Skip to content

Latest commit

 

History

History
105 lines (97 loc) · 4.02 KB

Documentation.md

File metadata and controls

105 lines (97 loc) · 4.02 KB
Raw

NeuralScaleID

Model class (NeuralNet.py)

Instantiate a model via:

import NeuralNet as NN
model = NN.Model(architecture=[10,64,64,64,2],loss='KL',softmax=True)

Methods

train

train(x,y,batch_size=32,epochs=1,learning_rate=0.001,lr_scheduler = None,verbose=2,asymp_steps=1000,skip=100)

trains the model.

Arguments

  • x: training input, numpy array of shape [batch size,input dimension]
  • y: training output, numpy array of shape [batch size,output dimension]
  • batch_size: Integer.
  • learning_rate: Adam learning rate.
  • lr_scheduler: Learning rate scheduler. This is a function that takes in epoch and returns a float value for learning rate. The learning rate is updated at each training step and not just at the end of each epoch. If learning_rate and lr_scheduler are both specified, lr_scheduler overrides learning_rate.
  • verbose: Integer. 0,1 or 2. 0 = silent, 1 = epoch number, 2 = progress % within an epoch
  • asymp_steps: The method returns the training loss for the last asymp_steps steps of the training session (the last asymp_steps of the last epoch of the training session)
  • skip: The method returns training loss at every skip steps. set skip=1 for training loss for all training steps

predict

predict(x)

generates output predictions for input x

Arguments

  • x: input, numpy array of shape [batch size,input dimension]. The entire input array is evaluated as one batch

predict_layer

predict_layer(x,layer=None)

returns the predicted output of a specified layer for input x

Arguments

  • x: input, numpy array of shape [batch size,input dimension]. The entire input array is evaluated as one batch
  • layer: Integer. 1, 2... for 1st, 2nd... hidden layer, or 0, -1,... for output layer, prefinal layer... etc. If None, returns a dictionary of all layer outputs.

evaluate

evaluate(x,y)

evaluates loss given input x and output y

Arguments

  • x: input, numpy array of shape [batch size,input dimension]. The entire input array is evaluated as one batch
  • y: output, numpy array of shape [batch size,output dimension]

get_weights

get_weights()

returns W and b (weights and biases) as lists of numpy arrays

save_model

save_model(name,location='.')

saves the model

Arguments

  • name: String. The name under which you want to save the model
  • location String. The directory in which you want to save your model.

Other methods in NeuralNet.py

load_model

load_model(name,folder='.')

loads the model saved via save_model in Model class

Arguments

  • name: String. The name of the saved model
  • folder: String. The directory containing the model

Teacher class (teacher.py)

Instantiate a teacher via:

import teacher
t = teacher.Teacher(architecture=[4,256,256,8],scale=[1,1,0,0])

Methods

generate_input

generate_input(batch_size=1)

returns a randomly generated input

Arguments

  • batch_size: integer, returns numpy array of shape [batch size,input dimension] . Here input dimension is the first element of the global variable architecture

predict

predict(x_input)

returns the predicted value for input x_input

Arguments

  • x_input: input, numpy array of shape [batch size,input dimension]. The entire input array is evaluated as one batch. Here input dimension is the first element of the global variable architecture

save_weights

save_weights(folder,W=None,b=None)

saves W and b (weights and biases) as csv files.

Arguments

  • folder: directory to save weights and biases
  • W: Weights to be saved (list of numpy arrays). If None, defaults to global variable W
  • b: Biases to be saved (list of numpy arrays). If None, defaults to global variable b