froog 0.4.2

a toy tensor library with opencl 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

More information on downloading froog in the installation docs.

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
• and a bunch more

Sneak Peek

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

Overview

The most fundamental concept in all of froog and machine learning frameworks is the Tensor. A tensor is simply a matrix of matrices (more accurately a multi-dimensional array).

You can create a Tensor in froog with:

import numpy as np
from froog.tensor import Tensor
my_tensor = Tensor([1,2,3])

Notice how we had to import NumPy. If you want to create a Tensor manually, make sure that it is a NumPy array!

Tensors

Tensors are the fundamental datatype in froog, and one of the two main classes.

• def __init__(self, data):

  • Tensor takes in one param, which is the data. Since froog has a NumPy backend, the input data into tensors has to be a NumPy array.
  • Tensor has a self.data state that it holds. this contains the data inside of the tensor.
  • In addition, it has self.grad. this is to hold what the gradients of the tensor is.
  • Lastly, it has self._ctx. These are the internal variables used for autograd graph construction. This is where the backward gradient computations are saved.

Properties

• shape(self): this returns the tensor shape

Methods

• def zeros(*shape): this returns a tensor full of zeros with any shape that you pass in. Defaults to np.float32
• def ones(*shape): this returns a tensor full of ones with any shape that you pass in. Defaults to np.float32
• def randn(*shape):: this returns a randomly initialized Tensor of *shape

Gradient calculations

• froog computes gradients automatically through a process called automatic differentiation. it has a variable _ctx, which stores the chain of operations. It will take the current operation, let's say a dot product, and go to the dot product definition in froog/ops.py, which contains a backward pass specifically for dot products. all methods, from add to 2x2 maxpools, have this backward pass implemented.

Functions

The other base class in froog is the class Function. It keeps track of input tensors and tensors that need to be saved for backward passes

• def __init__(self, *tensors): takes in an argument of tensors, which are then saved.
• def save_for_backward(self, *x): saves Tensors that are necessary to compute for the computation of gradients in the backward pass.
• def apply(self, arg, *x): takes care of the forward pass, applying the operation to the inputs.

Register

• def register(name, fxn): allows you to add a method to a Tensor. This allows you to chain any operations, e.g. x.dot(w).relu(), where w is a tensor

Creating a model

Okay cool, so now you know that froog's main datatype is a Tensor and uses NumPy in the background. How do I actually build a model?

Here's an example of how to create an MNIST multi-layer perceptron (MLP). We wanted to make it as simple as possible for you to do so it resembles very basic Python concepts like classes. There are really only two methods you need to define:

1. __init__ that defines layers of the model (here we use Linear)
2. forward which defines how the input should flow through your model. We use a simple dot product with a Linear layer with a ReLU activation.

To create an instance of the mnistMLP model, do the same as you would in Python: model = mnistMLP().

We support a few different optimizers, here which include:

• Stochastic Gradient Descent (SGD)
• Adaptive Moment Estimation (Adam)
• Root Mean Square Propagation (RMSProp)

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

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

  def forward(self, x):
    return x.dot(self.l1).relu().dot(self.l2).logsoftmax()

model = mnistMLP()
optim = optim.SGD([model.l1, model.l2], lr=0.001)

You can also create a convolutional neural net by

class SimpleConvNet:
  def __init__(self):
    conv_size = 5
    channels = 17
    self.c1 = Tensor(Linear(channels,1,conv_size,conv_size))     # (num_filters, color_channels, kernel_h, kernel_w)
    self.l1 = Tensor(Linear((28-conv_size+1)**2*channels, 128))  # (28-conv+1)(28-conv+1) since kernel isn't padded
    self.l2 = Tensor(Linear(128, 10))                            # MNIST output is 10 classes

  def forward(self, x):
    x.data = x.data.reshape((-1, 1, 28, 28))                     # get however many number of imgs in batch
    x = x.conv2d(self.c1).relu()                                 # pass through conv first
    x = x.reshape(shape=(x.shape[0], -1))
    return x.dot(self.l1).relu().dot(self.l2).logsoftmax()

So there are two quick examples to get you up and running. You might have noticed some operations like reshape and were wondering what else you can do with froog. We have many more operations that you can apply on tensors:

• .add()
• .sub()
• .mul()
• .sum()
• .pow()
• .dot()
• .relu()
• .sigmoid()
• .reshape()
• .pad2d()
• .logsoftmax()
• .conv2d()
• .im2col2dconv()
• .max_pool2d()
• .avg_pool2d()

GPU Support

Have a GPU and need a speedup? You're in good luck because we have GPU support via OpenCL for our operations defined in ops_gpu.py.

Here's how you can send data to the GPU during a forward pass and bring it back to the CPU.

# ...
GPU = os.getenv("GPU", None) is not None
if GPU:
  out = model.forward(Tensor(img).to_gpu()).cpu()

EfficientNet in froog!

We have a really cool finished implementation of EfficientNet built entirely in froog!

In order to run EfficientNet inference:

VIZ=1 python3 models/efficientnet.py <https://put_your_image_url_here>

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.