3D Denoising with Noise2Map

Overview

Noise2Map is a command-line tool included in Warp's installation directory for training a denoising neural network on volumetric data, based on the noise2noise principle. It requires adding the installation directory to your PATH for easier execution.

System Requirements

• At least two GPUs, or one GPU with a minimum of 16 GB memory.

Data Types Supported

• Half-map reconstructions from single particle data.
• Tomograms.

Usage

Denoising Half-Maps

• Data Setup: Place the first map of each pair in one folder (e.g., "odd"), and the second in another (e.g., "even"). Ensure file names are identical across folders.
• Execution: Use arguments like --observation1 and --observation2 to specify folder paths.
• Additional Parameters:
  • --mask: Path to a binary mask to help balance training samples.
  • --dont_flatten_spectrum: Disable amplitude spectrum flattening.
  • --overflatten_factor: Adjusts the degree of flattening or sharpening.
  • --lowpass and --angpix: Used to set filtering based on resolution and pixel size.
• Training: Typical training involves 600 iterations but can be adjusted based on GPU memory constraints.

Denoising Tomograms

• Data Preparation: Utilize Warp's UI to separate odd/even tilt images and potentially pre-deconvolve them.
• Key Differences in Parameters:
  • Do not use a mask.
  • Set --dont_flatten_spectrum.
  • Typically requires more than 10,000 iterations.

Output

• Denoised maps are saved in a "denoised" folder. Separate denoising of half-maps can be specified.

Note

• If reusing a model (--old_model), ensure that the filtering parameters match those used during its training to avoid invalidating the model.