From 6d52392f4964420f6074376e3dddd50ca2c9886f Mon Sep 17 00:00:00 2001
From: Alex Rubinsteyn <alex.rubinsteyn@gmail.com>
Date: Mon, 18 Apr 2016 19:25:32 -0400
Subject: [PATCH] adding options to training script and updating to work with
new predictor class
---
.../class1_allele_specific_hyperparameters.py | 49 ++++++++++++++++++-
.../train-class1-allele-specific-models.py | 37 ++++++--------
test/test_known_class1_epitopes.py | 0
3 files changed, 63 insertions(+), 23 deletions(-)
create mode 100644 test/test_known_class1_epitopes.py
diff --git a/mhcflurry/class1_allele_specific_hyperparameters.py b/mhcflurry/class1_allele_specific_hyperparameters.py
index 6d2a4f2d..c39cc60f 100644
--- a/mhcflurry/class1_allele_specific_hyperparameters.py
+++ b/mhcflurry/class1_allele_specific_hyperparameters.py
@@ -12,7 +12,6 @@
# See the License for the specific language governing permissions and
# limitations under the License.
-N_PRETRAIN_EPOCHS = 5
N_EPOCHS = 250
ACTIVATION = "tanh"
INITIALIZATION_METHOD = "lecun_uniform"
@@ -20,3 +19,51 @@ EMBEDDING_DIM = 32
HIDDEN_LAYER_SIZE = 200
DROPOUT_PROBABILITY = 0.25
MAX_IC50 = 50000.0
+
+def add_hyperparameter_arguments_to_parser(parser):
+ """
+ Extend an argument parser with the following options:
+ --training-epochs
+ --activation
+ --initialization
+ --embedding-size
+ --hidden-layer-size
+ --dropout
+ --max-ic50
+ """
+ parser.add_argument(
+ "--training-epochs",
+ default=N_EPOCHS,
+ help="Number of training epochs")
+
+ parser.add_argument(
+ "--initialization",
+ default=INITIALIZATION_METHOD,
+ help="Initialization for neural network weights")
+
+ parser.add_argument(
+ "--activation",
+ default=ACTIVATION,
+ help="Activation function for neural network layers")
+
+ parser.add_argument(
+ "--embedding-size",
+ default=EMBEDDING_DIM,
+ help="Size of vector representations for embedding amino acids")
+
+ parser.add_argument(
+ "--hidden-layer-size",
+ default=HIDDEN_LAYER_SIZE,
+ help="Size of hidden neural network layer")
+
+ parser.add_argument(
+ "--dropout",
+ default=DROPOUT_PROBABILITY,
+ help="Dropout probability after neural network layers")
+
+ parser.add_argument(
+ "--max-ic50",
+ default=MAX_IC50,
+ help="Largest IC50 represented by neural network output")
+
+ return parser
diff --git a/scripts/train-class1-allele-specific-models.py b/scripts/train-class1-allele-specific-models.py
index 6342e05e..5b25db86 100755
--- a/scripts/train-class1-allele-specific-models.py
+++ b/scripts/train-class1-allele-specific-models.py
@@ -47,14 +47,7 @@ from mhcflurry.common import normalize_allele_name
from mhcflurry.feedforward import make_network
from mhcflurry.data_helpers import load_allele_datasets
from mhcflurry.class1_allele_specific_hyperparameters import (
- N_PRETRAIN_EPOCHS,
- N_EPOCHS,
- ACTIVATION,
- INITIALIZATION_METHOD,
- EMBEDDING_DIM,
- HIDDEN_LAYER_SIZE,
- DROPOUT_PROBABILITY,
- MAX_IC50
+ add_hyperparamer_arguments_to_parser
)
from mhcflurry.paths import (
CLASS1_MODEL_DIRECTORY,
@@ -70,6 +63,7 @@ parser.add_argument(
default=CLASS1_MODEL_DIRECTORY,
help="Output directory for allele-specific predictor HDF weights files")
+
parser.add_argument(
"--overwrite",
default=False,
@@ -87,6 +81,9 @@ parser.add_argument(
help="Don't train predictors for alleles with fewer samples than this",
type=int)
+# add options for neural network hyperparameters
+parser = add_hyperparamer_arguments_to_parser(parser)
+
if __name__ == "__main__":
args = parser.parse_args()
@@ -97,7 +94,7 @@ if __name__ == "__main__":
args.binding_data_csv_path,
peptide_length=9,
binary_encoding=False,
- max_ic50=MAX_IC50,
+ max_ic50=args.max_ic50,
sep=",",
peptide_column_name="peptide")
@@ -107,18 +104,15 @@ if __name__ == "__main__":
Y_all = np.concatenate([group.Y for group in allele_groups.values()])
print("Total Dataset size = %d" % len(Y_all))
- model = make_network(
- input_size=9,
- embedding_input_dim=20,
- embedding_output_dim=EMBEDDING_DIM,
- layer_sizes=(HIDDEN_LAYER_SIZE,),
- activation=ACTIVATION,
- init=INITIALIZATION_METHOD,
- dropout_probability=DROPOUT_PROBABILITY)
- print("Model config: %s" % (model.get_config(),))
- model.fit(X_all, Y_all, nb_epoch=N_PRETRAIN_EPOCHS)
- old_weights = model.get_weights()
for allele_name, allele_data in allele_groups.items():
+ model = make_network(
+ input_size=9,
+ embedding_input_dim=20,
+ embedding_output_dim=args.embedding_size,
+ layer_sizes=(args.hidden_layer_size,),
+ activation=args.activation,
+ init=args.initialization,
+ dropout_probability=args.dropout)
allele_name = normalize_allele_name(allele_name)
if allele_name.isdigit():
print("Skipping allele %s" % (allele_name,))
@@ -147,11 +141,10 @@ if __name__ == "__main__":
print("-- removing old weights file %s" % hdf_path)
remove(hdf_path)
- model.set_weights(old_weights)
model.fit(
allele_data.X,
allele_data.Y,
- nb_epoch=N_EPOCHS,
+ nb_epoch=args.training_epochs,
show_accuracy=True)
print("Saving model description for %s to %s" % (
allele_name, json_path))
diff --git a/test/test_known_class1_epitopes.py b/test/test_known_class1_epitopes.py
new file mode 100644
index 00000000..e69de29b
--
GitLab