https://github.com/voldien/ritsu
https://github.com/voldien/ritsu
Last synced: 7 months ago
JSON representation
- Host: GitHub
- URL: https://github.com/voldien/ritsu
- Owner: voldien
- License: mit
- Created: 2022-08-25T12:35:11.000Z (about 3 years ago)
- Default Branch: master
- Last Pushed: 2025-02-19T08:06:17.000Z (8 months ago)
- Last Synced: 2025-02-19T09:22:37.211Z (8 months ago)
- Language: C++
- Size: 577 KB
- Stars: 0
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# Ritsu - ăȘă
[](https://github.com/voldien/ritsu/actions/workflows/ci.yml)
[](https://opensource.org/licenses/MIT)
[](https://github.com/voldien/ritsu/releases)
A machine learning library, created for personal educational purposes. Is not recommended for product development. But only as educational material.
```cpp
Input input({2}, "input");
Dense dense0(2, false);
Dense outputDense(1, false);
RandomUniformInitializer random(0, 2, 10052);
Tensor dataX = random(Shape({128, 2}));
/* Sum. */
Tensor dataY({128, 1});
for (unsigned int i = 0; i < dataY.getNrElements(); i++) {
const float value = dataX.getValue({i, 0}) + dataX.getValue({i, 1});
dataY.getValue(i) = value;
}
Layer &output = outputDense(dense0(input));
SGD optimizer(0.0001, 0.0);
MetricAccuracy accuracy;
Model forwardModel = Model({&input}, {&output});
MeanSquareError mse_loss = MeanSquareError();
forwardModel.compile(&optimizer, mse_loss, {dynamic_cast(&accuracy)});
Model::History *result = &forwardModel.fit(8, dataX, dataY, 1, 0, false, false);
```
## Example Run Image Recognition MNIST
```bash
bin/autoencoder -E 8 -B 1 -l 0.00001
Loaded MNIST Data Set: [60000,28,28,1] Labels: [60000,1]
Train Object Size: [28,28,1] Expected result Size: [28,28,1]
input [28,28,1]
flatten0 [784] <-- [ input ] f
layer0 [128] <-- [ flatten0 ] f
relu0 [128] <-- [ layer0 ] f
layer1 [64] <-- [ relu0 ] f
relu1 [64] <-- [ layer1 ] f
layer2 [32] <-- [ relu1 ] f
relu2 [32] <-- [ layer2 ] f
latent [8] <-- [ relu2 ] f
layer3 [32] <-- [ latent ] f
relu3 [32] <-- [ layer3 ] f
layer4 [64] <-- [ relu3 ] f
relu4 [64] <-- [ layer4 ] f
layer5 [128] <-- [ relu4 ] f
relu5 [128] <-- [ layer5 ] f
layer6 [784] <-- [ relu5 ] f
sigmoid [784] <-- [ layer6 ] f
reshape [28,28,1] <-- [ sigmoid ] f
number of weights: 222936
Trainable in Bytes: 870 KB
None-Trainable in Bytes: 0 KB
Loss Function: MSE
Optimizer: adam
Epoch: 0 / 8
Batch: 54000/54000 ETA: 0 - lr: 1e-05 loss: 0.29203 - accuracy: 0.207908
Batch: 6000/6000 ETA: loss-val: 0.297233 - accuracy: 0.190051
Epoch: 1 / 8
Batch: 54000/54000 ETA: 0 - lr: 1e-05 loss: 0.264627 - accuracy: 0.357143
Batch: 6000/6000 ETA: loss-val: 0.269659 - accuracy: 0.331633
Epoch: 2 / 8
Batch: 54000/54000 ETA: 0 - lr: 1e-05 loss: 0.240133 - accuracy: 0.459184
Batch: 6000/6000 ETA: loss-val: 0.244998 - accuracy: 0.438776
Epoch: 3 / 8
Batch: 54000/54000 ETA: 0 - lr: 1e-05 loss: 0.218462 - accuracy: 0.571429
Batch: 6000/6000 ETA: loss-val: 0.223162 - accuracy: 0.561224
Epoch: 4 / 8
Batch: 54000/54000 ETA: 0 - lr: 1e-05 loss: 0.199439 - accuracy: 0.665816
Batch: 6000/6000 ETA: loss-val: 0.203973 - accuracy: 0.655612
Epoch: 5 / 8
Batch: 54000/54000 ETA: 0 - lr: 1e-05 loss: 0.182833 - accuracy: 0.778061
Batch: 6000/6000 ETA: loss-val: 0.187198 - accuracy: 0.765306
Epoch: 6 / 8
Batch: 54000/54000 ETA: 0 - lr: 1e-05 loss: 0.168387 - accuracy: 0.817602
Batch: 6000/6000 ETA: loss-val: 0.172584 - accuracy: 0.797194
Epoch: 7 / 8
Batch: 54000/54000 ETA: 0 - lr: 1e-05 loss: 0.155844 - accuracy: 0.821429
Batch: 6000/6000 ETA: loss-val: 0.159869 - accuracy: 0.808673
Average Test Loss: 0.16282
Accuracy Test: 0.828147
```
## Example Run AutoEncoder MNIST
```bash
bin/mnist -B 16 -E 8 -l 0.00000001
Loaded MNIST Data Set: [60000,28,28,1] Labels: [60000,1]
Train Object Size: [28,28,1] Expected result Size: [10]
input [28,28,1]
flatten0 [784] <-- [ input ] f
layer0 [32] <-- [ flatten0 ] f
relu0 [32] <-- [ layer0 ] f
layer1 [16] <-- [ relu0 ] f
relu1 [16] <-- [ layer1 ] f
layer2 [10] <-- [ relu1 ] f
number of weights: 25818
Trainable in Bytes: 100 KB
None-Trainable in Bytes: 0 KB
Loss Function: Categorical Crossentropy
Optimizer: sgd
Epoch: 0 / 8
Batch: 3375/3375 ETA: 0 - lr: 1e-08 loss: 0.00194646 - accuracy: 0.3
Batch: 375/375 ETA: loss-val: 0.00136129 - accuracy: 0.25625
Epoch: 1 / 8
Batch: 3375/3375 ETA: 0 - lr: 1e-08 loss: 0.00134664 - accuracy: 0.475
Batch: 375/375 ETA: loss-val: 0.00102787 - accuracy: 0.38125
Epoch: 2 / 8
Batch: 3375/3375 ETA: 0 - lr: 1e-08 loss: 0.00113783 - accuracy: 0.5625
Batch: 375/375 ETA: loss-val: 0.000954269 - accuracy: 0.4625
Epoch: 3 / 8
Batch: 3375/3375 ETA: 0 - lr: 1e-08 loss: 0.0010338 - accuracy: 0.58125
Batch: 375/375 ETA: loss-val: 0.000924363 - accuracy: 0.525
Epoch: 4 / 8
Batch: 3375/3375 ETA: 0 - lr: 1e-08 loss: 0.000975035 - accuracy: 0.60625
Batch: 375/375 ETA: loss-val: 0.00090925 - accuracy: 0.54375
Epoch: 5 / 8
Batch: 3375/3375 ETA: 0 - lr: 1e-08 loss: 0.00093868 - accuracy: 0.61875
Batch: 375/375 ETA: loss-val: 0.000900569 - accuracy: 0.5625
Epoch: 6 / 8
Batch: 3375/3375 ETA: 0 - lr: 1e-08 loss: 0.000916001 - accuracy: 0.64375
Batch: 375/375 ETA: loss-val: 0.000894858 - accuracy: 0.5875
Epoch: 7 / 8
Batch: 3375/3375 ETA: 0 - lr: 1e-08 loss: 0.000901652 - accuracy: 0.6625
Batch: 375/375 ETA: loss-val: 0.000890853 - accuracy: 0.59375
Average Test Loss: 0.00201005
Accuracy Test: 0.67083,
```
## Supported Precisions
- **Float**
- **Double**
- **Integar**
## Layer Support
### Activation
- **Relu**
- **Leaky Relu**
- **Sigmoid**
- **Tanh**
- **SoftMax**
- **Linear**
- **Swish**
- **ExpLinear**
- **Linear**
### Layer
- **Dense**
- **Batch Normalization** - WIP
- **Dropout** - WIP
- **GuassianNoise** - WIP
- **Flatten**
- **Reshape**
## Loss Function
- **MSE** - Mean Square Error.
- **MSA** - Mean Square Absolute.
- **CrossEntropy** -
- **CatagorialCrossEntropy** -
## Optimizer
- **SGD** - Stochastic Gradient Descent + Momentum
- **Adam** - Adaptive Moment Estimation
## Dependencies
In order to compile the program on Linux based machine, the following Debian packages are required.
```bash
sudo apt-get install cmake g++ libgtest-dev googletest libomp-dev libjemalloc-dev
```
## Building
The OpenMP can be built with other graphic framework, to be executed on the graphic devices. However, no supported yet added.
### AMD - ROCM
### NVIDIA - CUDA
### Intel - OneAPI
## Installation
## Testing
```bash
ctest -vv
```
```bash
ctest -O failure.txt --output-on-failure
```
## License
This project is licensed under the MIT License - see the [LICENSE](LICENSE) file for details