Question

下面的代码适用于输入和标签的形状，但是当我使用它们进行估算时，它不起作用（请参见下文）。

x1 = pickle.load(open(data_folder+'train_sen1.p', 'rb'))
x2 = pickle.load(open(data_folder+'train_sen2.p', 'rb'))
y = pickle.load(open(data_folder+'y_train.p', 'rb'))
x_train = np.concatenate([np.expand_dims(train_sen1, axis=1),np.expand_dims(train_sen2, axis=1)], axis=1)

print(x_train.shape)
print(y.shape)
#(11118, 2, 50)
#(11118,)

def build_model(vectors, max_length, num_classes, projected_dim,  
                num_hidden=200, dropout_rate1=0.2, dropout_rate2=0.2, 
                dropout_rate3=0.2, learn_rate=0.0001, optimizer='nadam'):
    K.clear_session()
    #input1 = layers.Input(shape=(max_length,), dtype='int32', name='words1')
    #input2 = layers.Input(shape=(max_length,), dtype='int32', name='words2')

    # modified model to take single input instead of two in order to work with hyperas specs.
    model_input = layers.Input(shape=(2, max_length), dtype='int32')
    input1 = layers.Lambda(get_input_slice(1))(model_input)
    input2 = layers.Lambda(get_input_slice(2))(model_input)

    # embeddings (projected)
    embed = create_embedding(vectors, max_length, projected_dim)
    a = embed(input1)
    b = embed(input2)     

    # step 1: attend
    F = create_feedforward(num_hidden, dropout_rate=dropout_rate1)
    att_weights = layers.dot([F(a), F(b)], axes=-1, normalize=True)

    G = create_feedforward(num_hidden, dropout_rate=dropout_rate2)    

    norm_weights_a = layers.Lambda(normalizer(1))(att_weights)
    norm_weights_b = layers.Lambda(normalizer(2))(att_weights)
    alpha = layers.dot([norm_weights_a, a], axes=1)
    beta  = layers.dot([norm_weights_b, b], axes=1)

    # step 2: compare
    comp1 = layers.concatenate([a, beta])
    comp2 = layers.concatenate([b, alpha])
    v1 = layers.TimeDistributed(G)(comp1)
    v2 = layers.TimeDistributed(G)(comp2)

    # step 3: aggregate
    v1_sum = layers.Lambda(sum_word)(v1)
    v2_sum = layers.Lambda(sum_word)(v2)
    concat = layers.concatenate([v1_sum, v2_sum])

    H = create_feedforward(num_hidden, dropout_rate=dropout_rate3)
    out = H(concat)
    out = layers.Dense(num_classes, activation='sigmoid', use_bias=True)(out)

    model = Model(model_input, out)
    if optimizer == 'sgd':
        opt = SGD(lr=learn_rate)
    elif optimizer == 'adam':
        opt = Adam(lr=learn_rate)
    elif optimizer == 'rmsprop':
        opt = RMSprop(lr=learn_rate)
    else:
        opt = Nadam(lr=learn_rate)

    model.compile(optimizer=opt,
    #model.compile(optimizer=optimizer,
                  loss='binary_crossentropy',
                  metrics=["accuracy",threshold_acc,precision_threshold(0.8),recall_threshold(0.8)])
    return model


model = build_model(vectors=embedding_matrix, max_length=50, projected_dim=200, num_classes=1, 
                    num_hidden=200, dropout_rate1=0.1, 
                    dropout_rate2=0.3, dropout_rate3=0.3,
                    optimizer="adam",
                    learn_rate=0.001)

result=model.fit(x_train, y_train, batch_size= 50, epochs=1)

此代码适用于输入和标签的形状。

但是，如果我想做一个估算器：

train_input_fn = tf.estimator.inputs.numpy_input_fn(
x={model.input_names[0]: x_train},  # input_names[0] would be 'main_input'
y=y,
batch_size=100,
num_epochs=None,
shuffle=True)  



estimator = tf.keras.estimator.model_to_estimator(model)
estimator.train(input_fn=train_input_fn, steps=1)

形状错误：

ValueError: logits and labels must have the same shape ((100, 1) vs (100,))

我可以将标签的形状调整为：

np.asarray(y_train).astype('float32').reshape((-1,1))

但是我不知道为什么这对于估计量是必需的？

tf.keras与估算器的形状差异从何而来？

tf.keras与估算器需要不同的形状？

0 个答案: