gradients 0.0.2

Gradient Checker for Custom built PyTorch Models

Build your deep learning models with confidence

Medium article is under work

Gradients provide a unit test function to perform gradient checking on your deep learning models. It uses centered finite difference approximation to check the difference between analytical and numerical gradients and report if the check fails on any parameters of your model. Currently the library supports only PyTorch models built with custom layers, custom loss functions, activation functions and any neural network function subclassing AutoGrad.

Optimizing deep learning models is a two step process:

1. Compute gradients with respect to parameters

2. Update the parameters given the gradients

In PyTorch, step 1 is done by the type-based automatic differentiation system torch.nn.autograd and 2 by the package implementing optimization algorithms torch.optim. Using them, we can develop fully customized deep learning models with torch.nn.Module and test them using Gradients as follows;

Activation function with backward

class MySigmoid(torch.autograd.Function):

    @staticmethod
    def forward(ctx, input):
        output = ctx.save_for_backward(output)
        ctx.save_for_backward(output)
        return output

    @staticmethod
    def backward(ctx, grad_output):
        input, = ctx.saved_tensors
        return input*(1-input)*grad_output

Loss function with autograd backward

class MSELoss(torch.autograd.Function):

    @staticmethod
    def forward(ctx, y_pred, y):
        ctx.save_for_backward(y_pred, y)
        return ((y_pred-y)**2).sum()/y_pred.shape[0]

    @staticmethod
    def backward(ctx, grad_output):
        y_pred, y = ctx.saved_tensors
        grad_input = 2 * (y_pred-y)/y_pred.shape[0]
        return grad_input, None

Pytorch Model

class MyModel(torch.nn.Module):
    def __init__(self,D_in, D_out):
        super(MyModel,self).__init__()
        self.w1 = torch.nn.Parameter(torch.randn(D_in, D_out), requires_grad=True)
    def forward(self,x):
        y_pred = mysigmoid(x.mm(self.w1))
        return y_pred

Optimizer

class SGD(torch.optim.Optimizer):
    """Reference: http://pytorch.org/docs/master/_modules/torch/optim/sgd.html#SGD"""
    def __init__(self, params, lr=1e-3):
        defaults = dict(lr=lr)
        super(SGD,self).__init__(params,defaults)

    def __setstate__(self, state):
        super(SGD, self).__setstate__(state)

    def step(self, closure=None):

        loss = None
        if closure is not None:
            loss = closure()

        for group in self.param_groups:

            for p in group['params']:
                if p.grad is None:
                    continue
                d_p = p.grad.data
                p.data.add_(-group['lr'], d_p)
        return loss

TODO

Instantiate, gradcheck and train the model