tensorflow2 自定義損失函式使用的隱藏坑

geeks_reign 發表於 2021-07-26

Keras的核心原則是逐步揭示覆雜性,可以在保持相應的高階便利性的同時,對操作細節進行更多控制。當我們要自定義fit中的訓練演算法時,可以重寫模型中的train_step方法,然後呼叫fit來訓練模型。

這裡以tensorflow2官網中的例子來說明:

import numpy as np
import tensorflow as tf
from tensorflow import keras
x = np.random.random((1000, 32))
y = np.random.random((1000, 1))
class CustomModel(keras.Model):
    tf.random.set_seed(100)
    def train_step(self, data):
        # Unpack the data. Its structure depends on your model and
        # on what you pass to `fit()`.
        x, y = data

        with tf.GradientTape() as tape:
            y_pred = self(x, training=True)  # Forward pass
            # Compute the loss value
            # (the loss function is configured in `compile()`)
            loss = self.compiled_loss(y, y_pred, regularization_losses=self.losses)

        # Compute gradients
        trainable_vars = self.trainable_variables
        gradients = tape.gradient(loss, trainable_vars)
        # Update weights
        self.optimizer.apply_gradients(zip(gradients, trainable_vars))
        # Update metrics (includes the metric that tracks the loss)
        self.compiled_metrics.update_state(y, y_pred)
        # Return a dict mapping metric names to current value
        return {m.name: m.result() for m in self.metrics}
    


# Construct and compile an instance of CustomModel
inputs = keras.Input(shape=(32,))
outputs = keras.layers.Dense(1)(inputs)
model = CustomModel(inputs, outputs)
model.compile(optimizer="adam", loss=tf.losses.MSE, metrics=["mae"])

# Just use `fit` as usual

model.fit(x, y, epochs=1, shuffle=False)
32/32 [==============================] - 0s 1ms/step - loss: 0.2783 - mae: 0.4257

<tensorflow.python.keras.callbacks.History at 0x7ff7edf6dfd0>

這裡的loss是tensorflow庫中實現了的損失函式,如果想自定義損失函式,然後將損失函式傳入model.compile中,能正常按我們預想的work嗎?

答案竟然是否定的,而且沒有錯誤提示,只是loss計算不會符合我們的預期。

def custom_mse(y_true, y_pred):
    return tf.reduce_mean((y_true - y_pred)**2, axis=-1)
a_true = tf.constant([1., 1.5, 1.2])
a_pred = tf.constant([1., 2, 1.5])
custom_mse(a_true, a_pred)
<tf.Tensor: shape=(), dtype=float32, numpy=0.11333332>
tf.losses.MSE(a_true, a_pred)
<tf.Tensor: shape=(), dtype=float32, numpy=0.11333332>

以上結果證實了我們自定義loss的正確性,下面我們直接將自定義的loss置入compile中的loss引數中,看看會發生什麼。

my_model = CustomModel(inputs, outputs)
my_model.compile(optimizer="adam", loss=custom_mse, metrics=["mae"])
my_model.fit(x, y, epochs=1, shuffle=False)
32/32 [==============================] - 0s 820us/step - loss: 0.1628 - mae: 0.3257

<tensorflow.python.keras.callbacks.History at 0x7ff7edeb7810>

我們看到,這裡的loss與我們與標準的tf.losses.MSE明顯不同。這說明我們自定義的loss以這種方式直接傳遞進model.compile中,是完全錯誤的操作。

正確運用自定義loss的姿勢是什麼呢?下面揭曉。

loss_tracker = keras.metrics.Mean(name="loss")
mae_metric = keras.metrics.MeanAbsoluteError(name="mae")

class MyCustomModel(keras.Model):
    tf.random.set_seed(100)
    def train_step(self, data):
        # Unpack the data. Its structure depends on your model and
        # on what you pass to `fit()`.
        x, y = data

        with tf.GradientTape() as tape:
            y_pred = self(x, training=True)  # Forward pass
            # Compute the loss value
            # (the loss function is configured in `compile()`)
            loss = custom_mse(y, y_pred)
            # loss += self.losses

        # Compute gradients
        trainable_vars = self.trainable_variables
        gradients = tape.gradient(loss, trainable_vars)
        # Update weights
        self.optimizer.apply_gradients(zip(gradients, trainable_vars))
        
        # Compute our own metrics
        loss_tracker.update_state(loss)
        mae_metric.update_state(y, y_pred)
        return {"loss": loss_tracker.result(), "mae": mae_metric.result()}
    
    @property
    def metrics(self):
        # We list our `Metric` objects here so that `reset_states()` can be
        # called automatically at the start of each epoch
        # or at the start of `evaluate()`.
        # If you don't implement this property, you have to call
        # `reset_states()` yourself at the time of your choosing.
        return [loss_tracker, mae_metric]
    
# Construct and compile an instance of CustomModel
inputs = keras.Input(shape=(32,))
outputs = keras.layers.Dense(1)(inputs)
my_model_beta = MyCustomModel(inputs, outputs)
my_model_beta.compile(optimizer="adam")

# Just use `fit` as usual

my_model_beta.fit(x, y, epochs=1, shuffle=False)
32/32 [==============================] - 0s 960us/step - loss: 0.2783 - mae: 0.4257

<tensorflow.python.keras.callbacks.History at 0x7ff7eda3d810>

終於,通過跳過在 compile() 中傳遞損失函式,而在 train_step 中手動完成所有計算內容,我們獲得了與之前預設tf.losses.MSE完全一致的輸出,這才是我們想要的結果。

總結一下,當我們在模型中想用自定義的損失函式,不能直接傳入fit函式,而是需要在train_step中手動傳入,完成計算過程。