v0.3.0

pixtreme-train v0.3.0

Training from a YAML configuration:

• register and lookup select networks, losses, training steps, optimizers, schedulers and datasets by name. The default step, the adam, adamw and sgd optimizers, the warmup_cosine scheduler, the l1, mse and charbonnier losses, and the image_degradation and image_evaluation datasets are registered.
• TrainingStep is the base class for one update; a trainer overrides step and, when needed, the optimizer construction, state, configuration checks and evaluation. DefaultStep trains one main network with weighted losses, a scheduler and optional EMA.
• With train.precision: bf16, only the network forward runs under bfloat16 autocast on CUDA; the loss, weights, gradients, optimizer state, EMA weights, evaluation and checkpoints stay float32.
• train runs the training loop: resume from a checkpoint, checkpoint and evaluation intervals, one JSON line per update, and SIGINT / SIGTERM handling that saves a checkpoint before exiting.
• The pixtreme-train train CONFIG --output DIR [--resume CHECKPOINT] command runs train and prints the path of the last checkpoint.
• load_network rebuilds a trained network from a checkpoint, with its EMA or training weights.

Image data and on-the-fly LQ generation:

• DegradationPipeline and load_pipeline describe named, seeded sequences of resize and encode steps (jpeg, prores_lt, h264, hevc) with fixed values, ranges or candidates.
• ImageDegradationDataset crops aligned GT / LQ patches with randomness derived from the sample number, so samples do not depend on DataLoader workers; NumberedSampler and sample_generator provide those numbers and streams. augment_images flips and transposes paired images.
• ImageEvaluationDataset degrades full evaluation images; image_paths selects image files.

The prores_lt, h264 and hevc steps, and reading 16-bit RGB images, need an ffmpeg command on PATH. Importing pixtreme.train.common still does not load torch.

The wheel attached here is the reference copy for the PyPI distribution of pixtreme-train. Verify it with sha256sum -c SHA256SUMS.

v0.2.0

pixtreme-train v0.2.0

New training components in pixtreme.train.common:

• write_text and staging_directory replace a file or a directory only after the new content is fully written.
• training_signals turns SIGINT and SIGTERM into a stop request for a training loop, with defer() for sections that must finish.
• worker_signals and stop_loader_workers handle signals in DataLoader worker processes and stop and reap the workers of an iterator without waiting for prefetched batches.
• rng_state and restore_rng_state save and restore the Python, NumPy, torch CPU and CUDA, and named torch.Generator random states.
• warmup_cosine_factor, update_ema, check_finite_loss, check_finite_gradients, release_gpu_memory and EndlessSampler provide small training parts.

Importing pixtreme.train.common still does not load torch; each torch-based name loads its module on first access.

The wheel attached here is the reference copy for the PyPI distribution of pixtreme-train. Verify it with sha256sum -c SHA256SUMS.

v0.1.2

pixtreme-train v0.1.2

• MaterialMetrics and EvaluationMetrics with no frames now raise ValueError from psnr_mean, ssim_mean and to_dict. For a material, the message names its path. The means stay float.

The wheel attached here is the reference copy for the PyPI distribution of pixtreme-train. Verify it with sha256sum -c SHA256SUMS.

v0.1.1

pixtreme-train v0.1.1

• Importing pixtreme.train.common or pixtreme.train.common.configuration no longer loads torch. The metrics and checkpoint components are loaded when one of their names is first used. All public names stay the same.

The wheel attached here is the reference copy for the PyPI distribution of pixtreme-train. Verify it with sha256sum -c SHA256SUMS.

v0.1.0

pixtreme-train v0.1.0: the first release of pixtreme.train.common, the shared training components for pixtreme model trainers.

• Option values and reusable YAML configuration (Option, plain, write_yaml)
• PSNR and SSIM measurement with pixtreme and frame-weighted aggregation (rgb_frame, measure, FrameMetrics, MaterialMetrics, EvaluationMetrics, json_metric)
• Checkpoint saving that keeps the previous file when writing fails (save_checkpoint)

The wheel attached here is the reference copy for the PyPI distribution of pixtreme-train. Verify it with sha256sum -c SHA256SUMS.