Question

我在TensorFlow 2.0和Python 3.X中有一个用于二进制分类任务的神经网络，使用的是“心脏”数据集，该数据集可以从以下网址下载：

神经网络使用“ ReLU”激活功能和Adam梯度下降优化器具有两个隐藏层。我使用的自定义循环使用的是“ apply_gradients（）”和“ GradientTape（）”方法，而不是“ fit（）”方法。

此外，原始的CSV文件使用“ pandas”（pd）读取，然后再使用API加载到TensorFlow中。

我的代码如下：

# Load CSV file-
data = pd.read_csv("heart.csv")

# Get shape/dimension of data-
data.shape
# (303, 14)

# Check for missing values-
data.isnull().values.any()
# False

# Convert 'thal' column which is an object in the dataframe to a discrete
# numerical value-
data['thal'] = pd.Categorical(data['thal'])
data['thal'] = data.thal.cat.codes

# Remove 'target' attribute-
target = data.pop('target')

# Load pandas DataFrame to TensorFlow-
dataset = tf.data.Dataset.from_tensor_slices((data.values, target.values))

# Shuffle and batch the dataset-
train_dataset = dataset.shuffle(buffer_size=len(data)).batch(batch_size=1, drop_remainder=False)




# Create a neural network:
model = tf.keras.Sequential([
    tf.keras.layers.InputLayer(input_shape = (13,), dtype = tf.float64),
    tf.keras.layers.Dense(10, activation = 'relu'),
    tf.keras.layers.Dense(10, activation = 'relu'),
    tf.keras.layers.Dense(1, activation = 'sigmoid')
    ])


def cost_fn(model, x, y):
    '''
    Cost function to compute binary cross-entropy
    '''
    y_ = model(x)
    # return tf.keras.losses.binary_crossentropy(y_true = y, y_pred = y_)

    bce = tf.keras.losses.BinaryCrossentropy(from_logits = True)
    loss = bce(y_true=y, y_pred=y_)

    return loss


def compute_gradients(model, inputs, targets):
    '''
    Function to compute the gradients for optimizing
    model
    '''
    with tf.GradientTape() as t:
        # t.watch(model.trainable_variables)
        cost_val = cost_fn(model, inputs, targets)

    return cost_val, t.gradient(cost_val, model.trainable_variables)


# Create Adam Gradient Descent optimizer-
optimizer = tf.keras.optimizers.Adam(learning_rate=0.01)


# Training loop:

# Used for plotting-
train_loss_results = []
train_accuracy_results = []

# An epoch is one pass through the dataset
num_epochs = 1000

for epoch in range(num_epochs):
    epoch_loss_avg = tf.keras.metrics.Mean()
    # epoch_accuracy = tf.keras.metrics.SparseCategoricalAccuracy()
    epoch_accuracy = tf.keras.metrics.BinaryAccuracy()


    # Training loop - using batches of one example-
    for x, y in train_dataset:
        # Optimize model-
        cost_val, grads = compute_gradients(model, x, y)
        optimizer.apply_gradients(zip(grads, model.trainable_variables))

        # Keep track of progress-
        epoch_loss_avg(cost_val)    # add current batch cost/loss

        # Compare accuracy of predicted label to actual label-
        epoch_accuracy(y, model(x))

    # End of each epoch-
    train_loss_results.append(epoch_loss_avg.result())
    train_accuracy_results.append(epoch_accuracy.result())

    if epoch % 50 == 0:
        print("\nepoch = {0}, cost = {1:.4f} & accuracy = {2:.4f}\n".format(epoch, epoch_loss_avg.result(), epoch_accuracy.result()))

上述代码存在的问题是：

损失/成本值不会随着时间的推移而改善
在每个时期使用“ for”循环进行训练需要更多的时间，而不是如果我不使用下面的代码

model.compile（optimizer ='adam'，loss ='binary_crossentropy'，metrics = ['accuracy']）

model.fit（train_dataset，epochs = 50）

成本价值没有提高-神经网络

0 个答案: