tensoflow和pytorch之间函数的转换

tf 1.x 对应的 pytorch 3.6+ 函数

1. tf.nn.l2_loss（变量） <–> torch.norm（tensor）** 2 / 2

    a = tf.Variable([[3.0, 4.0], [3.0, 4.0]])
    a_l2_loss = tf.nn.l2_loss(a)  # a中每个元素平方和除以2， 是一个scalar
    a_tor = torch.tensor([[3.0, 4.0], [3.0, 4.0]], dtype=torch.float32)
    a_torch = torch.norm(a_tor) ** 2 / 2
    print(a_l2_loss)
    print(a_torch)

2. tf.nn.embedding_lookup < – > torch.index_select

    a = tf.Variable([[3.0, 4.0], [5.0, 6.0]])
    a_tor = torch.tensor([[3.0, 4.0], [5.0, 6.0]], dtype=torch.float32)
    index = torch.tensor([1])
    print(a)
    print(a_tor)
    a_select = tf.nn.embedding_lookup(a, index)
    b_select = torch.index_select(a_tor, 0, index)
    print(a_select)
    print(b_select)

3. 大坑，由于这个我找了4h的bug，o(╥﹏╥)o
   keep_prob = prob
   tf.nn.dropout(keep_prob) < – > torch.nn.Dropout(prob)(tensor)
   =torch.nn.F.dropout(tensor, prob)