A library for pytorch layer testing.
Project description
tencheck
tencheck
provides a simple set of utilities for analyzing and validating the properties of layers of deep neural nets.
It's typically quite difficult to validate that a layer "behaves properly" (oftentimes, the final barometer is simply how well a model performs with the layer included), and many brittle unit tests have been written involving randomly instantiated tensors and is_close
checks. We believe a good "first line of defense" for neural nets is to create a suite of properties that can be asserted about a layer, while requiring minimal effort per layer in order to do so.
We think there are two aspects of property-based testing that are quite useful to take inspiration from:
- Automatically generating inputs (and generating inputs of variable sizes and values to elucidate properties of interest).
- Evaluating properties based on the maintenance of invariants instead of attempting to exactly match values (which is particularly difficult to interpret in deep neural nets).
However, an important difference is that the properties of interest are generally fairly generic and often shared between layers, while the input generation strategies are pretty similar (they're all tensors). So the focus of tencheck
is to provide:
- An (attempted) universal input generation harness.
- A variety of interesting properties.
- Three modalities: assertion, analysis, and profiling.
The following requirements need to be met for tencheck
to work:
- Your layers are implemented in torch.
- The
.forward()
method is annotated with jaxtyping.
Backlog
- For profiling, use a grid of input sizes to generate performance curves.
- Pick a flop counter and use for profiling
- Tensor container types include more options like dataclasses.
- Auto-generate simple hyperparameters for layer instantiation.
- Refine dtype mapping and coherence.
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
File details
Details for the file tencheck-0.0.4.tar.gz
.
File metadata
- Download URL: tencheck-0.0.4.tar.gz
- Upload date:
- Size: 11.5 kB
- Tags: Source
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/5.1.1 CPython/3.11.4
File hashes
Algorithm | Hash digest | |
---|---|---|
SHA256 | 6c42d61da716917a70f57db4074eb9ca3082c56458e5a79f4b180273849a31d9 |
|
MD5 | 89ffb237d434901c94d59f0f352da885 |
|
BLAKE2b-256 | 45109a27bd617c76c89863ab940229f3d7671c5b2356e26989ab4067cb7facf3 |
File details
Details for the file tencheck-0.0.4-py3-none-any.whl
.
File metadata
- Download URL: tencheck-0.0.4-py3-none-any.whl
- Upload date:
- Size: 11.7 kB
- Tags: Python 3
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/5.1.1 CPython/3.11.4
File hashes
Algorithm | Hash digest | |
---|---|---|
SHA256 | 1fdb121ae2af14848d6dd6a18349c95b51101680f055beaf5b292615c7a13158 |
|
MD5 | 929ac2b069713eaefdbafd9e7f1cfbc5 |
|
BLAKE2b-256 | b9084fd9861deaff70ce1bf22271769bd417aa894aee3af5cb4b0a6db7c89157 |