GradienTape 收敛比 Keras.model.fit 慢得多

GradienTape convergence much slower than Keras.model.fit

我目前正在尝试获取 TF2.0 api,但当我将 GradientTape 与常规 [=33] 进行比较时=]keras.Model.fit 我注意到:

  1. 它 运行 较慢(可能是由于 Eager Execution)

  2. 它收敛得更慢(我不确定为什么)。

+--------+--------------+--------------+------------------+
|  Epoch | GradientTape | GradientTape | keras.Model.fit  |
|        |              |  shuffling   |                  |
+--------+--------------+--------------+------------------+
|    1   |     0.905    |     0.918    |      0.8793      |
+--------+--------------+--------------+------------------+
|    2   |     0.352    |     0.634    |      0.2226      |
+--------+--------------+--------------+------------------+
|    3   |     0.285    |     0.518    |      0.1192      |
+--------+--------------+--------------+------------------+
|    4   |     0.282    |     0.458    |      0.1029      |
+--------+--------------+--------------+------------------+
|    5   |     0.275    |     0.421    |      0.0940      |
+--------+--------------+--------------+------------------+

这是我在 GradientTape 中使用的训练循环:


optimizer = keras.optimizers.Adam()
glove_model = GloveModel(vocab_size=len(labels))
train_loss = keras.metrics.Mean(name='train_loss')

@tf.function
def train_step(examples, labels):
    with tf.GradientTape() as tape:
        predictions = glove_model(examples)
        loss = glove_model.glove_loss(labels, predictions)

    gradients = tape.gradient(loss, glove_model.trainable_variables)
    optimizer.apply_gradients(zip(gradients, glove_model.trainable_variables))

    train_loss(loss)



total_step = 0
for epoch in range(epochs_number):

    pbar = tqdm(train_ds.enumerate(), total=int(len(index_data) / batch_size) + 1)

    for ix, (examples, labels) in pbar:

        train_step(examples, labels)


    print(f"Epoch {epoch + 1}, Loss {train_loss.result()}")

    # Reset the metrics for the next epoch
    train_loss.reset_states()

这里是 Keras.Model.fit 培训:

glove_model.compile(optimizer, glove_model.glove_loss)
glove_model.fit(train_ds, epochs=epochs_number)

这是tf.data.Dataset来源

train_ds = data.Dataset.from_tensor_slices(
    (np.hstack([index_rows.reshape(-1, 1), index_cols.reshape(-1, 1)]), index_data)
).shuffle(100000).batch(batch_size, drop_remainder=True)

这是模型。

class GloveModel(keras.Model):

    def __init__(self, vocab_size, dim=100, a=3/4, x_max=100):
        super(GloveModel, self).__init__()

        self.vocab_size = vocab_size
        self.dim = dim
        self.a = a
        self.x_max = x_max

        self.target_embedding = layers.Embedding(
            input_dim=self.vocab_size, output_dim=self.dim, input_length=1, name="target_embedding"
        )
        self.target_bias = layers.Embedding(
            input_dim=self.vocab_size, output_dim=1, input_length=1, name="target_bias"
        )

        self.context_embedding = layers.Embedding(
            input_dim=self.vocab_size, output_dim=self.dim, input_length=1, name="context_embedding"
        )
        self.context_bias = layers.Embedding(
            input_dim=self.vocab_size, output_dim=1, input_length=1, name="context_bias"
        )

        self.dot_product = layers.Dot(axes=-1, name="dot")

        self.prediction = layers.Add(name="add")
        self.step = 0

    def call(self, inputs):

        target_ix = inputs[:, 0]
        context_ix = inputs[:, 1]

        target_embedding = self.target_embedding(target_ix)
        target_bias = self.target_bias(target_ix)

        context_embedding = self.context_embedding(context_ix)
        context_bias = self.context_bias(context_ix)

        dot_product = self.dot_product([target_embedding, context_embedding])
        prediction = self.prediction([dot_product, target_bias, context_bias])

        return prediction

    def glove_loss(self, y_true, y_pred):

        weight = tf.math.minimum(
            tf.math.pow(y_true/self.x_max, self.a), 1.0
        )
        loss_value = tf.math.reduce_mean(weight * tf.math.pow(y_pred - tf.math.log(y_true), 2.0))

        return loss_value



我尝试了多种配置和优化器,但似乎没有任何改变收敛速度。

Dataset.shuffle() 只打乱每个 minibatch,所以每个 epoch 都有相同的顺序。 Keras .fit() 使用一些魔法在每个纪元之前对整个数据集进行洗牌。要在 TF 中执行此操作,您需要使用数据集 .repeat(epochs_number).shuffle(..., reshuffle_each_iteration=True):

train_ds = data.Dataset.from_tensor_slices(
    (np.hstack([index_rows.reshape(-1, 1), index_cols.reshape(-1, 1)]), index_data)
    ).shuffle(100000, reshuffle_each_iteration=True
    ).batch(batch_size, drop_remainder=True
    ).repeat(epochs_number)

for ix, (examples, labels) in train_ds.enumerate():
    train_step(examples, labels)
    current_epoch = ix // (len(index_data) // batch_size)

这个解决方法既不美观也不自然,目前您可以使用它来随机播放每个 epoch。这是一个已知问题,将得到修复,将来您可以使用 for epoch in range(epochs_number) 而不是 .repeat()

问题来自使用 tf.Dataset 方法的 shuffling。它当时只在数据集中洗牌了一个桶。使用 Keras.Model.fit 会产生更好的结果,因为它可能会增加另一个洗牌。

我用 numpy.random.shuffle 添加了一个改组,它提高了两种训练方法的性能:

现在生成的数据集是:

numpy_data = np.hstack([index_rows.reshape(-1, 1), index_cols.reshape(-1, 1), index_data.reshape(-1, 1)])

np.random.shuffle(numpy_data)

indexes = np.array(numpy_data[:, :2], dtype=np.uint32)
labels = np.array(numpy_data[:, 2].reshape(-1, 1), dtype=np.float32)

train_ds = data.Dataset.from_tensor_slices(
    (indexes, labels)
).shuffle(100000).batch(batch_size, drop_remainder=True)

结果是:

+--------+--------------+------------------+
|  Epoch | GradientTape |  keras.Model.fit |
+--------+--------------+------------------+
|    1   |     0.294    |      0.294       |
+--------+--------------+------------------+
|    2   |     0.111    |      0.110       |
+--------+--------------+------------------+
|    3   |     0.089    |      0.089       |
+--------+--------------+------------------+
|    4   |     0.074    |      0.075       |
+--------+--------------+------------------+
|    5   |     0.063    |      0.063       |
+--------+--------------+------------------+

每个 epoch 的训练类型大致相同,每个 epoch 2 分钟