The SimpleMind class is a powerful yet straightforward implementation of a neural network in JAX. It supports various activation functions, optimizers, and regularization techniques, making it versatile for different machine learning tasks. With parallel backpropagation and detailed logging, it provides an efficient and transparent framework for neural network training.
The python SimpleMind class represents a neural network implemented in JAX, designed to be versatile and efficient for various machine learning tasks. Below is a detailed explanation of the class and its components.
Overview
SimpleMind is a neural network class that supports initialization, forward propagation, backpropagation, and training using JAX. It supports different activation functions and optimizers, and it also includes regularization for better generalization.
Components of SimpleMind
Initialization
The constructor initializes the network parameters and sets up logging.
class SimpleMind:
def init(self, input_size, hidden_sizes, output_size, activation='relu', optimizer='adam', learning_rate=0.001, regularization=None, reg_lambda=0.01):
# Initialization code
input_size: Size of the input layer.
hidden_sizes: List of sizes for hidden layers.
output_size: Size of the output layer.
activation: Activation function ('relu', 'sigmoid', 'tanh').
optimizer: Optimizer to use ('adam', 'sgd').
learning_rate: Learning rate for the optimizer.
regularization: Regularization method ('l2').
reg_lambda: Regularization strength.
Parameter Initialization
The _initialize_parameters method initializes the weights and biases for each layer.
def _initialize_parameters(self):
layer_sizes = [self.input_size] + self.hidden_sizes + [self.output_size]
params = {}
for i in range(len(layer_sizes) - 1):
self.rng, layer_rng = random.split(self.rng)
params[f'W{i}'] = random.normal(layer_rng, (layer_sizes[i], layer_sizes[i+1])) * 0.01
params[f'b{i}'] = jnp.zeros(layer_sizes[i+1])
return params
Optimizer Initialization
The _initialize_optimizer method sets up the optimizer state.
def _initialize_optimizer(self):
if self.optimizer_name == 'adam':
optimizer = optax.adam(self.learning_rate)
elif self.optimizer_name == 'sgd':
optimizer = optax.sgd(self.learning_rate)
else:
raise ValueError(f"Unsupported optimizer: {self.optimizer_name}")
return optimizer.init(self.params)
Activation Functions
The _activation_function method applies the chosen activation function.
def _activation_function(self, s):
if self.activation == 'sigmoid':
return 1 / (1 + jnp.exp(-s))
elif self.activation == 'tanh':
return jnp.tanh(s)
elif self.activation == 'relu':
return jnp.maximum(0, s)
else:
raise ValueError("Unsupported activation function.")
Forward Propagation
The forward method performs a forward pass through the network.
def forward(self, X):
out = X
for i in range(len(self.hidden_sizes)):
W, b = self.params[f'W{i}'], self.params[f'b{i}']
out = jnp.dot(out, W) + b
out = self._activation_function(out)
W, b = self.params[f'W{len(self.hidden_sizes)}'], self.params[f'b{len(self.hidden_sizes)}']
out = jnp.dot(out, W) + b
return out
Backpropagation
The backpropagate method calculates the gradients and updates the parameters.
grads = grad(loss)(self.params)
updates, self.opt_state = self.optimizer.update(grads, self.opt_state)
self.params = optax.apply_updates(self.params, updates)
@jit
def backpropagate(self, X, y):
def loss(params):
predictions = self.forward(X)
loss_value = jnp.mean((predictions - y) ** 2)
if self.regularization == 'l2':
l2_penalty = sum(jnp.sum(jnp.square(params[f'W{i}'])) for i in range(len(self.hidden_sizes) + 1))
loss_value += self.reg_lambda * l2_penalty / 2
return loss_value
Training
The train method trains the network for a specified number of epochs.
def train(self, X, y, epochs):
for epoch in range(epochs):
self.params, self.opt_state = self._parallel_backpropagate(X, y)
if epoch % 100 == 0:
loss_value = self._calculate_loss(X, y)
logging.info(f"Epoch {epoch}, Loss: {loss_value}")
Parallel Backpropagation
The _parallel_backpropagate method performs backpropagation in parallel using multiple threads.
def _parallel_backpropagate(self, X, y):
with ThreadPoolExecutor() as executor:
futures = [executor.submit(self.backpropagate, X[i], y[i]) for i in range(len(X))]
for future in as_completed(futures):
params, opt_state = future.result()
return params, opt_state
Loss Calculation
The _calculate_loss method calculates the network’s loss.
@jit
def _calculate_loss(self, X, y):
output = self.forward(X)
loss_value = jnp.mean(jnp.square(y - output))
if self.regularization == 'l2':
loss_value += self.reg_lambda / 2 * sum(jnp.sum(jnp.square(self.params[f'W{i}'])) for i in range(len(self.hidden_sizes) + 1))
return loss_value
using the SimpleMind class:
if name == "main":
input_size = 4
hidden_sizes = [10, 10]
output_size = 1
learning_rate = 0.001
epochs = 1000
X = jnp.array([[1.0, 2.0, 3.0, 4.0]])
y = jnp.array([[1.0]])
simple_mind = SimpleMind(input_size, hidden_sizes, output_size, activation='relu', optimizer='adam', learning_rate=learning_rate, regularization='l2', reg_lambda=0.01)
simple_mind.train(X, y, epochs)
print("Final Output:", simple_mind.forward(X))
The SimpleMind class is a powerful yet straightforward implementation of a neural network written in python using JAX. It supports various activation functions, o
General Laptop Optimization: This setting balances the workload between CPU and GPU, using moderately sized hidden layers and the adam
optimizer for robust performance across different hardware configurations.GPU-Optimized: This setting takes advantage of GPU parallelism with larger hidden layers and a slightly lower learning rate for better convergence. The adam
optimizer is well-suited for GPUs.CPU-Optimized: This setting reduces the workload on the CPU with smaller hidden layers and uses the sgd
optimizer, which is lightweight and efficient for CPUs. The tanh
activation function is chosen for its efficiency on CPUs. A higher learning rate is used to speed up training on the CPU.
# General laptop optimization settings
simple_mind_general = SimpleMind(
input_size=10, # Adjust based on your specific input data
hidden_sizes=[64, 32], # Moderately sized hidden layers
output_size=1, # Adjust based on your specific output requirements
activation='relu', # Efficient activation function
optimizer='adam', # Robust optimizer for varied hardware
learning_rate=0.001, # Standard learning rate
regularization='l2', # Use L2 regularization to prevent overfitting
reg_lambda=0.01 # Regularization strength
)
# GPU-optimized settings
simple_mind_gpu = SimpleMind(
input_size=10, # Adjust based on your specific input data
hidden_sizes=[128, 64, 32],# Larger hidden layers to leverage GPU parallelism
output_size=1, # Adjust based on your specific output requirements
activation='relu', # Efficient activation function for GPUs
optimizer='adam', # Optimizer that works well with GPUs
learning_rate=0.0005, # Slightly lower learning rate for better convergence
regularization='l2', # Use L2 regularization to prevent overfitting
reg_lambda=0.01 # Regularization strength
)
# CPU-optimized settings
simple_mind_cpu = SimpleMind(
input_size=10, # Adjust based on your specific input data
hidden_sizes=[32, 16], # Smaller hidden layers to reduce CPU load
output_size=1, # Adjust based on your specific output requirements
activation='tanh', # Efficient and fast activation function for CPUs
optimizer='sgd', # Lightweight optimizer for CPU
learning_rate=0.01, # Higher learning rate for faster convergence on CPU
regularization='l2', # Use L2 regularization to prevent overfitting
reg_lambda=0.01 # Regularization strength
)