def validate(self, validX, validY): # Input # validX : Validation Input Data # validY : Validation Labels # Returns the validation accuracy evaluated over the current neural network model valActivations = self.feedforward(validX) pred = np.argmax(valActivations[-1], axis=1) validPred = oneHotEncodeY(pred, self.outputSize) validAcc = self.computeAccuracy(validY, validPred) return pred, validAcc
def train(self, trainX, trainY, printTrainStats=True, saveModel=False, modelName=None):
for epoch in range(self.epochs):
if printTrainStats:
print("Epoch: ", epoch)
X = np.asarray(trainX)
Y = np.asarray(trainY)
perm = np.arange(X.shape[0])
np.random.shuffle(perm)
X = X[perm]
Y = Y[perm]
trainLoss = 0
numBatches = int(np.ceil(float(X.shape[0]) / self.batchSize))
for batchNum in range(numBatches):
XBatch = np.asarray(X[batchNum*self.batchSize: (batchNum+1)*self.batchSize])
YBatch = np.asarray(Y[batchNum*self.batchSize: (batchNum+1)*self.batchSize])
activations = self.feedforward(XBatch)
loss = self.computeLoss(YBatch, activations)
trainLoss += loss
predLabels = oneHotEncodeY(np.argmax(activations[-1], axis=1), self.out_nodes)
acc = self.computeAccuracy(YBatch, predLabels)
trainAcc += acc
self.backpropagate(activations, YBatch)
trainAcc /= numBatches
if printTrainStats:
print("Epoch ", epoch, " Training Loss=", loss, " Training Accuracy=", trainAcc)
if saveModel:
model = []
for l in self.layers:
if type(l).__name__ != "FlattenLayer":
model.append(l.weights)
model.append(l.biases)
np.save(modelName, model)
print("Model Saved... ")
def train(self, trainX, trainY, validX=None, validY=None, printTrainStats=True, printValStats=True):
# Method for training the Neural Network
# Input
# trainX - A list of training input data to the neural network
# trainY - Corresponding list of training data labels
# validX - A list of validation input data to the neural network
# validY - Corresponding list of validation data labels
# printTrainStats - Print training loss and accuracy for each epoch
# printValStats - Prints validation set accuracy after each epoch of training
# The methods trains the weights and baises using the training data(trainX, trainY)
# and evaluates the validation set accuracy after each epoch of training
for epoch in range(self.epochs):
# A Training Epoch
if printTrainStats or printValStats:
print("Epoch: ", epoch)
# Shuffle the training data for the current epoch
X = np.asarray(trainX)
Y = np.asarray(trainY)
perm = np.arange(X.shape[0])
np.random.shuffle(perm)
X = X[perm]
Y = Y[perm]
# Initializing training loss and accuracy
trainLoss = 0
trainAcc = 0
# Divide the training data into mini-batches
numBatches = int(np.ceil(float(X.shape[0]) / self.batchSize))
for batchNum in range(numBatches):
XBatch = np.asarray(X[batchNum*self.batchSize: (batchNum+1)*self.batchSize])
YBatch = np.asarray(Y[batchNum*self.batchSize: (batchNum+1)*self.batchSize])
# Calculate the activations after the feedforward pass
activations = self.feedforward(XBatch)
# Compute the loss
loss = self.computeLoss(YBatch, activations)
trainLoss += loss
# Estimate the one-hot encoded predicted labels after the feedword pass
predLabels = oneHotEncodeY(np.argmax(activations[-1], axis=1), self.outputSize)
# Calculate the training accuracy for the current batch
acc = self.computeAccuracy(YBatch, predLabels)
trainAcc += acc
# Backpropagation Pass to adjust weights and biases of the neural network
self.backpropagate(activations, YBatch)
# Print Training loss and accuracy statistics
trainAcc /= numBatches
if printTrainStats:
print("Epoch ", epoch, " Training Loss=", loss, " Training Accuracy=", trainAcc)
# Estimate the prediction accuracy over validation data set
if validX is not None and validY is not None and printValStats:
_, validAcc = self.validate(validX, validY)
print("Validation Set Accuracy: ", validAcc, "%")
def train(self, trainX, trainY, validX=None, validY=None, printTrainStats=True, printValStats=True, saveModel=False, loadModel=False, modelName=None):
# Method for training the Neural Network
# Input
# trainX - A list of training input data to the neural network
# trainY - Corresponding list of training data labels
# validX - A list of validation input data to the neural network
# validY - Corresponding list of validation data labels
# printTrainStats - Print training loss and accuracy for each epoch
# printValStats - Prints validation set accuracy after each epoch of training
# saveModel - True -> Saves model in "modelName" file after each epoch of training
# loadModel - True -> Loads model from "modelName" file before training
# modelName - Name of the model from which the funtion loads and/or saves the neural net
# The methods trains the weights and baises using the training data(trainX, trainY)
# and evaluates the validation set accuracy after each epoch of training
if loadModel:
model = np.load(modelName)
k,i = 0,0
for l in self.layers:
if type(l).__name__ != "AvgPoolingLayer" and type(l).__name__ != "FlattenLayer":
self.layers[i].weights = model[k]
self.layers[i].biases = model[k+1]
k+=2
i+=1
print("Model Loaded... ")
for epoch in range(self.epochs):
# A Training Epoch
if printTrainStats or printValStats:
print("Epoch: ", epoch)
# Shuffle the training data for the current epoch
X = np.asarray(trainX)
Y = np.asarray(trainY)
perm = np.arange(X.shape[0])
np.random.shuffle(perm)
X = X[perm]
Y = Y[perm]
# Initializing training loss and accuracy
trainLoss = 0
trainAcc = 0
# Divide the training data into mini-batches
numBatches = int(np.ceil(float(X.shape[0]) / self.batchSize))
for batchNum in range(numBatches):
# print(batchNum,"/",numBatches) # UNCOMMENT if model is taking too long to train
XBatch = np.asarray(X[batchNum*self.batchSize: (batchNum+1)*self.batchSize])
YBatch = np.asarray(Y[batchNum*self.batchSize: (batchNum+1)*self.batchSize])
# Calculate the activations after the feedforward pass
activations = self.feedforward(XBatch)
# Compute the loss
loss = self.computeLoss(YBatch, activations)
trainLoss += loss
# Estimate the one-hot encoded predicted labels after the feedword pass
predLabels = oneHotEncodeY(np.argmax(activations[-1], axis=1), self.out_nodes)
# Calculate the training accuracy for the current batch
acc = self.computeAccuracy(YBatch, predLabels)
trainAcc += acc
# Backpropagation Pass to adjust weights and biases of the neural network
self.backpropagate(activations, YBatch)
# Print Training loss and accuracy statistics
trainAcc /= numBatches
if printTrainStats:
print("Epoch ", epoch, " Training Loss=", loss, " Training Accuracy=", trainAcc)
if saveModel:
model = []
for l in self.layers:
print(type(l).__name__)
if type(l).__name__ != "AvgPoolingLayer" and type(l).__name__ != "FlattenLayer":
model.append(l.weights)
model.append(l.biases)
np.save(modelName, model)
print("Model Saved... ")
# Estimate the prediction accuracy over validation data set
if validX is not None and validY is not None and printValStats:
_, validAcc = self.validate(validX, validY)
print("Validation Set Accuracy: ", validAcc, "%")
def test(self, testX, testY): testActivations = self.feedforward(testX) pred = np.argmax(testActivations[-1], axis=1) testPred = oneHotEncodeY(pred, self.out_nodes) testAcc = self.computeAccuracy(testY, testPred) return testAcc
