Tensorflow 2.0-GPU Windows在CPU上运行训练代码

时间:2020-01-16 11:56:04

标签: python tensorflow deep-learning

当前,我在 Tensorflow 2.0 上使用。这是我的电脑:

  • CPU:i5-4690 3.5Ghz。
  • RAM:16GB。
  • GPU:NVIDIA GeForce 1050Ti 4GB。
  • 操作系统:Windows 10 Pro 64位。
  • CUDA 10.0和cuDNN 7.4。

我的项目是使用 ResNet50 CIFAR100数据集的图像分类项目。

我使用子类构建网络(代码段太长,因此我没有在这个问题上附加它),并使用tf.data.Dataset.from_tensor_slices加载了数据:

def load_cifar100(batch_size, num_classes=100):
    (x_train, y_train), (x_test, y_test) = cifar100.load_data()

    x_train, x_test = x_train.astype('float32') / 255, x_test.astype('float32') / 255

    x_val, y_val = x_train[-10000:], y_train[-10000:]
    x_train, y_train = x_train[:-10000], y_train[:-10000]

    y_train = to_categorical(y_train, num_classes)
    y_test = to_categorical(y_test, num_classes)
    y_val = to_categorical(y_val, num_classes)

    train_dataset = tf.data.Dataset.from_tensor_slices((x_train, y_train))
    train_dataset = train_dataset.shuffle(buffer_size=1024).batch(batch_size)

    val_dataset = tf.data.Dataset.from_tensor_slices((x_val, y_val))
    val_dataset = val_dataset.batch(batch_size)

    test_dataset = tf.data.Dataset.from_tensor_slices((x_test, y_test))
    test_dataset = test_dataset.batch(batch_size)

    return train_dataset, val_dataset, test_dataset

我使用GradientTape来设置训练过程:

def training(x_batch, y_batch):
    with tf.GradientTape() as tape:
        logits = model(x_batch, training=True)
        loss_val = loss(y_batch, logits)

    grads = tape.gradient(loss_val, model.trainable_weights)
    optimizer.apply_gradients(zip(grads, model.trainable_weights))
    train_acc_metric(y_batch, logits)


for epoch in range(epochs):
    train_acc_metric.reset_states()
    for step, (x_batch_train, y_batch_train) in enumerate(train_dataset):
        training(x_batch_train, y_batch_train)

    train_acc = train_acc_metric.result()
    template = 'Epoch {}, Train_Acc: {}'
    print(template.format(epoch + 1,
                          train_acc))

在培训期间,我发现GPU根本无法工作[Pic1],即使我打开了调试配置{{1},所有培训过程也都放入了CPU中},似乎所有图层都已加载到tf.debugging.set_log_device_placement(True) [Pic2]。

GPU not working

Tensorflow logs

更新:

这是当我更改为GPU函数时GPU的外观。而且每个时期的训练时间比model.fit快得多: RUn on model.fit

1 个答案:

答案 0 :(得分:1)

在开始训练过程之前检查Tensorflow(TF2)是否使用GPU很有帮助,

assert tf.test.is_gpu_available()
assert tf.test.is_built_with_cuda()