Freelance data scientist · All case studies

Conv, pool, linear, batch-norm — without nn.*.

Custom forward passes for Conv2D, MaxPool2D, Linear, and BatchNorm2d. Vectorised. No python for-loops over batch or spatial axes. The constraint was the point.

Client: Deep learning internals. Built by Dilshad Raza.

The brief banned the usual modules. Forward passes had to match the math — im2col-style or broadcast convolution, pooling windows, affine maps, batch-norm statistics — and stay vectorised so a batch of images did not become a Python loop.

I implemented vectorised forwards for the four layers under those constraints — convolution and pooling as batched tensor rearrangements, linear as matmul, batch-norm from batch moments — and checked them against standard PyTorch modules.

Hire Dilshad Raza for similar freelance data science, machine learning, and AI automation in the UK, United States, and Australia.