https://github.com/sidray-infinity/dumbtf
A deep learning library for dumb people.
https://github.com/sidray-infinity/dumbtf
deep-learning flask machine-learning material-ui reactjs
Last synced: 12 months ago
JSON representation
A deep learning library for dumb people.
- Host: GitHub
- URL: https://github.com/sidray-infinity/dumbtf
- Owner: Sidray-Infinity
- License: gpl-3.0
- Created: 2020-04-23T14:43:09.000Z (about 6 years ago)
- Default Branch: master
- Last Pushed: 2024-05-03T20:54:44.000Z (about 2 years ago)
- Last Synced: 2025-04-10T23:51:44.568Z (about 1 year ago)
- Topics: deep-learning, flask, machine-learning, material-ui, reactjs
- Language: Python
- Homepage:
- Size: 653 KB
- Stars: 2
- Watchers: 1
- Forks: 1
- Open Issues: 3
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
DumbTF
---
A deep learning library for dumb people.
Focus
---
- Implement additional layers, to be used along with convolutions
- Explore LAMB : https://towardsdatascience.com/an-intuitive-understanding-of-the-lamb-optimizer-46f8c0ae4866
- Convolutions are still slow, despite using the numpy stride trick and einsum.
## Installation
`pip install Dumb-TF`
## Status
---
- ### Regression (Boston Housing Dataset)
Network Architecture
```python
model = Model()
model.add(Dense(128, input_shape=13, activation='relu'))
model.add(Dense(64, input_shape=128, activation='relu'))
model.add(Dense(1, input_shape=64, activation='linear'))
model.compile(loss="mse", optimizer="mini_batch_gd", lr=0.01)
loss, loss_pe = model.fit(x_train, y_train, epochs=20, batch_size=64)
```
- ### Binary Classification (Breast Cancer Dataset)
Network Architecture
```python
model = Model()
model.add(Dense(64, input_shape=30, activation='relu'))
model.add(Dense(32, input_shape=64, activation='relu'))
model.add(Dense(1, input_shape=32, activation='sigmoid'))
model.compile(loss="bce", optimizer="mini_batch_gd", lr=0.01)
loss, loss_pe = model.fit(x, y, epochs=150, batch_size=64)
```
- ### Multiclass Classification (MNIST)
Network Architecture
```python
model = Model()
model.add(Dense(32, input_shape=784, activation='relu'))
model.add(Dense(10, input_shape=32, activation='softmax'))
model.compile(loss="cce", optimizer="mini_batch_gd", lr=0.01)
loss, loss_pe = model.fit(x_train, y_train, epochs=10, batch_size=64)
```
- ### Convolution Network (MNIST)
Network Architecture
```python
model = Model()
model.add(Conv2D(32, 3, 'relu', (28, 28, 1)))
model.add(Conv2D(32, 3, 'relu', (26, 26, 32)))
model.add(Flatten())
model.add(Dense(10, input_shape=26*26*32, activation='softmax'))
model.compile(loss='cce', optimizer='mini_batch_gd')
loss, loss_pe = model.fit(x_train, y_train, epochs=10, batch_size=64)
```
## References
- http://neuralnetworksanddeeplearning.com/chap2.html
- https://brilliant.org/wiki/backpropagation/#$
- https://towardsdatascience.com/batch-mini-batch-stochastic-gradient-descent-
- https://www.youtube.com/watch?v=i94OvYb6noo
- https://aimatters.wordpress.com/2020/06/14/derivative-of-softmax-layer/
- https://towardsdatascience.com/deriving-the-backpropagation-equations-from-scratch-part-1-343b300c585a
### For Convolutions
* https://www.sicara.ai/blog/2019-10-31-convolutional-layer-convolution-kernel
* https://towardsdatascience.com/beginners-guide-to-understanding-convolutional-neural-networks-ae9ed58bb17d
* https://www.youtube.com/watch?v=8rrHTtUzyZA
(FAST CONVOLUTIONS)
* https://laurentperrinet.github.io/sciblog/posts/2017-09-20-the-fastest-2d-convolution-in-the-world.html
* https://towardsdatascience.com/how-are-convolutions-actually-performed-under-the-hood-226523ce7fbf
* https://jessicastringham.net/2017/12/31/stride-tricks/
(Backpropagation in convolutions)
* https://medium.com/@pavisj/convolutions-and-backpropagations-46026a8f5d2c
* https://www.jefkine.com/general/2016/09/05/backpropagation-in-convolutional-neural-networks/
(MaxPool2D)
* https://wiseodd.github.io/techblog/2016/07/18/convnet-maxpool-layer/#:~:text=Maxpool%20backward&text=We%20let%20the%20gradient%20pass,max%20operation%20do%20in%20backpropagation
.