def main(model_name,
batch_size=64,
model_params=None,
preprocessing_params=None,
test_run=False):
model_wrapper = model_dict[model_name](*model_params)
preprocessor = Preprocessor(**preprocessing_params)
model_dir = os.path.join('test_runs' if test_run else 'output',
model_wrapper.path(), preprocessor.path())
success_path = os.path.join(model_dir, 'history')
if not os.path.exists(success_path):
best_model_path = os.path.join(model_dir, 'model-best.hdf5')
shutil.rmtree(model_dir, ignore_errors=True)
os.makedirs(model_dir)
model = model_wrapper.get_model()
callbacks = [
ModelCheckpoint(best_model_path,
monitor='val_loss',
verbose=1,
save_best_only=True,
mode='min'),
EarlyStopping(patience=1 if test_run else 5, verbose=1)
]
if test_run:
print 'test run'
train_generator = preprocessor.get_train_generator(
vs.TINY_TRAIN_DIR, batch_size)
val_generator = preprocessor.get_test_generator(
vs.TINY_VALIDATION_DIR, batch_size)
test_generator = preprocessor.get_test_generator(
vs.TINY_TEST_DIR, batch_size)
print 'class indices'
print train_generator.class_indices
else:
print 'true run'
train_generator = preprocessor.get_train_generator(
vs.TRAIN_DIR, batch_size)
val_generator = preprocessor.get_test_generator(
vs.VALIDATION_DIR, batch_size)
test_generator = preprocessor.get_test_generator(
vs.TEST_DIR, batch_size)
train_samples = train_generator.samples
val_samples = val_generator.samples
history = model.fit_generator(
train_generator,
train_samples // batch_size,
validation_data=val_generator,
validation_steps=val_samples // batch_size + 1,
epochs=100,
callbacks=callbacks).history
with open(success_path, 'wb') as f:
pickle.dump(history, f)
model = load_model(best_model_path)
make_submission(model, test_generator,
os.path.join(model_dir, 'submission.csv'))
with open(success_path, 'rb') as f:
history = pickle.load(f)
print model_wrapper
print preprocessor
print 'best loss %.3f' % min(history['val_loss'])
print 'best accuracy %.3f' % max(history['val_acc'])
