https://github.com/vigneswaran-chandrasekaran/flipflop_experiments
Experiments with Flip flop networks on a variety of tasks to check and showcase its generalization abilities
https://github.com/vigneswaran-chandrasekaran/flipflop_experiments
rnn-tensorflow sequential-models
Last synced: 4 months ago
JSON representation
Experiments with Flip flop networks on a variety of tasks to check and showcase its generalization abilities
- Host: GitHub
- URL: https://github.com/vigneswaran-chandrasekaran/flipflop_experiments
- Owner: vigneswaran-chandrasekaran
- Created: 2020-12-18T12:01:45.000Z (over 4 years ago)
- Default Branch: main
- Last Pushed: 2021-12-21T15:41:51.000Z (over 3 years ago)
- Last Synced: 2025-01-12T16:48:56.391Z (5 months ago)
- Topics: rnn-tensorflow, sequential-models
- Language: Jupyter Notebook
- Homepage:
- Size: 13.2 MB
- Stars: 1
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# Flipflop Experiments
Repository contains list of experiments in various applications to showcase
the generalization capacity of flipflop network
## List of experiments
- [Shakespeare text generation](#shakespeare-text-generation)
- [Handwriting generation](#handwriting-generation)
- [UCF 11/101 action recognition](#ucf11-action-recognition)
- [Moving squares next frame prediction](#moving-squares)
- Signal reconstruction
- Predict lung volume from CT scans
- Predict next frame in moving MNIST
- Phoneme recognition
- Name origin classification
- Magenta sketch
---
### Shakespeare text generation
Text generation model trained at character-level.
Code directory: `shakespeare/`
Sample output of flipflop network,
```
ROMEO:
Is not it! Voluble; give your general saint;
She were all: this is come brush'd alove,
And traitor him for rude, that back again; yet sure,
Yet a mind bid me with repity here.
RICHOOS:
I have becamed done, as you perchance and tell thereof,
Both in the beggar hate, conlein on thy brother:
Yet beggar to the faith, yet he, as he' whom was
That I, unreasons, with a leave, and blunt goodly.
KING EDWARD IV:
Now it that violent cuttain
Presence; and weigh affection,
His troth?--'tod God, I!
Shall I am never speak, that I have gone.
Ventious,
She have is lates of that say the cincer:
Under the sortuness joy is fooler;
And rather mother from the banish.
Very go twal liberty, who comest thou
deny to and time, and soil-whum doublied mine.
Because hath he shall have entertatching dam
Our pace of darmward in govern wing
Of fly.
First Keeper: I willow'd with chamber, to been the people.
Call, not breathed more authority to with some
jefty malice; thou art dangerous are not news
Is news Petruch
```
**Performance comparison with LSTM**
### Handwriting generation
Handwriting generation model trained on IAM online handwriting data. The
architecture used is similar to
[Generating Sequences With Recurrent Neural Networks](https://arxiv.org/abs/1308.0850)
The data preprocessing and config files are taken from
[github.com/snowkylin/rnn-handwriting-generation](https://github.com/snowkylin/rnn-handwriting-generation)
Code directory: `handwriting-gen/ff` contains flipflop's and `handwriting-gen/lstm`
contains LSTM's. Saved models for both are available
**Performance comparison with LSTM**
Flipflop output - 'how are you?'
LSTM output - 'how are you?'
Flipflop output - 'sometime'
LSTM output - 'sometime'
Flipflop output - 'Bonjour'
LSTM output - 'Bonjour'
### UCF11 Action Recognition
Action recognition model trained on UCF 11 dataset. Convolutional Flipflop
layer is used in flipflop model. Convolutional flipflop is similar to `tensorflow.keras.layers.ConvLSTM2D`.
The classification accuracy of flipflop was 99.3% whereas 98.5% for LSTM
Code dir: `ucf11/`
### Moving squares
Example taken from [Keras docs example for ConvLSTM2D]. Model should predict the
next frame in a movie of moving squares
Sample output of ConvFF and ConvLSTM:
### Signal reconstruction
Check sequence generation capacity of the model. Two types of signals (of labels `001` and `011`)
should be generated with MSE as the loss
Sample signal images,
