From ab85ec989c1974436bd07f3079eff2d90e071623 Mon Sep 17 00:00:00 2001
From: Tim O'Donnell <timodonnell@gmail.com>
Date: Thu, 4 Jul 2019 12:15:42 -0400
Subject: [PATCH] fixes
---
mhcflurry/class1_neural_network.py | 29 +++-
mhcflurry/select_pan_allele_models_command.py | 148 +-----------------
mhcflurry/train_pan_allele_models_command.py | 29 +++-
test/test_train_pan_allele_models_command.py | 10 +-
4 files changed, 58 insertions(+), 158 deletions(-)
diff --git a/mhcflurry/class1_neural_network.py b/mhcflurry/class1_neural_network.py
index d2d36761..9d9f00bd 100644
--- a/mhcflurry/class1_neural_network.py
+++ b/mhcflurry/class1_neural_network.py
@@ -458,6 +458,9 @@ class Class1NeuralNetwork(object):
"""
import keras
+ from keras import backend as K
+
+ fit_info = collections.defaultdict(list)
loss = get_loss(self.hyperparameters['loss'])
@@ -478,6 +481,13 @@ class Class1NeuralNetwork(object):
network._make_predict_function()
self.set_allele_representations(allele_representations)
+ if self.hyperparameters['learning_rate'] is not None:
+ K.set_value(
+ self.network().optimizer.lr,
+ self.hyperparameters['learning_rate'])
+ fit_info["learning_rate"] = float(
+ K.get_value(self.network().optimizer.lr))
+
validation_x_dict = {
'peptide': self.peptides_to_network_input(
validation_peptide_encoding),
@@ -513,7 +523,8 @@ class Class1NeuralNetwork(object):
yielded_values_box[0] += len(affinities)
start = time.time()
- result = network.fit_generator(
+
+ fit_history = network.fit_generator(
wrapped_generator(),
steps_per_epoch=steps_per_epoch,
epochs=epochs,
@@ -526,10 +537,12 @@ class Class1NeuralNetwork(object):
patience=patience,
verbose=verbose)]
)
- if verbose > 0:
- print("fit_generator completed in %0.2f sec (%d total points)" % (
- time.time() - start, yielded_values_box[0]))
- return result
+ for (key, value) in fit_history.history.items():
+ fit_info[key].extend(value)
+
+ fit_info["time"] = time.time() - start
+ fit_info["num_points"] = yielded_values_box[0]
+ self.fit_info.append(dict(fit_info))
def fit(
@@ -585,8 +598,10 @@ class Class1NeuralNetwork(object):
How often (in seconds) to print progress update. Set to None to
disable.
"""
+ from keras import backend as K
encodable_peptides = EncodableSequences.create(peptides)
peptide_encoding = self.peptides_to_network_input(encodable_peptides)
+ fit_info = collections.defaultdict(list)
length_counts = (
pandas.Series(encodable_peptides.sequences)
@@ -687,10 +702,11 @@ class Class1NeuralNetwork(object):
loss=loss.loss, optimizer=self.hyperparameters['optimizer'])
if self.hyperparameters['learning_rate'] is not None:
- from keras import backend as K
K.set_value(
self.network().optimizer.lr,
self.hyperparameters['learning_rate'])
+ fit_info["learning_rate"] = float(
+ K.get_value(self.network().optimizer.lr))
if loss.supports_inequalities:
# Do not sample negative affinities: just use an inequality.
@@ -762,7 +778,6 @@ class Class1NeuralNetwork(object):
min_val_loss_iteration = None
min_val_loss = None
- fit_info = collections.defaultdict(list)
start = time.time()
last_progress_print = None
x_dict_with_random_negatives = {}
diff --git a/mhcflurry/select_pan_allele_models_command.py b/mhcflurry/select_pan_allele_models_command.py
index e44cdcdf..510016bf 100644
--- a/mhcflurry/select_pan_allele_models_command.py
+++ b/mhcflurry/select_pan_allele_models_command.py
@@ -43,13 +43,6 @@ parser.add_argument(
help=(
"Model selection data CSV. Expected columns: "
"allele, peptide, measurement_value"))
-parser.add_argument(
- "--exclude-data",
- metavar="FILE.csv",
- required=False,
- help=(
- "Data to EXCLUDE from model selection. Useful to specify the original "
- "training data used"))
parser.add_argument(
"--models-dir",
metavar="DIR",
@@ -60,24 +53,6 @@ parser.add_argument(
metavar="DIR",
required=True,
help="Directory to write selected models")
-parser.add_argument(
- "--out-unselected-predictions",
- metavar="FILE.csv",
- help="Write predictions for validation data using unselected predictor to "
- "FILE.csv")
-parser.add_argument(
- "--unselected-accuracy-scorer",
- metavar="SCORER",
- default="combined:mass-spec,mse")
-parser.add_argument(
- "--unselected-accuracy-scorer-num-samples",
- type=int,
- default=1000)
-parser.add_argument(
- "--unselected-accuracy-percentile-threshold",
- type=float,
- metavar="X",
- default=95)
parser.add_argument(
"--min-models",
type=int,
@@ -122,15 +97,14 @@ def run(argv=sys.argv[1:]):
print("Loaded: %s" % input_predictor)
alleles = input_predictor.supported_alleles
+ (min_peptide_length, max_peptide_length) = (
+ input_predictor.supported_peptide_lengths)
metadata_dfs = {}
df = pandas.read_csv(args.data)
print("Loaded data: %s" % (str(df.shape)))
- (min_peptide_length, max_peptide_length) = (
- input_predictor.supported_peptide_lengths)
-
- df = df.ix[
+ df = df.loc[
(df.peptide.str.len() >= min_peptide_length) &
(df.peptide.str.len() <= max_peptide_length)
]
@@ -141,26 +115,10 @@ def run(argv=sys.argv[1:]):
# Allele names in data are assumed to be already normalized.
df = df.loc[df.allele.isin(alleles)].dropna()
- print("Selected %d alleles: %s" % (len(alleles), ' '.join(alleles)))
-
- if args.exclude_data:
- exclude_df = pandas.read_csv(args.exclude_data)
- metadata_dfs["model_selection_exclude"] = exclude_df
- print("Loaded exclude data: %s" % (str(df.shape)))
-
- df["_key"] = df.allele + "__" + df.peptide
- exclude_df["_key"] = exclude_df.allele + "__" + exclude_df.peptide
- df["_excluded"] = df._key.isin(exclude_df._key.unique())
- print("Excluding measurements per allele (counts): ")
- print(df.groupby("allele")._excluded.sum())
+ print("Subselected to supported alleles: %s" % str(df.shape))
- print("Excluding measurements per allele (fractions): ")
- print(df.groupby("allele")._excluded.mean())
- df = df.loc[~df._excluded]
- del df["_excluded"]
- del df["_key"]
- print("Reduced data to: %s" % (str(df.shape)))
+ print("Selected %d alleles: %s" % (len(alleles), ' '.join(alleles)))
metadata_dfs["model_selection_data"] = df
@@ -168,101 +126,9 @@ def run(argv=sys.argv[1:]):
args.mass_spec_regex)
- if args.out_unselected_predictions:
- df["unselected_prediction"] = input_predictor.predict(
- alleles=df.allele.values,
- peptides=df.peptide.values)
- df.to_csv(args.out_unselected_predictions)
- print("Wrote: %s" % args.out_unselected_predictions)
-
- selectors = {}
- selector_to_model_selection_kwargs = {}
-
- def make_selector(
- scoring,
- combined_min_contribution_percent=args.combined_min_contribution_percent):
- if scoring in selectors:
- return (
- selectors[scoring], selector_to_model_selection_kwargs[scoring])
-
- start = time.time()
- if scoring.startswith("combined:"):
- model_selection_kwargs = {
- 'min_models': args.combined_min_models,
- 'max_models': args.combined_max_models,
- }
- component_selectors = []
- for component_selector in scoring.split(":", 1)[1].split(","):
- component_selectors.append(
- make_selector(
- component_selector)[0])
- selector = CombinedModelSelector(
- component_selectors,
- min_contribution_percent=combined_min_contribution_percent)
- elif scoring == "mse":
- model_selection_kwargs = {
- 'min_models': args.mse_min_models,
- 'max_models': args.mse_max_models,
- }
- min_measurements = args.mse_min_measurements
- selector = MSEModelSelector(
- df=df.loc[~df.mass_spec],
- predictor=input_predictor,
- min_measurements=min_measurements)
- elif scoring == "mass-spec":
- mass_spec_df = df.loc[df.mass_spec]
- model_selection_kwargs = {
- 'min_models': args.mass_spec_min_models,
- 'max_models': args.mass_spec_max_models,
- }
- min_measurements = args.mass_spec_min_measurements
- selector = MassSpecModelSelector(
- df=mass_spec_df,
- predictor=input_predictor,
- min_measurements=min_measurements)
- elif scoring == "consensus":
- model_selection_kwargs = {
- 'min_models': args.consensus_min_models,
- 'max_models': args.consensus_max_models,
- }
- selector = ConsensusModelSelector(
- predictor=input_predictor,
- num_peptides_per_length=args.consensus_num_peptides_per_length)
- else:
- raise ValueError("Unsupported scoring method: %s" % scoring)
- print("Instantiated model selector %s in %0.2f sec." % (
- scoring, time.time() - start))
- return (selector, model_selection_kwargs)
-
- for scoring in args.scoring:
- (selector, model_selection_kwargs) = make_selector(scoring)
- selectors[scoring] = selector
- selector_to_model_selection_kwargs[scoring] = model_selection_kwargs
-
- unselected_accuracy_scorer = None
- if args.unselected_accuracy_scorer:
- # Force running all selectors by setting combined_min_contribution_percent=0.
- unselected_accuracy_scorer = make_selector(
- args.unselected_accuracy_scorer,
- combined_min_contribution_percent=0.0)[0]
- print("Using unselected accuracy scorer: %s" % unselected_accuracy_scorer)
- GLOBAL_DATA["unselected_accuracy_scorer"] = unselected_accuracy_scorer
- print("Selectors for alleles:")
- allele_to_selector = {}
- allele_to_model_selection_kwargs = {}
- for allele in alleles:
- selector = None
- for possible_selector in args.scoring:
- if selectors[possible_selector].usable_for_allele(allele=allele):
- selector = selectors[possible_selector]
- print("%20s %s" % (allele, selector.plan_summary(allele)))
- break
- if selector is None:
- raise ValueError("No selectors usable for allele: %s" % allele)
- allele_to_selector[allele] = selector
- allele_to_model_selection_kwargs[allele] = (
- selector_to_model_selection_kwargs[possible_selector])
+
+
GLOBAL_DATA["args"] = args
GLOBAL_DATA["input_predictor"] = input_predictor
diff --git a/mhcflurry/train_pan_allele_models_command.py b/mhcflurry/train_pan_allele_models_command.py
index bafb6f72..1abf6712 100644
--- a/mhcflurry/train_pan_allele_models_command.py
+++ b/mhcflurry/train_pan_allele_models_command.py
@@ -9,6 +9,7 @@ import time
import traceback
import random
import pprint
+import hashlib
from functools import partial
import numpy
@@ -130,6 +131,7 @@ add_worker_pool_args(parser)
def assign_folds(df, num_folds, held_out_fraction, held_out_max):
result_df = pandas.DataFrame(index=df.index)
+
for fold in range(num_folds):
result_df["fold_%d" % fold] = True
for (allele, sub_df) in df.groupby("allele"):
@@ -172,6 +174,9 @@ def assign_folds(df, num_folds, held_out_fraction, held_out_max):
print("Test points per fold")
print((~result_df).sum())
+ result_df["allele"] = df["allele"]
+ result_df["peptide"] = df["peptide"]
+
return result_df
@@ -422,8 +427,6 @@ def train_model(
progress_print_interval,
predictor,
save_to):
- import keras.backend as K
- import keras
df = GLOBAL_DATA["train_data"]
folds_df = GLOBAL_DATA["folds_df"]
@@ -484,10 +487,10 @@ def train_model(
epochs=pretrain_max_epochs,
verbose=verbose,
)
- if model.hyperparameters['learning_rate']:
- model.hyperparameters['learning_rate'] /= 10
- else:
- model.hyperparameters['learning_rate'] = 0.0001
+
+ # Use a smaller learning rate for training on real data
+ learning_rate = model.fit_info[-1]["learning_rate"]
+ model.hyperparameters['learning_rate'] = learning_rate / 10
model.fit(
peptides=train_peptides,
@@ -500,6 +503,20 @@ def train_model(
progress_print_interval=progress_print_interval,
verbose=verbose)
+ # Save model-specific training info
+ train_peptide_hash = hashlib.sha1()
+ for peptide in train_data.peptide.values:
+ train_peptide_hash.update(peptide.encode())
+ model.fit_info[-1]["training_info"] = {
+ "fold_num": fold_num,
+ "num_folds": num_folds,
+ "replicate_num": replicate_num,
+ "num_replicates": num_replicates,
+ "architecture_num": architecture_num,
+ "num_architectures": num_architectures,
+ "train_peptide_hash": train_peptide_hash.hexdigest(),
+ }
+
numpy.testing.assert_equal(
predictor.manifest_df.shape[0], len(predictor.class1_pan_allele_models))
predictor.add_pan_allele_model(model, models_dir_for_save=save_to)
diff --git a/test/test_train_pan_allele_models_command.py b/test/test_train_pan_allele_models_command.py
index f4142a1d..f98599ff 100644
--- a/test/test_train_pan_allele_models_command.py
+++ b/test/test_train_pan_allele_models_command.py
@@ -101,7 +101,7 @@ HYPERPARAMETERS_LIST = [
][1:]
-def run_and_check(n_jobs=0):
+def run_and_check(n_jobs=0, delete=True):
models_dir = tempfile.mkdtemp(prefix="mhcflurry-test-models")
hyperparameters_filename = os.path.join(
models_dir, "hyperparameters.yaml")
@@ -140,8 +140,9 @@ def run_and_check(n_jobs=0):
alleles=["HLA-A*02:01"])
print(df)
- print("Deleting: %s" % models_dir)
- shutil.rmtree(models_dir)
+ if delete:
+ print("Deleting: %s" % models_dir)
+ shutil.rmtree(models_dir)
if os.environ.get("KERAS_BACKEND") != "theano":
@@ -153,5 +154,6 @@ if os.environ.get("KERAS_BACKEND") != "theano":
def test_run_serial():
run_and_check(n_jobs=0)
+
if __name__ == "__main__":
- test_run_serial()
\ No newline at end of file
+ run_and_check(n_jobs=0, delete=False)
--
GitLab