python - Tensorflow 类型错误 : Using a `tf.Tensor` as a Python `bool` is not allowed.

Question

我正在使用 tensorflow 训练“显示和讲述”模型，其中模型自动生成图像的标题。我怎么会收到这个错误。

这是回溯:

TypeError                                 Traceback (most recent call 
last)
<ipython-input-15-b6da0a27b701> in <module>()
  1 try:
  2     #train(.001,False,False) #train from scratch
----> 3     train(.001,True,True)    #continue training from pretrained weights @epoch500
  4     #train(.001)  #train from previously saved weights
  5 except KeyboardInterrupt:

<ipython-input-14-39693d0edd0a> in train(learning_rate, continue_training, transfer)
 23     n_words = len(wordtoix)
 24     maxlen = np.max( [x for x in map(lambda x: len(x.split(' ')), captions) ] )
---> 25     caption_generator = Caption_Generator(dim_in, dim_hidden, dim_embed, batch_size, maxlen+2, n_words, init_b)
 26 
 27     loss, image, sentence, mask = caption_generator.build_model()

<ipython-input-12-7ef491a16183> in __init__(self, dim_in, dim_embed, dim_hidden, batch_size, n_lstm_steps, n_words, init_b)
 11         # declare the variables to be used for our word embeddings
 12         with tf.device("/cpu:0"):
---> 13             self.word_embedding = tf.get_variable("word_embedding", tf.random_uniform([self.n_words, self.dim_embed], -0.1, 0.1))
 14 
 15             self.embedding_bias = tf.get_variable("embedding_bias", tf.zeros([dim_embed]))

/home/niraj/anaconda2/lib/python2.7/site-packages/tensorflow/python/ops/variable_scope.pyc in get_variable(name, shape, dtype, initializer, regularizer, trainable, collections, caching_device, partitioner, validate_shape, use_resource, custom_getter)
1063       collections=collections, caching_device=caching_device,
1064       partitioner=partitioner, validate_shape=validate_shape,
-> 1065       use_resource=use_resource, custom_getter=custom_getter)
1066 get_variable_or_local_docstring = (
1067     """%s

/home/niraj/anaconda2/lib/python2.7/site-packages/tensorflow/python/ops/variable_scope.pyc in get_variable(self, var_store, name, shape, dtype, initializer, regularizer, reuse, trainable, collections, caching_device, partitioner, validate_shape, use_resource, custom_getter)
960           collections=collections, caching_device=caching_device,
961           partitioner=partitioner, validate_shape=validate_shape,
--> 962           use_resource=use_resource, custom_getter=custom_getter)
963 
964   def _get_partitioned_variable(self,

/home/niraj/anaconda2/lib/python2.7/site-packages/tensorflow/python/ops/variable_scope.pyc in get_variable(self, name, shape, dtype, initializer, regularizer, reuse, trainable, collections, caching_device, partitioner, validate_shape, use_resource, custom_getter)
365           reuse=reuse, trainable=trainable, collections=collections,
366           caching_device=caching_device, partitioner=partitioner,
--> 367           validate_shape=validate_shape, use_resource=use_resource)
368 
369   def _get_partitioned_variable(

/home/niraj/anaconda2/lib/python2.7/site-packages/tensorflow/python/ops/variable_scope.pyc in _true_getter(name, shape, dtype, initializer, regularizer, reuse, trainable, collections, caching_device, partitioner, validate_shape, use_resource)
301                      trainable=True, collections=None, caching_device=None,
302                      partitioner=None, validate_shape=True, use_resource=None):
--> 303       is_scalar = shape is not None and not shape
304       # Partitioned variable case
305       if partitioner is not None and not is_scalar:

/home/niraj/anaconda2/lib/python2.7/site-packages/tensorflow/python/framework/ops.pyc in __nonzero__(self)
511       `TypeError`.
512     """
--> 513     raise TypeError("Using a `tf.Tensor` as a Python `bool` is not allowed. "
514                     "Use `if t is not None:` instead of `if t:` to test if a "
515                     "tensor is defined, and use TensorFlow ops such as "

TypeError:不允许将 tf.Tensor 用作 Python bool。使用 if t is not None: 而不是 if t: 来测试是否定义了张量，并使用 TensorFlow ops 如 tf.cond 来执行以值为条件的子图张量的。

代码如下:

def preProBuildWordVocab(sentence_iterator, word_count_threshold=30): # function from Andre Karpathy's NeuralTalk
print('preprocessing %d word vocab' % (word_count_threshold, ))
word_counts = {}
nsents = 0
for sent in sentence_iterator:
  nsents += 1
  for w in sent.lower().split(' '):
    word_counts[w] = word_counts.get(w, 0) + 1
vocab = [w for w in word_counts if word_counts[w] >= word_count_threshold]
print('preprocessed words %d -> %d' % (len(word_counts), len(vocab)))


ixtoword = {}
ixtoword[0] = '.'  
wordtoix = {}
wordtoix['#START#'] = 0 
ix = 1
for w in vocab:
  wordtoix[w] = ix
  ixtoword[ix] = w
  ix += 1

word_counts['.'] = nsents
bias_init_vector = np.array([1.0*word_counts[ixtoword[i]] for i in ixtoword])
bias_init_vector /= np.sum(bias_init_vector) 
bias_init_vector = np.log(bias_init_vector)
bias_init_vector -= np.max(bias_init_vector) 
return wordtoix, ixtoword, bias_init_vector.astype(np.float32)

class Caption_Generator():
def __init__(self, dim_in, dim_embed, dim_hidden, batch_size, n_lstm_steps, n_words, init_b):

    self.dim_in = dim_in
    self.dim_embed = dim_embed
    self.dim_hidden = dim_hidden
    self.batch_size = batch_size
    self.n_lstm_steps = n_lstm_steps
    self.n_words = n_words

    # declare the variables to be used for our word embeddings
    with tf.device("/cpu:0"):
        self.word_embedding = tf.get_variable("word_embedding", tf.random_uniform([self.n_words, self.dim_embed], -0.1, 0.1))

        self.embedding_bias = tf.get_variable("embedding_bias", tf.zeros([dim_embed]))

    # declare the LSTM itself
        self.lstm = tf.contrib.rnn.BasicLSTMCell(dim_hidden)

    # declare the variables to be used to embed the image feature embedding to the word embedding space
        self.img_embedding = tf.get_variable("img_embedding", tf.random_uniform([dim_in, dim_hidden], -0.1, 0.1))
        self.img_embedding_bias = tf.get_variable("img_embedding_bias", tf.zeros([dim_hidden]))

    # declare the variables to go from an LSTM output to a word encoding output
        self.word_encoding = tf.get_variable("word_encoding", tf.random_uniform([dim_hidden, n_words], -0.1, 0.1))
    # initialize this bias variable from the preProBuildWordVocab output
        self.word_encoding_bias = tf.get_variable("word_encoding_bias", init_b)

def build_model(self):
    # declaring the placeholders for our extracted image feature vectors, our caption, and our mask
    # (describes how long our caption is with an array of 0/1 values of length `maxlen`  
    img = tf.placeholder(tf.float32, [self.batch_size, self.dim_in])
    caption_placeholder = tf.placeholder(tf.int32, [self.batch_size, self.n_lstm_steps])
    mask = tf.placeholder(tf.float32, [self.batch_size, self.n_lstm_steps])

    # getting an initial LSTM embedding from our image_imbedding
    image_embedding = tf.matmul(img, self.img_embedding) + self.img_embedding_bias

    # setting initial state of our LSTM
    state = self.lstm.zero_state(self.batch_size, dtype=tf.float32)

    total_loss = 0.0
    with tf.variable_scope("RNN"):
        for i in range(self.n_lstm_steps): 
            if i > 0:
               #if this isn’t the first iteration of our LSTM we need to get the word_embedding corresponding
               # to the (i-1)th word in our caption 
                with tf.device("/cpu:0"):
                    current_embedding = tf.nn.embedding_lookup(self.word_embedding, caption_placeholder[:,i-1]) + self.embedding_bias
            else:
                 #if this is the first iteration of our LSTM we utilize the embedded image as our input 
                current_embedding = image_embedding
            if i > 0: 
                # allows us to reuse the LSTM tensor variable on each iteration
                tf.get_variable_scope().reuse_variables()

                out, state = self.lstm(current_embedding, state)
                    #out, state = self.tf.nn.dynamic_rnn(current_embedding, state)


            if i > 0:
                #get the one-hot representation of the next word in our caption 
                labels = tf.expand_dims(caption_placeholder[:, i], 1)
                ix_range=tf.range(0, self.batch_size, 1)
                ixs = tf.expand_dims(ix_range, 1)
                concat = tf.concat([ixs, labels],1)
                onehot = tf.sparse_to_dense(
                concat, tf.stack([self.batch_size, self.n_words]), 1.0, 0.0)


                #perform a softmax classification to generate the next word in the caption
                logit = tf.matmul(out, self.word_encoding) + self.word_encoding_bias
                xentropy = tf.nn.softmax_cross_entropy_with_logits(logits=logit, labels=onehot)
                xentropy = xentropy * mask[:,i]

                loss = tf.reduce_sum(xentropy)
                total_loss += loss

        total_loss = total_loss / tf.reduce_sum(mask[:,1:])
        return total_loss, img,  caption_placeholder, mask

### Parameters ###
dim_embed = 256
dim_hidden = 256
dim_in = 4096
batch_size = 128
momentum = 0.9
n_epochs = 150

def train(learning_rate=0.001, continue_training=False, transfer=True):

tf.reset_default_graph()

feats, captions = get_data(annotation_path, feature_path)
wordtoix, ixtoword, init_b = preProBuildWordVocab(captions)

np.save('data/ixtoword', ixtoword)

index = (np.arange(len(feats)).astype(int))
np.random.shuffle(index)


sess = tf.InteractiveSession()
n_words = len(wordtoix)
maxlen = np.max( [x for x in map(lambda x: len(x.split(' ')), captions) ] )
caption_generator = Caption_Generator(dim_in, dim_hidden, dim_embed, batch_size, maxlen+2, n_words, init_b)

loss, image, sentence, mask = caption_generator.build_model()

saver = tf.train.Saver(max_to_keep=100)
global_step=tf.Variable(0,trainable=False)
learning_rate = tf.train.exponential_decay(learning_rate, global_step,
                                   int(len(index)/batch_size), 0.95)
train_op = tf.train.AdamOptimizer(learning_rate).minimize(loss)
tf.global_variables_initializer().run()

if continue_training:
    if not transfer:
        saver.restore(sess,tf.train.latest_checkpoint(model_path))
    else:
        saver.restore(sess,tf.train.latest_checkpoint(model_path_transfer))
losses=[]
for epoch in range(n_epochs):
    for start, end in zip( range(0, len(index), batch_size), range(batch_size, len(index), batch_size)):

        current_feats = feats[index[start:end]]
        current_captions = captions[index[start:end]]
        current_caption_ind = [x for x in map(lambda cap: [wordtoix[word] for word in cap.lower().split(' ')[:-1] if word in wordtoix], current_captions)]

        current_caption_matrix = sequence.pad_sequences(current_caption_ind, padding='post', maxlen=maxlen+1)
        current_caption_matrix = np.hstack( [np.full( (len(current_caption_matrix),1), 0), current_caption_matrix] )

        current_mask_matrix = np.zeros((current_caption_matrix.shape[0], current_caption_matrix.shape[1]))
        nonzeros = np.array([x for x in map(lambda x: (x != 0).sum()+2, current_caption_matrix )])

        for ind, row in enumerate(current_mask_matrix):
            row[:nonzeros[ind]] = 1

        _, loss_value = sess.run([train_op, loss], feed_dict={
            image: current_feats.astype(np.float32),
            sentence : current_caption_matrix.astype(np.int32),
            mask : current_mask_matrix.astype(np.float32)
            })

        print("Current Cost: ", loss_value, "\t Epoch {}/{}".format(epoch, n_epochs), "\t Iter {}/{}".format(start,len(feats)))
    print("Saving the model from epoch: ", epoch)
    saver.save(sess, os.path.join(model_path, 'model'), global_step=epoch)

Answer 1

问题源于将 tf.Tensor 作为 tf.get_variable(name, shape=None, ...) 的 shape 参数传递在这条线上:

self.word_embedding = tf.get_variable("word_embedding", tf.random_uniform([self.n_words, self.dim_embed], -0.1, 0.1))

我怀疑您打算将随机张量作为 initializer 参数传递。解决这个问题的最简单方法是为参数指定一个名称:

self.word_embedding = tf.get_variable(
    "word_embedding",
    initializer=tf.random_uniform([self.n_words, self.dim_embed], -0.1, 0.1))

看起来您对 tf.get_variable() 的所有调用都需要类似的修复。