Configure the number of image samples logged at each epoch and batch - #49

…49)

* log sample size at epoch and batch levels

* update example configs

* pixelshuffle decoder

* Allow sampling multiple patches from the same stack (#35)

* sample multiple patches from one stack

* do not use type annotations from future
it breaks jsonargparse

* fix channel stacking for non-training samples

* remove batch size from model
the metrics will be automatically reduced by lightning

* add flop counting script

* 3d ouput head

* add datamodule target dims mode

* remove unused argument and configure drop path

* move architecture argument to model level

* DLMBL 2023 excercise (#36)

* updated intro and paths

* updated figures, tested data loader

* setup.sh fetches correct dataset

* finalized the exercise outline

* semi-final exercise

* parts 1 and 2 tested, part 3 outline ready

* clearer variables, train with larger patch size

* fix typo

* clarify variable names

* trying to log graph

* match example size with training

* reuse globals

* fix reference

* log sample images from the first batch

* wider model

* low LR solution

* fix path

* seed everything

* fix test dataset without masks

* metrics solution
this needs a new test dataset

* fetch test data, compute metrics

* byass cellpose import error due to numpy version conflicts

* final exercise

* moved files

* fixed formatting - ready for review

* viscy -> VisCy (#34) (#39)

Introducing capitalization to highlight vision and single-cell aspects of the pipeline.

* trying to log graph

* log graph

* black

---------

Co-authored-by: Shalin Mehta <shalin.mehta@gmail.com>
Co-authored-by: Shalin Mehta <shalin.mehta@czbiohub.org>

* fix channel dimension size for example input

#40

* fix argument linking

* 3D prediction writer
sliding windows are blended with uniform average

* update network diagram

* upgrade flop counting

* shallow 3D (2.5D) SSIM metric

* ms-ssim

* mixed loss

* fix arguments

* fix inheritance

* fix weight checking

* squeeze metric

* aggregate metrics

* optinal clamp to stabilize gradient of MS-SSIM

* fix calling

* increase epsilon

* disable autocast for loss

* shuffle validation data for logging

this hurts cache hit rate, but can avoid logging neighboring windows

* simplify decoder structure

* pop-head

* fix head expansion

* init conv weights

* update diagnostic scripts

* fix center slice metrics for 3D output (#51)

* Configure the number of  image samples logged at each epoch and batch (#49)

* log sample size at epoch and batch levels

* update example configs

* do not shuffle validation dataset

* fix upsampling weight initialization

* fix merge

* fix merge error

* fix formatting

---------

Co-authored-by: Shalin Mehta <shalin.mehta@gmail.com>
Co-authored-by: Shalin Mehta <shalin.mehta@czbiohub.org>