Prometheus : A Neural Network Kernel in C
The neural network kernel is written in C and iterfaced with python. The python module also takes care input-output and train-test data parsing and analysis
BUILD
To build the neural network C library that will be interface with python
make clean; make
RUN
The python driver app is required to pass a json file containg the configurations to prometheus module
./prometheus/test.py -i examples/iris/iris.json
Cfg file
The input config file consists of two major configurations
"neural_nw_lib_path": Path to the C shared library containing the neural network kernel-
"neural_nw_config": Contains neural network configuration infoConfig Description "neural_nw_type"can either be "REGRESS"or"CLASSIFY""num_input"number of input "num_output"number of output "learning_rate"learning rate of the neural newtwork "oactv"output activation function required only if "neural_nw_type"="CLASSIFY""normalization""L1"or"L2"(Note curently L1 is supported L2 needs to be implemeneted)"num_hidden_layers"number of hidden layers and for each hidden layer define proporties "hl_prop"list of all hidden layers properties “hl_prop” list of all hidden layers properties to be define for each hiden layer "actv_fn""LINEAR","RELU","SIGMOID""size"size of the hidden layer ,i.e, No of neurons "neural_network_test": Contains neural network input-output and train-test info
| Config | Description |
|---|---|
"test_file" |
Path to the file containg input-outputs used for training and testing usually a .csv or .xlsx (Note: current support is for .csv) |
"delim" |
Delimiter in case of csv file |
"inputs" |
List of label names that has to be considered as inputs (Note: size of lablel list should be same as “num_input”) |
"outputs" |
List of label names that has to be considered as outputs (Note: size of lablel list should be same as “num_output”) |
"ntrain" |
Percentage of the data to be used for training and the remaining is used for testing |
"shuffle" |
Shuffle the data set if 1 do not shuffle if 0 |
"epochs" |
no of training epochs (Note: Currently only 1 is supported) |
"training_method" |
Training method "Batch" is only supported currently |
"mean_std_scale" |
optional field mean scale the given labels |
"min_max_scale" |
optional field min max scale the given labels |
"stdv_scale" |
optional field standard deviation based scale the given labels |
"hot_encode" |
optional field if Categorical variable present hot encode the outputs. more info refer examples/iris/iris.json |
Folders
| Folder | Description |
|---|---|
| common | os includes |
| neural_network | Contains sources that define the neural network in C |
| prometheus | contains the python wrapper module |
| docs | contains Code documentation |
| example | example configs tested bc_diag, tree_wilt and iris |
Using prometheus module in your python code
Below is a code snippet that shows how to use and call the prometheus module
import sys, getopt
from prometheus import prometheus
p1 = prometheus()
p1.fetch_configs(sys.argv[1:])
p1.create_brain()
p1.teach_brain()
p1.sentient_brain()
p1.destroy_brain()
Future Scope (Highly optimistic)
- Graphs and plots to analyse input/output
- GUI based cfg selection
- Adding Convolutional layer
- Neural network(Scaling,bounding/probalistic,)
- Training-strategy(loss Index and Optimization)
- Data-analysis(data validity and analysis)
- Model-selection(regress to get optimized model)
- Testing-analysis(How good did we do?)
Current Scope
We have classification working but regression has some problems and seem to be beyond my scope of understanding any help will be appreciated
Contributors
If interested in developing further join our discord