The PoolFormer model was proposed in MetaFormer is Actually What You Need for Vision by Sea AI Labs. Instead of designing complicated token mixer to achieve SOTA performance, the target of this work is to demonstrate the competence of transformer models largely stem from the general architecture MetaFormer.
The abstract from the paper is the following:
Transformers have shown great potential in computer vision tasks. A common belief is their attention-based token mixer module contributes most to their competence. However, recent works show the attention-based module in transformers can be replaced by spatial MLPs and the resulted models still perform quite well. Based on this observation, we hypothesize that the general architecture of the transformers, instead of the specific token mixer module, is more essential to the model’s performance. To verify this, we deliberately replace the attention module in transformers with an embarrassingly simple spatial pooling operator to conduct only the most basic token mixing. Surprisingly, we observe that the derived model, termed as PoolFormer, achieves competitive performance on multiple computer vision tasks. For example, on ImageNet-1K, PoolFormer achieves 82.1% top-1 accuracy, surpassing well-tuned vision transformer/MLP-like baselines DeiT-B/ResMLP-B24 by 0.3%/1.1% accuracy with 35%/52% fewer parameters and 48%/60% fewer MACs. The effectiveness of PoolFormer verifies our hypothesis and urges us to initiate the concept of “MetaFormer”, a general architecture abstracted from transformers without specifying the token mixer. Based on the extensive experiments, we argue that MetaFormer is the key player in achieving superior results for recent transformer and MLP-like models on vision tasks. This work calls for more future research dedicated to improving MetaFormer instead of focusing on the token mixer modules. Additionally, our proposed PoolFormer could serve as a starting baseline for future MetaFormer architecture design.
The figure below illustrates the architecture of PoolFormer. Taken from the original paper.
This model was contributed by heytanay. The original code can be found here.
| Model variant | Depths | Hidden sizes | Params (M) | ImageNet-1k Top 1 |
|---|---|---|---|---|
| s12 | [2, 2, 6, 2] | [64, 128, 320, 512] | 12 | 77.2 |
| s24 | [4, 4, 12, 4] | [64, 128, 320, 512] | 21 | 80.3 |
| s36 | [6, 6, 18, 6] | [64, 128, 320, 512] | 31 | 81.4 |
| m36 | [6, 6, 18, 6] | [96, 192, 384, 768] | 56 | 82.1 |
| m48 | [8, 8, 24, 8] | [96, 192, 384, 768] | 73 | 82.5 |
A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with PoolFormer.
If you’re interested in submitting a resource to be included here, please feel free to open a Pull Request and we’ll review it! The resource should ideally demonstrate something new instead of duplicating an existing resource.
( num_channels = 3 patch_size = 16 stride = 16 pool_size = 3 mlp_ratio = 4.0 depths = [2, 2, 6, 2] hidden_sizes = [64, 128, 320, 512] patch_sizes = [7, 3, 3, 3] strides = [4, 2, 2, 2] padding = [2, 1, 1, 1] num_encoder_blocks = 4 drop_path_rate = 0.0 hidden_act = 'gelu' use_layer_scale = True layer_scale_init_value = 1e-05 initializer_range = 0.02 **kwargs )
Parameters
int, optional, defaults to 3) —
The number of channels in the input image. int, optional, defaults to 16) —
The size of the input patch. int, optional, defaults to 16) —
The stride of the input patch. int, optional, defaults to 3) —
The size of the pooling window. float, optional, defaults to 4.0) —
The ratio of the number of channels in the output of the MLP to the number of channels in the input. list, optional, defaults to [2, 2, 6, 2]) —
The depth of each encoder block. list, optional, defaults to [64, 128, 320, 512]) —
The hidden sizes of each encoder block. list, optional, defaults to [7, 3, 3, 3]) —
The size of the input patch for each encoder block. list, optional, defaults to [4, 2, 2, 2]) —
The stride of the input patch for each encoder block. list, optional, defaults to [2, 1, 1, 1]) —
The padding of the input patch for each encoder block. int, optional, defaults to 4) —
The number of encoder blocks. float, optional, defaults to 0.0) —
The dropout rate for the dropout layers. str, optional, defaults to "gelu") —
The activation function for the hidden layers. bool, optional, defaults to True) —
Whether to use layer scale. float, optional, defaults to 1e-05) —
The initial value for the layer scale. float, optional, defaults to 0.02) —
The initializer range for the weights. This is the configuration class to store the configuration of PoolFormerModel. It is used to instantiate a PoolFormer model according to the specified arguments, defining the model architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of the PoolFormer sail/poolformer_s12 architecture.
Configuration objects inherit from PretrainedConfig and can be used to control the model outputs. Read the documentation from PretrainedConfig for more information.
Example:
>>> from transformers import PoolFormerConfig, PoolFormerModel
>>> # Initializing a PoolFormer sail/poolformer_s12 style configuration
>>> configuration = PoolFormerConfig()
>>> # Initializing a model (with random weights) from the sail/poolformer_s12 style configuration
>>> model = PoolFormerModel(configuration)
>>> # Accessing the model configuration
>>> configuration = model.configPreprocess an image or a batch of images.
( do_resize: bool = True size: Dict = None crop_pct: int = 0.9 resample: Resampling = <Resampling.BICUBIC: 3> do_center_crop: bool = True crop_size: Dict = None rescale_factor: Union = 0.00392156862745098 do_rescale: bool = True do_normalize: bool = True image_mean: Union = None image_std: Union = None **kwargs )
Parameters
bool, optional, defaults to True) —
Whether to resize the image’s (height, width) dimensions to the specified size. Can be overridden by
do_resize in the preprocess method. Dict[str, int] optional, defaults to {"shortest_edge" -- 224}):
Size of the image after resizing. Can be overridden by size in the preprocess method. If crop_pct is
unset:
{"height": h, "width": w}: the image is resized to (h, w).{"shortest_edge": s}: the shortest edge of the image is resized to s whilst maintaining the
aspect ratio.If crop_pct is set:
{"height": h, "width": w}: the image is resized to (int(floor(h/crop_pct)), int(floor(w/crop_pct))){"height": c, "width": c}: the shortest edge of the image is resized to int(floor(c/crop_pct)
whilst maintaining the aspect ratio.{"shortest_edge": c}: the shortest edge of the image is resized to int(floor(c/crop_pct)
whilst maintaining the aspect ratio.float, optional, defaults to 0.9) —
Percentage of the image to crop from the center. Can be overridden by crop_pct in the preprocess
method. PILImageResampling, optional, defaults to Resampling.BICUBIC) —
Resampling filter to use if resizing the image. Can be overridden by resample in the preprocess method. bool, optional, defaults to True) —
Whether to center crop the image. If the input size is smaller than crop_size along any edge, the image
is padded with 0’s and then center cropped. Can be overridden by do_center_crop in the preprocess
method. Dict[str, int], optional, defaults to {"height" -- 224, "width": 224}):
Size of the image after applying center crop. Only has an effect if do_center_crop is set to True. Can
be overridden by the crop_size parameter in the preprocess method. int or float, optional, defaults to 1/255) —
Scale factor to use if rescaling the image. Can be overridden by the rescale_factor parameter in the
preprocess method. bool, optional, defaults to True) —
Whether to rescale the image by the specified scale rescale_factor. Can be overridden by the do_rescale
parameter in the preprocess method. bool, optional, defaults to True) —
Controls whether to normalize the image. Can be overridden by the do_normalize parameter in the
preprocess method. float or List[float], optional, defaults to IMAGENET_STANDARD_MEAN) —
Mean to use if normalizing the image. This is a float or list of floats the length of the number of
channels in the image. Can be overridden by the image_mean parameter in the preprocess method. float or List[float], optional, defaults to IMAGENET_STANDARD_STD) —
Standard deviation to use if normalizing the image. This is a float or list of floats the length of the
number of channels in the image. Can be overridden by the image_std parameter in the preprocess method. Constructs a PoolFormer image processor.
( images: Union do_resize: bool = None size: Dict = None crop_pct: int = None resample: Resampling = None do_center_crop: bool = None crop_size: Dict = None do_rescale: bool = None rescale_factor: float = None do_normalize: bool = None image_mean: Union = None image_std: Union = None return_tensors: Union = None data_format: ChannelDimension = <ChannelDimension.FIRST: 'channels_first'> input_data_format: Union = None **kwargs )
Parameters
ImageInput) —
Image to preprocess. Expects a single or batch of images with pixel values ranging from 0 to 255. If
passing in images with pixel values between 0 and 1, set do_rescale=False. bool, optional, defaults to self.do_resize) —
Whether to resize the image. Dict[str, int], optional, defaults to self.size) —
Size of the image after applying resize. float, optional, defaults to self.crop_pct) —
Percentage of the image to crop. Only has an effect if do_resize is set to True. int, optional, defaults to self.resample) —
Resampling filter to use if resizing the image. This can be one of the enum PILImageResampling, Only
has an effect if do_resize is set to True. bool, optional, defaults to self.do_center_crop) —
Whether to center crop the image. Dict[str, int], optional, defaults to self.crop_size) —
Size of the image after applying center crop. bool, optional, defaults to self.do_rescale) —
Whether to rescale the image values between [0 - 1]. float, optional, defaults to self.rescale_factor) —
Rescale factor to rescale the image by if do_rescale is set to True. bool, optional, defaults to self.do_normalize) —
Whether to normalize the image. float or List[float], optional, defaults to self.image_mean) —
Image mean. float or List[float], optional, defaults to self.image_std) —
Image standard deviation. str or TensorType, optional) —
The type of tensors to return. Can be one of:np.ndarray.TensorType.TENSORFLOW or 'tf': Return a batch of type tf.Tensor.TensorType.PYTORCH or 'pt': Return a batch of type torch.Tensor.TensorType.NUMPY or 'np': Return a batch of type np.ndarray.TensorType.JAX or 'jax': Return a batch of type jax.numpy.ndarray.ChannelDimension or str, optional, defaults to ChannelDimension.FIRST) —
The channel dimension format for the output image. Can be one of:ChannelDimension.FIRST: image in (num_channels, height, width) format.ChannelDimension.LAST: image in (height, width, num_channels) format.ChannelDimension or str, optional) —
The channel dimension format for the input image. If unset, the channel dimension format is inferred
from the input image. Can be one of:"channels_first" or ChannelDimension.FIRST: image in (num_channels, height, width) format."channels_last" or ChannelDimension.LAST: image in (height, width, num_channels) format."none" or ChannelDimension.NONE: image in (height, width) format.Preprocess an image or batch of images.
( config )
Parameters
The bare PoolFormer Model transformer outputting raw hidden-states without any specific head on top. This model is a PyTorch torch.nn.Module sub-class. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior.
( pixel_values: Optional = None output_hidden_states: Optional = None return_dict: Optional = None ) → transformers.modeling_outputs.BaseModelOutputWithNoAttention or tuple(torch.FloatTensor)
Parameters
torch.FloatTensor of shape (batch_size, num_channels, height, width)) —
Pixel values. Pixel values can be obtained using AutoImageProcessor. See
PoolFormerImageProcessor.call() for details. Returns
transformers.modeling_outputs.BaseModelOutputWithNoAttention or tuple(torch.FloatTensor)
A transformers.modeling_outputs.BaseModelOutputWithNoAttention or a tuple of
torch.FloatTensor (if return_dict=False is passed or when config.return_dict=False) comprising various
elements depending on the configuration (PoolFormerConfig) and inputs.
last_hidden_state (torch.FloatTensor of shape (batch_size, num_channels, height, width)) — Sequence of hidden-states at the output of the last layer of the model.
hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of torch.FloatTensor (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape (batch_size, num_channels, height, width).
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
The PoolFormerModel forward method, overrides the __call__ special method.
Although the recipe for forward pass needs to be defined within this function, one should call the Module
instance afterwards instead of this since the former takes care of running the pre and post processing steps while
the latter silently ignores them.
Example:
>>> from transformers import AutoImageProcessor, PoolFormerModel
>>> import torch
>>> from datasets import load_dataset
>>> dataset = load_dataset("huggingface/cats-image")
>>> image = dataset["test"]["image"][0]
>>> image_processor = AutoImageProcessor.from_pretrained("sail/poolformer_s12")
>>> model = PoolFormerModel.from_pretrained("sail/poolformer_s12")
>>> inputs = image_processor(image, return_tensors="pt")
>>> with torch.no_grad():
... outputs = model(**inputs)
>>> last_hidden_states = outputs.last_hidden_state
>>> list(last_hidden_states.shape)
[1, 512, 7, 7]( config )
Parameters
PoolFormer Model transformer with an image classification head on top
This model is a PyTorch torch.nn.Module sub-class. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior.
( pixel_values: Optional = None labels: Optional = None output_hidden_states: Optional = None return_dict: Optional = None ) → transformers.modeling_outputs.ImageClassifierOutputWithNoAttention or tuple(torch.FloatTensor)
Parameters
torch.FloatTensor of shape (batch_size, num_channels, height, width)) —
Pixel values. Pixel values can be obtained using AutoImageProcessor. See
PoolFormerImageProcessor.call() for details. torch.LongTensor of shape (batch_size,), optional) —
Labels for computing the image classification/regression loss. Indices should be in [0, ..., config.num_labels - 1]. If config.num_labels == 1 a regression loss is computed (Mean-Square loss), If
config.num_labels > 1 a classification loss is computed (Cross-Entropy). Returns
transformers.modeling_outputs.ImageClassifierOutputWithNoAttention or tuple(torch.FloatTensor)
A transformers.modeling_outputs.ImageClassifierOutputWithNoAttention or a tuple of
torch.FloatTensor (if return_dict=False is passed or when config.return_dict=False) comprising various
elements depending on the configuration (PoolFormerConfig) and inputs.
torch.FloatTensor of shape (1,), optional, returned when labels is provided) — Classification (or regression if config.num_labels==1) loss.torch.FloatTensor of shape (batch_size, config.num_labels)) — Classification (or regression if config.num_labels==1) scores (before SoftMax).tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of torch.FloatTensor (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each stage) of shape (batch_size, num_channels, height, width). Hidden-states (also
called feature maps) of the model at the output of each stage.The PoolFormerForImageClassification forward method, overrides the __call__ special method.
Although the recipe for forward pass needs to be defined within this function, one should call the Module
instance afterwards instead of this since the former takes care of running the pre and post processing steps while
the latter silently ignores them.
Example:
>>> from transformers import AutoImageProcessor, PoolFormerForImageClassification
>>> import torch
>>> from datasets import load_dataset
>>> dataset = load_dataset("huggingface/cats-image")
>>> image = dataset["test"]["image"][0]
>>> image_processor = AutoImageProcessor.from_pretrained("sail/poolformer_s12")
>>> model = PoolFormerForImageClassification.from_pretrained("sail/poolformer_s12")
>>> inputs = image_processor(image, return_tensors="pt")
>>> with torch.no_grad():
... logits = model(**inputs).logits
>>> # model predicts one of the 1000 ImageNet classes
>>> predicted_label = logits.argmax(-1).item()
>>> print(model.config.id2label[predicted_label])
tabby, tabby cat