Skip to content

examples.md

Latest commit

 

History

History
191 lines (128 loc) · 5.24 KB

examples.md

File metadata and controls

191 lines (128 loc) · 5.24 KB
Raw

Examples

DeepRacerLog is the main entry point in deepracer-utils. It supports local folders, S3 buckets, and .tar.gz exports from the DeepRacer console via TarFileHandler.

Training

Local model folder (trace CSV)

Use this when you have an extracted model folder with sim-trace/ or training-simtrace/ data.

from deepracer.logs import AnalysisUtils, DeepRacerLog

log = DeepRacerLog(model_folder="./deepracer/logs/sample-console-logs")
log.load_training_trace()
df = log.dataframe()

simulation_agg = AnalysisUtils.simulation_agg(df)

df contains per-step data. simulation_agg is per-episode aggregated data.

For continuous-action logs, the action column is set to -1 (no discrete action index). The raw steering and speed values are still available in steering_angle and speed.

Local .tar.gz export (new)

If you downloaded console logs as a .tar.gz, you can read them directly without extracting.

from deepracer.logs import AnalysisUtils, DeepRacerLog, TarFileHandler

fh = TarFileHandler("./deepracer/logs/sample-droa-solution-logs.tar.gz")
log = DeepRacerLog(filehandler=fh)

# New console/DROA archives may not include metadata files.
log.load_training_trace(ignore_metadata=True)
df = log.dataframe()

simulation_agg = AnalysisUtils.simulation_agg(df)

Auto-load shortcut

load() picks the correct training loader based on detected folder type.

from deepracer.logs import DeepRacerLog, TarFileHandler

log = DeepRacerLog(filehandler=TarFileHandler("./my-model-logs.tar.gz"))
log.load(ignore_metadata=True)
df = log.dataframe()

Verbose output (new)

Pass verbose=True to print helpful load progress to stdout.

from deepracer.logs import DeepRacerLog, TarFileHandler

log = DeepRacerLog(
    filehandler=TarFileHandler("./my-model-logs.tar.gz"),
    verbose=True,
)
log.load_training_trace(ignore_metadata=True)

Typical output:

Folder type detected: DROA_SOLUTION_LOGS
Loaded training trace: 846 steps, 40 episodes, 2 iterations

When verbose=True, load methods print a one-line summary:

  • load_training_trace() -> Loaded training trace: ...
  • load_evaluation_trace() -> Loaded evaluation trace: ...
  • load_robomaker_logs() -> Loaded robomaker logs: ...

Each summary reports steps, episodes, and iterations.

S3 bucket

from deepracer.logs import AnalysisUtils, DeepRacerLog, S3FileHandler

fh = S3FileHandler(bucket="<my_bucket>", prefix="<my_prefix>")
log = DeepRacerLog(filehandler=fh)
log.load_training_trace()
df = log.dataframe()

simulation_agg = AnalysisUtils.simulation_agg(df)

Local S3 / MinIO

from deepracer.logs import AnalysisUtils, DeepRacerLog, S3FileHandler

fh = S3FileHandler(
    bucket="<my_bucket>",
    prefix="<my_prefix>",
    profile="<awscli_profile>",
    s3_endpoint_url="<minio_url>",
)
log = DeepRacerLog(filehandler=fh)
log.load_training_trace()
df = log.dataframe()

simulation_agg = AnalysisUtils.simulation_agg(df)

Robomaker / simulation logs

For console-style folders, you can also load training from the simulation log file.

from deepracer.logs import AnalysisUtils, DeepRacerLog, TarFileHandler

log = DeepRacerLog(filehandler=TarFileHandler("./my-model-logs.tar.gz"))
log.load_robomaker_logs()
df = log.dataframe()

simulation_agg = AnalysisUtils.simulation_agg(df)

Evaluation

Evaluation data can include multiple runs; they are separated with the stream column.

from deepracer.logs import AnalysisUtils, DeepRacerLog, TarFileHandler

log = DeepRacerLog(filehandler=TarFileHandler("./my-model-logs.tar.gz"))
log.load_evaluation_trace(ignore_metadata=True)
df = log.dataframe()

simulation_agg = AnalysisUtils.simulation_agg(df, firstgroup="stream", is_eval=True)

Evaluation-only archive (new)

Some exports contain only evaluation data. load_evaluation_trace() works the same way:

from deepracer.logs import DeepRacerLog, TarFileHandler

log = DeepRacerLog(filehandler=TarFileHandler("./sample-droa-eval-only.tar.gz"))
log.load_evaluation_trace(ignore_metadata=True)
df = log.dataframe()

Evaluation from simulation logs

from deepracer.logs import AnalysisUtils, DeepRacerLog, LogType, TarFileHandler

log = DeepRacerLog(filehandler=TarFileHandler("./my-model-logs.tar.gz"))
log.load_robomaker_logs(type=LogType.EVALUATION)
df = log.dataframe()

simulation_agg = AnalysisUtils.simulation_agg(df, firstgroup="stream", is_eval=True)

Leaderboard submissions

Leaderboard submissions are available through simulation log files.

from deepracer.logs import AnalysisUtils, DeepRacerLog, LogType, TarFileHandler

log = DeepRacerLog(filehandler=TarFileHandler("./my-model-logs.tar.gz"))
log.load_robomaker_logs(type=LogType.LEADERBOARD)
df = log.dataframe()

simulation_agg = AnalysisUtils.simulation_agg(df, firstgroup="stream", is_eval=True)

Stability analysis (new)

You can analyze step timing stability from loaded simtrace data.

from deepracer.logs import DeepRacerLog, TarFileHandler

log = DeepRacerLog(filehandler=TarFileHandler("./my-model-logs.tar.gz"))
log.load_training_trace(ignore_metadata=True)

stability_df = log.stability.analyze()
print(stability_df[["iteration", "avg_ms", "p95_ms", "rtf"]].head())