CLI Reference¶

Command-line tools for training, evaluation, and benchmarking.

Training¶

The unified tinify CLI provides a standardized interface for training all compression models.

Basic Usage¶

# Train with config file
tinify train image --config configs/mbt2018-mean.yaml

# Train with CLI arguments
tinify train image -m mbt2018-mean -d /path/to/dataset --epochs 100

# Train video model
tinify train video --config configs/ssf2020-video.yaml

# List available models
tinify train list-models --domain image

Training Commands¶

Command	Description
`tinify train image`	Train image compression model
`tinify train video`	Train video compression model
`tinify train pointcloud`	Train point cloud compression model
`tinify train list-models`	List available models

Training Arguments¶

Argument	Description
`-c, --config`	Path to config file (YAML/JSON/TOML)
`-m, --model`	Model architecture name
`-q, --quality`	Quality level (1-8)
`-d, --dataset`	Path to training dataset
`-e, --epochs`	Number of training epochs
`--batch-size`	Training batch size
`--lambda`	Rate-distortion trade-off parameter
`-lr, --learning-rate`	Learning rate
`--cuda` / `--no-cuda`	Enable/disable CUDA
`--checkpoint`	Resume from checkpoint
`--save-dir`	Directory to save checkpoints
`--seed`	Random seed

Config File Format¶

# configs/mbt2018-mean.yaml
domain: image

model:
  name: mbt2018-mean
  quality: 3

dataset:
  path: /path/to/vimeo90k
  patch_size: [256, 256]
  num_workers: 4

training:
  epochs: 300
  batch_size: 16
  lmbda: 0.0067
  metric: mse  # or ms-ssim
  cuda: true
  save_dir: ./checkpoints

optimizer:
  net: {type: Adam, lr: 0.0001}
  aux: {type: Adam, lr: 0.001}

scheduler:
  type: ReduceLROnPlateau
  patience: 20

Available Models¶

Image Compression: - bmshj2018-factorized - Factorized prior (Ballé 2018) - bmshj2018-hyperprior - Scale hyperprior (Ballé 2018) - mbt2018-mean - Mean-scale hyperprior (Minnen 2018) - mbt2018 - Joint autoregressive (Minnen 2018) - cheng2020-anchor - Attention-based (Cheng 2020) - cheng2020-attn - Attention with GMM (Cheng 2020) - *-vbr - Variable bitrate variants

Video Compression: - ssf2020 - Scale-space flow (Agustsson 2020)

Training Examples¶

# Train mbt2018-mean on Vimeo90K
tinify train image \
    -m mbt2018-mean \
    -d /data/vimeo90k \
    --epochs 300 \
    --batch-size 16 \
    --lambda 0.0067 \
    --cuda

# Resume training from checkpoint
tinify train image \
    --config configs/mbt2018-mean.yaml \
    --checkpoint checkpoints/checkpoint.pth.tar

# Train with MS-SSIM metric
tinify train image \
    -m mbt2018-mean \
    -d /data/vimeo90k \
    --lambda 8.73 \
    --config configs/mbt2018-mean.yaml

Model Evaluation¶

Evaluate Trained Checkpoints¶

python -m tinify.utils.eval_model checkpoint /path/to/images/ \
    -a ARCHITECTURE \
    -p /path/to/checkpoint.pth.tar

Evaluate Pre-trained Models¶

python -m tinify.utils.eval_model pretrained /path/to/images/ \
    -a ARCHITECTURE \
    -q QUALITY_LEVELS

Arguments:

Argument	Description
`-a, --architecture`	Model architecture name
`-p, --path`	Path to checkpoint file
`-q, --quality`	Quality level(s) to evaluate
`--cuda`	Use GPU acceleration
`--half`	Use half precision (FP16)

Example:

python -m tinify.utils.eval_model pretrained /path/to/kodak/ \
    -a mbt2018-mean -q 1 2 3 4 5 6 7 8 --cuda

Codec Benchmarking¶

BPG Codec¶

python -m tinify.utils.bench bpg /path/to/images/ [OPTIONS]

VTM (VVC Reference)¶

python -m tinify.utils.bench vtm /path/to/images/ [OPTIONS]

General Options¶

python -m tinify.utils.bench --help

Plotting Results¶

Generate rate-distortion plots from evaluation results:

python -m tinify.utils.plot results.json [OPTIONS]

Options:

Option	Description
`--show`	Display plot interactively
`--output`	Save plot to file
`--metric`	Metric to plot (psnr, ms-ssim)

Video Compression¶

Evaluate Video Models¶

# Trained checkpoint
python -m tinify.utils.video.eval_model checkpoint /path/to/videos/ \
    -a ssf2020 -p /path/to/checkpoint.pth.tar

# Pre-trained model
python -m tinify.utils.video.eval_model pretrained /path/to/videos/ \
    -a ssf2020 -q QUALITY_LEVELS

Video Codec Benchmarks¶

# x265/HEVC
python -m tinify.utils.video.bench x265 /path/to/videos/

# VTM (VVC)
python -m tinify.utils.video.bench VTM /path/to/videos/

Video Plot¶

python -m tinify.utils.video.plot results.json --show

Model Update¶

Update entropy bottleneck parameters after training:

python -m tinify.utils.update_model \
    --architecture ARCHITECTURE \
    checkpoint.pth.tar

This updates the learned CDFs required for entropy coding.

Example Workflow¶

Train a model:

python examples/train.py -d /path/to/dataset/ \
    -a mbt2018-mean --epochs 300 --cuda --save

Update the model:

python -m tinify.utils.update_model \
    --architecture mbt2018-mean \
    checkpoint_best_loss.pth.tar

Evaluate:

python -m tinify.utils.eval_model checkpoint /path/to/kodak/ \
    -a mbt2018-mean -p checkpoint_best_loss.pth.tar --cuda

Plot results:

python -m tinify.utils.plot results.json --show