Tensorflow 2.0 implementation of Sinusodial Representation networks (SIREN) from the paper Implicit Neural Representations with Periodic Activation Functions.
- Pip install
$ pip install --upgrade tf_siren- Pip install (test support)
$ pip install --upgrade tf_siren[tests]For general usage equivalent to the paper, import and use either SinusodialRepresentationDense or SIRENModel.
from tf_siren import SinusodialRepresentationDense
from tf_siren import SIRENModel
# You can use SinusodialRepresentationDense exactly like you ordinarily use Dense layers.
ip = tf.keras.layers.Input(shape=[2])
x = SinusodialRepresentationDense(32,
activation='sine', # default activation function
w0=1.0)(ip) # w0 represents sin(w0 * x) where x is the input.
model = tf.keras.Model(inputs=ip, outputs=x)
# Or directly use the model class to build a multi layer SIREN
model = SIRENModel(units=256, final_units=3, final_activation='sigmoid',
num_layers=5, w0=1.0, w0_initial=30.0)For the (experimental) kernel scaled variants, import and use either ScaledSinusodialRepresentationDense or ScaledSIRENModel.
from tf_siren import ScaledSinusodialRepresentationDense
from tf_siren import ScaledSIRENModel
# You can use SinusodialRepresentationDense exactly like you ordinarily use Dense layers.
ip = tf.keras.layers.Input(shape=[2])
x = ScaledSinusodialRepresentationDense(32,
scale=1.0 # scale value should be carefully chosen in range [1, 2]
activation='sine', # default activation function
w0=1.0)(ip) # w0 represents sin(w0 * x) where x is the input.
model = tf.keras.Model(inputs=ip, outputs=x)
# Or directly use the model class to build a multi layer Scaled SIREN
model = ScaledSIRENModel(units=256, final_units=3, final_activation='sigmoid', scale=1.0,
num_layers=5, w0=1.0, w0_initial=30.0)A partial implementation of the image inpainting task is available as the train_inpainting_siren.py and eval_inpainting_siren.py scripts inside the scripts directory.
Weight files are made available in the repository under the Release tab of the project. Extract the weights and place the checkpoints folder at the scripts directory
These weights generates the following output after 5000 epochs of training with batch size 8192 while using only 10% of the available pixels in the image during training phase.
If we train for using only 20% of the available pixels in the image during training phase -
If we train for using only 30% of the available pixels in the image during training phase -
We can use a Hyper Network in order to encode an entire dataset into the weights of a SIREN model. The weights for the SIREN model are generated by this hyper network, which computes these weights based on an encoded representation.
Support for the Hyper Network is available by using NeuralProcessHyperNet, which uses the SetEncoder from the paper as the encoder.
Training on the CIFAR 10 dataset is available inside the scripts directory - train_cifar_inpainting_siren.py and eval_cifar_inpainting_siren.py.
Pre-trained weights are available in the Release tab under assets.
On evaluating on the test set with 1000 context pixels, this model gets an average MSE of 0.009. Using 100 context pixels, the MSE increases to 0.019.
The following image is using 1000 context pixels on the test set :
The kernel scaled variants of the model converge faster than the original SIREN under certain circumstances. All the models below are trained with Adam optimizer with constant learning rate of 5e-5 for 5000 epochs and batch size of 8192 on the same image pixels (10% of the celtic spiral image).
The tensorboard logs can be found here -
@inproceedings{sitzmann2019siren,
author = {Sitzmann, Vincent
and Martel, Julien N.P.
and Bergman, Alexander W.
and Lindell, David B.
and Wetzstein, Gordon},
title = {Implicit Neural Representations
with Periodic Activation Functions},
booktitle = {arXiv},
year={2020}
}
- Tensorflow 2.0+
- Matplotlib to visualize eval result
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent