前言

Tensor 在PyTorch和Tensorflow中都用于存储多维数据,两个框架的Tensor有许多相同之处.

Requirement

OS: Ubuntu 18.04
Tensorflow >= 2.0.0

Tensorflow

A Tensor is a multi-dimensional array. Similar to NumPy ndarray objects, tf.Tensor objects have a data type and a shape. Additionally, tf.Tensors can reside in accelerator memory (like a GPU). TensorFlow offers a rich library of operations (tf.add, tf.matmul, tf.linalg.inv etc.) that consume and produce tf.Tensors. These operations automatically convert native Python types.

tf.Tensor 与numpy中的ndarray和相似,但是也存在不同:

numpy只能在CPU上计算,numpy没有实现GPU加速计算
numpy.ndarray可以修改(元素), 但是tf.Tensor的数据不可改

Tensors can be backed by accelerator memory (like GPU, TPU).

Tensors are immutable.

# Tensor 创建
    var1 = tf.ones(shape=[3,4], dtype=tf.float32)
    print(var1)
    print(var1.shape)
    print(var1.dtype)

    print(var1[0,:])    # return a tensor
    # var1[0, :] = 10     # 错误, tf.Tensor不可改(Tensors are immutable.), TypeError: 'tensorflow.python.framework.ops.EagerTensor' object does not support item assignment

# tensorflow
    print('Tensorflow')
    x = tf.ones([3,3])
    print(id(x))
    x += 10                        # 这里新建了一个Tensor, 并不是in-place add
    print(id(x))

    # PyTorch
    print('torch')
    x = torch.ones([3,3])
    print(id(x))
    x += 10
    print(id(x))

    # numpy
    print('numpy')
    x = np.ones([3,3])
    print(id(x))
    x += 10
    print(id(x))

图片.png

tf.Tensor与np.ndarray相互转换


    # tf.Tensor 与numpy转换
    # tf.Tensor 与numpy 可以自动转换

    var2 = tf.zeros([3,3])
    var3 = var2.numpy()     # tensor--->numpy
    print(type(var3))

    # numpy ---> tensor
    var4 = tf.convert_to_tensor(np.arange(12).reshape([3,4]))
    print(var4)

tf.Tensor指定计算设备(CPU/GPU)

# CPU/GPU
    # 查询CPU/GPU
    print(tf.config.experimental.list_physical_devices('CPU'))
    print(tf.config.experimental.list_physical_devices('GPU'))

    # 判读Tensor是否在GPU/CPU
    var5 = tf.random.uniform(shape=[3,3])
    print(var5.device)      # 判断Tensor在CPU/GPU
    print(var5.device.endswith('GPU:0'))    # 判断是否GPU Tensor

    # 在CPU上计算
    with tf.device('CPU:0'):
        x1 = tf.random.normal(shape=[128, 64])
        x2 = tf.random.normal(shape=[64, 128])
        y = tf.matmul(a=x1,b=x2)
        assert y.device.endswith('CPU:0')
        print(y.shape)