prismpyp train

Purpose

Run SimSiam training on either the real-space inputs or the Fourier-space inputs.

When executed, this command creates an output subdirectory named checkpoints that contains:

• Best and last model weights
• Total loss function plot
• Collapse plot
• training_config.yaml — a record of all command-line arguments used during the run

Usage

usage: prismpyp train [-h]
                      [--metadata-path METADATA_PATH]
                      [--output-path OUTPUT_PATH]
                      [-a {resnet18,resnet34,resnet50}]
                      [--epochs N]
                      [-b N, --batch-size N]
                      [--lr LR]
                      [--resume PATH]
                      [--dist-url DIST_URL]
                      ...

Named Arguments

Commonly Changed

Argument	Description	Default
--output-path	Path to output directory	—
--metadata-path	Path to metadata file	—
-a, --arch	Model architecture. Options: resnet18, resnet34, resnet50	resnet50
--epochs N	Number of total epochs to run	100
-b N, --batch-size N	Mini-batch size (total across GPUs)	512
--lr LR, --learning-rate LR	Initial (base) learning rate	0.05
--resume PATH	Path to latest checkpoint	none
--dist-url DIST_URL	URL used to set up distributed training	tcp://localhost:10058

Less Commonly Changed

Argument	Description	Default
-j, --workers	Number of data loading workers	1
--momentum	Momentum of SGD solver	0.9
--wd W, --weight-decay W	Weight decay	1e-4
-p N, --print-freq N	Print frequency	10
--feature-extractor-weights PATH	Path to pre-trained feature extractor weights	none
--classifier-weights PATH	Path to pre-trained classifier weights	none
--world-size WORLD_SIZE	Number of nodes for distributed training	1
--rank RANK	Node rank for distributed training	0
--dist-backend DIST_BACKEND	Distributed backend	nccl
--seed SEED	Random seed	—
--gpu GPU	GPU ID to use	0
--multiprocessing-distributed	Use multi-processing distributed training	True
--dim DIM	Feature dimension	512
--pred-dim PRED_DIM	Hidden dimension of the predictor	256
--fix-pred-lr	Fix learning rate for predictor	True
--use-fft	Use FFT of the image as input	—
--downsample DOWNSAMPLE	Downsample the image	—
--pixel-size PIXEL_SIZE	Pixel size of the image	—
--size SIZE	Size of the image in pixels (before downsampling)	—
--conf-thresh CONF_THRESH	Confidence threshold for filtering	—
--add-datetime	Append datetime to output directory name	—
--evaluate	Evaluate model on validation set	True
--n-clusters N_CLUSTERS	Number of clusters for KMeans	—
--num-neighbors NUM_NEIGHBORS	Number of neighbors for UMAP	—
--min-dist-umap MIN_DIST_UMAP	Minimum distance for UMAP	0
--n-components N_COMPONENTS	Number of UMAP components	—
--nextpyp-preproc NEXTPYP_PREPROC	Path to nextPYP project pre-processing directory	—
--zip-images	Save zipped image thumbnails	—