Debug Julia Flux Satellite Inference
Budget: $10 – $30 USD
I am working on cloud-free reconstructions of GOES-R satellite imagery, following the methodology outlined in https://doi.org/10.5194/os-20-1567-2024 and the companion package at https://github.com/gher-uliege/DINDiff.jl. My inference pipeline is written entirely in Julia and built on Flux.jl. Training behaves as expected, but when I load the pretrained weights and run inference the predicted values fall far outside the physical data range, so the resulting images are unusable.
You will receive:
• the current inference script,
• the pretrained model weights,
• a small sample of GOES-R test data, and
• the original model definition.
Your task is to trace and fix whatever is causing the incorrect predictions—whether that is due to mishandled normalisation, wrongly ordered dimensions, missed RNG state, or something else—so that the output fields match the expected radiometric range shown in the paper. Please return a patched, well-commented script that reproduces correct results on my sample set, along with a short note summarising the changes and why they work.
You will receive:
• the current inference script,
• the pretrained model weights,
• a small sample of GOES-R test data, and
• the original model definition.
Your task is to trace and fix whatever is causing the incorrect predictions—whether that is due to mishandled normalisation, wrongly ordered dimensions, missed RNG state, or something else—so that the output fields match the expected radiometric range shown in the paper. Please return a patched, well-commented script that reproduces correct results on my sample set, along with a short note summarising the changes and why they work.