froog 0.5.0

tensor library with opencl and metal support

froog

froog: fast real-time optimization of gradients
a beautifully compact tensor library
homepage | documentation | pip

froog is an easy-to-read tensor library (25k pip installs!) with OpenCL support for GPU acceleration. Inspired by pytorch, tinygrad, and micrograd.

Installation

pip install froog

Features

• Custom Tensors
  • Backpropagation
  • Automatic Differentiation (autograd)
    • Forward and backward passes
• ML Operations
  • 2D Convolutions (im2col)
  • Numerical gradient checking
  • Acceleration methods (Adam)
  • Avg & Max pooling
• EfficientNet inference
• GPU Support
• Configuration via Environment Variables
• and a bunch more

Example

Here's how you set up a simple multilayer perceptron for classification on MNIST. Looks pretty similar to pytorch, right?

from froog.tensor import Tensor
from froog.nn import Linear
import froog.optim as optim

class mnistMLP:
  def __init__(self):
    self.l1 = Tensor(Linear(784, 128)) # layer 1
    self.l2 = Tensor(Linear(128, 10))  # layer 2

  def forward(self, x):
    # forward pass through both layers and softmax for output probabilities
    return x.dot(self.l1).relu().dot(self.l2).logsoftmax() 

model = mnistMLP() # create model
optim = optim.SGD([model.l1, model.l2], lr=0.001) # stochastic gradient descent optimizer

GPU Support

Device management is handled transparently and will automatically select one of [METAL, OPENCL, CPU]. To use the GPU:

from froog.tensor import Tensor
from froog import get_device
# Check if GPU is available
has_gpu = get_device() is not None and get_device().name != "CPU"
# Create a tensor
x = Tensor([1, 2, 3])
# Push to GPU if available
if has_gpu: x = x.to_gpu()
# Operations run on GPU automatically
y = x + x
z = y * y
# Bring back to CPU when needed
result = z.to_cpu()
print(result.data)

You can also check what devices are available:

from froog import get_available_devices
available_devices = get_available_devices()
print(f"Available devices: {available_devices}")

Or set a specific device:

from froog import set_device
set_device("METAL")  # or "OPENCL"

EfficientNet in froog!

We have an implementation of EfficientNet v2 built entirely in froog using the official PyTorch weights! Run inference with:

python3 models/efficientnet.py <https://optional_image_url>

***********output*************
inference 4.34 s

imagenet class: 254
prediction    : pug, pug-dog
probability   : 0.8902361
******************************

I would recommend checking out the code, it's highly documented and pretty cool.

Contributing

Pull requests will be merged if they:

• increase simplicity
• increase functionality
• increase efficiency

More info on contributing. Make sure to run python -m pytest before creating a PR.

API

Basic Math Operations

• .add(y) - Addition with y
• .sub(y) - Subtraction with y
• .mul(y) - Multiplication with y
• .div(y) - Division by y
• .pow(y) - Power function (raise to power y)
• .sum() - Sum all elements
• .mean() - Mean of all elements
• .sqrt() - Square root

Linear Algebra Operations

• .dot(y) - Matrix multiplication with y
• .matmul(y) - Alias for dot

Neural Network Operations

• .relu() - Rectified Linear Unit activation
• .sigmoid() - Sigmoid activation
• .dropout(p=0.5, training=True) - Dropout regularization
• .logsoftmax() - Log softmax function
• .swish() - Swish activation function (x * sigmoid(x))
• .conv2d(w, stride=1, groups=1) - 2D convolution
• .im2col2dconv(w) - Image to column for convolution

Pooling Operations

• .max_pool2d(kernel_size=(2,2)) - 2D max pooling
• .avg_pool2d(kernel_size=(2,2)) - 2D average pooling

Tensor Manipulation

• .reshape(*shape) - Change tensor shape
• .view(*shape) - Alternative to reshape
• .pad2d(padding=None) - Pad 2D tensors
• .flatten() - Returns a flattened 1D copy of the tensor
• .unsqueeze(dim) - Add dimension of size 1 at specified position
• .squeeze(dim=None) - Remove dimensions of size 1
• .detach() - Returns a tensor detached from computation graph
• .assign(x) - Assign values from tensor x to this tensor

Tensor Properties

• .shape - The shape of the tensor as a tuple
• .size - Total number of elements in the tensor
• .ndim - Number of dimensions (rank) of the tensor
• .transpose - Transpose of the tensor
• .dtype - Data type of the tensor
• .is_gpu - Whether tensor is on GPU
• .grad - Gradient of tensor with respect to some scalar value
• .data - Underlying NumPy array (or GPU buffer)

Device Management

• .to_cpu() - Moves tensor to CPU
• .to_gpu() - Moves tensor to GPU
• .gpu_() - In-place GPU conversion (modifies tensor)

Data Type Conversion

• .to_float() - Converts tensor to float32 data type
• .to_int() - Converts tensor to int32 data type
• .to_bool() - Converts tensor to boolean data type

Autograd Operations

• .backward(allow_fill=True) - Performs backpropagation

Tensor Creation Methods

• Tensor.zeros(*shape) - Create tensor of zeros
• Tensor.ones(*shape) - Create tensor of ones
• Tensor.randn(*shape) - Create tensor with random normal values
• Tensor.eye(dim) - Create identity matrix
• Tensor.arange(start, stop=None, step=1) - Create tensor with evenly spaced values