# Show HN: I created a new Deep Learning Framework

> Source: <https://picodl.vercel.app>
> Published: 2026-08-05 08:53:07+00:00

// numpy-only · zero dependencies

picodl is a from-scratch autograd engine and neural network stack - every gradient traced by hand, every op built on plain numpy. No hidden framework underneath.

``` bash
$ pip install picodl-nn
```

// what's inside

Each piece does one job and hands a plain `Tensor`

to the next. Nothing is generated for you - every layer, loss, and optimizer is code you can read start to finish.

tensor.py

Tracks every op into a graph, topologically sorts on `.backward()`

. 25+ differentiable ops: matmul, conv2d, softmax, attention primitives, and more.

layers.py

Linear, Conv2D, Embedding, LayerNorm, BatchNorm2D, MaxPool2D, AvgPool2D, GlobalAvgPool, Dropout, GELU, TiedLinear.

loss.py

MSE, BinaryCrossEntropy, NLLLoss, CrossEntropyLoss - built from the same primitive ops as everything else.

optim.py

SGD, RMSprop, Adam, AdamW - decoupled weight decay included, per-parameter state tracked by identity.

nn.py

Chains layers, exposes params for the optimizer, saves and loads weights to any file extension you like.

data.py / train.py

BatchIterator for mini-batches, a train loop that accepts raw numpy or Tensor input directly.

// thirty seconds in

A small classifier - GELU hidden layer, Adam optimizer, saved to disk when done.

``` python
from picodl.nn import NeuralNet
from picodl.layers import Linear, GELU
from picodl.loss import CrossEntropyLoss
from picodl.optim import AdamW
from picodl.train import train

net = NeuralNet([
    Linear(784, 128),
    GELU(),
    Linear(128, 10),
])

train(net, x_train, y_train,
      num_epochs=10,
      loss=CrossEntropyLoss(),
      optimizer=AdamW(lr=0.001, weight_decay=0.01))

net.save("model.picodl")
```