Skip to content

Transforms

TorchRL transforms are composable modules that modify environment observations, rewards, or actions. TorchTrade extends TorchRL's transform system with domain-specific transforms for trading.

Available Transforms

Transform Purpose Use Case
CoverageTracker Track dataset coverage during training Monitor exploration, detect overfitting
ChronosEmbeddingTransform Embed time series with Chronos T5 models Replace raw OHLCV with learned representations
TimestampTransform Add Unix timestamps to TensorDicts Create offline datasets from live trading runs

CoverageTracker

Tracks which portions of the dataset are visited during training with random episode resets. Monitors both episode start diversity (reset coverage) and full trajectory coverage (state coverage). Auto-detects environment settings and only activates when random_start=True.

Metric Range Meaning
reset_coverage [0, 1] Fraction of positions used as episode starts
state_coverage [0, 1] Fraction of all states visited during episodes
reset_entropy / state_entropy [0, log(N)] Uniformity of visit distribution 
from torchtrade.envs.transforms import CoverageTracker
from torchrl.collectors import SyncDataCollector

# Use as postproc in collector (NOT in environment transform chain)
coverage_tracker = CoverageTracker()

collector = SyncDataCollector(
    env, policy,
    frames_per_batch=1000,
    total_frames=100000,
    postproc=coverage_tracker,
)

for batch in collector:
    # ... train on batch ...
    stats = coverage_tracker.get_coverage_stats()
    if stats["enabled"]:
        logger.log({
            "train/reset_coverage": stats["reset_coverage"],
            "train/state_coverage": stats["state_coverage"],
        })

ChronosEmbeddingTransform

Embeds time series observations using pretrained Chronos T5 forecasting models. Replaces raw OHLCV data with learned representations. Model is lazy-loaded on first use.

Model Parameters Embedding Dim Memory
chronos-t5-tiny 8M 512 ~1GB
chronos-t5-mini 20M 512 ~2GB
chronos-t5-small 46M 768 ~3GB
chronos-t5-base 200M 768 ~5GB
chronos-t5-large 710M 1024 ~12GB
Parameter Default Description
in_keys Required Market data keys to embed
out_keys Required Output embedding keys
model_name "amazon/chronos-t5-large" HuggingFace model ID
aggregation "mean" "mean", "max", or "concat" for multi-feature embeddings
del_keys True Remove input keys after transformation 
from torchrl.envs import TransformedEnv, Compose
from torchtrade.envs.transforms import ChronosEmbeddingTransform

env = TransformedEnv(
    base_env,
    Compose(
        ChronosEmbeddingTransform(
            in_keys=["market_data_1Minute_12"],
            out_keys=["chronos_embedding"],
            model_name="amazon/chronos-t5-base",
            aggregation="mean",
        ),
    )
)
# Observation: {"market_data_1Minute_12": (12, 5)} → {"chronos_embedding": (768,)}

pip install git+https://github.com/amazon-science/chronos-forecasting.git

TimestampTransform

Adds Unix timestamps to TensorDicts on reset and step. Useful for creating offline datasets from live trading runs, debugging latency, and correlating trading decisions with real-world events.

Parameter Default Description
out_key "timestamp" Key to store the timestamp 
from torchrl.envs import TransformedEnv
from torchtrade.envs.transforms import TimestampTransform

env = TransformedEnv(base_env, TimestampTransform())

td = env.reset()           # td["timestamp"] = 1738617600.123
td = env.step(td)          # td["next", "timestamp"] = 1738617600.456

# Convert to datetime
from datetime import datetime
dt = datetime.fromtimestamp(td["timestamp"])

Note: Timestamps are wall-clock time (when step()/reset() is called), not execution timeframe timestamps. For execution timestamps, use the market data timestamps in observations.

See Also