From 347835717dc714c96bd4466486867438e6ceb229 Mon Sep 17 00:00:00 2001
From: Nikolai Hartmann <nikolai.hartmann@gmx.de>
Date: Thu, 29 Nov 2018 10:19:10 +0100
Subject: [PATCH] use keras callbacks in adversarial training
---
toolkit.py | 43 ++++++++++++++++++++++++++++++++++++++++++-
1 file changed, 42 insertions(+), 1 deletion(-)
diff --git a/toolkit.py b/toolkit.py
index 9e903ca..0403f9b 100755
--- a/toolkit.py
+++ b/toolkit.py
@@ -37,7 +37,7 @@ from sklearn.utils.extmath import stable_cumsum
from sklearn.model_selection import KFold
from keras.models import Sequential, Model, model_from_json
from keras.layers import Dense, Dropout, Input, Masking, GRU, LSTM, concatenate, SimpleRNN
-from keras.callbacks import History, EarlyStopping, CSVLogger, ModelCheckpoint, TensorBoard
+from keras.callbacks import History, EarlyStopping, CSVLogger, ModelCheckpoint, TensorBoard, CallbackList
from keras.optimizers import SGD
import keras.optimizers
from keras.utils.vis_utils import model_to_dot
@@ -2220,29 +2220,70 @@ class ClassificationProjectDecorr(ClassificationProject):
def train(self, epochs=10):
+ """
+ Train classifier and adversary concurrently. Most of the garbage in this
+ code block is just organising stuff to get all the keras callbacks
+ right. That code is mostly copied over from keras `engine/training_generator.py`.
+ """
+
batch_generator = self.yield_batch()
metric_list = []
+ out_labels = self.model.metrics_names
+ callback_metrics = out_labels + ['val_' + n for n in out_labels]
+ callbacks = CallbackList(self.callbacks_list)
+ callbacks.set_model(self.model)
+ callbacks.set_params({
+ 'epochs': epochs,
+ 'steps': self.steps_per_epoch,
+ 'verbose': self.verbose,
+ #'do_validation': do_validation,
+ 'do_validation': False,
+ 'metrics': callback_metrics,
+ })
+ self.model.stop_training = False
+ callbacks.on_train_begin()
+ epoch_logs = {}
for epoch in range(epochs):
+ callbacks.on_epoch_begin(epoch)
logger.info("Fitting epoch {}".format(epoch))
metrics = None
avg_metrics = None
for batch_id in tqdm(range(self.steps_per_epoch)):
x, y, w = next(batch_generator)
+ batch_logs = {}
+ batch_logs['batch'] = batch_id
+ batch_logs['size'] = len(x)
+ callbacks.on_batch_begin(batch_id, batch_logs)
+
# fit the classifier
batch_metrics = self.model.train_on_batch(
x, y, sample_weight=w
)
+
# fit the adversary
self.model_adv.train_on_batch(
x, y[1:], sample_weight=w[1:]
)
+
batch_metrics = np.array(batch_metrics).reshape(1, len(batch_metrics))
if metrics is None:
metrics = batch_metrics
else:
metrics = np.concatenate([metrics, batch_metrics])
avg_metrics = np.mean(metrics, axis=0)
+ outs = list(batch_metrics)
+ for l, o in zip(out_labels, outs):
+ batch_logs[l] = o
+ callbacks.on_batch_end(batch_id, batch_logs)
metric_list.append(avg_metrics)
+ val_metrics = self.model.test_on_batch(*self.validation_data)
+ val_outs = list(val_metrics)
+ for l, o in zip(out_labels, val_outs):
+ epoch_logs['val_' + l] = o
+ callbacks.on_epoch_end(epoch, epoch_logs)
+ if self.model.stop_training:
+ break
+ callbacks.on_train_end()
return metric_list
--
GitLab