https://github.com/vidursatija/fasterconv3d
3D Convolution to 2D Convolution
https://github.com/vidursatija/fasterconv3d
cnn convolutional-neural-networks machine-learning mobile python tensorflow
Last synced: 2 months ago
JSON representation
3D Convolution to 2D Convolution
- Host: GitHub
- URL: https://github.com/vidursatija/fasterconv3d
- Owner: vidursatija
- Created: 2020-09-13T13:03:48.000Z (almost 6 years ago)
- Default Branch: master
- Last Pushed: 2020-09-13T13:06:02.000Z (almost 6 years ago)
- Last Synced: 2025-10-06T22:34:34.799Z (9 months ago)
- Topics: cnn, convolutional-neural-networks, machine-learning, mobile, python, tensorflow
- Language: Jupyter Notebook
- Homepage:
- Size: 3.91 KB
- Stars: 0
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# Comparison between 1 Conv3D op vs 3 Conv2D ops
This repo serves as dummy code which shows that 3 Conv2Ds are actually faster than 1 Conv3D operation in tensorflow. Obviously this result will vary with the image size, kernel size and hardware stuff.
| X | Conv3D | Conv2D |
|------------|--------|--------|
| Parameters | `k*k*k*i*o` | `3*k*1*i*o` |
| Speed | `h*w*d*k*k*k*i*o` | `h*w*d*3*k*1*i*o` |
```
i = input channels
o = output channels
h, w, d = input tensor dimensions
k = kernel size
```
## Use cases
1. Mobiles which don't support 3D convolutions and 3D tensors but need such operations
2. To save space and improve performance of the model
3. ?