diff --git a/docs/Makefile b/docs/Makefile
index 3dfdf85d46949236d6a6fcd33a897ee201828804..8c1bf9d5c429ffa198fed21a652998f6f5f528e5 100644
--- a/docs/Makefile
+++ b/docs/Makefile
@@ -56,7 +56,6 @@ generate:
mhcflurry-downloads fetch models_class1 cross_validation_class1
python generate.py \
--out-models-cv-rst _build/_models_cv.rst \
- --out-models-architecture-png _build/_models_architecture.png \
--out-models-info-rst _build/_models_info.rst \
--out-models-supported-alleles-rst _build/_models_supported_alleles.rst
diff --git a/docs/commandline_tutorial.rst b/docs/commandline_tutorial.rst
index 6dc803c81b40657c2be83f8d5106a46654fb1c73..ade306d6f4a781df3080cad53547a5e6eee37f29 100644
--- a/docs/commandline_tutorial.rst
+++ b/docs/commandline_tutorial.rst
@@ -111,7 +111,7 @@ training data. The data we use for our released predictors can be downloaded wit
It looks like this:
.. command-output::
- bzcat "$(mhcflurry-downloads path data_curated)/curated_training_data.csv.bz2" | head -n 3
+ bzcat "$(mhcflurry-downloads path data_curated)/curated_training_data.no_mass_spec.csv.bz2" | head -n 3
:shell:
:nostderr:
diff --git a/docs/conf.py b/docs/conf.py
index be1d9ae6aa14dab0664790112c8df8957fdd7682..98df4244ca721115c27b62f72aea8c2bf9f6613a 100644
--- a/docs/conf.py
+++ b/docs/conf.py
@@ -86,7 +86,7 @@ author = 'Timothy O\'Donnell'
# The short X.Y version.
# Added by Tim: reading version from mhcflurry __init__.py as in setup.py
-with open('../mhcflurry/__init__.py', 'r') as f:
+with open('../mhcflurry/version.py', 'r') as f:
version = re.search(
r'^__version__\s*=\s*[\'"]([^\'"]*)[\'"]',
f.read(),
diff --git a/docs/generate.py b/docs/generate.py
index 848d04ce5a15c0d8391a4c1bd81e90b4a5229939..fb39ade21ba133c8775cf6a6d5c06bbdd9db3b60 100644
--- a/docs/generate.py
+++ b/docs/generate.py
@@ -4,11 +4,14 @@ Generate certain RST files used in documentation.
import sys
import argparse
+import json
from textwrap import wrap
+from collections import OrderedDict
import pypandoc
import pandas
from keras.utils.vis_utils import plot_model
+from tabulate import tabulate
from mhcflurry import __version__
from mhcflurry.downloads import get_path
@@ -89,18 +92,66 @@ def go(argv):
# Architecture information rst
if predictor is None:
predictor = Class1AffinityPredictor.load(args.class1_models_dir)
- network = predictor.neural_networks[0].network()
- lines = []
- network.summary(print_fn=lines.append)
+
+ representative_networks = OrderedDict()
+ for network in predictor.neural_networks:
+ config = json.dumps(network.hyperparameters)
+ if config not in representative_networks:
+ representative_networks[config] = network
+
+ all_hyperparameters = [
+ network.hyperparameters for network in representative_networks.values()
+ ]
+ hyperparameter_keys = all_hyperparameters[0].keys()
+ assert all(
+ hyperparameters.keys() == hyperparameter_keys
+ for hyperparameters in all_hyperparameters)
+
+ constant_hyperparameter_keys = [
+ k for k in hyperparameter_keys
+ if all([
+ hyperparameters[k] == all_hyperparameters[0][k]
+ for hyperparameters in all_hyperparameters
+ ])
+ ]
+ constant_hypeparameters = dict(
+ (key, all_hyperparameters[0][key])
+ for key in sorted(constant_hyperparameter_keys)
+ )
+
+ def write_hyperparameters(fd, hyperparameters):
+ rows = []
+ for key in sorted(hyperparameters.keys()):
+ rows.append((key, json.dumps(hyperparameters[key])))
+ fd.write("\n")
+ fd.write(
+ tabulate(rows, ["Hyperparameter", "Value"], tablefmt="grid"))
with open(args.out_models_info_rst, "w") as fd:
- fd.write("Layers and parameters summary: ")
- fd.write("\n\n::\n\n")
- for line in lines:
- fd.write(" ")
- fd.write(line)
+ fd.write("Hyperparameters shared by all %d architectures:\n" %
+ len(representative_networks))
+ write_hyperparameters(fd, constant_hypeparameters)
+ fd.write("\n")
+ for (i, network) in enumerate(representative_networks.values()):
+ lines = []
+ network.network().summary(print_fn=lines.append)
+
+ fd.write("Architecture %d / %d:\n" % (
+ (i + 1, len(representative_networks))))
+ fd.write("+" * 40)
fd.write("\n")
- print("Wrote: %s" % args.out_models_info_rst)
+ write_hyperparameters(
+ fd,
+ dict(
+ (key, value)
+ for (key, value) in network.hyperparameters.items()
+ if key not in constant_hypeparameters))
+ fd.write("\n\n::\n\n")
+ for line in lines:
+ fd.write(" ")
+ fd.write(line)
+ fd.write("\n")
+ print("Wrote: %s" % args.out_models_info_rst)
if args.out_models_cv_rst:
# Models cv output
diff --git a/docs/models.rst b/docs/models.rst
index 4f7dee9edb93dcf686b830795aad79bf93c74bd7..96f2d1243e8975a3d6a3b31d7a88bbdf903680f8 100644
--- a/docs/models.rst
+++ b/docs/models.rst
@@ -1,35 +1,28 @@
Details on the released models
===============================
-The released MHCflurry predictor consists of an ensemble of eight models for each
-supported allele. Each model in the ensemble was trained on a random 80% sample
-of the data for the allele, and the remaining 20% was used for early stopping.
-All models use the same architecture. The predictions are taken to be the geometric
-mean of the nM binding affinity predictions of the individual models. The script
-we run to train these models is in "downloads-generation/models_class1/GENERATE.sh"
-in the repository.
-
-Neural network architecture
+The released MHCflurry predictor consists of an ensemble of models for each
+supported allele. Each model in the ensemble was trained on a random 90% sample
+of the data for the allele, and the remaining data was used for early stopping.
+The predictions are taken to be the geometric mean of the nM binding affinity
+predictions of the individual models whose predictions fall in the middle 50% of
+values for a given prediction. The script we run to train these models is in
+"downloads-generation/models_class1/GENERATE.sh" in the repository.
+
+Neural network architectures
-------------------------------------------------------------
-The neural network architecture is quite simple, consisting of a locally
-connected layer, a dense layer, and a sigmoid output.
-
.. include:: /_build/_models_info.rst
-Architecture diagram:
-
-.. image:: /_build/_models_architecture.png
-
-Cross validation performance
--------------------------------------------------------------
+.. Cross validation performance
+.. -------------------------------------------------------------
-The accuracy of the MHCflurry downloadable models was estimated using 5-fold cross
-validation on the training data. The values shown here are the mean cross validation
-scores across folds.
+.. The accuracy of the MHCflurry downloadable models was estimated using 5-fold cross
+.. validation on the training data. The values shown here are the mean cross validation
+.. scores across folds.
-The AUC and F1 estimates use a 500 nM cutoff for distinguishing strong-binders
-from weak- or non-binders. The Kendall Tau score gives the rank correlation
-between the predicted and measured affinities; it uses no cutoff.
+.. The AUC and F1 estimates use a 500 nM cutoff for distinguishing strong-binders
+.. from weak- or non-binders. The Kendall Tau score gives the rank correlation
+.. between the predicted and measured affinities; it uses no cutoff.
-.. include:: /_build/_models_cv.rst
+.. .. include:: /_build/_models_cv.rst
diff --git a/docs/python_tutorial.rst b/docs/python_tutorial.rst
index 9b3d8b882a91bea6899cb6583382d691f717c7e8..5840db46484f64b430a96fb9ec86694af0328a01 100644
--- a/docs/python_tutorial.rst
+++ b/docs/python_tutorial.rst
@@ -65,7 +65,7 @@ We can get the path to this data from Python using `mhcflurry.downloads.get_path
.. runblock:: pycon
>>> from mhcflurry.downloads import get_path
- >>> data_path = get_path("data_curated", "curated_training_data.csv.bz2")
+ >>> data_path = get_path("data_curated", "curated_training_data.no_mass_spec.csv.bz2")
>>> data_path
Now let's load it with pandas and filter to reasonably-sized peptides:
diff --git a/docs/requirements.txt b/docs/requirements.txt
index a88547504c1d7ede3af6ca6d67010fc0614ade96..af205bbbf1b5613450cffece92b58334dd06ea34 100644
--- a/docs/requirements.txt
+++ b/docs/requirements.txt
@@ -7,3 +7,4 @@ numpydoc
pypandoc
mhctools
pydot
+tabulate
diff --git a/mhcflurry/class1_neural_network.py b/mhcflurry/class1_neural_network.py
index c48b051ff22600e593676190226841eeb2f6df56..936e75119f6974fb6b619ad4f57461a9a7e50339 100644
--- a/mhcflurry/class1_neural_network.py
+++ b/mhcflurry/class1_neural_network.py
@@ -249,6 +249,7 @@ class Class1NeuralNetwork(object):
result = dict(self.__dict__)
result['_network'] = None
result['network_weights'] = None
+ result['network_weights_loader'] = None
return result
@classmethod
@@ -277,6 +278,12 @@ class Class1NeuralNetwork(object):
return instance
def load_weights(self):
+ """
+ Load weights by evaluating self.network_weights_loader, if needed.
+
+ After calling this, self.network_weights_loader will be None and
+ self.network_weights will be the weights list, if available.
+ """
if self.network_weights_loader:
self.network_weights = self.network_weights_loader()
self.network_weights_loader = None