The layer can be useful, when the input data are XYZ coordinates, as in Molecular Dynamics, Robotics, LIDAR, autopilot driving.
The AngleComparison layer is an analogy for angle calculation using cosine formula using inner product. It is used in the case when we use X, Y, Z coordinates and want to include information between the features, but we don't want to do it explicitly.
Sure, here's an updated Github description for the AngleComparison layer that includes the additional information you requested:
AngleComparison Layer
The AngleComparison layer is a custom layer for comparing two outputs of a 1D CNN layer in TensorFlow and PyTorch. The layer works by passing an input tensor through two different 1D CNN layers with the same filter size and kernel size. The output of each 1D CNN layer is then compared using a dot product computed with respect to the channel dimension. The dot product is normalized by the product of the L2 norms of the two 1D CNN layers. The final output is a tensor that concatenates the angle tensor with the input tensor.
Figure 1. Analogy to the angle calculation using dot products.
The AngleComparison layer is an analogy for angle calculation using cosine formula using inner product. It is used in the case when we use X, Y, Z coordinates and want to include information between the features, but we don't want to do it explicitly. So we calculate them for the higher level of abstraction.
The AngleComparison layer is useful for applications where we want to compare two outputs of a 1D CNN layer and extract a measure of similarity between them. This layer can be used in many different applications, such as natural language processing, speech processing, and time series analysis.
Some people can find it similar to self attention. The are several major differences:
- The result of the comparison is added to the features, not used for weighting some sort of input
- The dot product is normalized with respect to the L2 norm, so that the size of the vectors doesn't affect its results
- The output is not weighted with SoftMax function to obtain a probability distribution
Here's an example of how to use the AngleComparison layer in TensorFlow:
import tensorflow as tf
from angle_comparison_tensorflow import AngleComparison
inputs = tf.keras.Input(shape=(100, 10))
x = tf.keras.layers.Conv1D(filters=32, kernel_size=3, padding='same')(inputs)
x = AngleComparison(filters=32, kernel_size=3)(x)
model = tf.keras.Model(inputs=inputs, outputs=x)
And here's an example of how to use the AngleComparison layer in PyTorch:
import torch
import torch.nn as nn
from angle_comparison_pytorch import AngleComparison
inputs = torch.randn(32, 100)
x = nn.Conv1d(in_channels=1, out_channels=32, kernel_size=3, padding=1)(inputs.unsqueeze(1))
x = AngleComparison(filters=32, kernel_size=3)(x)
model = nn.Sequential(nn.Linear(101, 64), nn.ReLU(), nn.Linear(64, 10))
output = model(x)
You can install the AngleComparison layer by coping code or cloning this repository and importing respective module.
This project is licensed under the MIT License - see the LICENSE file for details.