Question

我已经在keras功能API上开发了以下网络，该网络接收形状为[35，15，1]的多个输入序列，并预测形状为[35，1]的单个输出：

# Model

def create_keras_model(window_size, n_inputs):

    region_input = tf.keras.Input(shape = (window_size), name = 'region')
    freight_rev_input = tf.keras.Input(shape = (window_size), name = 'freight_rev')
    meas_freight_expense_input = tf.keras.Input(shape = (window_size), name = 'meas_freight_expense')
    prod_recipe_input = tf.keras.Input(shape = (window_size), name = 'prod_recipe')
    item_qty_input = tf.keras.Input(shape = (window_size), name = 'item_qty')
    calc_qty_input = tf.keras.Input(shape = (window_size), name = 'calc_qty')
    dlvy_qty_input = tf.keras.Input(shape = (window_size), name = 'dlvy_qty')
    #conv_input = tf.keras.Input(shape = (), name = 'conv')

    # Reshape features for LSTM
    region_feature = tf.keras.layers.Reshape([window_size, 1])(region_input)
    freight_rev_feature = tf.keras.layers.Reshape([window_size, 1])(freight_rev_input)
    meas_freight_expense_feature = tf.keras.layers.Reshape([window_size, 1])(meas_freight_expense_input)
    prod_recipe_feature = tf.keras.layers.Reshape([window_size, 1])(prod_recipe_input)
    item_qty_feature = tf.keras.layers.Reshape([window_size, 1])(item_qty_input)
    calc_qty_feature = tf.keras.layers.Reshape([window_size, 1])(calc_qty_input)
    dlvy_qty_feature = tf.keras.layers.Reshape([window_size, 1])(dlvy_qty_input)
    #conv_feature = tf.keras.layers.Reshape([window_size, 1])(conv_input)

    # Concatenate features
    x = tf.keras.layers.concatenate([region_feature,freight_rev_feature,meas_freight_expense_feature,
                                     prod_recipe_feature,item_qty_feature,calc_qty_feature,dlvy_qty_feature,
                                    #conv_feature
                                    ])

    # Apply 1st LSTM
    x = tf.keras.layers.LSTM(n_inputs, return_sequences = True)(x)

    # Apply 2nd LSTM
    x = tf.keras.layers.LSTM(n_inputs, return_sequences = True)(x)

    # Apply 3rd LSTM
    x = tf.keras.layers.LSTM(n_inputs)(x)

    # Calculate Dense Output
    y = tf.keras.layers.Dense(1, activation = tf.nn.relu, name='prediction')(x)

    # Create Model
    model = tf.keras.Model(inputs = [region_input,freight_rev_input,meas_freight_expense_input,
                                     prod_recipe_input,item_qty_input,calc_qty_input,dlvy_qty_input,
                                    # conv_input
                                    ],
    outputs = [y])

    model.compile(optimizer = tf.keras.optimizers.Adam(learning_rate=0.001)
                  ,loss = tf.keras.losses.MSE
                  ,metrics=[tf.keras.losses.MSE])

    return model

在model.fit（features，target）上运行它，没关系；当我将其转换为tf.estimator时，在节点'loss / prediction_loss / SquaredDifference'上出现不兼容的形状问题：

#Build Keras model
    model = create_keras_model(lookback, num_features)

#Convert Keras model to an Estimator
estimator = tf.keras.estimator.model_to_estimator(
            keras_model = model, 
            model_dir = output_dir, 
            config = tf.estimator.RunConfig(save_checkpoints_secs = min_eval_frequency))

当调用train_and_evaluate时，出现此错误：

InvalidArgumentError: Incompatible shapes: [35,1] vs. [0,1]
     [[node loss/prediction_loss/SquaredDifference (defined at <ipython-input-67-b45d1de25e2a>:47) ]]

Errors may have originated from an input operation.
Input Source operations connected to node loss/prediction_loss/SquaredDifference:
 ExpandDims (defined at <ipython-input-8-ea21aacd12c5>:18)

我相信它与在净输出上定义的固定张量形状有关，但是我找不到解决它的方法。有人可以帮我吗？

InvalidArgumentError：将tf.keras.Model模型转换为tf.estimator.Estimator时，损失函数上的形状不兼容

0 个答案: