ATC-ASR Spectrogram Classifier

A from-scratch implementation of a Convolutional Neural Network (CNN) for classifying Air Traffic Control (ATC) audio. Bypasses high-level frameworks to implement the underlying linear algebra and calculus. Optimize with GPU kernels.

Architecture & Design

Input: (N, 1, 64, 400)

Block 1 Conv2D: 1 → 16, 3×3, stride 1, pad 1 → (N, 16, 64, 400) ReLU MaxPool2D: 2×2 → (N, 16, 32, 200)

Block 2 Conv2D: 16 → 32, 3×3, pad 1 → (N, 32, 32, 200) ReLU MaxPool2D: 2×2 → (N, 32, 16, 100)

Block 3 Conv2D: 32 → 64, 3×3, pad 1 → (N, 64, 16, 100) ReLU MaxPool2D: 2×2 → (N, 64, 8, 50)

Classifier Flatten → (N, 64 × 8 × 50 = 25600) Dense: 25600 → 128, ReLU Dense: 128 → 8 Softmax

Optimization Roadmap

Phase 1: Mathematical Optimization

• im2col: Lowering convolutions to dense matrix multiplications (GEMM) to improve cache locality.
• https://ieeexplore.ieee.org/document/9342343
• Quantization: Reducing precision to INT8 for increased arithmetic throughput.

Phase 2: CUDA Implementation (A100 Target)

• Memory Management: Using constant memory for kernels and shared memory for tiling.
• Precision: Implementing __half2 (FP16) to leverage Tensor Core acceleration.
• Concurrency: Overlapping H2D/D2H transfers with compute using CUDA streams.
• Profiling: Identifying compute vs. memory bottlenecks using nsys and ncu.

https://huggingface.co/datasets/jacktol/ATC-ASR-Dataset