writer.writeheader()
loss_val = []
for epoch in range(num_epochs):
print("EPOCH ", epoch, "/", num_epochs)
temp_loss_list = []
label_list = []
pred_list = []
temp_loss_val = []
label_val = []
pred_val = []
'Update weights on training set'
for i in range(training_generator.__len__()):
x_train,y_train=training_generator.__getitem__(i)
with tf.GradientTape() as tape:
logits = model(x_train, training = True)
loss_values = loss(y_train, logits)
'Follow training by printing the loss and prediction for each image'
#print("Batch ", i, "/",training_generator.__len__(),
# ", Loss: ", loss_values.numpy()[0])
#print("Prediction: ", logits.numpy()[0,:], ", Label: ", y_train[0])
temp_loss_list.append(loss_values.numpy().mean())
label_list.append(y_train[0])
pred_list.append(logits.numpy()[0,1])
grads = tape.gradient(loss_values, model.trainable_variables[-4:])
print("EPOCH ", epoch, "/", num_epochs)
"Reset all metrics"
train_loss.reset_states()
train_accuracy.reset_states()
test_loss.reset_states()
test_accuracy.reset_states()
label_list = []
pred_list = []
label_val = []
pred_val = []
pred_val_ema = []
"Iterate over all images of training set"
for i in range(training_generator.__len__()):
print("Batch ", i, "/",training_generator.__len__() - 1)
x_train, y_train = training_generator.__getitem__(i)
"Update model"
with tf.GradientTape() as tape:
predictions = model(x_train, training = True)
loss = loss_object(y_train, predictions)
gradients = tape.gradient(loss,
model.trainable_variables[-nb_train_layers:])
optimizer.apply_gradients(zip(gradients,
model.trainable_variables[-nb_train_layers:]))
"Collect labels and predictions"
train_loss(loss)
train_accuracy(y_train[0], predictions)
label_list.append(y_train[0])
mean_cm = cnf_matrix_frame
else:
mean_cm += cnf_matrix_frame
mean_train.append(history.history['acc'][-1])
mean_test.append(history.history['val_acc'][-1])
mean_train_loss.append(history.history['loss'][-1])
mean_test_loss.append(history.history['val_loss'][-1])
mean_test_3.append(history.history['val_top_3_accuracy'][-1])
mean_test_5.append(history.history['val_top_5_accuracy'][-1])
if config_params['logging']['enable']:
with open(LOGGING_FILE, 'a') as f:
f.write('{},{},{},{},{},{},{},{},{}\n'.format(
SUBJECT,
train_generator.__len__() * PARAMS_TRAIN_GENERATOR['batch_size'],
valid_generator.__len__(), mean_train_loss[-1], mean_train[-1],
mean_test_loss[-1], mean_test[-1], mean_test_3[-1],
mean_test_5[-1]))
metrics_dict = {
'mean_cm': mean_cm,
'mean_test': mean_test,
'mean_test_3': mean_test_3,
'mean_test_5': mean_test_5,
'mean_train': mean_train,
'mean_train_loss': mean_train_loss,
'mean_test_loss': mean_test_loss
}
scipy.io.savemat(METRICS_SAVE_FILE.format(SUBJECT), metrics_dict)
def train(hires=False,
filters=16,
max_depth=2,
kernel_size=2,
pool_size=2,
doubling=False,
algorithm='adam'):
'''
*** 学習時のパラメータを設定 ***
'''
input_length = 552
batch_size = 10
epochs = 10
learn_rate = 1.0e-3
'''
*** 学習と評価用のデータを選ぶ ***
'''
if hires is True:
# 80次元のデータ．こちらの方が詳細なデータ
input_dim = 80
train_data = 'train_hires.h5'
test_data = 'test_hires.h5'
else:
# 40次元のデータ．デフォルト
input_dim = 40
train_data = './train.h5'
test_data = './test.h5'
'''
*** 学習に使うアルゴリズム（最適化手法）の選択 ***
'''
if algorithm == 'sgd':
optimizer = tf.keras.optimizers.SGD()
elif algorithm == 'adadelta':
optimizer = tf.keras.optimizers.Adadelta()
elif algorithm == 'rmsprop':
optimizer = tf.keras.optimizers.RMSprop()
else:
optimizer = tf.keras.optimizers.Adam()
'''
*** 入力ベクトルの形状(次元)を指定 ***
( 特徴量の次元，時間(長さ)の次元，フィルタ数の次元 ) = numpyのshape
'''
input = tf.keras.layers.Input(shape=(input_dim, input_length, 1))
'''
*** ニューラルネットワークを作成 ***
どのようにして作っているかはcnn.pyを読むこと
'''
output = CNNClassifier()(input)
'''
*** モデルの設定 ***
ニューラルネットワークの入出力を指定し，学習・評価できるように設定する
'''
model = tf.keras.models.Model(input, output)
# 学習可能な計算グラフを作成
model.compile(optimizer=optimizer,
loss='categorical_crossentropy',
metrics=['categorical_accuracy'])
'''
*** 学習データの準備 ***
'''
training_generator = DataGenerator(train_data,
dim=(input_dim, input_length),
batch_size=batch_size)
# 入力データが正規分布にしたがうよう，偏りをなくす（正規化）
training_generator.compute_norm()
mean, var = training_generator.get_norm()
'''
*** 評価データの準備 ***
'''
validation_generator = DataGenerator(test_data,
dim=(input_dim, input_length),
batch_size=batch_size)
validation_generator.set_norm(mean, var)
'''
*** Tensorboardを使って学習状況のログを保存する***
'''
try:
shutil.rmtree('./logs')
except:
pass
os.makedirs('./logs')
tensorboard = tf.keras.callbacks.TensorBoard(log_dir='./logs')
'''
*** 学習 ***
tensorflowの以下の関数を呼び出して行う．
'''
model.fit(training_generator,
validation_data=validation_generator,
epochs=epochs,
callbacks=[tensorboard],
shuffle=True)
'''
*** 評価 ***
学習したニューラルネットワークを使って評価を行う
'''
# 混同行列の初期化
mat = np.zeros((ASC_CLASS, ASC_CLASS), dtype=int)
# 正解数の初期化
acc = 0
for bt in range(validation_generator.__len__()):
# データを10個取り出す
# x ... メルスペクトル特徴, y...ラベル
x, y = validation_generator.__getitem__(bt)
# 学習したモデルで予測する
# pred は5つの鳴き声の確率が入ったベクトルを10個ならべたもの
pred = model.predict_on_batch(x)
# 確率が最大となるラベルの番号を求める
y_pred = np.argmax(pred, axis=1)
y_true = np.argmax(y, axis=1)
# 混同行列を作る
mat += confusion_matrix(y_true, y_pred, labels=[0, 1, 2, 3, 4])
# 正解と予測ラベルを比較して，正解した数を加算
acc += np.sum(y_pred == y_true)
# 評価データの総数で割る
acc = float(acc) / validation_generator.__num_samples__()
return acc, mat
metrics=['accuracy'])
tbCallBack = keras.callbacks.TensorBoard(log_dir='../Graph',
histogram_freq=0,
write_graph=True,
write_images=True)
training_generator = DataGenerator(IMAGES_PER_CLASS,
CLASSES,
RATIO,
data_path,
batch_size=batch_size,
dim=(SIZE, SIZE))
validation_generator = DataGenerator(IMAGES_PER_CLASS,
CLASSES,
RATIO,
data_path,
batch_size=batch_size,
dim=(SIZE, SIZE),
mode="validation")
model.fit_generator(training_generator,
epochs=epochs,
steps_per_epoch=training_generator.__len__(),
validation_data=validation_generator,
validation_steps=validation_generator.__len__(),
callbacks=[tbCallBack])
model.save('../models/keras/model_{}.h5'.format(
len(next(os.walk("../models/keras"))[2])))
