Train all aligners
MetroEM is based on modelhouse, artificery, and scalenet packages. It's useful to understand these packages before proceeding.
Create a directory, with subdirectories for each module.
It's easiest to adapt an existing model.
For example, use modelhouse
to download the following model:
modelhouse load gs://corgie/models/pyramid_m4m6m9
Module subdirectories must follow the name convention
<STAGE>_mip<IN_MIP>in_<SUFFIX>
such that:
STAGE
(int): processing order of module (lowest-to-highest, 0-indexed)IN_MIP
(int): mip level of the images to be used as inputSUFFIX
(str): string for description only
Each module subdirectory must follow modelhouse
directory conventions.
Training sets consists of an image dataset as well as a field dataset.
An image dataset is a set of image pairs (source and target images), one set of image pairs for the MIP level required as input for each module.
A field dataset is a displacement field computed by a previous module, one field for each image in an image set.
Image datasets may be created from CloudVolumes.
For models with multiple modules, the image set of each module
must be aligned in physical space across the span of MIP levels.
Images should be normalized (mean=0, std=1) over a larger field
of view than the model.
See corgie
for a method to normalize CloudVolumes.
To create a dataset from a CloudVolume, use download_image.py
.
For example,
python download_image.py \
--dst_folder ~/data \
--z_start 8175 \
--z_end 8180 \
--x_offset 200960 \
--y_offset 136448 \
--mip 3 5 \
--patch_size 1536 1536 \
--cv_path <CloudVolume path>
Field datasets will be automatically created during the training procedure.
There may be times when a user would like to create a field dataset from
an existing field CloudVolume.
To do so, use download_field.py
.
For example:
python download_field.py \
--dst_folder ~/data/ \
--z_start 8175 \
--z_end 8180 \
--x_offset 200960 \
--y_offset 136448 \
--src_mip 8 \
--dst_mip 5 \
--patch_size 1536 \
--suffix precomputed \
--cv_path <CloudVolume path>
Once downloaded, a field dataset must be used as input when creating a new image dataset. For example:
python download_image.py \
--dst_folder ~/data \
--z_start 8175 \
--z_end 8180 \
--x_offset 200960 \
--y_offset 136448 \
--mip 5 \
--patch_size 1536 \
--cv_path <CloudVolume path> \
--field_dset ~/data/field_0_x200960_y136448_z8175_MIP5_precomputed.h5 \
Each downloaded field is adjusted by its average translation, and the bounding box of the associated image must also be adjusted by the translation.
With a model & dataset in place, it's time to train the model with train.py
.
For example:
python train.py \
--pyramid_path ~/model/m4m6m8 \
--dataset_path ~/data \
--train_mode scratch \
--train_stages 1 \
--generate_field_stages 1 \
--checkpoint_name precomputed \
--no_redirect_stdout