Implementation:Danijar Dreamerv3 Logger And Report

Knowledge Sources	DreamerV3 elements library
Domains	Reinforcement_Learning, Monitoring
Last Updated	2026-02-15 09:00 GMT

Overview

Concrete tool for multi-backend metric logging and open-loop video prediction reporting provided by the DreamerV3 agent and elements library.

Description

The logging system consists of:

• make_logger() in dreamerv3/main.py — Factory that creates an elements.Logger with configured output backends (terminal, JSONL, TensorBoard, WandB, Expa, Scope)
• Agent.report() in dreamerv3/agent.py — Generates open-loop predictions by observing the first half of sequences through the RSSM, then imagining the second half and decoding to pixel space

The logger uses elements.Agg for metric aggregation with configurable reduction (sum, avg, max, stack) and elements.FPS for throughput tracking.

Usage

make_logger is called once during initialization. Agent.report() is called periodically during training to generate diagnostic videos. Metric logging happens at log_every intervals; reporting at report_every intervals.

Code Reference

Source Location

• Repository: dreamerv3
• File: dreamerv3/main.py (make_logger), dreamerv3/agent.py (Agent.report)
• Lines: dreamerv3/main.py L152-180, dreamerv3/agent.py L247-310

Signature

def make_logger(config):
    """
    Create a Logger with configured output backends.

    Args:
        config: elements.Config with config.logger.outputs (list of backend names),
                config.logger.filter (terminal filter), config.logger.fps (video FPS).

    Returns:
        elements.Logger: Logger with step counter and output backends.
    """

def Agent.report(self, carry, data):
    """
    Generate open-loop prediction report.

    Args:
        carry: Tuple of (enc_carry, dyn_carry, dec_carry, prevact).
        data: Dict of batched sequences from replay buffer.

    Returns:
        carry: Updated carry state.
        metrics: Dict containing 'openloop/{key}' video grids (uint8 arrays).
    """

Import

from dreamerv3.main import make_logger

I/O Contract

Inputs

Name	Type	Required	Description
config	elements.Config	Yes	Logger configuration with outputs list, filter, fps settings
carry	tuple	Yes	Agent carry state for report generation
data	dict	Yes	Batched sequences from replay for open-loop evaluation

Outputs

Name	Type	Description
logger	elements.Logger	Logger instance with add(metrics, prefix), write(), close() methods
metrics	dict	Report metrics including 'openloop/{key}' video grids as uint8 arrays of shape (T, H, B*W, C)

Usage Examples

Logger Setup

from dreamerv3.main import make_logger

logger = make_logger(config)

# Add metrics and write
logger.add({'loss/dyn': 0.5, 'loss/rep': 0.3}, prefix='train')
logger.add({'score': 42.0, 'length': 1000}, prefix='episode')
logger.write()

# Close when done
logger.close()

Open-Loop Reporting

# During training loop
if should_report(step):
    agg = elements.Agg()
    for _ in range(args.report_batches):
        carry_report, mets = agent.report(carry_report, next(stream_report))
        agg.add(mets)
    logger.add(agg.result(), prefix='report')

Related Pages

Implements Principle

• Principle:Danijar_Dreamerv3_Logging_And_Reporting