Model Zoo Downloader

A command-line tool for downloading checkpoints and metadata developed by the AIML Lab.

This package provides a simple CLI to fetch artifacts from the Core, EuroSAT, and ViT model zoos.

Installation

It is recommended to install the package inside a virtual environment:

python3 -m venv .venv
source .venv/bin/activate   # On Windows: .venv\Scripts\activate
pip install wsl-modelzoo

This tool currently also requires the wget executable to be available on your system.

• macOS: brew install wget
• Ubuntu/Debian: sudo apt install wget

Model Zoos

Model Zoo	Available	Paper
Core	✅	Link
EuroSAT	✅	Link
ViT	✅	Link
Phase Transitions	❌	Link

Download directory

The CLI supports a --dir argument that lets you specify the root directory where checkpoints and metadata will be downloaded.

If --dir is not provided, the tool creates a downloads directory in the package root and stores all downloaded files there.

Core Model Zoo

Arguments

Argument	Required	Description
--zoo	✅	Select the model zoo (core)
--arch	✅	Model architecture
--dataset	✅	Dataset
--config	✅	Configuration
--seed	❌	Model seed(s)
--ckpts	❌	Checkpoints indices to fetch
--dropout	❌	Dropout
--init	❌	Initialization Type
--activation	❌	Activation Function
--lr	❌	Learning Rate
--optimizer	❌	Optimizer
--wd	❌	Weight Decay

Core Model Zoo: valid combinations of architectures and datasets

Datasets marked with * also have a sparsified variant, for example mnist_sparsified.

Architecture	Dataset
cnn-small	mnist*, fmnist*, svhn*, usps*, cifar10*, stl10*
cnn-large	cifar10*, stl10*
resnet18	cifar10, cifar100, tinyimagenet

Valid config combinations

Architecture	Dataset	Supported configs
cnn-small	mnist*, fmnist*, svhn*, usps, cifar10, stl10	uniform, fixed, random
cnn-small	usps_sparsified, cifar10_sparsified, stl10_sparsified	uniform
cnn-large	cifar10*, stl10*	uniform, fixed, random
resnet18	cifar10, cifar100, tinyimagenet	uniform

Extra hyperparameter filters

These filters only apply to selected cnn-small and cnn-large combinations with fixed or random configs.

Currently, extra filters accept only one value per argument unlike --seed and --ckpts (see below)

Architecture	Dataset	Config	Supported extra filters
cnn-small	mnist, fmnist, svhn usps, cifar10, stl10	fixed, random	--dropout, --init, --activation, --lr, --optimizer, --wd
cnn-large	cifar10, stl10	fixed, random	--dropout, --init, --activation, --lr, --optimizer, --wd

Allowed values for extra hyperparameter filters

Filter	Applies to	Allowed values
--init	all supported extra-filter combinations	uniform, normal, kaiming_uniform, kaiming_normal
--activation	all supported extra-filter combinations	tanh, relu, gelu, sigmoid
--optimizer	most supported extra-filter combinations	adam, sgd
--optimizer	cnn-small + svhn + (fixed or random)	sgd
--dropout	most supported extra-filter combinations	0, 0.5
--dropout	cnn-small + svhn + (fixed or random)	0, 0.3, 0.5
--lr	most supported extra-filter combinations	0.001, 0.0001
--lr	cnn-small + svhn + (fixed or random)	0.1, 0.001, 0.0001
--lr	cnn-small + cifar10 + random	0.001
--lr	cnn-large + cifar10 + (fixed or random)	0.001
--wd	most supported extra-filter combinations	0.01, 0.001
--wd	cnn-small + svhn + (fixed or random)	0, 0.001, 0.1

--seed cannot be used with fixed or random configs.

Seed and checkpoint ranges

--seed and --ckpts accept one or more values within their allowed ranges.

You can pass:

• individual values, for example 1 10 25
• inclusive ranges, for example 25-30
• the special value all to download every available seed or checkpoint

By default, all the available seeds and checkpoints are retrieved if the user skips one of the two arguments above.

Parameter	Range (inclusive)
--seed	1..1000

Architecture / dataset	Checkpoint range (inclusive)
cnn-small + mnist	0..50
cnn-small + mnist_sparsified	0..25
cnn-small + fmnist	0..50
cnn-small + fmnist_sparsified	0..25
cnn-small + svhn	0..50
cnn-small + svhn_sparsified	0..25
cnn-small + usps	0..50
cnn-small + usps_sparsified	0..25
cnn-small + cifar10	0..50
cnn-small + cifar10_sparsified	0..25
cnn-small + stl10	0..50
cnn-small + stl10_sparsified	0..25
cnn-large + cifar10	0..50
cnn-large + cifar10_sparsified	0..25
cnn-large + stl10	0..50
cnn-large + stl10_sparsified	0..25
resnet18 + cifar10	0..50
resnet18 + cifar100	0..60
resnet18 + tinyimagenet	0..60

Examples

modelzoo fetch --zoo core --arch cnn-small --dataset mnist --config uniform --seed all --ckpts 0 10 20 30-45 --dir /your/destination/path

modelzoo fetch --zoo core --arch cnn-small --dataset mnist --config fixed --init kaiming_uniform --dropout 0.5 --activation relu --lr 0.001 --optim adam --wd 0.001 --ckpts 0 10 20

modelzoo fetch --zoo core --arch cnn-small --dataset mnist_sparsified --config uniform --ckpts 0 5 10

modelzoo fetch --zoo core --arch cnn-large --dataset stl10 --config fixed --init kaiming_normal --dropout 0.5 --activation tanh --lr 0.0001 --optim adam --wd 0.001 --ckpts 0 10 20

modelzoo fetch --zoo core --arch resnet18 --dataset cifar100 --config uniform --seed 42 --ckpts all

EuroSAT Model Zoo

Arguments

Argument	Required	Description
--zoo	✅	Select the model zoo (eurosat)
--arch	✅	Model architecture
--dataset	✅	Dataset
--config	❌	Sparsified configuration
--seed	❌	Model seed(s)
--ckpts	❌	Checkpoint indices to fetch

--config is required when --dataset eurosat_sparsified is used.
--seed is not available when --dataset eurosat_sparsified is used. All the available seeds will be automatically downloaded

Valid architecture / dataset combinations

Architecture	Supported datasets
cnn-small	eurosat, eurosat_sparsified

Valid config combinations

Currently, --config accepts only a single value

Dataset	Supported configs
eurosat	no extra config required
eurosat_sparsified	ard_vd, magn_10, magn_20, magn_30, magn_40, magn_50, magn_60, magn_70, magn_80, magn_90

Seed and checkpoint ranges

--seed and --ckpts accept one or more values within their allowed ranges.

You can pass:

• individual values, for example 1 10 25
• ranges, for example 25-30
• the special value all to download every available seed or checkpoint

By default, all the available seeds and checkpoints are retrieved if the user skips one of the two arguments above.

Parameter	Range (inclusive)
--seed	1..1000

Dataset / config	Checkpoint range (inclusive)
eurosat	0..50
eurosat_sparsified + ard_vd	0..25
eurosat_sparsified + magn_10	0..15
eurosat_sparsified + magn_20	0..15
eurosat_sparsified + magn_30	0..15
eurosat_sparsified + magn_40	0..15
eurosat_sparsified + magn_50	0..15
eurosat_sparsified + magn_60	0..15
eurosat_sparsified + magn_70	0..15
eurosat_sparsified + magn_80	0..15
eurosat_sparsified + magn_90	0..15

Examples

modelzoo fetch --zoo eurosat --arch cnn-small --dataset eurosat --seed 1 --ckpts 0 5 10 --dir /your/destination/path

modelzoo fetch --zoo eurosat --arch cnn-small --dataset eurosat_sparsified --config ard_vd --ckpts 0-10

modelzoo fetch --zoo eurosat --arch cnn-small --dataset eurosat_sparsified --config magn_50 --ckpts all

ViT Model Zoo

The ViT model zoo supports:

• pretrained models
• pretrained + corresponding finetuned models

To download only pretrained models, provide only the pretraining arguments and omit all finetuning arguments.

It is not possible to download only finetuned models. Finetuned models are always selected starting from a matching pretrained model.

Arguments

Pretraining arguments

Argument	Required	Description
--zoo	✅	Select the model zoo (vit)
--dataset_pt	✅	Pretraining dataset
--config	✅	Pretraining regime
--seed_pt	❌	Pretraining seed(s)
--ckpts_pt	❌	Pretraining checkpoint indices
--ddp	❌	Use the DDP supervised-pretraining variant (true or false)

Finetuning arguments

Argument	Required	Description
--dataset_ft	❌	Finetuning dataset
--seed_ft	❌	Finetuning seed(s)
--lr_ft	❌	Finetuning learning rate
--optim_ft	❌	Finetuning optimizer
--hd_ft	❌	Finetuning head hidden dimension
--fcmlp_ft	❌	Whether the finetuning head uses an MLP
--ckpts_ft	❌	Finetuning checkpoint indices

Valid pretraining combinations

Argument	Allowed values
--dataset_pt	imagenet
--config	sl, cl
--ddp	true, false (default)

Valid pretraining seeds

Pretraining mode	Allowed seeds
sl + --ddp false	42, 6312
sl + --ddp true	1957, 2517, 3753, 6312, 7853
cl	42, 2517, 6312

Pretraining checkpoint range

Parameter	Range (inclusive)
--ckpts_pt	0..89

• If --seed_pt is omitted, all available pretraining seeds are used.
• If --ckpts_pt is omitted, all available pretraining checkpoints are used.

You can pass:

• individual values, for example 1 10 25
• ranges, for example 25-30
• the special value all to download every available seed or checkpoint

By default, all the available seeds and checkpoints are retrieved if the user skips one of the two arguments above.

Valid finetuning combinations

Finetuning is only available on top of a selected pretrained model.

Argument	Allowed values
--dataset_ft	cifar100
--seed_ft	42, 6312
--lr_ft	0.003, 0.0001, 0.001
--optim_ft	sgd, adamw
--hd_ft	256, none
--fcmlp_ft	true, false

Finetuning checkpoint range

Parameter	Range (inclusive)
--ckpts_ft	0..49

Important usage notes

Use case	What to do
Download only pretrained models	Provide only --dataset_pt, --config, and optional pretraining arguments such as --seed_pt, --ckpts_pt, --ddp
Download pretrained and finetuned models	Provide the pretraining arguments and the finetuning arguments
Download only finetuned models	Not supported

Examples

modelzoo fetch --zoo vit --dataset_pt imagenet --config sl --ddp false --seed_pt 42 --ckpts_pt 0 10 20 --dir /your/destination/path

modelzoo fetch --zoo vit --dataset_pt imagenet --config sl --ddp true --seed_pt 1957 --ckpts_pt 0 10

modelzoo fetch --zoo vit --dataset_pt imagenet --config cl --ddp false --seed_pt 2517 --ckpts_pt 0 20

modelzoo fetch --zoo vit --dataset_pt imagenet --config sl --ddp false --seed_pt 42 --ckpts_pt 0 10 --dataset_ft cifar100 --seed_ft 42 --lr_ft 0.001 --optim_ft adamw --hd_ft 256 --fcmlp_ft true --ckpts_ft 0 10

modelzoo fetch --zoo vit --dataset_pt imagenet --config cl --ddp false --seed_pt 6312 --ckpts_pt 0 20 --dataset_ft cifar100 --seed_ft 6312 --lr_ft 0.003 --optim_ft sgd --hd_ft none --fcmlp_ft false --ckpts_ft 0 5 10

Get in touch!

If there are any questions, reach out to us! We'll respond to issues here, or via email salvatore.pascarella@unisg.ch, leo.meynent@unisg.ch.

License

This project is licensed under the MIT License — see the LICENSE file for details.