122 lines
4.3 KiB
Python
122 lines
4.3 KiB
Python
# Copyright (c) OpenMMLab. All rights reserved.
|
|
import math
|
|
from collections import OrderedDict
|
|
|
|
import torch.nn as nn
|
|
import torch.nn.functional as F
|
|
from mmcv.cnn import build_activation_layer
|
|
from mmcv.cnn.utils.weight_init import trunc_normal_
|
|
from mmcv.runner import Sequential
|
|
|
|
from ..builder import HEADS
|
|
from .cls_head import ClsHead
|
|
|
|
|
|
@HEADS.register_module()
|
|
class VisionTransformerClsHead(ClsHead):
|
|
"""Vision Transformer classifier head.
|
|
|
|
Args:
|
|
num_classes (int): Number of categories excluding the background
|
|
category.
|
|
in_channels (int): Number of channels in the input feature map.
|
|
hidden_dim (int): Number of the dimensions for hidden layer. Only
|
|
available during pre-training. Default None.
|
|
act_cfg (dict): The activation config. Only available during
|
|
pre-training. Defaults to Tanh.
|
|
"""
|
|
|
|
def __init__(self,
|
|
num_classes,
|
|
in_channels,
|
|
hidden_dim=None,
|
|
act_cfg=dict(type='Tanh'),
|
|
init_cfg=dict(type='Constant', layer='Linear', val=0),
|
|
*args,
|
|
**kwargs):
|
|
super(VisionTransformerClsHead, self).__init__(
|
|
init_cfg=init_cfg, *args, **kwargs)
|
|
self.in_channels = in_channels
|
|
self.num_classes = num_classes
|
|
self.hidden_dim = hidden_dim
|
|
self.act_cfg = act_cfg
|
|
|
|
if self.num_classes <= 0:
|
|
raise ValueError(
|
|
f'num_classes={num_classes} must be a positive integer')
|
|
|
|
self._init_layers()
|
|
|
|
def _init_layers(self):
|
|
if self.hidden_dim is None:
|
|
layers = [('head', nn.Linear(self.in_channels, self.num_classes))]
|
|
else:
|
|
layers = [
|
|
('pre_logits', nn.Linear(self.in_channels, self.hidden_dim)),
|
|
('act', build_activation_layer(self.act_cfg)),
|
|
('head', nn.Linear(self.hidden_dim, self.num_classes)),
|
|
]
|
|
self.layers = Sequential(OrderedDict(layers))
|
|
|
|
def init_weights(self):
|
|
super(VisionTransformerClsHead, self).init_weights()
|
|
# Modified from ClassyVision
|
|
if hasattr(self.layers, 'pre_logits'):
|
|
# Lecun norm
|
|
trunc_normal_(
|
|
self.layers.pre_logits.weight,
|
|
std=math.sqrt(1 / self.layers.pre_logits.in_features))
|
|
nn.init.zeros_(self.layers.pre_logits.bias)
|
|
|
|
def pre_logits(self, x):
|
|
if isinstance(x, tuple):
|
|
x = x[-1]
|
|
_, cls_token = x
|
|
if self.hidden_dim is None:
|
|
return cls_token
|
|
else:
|
|
x = self.layers.pre_logits(cls_token)
|
|
return self.layers.act(x)
|
|
|
|
def simple_test(self, x, softmax=True, post_process=True):
|
|
"""Inference without augmentation.
|
|
|
|
Args:
|
|
x (tuple[tuple[tensor, tensor]]): The input features.
|
|
Multi-stage inputs are acceptable but only the last stage will
|
|
be used to classify. Every item should be a tuple which
|
|
includes patch token and cls token. The cls token will be used
|
|
to classify and the shape of it should be
|
|
``(num_samples, in_channels)``.
|
|
softmax (bool): Whether to softmax the classification score.
|
|
post_process (bool): Whether to do post processing the
|
|
inference results. It will convert the output to a list.
|
|
|
|
Returns:
|
|
Tensor | list: The inference results.
|
|
|
|
- If no post processing, the output is a tensor with shape
|
|
``(num_samples, num_classes)``.
|
|
- If post processing, the output is a multi-dimentional list of
|
|
float and the dimensions are ``(num_samples, num_classes)``.
|
|
"""
|
|
x = self.pre_logits(x)
|
|
cls_score = self.layers.head(x)
|
|
|
|
if softmax:
|
|
pred = (
|
|
F.softmax(cls_score, dim=1) if cls_score is not None else None)
|
|
else:
|
|
pred = cls_score
|
|
|
|
if post_process:
|
|
return self.post_process(pred)
|
|
else:
|
|
return pred
|
|
|
|
def forward_train(self, x, gt_label, **kwargs):
|
|
x = self.pre_logits(x)
|
|
cls_score = self.layers.head(x)
|
|
losses = self.loss(cls_score, gt_label, **kwargs)
|
|
return losses
|