_everything(seed=config["random_seed"], workers=True)
 
+    # Download model weights if these were not specified (except when training).
+    if model is None and mode != "train":
+        try:
+            model = _get_model_weights()
+        except github.RateLimitExceededException:
+            logger.error(
+                "GitHub API rate limit exceeded while trying to download the "
+                "model weights. Please download compatible model weights "
+                "manually from the official Casanovo code website "
+                "(https://github.com/Noble-Lab/casanovo) and specify these "
+                "explicitly using the `--model` parameter when running "
+                "Casanovo."
+            )
+            raise PermissionError(
+                "GitHub API rate limit exceeded while trying to download the "
+                "model weights"
+            ) from None
+
     # Log the active configuration.
     logger.info("Casanovo version %s", str(__version__))
     logger.debug("mode = %s", mode)
@@ -198,5 +228,100 @@ def main(
         model_runner.train(peak_path, peak_path_val, model, config)
 
 
+def _get_model_weights() -> str:
+    """
+    Use cached model weights or download them from GitHub.
+
+    If no weights file (extension: .ckpt) is available in the cache directory,
+    it will be downloaded from a release asset on GitHub.
+    Model weights are retrieved by matching release version. If no model weights
+    for an identical release (major, minor, patch), alternative releases with
+    matching (i) major and minor, or (ii) major versions will be used.
+    If no matching release can be found, no model weights will be downloaded.
+
+    Note that the GitHub API is limited to 60 requests from the same IP per
+    hour.
+
+    Returns
+    -------
+    str
+        The name of the model weights file.
+    """
+    cache_dir = appdirs.user_cache_dir("casanovo", False, opinion=False)
+    os.makedirs(cache_dir, exist_ok=True)
+    version = utils.split_version(__version__)
+    version_match: Tuple[Optional[str], Optional[str], int] = None, None, 0
+    # Try to find suitable model weights in the local cache.
+    for filename in os.listdir(cache_dir):
+        root, ext = os.path.splitext(filename)
+        if ext == ".ckpt":
+            file_version = tuple(
+                g for g in re.match(r".*_v(\d+)_(\d+)_(\d+)", root).groups()
+            )
+            match = sum([i == j for i, j in zip(version, file_version)])
+            if match > version_match[2]:
+                version_match = os.path.join(cache_dir, filename), None, match
+    # Provide the cached model weights if found.
+    if version_match[2] > 0:
+        logger.info(
+            "Model weights file %s retrieved from local cache",
+            version_match[0],
+        )
+        return version_match[0]
+    # Otherwise try to find compatible model weights on GitHub.
+    else:
+        repo = github.Github().get_repo("Noble-Lab/casanovo")
+        # Find the best matching release with model weights provided as asset.
+        for release in repo.get_releases():
+            rel_version = tuple(
+                g
+                for g in re.match(
+                    r"v(\d+)\.(\d+)\.(\d+)", release.tag_name
+                ).groups()
+            )
+            match = sum([i == j for i, j in zip(version, rel_version)])
+            if match > version_match[2]:
+                for release_asset in release.get_assets():
+                    fn, ext = os.path.splitext(release_asset.name)
+                    if ext == ".ckpt":
+                        version_match = (
+                            os.path.join(
+                                cache_dir,
+                                f"{fn}_v{'_'.join(map(str, rel_version))}{ext}",
+                            ),
+                            release_asset.browser_download_url,
+                            match,
+                        )
+                        break
+        # Download the model weights if a matching release was found.
+        if version_match[2] > 0:
+            filename, url, _ = version_match
+            logger.info(
+                "Downloading model weights file %s from %s", filename, url
+            )
+            r = requests.get(url, stream=True, allow_redirects=True)
+            r.raise_for_status()
+            file_size = int(r.headers.get("Content-Length", 0))
+            desc = "(Unknown total file size)" if file_size == 0 else ""
+            r.raw.read = functools.partial(r.raw.read, decode_content=True)
+            with tqdm.tqdm.wrapattr(
+                r.raw, "read", total=file_size, desc=desc
+            ) as r_raw, open(filename, "wb") as f:
+                shutil.copyfileobj(r_raw, f)
+            return filename
+        else:
+            logger.error(
+                "No matching model weights for release v%s found, please "
+                "specify your model weights explicitly using the `--model` "
+                "parameter",
+                __version__,
+            )
+            raise ValueError(
+                f"No matching model weights for release v{__version__} found, "
+                f"please specify your model weights explicitly using the "
+                f"`--model` parameter"
+            )
+
+
 if __name__ == "__main__":
     main()

casanovo/config.yaml

@@ -8,7 +8,7 @@ random_seed: 454
 
 # Spectrum processing options.
 n_peaks: 150
-min_mz: 50.52564895  # 1.0005079 * 50.5
+min_mz: 50.0
 max_mz: 2500.0
 min_intensity: 0.01
 remove_precursor_tol: 2.0  # Da
@@ -21,7 +21,7 @@ dim_model: 512
 n_head: 8
 dim_feedforward: 1024
 n_layers: 9
-dropout: 0
+dropout: 0.0
 dim_intensity:
 custom_encoder:
 max_length: 100

casanovo/denovo/dataloaders.py

@@ -54,7 +54,7 @@ def __init__(
         test_index: Optional[AnnotatedSpectrumIndex] = None,
         batch_size: int = 128,
         n_peaks: Optional[int] = 150,
-        min_mz: float = 140.0,
+        min_mz: float = 50.0,
         max_mz: float = 2500.0,
         min_intensity: float = 0.01,
         remove_precursor_tol: float = 2.0,

casanovo/denovo/model.py

@@ -79,10 +79,10 @@ class Spec2Pep(pl.LightningModule, ModelMixin):
 
     def __init__(
         self,
-        dim_model: int = 128,
+        dim_model: int = 512,
         n_head: int = 8,
         dim_feedforward: int = 1024,
-        n_layers: int = 1,
+        n_layers: int = 9,
         dropout: float = 0.0,
         dim_intensity: Optional[int] = None,
         custom_encoder: Optional[SpectrumEncoder] = None,