CalFlops

0.3.2 verified Mon Apr 27 auth: no python

CalFlops is a PyTorch-based FLOPs, MACs, and parameter counter for neural networks including CNNs, RNNs, GCNs, and Transformers (e.g., BERT, LLaMA). Version 0.3.2 supports custom models. Released as needed on GitHub.

pip install calflops

Common errors

error ModuleNotFoundError: No module named 'calflops' ↓

cause Package not installed or installed in a different environment.

fix

Run 'pip install calflops' in the correct Python environment.

error AttributeError: module 'calflops' has no attribute 'calculate_flops' ↓

cause Outdated version of calflops or wrong import path.

fix

Update calflops: 'pip install --upgrade calflops' and use 'from calflops import calculate_flops'.

error RuntimeError: Expected all tensors to be on the same device, but found at least two devices ↓

cause Model and input tensors are on different devices (CPU vs GPU).

fix

Ensure both model and input tensors are on the same device: model.to(device), input_tensor = input_tensor.to(device).

Warnings

gotcha Torchvision models may include non-parameter operations (like softmax) that are counted differently. Ensure model is in eval mode. ↓

fix Set model.eval() before calling calculate_flops to disable dropout/batchnorm effects.

gotcha If model has multiple inputs (e.g., encoders), use input_constructor instead of input_shape. ↓

fix Pass a callable that returns a tuple of tensors: input_constructor=lambda: (x,).

gotcha FLOPs counting for dynamic architectures (e.g., with torch.where) may be inaccurate due to static graph assumption. ↓

fix If using control flow, test with a simple forward pass first and verify counts are reasonable.

Imports

calculate_flops
wrong
```
from calflops.flops_counter import calculate_flops
```
correct
```
from calflops import calculate_flops
```
Older documentation suggested deep import; the top-level import is simpler and stable.
FlopsCounter
wrong
```
from calflops.counter import FlopsCounter
```
correct
```
from calflops import FlopsCounter
```
Common mistake: FlopsCounter is exposed at package level, not in a submodule.

Quickstart

Compute FLOPs, MACs, and parameters for a ResNet-18 model.

import torch
import torchvision.models as models
from calflops import calculate_flops

model = models.resnet18()
batch_size = 1
input_shape = (batch_size, 3, 224, 224)
flops, macs, params = calculate_flops(
    model=model,
    input_shape=input_shape,
    output_as_string=True,
    output_precision=4
)
print(f"FLOPs: {flops}, MACs: {macs}, Params: {params}")