CHANGELOG.md

Changelog

All notable changes to this project will be documented in this file.

1.6.1

Added

• Native Action Recognition video uploads with Project.upload_video (#535). Upload original MP4/MOV bytes with batch, tags, metadata, and split; read processing state with get_video_upload_status or use bounded polling with wait_for_video_upload / upload_video(..., wait=True). Use the final uploaded response's videoId: deduplication can change the canonical Source ID and return resolvedBatch: null. See the public upload example.

Fixed

• CLI project creation honors --api-key and the selected workspace's stored credentials (#536).
• roboflow train start --type cosmos3-edge returns a trainingId through the v2 training API even without a custom recipe (#537).

1.6.0

Added

• EU region selection for the SDK and CLI (#513):
  • ROBOFLOW_REGION=us|eu (case-insensitive) selects the platform; an environment value takes precedence over the saved region.
  • roboflow auth login --region eu authenticates against the EU data-residency platform and records the credentials' region.
  • roboflow auth set-region <us|eu> saves the region; roboflow auth status reports the region and environment.
  • Explicit per-URL settings such as API_URL still take precedence over the region. Hosted semantic segmentation has no EU deployment and raises an error in the EU region instead of sending images to the US.
• project create --type action-recognition in the CLI (#532). Requires a Roboflow platform release that accepts the type; older releases answer HTTP 422 and the CLI adds a hint.
• Action Recognition training from the SDK (#533): Version.create_training(model_type="cosmos3-edge", ...) ensures the video-coco export and returns a Training handle. Version.train() is unchanged.
• add_to_dataset on Project.single_upload / Project.save_annotation (#531). Pass False to store an annotation without moving the image out of its batch into the Dataset. The default (None) keeps the API's current behavior.

1.5.1

Added

• Model evaluation comparison for a dataset version (#529):
  • Workspace.compare_model_evaluations(project, version, frontier_metric=None) — compare test-set accuracy and median latency, including Pareto frontier membership and reasons models are excluded.
  • roboflow --workspace <workspace> eval compare --project <project> --version <N> — display the comparison as a table; use --json for the public API response. Use --frontier-metric to choose the metric for frontier membership.
  • Both surfaces read existing results without starting evaluations.

Changed

• All roboflow eval commands now return exit code 2 for HTTP 401/403 access errors (previously 1).

1.5.0

Added

• Hosted auto-label support, matching the Roboflow MCP tools:
  • Workspace.autolabel_models() — list the foundation-model catalog (gpt-6-astra-boxes, sam3-rle, gemini-boxes, ...) with availability, guidance and credits per image.
  • Project.autolabel_preview(model, image, ontology=...) — free single-image preview to compare models before starting a job. image accepts an HTTPS URL, a local file path or a base64 string.
  • Project.autolabel(batch_id, model, model_type="foundational" | "roboflow", ...) — start a job over a batch; returns {jobId, annotationJobId}. The ontology is keyed by prompt ({"kitten": "cat", "tabby": "cat"}), so several prompts can share one output class. preserve_existing_annotations=True keeps annotations already on the images (the server default replaces them).
  • Project.autolabel_job(job_id) / Workspace.autolabel_job(job_id) — poll per-subjob progress.
  • roboflow autolabel models | preview | start | job CLI commands. start --preserve-existing mirrors the SDK flag; job -p ws/project resolves the workspace the same way start does.

Changed

• HEIC/HEIF decoding is an optional extra: pip install "roboflow[heic]" installs pillow-heif>=1.7.0, and import roboflow registers its Pillow opener when it is installed. The default install no longer depends on pi-heif, which is discontinued upstream: its final release, 1.4.0, bundles libheif 1.23.0, which is affected by the security advisories fixed in libheif 1.23.2 and 1.23.3 (including CVE-2026-84383). roboflow no longer registers pi-heif even when it is still installed.
  • Uploads and Project.check_valid_image() handle HEIC without the extra. Decoding a local HEIC file, for example with model.predict("photo.heic"), needs it.
  • The extra is opt-in because pillow-heif's binary wheels bundle the x265 encoder, which makes them GPL-2.0 (#398).

1.4.1

Added

• Custom train recipes on v2 trainings (#510):
  • Version.describe_train_recipe(model_type) — fetch the tunable hyperparameter schema, allowed online augmentation/preprocessing steps, and a ready-to-edit recipe template for a model type.
  • train_recipe on Version.create_training(...) — pass an edited describe_train_recipe template for custom hyperparameters/online augmentation (the server dense-fills omitted defaults). A top-level epochs is folded into the recipe's hyperparameters (the server resolves recipe epochs ahead of the body value). train_recipe requires model_type — recipes are minted per model type, and without one the platform would train the project's default architecture.
  • roboflow train recipe -p <project> -v <N> -m <model_type> — print the recipe schema and template as JSON.
  • roboflow train start --train-recipe '<json>' — create the training through the v2 API and print the new trainingId. Accepts inline JSON or a curl-style file reference (--train-recipe @train_recipe.json).

1.4.0

Added — Support for multiple models per version

A dataset version can now own many trainings, and a training can produce many models (e.g. a NAS sweep). New object types expose this:

SDK (roboflow/core/training.py, roboflow/core/version.py):

• Version.trainings() — list the version's training runs as Training objects.
• Version.models() — every trained model for the version (the union across its trainings), as TrainedModel objects. This is now the canonical way to get a version's models.
• Version.create_training(speed=, model_type=, checkpoint=, epochs=) — launch a run without blocking, returning a Training.
• Training — .models, .refresh(), .cancel(), .stop(), plus .training_id / .status / .model_type.
• TrainedModel — .predict(), .predict_video(), .download(), plus .model_id / .model_type / .metrics. A TrainedModel does everything the old version.model could; you just reach it through version.models().

Adapters (roboflow/adapters/rfapi.py): v2 trainings endpoints — list_trainings_for_version, get_training, create_training_v2, cancel_training_v2, stop_training_v2, get_model_weights_url.

Added

• workspace.update_image_metadata() and workspace.batch_update_image_metadata() (plus a project.update_image_metadata() convenience alias) — public SDK wrappers for updating metadata and tags on existing images, previously only reachable via the internal rfapi adapter or the CLI. The batch method accepts wait=True to poll the async task until completion and return per-image results.
• Upload raw rf-detr PyTorch-Lightning checkpoints (e.g. checkpoint_best_ema.pth): upload_model detects them and rebuilds a deploy-ready bundle via rf-detr's export_for_roboflow (requires rfdetr>=1.8.0) (#488)

Changed

• Keypoint detection inference now reports its prediction type correctly (previously mislabeled as classification), fixing rendering/plotting of keypoint predictions.

Deprecated

• version.model (the singular attribute) is deprecated and emits a DeprecationWarning. It cannot represent a version with multiple models; use version.models() instead.

1.3.11

Added

• roboflow api-key CLI command group and SDK methods to create, list, get, update, protect, and revoke workspace API keys — including scoped keys, folder restrictions, and custom metadata (scoping/metadata require the Advanced API Keys plan feature).

1.3.10

Added

• Weight upload support for yolo26-sem semantic segmentation models via version.deploy() and workspace.deploy_model()

1.3.9

Added — Model evaluations SDK & CLI

Wraps the public /{workspace}/model-evals REST surface (roboflow/roboflow#11636) so users can read evaluation results — mAP, confidence sweep, per-class performance, confusion matrix, vector clusters, per-image stats, recommendations — from Python and from the CLI without hitting the API directly. Companion docs: roboflow-dev-reference#18.

SDK (roboflow/core/model_eval.py):

• Workspace.evals(project=None, version=None, model=None, status=None, limit=None) — list evals as ModelEval instances pre-populated with metadata from the list response.
• Workspace.eval(eval_id) — fetch a single eval (returns a ModelEval with .summary populated when status is done).
• ModelEval.refresh() — re-fetch the eval header.
• ModelEval.map_results(), .confidence_sweep(), .performance_by_class(split=None), .confusion_matrix(split=None, confidence=None), .vector_analysis(confidence=None), .image_predictions(split=None, confidence=None, limit=None, offset=None), .recommendations() — one method per panel; each returns the raw JSON dict.

CLI (roboflow/cli/handlers/eval.py):

• roboflow eval list [--project P] [--version V] [--model M] [--status S] [--limit N]
• roboflow eval get <eval_id>
• roboflow eval map-results <eval_id>
• roboflow eval confidence-sweep <eval_id>
• roboflow eval performance-by-class <eval_id> [--split S]
• roboflow eval confusion-matrix <eval_id> [--split S] [--confidence N]
• roboflow eval vector-analysis <eval_id> [--confidence N]
• roboflow eval image-predictions <eval_id> [--split S] [--confidence N] [--limit N] [--offset N]
• roboflow eval recommendations <eval_id>

Exit codes are stable per error class so shell scripts and AI agents can react without parsing message strings: 3 for model_eval_not_found (404), 4 for model_eval_not_done (409), 5 for invalid_split / invalid_confidence (400). Every command supports --json for structured output.

Low-level (roboflow.adapters.rfapi):

• list_model_evals, get_model_eval, get_model_eval_map_results, get_model_eval_confidence_sweep, get_model_eval_performance_by_class, get_model_eval_confusion_matrix, get_model_eval_vector_analysis, get_model_eval_image_predictions, get_model_eval_recommendations.
• New typed exceptions ModelEvalNotFoundError, ModelEvalNotDoneError, InvalidSplitError, InvalidConfidenceError (all subclasses of RoboflowError) so callers can distinguish "eval doesn't exist" from "eval still running" from "bad argument" without parsing strings.

The endpoints require the model-eval:read scope. The base URL is configurable via API_URL (set to https://localapi.roboflow.one to test against a local API server).

Fixed

• rf-detr model upload: accept checkpoints whose args is a plain dict (e.g. EMA checkpoints) when extracting class names, instead of raising TypeError from vars().

Changed

• Pin typer<0.26 and declare click explicitly: typer 0.26 vendors its own click and drops the external dependency, which broke the CLI and its type checks.

1.3.7

Added — Soft-delete / Trash support

Mirrors the soft-delete and Trash features added to the Roboflow web app (roboflow/roboflow#11131). Deleting a project, version, or workflow now moves it to Trash with a 30-day retention window (and cancels any in-flight training jobs); items can be restored within that window. Companion docs: roboflow-dev-reference#5.

SDK (roboflow/):

• Project.delete() / Project.restore() — soft-delete and restore by slug.
• Version.delete() / Version.restore() — same shape on a version handle.
• Workspace.trash() — list everything currently in a workspace's Trash, grouped by projects / versions / workflows.
• Workspace.restore_from_trash(item_type, item_id, parent_id=None) — restore an item by id when you don't have a live SDK handle (or for workflows, which don't have a first-class object yet).

CLI (roboflow/cli/):

• roboflow project delete / roboflow project restore
• roboflow version delete / roboflow version restore
• roboflow workflow delete / roboflow workflow restore
• roboflow trash list

Destructive commands prompt for confirmation interactively and accept --yes / -y for scripted use. Every command supports --json for structured output and emits actionable error hints with stable exit codes.

Low-level (roboflow.adapters.rfapi):

• delete_project, delete_version, delete_workflow, list_trash, restore_trash_item.
• RoboflowError messages now extract the error field from JSON response bodies (e.g. "Not authorized to view trash") instead of the raw response text.

Permanent deletion is intentionally web-UI-only. Emptying Trash or immediately deleting a single Trash item destroys data irrecoverably, so those actions are not exposed on the SDK or CLI — they live only in the Roboflow app's Trash view, which has an explicit confirmation dialog. Items left in Trash are cleaned up automatically after 30 days.

Fixed — Workflows created via SDK/CLI now execute successfully

Workspace.create_workflow() and roboflow workflow create --definition auto-wrap bare workflow definitions in {"specification": ...} before POSTing to the backend, matching what the web app does (#460). Previously, the user-facing flat shape ({version, inputs, steps, outputs}) was sent verbatim, so POST /infer/workflows/... against the resulting workflow returned HTTP 502 with MalformedWorkflowResponseError: Workflow specification not found in Roboflow API response.

Workflows already wrapped (top-level specification key) are passed through unchanged. Non-workflow dicts and non-JSON strings are also passed through verbatim so custom payloads aren't second-guessed.

Note: workflows that were stored with the bare shape before this fix will still 502 until re-saved. Run roboflow workflow update <url> --definition <file> once per affected workflow to migrate.

Changed — Image upload no longer re-encodes images

upload_image now uploads original image bytes instead of re-encoding to JPEG client-side (#464).

Backward compatibility

Purely additive on the public API surface. The new endpoints require project:update, version:update, or workflow:update scopes — most existing keys already have these.

1.1.50

• Added support for Palligema2 model uploads via upload_model command with the following model types:
  • paligemma2-3b-pt-224
  • paligemma2-3b-pt-448
  • paligemma2-3b-pt-896