basecls.models

basecls.models.build_model(cfg)

The factory function to build model.

备注

if cfg.model does not have the attr head, this function will build model with the default head. Otherwise if cfg.model.head is None, this function will build model without any head.

备注

if cfg.model.head does not have the attr w_out and cfg.num_classes exists, w_out will be overridden by cfg.num_classes.

参数

cfg (ConfigDict) – config for building model.

返回类型

Module

返回

A model.

basecls.models.load_model(model, weight_path, strict=True)

Load model weights.

参数

• model (Module) – model for loading weights.

• weight_path (str) – weight path, both local path and OSS path are supported.

• strict (bool) – load weights in strict mode or not. Default: True

basecls.models.sync_model(model)

Sync parameters and buffers.

参数

model (Module) – model for syncing.

class basecls.models.EffNet(stem_w, block_name, depths, widths, strides, kernels, exp_rs=1.0, se_rs=0.0, drop_path_prob=0.0, depth_mult=1.0, width_mult=1.0, omit_mult=False, norm_name='BN', act_name='silu', head=None)

基类：Module

EfficientNet model.

参数

• stem_w (int) – stem width.

• block_name (Union[str, Callable, Sequence[Union[str, Callable]]]) – block name.

• depths (Sequence[int]) – depth for each stage (number of blocks in the stage).

• widths (Sequence[int]) – width for each stage (width of each block in the stage).

• strides (Sequence[int]) – strides for each stage (applies to the first block of each stage).

• kernels (Sequence[int]) – kernel sizes for each stage.

• exp_rs (Union[float, Sequence[Union[float, Sequence[float]]]]) – expansion ratios for MBConv blocks in each stage.

• se_r – Squeeze-and-Excitation (SE) ratio. Default: 0.25

• drop_path_prob (float) – drop path probability. Default: 0.0

• depth_mult (float) – depth multiplier. Default: 1.0

• width_mult (float) – width multiplier. Default: 1.0

• omit_mult (bool) – omit multiplier for stem width, head width, the first stage depth and the last stage depth, enabled in EfficientNet-Lite. Default: False

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

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

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

forward(x)

static get_block_func(name)

Retrieves the block function by name.

class basecls.models.HRNet(stage_modules, stage_blocks, stage_block_names, stage_channels, w_stem=64, multi_scale_output=True, merge_block_name='bottleneck', merge_channels=[32, 64, 128, 256], norm_name='BN', act_name='relu', head=None, **kwargs)

基类：Module

HRNet model.

参数

• stage_modules (List[int]) – Number of modules for each stage.

• stage_blocks (List[List[int]]) – Number of blocks for each module in stages.

• stage_block_names (List[str]) – Branch block types for each stage.

• stage_channels (List[List[int]]) – Number of channels for each stage.

• w_stem (int) – Stem width. Default: 64

• multi_scale_output (bool) – Whether output multi-scale features. Default: True

• merge_block_name (str) – Merge block type. Default: "bottleneck"

• merge_channels (List[int]) – Channels of each scale in merge block. Default: [32, 64, 128, 256]

• norm_name (str) – Normalization layer. Default: "BN"

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

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

forward(x)

class basecls.models.MBNet(stem_w, depths, widths, strides, kernels, exp_rs=1.0, se_rs=0.0, stage_act_names=None, has_proj_act=False, has_skip=True, drop_path_prob=0.0, width_mult=1.0, norm_name='BN', act_name='relu6', head=None)

基类：Module

MobileNet model.

参数

• stem_w (int) – stem width.

• depths (Sequence[int]) – depth for each stage (number of blocks in the stage).

• widths (Sequence[int]) – width for each stage (width of each block in the stage).

• strides (Sequence[int]) – strides for each stage (applies to the first block of each stage).

• kernels (Sequence[int]) – kernel sizes for each stage.

• exp_rs (Union[float, Sequence[Union[float, Sequence[float]]]]) – expansion ratios for MobileNet basic blocks in each stage. Default: 1.0

• se_rs (Union[float, Sequence[Union[float, Sequence[float]]]]) – Squeeze-and-Excitation (SE) ratios. Default: 0.0

• stage_act_names (Optional[Sequence[str]]) – activation function for stages. Default: None

• has_proj_act (bool) – whether apply activation to output. Default: False

• has_skip (bool) – whether apply skip connection. Default: True

• drop_path_prob (float) – drop path probability. Default: 0.0

• width_mult (float) – width multiplier. Default: 1.0

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

• act_name (str) – activation function. Default: "relu6"

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

forward(x)

class basecls.models.RegNet(stem_name, stem_w, block_name, depth, w0, wa, wm, group_w, stride=2, bot_mul=1.0, se_r=0.0, drop_path_prob=0.0, zero_init_final_gamma=False, norm_name='BN', act_name='relu', head=None)

基类：ResNet

RegNet model.

参数

• stem_name (Union[str, Callable]) – stem name.

• stem_w (int) – stem width.

• block_name (Union[str, Callable]) – block name.

• depth (int) – depth.

• w0 (int) – initial width.

• wa (float) – slope.

• wm (float) – quantization.

• group_w (int) – group width for each stage (applies to bottleneck block).

• stride (int) – stride for each stage (applies to the first block of each stage). Default: 2

• bot_mul (float) – bottleneck multiplier for each stage (applies to bottleneck block). Default: 1.0

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

• drop_path_prob (float) – drop path probability. Default: 0.0

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

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

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

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

class basecls.models.RepVGG(num_blocks, width_multiplier, head=None, groups=1, se_r=0.0, act_name='relu', deploy=False)

基类：Module

RepVGG Model.

Use RepVGG.convert_to_deploy() to convert a training RepVGG to deploy:

model = RepVGG(..., deploy=False)
model.load_state_dict(...)
_ = RepVGG.convert_to_deploy(model)

参数

• num_blocks (Sequence[int]) – RepVGG depths.

• width_multiplier (Sequence[int]) – RepVGG widths, base_width is [64, 128, 256, 512].

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

• groups (Union[int, List[Union[int, List[int]]]]) – number of groups for blocks. Default: 1

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

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

• deploy (bool) – switch a reparamed RepVGG into deploy mode. Default: False

forward(x)

classmethod convert_to_deploy(module)

返回类型

Module

class basecls.models.ResMLP(img_size=224, patch_size=16, in_chans=3, embed_dim=768, depth=12, drop_rate=0.0, drop_path_rate=0.0, embed_layer=PatchEmbed, init_scale=1e-4, ffn_ratio=4.0, act_name='gelu', num_classes=1000, **kwargs)

基类：Module

ResMLP model.

参数

• img_size (int) – Input image size. Default: 224

• patch_size (int) – Patch token size. Default: 16

• in_chans (int) – Number of input image channels. Default: 3

• embed_dim (int) – Number of linear projection output channels. Default: 768

• depth (int) – Depth of Transformer Encoder layer. Default: 12

• drop_rate (float) – Dropout rate. Default: 0.0

• drop_path_rate (float) – Stochastic depth rate. Default: 0.0

• embed_layer (Module) – Patch embedding layer. Default: PatchEmbed

• init_scale (float) – Initial value for LayerScale. Default: 1e-4

• ffn_ratio (float) – Ratio of ffn hidden dim to embedding dim. Default: 4.0

• act_name (str) – Activation function. Default: "gelu"

• num_classes (int) – Number of classes. Default: 1000

forward(x)

class basecls.models.ResNet(stem_name, stem_w, block_name, depths, widths, strides, bot_muls=1.0, group_ws=None, se_r=0.0, avg_down=False, drop_path_prob=0.0, zero_init_final_gamma=False, norm_name='BN', act_name='relu', head=None)

基类：Module

ResNet model.

参数

• stem_name (Union[str, Callable]) – stem name.

• stem_w (int) – stem width.

• block_name (Union[str, Callable]) – block name.

• depths (Sequence[int]) – depth for each stage (number of blocks in the stage).

• widths (Sequence[int]) – width for each stage (width of each block in the stage).

• strides (Sequence[int]) – strides for each stage (applies to the first block of each stage).

• bot_muls (Union[float, Sequence[float]]) – bottleneck multipliers for each stage (applies to bottleneck block). Default: 1.0

• group_ws (Optional[Sequence[int]]) – group widths for each stage (applies to bottleneck block). Default: None

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

• drop_path_prob (float) – drop path probability. Default: 0.0

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

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

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

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

forward(x)

static get_stem_func(name)

Retrieves the stem function by name.

static get_block_func(name)

Retrieves the block function by name.

class basecls.models.SNetV2(block, stem_w, depths, widths, strides, kernels, use_maxpool=True, se_r=0.0, drop_path_prob=0.0, norm_name='BN', act_name='relu', head=None)

基类：Module

ShufflenetV2 model.

参数

• block (Callable) – building block to use, SNV2XceptionBlock for v2+.

• stem_w (int) – width for stem layer.

• depths (Sequence[int]) – depth for each stage (number of blocks in the stage).

• widths (Sequence[int]) – width for each stage (width of each block in the stage).

• strides (Sequence[int]) – strides for each stage (applies to the first block of each stage).

• kernels (Sequence[int]) – kernel sizes for each stage.

• use_maxpool (bool) – whether use maxpool stride 2 after stem. Default: True

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

• drop_path_prob (float) – drop path probability. Default: 0.0

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

• act_name (str) – activation function. Default: "relu6"

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

forward(x)

class basecls.models.SwinTransformer(img_size=224, patch_size=4, in_chans=3, embed_dim=96, depths=[2, 2, 6, 2], num_heads=[3, 6, 12, 24], window_size=7, ffn_ratio=4.0, qkv_bias=True, qk_scale=None, ape=False, patch_norm=True, drop_rate=0.0, attn_drop_rate=0.0, drop_path_rate=0.1, embed_layer=PatchEmbed, norm_name='LN', act_name='gelu', num_classes=1000, **kwargs)

基类：Module

Swin Transformer

A PyTorch impl of :

Swin Transformer: Hierarchical Vision Transformer using Shifted Windows - https://arxiv.org/pdf/2103.14030

参数

• img_size (int) – Input image size. Default: 224

• patch_size (int) – Patch size. Default: 4

• in_chans (int) – Number of input image channels. Default: 3

• embed_dim (int) – Patch embedding dimension. Default: 96

• depths (Sequence[int]) – Depth of each Swin Transformer layer.

• num_heads (Sequence[int]) – Number of attention heads in different layers.

• window_size (int) – Window size. Default: 7

• ffn_ratio (float) – Ratio of ffn hidden dim to embedding dim. Default: 4.0

• qkv_bias (bool) – If True, add a learnable bias to query, key, value. Default: True

• qk_scale (Optional[float]) – Override default qk scale of head_dim ** -0.5 if set. Default: None

• ape (bool) – If True, add absolute position embedding to the patch embedding. Default: False

• patch_norm (bool) – If True, add normalization after patch embedding. Default: True

• drop_rate (float) – Dropout rate. Default: 0

• attn_drop_rate (float) – Attention dropout rate. Default: 0

• drop_path_rate (float) – Stochastic depth rate. Default: 0.1

• embed_layer (Module) – Patch embedding layer. Default: PatchEmbed

• norm_name (str) – Normalization layer. Default: "LN"

• act_name (str) – Activation layer. Default: "gelu"

• num_classes (int) – Number of classes for classification head. Default: 1000

forward(x)

class basecls.models.VGG(depths, widths, norm_name=None, act_name='relu', head=None)

基类：Module

VGG model.

参数

• depths (Sequence[int]) – depth for each stage (number of blocks in the stage).

• widths (Sequence[int]) – width for each stage (width of each block in the stage).

• norm_name (Optional[str]) – normalization function. Default: None

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

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

forward(x)

返回类型

Tensor

class basecls.models.ViT(img_size=224, patch_size=16, in_chans=3, embed_dim=768, depth=12, num_heads=12, ffn_ratio=4.0, qkv_bias=True, qk_scale=None, representation_size=None, distilled=False, drop_rate=0.0, attn_drop_rate=0.0, drop_path_rate=0.0, embed_layer=PatchEmbed, norm_name='LN', act_name='gelu', num_classes=1000, **kwargs)

基类：Module

ViT model.

参数

• img_size (int) – Input image size. Default: 224

• patch_size (int) – Patch token size. Default: 16

• in_chans (int) – Number of input image channels. Default: 3

• embed_dim (int) – Number of linear projection output channels. Default: 768

• depth (int) – Depth of Transformer Encoder layer. Default: 12

• num_heads (int) – Number of attention heads. Default: 12

• ffn_ratio (float) – Ratio of ffn hidden dim to embedding dim. Default: 4.0

• qkv_bias (bool) – If True, add a learnable bias to query, key, value. Default: True

• qk_scale (Optional[float]) – Override default qk scale of head_dim ** -0.5 if set. Default: None

• representation_size (Optional[int]) – Size of representation layer (pre-logits). Default: None

• distilled (bool) – Includes a distillation token and head. Default: False

• drop_rate (float) – Dropout rate. Default: 0.0

• attn_drop_rate (float) – Attention dropout rate. Default: 0.0

• drop_path_rate (float) – Stochastic depth rate. Default: 0.0

• embed_layer (Module) – Patch embedding layer. Default: PatchEmbed

• norm_name (str) – Normalization layer. Default: "LN"

• act_name (str) – Activation function. Default: "gelu"

• num_classes (int) – Number of classes. Default: 1000

init_weights()

forward(x)

load_state_dict(state_dict, strict=True)

Loads a given dictionary created by state_dict() into this module. If strict is True, the keys of state_dict() must exactly match the keys returned by state_dict().

Users can also pass a closure: Function[key: str, var: Tensor] -> Optional[np.ndarray] as a state_dict, in order to handle complex situations. For example, load everything except for the final linear classifier:

state_dict = {...}  #  Dict[str, np.ndarray]
model.load_state_dict({
    k: None if k.startswith('fc') else v
    for k, v in state_dict.items()
}, strict=False)

Here returning None means skipping parameter k.

To prevent shape mismatch (e.g. load PyTorch weights), we can reshape before loading:

state_dict = {...}
def reshape_accordingly(k, v):
    return state_dict[k].reshape(v.shape)
model.load_state_dict(reshape_accordingly)

We can also perform inplace re-initialization or pruning:

def reinit_and_pruning(k, v):
    if 'bias' in k:
        M.init.zero_(v)
    if 'conv' in k:

• basecls.models.build
  • build_model()
  • load_model()
  • sync_model()
• basecls.models.effnet
  • FuseMBConv
    • FuseMBConv.forward()
  • EffNet
    • EffNet.forward()
    • EffNet.get_block_func()
• basecls.models.hrnet
  • UpsampleNearest
    • UpsampleNearest.forward()
  • HRFusion
    • HRFusion.forward()
  • HRModule
    • HRModule.forward()
  • HRTrans
    • HRTrans.forward()
  • HRStage
    • HRStage.forward()
  • HRMerge
    • HRMerge.forward()
  • HRNet
    • HRNet.forward()
• basecls.models.mbnet
  • MBConv
    • MBConv.forward()
  • MBStage
    • MBStage.forward()
  • MBNet
    • MBNet.forward()
• basecls.models.regnet
  • RegBottleneckBlock
    • RegBottleneckBlock.forward()
  • RegNet
• basecls.models.repvgg
  • RepVGGBlock
    • RepVGGBlock.forward()
    • RepVGGBlock.convert_to_deploy()
  • RepVGG
    • RepVGG.forward()
    • RepVGG.convert_to_deploy()
• basecls.models.resmlp
  • Affine
    • Affine.forward()
  • ResMLPBlock
    • ResMLPBlock.forward()
  • ResMLP
    • ResMLP.forward()
• basecls.models.resnet
  • ResBasicBlock
    • ResBasicBlock.forward()
  • ResBottleneckBlock
    • ResBottleneckBlock.forward()
  • ResDeepStem
    • ResDeepStem.forward()
  • ResStem
    • ResStem.forward()
  • SimpleStem
    • SimpleStem.forward()
  • AnyStage
    • AnyStage.forward()
  • ResNet
    • ResNet.forward()
    • ResNet.get_stem_func()
    • ResNet.get_block_func()
• basecls.models.snet
  • SNV2Block
    • SNV2Block.forward()
  • SNV2XceptionBlock
  • SNetV2
    • SNetV2.forward()
• basecls.models.swin
  • window_partition()
  • window_reverse()
  • WindowAttention
    • WindowAttention.forward()
  • PatchMerging
    • PatchMerging.forward()
  • SwinBlock
    • SwinBlock.forward()
  • SwinBasicLayer
    • SwinBasicLayer.forward()
  • SwinTransformer
    • SwinTransformer.forward()
• basecls.models.vgg
  • VGGStage
    • VGGStage.forward()
  • VGG
    • VGG.forward()
• basecls.models.vit
  • PatchEmbed
    • PatchEmbed.forward()
  • Attention
    • Attention.forward()
  • FFN
    • FFN.forward()
  • EncoderBlock
    • EncoderBlock.forward()
  • ViT
    • ViT.init_weights()
    • ViT.forward()
    • ViT.load_state_dict()