Practical overview
When it fits
Use it when building a PyTorch visual-recognition model and you want a documented convolution-and-attention layer that can be placed in a ResNet-style feature stack.
Who it is for
- PyTorch developers building image-classification or other visual-recognition models.
- Researchers and practitioners adapting ResNet-style backbones with a bottleneck transformer layer.
Documented uses
- Insert a BottleStack layer into a model that processes image feature maps.
- Replace part of a ResNet-50 backbone to assemble the documented BotNet-style image classifier.
- Configure channel dimensions, feature-map size, attention heads, downsampling and positional embeddings for the inserted layer.
Requirements & limits
The public GitHub repository lists an MIT license and provides a pip installation command.
- BottleStack operates on feature maps; the caller must configure dimensions to match the model stage.
- The README's standalone example uses absolute positional embeddings by setting rel_pos_emb to false, while the ResNet-50 example enables relative positional embeddings.
- The documented ResNet-50 integration uses torchvision and a 224-by-224 image input.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Add BottleStack to a ResNet-50 classifier
Construct the documented BotNet-style model that replaces part of a ResNet-50 backbone with BottleStack.
Before you start
- Python with PyTorch and torchvision available.
- The bottleneck-transformer-pytorch package installed.
- A ResNet-50 backbone and input images sized 224 by 224, as used in the example.
Steps
Import torch, torch.nn, torchvision.models.resnet50 and BottleStack.
Configure BottleStack with dim 256, fmap_size 56, dim_out 2048, proj_factor 4, downsample enabled, four heads, dim_head 128, relative positional embeddings enabled and ReLU activation.
Create resnet50, take its first five children, and place those layers followed by BottleStack, adaptive average pooling, flattening and a 2048-to-1000 linear layer in nn.Sequential.
Pass a batch-shaped image tensor with the example shape (2, 3, 224, 224) to the model.
Expected outcome
The README's example produces a prediction tensor shaped (2, 1000).
Guide limitations
- The example shows model assembly and a forward pass, not a training loop or dataset setup.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 16 October 2026
- Original website title
- lucidrains/bottleneck-transformer-pytorch · Official homepage title · retained in its source language