File size: 1,652 Bytes
f15d29e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
"""
Copyright (c) Facebook, Inc. and its affiliates.
Copyright (c) Microsoft Corporation.
Licensed under the MIT License.
Adapted from https://github.com/FAIR-Chem/fairchem/blob/main/src/fairchem/core/models/gemnet/initializers.py.
"""

import torch


# This function is not type annotated because mypy complains that axis could be either an integer or a tuple of integers,
# even though this is precicely how torch.var_mean works
def _standardize(kernel):
    """
    Makes sure that N*Var(W) = 1 and E[W] = 0
    """
    eps = 1e-6

    if len(kernel.shape) == 3:
        axis = (0, 1)  # last dimension is output dimension
    else:
        axis = 1

    var, mean = torch.var_mean(kernel, dim=axis, unbiased=True, keepdim=True)
    kernel = (kernel - mean) / (var + eps) ** 0.5
    return kernel


def he_orthogonal_init(tensor: torch.Tensor) -> torch.Tensor:
    """
    Generate a weight matrix with variance according to He (Kaiming) initialization.
    Based on a random (semi-)orthogonal matrix neural networks
    are expected to learn better when features are decorrelated
    (stated by eg. "Reducing overfitting in deep networks by decorrelating representations",
    "Dropout: a simple way to prevent neural networks from overfitting",
    "Exact solutions to the nonlinear dynamics of learning in deep linear neural networks")
    """
    tensor = torch.nn.init.orthogonal_(tensor)

    if len(tensor.shape) == 3:
        fan_in = tensor.shape[:-1].numel()
    else:
        fan_in = tensor.shape[1]

    with torch.no_grad():
        tensor.data = _standardize(tensor.data)
        tensor.data *= (1 / fan_in) ** 0.5

    return tensor