neowise 0.1.0

A Deep Learning library built from scratch using Python and NumPy

neowise Documentation

Steps to train your own Neural Network using neowise

1. Install and import neowise as nw

2. Get your data and pre-process it. Your data should be in the dimensions as (number_of_examples, number_of_features) and labels should have (number_of_output_units, number_of examples) as its dimensions. This is a must, any changes here would raise errors!

3. Create a model by calling model = nw.Model(your_train_data, your_train_labels, your_test_data, your_test_labels, your_crossval_data, your_crossval_labels) If you do not have Cross Validation data, enter None for the last two arguments.

4. Add layers to your model by model.add(layer_name,num_inputs,num_outputs,activation_function,dropout), where you give a unique name to each of your layers so that you know what type of layer it is. Example for dense layer, if it is your first layer, name it dense1. Enter the number of inputs to that layer, in num_inputs and number of units for that layer in num_outputs. For activation_function use any of the following supported activation functions ["relu", "sigmoid", "tanh", "softmax", "sine"]. To prevent overflows and nans, we suggest that if you use a softmax classifier, to set the activation of the previous layer of the output layer as "tanh" as it squishes values between -1 and 1, thus preventing catastrophe. If you want to use Dropout, set the dropout anywhere between 0 and 1. Else, the default value is taken as 1, i.e no Dropout.

5. Just to be sure of your architecture and to know the amount of parameters that'll be trained call model.summary() which uses prettytable to print out a summary of your architecture.

6. Train your model using model.fit(your_train_data, your_train_labels, learning_rate, number_of_iterations, optimizer, problem_type, mini_batch_size, regularization_type, lambda, learning_rate_decay), where you enter your training data, choose the learning rate, set the number of iterations to train for, choose which type of optimizer you want from ["GD" for Gradient Descent, "Momentum" for Gradient Descent using Momentum, "RMSprop" for Root Mean Square Propagation, "Adam" for Adaptive Moment Estimation] If you want to train using Batch Gradient Descent, choose "GD" as optimizer and set the mini_batch_size to the total number of examples in your training data and if you want to train using Stochastic Gradient Descent, choose "GD" as optimizer and set mini_batch_size to 1. For problem_type choose any of the currently supported tasks ["Binary" for Binary Classification Tasks, "Multi" for Multi Class Classification Tasks] Set mini_batch_size to the size you want each of your mini batches to be and be sure that this value is less than the total number of examples you have. If you want to use L1 or L2 regularization choose from ["L1" for L1 Regularization and "L2" for L2 Regularization] and set the regularization parameter lambda. If you want your learning_rate to not be constant and be decreasing as the training progresses, set alpha_decay to True

7. Test the model on your test data by calling model.test(your_test_data, your_test_labels, problem_type) on your training data and set problem_type as you did for model.fit. This displays a prettytable with precision, recall and f1 scores and its accuracy on the test data.

8. For plotting the costs and accuracy vs number of iterations, call model.plot(type_function, animate, directory, frequency) and set type_function to "Cost" if you want to plot costs vs number of iterations and "Accuracy" for accuracy vs number of iterations. If you want to create animated graphs, set animate to True and specify the directory in which you want to save the plots in directory and set the frequency with which the graphs should update in frequency, then feed those images to a GIF creator to create animated plots.

9. To save the model, call model.save_model(file_name) and specify the directory in which you want to save the model with the name of the model with the extension .h5 in filename

10. To load a previously saved model, create a new model be calling saved_model = nw.Model(your_train_data, your_train_labels, your_test_data, your_test_labels, your_crossval_data, your_crossval_labels), where these are the same data on which model was trained on. Call saved_model.load_model(file_name) to load the model from the directory specified in file_name

11. These are the functionalities that neowise offers. For detailed doc_strings visit Source Code to find more about the project!