Terminal
Terminal/Skill
Annotates passive acoustic recordings and builds custom species detectors for ecologists, using the perch-hoplite Python package.
By Google
Works with
npx skills add google-deepmind/agentic_ecology --skill agentic-ecology-bioacoustics[!CAUTION] DATABASE SAFETY AND INTEGRITY: Do NOT write, modify, or insert test annotations directly into the user's production databases. If you need to test database operations (such as saving annotations or training models), you MUST copy the database to a temporary location in
agent_workspace(e.g.,agent_workspace/test_db) and test against the copy. Never leave testing data in production databases.
Follow these sequential steps:
perch_v2, surfperch). Query the user if they have not specified one.AGENTS.md to evaluate options (Local vs. Colab
vs. GCP Dataflow) and await the user's decision.[!IMPORTANT] MANDATORY INGESTION PIPELINE: Always use
dataflow_embed.pyto extract audio embeddings into sharded Apache Parquet files, followed byingest_embeddings.pyto build the Hoplite database. Do NOT useEmbedWorkerdirectly.
--runner=DirectRunner, passing local audio paths for --input_glob and a local workspace directory (e.g., agent_workspace/embeddings/<dataset>) for --output_dir.databases/<dataset> using ingest_embeddings.py.AGENTS.md.--runner=DirectRunner on the Colab GPU/TPU runtime to generate
sharded Apache Parquet embeddings.databases/ directory. Ensure the database
audio source metadata points to the local recordings path so the
local web app can resolve and stream audio.AGENTS.md.DataflowRunner to extract embeddings and generate sharded
Apache Parquet files on GCS.google-cloud-storage-fuse) so the web app can stream audio
windows on demand without downloading full recordings.
Refer to GCP Dataflow Audio Embedding
for complete instructions and commands.perch-hoplite APIs, and use the classifier's weights to
score database rows and rerank them.For detailed API usage, implementation instructions, and code examples, see:
This reference covers:
dataflow_embed.py (with DirectRunner locally or DataflowRunner on GCP)ingest_embeddings.pyingest_embeddings.py