crYOLO is a deep-learning based particle picking software.
The crYOLO tutorial is available at: http://sphire.mpg.de/wiki/doku.php?id=pipeline:window:cryolo
Loading the environment
CrYOLO is installed using the CSSB anaconda.
It can be used in 2 ways:
- GPU-based: train and predict
- CPU-based: mainly for predicting since the training takes too long on CPUs (version 1.3, can be installed on request)
To use it at CSSB/DESY call:
and copy the CSSB templates as printed on the screen.
Replace ?:? with the actual CrYOLO version printed by the module show command or the version you would like to use.
CSSB Beginner's Notes for crYOLO >= 1.5
Since crYOLO 1.5 the software comes with a graphical user interface for generating the configuration files but also for running the training and prediction on-the-fly on the local node. The official crYOLO tutorial helps you in using this. Since crYOLO 1.4 it is recommended to the PhosaurusNet instead of the YOLO algorithm in crYOLO. Start it now with:
Sometimes it is recommended to use the JANNI denoising software. The settings can be seen in the screenshot below.
Use the pre-trained JANNI model from the following CSSB location and save the config file.
Instead of training your own model you can try a generic pre-trained model.
So you just have to provide a motion corrected micrograph folder like below:
The pretrained models are at the following CSSB location:
For JANNI use the *_nn_*.h5 file.
If you want to submit the job to the cluster an updated submission template can be found as follows:
Just copy & paste the GUI command using the Print cmd button into it and submit it to the cluster using the sbatch command.
CSSB Beginner's Notes for crYOLO <= v1.4
This section describes shortly the CSSB specific setup and an example of the integration with Relion. The directory naming convention follows the crYOLO tutorial. Please read the tutorial first.
Import the micrographs
crYOLO needs motion-corrected single frame micrographs as input. The Relion 3.x naming conventions can be confusing (e.g. Movies directories instead of the previous Micrographs)!
Copy all micrographs as symbolic links from the original location to the full_data folder (crYOLO v1.0, v1.1).
crYOLO v1.2+ reads only the files from the full_data directory and pre-processes them to a temporary tmp_filtered directory. So just linking the original directory is enough.
Prepare the training set
For training its recommended to use 10 micrographs or about 100 particles.
This statement copies the first 10 files in the full_data directory to the local train_image directory. You can also use a graphical file manager as alternative.
Pick particles in the training set
You can start picking directly (don't forget to load the environment):
Open the files in the folder train_image, choose your boxsize and save the picked coordinates as box files in the folder train_annotation.
Here is how to pick particles (from the tutorial):
LEFT MOUSE BUTTON: Place a box
HOLD LEFT MOUSE BUTTON: Move a box
CONTROL + LEFT MOUSE BUTTON: Remove a box
If you want to continue, open the files in the folder train_image and import the box files by choosing the folder train_annotation.
If you have already picked coordinates, copy them into the folder train_annotation and do as described in continue above.
Denoising your training data set
If you hardly see your particles for handpicking do the following in the activated crYOLO conda environment on a GPU node (crYOLO v1.4+):
Then start the cryolo_boxmanager.py as described above and use as input micrograph folder janni_denoised/train_image.
Please note that the JANNI micrographs should only be used for handpicking, for all other steps use the original micrographs.
Train and build a model file
Prepare the crYOLO configuration file and the CSSB submission script as printed in:
Edit the provided submission script and uncomment all lines with #cryolo_train.py by removing the # from the beginning of the line.
Comment all cryolo_predict.py by adding a # sign at the beginning of the line.
Save and submit the script to the cluster:
Once its done the trained model is in the model_yolo.h5 file. This you could save with the corresponding config_yolo.json configuration file for future prediction without training again.
Pick the full dataset with the trained model
Edit the provided submission script and uncomment one line with #cryolo_predict.py by removing the # from the beginning of the line. The various examples are with different thresholds so if the result is not satisfying rerun the picking with a different one.
Comment all lines with cryolo_train.py by adding a # sign at the beginning of the line.
Save and submit the script to the cluster:
Once the job is finished you could check the amount of picked particles by:
View the picked coordinates
You can use the crYOLO boxmanager to view the results. Use Open image folder: full_data and Import box files: predict_yolo/EMAN as input.
Copy the picked coordinates to your project folder
To import the coordinates in Relion you need to copy the files to your original Movies folder (not the MotionCorr/job002/Movies folder!):
Now use the Relion import function pattern: Movies/*.star, set your boxsize and extract the particles again.