Building a Stock Advisor on a Coral Dev Board: From Edge TPU Bugs to Working TPU Inference

Short summary

Author built a stock prediction system on Google Coral Edge TPU achieving 2.5ms latency. They debugged three critical issues: silent scaler refitting at inference, BatchNormalization causing subgraph splits, and misread quantization scales. Since the Edge TPU's operation set is frozen at 2019, LSTM and BERT patterns fallback to CPU, requiring Conv1D-based architecture for full TPU utilization.

• •Built working stock prediction on Edge TPU with 2.5ms inference latency
• •Debugged silent scaler refitting, BatchNormalization subgraph splits, and quantization scale misreads
• •Designed Conv1D architecture to stay on-chip given 2019 TPU operation constraints

Generated with AI, which can make mistakes.