Skip to content

Releases: didriknielsen/survae_flows

Pretrained models

22 Sep 16:24
@didriknielsen didriknielsen
v1.0.0
5d81a8e
This commit was created on GitHub.com and signed with GitHub’s verified signature. The key has expired.
GPG key ID: 4AEE18F83AFDEB23
Expired
Learn about vigilant mode.
Compare
Choose a tag to compare
Pretrained models Latest
Latest

The checkpoints contain model weights, optimizer state, etc.
For details, see the code for checkpoint saving and checkpoint loading.

Usage

  1. Extract checkpoint(s) in YOUR_PATH/survae_flows/experiments/image/log/.
  2. Sample from the models, evaluate the test log-likelihood or continue training using these scripts.

Sampling:

To sample from the models, use the eval_sample.py script.

CIFAR-10:

python eval_sample.py --model YOUR_PATH/survae_flows/experiments/image/log/cifar10_8bit/pool_flow/more/maxpool
python eval_sample.py --model YOUR_PATH/survae_flows/experiments/image/log/cifar10_8bit/pool_flow/more/nonpool

ImageNet 32x32:

python eval_sample.py --model YOUR_PATH/survae_flows/experiments/image/log/imagenet32_8bit/pool_flow/more/maxpool
python eval_sample.py --model YOUR_PATH/survae_flows/experiments/image/log/imagenet32_8bit/pool_flow/more/nonpool

ImageNet 64x64:

python eval_sample.py --model YOUR_PATH/survae_flows/experiments/image/log/imagenet64_8bit/pool_flow/more/maxpool
python eval_sample.py --model YOUR_PATH/survae_flows/experiments/image/log/imagenet64_8bit/pool_flow/more/nonpool

Log-likelihood:

To compute the test log-likelihood for the models, use the eval_loglik.py script.

CIFAR-10:

python eval_loglik.py --model YOUR_PATH/survae_flows/experiments/image/log/cifar10_8bit/pool_flow/more/maxpool --k 1000 --kbs 10
python eval_loglik.py --model YOUR_PATH/survae_flows/experiments/image/log/cifar10_8bit/pool_flow/more/nonpool --k 1000 --kbs 10

ImageNet 32x32:

python eval_loglik.py --model YOUR_PATH/survae_flows/experiments/image/log/imagenet32_8bit/pool_flow/more/maxpool
python eval_loglik.py --model YOUR_PATH/survae_flows/experiments/image/log/imagenet32_8bit/pool_flow/more/nonpool

ImageNet 64x64:

python eval_loglik.py --model YOUR_PATH/survae_flows/experiments/image/log/imagenet64_8bit/pool_flow/more/maxpool
python eval_loglik.py --model YOUR_PATH/survae_flows/experiments/image/log/imagenet64_8bit/pool_flow/more/nonpool

Continue Training:

To continue training for some additional epochs using a new, fixed learning rate, use the train_more.py script.

CIFAR-10:

python train_more.py --model YOUR_PATH/survae_flows/experiments/image/log/cifar10_8bit/pool_flow/more/maxpool --new_epochs 600 --new_lr 1e-5
python train_more.py --model YOUR_PATH/survae_flows/experiments/image/log/cifar10_8bit/pool_flow/more/nonpool --new_epochs 600 --new_lr 1e-5

ImageNet 32x32:

python train_more.py --model YOUR_PATH/survae_flows/experiments/image/log/imagenet32_8bit/pool_flow/more/maxpool --new_epochs 30 --new_lr 1e-5
python train_more.py --model YOUR_PATH/survae_flows/experiments/image/log/imagenet32_8bit/pool_flow/more/nonpool --new_epochs 30 --new_lr 1e-5

ImageNet 64x64:

python train_more.py --model YOUR_PATH/survae_flows/experiments/image/log/imagenet64_8bit/pool_flow/more/maxpool --new_epochs 25 --new_lr 1e-5
python train_more.py --model YOUR_PATH/survae_flows/experiments/image/log/imagenet64_8bit/pool_flow/more/nonpool --new_epochs 25 --new_lr 1e-5
Assets 8
Loading