def test(self, classifier, model=None):
du = DataUtils()
X_test, y_test = du.data_preprocess('test')
pred = self.predict(X_test, classifier, model)
y_pred = np.zeros(len(pred), dtype=int)
y_pred[pred[:, 1] > pred[:, 0]] = 1
score = metrics.accuracy_score(y_test[:, 1], y_pred)
logger_tc.info('test accuracy: %.3f' % score)
with h5py.File(self.model_output_path) as model_output:
if '%s_test_pred' % classifier not in model_output:
model_output.create_dataset('%s_test_pred' % classifier, data=pred)
class TransferLearner:
def __init__(self):
self.num_classes = 2
self.resnet50_weights = os.path.realpath('models/resnet50_weights_tf_dim_ordering_tf_kernels_notop.h5')
self.xception_weights = os.path.realpath('models/xception_weights_tf_dim_ordering_tf_kernels_notop.h5')
self.model_output_path = os.path.realpath('data/model_output.h5')
self.model_path = {'resnet50': os.path.realpath('data/model_resnet50.h5'),
'xception': os.path.realpath('data/model_xception.h5')}
self.transfer_classifiers = {'resnet50': (ResNet50, self.resnet50_weights),
'xception': (Xception, self.xception_weights)}
self.du = DataUtils()
# transfer learn from pre-trained cnn to extract the features; mode as 'train', 'valid', 'cross_val' or 'test'
# mode: 'train', 'valid', 'no_save' (not save result, used in cross validation and prediction for analysis and test)
def _get_transfer(self, X, mode, classifier):
path = '%s_%s' % (classifier, mode)
with h5py.File(self.model_output_path) as model_output:
if path in model_output:
return model_output[path][:]
# load pre-trained cnn without top, output 2 classes as 0: negative; 1: positive
clf, weights = self.transfer_classifiers[classifier]
base = clf(False, weights, Input(shape=(299, 299, 3)), classes=self.num_classes)
f = K.function([base.layers[0].input, K.learning_phase()], [base.layers[-1].output])
# split whole data set to 100 per batch due to memory issue
transfer_output = [f([X[i: i + 100], 0])[0] for i in range(0, len(X), 100)]
transfer_output = np.concatenate(transfer_output, axis=0)
if mode != 'no_save':
with h5py.File(self.model_output_path) as model_output:
model_output.create_dataset(path, data=transfer_output)
return transfer_output
# if cross_val is on, inputs should be X_train, y_train, X_valid, y_valid which have been preprocessed
# mode: None (general model, save validation prediction), 'cross_val', 'model': output trained model
def train(self, classifier, mode=None, *inputs):
if mode != 'cross_val':
# data preprocess
X_train, y_train, X_valid, y_valid = self.du.data_preprocess('train')
# get pre-trained cnn's result as extracted features
logger_tc.info('start transfer learning')
transfer_train_output = self._get_transfer(X_train, 'train', classifier)
transfer_valid_output = self._get_transfer(X_valid, 'valid', classifier)
else:
X_train, y_train, X_valid, y_valid = inputs
logger_tc.info('start transfer learning')
transfer_train_output = self._get_transfer(X_train, 'no_save', classifier)
transfer_valid_output = self._get_transfer(X_valid, 'no_save', classifier)
logger_tc.info('transfer learning finished')
# parameter tuning
b_score, b_r1, b_r2, b_pred = 0, 0, 0, np.array([])
logger_tc.info('parameter tuning')
# for r1, r2 in permutations([1e-2, 1e-3, 1e-4], 2):
# for r1, r2 in [(np.random.randint(1, 10) / 10000., np.random.randint(1, 10) / 10000.) for _ in range(5)]:
for r1, r2 in [(1e-3, 3e-4)]:
# input extracted features to 2 FC layers (2048 -(RELU)-> 1024 -(SOFTMAX)-> 2) to get result
input_tensor = Input(shape=(1, 1, 2048))
X = Flatten()(input_tensor)
X = Dense(1024, activation='relu', kernel_regularizer=regularizers.l2(r1))(X)
predictions = Dense(self.num_classes, activation='softmax', kernel_regularizer=regularizers.l2(r2))(X)
model = Model(inputs=input_tensor, outputs=predictions)
model.compile(loss='categorical_crossentropy', optimizer='Adam', metrics=['accuracy'])
model.fit(transfer_train_output, y_train, epochs=30, batch_size=128, verbose=0)
# only save model when parameter and cross validation finished
# model.save(self.model_path[classifier])
# validate
pred = model.predict(transfer_valid_output, batch_size=32)
y_pred = np.zeros(len(pred), dtype=int)
y_pred[pred[:, 1] > pred[:, 0]] = 1
score = metrics.accuracy_score(y_valid[:, 1], y_pred)
if score > b_score:
b_score, b_r1, b_r2, b_pred = score, r1, r2, pred
logger_tc.info('cur: acc %.3f, r1 %s, r2 %s; best: acc %.3f, r1 %s, r2 %s' % (score, r1, r2, b_score, b_r1, b_r2))
if mode == 'model':
return model
if mode == 'cross_val':
return b_score
with h5py.File(self.model_output_path) as model_output:
if '%s_valid_pred' % classifier not in model_output:
model_output.create_dataset('%s_valid_pred' % classifier, data=b_pred)
def cross_validation(self, classifier, fold=5):
scores = []
for i in range(1, fold + 1):
logger_tc.info('cross validation fold %s start' % i)
X_train, y_train, X_valid, y_valid = self.du.data_extract('cross_val')
X_train, X_valid, y_train, y_valid= self.du.augmentation(X_train), self.du.augmentation(X_valid),\
np.tile(y_train, (8, 1)), np.tile(y_valid, (8, 1))
train_mean, train_std = np.mean(X_train, axis=0), np.std(X_train, axis=0)
X_train, X_valid = (X_train - train_mean) / train_std, (X_valid - train_mean) / train_std
y_train = np_utils.to_categorical(y_train[:, 1], self.num_classes)
scores += self.train(classifier, 'cross_val', X_train, y_train, X_valid, y_valid),
logger_tc.info('cross validation fold %s end\n\n' % i)
logger_tc.info(scores)
logger_tc.info('%s %s fold cross val: avg acc: %s, std: %s' % (classifier, fold, np.mean(scores), np.std(scores)))
def predict(self, X, classifier, model=None):
X = self._get_transfer(X, 'no_save', classifier)
return (model or load_model(self.model_path[classifier])).predict(X, batch_size=32)
def test(self, classifier, model=None):
du = DataUtils()
X_test, y_test = du.data_preprocess('test')
pred = self.predict(X_test, classifier, model)
y_pred = np.zeros(len(pred), dtype=int)
y_pred[pred[:, 1] > pred[:, 0]] = 1
score = metrics.accuracy_score(y_test[:, 1], y_pred)
logger_tc.info('test accuracy: %.3f' % score)
with h5py.File(self.model_output_path) as model_output:
if '%s_test_pred' % classifier not in model_output:
model_output.create_dataset('%s_test_pred' % classifier, data=pred)