Pretrained models for Pytorch

0.7.4 verified Mon Apr 27 auth: no python maintenance

A library providing pretrained models for PyTorch, including architectures like ResNet, DenseNet, Inception, and more. Version 0.7.4 is the latest release, though the project is in maintenance mode with infrequent updates. It is commonly used for transfer learning and feature extraction, but users should prefer torchvision's model zoo for active support.

pip install pretrainedmodels

Common errors

error AttributeError: module 'torch' has no attribute 'irfft' ↓

cause PyTorch 1.9 removed deprecated torch.irfft; used by some models.

fix

Use an older PyTorch version (<=1.8) or avoid models that use irfft (e.g., pnasnet).

error ImportError: cannot import name 'TransformImage' from 'pretrainedmodels.utils' ↓

cause Missing torchvision dependency or incorrect import path.

fix

Install torchvision: pip install torchvision, then from pretrainedmodels import utils.

error RuntimeError: CUDA error: out of memory ↓

cause Model too large for GPU memory or batch size too big.

fix

Reduce batch size, use model.half() for half precision, or switch to CPU.

error SSLError: [SSL: CERTIFICATE_VERIFY_FAILED] certificate verify failed ↓

cause SSL issues when downloading pretrained weights from the internet.

fix

Set environment variable CURL_CA_BUNDLE to a valid cert bundle or use pretrainedmodels.models.__dict__['model_name'](pretrained=False) to skip download.

Warnings

deprecated pretrainedmodels is in maintenance mode; many models are outdated and may not work with newer PyTorch versions. Prefer torchvision's model zoo or timm. ↓

fix Switch to torchvision.models or the timm library.

gotcha The pretrained='imagenet' argument may fail silently if checkpoint not found. Always verify the model downloads correctly. ↓

fix Check network connectivity or manually download pretrained weights from the official repository.

gotcha Model input sizes vary; not all models accept 224x224 images. Use utils.TransformImage to get the correct preprocessing. ↓

fix Always use pretrainedmodels.utils.TransformImage(model) to get the right transforms.

breaking PyTorch 1.9+ may cause compatibility issues with older model definitions. You may see AttributeError: module 'torch' has no attribute 'irfft'. ↓

fix Use torchvision or timm instead. Or pin PyTorch to 1.8.x.

Imports

pretrainedmodels
wrong
```
from pretrainedmodels import ...
```
correct
```
import pretrainedmodels
```
Direct import is fine, but common mistake is to import submodules incorrectly; use pretrainedmodels.__dict__ to list models.
models
```
from pretrainedmodels import models
```
To access model constructors like models.resnet101()

Quickstart

Load a pretrained ResNet-101 model and run inference.

import torch
import pretrainedmodels

model = pretrainedmodels.__dict__['resnet101'](pretrained='imagenet')
model.eval()
print(model)

# Example inference
from pretrainedmodels import utils
import torchvision.transforms as transforms

tf = utils.TransformImage(model)
input_tensor = torch.randn(1, 3, 224, 224)
out = model(input_tensor)
print(out.shape)