Skip to content

armgit / camel

Public Imported from GitHub
0

🐫 CAMEL: The first and the best multi-agent framework. Finding the Scaling Law of Agents. https://www.camel-ai.org

enhance: Non deterministic Reset method in single step env for BaseGe… (#3300)

Wendong-Fane38a3315d7a7
diff --git a/camel/datasets/base_generator.py b/camel/datasets/base_generator.py
index 31f18f5af..ec443c4fb 100644
--- a/camel/datasets/base_generator.py
+++ b/camel/datasets/base_generator.py
@@ -62,6 +62,7 @@ class BaseGenerator(abc.ABC, IterableDataset):
self._buffer = buffer
self._data: List[DataPoint] = []
self._batch_to_save: List[DataPoint] = []
+ self._iter_position: int = 0
if data_path:
file_path = Path(data_path)
@@ -103,9 +104,9 @@ class BaseGenerator(abc.ABC, IterableDataset):
r"""Async iterator that yields datapoints dynamically.
If a `data_path` was provided during initialization, those datapoints
- are yielded first. When self._data is empty, 20 new datapoints
- are generated. Every 100 yields, the batch is appended to the
- JSONL file or discarded if `cache` is None.
+ are yielded first. When self._iter_position reaches the end of _data,
+ new datapoints are generated. Every 100 yields, the batch is appended
+ to the JSONL file or discarded if `cache` is None.
Yields:
DataPoint: A single datapoint.
@@ -113,9 +114,10 @@ class BaseGenerator(abc.ABC, IterableDataset):
async def generator():
while True:
- if not self._data:
+ if self._iter_position >= len(self._data):
await self.generate_new(self._buffer)
- datapoint = self._data.pop(0)
+ datapoint = self._data[self._iter_position]
+ self._iter_position += 1
yield datapoint
self._batch_to_save.append(datapoint)
if len(self._batch_to_save) == 100:
@@ -132,9 +134,9 @@ class BaseGenerator(abc.ABC, IterableDataset):
r"""Synchronous iterator for PyTorch IterableDataset compatibility.
If a `data_path` was provided during initialization, those datapoints
- are yielded first. When self._data is empty, 20 new datapoints
- are generated. Every 100 yields, the batch is appended to the
- JSONL file or discarded if `cache` is None.
+ are yielded first. When self._iter_position reaches the end of _data,
+ new datapoints are generated. Every 100 yields, the batch is appended
+ to the JSONL file or discarded if `cache` is None.
Yields:
DataPoint: A single datapoint.
@@ -150,9 +152,10 @@ class BaseGenerator(abc.ABC, IterableDataset):
raise
while True:
- if not self._data:
+ if self._iter_position >= len(self._data):
asyncio.run(self.generate_new(self._buffer))
- datapoint = self._data.pop(0)
+ datapoint = self._data[self._iter_position]
+ self._iter_position += 1
yield datapoint
self._batch_to_save.append(datapoint)
if len(self._batch_to_save) == 100:
@@ -248,6 +251,7 @@ class BaseGenerator(abc.ABC, IterableDataset):
self.save_to_jsonl(file_path)
self._data = []
+ self._iter_position = 0
logger.info(f"Data flushed to {file_path} and cleared from the memory")
def _init_from_jsonl(self, file_path: Path) -> List[Dict[str, Any]]:
diff --git a/camel/environments/single_step.py b/camel/environments/single_step.py
index 07f403e50..6a300f118 100644
--- a/camel/environments/single_step.py
+++ b/camel/environments/single_step.py
@@ -218,11 +218,33 @@ class SingleStepEnv:
return observations[0] if batch_size == 1 else observations
elif isinstance(self.dataset, BaseGenerator):
+ # Generate more data if needed
if batch_size > len(self.dataset):
new_datapoints_needed = batch_size - len(self.dataset)
await self.dataset.generate_new(n=new_datapoints_needed)
- random_indices = rng.sample(range(len(self.dataset)), batch_size)
- self._states = [self.dataset[ind] for ind in random_indices]
+
+ # Verify that enough data was generated
+ if len(self.dataset) < batch_size:
+ raise RuntimeError(
+ f"Failed to generate enough datapoints. "
+ f"Requested {batch_size}, but only "
+ f"{len(self.dataset)} available after generation."
+ )
+
+ # Choose sampling strategy based on whether seed is provided
+ if seed is not None:
+ # Deterministic random sampling when seed is provided
+ random_indices = rng.sample(
+ range(len(self.dataset)), batch_size
+ )
+ self._states = [self.dataset[ind] for ind in random_indices]
+ else:
+ # Sequential sampling when no seed (backward compatible)
+ # Use async_sample to maintain sequential behavior
+ self._states = [
+ await self.dataset.async_sample()
+ for _ in range(batch_size)
+ ]
self.current_batch_size = batch_size
self._states_done = [False] * batch_size