torchextractor: PyTorch Intermediate Feature Extraction

Introduction

Too many times some model definitions get remorselessly copy-pasted just because the forward function does not return what the person expects. You provide module names and torchextractor takes care of the extraction for you.It's never been easier to extract feature, add an extra loss or plug another head to a network. Ler us know what amazing things you build with torchextractor!

Installation

pip install torchextractor  # stable
pip install git+https://github.com/antoinebrl/torchextractor.git  # latest

Requirements:

• Python >= 3.6+
• torch >= 1.4.0

Usage

import torch
import torchvision
import torchextractor as tx

model = torchvision.models.resnet18(pretrained=True)
model = tx.Extractor(model, ["layer1", "layer2", "layer3", "layer4"])
dummy_input = torch.rand(7, 3, 224, 224)
model_output, features = model(dummy_input)
feature_shapes = {name: f.shape for name, f in features.items()}
print(feature_shapes)

# {
#   'layer1': torch.Size([1, 64, 56, 56]),
#   'layer2': torch.Size([1, 128, 28, 28]),
#   'layer3': torch.Size([1, 256, 14, 14]),
#   'layer4': torch.Size([1, 512, 7, 7]),
# }

See more examples

Read the documentation

FAQ

• How do I know the names of the modules?

You can print all module names like this:

tx.list_module_names(model)

# OR

for name, module in model.named_modules():
    print(name)

• Why do some operations not get listed?

It is not possible to add hooks if operations are not defined as modules. Therefore, F.relu cannot be captured but nn.Relu() can.

• How can I avoid listing all relevant modules?

You can specify a custom filtering function to hook the relevant modules:

# Hook everything !
module_filter_fn = lambda module, name: True

# Capture of all modules inside first layer
module_filter_fn = lambda module, name: name.startswith("layer1")

# Focus on all convolutions
module_filter_fn = lambda module, name: isinstance(module, torch.nn.Conv2d)

model = tx.Extractor(model, module_filter_fn=module_filter_fn)

• Is it compatible with ONNX?

tx.Extractor is compatible with ONNX! This means you can also access intermediate features maps after the export.

Pro-tip: name the output nodes by using output_names when calling torch.onnx.export.

• Is it compatible with TorchScript?

Bad news, TorchScript cannot take variable number of arguments and keyword-only arguments.

Good news, there is a workaround! The solution is to overwrite the forward function of tx.Extractor to replicate the interface of the model.

import torch
import torchvision
import torchextractor as tx

class MyExtractor(tx.Extractor):
    def forward(self, x1, x2, x3):
        # Assuming the model takes x1, x2 and x3 as input
        output = self.model(x1, x2, x3)
        return output, self.feature_maps

model = torchvision.models.resnet18(pretrained=True)
model = MyExtractor(model, ["layer1", "layer2", "layer3", "layer4"])
model_traced = torch.jit.script(model)

• "One more thing!" :wink: By default we capture the latest output of the relevant modules, but you can specify your own custom operations.

For example, to accumulate features over 10 forward passes you can do the following:

import torch
import torchvision
import torchextractor as tx

model = torchvision.models.resnet18(pretrained=True)

def capture_fn(module, input, output, module_name, feature_maps):
    if module_name not in feature_maps:
        feature_maps[module_name] = []
    feature_maps[module_name].append(output)

extractor = tx.Extractor(model, ["layer3", "layer4"], capture_fn=capture_fn)

for i in range(20):
    for i in range(10):
        x = torch.rand(7, 3, 224, 224)
        model(x)
    feature_maps = extractor.collect()

    # Do your stuffs here

    # Discard collected elements
    extractor.clear_placeholder()

Contributing

All feedbacks and contributions are welcomed. Feel free to report an issue or to create a pull request!

If you want to get hands-on:

1. (Fork and) clone the repo.
2. Create a virtual environment: virtualenv -p python3 .venv && source .venv/bin/activate
3. Install dependencies: pip install -r requirements.txt && pip install -r requirements-dev.txt
4. Hook auto-formatting tools: pre-commit install
5. Hack as much as you want!
6. Run tests: python -m unittest discover -vs ./tests/
7. Share your work and create a pull request.

To Build documentation:

cd docs
pip install requirements.txt
make html