Step	Problem	Possible reason	Possible solution
4	Out of memory error when running cryodrgn downsample	The particle stack is too large to fit into memory	Add the --chunk flag to the cryodrgn downsample command
5	Back projection is noisy, discontinuous, or does not resemble consensus refinement	Incorrect pose and CTF metadata were supplied, or pose and CTF parameters were incorrectly mapped to particles. Noisy maps may also result from using a small number of particles used in the back projection (default: 10,000). Users may also have not applied the correct --uninvert-data convention, which determines whether the data is light-on-dark or dark-on-light	Verify that the correct pose and CTF parameters were supplied during parsing and that the particle stack originated from, and contains the same particle index/order as the pose and CTF parameter metadata. If the volume is very noisy, re-run cryodrgn backproject_voxel with a larger number of particles using --first flag. Check whether the correct --uninvert-data convention for the dataset is followed by running cryodrgn backproject_voxel with and without --uninvert-data
6, 18	Out of memory error shortly after starting cryodrgn train_vae	The particle stack is too large to preload into memory	Append --lazy to the cryodrgn train_vae command to allow on-the-fly image loading, further downsample particles, or train on a subset of the particle stack.
6, 18	CUDA out of memory error during cryodrgn train_vae	Batch size may be set too large for your GPU’s memory capacity	Manually decrease batchsize with the --batchsize flag in the cryodrgn train_vae command
6, 18	Assertion error during cryodrgn train_vae similar to assert (coords[…,0:3].abs() - 0.5 < 1e-4).all()	Infrequent issues with numerical instability using --amp may cause this assertion to fail	Restart cryodrgn train_vae without --amp
7, 9, 19, 22–23	Volumes generated after training appear non-continuous or hollow in the center of the box	Users failed to apply correct --uninvert-data flag	Run cryoDRGN backproject_voxel with and without the --uninvert-data flag and determine which convention is applicable, then re-run cryodrgn train_vae as necessary
7, 9, 19, 22–23	Volumes generated after training all resemble junk	Volumes may be displayed at too permissive an isosurface. Alternatively, data may have been parsed incorrectly in the preprocessing steps.	Increase the isosurface threshold for display. Run cryodrgn backproject_voxel to determine whether poses and CTF parameters were correctly parsed.
7, 9, 19, 22–23	Volumes generated after training all appear homogeneous	For datasets other than EMPIAR-10076, this may be caused by too much upstream filtering prior to cryoDRGN training.	Restart cryoDRGN training with an unfiltered dataset.
All	Jupyter notebooks aren’t behaving as described in the protocol	Cells may have been run out of order or may reference outdated variables	Restart the kernel and run the notebook again in order from top to bottom