tf-keras-vis is a visualization toolkit for debugging Keras models with Tensorflow2, but not original Keras.
The features of tf-keras-vis are based on keras-vis, but tf-keras-vis's APIs doesn't have compatibility with keras-vis's because, instead of getting it, we prioritized to get following features.
- Support processing multipul images at a time as a batch
- Support tf.keras.Model that has multipul inputs (and, of course, multipul outpus too)
- Allow to use optimizers that embeded in tf.keras
And then we will add some algorisms such as below.
- SmoothGrad: removing noise by adding noise (DONE)
- Deep Dream
- Style transfer
- Python 3.5, 3.6, 3.7 or 3.8
- tensorflow>=2.0.0
- PyPI
$ pip install tf-keras-vis tensorflow
- Docker
$ docker pull keisen/tf-keras-vis:0.2.4
You can find other images (that's nvidia-docker images) at dockerhub.
Please see examples/attentions.ipynb, examples/visualize_dense_layer.ipynb and examples/visualize_conv_filters.ipynb.
- Run Jupyter notebooks on Docker
$ docker run -itd -v /PATH/TO/tf-keras-vis:/tf-keras-vis -p 8888:8888 keisen/tf-keras-vis:0.2.4
Or, if you have GPU processors,
$ docker run -itd --runtime=nvidia -v /PATH/TO/tf-keras-vis:/tf-keras-vis -p 8888:8888 keisen/tf-keras-vis:0.2.4-gpu
T.B.D.
- With InceptionV3, ActivationMaximization doesn't work well, that's, it might generate meanninglessly bulr image.
- With cascading model, Gradcam doesn't work well, that's, it might occur some error.
- Unsupported
channels-first
models and datas.