From 53496d261f53cf375fa59cba2a0952499c56bf42 Mon Sep 17 00:00:00 2001
From: Alex Rubinsteyn <alex.rubinsteyn@gmail.com>
Date: Thu, 2 Jun 2016 12:54:11 -0400
Subject: [PATCH] pass dropout to Embedding instead of creating Dropout layer
---
mhcflurry/dataset.py | 10 +++++++++-
mhcflurry/feedforward.py | 9 ++++-----
2 files changed, 13 insertions(+), 6 deletions(-)
diff --git a/mhcflurry/dataset.py b/mhcflurry/dataset.py
index 63d892c9..a59d825b 100644
--- a/mhcflurry/dataset.py
+++ b/mhcflurry/dataset.py
@@ -33,6 +33,7 @@ from .imputation_helpers import (
check_dense_pMHC_array,
prune_dense_matrix_and_labels,
dense_pMHC_matrix_to_nested_dict,
+ imputer_from_name
)
@@ -649,7 +650,8 @@ class Dataset(object):
self,
test_allele=None,
n_folds=3,
- shuffle=True):
+ shuffle=True,
+ stratify_fn=None):
"""
Yields a sequence of training/test splits of this dataset.
@@ -657,6 +659,9 @@ class Dataset(object):
only split the measurements of the specified allele (other alleles
will then always be included in the training datasets).
"""
+ if stratify_fn is not None:
+ raise ValueError("StratifiedKFold not yet implemented!")
+
if test_allele is None:
candidate_test_indices = np.arange(len(self))
@@ -809,6 +814,9 @@ class Dataset(object):
Returns Dataset with original pMHC affinities and additional
synthetic samples.
"""
+ if isinstance(imputation_method, string_types):
+ imputation_method = imputer_from_name(imputation_method)
+
X_incomplete, peptide_list, allele_list = self.to_dense_pMHC_affinity_matrix(
min_observations_per_peptide=min_observations_per_peptide,
min_observations_per_allele=min_observations_per_allele)
diff --git a/mhcflurry/feedforward.py b/mhcflurry/feedforward.py
index eae30497..47ba4218 100644
--- a/mhcflurry/feedforward.py
+++ b/mhcflurry/feedforward.py
@@ -65,18 +65,17 @@ def make_network(
input_dim=embedding_input_dim,
output_dim=embedding_output_dim,
input_length=input_size,
- weights=[initial_embedding_weights]))
+ weights=[initial_embedding_weights],
+ dropout=dropout_probability))
else:
model.add(Embedding(
input_dim=embedding_input_dim,
output_dim=embedding_output_dim,
input_length=input_size,
- init=embedding_init_method))
+ init=embedding_init_method,
+ dropout=dropout_probability))
model.add(Flatten())
- if dropout_probability > 0:
- model.add(Dropout(dropout_probability))
-
input_size = input_size * embedding_output_dim
layer_sizes = (input_size,) + tuple(layer_sizes)
--
GitLab