Неправильный расчет потерь при работе с tf.keras. После построения модели tf.keras.fit_generator должен принять (inputs, targets, sample_weights) в качестве входных данных. Однако, если я умножу sample_weights на 10000, потери не изменятся.
Ошибка, по-видимому, появляется в версии Tensorflow 1.10 и выше, например (1.11, 1.12)
Код для воспроизведения
import numpy as np
import tensorflow as tf
WEIGHT_VARIABLE = 1
no_of_features = 10
timesteps = 3
batch_size = 32
def data_gen():
while True:
numerical = np.random.randint(5, size=(batch_size, timesteps, no_of_features))
y = np.random.randint(2, size=batch_size)
w = np.ones(batch_size) * WEIGHT_VARIABLE
yield {'numeric_input': numerical}, y, w
def build_model():
numerical_input = tf.keras.layers.Input(shape=(timesteps, no_of_features), name='numeric_input')
rnn_out = tf.keras.layers.GRU(32, return_sequences=False)(numerical_input)
dense = tf.keras.layers.Dense(1, activation='sigmoid', name='main_output')(rnn_out)
model = tf.keras.models.Model(numerical_input, dense)
params = {
'loss': 'binary_crossentropy',
'optimizer': tf.keras.optimizers.Adam(),
'metrics': [tf.keras.metrics.binary_crossentropy, tf.keras.metrics.binary_accuracy]
}
model.compile(**params)
return model
def train_model():
gen1 = data_gen()
model = build_model()
model.fit_generator(gen1, epochs=30, steps_per_epoch=10)
if __name__ == '__main__':
train_model()
В приведенном выше коде вам просто нужно изменить WEIGHT_VARIABLE = 1 с 1 на 100000 и перезапустить файл.
Журналы
v1.10
WEIGHT_VARIABLE = 1
Epoch 1/5 10/10 [==============================] -
1s 128ms/step - loss: 0.7407 - binary_crossentropy: 0.7407 - binary_accuracy: 0.5031
Epoch 2/5 10/10 [==============================] -
0s 4ms/step - loss: 0.7043 - binary_crossentropy: 0.7043 - binary_accuracy: 0.5125
Epoch 3/5 10/10 [==============================] -
0s 4ms/step - loss: 0.7055 - binary_crossentropy: 0.7055 - binary_accuracy: 0.5219
Epoch 4/5 10/10 [==============================] -
0s 4ms/step - loss: 0.7002 - binary_crossentropy: 0.7002 - binary_accuracy: 0.5250
Epoch 5/5 10/10 [==============================] -
0s 4ms/step - loss: 0.6944 - binary_crossentropy: 0.6944 - binary_accuracy: 0.5375
WEIGHT_VARIABLE = 10000
Epoch 1/5 10/10 [==============================] -
1s 131ms/step - loss: 7235.5976 - binary_crossentropy: 0.7236 - binary_accuracy: 0.4562
Epoch 2/5 10/10 [==============================] -
0s 4ms/step - loss: 7271.9184 - binary_crossentropy: 0.7272 - binary_accuracy: 0.4844
Epoch 3/5 10/10 [==============================] -
0s 4ms/step - loss: 7276.9147 - binary_crossentropy: 0.7277 - binary_accuracy: 0.4500
Epoch 4/5 10/10 [==============================] -
0s 4ms/step - loss: 7052.0121 - binary_crossentropy: 0.7052 - binary_accuracy: 0.4625
Epoch 5/5 10/10 [==============================] -
0s 4ms/step - loss: 7187.0285 - binary_crossentropy: 0.7187 - binary_accuracy: 0.4969
v1.12
WEIGHT_VARIABLE = 1
Epoch 1/5 10/10 [==============================] -
1s 68ms/step - loss: 0.7188 - binary_crossentropy: 0.7188 - binary_accuracy: 0.5312
Epoch 2/5 10/10 [==============================] -
0s 4ms/step - loss: 0.7044 - binary_crossentropy: 0.7044 - binary_accuracy: 0.4969
Epoch 3/5 10/10 [==============================] -
0s 4ms/step - loss: 0.7086 - binary_crossentropy: 0.7086 - binary_accuracy: 0.4844
Epoch 4/5 10/10 [==============================] -
0s 4ms/step - loss: 0.7075 - binary_crossentropy: 0.7075 - binary_accuracy: 0.4500
Epoch 5/5 10/10 [==============================] -
0s 4ms/step - loss: 0.6950 - binary_crossentropy: 0.6950 - binary_accuracy: 0.5187
WEIGHT_VARIABLE = 10000
Epoch 1/5 10/10 [==============================] -
1s 74ms/step - loss: 0.9084 - binary_crossentropy: 0.9084 - binary_accuracy: 0.4719
Epoch 2/5 10/10 [==============================] -
0s 4ms/step - loss: 0.7120 - binary_crossentropy: 0.7120 - binary_accuracy: 0.5062
Epoch 3/5 10/10 [==============================] -
0s 4ms/step - loss: 0.7024 - binary_crossentropy: 0.7024 - binary_accuracy: 0.5344
Epoch 4/5 10/10 [==============================] -
0s 4ms/step - loss: 0.7257 - binary_crossentropy: 0.7257 - binary_accuracy: 0.4500
Epoch 5/5 10/10 [==============================] -
0s 4ms/step - loss: 0.7013 - binary_crossentropy: 0.7013 - binary_accuracy: 0.4844
Ссылка на выпуск Github