From d47ba9ca4ed120aab1fef103b182cdde464f2046 Mon Sep 17 00:00:00 2001
From: Tim O'Donnell <timodonnell@gmail.com>
Date: Tue, 9 Jul 2019 12:02:08 -0400
Subject: [PATCH] fixes
---
mhcflurry/class1_neural_network.py | 4 ----
mhcflurry/select_pan_allele_models_command.py | 23 ++++++++++++++++---
mhcflurry/train_pan_allele_models_command.py | 20 +++++++---------
3 files changed, 28 insertions(+), 19 deletions(-)
diff --git a/mhcflurry/class1_neural_network.py b/mhcflurry/class1_neural_network.py
index 9d9f00bd..d1add2b7 100644
--- a/mhcflurry/class1_neural_network.py
+++ b/mhcflurry/class1_neural_network.py
@@ -10,7 +10,6 @@ import pandas
from .hyperparameters import HyperparameterDefaults
from .encodable_sequences import EncodableSequences, EncodingError
-from .amino_acid import available_vector_encodings, vector_encoding_length
from .regression_target import to_ic50, from_ic50
from .common import random_peptides, amino_acid_distribution
from .custom_loss import get_loss
@@ -156,7 +155,6 @@ class Class1NeuralNetwork(object):
hyperparameters[to_name] = value
return hyperparameters
-
def __init__(self, **hyperparameters):
self.hyperparameters = self.hyperparameter_defaults.with_defaults(
self.apply_hyperparameter_renames(hyperparameters))
@@ -420,7 +418,6 @@ class Class1NeuralNetwork(object):
allele_encoding.allele_representations(
self.hyperparameters['allele_amino_acid_encoding']))
-
def fit_generator(
self,
generator,
@@ -544,7 +541,6 @@ class Class1NeuralNetwork(object):
fit_info["num_points"] = yielded_values_box[0]
self.fit_info.append(dict(fit_info))
-
def fit(
self,
peptides,
diff --git a/mhcflurry/select_pan_allele_models_command.py b/mhcflurry/select_pan_allele_models_command.py
index eb3bca3d..c564e7c1 100644
--- a/mhcflurry/select_pan_allele_models_command.py
+++ b/mhcflurry/select_pan_allele_models_command.py
@@ -232,19 +232,31 @@ def run(argv=sys.argv[1:]):
chunksize=1)
models_by_fold = {}
+ summary_dfs = []
for result in tqdm.tqdm(results, total=len(work_items)):
pprint(result)
fold_num = result['fold_num']
+ (all_models_for_fold, _) = folds_to_predictors[fold_num]
models = [
- folds_to_predictors[fold_num][0][i]
+ all_models_for_fold[i]
for i in result['selected_indices']
]
+ summary_df = result['summary'].copy()
+ summary_df.index = summary_df.index.map(
+ lambda idx: all_models_for_fold[idx])
+ summary_dfs.append(summary_df)
+
print("Selected %d models for fold %d: %s" % (
len(models), fold_num, result['selected_indices']))
models_by_fold[fold_num] = models
for model in models:
result_predictor.add_pan_allele_model(model)
+ summary_df = pandas.concat(summary_dfs, ignore_index=False)
+ summary_df["model_config"] = summary_df.index.map(lambda m: m.get_config())
+ result_predictor.metadata_dataframes["model_selection_summary"] = (
+ summary_df.reset_index(drop=True))
+
result_predictor.save(args.out_models_dir)
model_selection_time = time.time() - start
@@ -312,11 +324,16 @@ def model_select(fold_num, models, min_models, max_models):
break
assert selected
+
+ summary_df = pandas.Series(individual_model_scores)[
+ numpy.arange(len(models))
+ ].to_frame()
+ summary_df.columns = ['mse_score']
+
return {
'fold_num': fold_num,
'selected_indices': selected,
- 'individual_model_scores': pandas.Series(
- individual_model_scores)[numpy.arange(len(models))],
+ 'summary': summary_df, # indexed by model index
}
diff --git a/mhcflurry/train_pan_allele_models_command.py b/mhcflurry/train_pan_allele_models_command.py
index 0bee23e8..1ec1118c 100644
--- a/mhcflurry/train_pan_allele_models_command.py
+++ b/mhcflurry/train_pan_allele_models_command.py
@@ -15,8 +15,6 @@ from functools import partial
import numpy
import pandas
import yaml
-from sklearn.metrics.pairwise import cosine_similarity
-from sklearn.model_selection import StratifiedKFold
from mhcnames import normalize_allele_name
import tqdm # progress bar
tqdm.monitor_interval = 0 # see https://github.com/tqdm/tqdm/issues/481
@@ -28,11 +26,8 @@ from .parallelism import (
add_worker_pool_args,
worker_pool_with_gpu_assignments_from_args,
call_wrapped_kwargs)
-from .hyperparameters import HyperparameterDefaults
from .allele_encoding import AlleleEncoding
from .encodable_sequences import EncodableSequences
-from .regression_target import to_ic50, from_ic50
-from .import custom_loss
# To avoid pickling large matrices to send to child processes when running in
@@ -173,10 +168,6 @@ 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
@@ -304,10 +295,13 @@ def main(args):
predictor = Class1AffinityPredictor(
allele_to_sequence=allele_encoding.allele_to_sequence,
metadata_dataframes={
- 'train_data': df,
- 'training_folds': folds_df,
+ 'train_data': pandas.merge(
+ df,
+ folds_df,
+ left_index=True,
+ right_index=True)
})
- serial_run = args.num_jobs == 1
+ serial_run = args.num_jobs == 0
work_items = []
for (h, hyperparameters) in enumerate(hyperparameters_lst):
@@ -353,6 +347,7 @@ def main(args):
if worker_pool:
print("Processing %d work items in parallel." % len(work_items))
+ assert not serial_run
results_generator = worker_pool.imap_unordered(
partial(call_wrapped_kwargs, train_model),
@@ -389,6 +384,7 @@ def main(args):
# which it adds models to, so no merging is required. It also saves
# as it goes so no saving is required at the end.
print("Processing %d work items in serial." % len(work_items))
+ assert serial_run
for _ in tqdm.trange(len(work_items)):
item = work_items.pop(0) # want to keep freeing up memory
work_predictor = train_model(**item)
--
GitLab