Note

You are reading the documentation for MMClassification 0.x, which will soon be deprecated at the end of 2022. We recommend you upgrade to MMClassification 1.0 to enjoy fruitful new features and better performance brought by OpenMMLab 2.0. Check the installation tutorial, migration tutorial and changelog for more details.

EfficientFormer

class mmcls.models.EfficientFormer(arch='l1', in_channels=3, pool_size=3, mlp_ratios=4, reshape_last_feat=False, out_indices=- 1, frozen_stages=- 1, act_cfg={'type': 'GELU'}, drop_rate=0.0, drop_path_rate=0.0, use_layer_scale=True, init_cfg=None)

EfficientFormer.

A PyTorch implementation of EfficientFormer introduced by: EfficientFormer: Vision Transformers at MobileNet Speed

Modified from the official repo <https://github.com/snap-research/EfficientFormer>.

Parameters

• arch (str | dict) –

  The model’s architecture. If string, it should be one of architecture in EfficientFormer.arch_settings. And if dict,

  it should include the following 4 keys:

  • layers (list[int]): Number of blocks at each stage.

  • embed_dims (list[int]): The number of channels at each stage.

  • downsamples (list[int]): Has downsample or not in the four stages.

  • vit_num (int): The num of vit blocks in the last stage.

  Defaults to ‘l1’.

• in_channels (int) – The num of input channels. Defaults to 3.

• pool_size (int) – The pooling size of Meta4D blocks. Defaults to 3.

• mlp_ratios (int) – The dimension ratio of multi-head attention mechanism in Meta4D blocks. Defaults to 3.

• reshape_last_feat (bool) –

  Whether to reshape the feature map from (B, N, C) to (B, C, H, W) in the last stage, when the vit-num

  in arch is not 0. Defaults to False. Usually set to True in downstream tasks.

• out_indices (Sequence[int]) – Output from which stages. Defaults to -1.

• frozen_stages (int) – Stages to be frozen (stop grad and set eval mode). -1 means not freezing any parameters. Defaults to -1.

• act_cfg (dict) – The config dict for activation between pointwise convolution. Defaults to dict(type='GELU').

• drop_rate (float) – Dropout rate. Defaults to 0.

• drop_path_rate (float) – Stochastic depth rate. Defaults to 0.

• use_layer_scale (bool) – Whether to use use_layer_scale in MetaFormer block. Defaults to True.

• init_cfg (dict, optional) – Initialization config dict. Defaults to None.

• Example –

  >>> from mmcls.models import EfficientFormer
  >>> import torch
  >>> inputs = torch.rand((1, 3, 224, 224))
  >>> # build EfficientFormer backbone for classification task
  >>> model = EfficientFormer(arch="l1")
  >>> model.eval()
  >>> level_outputs = model(inputs)
  >>> for level_out in level_outputs:
  ...     print(tuple(level_out.shape))
  (1, 448, 49)
  >>> # build EfficientFormer backbone for downstream task
  >>> model = EfficientFormer(
  >>>    arch="l3",
  >>>    out_indices=(0, 1, 2, 3),
  >>>    reshape_last_feat=True)
  >>> model.eval()
  >>> level_outputs = model(inputs)
  >>> for level_out in level_outputs:
  ...     print(tuple(level_out.shape))
  (1, 64, 56, 56)
  (1, 128, 28, 28)
  (1, 320, 14, 14)
  (1, 512, 7, 7)
  

forward(x)

Forward computation.

Parameters

x (tensor | tuple[tensor]) – x could be a Torch.tensor or a tuple of Torch.tensor, containing input data for forward computation.

train(mode=True)

Set module status before forward computation.

Parameters

mode (bool) – Whether it is train_mode or test_mode