# -*- coding: utf-8 -*- # Copyright (c) Alibaba, Inc. and its affiliates. # _no_grad_trunc_normal_ & trunc_normal_ & variance_scaling_ & lecun_normal_ # Modified from https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/layers/weight_init.py. # Copyright 2019 Ross Wightman # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # _init_transformer_weights & c2_msra_fill & _init_convnet_weights # Modified from https://github.com/facebookresearch/SlowFast/blob/main/slowfast/utils/weight_init_helper.py. # Copyright 2019, Facebook, Inc # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import math import warnings import torch import torch.nn as nn from torch.nn.init import _calculate_fan_in_and_fan_out def _no_grad_trunc_normal_(tensor, mean, std, a, b): # Cut & paste from PyTorch official master until it's in a few official releases - RW # Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf def norm_cdf(x): # Computes standard normal cumulative distribution function return (1. + math.erf(x / math.sqrt(2.))) / 2. if (mean < a - 2 * std) or (mean > b + 2 * std): warnings.warn( 'mean is more than 2 std from [a, b] in nn.init.trunc_normal_. ' 'The distribution of values may be incorrect.', stacklevel=2) with torch.no_grad(): # Values are generated by using a truncated uniform distribution and # then using the inverse CDF for the normal distribution. # Get upper and lower cdf values le = norm_cdf((a - mean) / std) u = norm_cdf((b - mean) / std) # Uniformly fill tensor with values from [l, u], then translate to # [2l-1, 2u-1]. tensor.uniform_(2 * le - 1, 2 * u - 1) # Use inverse cdf transform for normal distribution to get truncated # standard normal tensor.erfinv_() # Transform to proper mean, std tensor.mul_(std * math.sqrt(2.)) tensor.add_(mean) # Clamp to ensure it's in the proper range tensor.clamp_(min=a, max=b) return tensor def trunc_normal_(tensor, mean=0., std=1., a=-2., b=2.): # type: (torch.Tensor, float, float, float, float) -> torch.Tensor r"""Fills the input Tensor with values drawn from a truncated normal distribution. The values are effectively drawn from the normal distribution :math:`\mathcal{N}(\text{mean}, \text{std}^2)` with values outside :math:`[a, b]` redrawn until they are within the bounds. The method used for generating the random values works best when :math:`a \leq \text{mean} \leq b`. Args: tensor: an n-dimensional `torch.Tensor` mean: the mean of the normal distribution std: the standard deviation of the normal distribution a: the minimum cutoff value b: the maximum cutoff value Examples: >>> w = torch.empty(3, 5) >>> nn.init.trunc_normal_(w) """ return _no_grad_trunc_normal_(tensor, mean, std, a, b) def variance_scaling_(tensor, scale=1.0, mode='fan_in', distribution='normal'): fan_in, fan_out = _calculate_fan_in_and_fan_out(tensor) if mode == 'fan_in': denom = fan_in elif mode == 'fan_out': denom = fan_out elif mode == 'fan_avg': denom = (fan_in + fan_out) / 2 variance = scale / denom if distribution == 'truncated_normal': # constant is stddev of standard normal truncated to (-2, 2) trunc_normal_(tensor, std=math.sqrt(variance) / .87962566103423978) elif distribution == 'normal': tensor.normal_(std=math.sqrt(variance)) elif distribution == 'uniform': bound = math.sqrt(3 * variance) tensor.uniform_(-bound, bound) else: raise ValueError(f'invalid distribution {distribution}') def lecun_normal_(tensor): variance_scaling_(tensor, mode='fan_in', distribution='truncated_normal') def _init_transformer_weights(m): if isinstance(m, nn.Linear): trunc_normal_(m.weight, std=.02) if m.bias is not None: nn.init.zeros_(m.bias) elif isinstance(m, nn.LayerNorm): nn.init.zeros_(m.bias) nn.init.ones_(m.weight) def c2_msra_fill(module: nn.Module) -> None: """ Initialize `module.weight` using the "MSRAFill" implemented in Caffe2. Also initializes `module.bias` to 0. Args: module (torch.nn.Module): module to initialize. """ # pyre-ignore nn.init.kaiming_normal_(module.weight, mode='fan_out', nonlinearity='relu') if module.bias is not None: # pyre-ignore nn.init.constant_(module.bias, 0) def _init_convnet_weights(model, fc_init_std=0.01, zero_init_final_bn=True): """ Performs ResNet style weight initialization. Args: fc_init_std (float): the expected standard deviation for fc layer. zero_init_final_bn (bool): if True, zero initialize the final bn for every bottleneck. """ for m in model.modules(): if hasattr(m, 'skip_init'): continue if isinstance(m, nn.Conv3d) and not hasattr(m, 'linear'): """ Follow the initialization method proposed in: {He, Kaiming, et al. "Delving deep into rectifiers: Surpassing human-level performance on imagenet classification." arXiv preprint arXiv:1502.01852 (2015)} """ c2_msra_fill(m) elif isinstance(m, nn.BatchNorm3d): if (hasattr(m, 'transform_final_bn') and m.transform_final_bn and zero_init_final_bn): batchnorm_weight = 0.0 else: batchnorm_weight = 1.0 if m.weight is not None: m.weight.data.fill_(batchnorm_weight) if m.bias is not None: m.bias.data.zero_() if isinstance(m, nn.Linear) or hasattr(m, 'linear'): m.weight.data.normal_(mean=0.0, std=fc_init_std) if m.bias is not None: m.bias.data.zero_()