将 Keras 的后端从 Theano 更改为 Tensorflow 时如何避免形状错误？

Question

我尝试使用 Tensorflow 后端在 Keras 上使用 Deep Semantic Similarity Model(DSSM): https://github.com/airalcorn2/Deep-Semantic-Similarity-Model/blob/master/deep_semantic_similarity_keras.py。该代码适用于使用 Theano 后端的 Keras，并且可以正常工作。

为了让它在使用 Tensorflow 后端的 Keras 上正确运行，我将一些 Tensor 属性修改为：

-: backend.dot(x, backend.transpose(y)) / (x.norm(2) * y.norm(2)) 
+: backend.dot(backend.l2_normalize(x, 0), backend.transpose(backend.l2_normalize(y, 0)))  

和

-: x.max(axis = 1) 
+: backend.max(x, 1)

当我编译修改后的模型时，ValueError: logits and targets must have the same shape (() vs (?, ?))发生。如何避免错误？

这是修改后的代码

import keras
import numpy as np

from keras import backend
from keras.layers import Input, merge
from keras.layers.core import Dense, Lambda, Reshape
from keras.layers.convolutional import Convolution1D
from keras.models import Model

def R(vects):
    (x, y) = vects
    return backend.dot(backend.l2_normalize(x, 0), backend.transpose(backend.l2_normalize(y, 0)))

LETTER_GRAM_SIZE = 3
WINDOW_SIZE = 3 
TOTAL_LETTER_GRAMS = int(3 * 1e4)
WORD_DEPTH = WINDOW_SIZE * TOTAL_LETTER_GRAMS
K = 300
L = 128
J = 4
FILTER_LENGTH = 1

query = Input(shape = (None, WORD_DEPTH))
pos_doc = Input(shape = (None, WORD_DEPTH))
neg_docs = [Input(shape = (None, WORD_DEPTH)) for j in range(J)]

query_conv = Convolution1D(K, FILTER_LENGTH, border_mode = "same", input_shape = (None, WORD_DEPTH), activation = "tanh")(query) # See equation (2).

query_max = Lambda(lambda x: backend.max(x, 1), output_shape = (K, ))(query_conv) 

query_sem = Dense(L, activation = "tanh", input_dim = K)(query_max) 

doc_conv = Convolution1D(K, FILTER_LENGTH, border_mode = "same", input_shape = (None, WORD_DEPTH), activation = "tanh")
doc_max = Lambda(lambda x: backend.max(x, 1), output_shape = (K, ))
doc_sem = Dense(L, activation = "tanh", input_dim = K)

pos_doc_conv = doc_conv(pos_doc)
neg_doc_convs = [doc_conv(neg_doc) for neg_doc in neg_docs]

pos_doc_max = doc_max(pos_doc_conv)
neg_doc_maxes = [doc_max(neg_doc_conv) for neg_doc_conv in neg_doc_convs]

pos_doc_sem = doc_sem(pos_doc_max)
neg_doc_sems = [doc_sem(neg_doc_max) for neg_doc_max in neg_doc_maxes]

R_layer = Lambda(R, output_shape = (1, )) 

R_Q_D_p = R_layer([query_sem, pos_doc_sem])
R_Q_D_ns = [R_layer([query_sem, neg_doc_sem]) for neg_doc_sem in neg_doc_sems]

concat_Rs = merge([R_Q_D_p] + R_Q_D_ns, mode = "concat")
concat_Rs = Reshape((J + 1, 1))(concat_Rs)

weight = np.array([1]).reshape(1, 1, 1, 1)
with_gamma = Convolution1D(1, 1, border_mode = "same", input_shape = (J + 1, 1), activation = "linear", bias = False, weights = [weight])(concat_Rs) 

exponentiated = Lambda(lambda x: backend.exp(x), output_shape = (J + 1, ))(with_gamma) # See equation (5).
exponentiated = Reshape((J + 1, ))(exponentiated)

prob = Lambda(lambda x: x[0][0] / backend.sum(x[0]), output_shape = (1, ))(exponentiated) 

model = Model(input = [query, pos_doc] + neg_docs, output = prob)
model.compile(optimizer = "adadelta", loss = "binary_crossentropy")

和错误信息。

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
/Users/jun/anaconda/lib/python3.5/site-packages/tensorflow/python/framework/tensor_shape.py in merge_with(self, other)
    571       try:
--> 572         self.assert_same_rank(other)
    573         new_dims = []

/Users/jun/anaconda/lib/python3.5/site-packages/tensorflow/python/framework/tensor_shape.py in assert_same_rank(self, other)
    617         raise ValueError(
--> 618             "Shapes %s and %s must have the same rank" % (self, other))
    619 

ValueError: Shapes (?, ?) and () must have the same rank

During handling of the above exception, another exception occurred:

ValueError                                Traceback (most recent call last)
/Users/jun/anaconda/lib/python3.5/site-packages/tensorflow/python/ops/nn.py in sigmoid_cross_entropy_with_logits(logits, targets, name)
    430     try:
--> 431       targets.get_shape().merge_with(logits.get_shape())
    432     except ValueError:

/Users/jun/anaconda/lib/python3.5/site-packages/tensorflow/python/framework/tensor_shape.py in merge_with(self, other)
    578         raise ValueError("Shapes %s and %s are not compatible" %
--> 579                          (self, other))
    580 

ValueError: Shapes (?, ?) and () are not compatible

During handling of the above exception, another exception occurred:

ValueError                                Traceback (most recent call last)
<ipython-input-6-a4001289be07> in <module>()
     61 
     62 model = Model(input = [query, pos_doc] + neg_docs, output = prob)
---> 63 model.compile(optimizer = "adadelta", loss = "binary_crossentropy")

/Users/jun/anaconda/lib/python3.5/site-packages/keras/engine/training.py in compile(self, optimizer, loss, metrics, loss_weights, sample_weight_mode, **kwargs)
    628             loss_weight = loss_weights_list[i]
    629             output_loss = weighted_loss(y_true, y_pred,
--> 630                                         sample_weight, mask)
    631             if len(self.outputs) > 1:
    632                 self.metrics_tensors.append(output_loss)

/Users/jun/anaconda/lib/python3.5/site-packages/keras/engine/training.py in weighted(y_true, y_pred, weights, mask)
    330     def weighted(y_true, y_pred, weights, mask=None):
    331         # score_array has ndim >= 2
--> 332         score_array = fn(y_true, y_pred)
    333         if mask is not None:
    334             # Cast the mask to floatX to avoid float64 upcasting in theano

/Users/jun/anaconda/lib/python3.5/site-packages/keras/objectives.py in binary_crossentropy(y_true, y_pred)
     46 
     47 def binary_crossentropy(y_true, y_pred):
---> 48     return K.mean(K.binary_crossentropy(y_pred, y_true), axis=-1)
     49 
     50 

/Users/jun/anaconda/lib/python3.5/site-packages/keras/backend/tensorflow_backend.py in binary_crossentropy(output, target, from_logits)
   1464         output = tf.clip_by_value(output, epsilon, 1 - epsilon)
   1465         output = tf.log(output / (1 - output))
-> 1466     return tf.nn.sigmoid_cross_entropy_with_logits(output, target)
   1467 
   1468 

/Users/jun/anaconda/lib/python3.5/site-packages/tensorflow/python/ops/nn.py in sigmoid_cross_entropy_with_logits(logits, targets, name)
    432     except ValueError:
    433       raise ValueError("logits and targets must have the same shape (%s vs %s)"
--> 434                        % (logits.get_shape(), targets.get_shape()))
    435 
    436     # The logistic loss formula from above is

ValueError: logits and targets must have the same shape (() vs (?, ?))

Answer 1

（应@dga 的要求从评论转换为答案）

我询问了我们的 Keras 常驻专家，他的回答是：查看错误消息和代码，可以推断出最终输出概率是一个标量，而它应该是一个二维数组（每个批次条目一个标量概率） .问题很可能出在一行

prob = Lambda(lambda x: x[0][0] / backend.sum(x[0]), output_shape = (1, ))(exponentiated)