basecls.layers

class basecls.layers.ELU(alpha=1.0, name=None)

基类：Module

ELU activation function.

\[\begin{split}\text{ELU}(x) = \begin{cases} x, & \text{if } x > 0, \\ \alpha \left( \exp(x) - 1 \right), & \text{if } x \le 0 \end{cases}\end{split}\]

参数

alpha (float) – the \(\alpha\) value for the ELU formulation. Default: 1.0

forward(x)

返回类型

Tensor

class basecls.layers.HSigmoid(name=None)

基类：Module

Hard sigmoid activation function.

\[\begin{split}\text{HSigmoid}(x) = \begin{cases} 0 & \text{if } x \le -3, \\ 1 & \text{if } x \ge 3, \\ x / 6 + 1 / 2 & \text{otherwise} \end{cases}\end{split}\]

forward(x)

返回类型

Tensor

class basecls.layers.HSwish(name=None)

基类：Module

Hard swish activation function.

\[\begin{split}\text{HSwish}(x) = \begin{cases} 0 & \text{if } x \le -3, \\ x & \text{if } x \ge 3, \\ x (x + 3) / 6 & \text{otherwise} \end{cases}\end{split}\]

forward(x)

返回类型

Tensor

class basecls.layers.ReLU6(name=None)

基类：Module

ReLU6 activation function.

\[\text{ReLU6}(x) = \min \left( \max(0, x), 6 \right)\]

forward(x)

返回类型

Tensor

class basecls.layers.Tanh(name=None)

基类：Module

Tanh activation function.

\[\text{Tanh}(x) = \text{tanh}(x) = \frac{\exp(x) - \exp(-x)}{\exp(x) + \exp(-x)}\]

forward(x)

返回类型

Tensor

basecls.layers.activation(name, **kwargs)

Helper for building an activation layer.

参数

name (Union[str, Callable]) – activation name, supports "elu", "gelu", "hsigmoid", "hswish", "leaky_relu", "relu", "relu6", "prelu", "silu" and "tanh".

返回类型

Module

返回

An activation module.

class basecls.layers.ClsHead(w_in, w_out=1000, width=0, dropout_prob=0.0, norm_name='BN', act_name='relu', bias=True)

基类：Module

Cls head: Conv, BN, Act, AvgPool, FC.

参数

• w_in (int) – input width.

• w_out (int) – output width, normally the number of classes. Default: 1000

• width (int) – width for first conv in head, conv will be omitted if set to 0. Default: 0

• dropout_prob (float) – dropout probability. Default: 0.0

• norm_name (str) – normalization function. Default: "BN"

• act_name (str) – activation function. Default: "relu"

• bias (bool) – whether fc has bias. Default: True

forward(x)

返回类型

Tensor

class basecls.layers.MBV3Head(w_in, w_out=1000, width=960, w_h=1280, dropout_prob=0.0, se_r=0.0, norm_name='BN', act_name='hswish', bias=True)

基类：Module

MobileNet V3 head: Conv, BN, Act, AvgPool, SE, FC, Act, FC.

参数

• w_in (int) – input width.

• w_out (int) – output width, normally the number of classes.

• width (int) – width for first conv in head.

• w_h (int) – width for first linear in head.

• dropout_prob (float) – dropout probability. Default: 0.0

• se_r (float) – Squeeze-and-Excitation (SE) ratio. Default: 0.0

• norm_name (str) – normalization function. Default: "BN"

• act_name (str) – activation function. Default: "hswish"

• bias (bool) – whether fc has bias. Default: True

forward(x)

返回类型

Tensor

class basecls.layers.VGGHead(w_in, w_out=1000, width=4096, dropout_prob=0.5, act_name='relu', **kwargs)

基类：Module

VGG head: AvgPool, [FC, Act, Dropout] x2, FC.

参数

• w_in (int) – input width.

• w_out (int) – output width, normally the number of classes. Default: 1000

• width (int) – width for linear in head. Default: 4096

• dropout_prob (float) – dropout probability. Default: 0.5

• act_name (str) – activation function. Default: "relu"

forward(x)

返回类型

Tensor

basecls.layers.build_head(w_in, head_args=None, norm_name='BN', act_name='relu')

The factory function to build head.

备注

if head_args is None or head_args["name"] is None, this function will do nothing and return None.

参数

• w_in (int) – input width.

• head_args (Optional[Mapping[str, Any]]) – head args. Default: None

• norm_name (str) – default normalization function, will be overridden by the same key in head_args. Default: "BN"

• act_name (str) – default activation function, will be overridden by the same key in head_args. Default: "relu"

返回类型

Module

返回

A head.

class basecls.layers.BinaryCrossEntropy(**kwargs)

基类：Module

The module for binary cross entropy.

See binary_cross_entropy() for more details.

forward(x, y)

返回类型

Tensor

class basecls.layers.CrossEntropy(axis=1, label_smooth=0.0)

基类：Module

The module for cross entropy.

It supports both categorical labels and one-hot labels. See cross_entropy() for more details.

参数

• axis (int) – reduced axis. Default: 1

• label_smooth (float) – label smooth factor. Default: 0.0

forward(x, y)

返回类型

Tensor

basecls.layers.build_loss(cfg)

The factory function to build loss.

参数

cfg (ConfigDict) – config for building loss function.

返回类型

Module

返回

A loss function.

class basecls.layers.SE(w_in, w_se, act_name, approx_sigmoid=False)

基类：Module

Squeeze-and-Excitation (SE) block: AvgPool, FC, Act, FC, Sigmoid.

参数

• w_in (int) – input width.

• w_se (int) – se width.

• act_name (str) – activation name.

• approx_sigmoid (bool) – approximated sigmoid function.

avg_pool

gad2d layer.

f_ex

sequantial which conbines conv2d -> act -> conv2d -> sigmoid.

forward(x)

返回类型

Tensor

class basecls.layers.DropPath(drop_prob=0.0, **kwargs)

基类：Dropout

DropPath block.

参数

drop_prob – the probability to drop (set to zero) each path.

forward(x)

basecls.layers.conv2d(w_in, w_out, k, *, stride=1, dilation=1, groups=1, bias=False)

Helper for building a conv2d layer.

It will calculate padding automatically.

参数

• w_in (int) – input width.

• w_out (int) – output width.

• k (int) – kernel size.

• stride (int) – stride. Default: 1

• dilation (int) – dilation. Default: 1

• groups (int) – groups. Default: 1

• bias (bool) – enable bias or not. Default: False

返回类型

Conv2d

返回

A conv2d module.

basecls.layers.gap2d(shape=1)

Helper for building a gap2d layer.

参数

shape – output shape. Default: 1

返回类型

AdaptiveAvgPool2d

返回

A gap2d module.

basecls.layers.linear(w_in, w_out, *, bias=False)

Helper for building a linear layer.

参数

• w_in (int) – input width.

• w_out (int) – output width.

• bias (bool) – enable bias or not. Default: False

返回类型

Linear

返回

A linear module.

basecls.layers.norm2d(name, w_in, **kwargs)

Helper for building a norm2d layer.

参数

• norm_name – normalization name, supports None, "BN", "GN", "IN", "LN" and "SyncBN".

• w_in (int) – input width.

返回类型

Module

返回

A norm2d module.

basecls.layers.pool2d(k, *, stride=1, name='max')

Helper for building a pool2d layer.

参数

• k (int) – kernel size.

• stride (int) – stride. Default: 1

• name (str) – pooling name, supports "avg" and "max".

返回类型

Module

返回

A pool2d module.

class basecls.layers.Preprocess(mean, std)

基类：Module

forward(inputs)

返回类型

Tuple[Tensor, Tensor]

basecls.layers.adjust_block_compatibility(ws, bs, gs)

Adjusts the compatibility of widths, bottlenecks and groups.

参数

• ws (Sequence[int]) – widths.

• bs (Sequence[float]) – bottleneck multipliers.

• gs (Sequence[int]) – group widths.

返回类型

Tuple[List[int], ...]

返回

The adjusted widths, bottlenecks and groups.

basecls.layers.compute_precise_bn_stats(cfg, model, dataloader)

Computes precise BN stats on training data.

References: facebookresearch/pycls

参数

• cfg (ConfigDict) – config for precising BN.

• model (Module) – model for precising BN.

• dataloader (Union[DataLoader, FakeDataLoader]) – dataloader for precising BN.

basecls.layers.init_vit_weights(module)

Initialization for Vision Transformer (ViT).

References: rwightman/pytorch-image-models

参数

m – module to be initialized.

basecls.layers.init_weights(m, pytorch_style=False, zero_init_final_gamma=False)

Performs ResNet-style weight initialization.

About zero-initialize: Zero-initialize the last BN in each residual branch, so that the residual branch starts with zeros, and each residual block behaves like an identity. This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677.

References: facebookresearch/pycls

参数

• m (Module) – module to be initialized.

• pytorch_style (bool) – utilize pytorch style init for group conv. Default: False

• zero_init_final_gamma (bool) – enable zero-initialize or not. Default: False

basecls.layers.lecun_normal_(tensor)

basecls.layers.make_divisible(value, divisor=8, min_value=None, round_limit=0.0)

basecls.layers.trunc_normal_(tensor, mean=0.0, std=1.0, a=-2.0, b=2.0)

• basecls.layers.activations
  • activation()
  • ELU
    • ELU.forward()
  • HSigmoid
    • HSigmoid.forward()
  • HSwish
    • HSwish.forward()
  • ReLU6
    • ReLU6.forward()
  • Tanh
    • Tanh.forward()
• basecls.layers.heads
  • build_head()
  • ClsHead
    • ClsHead.forward()
  • MBV3Head
    • MBV3Head.forward()
  • VGGHead
    • VGGHead.forward()
• basecls.layers.losses
  • build_loss()
  • BinaryCrossEntropy
    • BinaryCrossEntropy.forward()
  • CrossEntropy
    • CrossEntropy.forward()
• basecls.layers.modules
  • conv2d()
  • norm2d()
  • pool2d()
  • gap2d()
  • linear()
  • SE
    • SE.avg_pool
    • SE.f_ex
    • SE.forward()
  • DropPath
    • DropPath.forward()
• basecls.layers.wrapper
  • Preprocess
    • Preprocess.forward()
  • adjust_block_compatibility()
  • calculate_fan_in_and_fan_out()
  • compute_precise_bn_stats()
  • init_weights()
  • init_vit_weights()
  • trunc_normal_()
  • lecun_normal_()
  • make_divisible()