We Keep Coding

Gradient and Loss function

I could not understand well especially how gradients were computed with regards to matrix transposes. My question is for DW2 but if you want also to discuss about the computation of the other gradients and extend my question I am open to discussion. Mathematically things seem a little bit different but this code is reliable and on github so I trust this code.
from __future__ import print_function
from builtins import range
from builtins import object
import numpy as np
import matplotlib.pyplot as plt
from past.builtins import xrange
class TwoLayerNet(object):
"""
A two-layer fully-connected neural network. The net has an input dimension of
D* (correction), a hidden layer dimension of H, and performs classification over C classes.
We train the network with a softmax loss function and L2 regularization on the
weight matrices. The network uses a ReLU nonlinearity after the first fully
connected layer.
In other words, the network has the following architecture:
input - fully connected layer - ReLU - fully connected layer - softmax
The outputs of the second fully-connected layer are the scores for each class.
"""
def __init__(self, input_size, hidden_size, output_size, std=1e-4):
"""
Initialize the model. Weights are initialized to small random values and
biases are initialized to zero. Weights and biases are stored in the
variable self.params, which is a dictionary with the following keys:
W1: First layer weights; has shape (D, H)
b1: First layer biases; has shape (H,)
W2: Second layer weights; has shape (H, C)
b2: Second layer biases; has shape (C,)
Inputs:
- input_size: The dimension D of the input data.
- hidden_size: The number of neurons H in the hidden layer.
- output_size: The number of classes C.
"""
self.params = {}
self.params['W1'] = std * np.random.randn(input_size, hidden_size)
self.params['b1'] = np.zeros(hidden_size)
self.params['W2'] = std * np.random.randn(hidden_size, output_size)
self.params['b2'] = np.zeros(output_size)
def loss(self, X, y=None, reg=0.0):
"""
Compute the loss and gradients for a two layer fully connected neural
network.
Inputs:
- X: Input data of shape (N, D). Each X[i] is a training sample.
- y: Vector of training labels. y[i] is the label for X[i], and each y[i] is
an integer in the range 0 <= y[i] < C. This parameter is optional; if it
is not passed then we only return scores, and if it is passed then we
instead return the loss and gradients.
- reg: Regularization strength.
Returns:
If y is None, return a matrix scores of shape (N, C) where scores[i, c] is
the score for class c on input X[i].
If y is not None, instead return a tuple of:
- loss: Loss (data loss and regularization loss) for this batch of training
samples.
- grads: Dictionary mapping parameter names to gradients of those parameters
with respect to the loss function; has the same keys as self.params.
"""
# Unpack variables from the params dictionary
W1, b1 = self.params['W1'], self.params['b1']
W2, b2 = self.params['W2'], self.params['b2']
N, D = X.shape
# Compute the forward pass
scores = None
#############################################################################
# TODO: Perform the forward pass, computing the class scores for the input. #
# Store the result in the scores variable, which should be an array of #
# shape (N, C). #
#############################################################################
# *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)*****
# perform the forward pass and compute the class scores for the input
# input - fully connected layer - ReLU - fully connected layer - softmax
# define lamba function for relu
relu = lambda x: np.maximum(0, x)
# a1 = X x W1 = (N x D) x (D x H) = N x H
a1 = relu(X.dot(W1) + b1) # activations of fully connected layer #1
# store the result in the scores variable, which should be an array of
# shape (N, C).
# scores = a1 x W2 = (N x H) x (H x C) = N x C
scores = a1.dot(W2) + b2 # output of softmax
# *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)*****
# If the targets are not given then jump out, we're done
if y is None:
return scores
# Compute the loss
loss = None
#############################################################################
# TODO: Finish the forward pass, and compute the loss. This should include #
# both the data loss and L2 regularization for W1 and W2. Store the result #
# in the variable loss, which should be a scalar. Use the Softmax #
# classifier loss. #
#############################################################################
# *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)*****
# shift values for 'scores' for numeric reasons (over-flow cautious)
# figure out the max score across all classes
# scores.shape is N x C
scores -= scores.max(axis = 1, keepdims = True)
# probs.shape is N x C
probs = np.exp(scores)/np.sum(np.exp(scores), axis = 1, keepdims = True)
loss = -np.log(probs[np.arange(N), y])
# loss is a single number
loss = np.sum(loss)
# Right now the loss is a sum over all training examples, but we want it
# to be an average instead so we divide by N.
loss /= N
# Add regularization to the loss.
loss += reg * (np.sum(W1 * W1) + np.sum(W2 * W2))
# *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)*****
# Backward pass: compute gradients
grads = {}
#############################################################################
# TODO: Compute the backward pass, computing the derivatives of the weights #
# and biases. Store the results in the grads dictionary. For example, #
# grads['W1'] should store the gradient on W1, and be a matrix of same size #
#############################################################################
# *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)*****
# since dL(i)/df(k) = p(k) - 1 (if k = y[i]), where f is a vector of scores for the given example
# i is the training sample and k is the class
dscores = probs.reshape(N, -1) # dscores is (N x C)
dscores[np.arange(N), y] -= 1
# since scores = a1.dot(W2), we get dW2 by multiplying a1.T and dscores
# W2 is H x C so dW2 should also match those dimensions
# a1.T x dscores = (H x N) x (N x C) = H x C
dW2 = np.dot(a1.T, dscores)
# Right now the gradient is a sum over all training examples, but we want it
# to be an average instead so we divide by N.
dW2 /= N
# b2 gradient: sum dscores over all N and C
db2 = dscores.sum(axis = 0)/N
# since a1 = X.dot(W1), we get dW1 by multiplying X.T and da1
# W1 is D x H so dW1 should also match those dimensions
# X.T x da1 = (D x N) x (N x H) = D x H
# first get da1 using scores = a1.dot(W2)
# a1 is N x H so da1 should also match those dimensions
# dscores x W2.T = (N x C) x (C x H) = N x H
da1 = dscores.dot(W2.T)
da1[a1 == 0] = 0 # set gradient of units that did not activate to 0
dW1 = X.T.dot(da1)
# Right now the gradient is a sum over all training examples, but we want it
# to be an average instead so we divide by N.
dW1 /= N
# b1 gradient: sum da1 over all N and H
db1 = da1.sum(axis = 0)/N
# Add regularization loss to the gradient
dW1 += 2 * reg * W1
dW2 += 2 * reg * W2
grads = {'W1': dW1, 'b1': db1, 'W2': dW2, 'b2': db2}
# *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)*****
return loss, grads
def train(self, X, y, X_val, y_val,
learning_rate=1e-3, learning_rate_decay=0.95,
reg=5e-6, num_iters=100,
batch_size=200, verbose=False):
"""
Train this neural network using stochastic gradient descent.
Inputs:
- X: A numpy array of shape (N, D) giving training data.
- y: A numpy array f shape (N,) giving training labels; y[i] = c means that
X[i] has label c, where 0 <= c < C.
- X_val: A numpy array of shape (N_val, D) giving validation data.
- y_val: A numpy array of shape (N_val,) giving validation labels.
- learning_rate: Scalar giving learning rate for optimization.
- learning_rate_decay: Scalar giving factor used to decay the learning rate
after each epoch.
- reg: Scalar giving regularization strength.
- num_iters: Number of steps to take when optimizing.
- batch_size: Number of training examples to use per step.
- verbose: boolean; if true print progress during optimization.
"""
num_train = X.shape[0]
iterations_per_epoch = max(num_train / batch_size, 1)
# Use SGD to optimize the parameters in self.model
loss_history = []
train_acc_history = []
val_acc_history = []
for it in range(num_iters):
X_batch = None
y_batch = None
#########################################################################
# TODO: Create a random minibatch of training data and labels, storing #
# them in X_batch and y_batch respectively. #
#########################################################################
# *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)*****
# generate random indices
indices = np.random.choice(num_train, batch_size)
X_batch, y_batch = X[indices], y[indices]
# *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)*****
# Compute loss and gradients using the current minibatch
loss, grads = self.loss(X_batch, y=y_batch, reg=reg)
loss_history.append(loss)
#########################################################################
# TODO: Use the gradients in the grads dictionary to update the #
# parameters of the network (stored in the dictionary self.params) #
# using stochastic gradient descent. You'll need to use the gradients #
# stored in the grads dictionary defined above. #
#########################################################################
# *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)*****
self.params['W1'] -= learning_rate * grads['W1']
self.params['W2'] -= learning_rate * grads['W2']
self.params['b1'] -= learning_rate * grads['b1']
self.params['b2'] -= learning_rate * grads['b2']
# *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)*****
if verbose and it % 100 == 0:
print('iteration %d / %d: loss %f' % (it, num_iters, loss))
# Every epoch, check train and val accuracy and decay learning rate.
if it % iterations_per_epoch == 0:
# Check accuracy
train_acc = (self.predict(X_batch) == y_batch).mean()
val_acc = (self.predict(X_val) == y_val).mean()
train_acc_history.append(train_acc)
val_acc_history.append(val_acc)
# Decay learning rate
learning_rate *= learning_rate_decay
return {
'loss_history': loss_history,
'train_acc_history': train_acc_history,
'val_acc_history': val_acc_history,
}
def predict(self, X):
"""
Use the trained weights of this two-layer network to predict labels for
data points. For each data point we predict scores for each of the C
classes, and assign each data point to the class with the highest score.
Inputs:
- X: A numpy array of shape (N, D) giving N D-dimensional data points to
classify.
Returns:
- y_pred: A numpy array of shape (N,) giving predicted labels for each of
the elements of X. For all i, y_pred[i] = c means that X[i] is predicted
to have class c, where 0 <= c < C.
"""
y_pred = None
###########################################################################
# TODO: Implement this function; it should be VERY simple! #
###########################################################################
# *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)*****
# define lamba function for relu
relu = lambda x: np.maximum(0, x)
# activations of fully connected layer #1
a1 = relu(X.dot(self.params['W1']) + self.params['b1'])
# output of softmax
# scores = a1 x W2 = (N x H) x (H x C) = N x C
scores = a1.dot(self.params['W2']) + self.params['b2']
y_pred = np.argmax(scores, axis = 1)
# *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)*****
return y_pred
With regards to above code, I could not understand how DW2 was computed well. I took picture of the point I need to clarify and need an explanation for the difference.enter image description here
My ideas

Why is softmax classifier gradient divided by batch size (CS231n)?

Question
In CS231 Computing the Analytic Gradient with Backpropagation which is first implementing a Softmax Classifier, the gradient from (softmax + log loss) is divided by the batch size (number of data being used in a cycle of forward cost calculation and backward propagation in the training).
Please help me understand why it needs to be divided by the batch size.
The chain rule to get the gradient should be below. Where should I incorporate the division?
Derivative of Softmax loss function
Code
N = 100 # number of points per class
D = 2 # dimensionality
K = 3 # number of classes
X = np.zeros((N*K,D)) # data matrix (each row = single example)
y = np.zeros(N*K, dtype='uint8') # class labels
#Train a Linear Classifier
# initialize parameters randomly
W = 0.01 * np.random.randn(D,K)
b = np.zeros((1,K))
# some hyperparameters
step_size = 1e-0
reg = 1e-3 # regularization strength
# gradient descent loop
num_examples = X.shape[0]
for i in range(200):
# evaluate class scores, [N x K]
scores = np.dot(X, W) + b
# compute the class probabilities
exp_scores = np.exp(scores)
probs = exp_scores / np.sum(exp_scores, axis=1, keepdims=True) # [N x K]
# compute the loss: average cross-entropy loss and regularization
correct_logprobs = -np.log(probs[range(num_examples),y])
data_loss = np.sum(correct_logprobs)/num_examples
reg_loss = 0.5*reg*np.sum(W*W)
loss = data_loss + reg_loss
if i % 10 == 0:
print "iteration %d: loss %f" % (i, loss)
# compute the gradient on scores
dscores = probs
dscores[range(num_examples),y] -= 1
dscores /= num_examples # <---------------------- Why?
# backpropate the gradient to the parameters (W,b)
dW = np.dot(X.T, dscores)
db = np.sum(dscores, axis=0, keepdims=True)
dW += reg*W # regularization gradient
# perform a parameter update
W += -step_size * dW
b += -step_size * db

It's because you are averaging the gradients instead of taking directly the sum of all the gradients.
You could of course not divide for that size, but this division has a lot of advantages. The main reason is that it's a sort of regularization (to avoid overfitting). With smaller gradients the weights cannot grow out of proportions.
And this normalization allows comparison between different configuration of batch sizes in different experiments (How can I compare two batch performances if they are dependent to the batch size?)
If you divide for that size the gradients sum it could be useful to work with greater learning rates to make the training faster.
This answer in the crossvalidated community is quite useful.

Came to notice that the dot in dW = np.dot(X.T, dscores) for the gradient at W is Σ over the num_sample instances. Since the dscore, which is probability (softmax output), was divided by the num_samples, did not understand that it was normalization for dot and sum part later in the code. Now understood divide by num_sample is required (may still work without normalization if the learning rate is trained though).
I believe the code below explains better.
# compute the gradient on scores
dscores = probs
dscores[range(num_examples),y] -= 1
# backpropate the gradient to the parameters (W,b)
dW = np.dot(X.T, dscores) / num_examples
db = np.sum(dscores, axis=0, keepdims=True) / num_examples

No gradients provided for any variable - Custom loss function with random weights depending on the Softmax output

I have difficulties writing a custom loss function that makes use of some random weights generated according to the class/state predicted by the Softmax output. The desired property is:
The model is a simple feedforward neural network with input-dimension as 1 and the output dimension as 6.
The activation function of the output layer is Softmax, which intends to estimate the actual number of classes or states using Argmax.
Note that the training data only consists of X (there is no Y).
The loss function is defined according to random weights (i.e., Weibull distribution) sampled based on the predicted state number for each input sample X.
As follows, I provided a minimal example for illustration. For simplification purposes, I only define the loss function based on the random weights for state/class-1. I get: "ValueError: No gradients provided for any variable: ['dense_41/kernel:0', 'dense_41/bias:0', 'dense_42/kernel:0', 'dense_42/bias:0']."
As indicated in the post below, I found out that argmax is not differntiable, and a softargmax function would help (as I implemented in the following code). However, I still get the same error.
Getting around tf.argmax which is not differentiable
import sys
import time
from tqdm import tqdm
import tensorflow as tf
import numpy as np
from tensorflow.keras import layers
from scipy.stats import weibull_min
###############################################################################################
# Generate Dataset
lb = np.array([2.0]) # Left boundary
ub = np.array([100.0]) # Right boundary
# Data Points - uniformly distributed
N_r = 50
X_r = np.linspace(lb, ub, N_r)
###############################################################################################
#Define Model
class DGM:
# Initialize the class
def __init__(self, X_r):
#Normalize training input data
self.Xmean, self.Xstd = np.mean(X_r), np.std(X_r)
X_r = (X_r - self.Xmean) / self.Xstd
self.X_r = X_r
#Input and output variable dimensions
self.X_dim = 1; self.Y_dim = 6
# Define tensors
self.X_r_tf = tf.convert_to_tensor(X_r, dtype=tf.float32)
#Learning rate
self.LEARNING_RATE=1e-4
#Feedforward neural network model
self.modelTest = self.test_model()
###############################################
# Initialize network weights and biases
def test_model(self):
input_shape = self.X_dim
dimensionality = self.Y_dim
model = tf.keras.Sequential()
model.add(layers.Input(shape=input_shape))
model.add(layers.Dense(64, kernel_initializer='glorot_uniform',bias_initializer='zeros'))
model.add(layers.Activation('tanh'))
model.add(layers.Dense(dimensionality))
model.add(layers.Activation('softmax'))
return model
##############################################
def compute_loss(self):
#Define optimizer
gen_opt = tf.keras.optimizers.Adam(lr=self.LEARNING_RATE, beta_1=0.0,beta_2=0.9)
with tf.GradientTape() as test_tape:
###### calculate loss
generated_u = self.modelTest(self.X_r_tf, training=True)
#number of data
n_data = generated_u.shape[0]
#initialize random weights assuming state-1 at all input samples
wt1 = np.zeros((n_data, 1),dtype=np.float32) #initialize weights
for b in range(n_data):
wt1[b] = weibull_min.rvs(c=2, loc=0, scale =4 , size=1)
wt1 = tf.reshape(tf.convert_to_tensor(wt1, dtype=tf.float32),shape=(n_data,1))
#print('-----------sampling done-----------')
#determine the actual state using softargmax
idst = self.softargmax(generated_u)
idst = tf.reshape(tf.cast(idst, tf.float32),shape=(n_data,1))
#index state-1
id1 = tf.constant(0.,dtype=tf.float32)
#assign weights if predicted state is state-1
wt1_final = tf.cast(tf.equal(idst, id1), dtype=tf.float32)*wt1
#final loss
test_loss = tf.reduce_mean(tf.square(wt1_final))
#print('-----------test loss calcuated-----------')
gradients_of_modelTest = test_tape.gradient(test_loss,
[self.modelTest.trainable_variables])
gen_opt.apply_gradients(zip(gradients_of_modelTest[0],self.modelTest.trainable_variables))
return test_loss
#reference: Getting around tf.argmax which is not differentiable
#https://stackoverflow.com/questions/46926809/getting-around-tf-argmax-which-is-not-differentiable
def softargmax(self, x, beta=1e10):
x = tf.convert_to_tensor(x)
x_range = tf.range(x.shape.as_list()[-1], dtype=x.dtype)
return tf.reduce_sum(tf.nn.softmax(x*beta,axis=1) * x_range, axis=-1)
##############################################
def train(self,training_steps=100):
train_start_time = time.time()
for step in tqdm(range(training_steps), desc='Training'):
start = time.time()
test_loss = self.compute_loss()
if (step + 1) % 10 == 0:
elapsed_time = time.time() - train_start_time
sec_per_step = elapsed_time / step
mins_left = ((training_steps - step) * sec_per_step)
tf.print("\nStep # ", step, "/", training_steps,
output_stream=sys.stdout)
tf.print("Current time:", elapsed_time, " time left:",
mins_left, output_stream=sys.stdout)
tf.print("Test Loss: ", test_loss, output_stream=sys.stdout)
###############################################################################################
#Define and train the model
model = DGM(X_r)
model.train(training_steps=100)

Autoencoder using Backpropagation

I'm trying to implement autoencoder using this resource which implements backpropagation algorithm. I'm using the same feed forward algorithm implemented there but however it gives me a large error. In Autoencoders, the sigmoid function to be applied to the hidden for encoding and again to the output for decoding.
def feedForwardPropagation(network, row, output=False):
currentInput = row
if not output:
layer = network[0]
else:
layer = network[1]
layer_output = []
for neuron in layer:
activation = neuron_activation(neuron['weights'], currentInput)
neuron['output'] = neuron_transfer(activation)
layer_output.append(neuron['output'])
currentInput = layer_output
return currentInput
def backPropagationNetworkErrorUpdate(network, expected):
for i in reversed(range(len(network))):
layer = network[i]
errors = list()
if i != len(network) - 1:
# Hidden Layers weight error compute
for j in range(len(layer)):
error = 0.0
for neuron in network[i + 1]: # It starts with computing weight error of output neuron.
error += (neuron['weights'][j] * neuron['delta'])
errors.append(error)
else:
# Output layer error computer
for j in range(len(layer)):
neuron = layer[j]
error = expected[j] - neuron['output']
errors.append(error)
for j in range(len(layer)):
neuron = layer[j]
transfer = neuron['output'] * (1.0 - neuron['output'])
neuron['delta'] = errors[j] * transfer
def updateWeights(network, row, l_rate, momentum=0.5):
for i in range(len(network)):
inputs = row[:-1]
if i != 0:
inputs = [neuron['output'] for neuron in network[i - 1]]
for neuron in network[i]:
for j in range(len(inputs)):
neuron['velocity'][j] = momentum * neuron['velocity'][j] + l_rate * neuron['delta'] * inputs[j]
neuron['weights'][j] += neuron['velocity'][j]
neuron['velocity'][-1] = momentum * neuron['velocity'][-1] + l_rate * neuron['delta'] * inputs[j]
neuron['weights'][-1] += neuron['velocity'][-1]
def trainNetwork(network, train, l_rate, n_epoch, n_outputs, test_set):
hitrate = list()
errorRate = list()
epoch_step = list()
for epoch in range(n_epoch):
sum_error = 0
np.random.shuffle(train)
for row in train:
outputs = feedForwardPropagation(network, row)
outputs = feedForwardPropagation(network, outputs)
expected = row
sum_error += sum([(expected[i] - outputs[i]) ** 2 for i in range(len(expected))])
backPropagationNetworkErrorUpdate(network, expected)
updateWeights(network, row, l_rate)
if epoch % 10 == 0:
errorRate.append(sum_error)
epoch_step.append(epoch)
log = '>epoch=%d, lrate=%.3f, error=%.3f' % (epoch, l_rate, sum_error)
print(log, n_epoch, len(network[1][0]['weights']) - 1, l_rate)
return epoch_step, errorRate
For autoencoding I use one hidden layer, n inputs and n outputs. I believe I have gone wrong with the feedforward implementation. Any suggestions will be greatly appreciated.
Edit: I tried computing the weights after first layer (continue commented in feedforward method) and then decoding the output using the sigmoid function commented in trainNetwork method. However, the error didn't change after 100 epochs

The characteristics of your problem (like error barely changing over 100 epochs, and remaining with a big error), suggest that the problem might be (and probably is) caused by the order of size of your input data, and the fact that you use sigmoids as activation function. I will give you a simple example:
Suppose I want to reconstruct the value x=100.
If I train it with an autoencoder on a single neuron, the reconstructed output will be given by r = sigmoid(w*x), where the error is the difference between the actual input and the reconstruction, i.e. e = x - r. Note, that since a sigmoid function is bounded between -1 and 1 , the minimum error you can get in this case is e = 100-1 = 99. No matter how good you train the weight w in this case, r=sigmoid(w*x) would always be bounded by one.
This means that the sigmoid activation function is not able to represent your data in this case.
To solve this problem, either:
Downscale or Normalize your input data to a size between -1 and 1, or
Change the sigmoid to another activation function, that can actually reconstruct the right size of order of your data.
Hope this helps.

Neural Network loss starts increasing while acc is increasing on both train/val data sets

For past few days I have been debugging my NN but I can't find an issue.
I've created total raw implementation of multi-layer perceptron for identifying MNIST dataset images.
Network seems to learn because after train cycle test data accuracy is above 94% accuracy. I have problem with loss function - it starts increasing after a while, when test/val accuracy reaches ~76%.
Can someone please check my forward/backprop math and tell me if my loss function is properly implemented, or suggest what might be wrong?
NN structure:
input layer: 758 nodes, (1 node per pixel)
hidden layer 1: 300 nodes
hidden layer 2: 75 nodes
output layer: 10 nodes
NN activation functions:
input layer -> hidden layer 1: ReLU
hidden layer 1 -> hidden layer 2: ReLU
hidden layer 2 -> output layer 3: Softmax
NN Loss function:
Categorial Cross-Entropy
Full CLEAN code available here as Jupyter Notebook.
Neural Network forward/backward pass:
def train(self, features, targets):
n_records = features.shape[0]
# placeholders for weights and biases change values
delta_weights_i_h1 = np.zeros(self.weights_i_to_h1.shape)
delta_weights_h1_h2 = np.zeros(self.weights_h1_to_h2.shape)
delta_weights_h2_o = np.zeros(self.weights_h2_to_o.shape)
delta_bias_i_h1 = np.zeros(self.bias_i_to_h1.shape)
delta_bias_h1_h2 = np.zeros(self.bias_h1_to_h2.shape)
delta_bias_h2_o = np.zeros(self.bias_h2_to_o.shape)
for X, y in zip(features, targets):
### forward pass
# input to hidden 1
inputs_to_h1_layer = np.dot(X, self.weights_i_to_h1) + self.bias_i_to_h1
inputs_to_h1_layer_activated = self.activation_ReLU(inputs_to_h1_layer)
# hidden 1 to hidden 2
h1_to_h2_layer = np.dot(inputs_to_h1_layer_activated, self.weights_h1_to_h2) + self.bias_h1_to_h2
h1_to_h2_layer_activated = self.activation_ReLU(h1_to_h2_layer)
# hidden 2 to output
h2_to_output_layer = np.dot(h1_to_h2_layer_activated, self.weights_h2_to_o) + self.bias_h2_to_o
h2_to_output_layer_activated = self.softmax(h2_to_output_layer)
# output
final_outputs = h2_to_output_layer_activated
### backpropagation
# output to hidden2
error = y - final_outputs
output_error_term = error.dot(self.dsoftmax(h2_to_output_layer_activated))
h2_error = np.dot(output_error_term, self.weights_h2_to_o.T)
h2_error_term = h2_error * self.activation_dReLU(h1_to_h2_layer_activated)
# hidden2 to hidden1
h1_error = np.dot(h2_error_term, self.weights_h1_to_h2.T)
h1_error_term = h1_error * self.activation_dReLU(inputs_to_h1_layer_activated)
# weight & bias step (input to hidden)
delta_weights_i_h1 += h1_error_term * X[:, None]
delta_bias_i_h1 = np.sum(h1_error_term, axis=0)
# weight & bias step (hidden1 to hidden2)
delta_weights_h1_h2 += h2_error_term * inputs_to_h1_layer_activated[:, None]
delta_bias_h1_h2 = np.sum(h2_error_term, axis=0)
# weight & bias step (hidden2 to output)
delta_weights_h2_o += output_error_term * h1_to_h2_layer_activated[:, None]
delta_bias_h2_o = np.sum(output_error_term, axis=0)
# update the weights and biases
self.weights_i_to_h1 += self.lr * delta_weights_i_h1 / n_records
self.weights_h1_to_h2 += self.lr * delta_weights_h1_h2 / n_records
self.weights_h2_to_o += self.lr * delta_weights_h2_o / n_records
self.bias_i_to_h1 += self.lr * delta_bias_i_h1 / n_records
self.bias_h1_to_h2 += self.lr * delta_bias_h1_h2 / n_records
self.bias_h2_to_o += self.lr * delta_bias_h2_o / n_records
Activation function implementation:
def activation_ReLU(self, x):
return x * (x > 0)
def activation_dReLU(self, x):
return 1. * (x > 0)
def softmax(self, x):
z = x - np.max(x)
return np.exp(z) / np.sum(np.exp(z))
def dsoftmax(self, x):
# TODO: vectorise math
vec_len = len(x)
J = np.zeros((vec_len, vec_len))
for i in range(vec_len):
for j in range(vec_len):
if i == j:
J[i][j] = x[i] * (1 - x[j])
else:
J[i][j] = -x[i] * x[j]
return J
Loss function implementation:
def categorical_cross_entropy(pred, target):
return (1/len(pred)) * -np.sum(target * np.log(pred))

I managed to find the problem.
Neural Network is large so I couldn't stick everything to this question. Though if you check my Jupiter Notebook you could see implementation of my Softmax activation function and how do I use it in train cycle.
Problem with Loss miscalculation was caused by the fact my Softmax implementation worked only for ndarray dim == 1.
During training step I have put only ndarray with dim 1 to activtion function so NN learned well, but my run() function was returning wrong predictions as I have inserted whole test data to it, not only single row of it in for loop. Because of that it calculated Softmax "matrix-wise" rather than "row-wise".
This is very fast fix for it:
def softmax(self, x):
# TODO: vectorise math to speed up computation
softmax_result = None
if x.ndim == 1:
z = x - np.max(x)
softmax_result = np.exp(z) / np.sum(np.exp(z))
return softmax_result
else:
softmax_result = []
for row in x:
z = row - np.max(row)
row_softmax_result = np.exp(z) / np.sum(np.exp(z))
softmax_result.append(row_softmax_result)
return np.array(softmax_result)
Yet this code should be vectorised to avoid for loops and ifs if possible because currently it's ugly and takes too much PC resources.