我正在尝试在 Tensorflow 2.0.0alpha0 中制作百度的Deep Speech 2模型。我在使用ctc_loss对象计算梯度时无法优化Tensorflow tf.GradientTape()。
我目前正在将形状为(batch_size, max_step, feats)的张量传递给我的模型，然后将计算出的对数传递给损失函数。我也尝试过传递稀疏张量，但这也行不通。
这是创建模型的代码
import tensorflow as tf
class DeepSpeech2(tf.keras.Model):
def __init__(self, vocab_size, conv_filters=[11], conv_kernel_sizes=[1280], conv_strides=[2],
recur_sizes=[100], rnn_type='gru', bidirect_rnn=False, batch_norm=True,
learning_rate=1e-3, name='DeepSpeech2'):
super(DeepSpeech2, self).__init__()
self._vocab_size = vocab_size
self._conv_filters = conv_filters
self._conv_kernel_sizes = conv_kernel_sizes
self._conv_strides = conv_strides
self._recur_sizes = recur_sizes
self._rnn_type = rnn_type
self._bidirect_rnn = bidirect_rnn
self._batch_norm = batch_norm
self._learning_rate = learning_rate
self._name = name
self._conv_batch_norm = None
with tf.name_scope(self._name):
self._convolution = [tf.keras.layers.Conv1D(filters=conv_filters[i],
kernel_size=conv_kernel_sizes[i], strides=conv_strides[i],
padding='valid', activation='relu',
name='conv1d_{}'.format(i)) for i in range(len(self._conv_filters))]
if self._batch_norm:
self._conv_batch_norm = tf.keras.layers.BatchNormalization(name='bn_conv_1d')
if self._rnn_type == 'gru':
rnn_init = tf.keras.layers.GRU
elif self._rnn_type == 'lstm':
rnn_init = tf.keras.layers.LSTM
else:
raise Exception("Invalid rnn_type: '{}' (must be 'lstm' or 'gru')"
.format(self._rnn_type))
self._rnn = []
for i, r in enumerate(self._recur_sizes):
layer = rnn_init(r, activation='relu', return_sequences=True,
name='{}_{}'.format(self._rnn_type, i))
if self._bidirect_rnn:
layer = tf.keras.layers.Bidirectional(layer)
self._rnn.append(layer)
if self._batch_norm:
self._rnn.append(tf.keras.layers.BatchNormalization())
self._fc = tf.keras.layers.TimeDistributed(tf.keras.layers.Dense(
self._vocab_size, name='fc', activation='linear'))
self._optimizer = tf.keras.optimizers.Adam(lr=self._learning_rate)
def __call__(self, specs):
with tf.name_scope(self._name):
feats = specs
for layer in self._convolution:
feats = layer(feats)
if self._conv_batch_norm:
feats = self._conv_batch_norm(feats)
rnn_outputs = feats
for layer in self._rnn:
rnn_outputs = layer(rnn_outputs)
outputs = self._fc(rnn_outputs)
return tf.transpose(outputs, (1, 0, 2))
@tf.function
def train_step(self, specs, spec_lengths, labels, label_lengths):
with tf.GradientTape() as tape:
logits = self.__call__(specs)
loss = tf.nn.ctc_loss(labels=labels, logits=logits,
label_length=label_lengths, logit_length=spec_lengths)
cost = tf.reduce_sum(loss)
decoded, neg_sum_logits = tf.nn.ctc_greedy_decoder(logits, label_lengths)
gradients = tape.gradient(cost, self.trainable_variables)
self._optimizer.apply_gradients(zip(gradients, self.trainable_variables))
return (decoded[0].indices, decoded[0].values, decoded[0].dense_shape), cost
我当前遇到以下错误
ValueError: No gradients provided for any variable: ['DeepSpeech2/conv1d_0/kernel:0', 'DeepSpeech2/conv1d_0/bias:0', 'DeepSpeech2/bn_conv_1d/gamma:0', 'DeepSpeech2/bn_conv_1d/beta:0', 'DeepSpeech2/gru_0/kernel:0', 'DeepSpeech2/gru_0/recurrent_kernel:0', 'DeepSpeech2/gru_0/bias:0', 'DeepSpeech2/batch_normalization_v2/gamma:0', 'DeepSpeech2/batch_normalization_v2/beta:0', 'DeepSpeech2/time_distributed/kernel:0', 'DeepSpeech2/time_distributed/bias:0'].
在将梯度应用于优化器的行上发生错误。当我打印出gradients变量时，它只是None
据我了解，此错误表示图中没有从变量到损失的路径，但我不确定为什么会得到这个。任何帮助将不胜感激！