Tensorflow 2:获取“警告：Tensorflow:对<function>的最近9次调用中有9次触发了tf.function retracting。跟踪代价高昂”

Question

我认为这个错误来自于形状的问题，但我不知道在哪里。完整的错误消息建议执行以下操作：

Also, tf.function has experimental_relax_shapes=True option that relaxes argument shapes that can avoid unnecessary retracing.

当我在函数decorator中输入此参数时，它确实起作用

@tf.function(experimental_relax_shapes=True)

原因可能是什么？以下是完整的代码：

import os
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
import tensorflow as tf
print(f'Tensorflow version {tf.__version__}')
from tensorflow import keras
from tensorflow.keras.layers import Dense, Conv1D, GlobalAveragePooling1D, Embedding
import tensorflow_datasets as tfds
from tensorflow.keras.models import Model

(train_data, test_data), info = tfds.load('imdb_reviews/subwords8k',
                                          split=[tfds.Split.TRAIN, tfds.Split.TEST],
                                          as_supervised=True, with_info=True)

padded_shapes = ([None], ())

train_dataset = train_data.shuffle(25000).\
    padded_batch(padded_shapes=padded_shapes, batch_size=16)
test_dataset = test_data.shuffle(25000).\
    padded_batch(padded_shapes=padded_shapes, batch_size=16)

n_words = info.features['text'].encoder.vocab_size


class ConvModel(Model):
    def __init__(self):
        super(ConvModel, self).__init__()
        self.embe = Embedding(n_words, output_dim=16)
        self.conv = Conv1D(32, kernel_size=6, activation='elu')
        self.glob = GlobalAveragePooling1D()
        self.dens = Dense(2)

    def call(self, x, training=None, mask=None):
        x = self.embe(x)
        x = self.conv(x)
        x = self.glob(x)
        x = self.dens(x)
        return x


conv = ConvModel()

conv(next(iter(train_dataset))[0])

loss_object = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)

train_loss = tf.keras.metrics.Mean()
test_loss = tf.keras.metrics.Mean()

train_acc = tf.keras.metrics.CategoricalAccuracy()
test_acc = tf.keras.metrics.CategoricalAccuracy()

optimizer = tf.keras.optimizers.Adam(learning_rate=1e-3)


@tf.function
def train_step(inputs, labels):
    with tf.GradientTape() as tape:
        logits = conv(inputs, training=True)
        loss = loss_object(labels, logits)
        train_loss(loss)
        train_acc(logits, labels)

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


@tf.function
def test_step(inputs, labels):
    logits = conv(inputs, training=False)
    loss = loss_object(labels, logits)
    test_loss(loss)
    test_acc(logits, labels)


def learn():
    train_loss.reset_states()
    test_loss.reset_states()
    train_acc.reset_states()
    test_acc.reset_states()

    for text, target in train_dataset:
        train_step(inputs=text, labels=target)

    for text, target in test_dataset:
        test_step(inputs=text, labels=target)


def main(epochs=2):
    for epoch in tf.range(1, epochs + 1):
        learn()
        template = 'TRAIN LOSS {:>5.3f} TRAIN ACC {:.2f} TEST LOSS {:>5.3f} TEST ACC {:.2f}'

        print(template.format(
            train_loss.result(),
            train_acc.result(),
            test_loss.result(),
            test_acc.result()
        ))

if __name__ == '__main__':
    main(epochs=1)

Answer 1

TF/DR：此错误的根本原因是train_data的形状因批次而异。固定train_data的大小/形状可以解决此跟踪警告。我更改了下面的行，然后一切都按预期进行。全部要点是here

padded_shapes = ([9000], ())#None.

详细信息：

如警告消息中所述

WARNING:tensorflow:10 out of the last 11 calls to <function train_step at 0x7f4825f6d400> triggered tf.function retracing. Tracing is expensive and the excessive number of tracings could be due to (1) creating @tf.function repeatedly in a loop, (2) passing tensors with different shapes, (3) passing Python objects instead of tensors. For (1), please define your @tf.function outside of the loop. For (2), @tf.function has experimental_relax_shapes=True option that relaxes argument shapes that can avoid unnecessary retracing.

发生此回溯警告是因为警告消息中提到的三个原因。原因（1）不是根本原因，因为@tf.function没有在循环中调用，原因（3）也不是根本原因，因为train_step和test_step的参数都是张量对象。因此，根本原因是警告中提到的原因（2）

当我打印train_data的大小时，它会打印不同的大小。所以我试着填充train_data，这样所有批次的形状都是一样的

 padded_shapes = ([9000], ())#None.  # this line throws tracing error as the shape of text is varying for each step in an epoch.
    # as the data size is varying, tf.function will start retracing it
    # For the demonstration, I used 9000 as max length, but please change it accordingly