lleisong / hrlcm Goto Github PK

• Make labels.
• Evaluate the labels.

#4
If use multiple GPU devices to train a DL model, it is necessary to use synchronized batch normalization to speed up the training and keep images consistent.

• Based on available device numbers or user setting, let model use multiple devices.
• Include synchronized batch normalization into training. Here is a repo that can be a good reference: Synchronized-BatchNorm-PyTorch

Just a reminder for myself:
A normalize transform is in need to reduce the time of coverage of DL algorithm, and meanwhile reduce the edge effects of tiles.

Steps:

• Add normalize transform.
• Calculate the "global" means and sds for all bands.

Train a DL model based on pixel-based labels.

Reference to paper "Smallholder maize area and yield mapping at national scales with Google Earth Engine" to gather an independent evaluation reference dataset.