nn API

Module

from MiniTorch.nn.module import Module

Base class for all neural network components.

parameters() → list[Variable]

Returns a flat list of all learnable parameters in the module (and any sub-modules).

model = Sequential(Linear(4, 8), Linear(8, 2))
params = model.parameters()   # [w1, b1, w2, b2]

__call__(x)

Calls forward(x). Build your computation graph here.

Custom Module Example

import numpy as np
from MiniTorch.core.variable import Variable
from MiniTorch.nn.module import Module
from MiniTorch.ops.relu import relu

class TwoLayerNet(Module):
    def __init__(self, in_dim, hidden, out_dim):
        super().__init__()
        self.fc1 = Linear(in_dim, hidden)
        self.fc2 = Linear(hidden, out_dim)

    def forward(self, x):
        return self.fc2(relu(self.fc1(x)))

---

Linear

from MiniTorch.nn.linear import Linear

Fully-connected layer: y = x @ W + b

Constructor

Linear(in_features: int, out_features: int, bias: bool = True)

Parameters are initialized with Kaiming uniform.

Attributes

Attribute	Shape	Description
weight	(in_features, out_features)	Weight matrix
bias	(out_features,)	Bias vector

Example

layer = Linear(128, 64)
out = layer(x)   # x: (batch, 128) → out: (batch, 64)

---

Sequential

from MiniTorch.nn.sequential import Sequential

Chains modules in order. Input is passed through each layer sequentially.

Constructor

Sequential(*layers)

Example

model = Sequential(
    Linear(784, 256),
    Linear(256, 128),
    Linear(128, 10),
)

logits = model(x_batch)