visionforge-studio 0.20.2__tar.gz → 0.22.0__tar.gz
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- visionforge_studio-0.22.0/PKG-INFO +878 -0
- visionforge_studio-0.22.0/README.md +783 -0
- visionforge_studio-0.22.0/THIRD_PARTY_NOTICES.md +145 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/pyproject.toml +76 -36
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/__main__.py +137 -11
- visionforge_studio-0.22.0/src/visionforge/blocks/__init__.py +62 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/blocks/anomaly.py +28 -6
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/blocks/classification.py +39 -12
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/blocks/cross_validation.py +35 -30
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/blocks/regression.py +14 -6
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/blocks/segmentation.py +16 -6
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/blocks/transfer_learning.py +10 -13
- visionforge_studio-0.22.0/src/visionforge/cli_strategies.py +307 -0
- visionforge_studio-0.22.0/src/visionforge/core/__init__.py +53 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/core/anomaly_trainer.py +123 -68
- visionforge_studio-0.22.0/src/visionforge/core/code_tour.py +1598 -0
- visionforge_studio-0.22.0/src/visionforge/core/comet_tracking.py +341 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/core/comparison.py +6 -6
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/core/data.py +29 -1
- visionforge_studio-0.22.0/src/visionforge/core/dataset_coco.py +527 -0
- visionforge_studio-0.22.0/src/visionforge/core/dataset_coco_import.py +954 -0
- visionforge_studio-0.22.0/src/visionforge/core/dataset_derived.py +378 -0
- visionforge_studio-0.22.0/src/visionforge/core/dataset_explorer.py +592 -0
- visionforge_studio-0.22.0/src/visionforge/core/dataset_incorporate.py +829 -0
- visionforge_studio-0.22.0/src/visionforge/core/dataset_merge.py +865 -0
- visionforge_studio-0.22.0/src/visionforge/core/dataset_resplit.py +925 -0
- visionforge_studio-0.22.0/src/visionforge/core/detection_images.py +34 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/core/detection_trainer.py +183 -37
- visionforge_studio-0.22.0/src/visionforge/core/dvc_export.py +490 -0
- visionforge_studio-0.22.0/src/visionforge/core/experiment_bundle.py +630 -0
- visionforge_studio-0.22.0/src/visionforge/core/github_export.py +251 -0
- visionforge_studio-0.22.0/src/visionforge/core/hub_export.py +271 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/core/latex_export.py +41 -6
- visionforge_studio-0.22.0/src/visionforge/core/live_confusion.py +50 -0
- visionforge_studio-0.22.0/src/visionforge/core/live_insights.py +281 -0
- visionforge_studio-0.22.0/src/visionforge/core/metric_vocabulary.py +107 -0
- visionforge_studio-0.22.0/src/visionforge/core/mlflow_tracking.py +254 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/core/plotter.py +123 -14
- visionforge_studio-0.22.0/src/visionforge/core/public_run_record.py +54 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/core/replicated_comparison.py +1 -1
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/core/replicates.py +96 -41
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/core/resume.py +19 -2
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/core/run_groups.py +457 -2
- visionforge_studio-0.22.0/src/visionforge/core/run_report.py +198 -0
- visionforge_studio-0.22.0/src/visionforge/core/run_trackers.py +79 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/core/segmentation_data.py +6 -3
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/core/significance.py +154 -20
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/core/sweep.py +12 -7
- visionforge_studio-0.22.0/src/visionforge/core/tracking_privacy.py +143 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/core/trainer.py +171 -41
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/core/training_health.py +59 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/core/training_loop.py +354 -73
- visionforge_studio-0.22.0/src/visionforge/core/wandb_tracking.py +348 -0
- visionforge_studio-0.22.0/src/visionforge/core/zenodo_export.py +188 -0
- visionforge_studio-0.22.0/src/visionforge/gui/api/coco_jobs.py +130 -0
- visionforge_studio-0.22.0/src/visionforge/gui/api/coco_routes.py +288 -0
- visionforge_studio-0.22.0/src/visionforge/gui/api/code_tour_routes.py +105 -0
- visionforge_studio-0.22.0/src/visionforge/gui/api/colab_connection.py +408 -0
- visionforge_studio-0.22.0/src/visionforge/gui/api/colab_drive.py +341 -0
- visionforge_studio-0.22.0/src/visionforge/gui/api/colab_jobs.py +641 -0
- visionforge_studio-0.22.0/src/visionforge/gui/api/colab_remote_runner.py +301 -0
- visionforge_studio-0.22.0/src/visionforge/gui/api/colab_routes.py +175 -0
- visionforge_studio-0.22.0/src/visionforge/gui/api/colab_run_routes.py +418 -0
- visionforge_studio-0.22.0/src/visionforge/gui/api/comet_routes.py +86 -0
- visionforge_studio-0.22.0/src/visionforge/gui/api/custom_task_routes.py +352 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/api/dataset_download.py +49 -2
- visionforge_studio-0.22.0/src/visionforge/gui/api/dataset_tools_routes.py +267 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/api/datasets_routes.py +58 -22
- visionforge_studio-0.22.0/src/visionforge/gui/api/derived_jobs.py +116 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/api/detection_export.py +10 -5
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/api/detection_testing.py +6 -7
- visionforge_studio-0.22.0/src/visionforge/gui/api/dvc_routes.py +169 -0
- visionforge_studio-0.22.0/src/visionforge/gui/api/error_messages.py +93 -0
- visionforge_studio-0.22.0/src/visionforge/gui/api/evaluation_routes.py +886 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/api/evidence_routes.py +5 -1
- visionforge_studio-0.22.0/src/visionforge/gui/api/execution.py +355 -0
- visionforge_studio-0.22.0/src/visionforge/gui/api/execution_routes.py +194 -0
- visionforge_studio-0.22.0/src/visionforge/gui/api/explorer_jobs.py +101 -0
- visionforge_studio-0.22.0/src/visionforge/gui/api/explorer_routes.py +222 -0
- visionforge_studio-0.22.0/src/visionforge/gui/api/github_routes.py +122 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/api/gpu_health.py +6 -3
- visionforge_studio-0.22.0/src/visionforge/gui/api/http_responses.py +75 -0
- visionforge_studio-0.22.0/src/visionforge/gui/api/hub_routes.py +135 -0
- visionforge_studio-0.22.0/src/visionforge/gui/api/incorporate_jobs.py +108 -0
- visionforge_studio-0.22.0/src/visionforge/gui/api/incorporate_routes.py +193 -0
- visionforge_studio-0.22.0/src/visionforge/gui/api/merge_jobs.py +115 -0
- visionforge_studio-0.22.0/src/visionforge/gui/api/merge_routes.py +199 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/api/model_card.py +43 -29
- visionforge_studio-0.22.0/src/visionforge/gui/api/profile_tokens.py +144 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/api/profiles.py +125 -2
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/api/profiles_routes.py +31 -28
- visionforge_studio-0.22.0/src/visionforge/gui/api/resplit_jobs.py +112 -0
- visionforge_studio-0.22.0/src/visionforge/gui/api/resplit_routes.py +199 -0
- visionforge_studio-0.22.0/src/visionforge/gui/api/routes.py +592 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/api/run_events.py +3 -1
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/api/run_history.py +8 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/api/run_queue.py +88 -2
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/api/runs_routes.py +68 -10
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/api/schemas.py +215 -9
- visionforge_studio-0.22.0/src/visionforge/gui/api/strategies.py +671 -0
- visionforge_studio-0.22.0/src/visionforge/gui/api/strategy_routes.py +387 -0
- visionforge_studio-0.22.0/src/visionforge/gui/api/submission.py +163 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/api/system_routes.py +5 -4
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/api/system_service.py +5 -1
- visionforge_studio-0.22.0/src/visionforge/gui/api/tracker_exports.py +90 -0
- visionforge_studio-0.22.0/src/visionforge/gui/api/training_routes.py +470 -0
- visionforge_studio-0.22.0/src/visionforge/gui/api/update_routes.py +106 -0
- visionforge_studio-0.22.0/src/visionforge/gui/api/wandb_routes.py +89 -0
- visionforge_studio-0.22.0/src/visionforge/gui/api/zenodo_routes.py +115 -0
- visionforge_studio-0.22.0/src/visionforge/gui/security.py +405 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/server.py +92 -11
- visionforge_studio-0.22.0/src/visionforge/gui/static/assets/ActionButton-DQM3s0XR.js +1 -0
- visionforge_studio-0.22.0/src/visionforge/gui/static/assets/BrandIcons-BfiNQMOx.js +1 -0
- visionforge_studio-0.22.0/src/visionforge/gui/static/assets/DatasetDownloadCard-DTfZHd92.js +1 -0
- visionforge_studio-0.22.0/src/visionforge/gui/static/assets/DatasetExplorer-NiO-yS6w.js +1 -0
- visionforge_studio-0.22.0/src/visionforge/gui/static/assets/DatasetPicker-B2LyRLfk.js +1 -0
- visionforge_studio-0.22.0/src/visionforge/gui/static/assets/DvcExportPanel-GDAl2GQR.js +1 -0
- visionforge_studio-0.22.0/src/visionforge/gui/static/assets/GitHubExportPanel-CSyDP73s.js +1 -0
- visionforge_studio-0.22.0/src/visionforge/gui/static/assets/HistoryOverlay-ByeYVAA8.js +3 -0
- visionforge_studio-0.22.0/src/visionforge/gui/static/assets/HubExportPanel-B66xdNEe.js +1 -0
- visionforge_studio-0.22.0/src/visionforge/gui/static/assets/LearnOverlay-wP9V71UT.js +1 -0
- visionforge_studio-0.22.0/src/visionforge/gui/static/assets/LineChart-DSwy7ZIj.js +1 -0
- visionforge_studio-0.22.0/src/visionforge/gui/static/assets/TrainingCodeTab-DBHTDofK.js +1 -0
- visionforge_studio-0.22.0/src/visionforge/gui/static/assets/ZenodoExportPanel-D8YP13hw.js +1 -0
- visionforge_studio-0.22.0/src/visionforge/gui/static/assets/client-AM4-ZEM8.js +9 -0
- visionforge_studio-0.22.0/src/visionforge/gui/static/assets/createLucideIcon-C3_nSUUv.js +1 -0
- visionforge_studio-0.22.0/src/visionforge/gui/static/assets/data-step-SQgXtDj5.webp +0 -0
- visionforge_studio-0.22.0/src/visionforge/gui/static/assets/dataset-identity-Cruczp6b.js +1 -0
- visionforge_studio-0.22.0/src/visionforge/gui/static/assets/drawings-JBuXyz7j.js +1 -0
- visionforge_studio-0.22.0/src/visionforge/gui/static/assets/flask-conical-GnVAw9W9.js +1 -0
- visionforge_studio-0.22.0/src/visionforge/gui/static/assets/folder-open-xw6q7Pi1.js +1 -0
- visionforge_studio-0.22.0/src/visionforge/gui/static/assets/index-Cj1vYMF2.js +6 -0
- visionforge_studio-0.22.0/src/visionforge/gui/static/assets/index-DrW5jBTE.css +2 -0
- visionforge_studio-0.22.0/src/visionforge/gui/static/assets/js-yaml-D1nxiHB1.js +32 -0
- visionforge_studio-0.22.0/src/visionforge/gui/static/assets/picker-feedback-BAGgLSIF.js +1 -0
- visionforge_studio-0.22.0/src/visionforge/gui/static/assets/react-XxnHDjaj.js +9 -0
- visionforge_studio-0.22.0/src/visionforge/gui/static/assets/results-BaqJIBvV.webp +0 -0
- visionforge_studio-0.22.0/src/visionforge/gui/static/assets/rolldown-runtime-Cr0BHlHC.js +1 -0
- visionforge_studio-0.22.0/src/visionforge/gui/static/assets/run-curves-CP1TnjEx.js +1 -0
- visionforge_studio-0.22.0/src/visionforge/gui/static/assets/task-accent-BJeXKC3r.js +1 -0
- visionforge_studio-0.22.0/src/visionforge/gui/static/assets/training-film-CvmtoL7Z.webp +0 -0
- visionforge_studio-0.22.0/src/visionforge/gui/static/assets/training-story-DBqd9SST.js +1 -0
- visionforge_studio-0.22.0/src/visionforge/gui/static/assets/use-viewing-DEjp04xa.js +1 -0
- visionforge_studio-0.22.0/src/visionforge/gui/static/assets/useT-CsmtVqSy.js +1 -0
- visionforge_studio-0.22.0/src/visionforge/gui/static/index.html +43 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/tasks/engine.py +75 -11
- visionforge_studio-0.22.0/src/visionforge/utils/comet_names.py +18 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/utils/config.py +41 -3
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/utils/cuda.py +17 -1
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/utils/doctor.py +121 -30
- visionforge_studio-0.22.0/src/visionforge/utils/messages.py +3290 -0
- visionforge_studio-0.22.0/src/visionforge/utils/third_party_privacy.py +234 -0
- visionforge_studio-0.22.0/src/visionforge/utils/wandb_names.py +16 -0
- visionforge_studio-0.22.0/src/visionforge_studio.egg-info/PKG-INFO +878 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge_studio.egg-info/SOURCES.txt +116 -3
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge_studio.egg-info/requires.txt +28 -8
- visionforge_studio-0.22.0/tests/test_capture_docs_images.py +239 -0
- visionforge_studio-0.22.0/tests/test_cli_strategies.py +539 -0
- visionforge_studio-0.22.0/tests/test_contributing_hooks.py +63 -0
- visionforge_studio-0.22.0/tests/test_decisions_format.py +72 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/tests/test_distribution_smoke.py +32 -0
- visionforge_studio-0.22.0/tests/test_docs_code_tab.py +124 -0
- visionforge_studio-0.22.0/tests/test_docs_exporters.py +68 -0
- visionforge_studio-0.22.0/tests/test_docs_pinned_facts.py +150 -0
- visionforge_studio-0.22.0/tests/test_docs_tokens.py +67 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/tests/test_example_configs.py +2 -0
- visionforge_studio-0.22.0/tests/test_helpers.py +58 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/tests/test_install_smoke.py +82 -3
- visionforge_studio-0.22.0/tests/test_main.py +116 -0
- visionforge_studio-0.22.0/tests/test_module_size.py +40 -0
- visionforge_studio-0.22.0/tests/test_public_claims.py +56 -0
- visionforge_studio-0.22.0/tests/test_readme_parity.py +90 -0
- visionforge_studio-0.22.0/tests/test_readme_pypi.py +64 -0
- visionforge_studio-0.22.0/tests/test_sync_site_release.py +184 -0
- visionforge_studio-0.22.0/tests/test_third_party_notices.py +76 -0
- visionforge_studio-0.20.2/PKG-INFO +0 -561
- visionforge_studio-0.20.2/README.md +0 -459
- visionforge_studio-0.20.2/src/visionforge/blocks/__init__.py +0 -23
- visionforge_studio-0.20.2/src/visionforge/core/__init__.py +0 -14
- visionforge_studio-0.20.2/src/visionforge/core/experiment_bundle.py +0 -330
- visionforge_studio-0.20.2/src/visionforge/gui/api/http_responses.py +0 -20
- visionforge_studio-0.20.2/src/visionforge/gui/api/routes.py +0 -3690
- visionforge_studio-0.20.2/src/visionforge/gui/static/assets/index-CGqpYm51.js +0 -53
- visionforge_studio-0.20.2/src/visionforge/gui/static/assets/index-cJ5jcHVX.css +0 -2
- visionforge_studio-0.20.2/src/visionforge/gui/static/index.html +0 -28
- visionforge_studio-0.20.2/src/visionforge/utils/messages.py +0 -1581
- visionforge_studio-0.20.2/src/visionforge_studio.egg-info/PKG-INFO +0 -561
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/LICENSE +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/setup.cfg +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/__init__.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/blocks/_cv_common.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/blocks/_search_utils.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/blocks/anomaly_runner.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/blocks/base.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/blocks/batch_prediction.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/blocks/classification_runner.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/blocks/detection.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/blocks/detection_runner.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/blocks/export_onnx.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/blocks/grid_search.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/blocks/model_comparison.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/blocks/random_search.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/blocks/registry.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/blocks/regression_cv.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/blocks/regression_runner.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/blocks/segmentation_cv.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/blocks/segmentation_runner.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/core/anomaly_data.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/core/batch_predict.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/core/cancellation.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/core/dataset_audit.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/core/dataset_fingerprint.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/core/detection_data.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/core/detection_dataset.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/core/detection_metrics.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/core/evaluator.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/core/gradcam.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/core/image_size.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/core/learning_rate.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/core/loader_lifecycle.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/core/materialized_dataset.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/core/metric_ci.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/core/onnx_export.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/core/preprocessing.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/core/regression_data.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/core/regression_trainer.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/core/resume_provenance.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/core/segmentation_trainer.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/core/task_runner.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/core/tracking.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/core/trial_name.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/__init__.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/api/__init__.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/api/folder_opener.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/api/model_notes.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/api/sample_dataset.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/api/torch_batch_predict.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/api/torch_gradcam.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/api/torch_onnx_export.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/api/updates.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/static/assets/bricolage-grotesque-latin-ext-wght-normal-CcLUaPy7.woff2 +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/static/assets/bricolage-grotesque-latin-wght-normal-DLoelf7F.woff2 +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/static/assets/bricolage-grotesque-vietnamese-wght-normal-BUzh504Q.woff2 +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/static/assets/familjen-grotesk-latin-ext-wght-normal-Cvgdx4cy.woff2 +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/static/assets/familjen-grotesk-latin-wght-normal-CfZa4wxB.woff2 +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/static/assets/familjen-grotesk-vietnamese-wght-normal-DKRsWTd7.woff2 +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/static/assets/ibm-plex-mono-cyrillic-400-normal-BSMlKf0J.woff2 +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/static/assets/ibm-plex-mono-cyrillic-400-normal-CEL4l2ZJ.woff +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/static/assets/ibm-plex-mono-cyrillic-500-normal-Ael50iVv.woff +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/static/assets/ibm-plex-mono-cyrillic-500-normal-Bq9vWWag.woff2 +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/static/assets/ibm-plex-mono-cyrillic-600-normal-CTOM6hUh.woff2 +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/static/assets/ibm-plex-mono-cyrillic-600-normal-fLZuRloM.woff +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/static/assets/ibm-plex-mono-cyrillic-ext-400-normal-DMdlQ8Kv.woff +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/static/assets/ibm-plex-mono-cyrillic-ext-400-normal-xuaO2J-f.woff2 +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/static/assets/ibm-plex-mono-cyrillic-ext-500-normal-BIfNGwUT.woff +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/static/assets/ibm-plex-mono-cyrillic-ext-500-normal-BqneJy0T.woff2 +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/static/assets/ibm-plex-mono-cyrillic-ext-600-normal-9HEixskS.woff +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/static/assets/ibm-plex-mono-cyrillic-ext-600-normal-V-xxqcpd.woff2 +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/static/assets/ibm-plex-mono-latin-400-normal-CvHOgSBP.woff +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/static/assets/ibm-plex-mono-latin-400-normal-DMJ8VG8y.woff2 +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/static/assets/ibm-plex-mono-latin-500-normal-CB9ihrfo.woff +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/static/assets/ibm-plex-mono-latin-500-normal-DSY6xOcd.woff2 +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/static/assets/ibm-plex-mono-latin-600-normal-BgSNZQsw.woff2 +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/static/assets/ibm-plex-mono-latin-600-normal-DWFSQ4vo.woff +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/static/assets/ibm-plex-mono-latin-ext-400-normal-BmRBH3aV.woff2 +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/static/assets/ibm-plex-mono-latin-ext-400-normal-D3D2R8hC.woff +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/static/assets/ibm-plex-mono-latin-ext-500-normal-CAhNIIs5.woff2 +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/static/assets/ibm-plex-mono-latin-ext-500-normal-CZ70TYgx.woff +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/static/assets/ibm-plex-mono-latin-ext-600-normal-D38SheWl.woff2 +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/static/assets/ibm-plex-mono-latin-ext-600-normal-DmB0ttJJ.woff +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/static/assets/ibm-plex-mono-vietnamese-400-normal-BulugwFq.woff2 +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/static/assets/ibm-plex-mono-vietnamese-400-normal-DDuiU_S-.woff +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/static/assets/ibm-plex-mono-vietnamese-500-normal-C8zxqsMH.woff +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/static/assets/ibm-plex-mono-vietnamese-500-normal-DZ4AoWbu.woff2 +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/static/assets/ibm-plex-mono-vietnamese-600-normal-D2EvbN8M.woff2 +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/static/assets/ibm-plex-mono-vietnamese-600-normal-iLQfcSjf.woff +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/static/favicon.svg +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/static/icons.svg +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/gui/updater.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/models/__init__.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/models/anomaly_factory.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/models/detection_factory.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/models/factory.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/models/registry.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/models/regression_factory.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/models/segmentation_factory.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/models/timm_source.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/tasks/__init__.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/tasks/base.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/tasks/manage.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/tasks/registry.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/tasks/runner.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/tasks/scaffold.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/utils/__init__.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/utils/anomaly_config.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/utils/credentials.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/utils/detection_config.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/utils/environment.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/utils/logger.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/utils/paths.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/utils/redact.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/utils/regression_config.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/utils/segmentation_config.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/utils/selftest.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/utils/selftest_data.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge/utils/workers.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge_studio.egg-info/dependency_links.txt +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge_studio.egg-info/entry_points.txt +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/src/visionforge_studio.egg-info/top_level.txt +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/tests/test_cli_evidence.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/tests/test_cli_run_dispatch.py +0 -0
- {visionforge_studio-0.20.2 → visionforge_studio-0.22.0}/tests/test_sonar_project_version.py +0 -0
|
@@ -0,0 +1,878 @@
|
|
|
1
|
+
Metadata-Version: 2.4
|
|
2
|
+
Name: visionforge-studio
|
|
3
|
+
Version: 0.22.0
|
|
4
|
+
Summary: Local-first computer-vision experimentation platform: five task families, multi-seed replicates with confidence intervals, K-fold CV, sweeps and full run provenance — on your own GPU.
|
|
5
|
+
Author-email: Marcus Reis <marcusvinicius0083@gmail.com>
|
|
6
|
+
License-Expression: MIT
|
|
7
|
+
Project-URL: Repository, https://github.com/marcus-vreis/VisionForge
|
|
8
|
+
Project-URL: Documentation, https://github.com/marcus-vreis/VisionForge/tree/main/docs
|
|
9
|
+
Project-URL: Changelog, https://github.com/marcus-vreis/VisionForge/blob/main/CHANGELOG.md
|
|
10
|
+
Project-URL: Issues, https://github.com/marcus-vreis/VisionForge/issues
|
|
11
|
+
Keywords: computer-vision,deep-learning,pytorch,experiment-tracking,reproducibility,statistical-testing,confidence-intervals,image-classification,object-detection,semantic-segmentation,anomaly-detection,gui
|
|
12
|
+
Classifier: Development Status :: 4 - Beta
|
|
13
|
+
Classifier: Environment :: GPU :: NVIDIA CUDA
|
|
14
|
+
Classifier: Environment :: Web Environment
|
|
15
|
+
Classifier: Intended Audience :: Science/Research
|
|
16
|
+
Classifier: Operating System :: Microsoft :: Windows
|
|
17
|
+
Classifier: Operating System :: POSIX :: Linux
|
|
18
|
+
Classifier: Programming Language :: Python :: 3
|
|
19
|
+
Classifier: Programming Language :: Python :: 3.12
|
|
20
|
+
Classifier: Programming Language :: Python :: 3.13
|
|
21
|
+
Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
|
|
22
|
+
Classifier: Topic :: Scientific/Engineering :: Image Recognition
|
|
23
|
+
Requires-Python: >=3.12
|
|
24
|
+
Description-Content-Type: text/markdown
|
|
25
|
+
License-File: LICENSE
|
|
26
|
+
License-File: THIRD_PARTY_NOTICES.md
|
|
27
|
+
Requires-Dist: pydantic>=2.0
|
|
28
|
+
Requires-Dist: pydantic-settings>=2.0
|
|
29
|
+
Requires-Dist: loguru>=0.7
|
|
30
|
+
Requires-Dist: scikit-learn>=1.4
|
|
31
|
+
Requires-Dist: matplotlib>=3.8
|
|
32
|
+
Requires-Dist: seaborn>=0.13
|
|
33
|
+
Requires-Dist: onnx>=1.21
|
|
34
|
+
Requires-Dist: pillow>=12.2
|
|
35
|
+
Requires-Dist: onnxruntime>=1.18
|
|
36
|
+
Requires-Dist: pyyaml>=6.0
|
|
37
|
+
Requires-Dist: rich>=13.0
|
|
38
|
+
Requires-Dist: fastapi>=0.115
|
|
39
|
+
Requires-Dist: uvicorn>=0.30
|
|
40
|
+
Requires-Dist: timm>=1.0
|
|
41
|
+
Requires-Dist: optuna>=3.0
|
|
42
|
+
Requires-Dist: tensorboard>=2.0
|
|
43
|
+
Requires-Dist: roboflow>=1.1
|
|
44
|
+
Requires-Dist: kaggle>=1.6
|
|
45
|
+
Requires-Dist: datasets>=2.0
|
|
46
|
+
Provides-Extra: cpu
|
|
47
|
+
Requires-Dist: torch>=2.13; extra == "cpu"
|
|
48
|
+
Requires-Dist: torchvision>=0.28; extra == "cpu"
|
|
49
|
+
Provides-Extra: cu118
|
|
50
|
+
Requires-Dist: torch>=2.3; extra == "cu118"
|
|
51
|
+
Requires-Dist: torchvision>=0.18; extra == "cu118"
|
|
52
|
+
Provides-Extra: cu121
|
|
53
|
+
Requires-Dist: torch>=2.3; extra == "cu121"
|
|
54
|
+
Requires-Dist: torchvision>=0.18; extra == "cu121"
|
|
55
|
+
Provides-Extra: cu124
|
|
56
|
+
Requires-Dist: torch>=2.3; extra == "cu124"
|
|
57
|
+
Requires-Dist: torchvision>=0.18; extra == "cu124"
|
|
58
|
+
Provides-Extra: cu126
|
|
59
|
+
Requires-Dist: torch>=2.13; extra == "cu126"
|
|
60
|
+
Requires-Dist: torchvision>=0.28; extra == "cu126"
|
|
61
|
+
Provides-Extra: cu128
|
|
62
|
+
Requires-Dist: torch>=2.10; extra == "cu128"
|
|
63
|
+
Requires-Dist: torchvision>=0.25; extra == "cu128"
|
|
64
|
+
Provides-Extra: cu130
|
|
65
|
+
Requires-Dist: torch>=2.13; extra == "cu130"
|
|
66
|
+
Requires-Dist: torchvision>=0.28; extra == "cu130"
|
|
67
|
+
Provides-Extra: mlflow
|
|
68
|
+
Requires-Dist: mlflow==3.18.0; extra == "mlflow"
|
|
69
|
+
Provides-Extra: wandb
|
|
70
|
+
Requires-Dist: wandb==0.31.0; extra == "wandb"
|
|
71
|
+
Provides-Extra: comet
|
|
72
|
+
Requires-Dist: comet-ml==3.58.12; extra == "comet"
|
|
73
|
+
Provides-Extra: dvc
|
|
74
|
+
Requires-Dist: dvc==3.67.1; extra == "dvc"
|
|
75
|
+
Provides-Extra: detection
|
|
76
|
+
Requires-Dist: ultralytics>=8.3; extra == "detection"
|
|
77
|
+
Provides-Extra: timm
|
|
78
|
+
Provides-Extra: optuna
|
|
79
|
+
Provides-Extra: tensorboard
|
|
80
|
+
Provides-Extra: roboflow
|
|
81
|
+
Provides-Extra: kaggle
|
|
82
|
+
Provides-Extra: huggingface
|
|
83
|
+
Provides-Extra: hub
|
|
84
|
+
Requires-Dist: huggingface-hub<2,>=1.16.1; extra == "hub"
|
|
85
|
+
Provides-Extra: dev
|
|
86
|
+
Requires-Dist: pytest>=8.0; extra == "dev"
|
|
87
|
+
Requires-Dist: pytest-cov>=5.0; extra == "dev"
|
|
88
|
+
Requires-Dist: ruff>=0.4; extra == "dev"
|
|
89
|
+
Requires-Dist: mypy==2.4.0; extra == "dev"
|
|
90
|
+
Requires-Dist: codespell>=2.3; extra == "dev"
|
|
91
|
+
Requires-Dist: pre-commit>=4.0; extra == "dev"
|
|
92
|
+
Requires-Dist: bump-my-version>=0.24; extra == "dev"
|
|
93
|
+
Requires-Dist: types-PyYAML>=6.0; extra == "dev"
|
|
94
|
+
Dynamic: license-file
|
|
95
|
+
|
|
96
|
+
# VisionForge
|
|
97
|
+
|
|
98
|
+
[](https://github.com/marcus-vreis/VisionForge/actions/workflows/ci.yml)
|
|
99
|
+
[](https://github.com/marcus-vreis/VisionForge/blob/main/LICENSE)
|
|
100
|
+
[](https://doi.org/10.5281/zenodo.23425765)
|
|
101
|
+
|
|
102
|
+
**Train, compare and defend computer-vision models on your own GPU.**
|
|
103
|
+
Five task families in one local-first interface, without notebooks — it runs on your machine; nothing leaves it unless you ask for it (datasets, Colab and exporters are opt-in, with a local preview).
|
|
104
|
+
|
|
105
|
+
> 🇧🇷 **Este README também está em português** — [role até a versão em
|
|
106
|
+
> português](#português-pt-br).
|
|
107
|
+
|
|
108
|
+

|
|
109
|
+
|
|
110
|
+
## Install
|
|
111
|
+
|
|
112
|
+
You need **Python 3.12+** and about five minutes. Nothing else — the published
|
|
113
|
+
package already carries the built interface.
|
|
114
|
+
|
|
115
|
+
A step-by-step guide per system (Windows, Linux, macOS), with getting Python
|
|
116
|
+
itself and the usual errors, is in [`docs/INSTALL.md`](https://github.com/marcus-vreis/VisionForge/blob/main/docs/INSTALL.md) (Portuguese).
|
|
117
|
+
On a Mac: Apple Silicon only (PyTorch has no Intel-Mac builds since 2.3),
|
|
118
|
+
training runs on the CPU, and `doctor --fix` is not needed.
|
|
119
|
+
|
|
120
|
+
```bash
|
|
121
|
+
mkdir my-research && cd my-research
|
|
122
|
+
python -m venv .venv
|
|
123
|
+
.venv\Scripts\activate
|
|
124
|
+
```
|
|
125
|
+
|
|
126
|
+
On Linux or macOS the last line is `source .venv/bin/activate` instead.
|
|
127
|
+
|
|
128
|
+
```bash
|
|
129
|
+
pip install visionforge-studio
|
|
130
|
+
```
|
|
131
|
+
|
|
132
|
+
Now let VisionForge look at your machine and install the matching PyTorch build
|
|
133
|
+
— it reads your GPU driver, names the wheel you need, and installs it after you
|
|
134
|
+
answer `y`:
|
|
135
|
+
|
|
136
|
+
```bash
|
|
137
|
+
visionforge doctor --fix
|
|
138
|
+
```
|
|
139
|
+
|
|
140
|
+
That is the whole install. Start the interface:
|
|
141
|
+
|
|
142
|
+
```bash
|
|
143
|
+
visionforge gui
|
|
144
|
+
```
|
|
145
|
+
|
|
146
|
+
It opens on <http://127.0.0.1:8000>. The first screen asks your name and offers
|
|
147
|
+
a short guided tour of the interface.
|
|
148
|
+
|
|
149
|
+
<details>
|
|
150
|
+
<summary>Why PyTorch is installed separately</summary>
|
|
151
|
+
|
|
152
|
+
Its build has to match your hardware, and no dependency resolver can choose
|
|
153
|
+
between the CPU and CUDA wheels for you — on Windows, PyPI only carries the CPU
|
|
154
|
+
one, so that is what a plain install brings. `doctor --fix` replaces it with
|
|
155
|
+
the build your driver calls for.
|
|
156
|
+
|
|
157
|
+
Already know you want CUDA? Install torch **first**, from the PyTorch index, and
|
|
158
|
+
the package after — that skips downloading the CPU build at all:
|
|
159
|
+
|
|
160
|
+
```bash
|
|
161
|
+
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu128
|
|
162
|
+
pip install visionforge-studio
|
|
163
|
+
```
|
|
164
|
+
|
|
165
|
+
Available in place of `cu128`: `cu130` (driver 580+) · `cu126` · `cpu`. The legacy
|
|
166
|
+
`cu118` · `cu121` · `cu124` builds stop at torch 2.7 and miss its security fixes
|
|
167
|
+
(ADR-198); use them only with a driver too old for CUDA 12.6.
|
|
168
|
+
`cu128` is the broadest — it spans Turing (sm_75) through Blackwell, and is the
|
|
169
|
+
only one that runs on an RTX 50-series card at all.
|
|
170
|
+
|
|
171
|
+
To update later, `pip install --upgrade visionforge-studio` — it keeps the
|
|
172
|
+
torch you have.
|
|
173
|
+
|
|
174
|
+
**On PyPI the distribution is `visionforge-studio`** — the bare `visionforge`
|
|
175
|
+
name belongs to an unrelated project. The import name, the CLI command and the
|
|
176
|
+
project itself are still `visionforge`.
|
|
177
|
+
|
|
178
|
+
</details>
|
|
179
|
+
|
|
180
|
+
<details>
|
|
181
|
+
<summary>What comes with it</summary>
|
|
182
|
+
|
|
183
|
+
Everything the five tasks need is in that one install — the torchvision
|
|
184
|
+
detection backend, the extra `timm` backbones, Optuna-guided sweeps,
|
|
185
|
+
TensorBoard scalars, and the Roboflow / Kaggle / Hugging Face dataset
|
|
186
|
+
downloaders. They used to be optional extras; the split cost more than it
|
|
187
|
+
saved, and they add about 83 MB against the gigabytes PyTorch already pulls.
|
|
188
|
+
The one exception is the YOLO / RT-DETR backend, Ultralytics, licensed
|
|
189
|
+
AGPL-3.0: `pip install "visionforge-studio[detection]"` adds it.
|
|
190
|
+
|
|
191
|
+
Choose the PyTorch build for your hardware. The optional tracking and
|
|
192
|
+
export integrations below are on `development` (**next release**): for local
|
|
193
|
+
MLflow tracking, install `visionforge-studio[mlflow]` and enable MLflow in the
|
|
194
|
+
profile picker. From the directory where the GUI started, run
|
|
195
|
+
`mlflow ui --backend-store-uri sqlite:///outputs/mlflow/mlflow.db` to inspect
|
|
196
|
+
training runs of every task, epoch metrics and the final `run.json`; if an export
|
|
197
|
+
fails, open the run in History to retry it. To follow them on
|
|
198
|
+
Weights & Biases, install `visionforge-studio[wandb]`, turn W&B on in the
|
|
199
|
+
profile picker and save an API key there; runs go to a private project. To push
|
|
200
|
+
a run's checkpoint, model card and metrics to a private Hugging Face Hub
|
|
201
|
+
repository, install `visionforge-studio[hub]`, save a write token in the
|
|
202
|
+
profile and review the local preview in History before sending
|
|
203
|
+
([`docs/gui/hugging-face-hub.md`](https://github.com/marcus-vreis/VisionForge/blob/main/docs/gui/hugging-face-hub.md)).
|
|
204
|
+
To cite a run by DOI, open it in History and create a Zenodo draft: the
|
|
205
|
+
bundle goes up without this machine's paths, and you fill in authors and
|
|
206
|
+
license and publish on Zenodo. To keep runs next to your code, send the
|
|
207
|
+
report to a private GitHub repository: one folder per run, without weights.
|
|
208
|
+
To version the weights with git, open the run in History and version it
|
|
209
|
+
with DVC in a repository on this computer (`visionforge-studio[dvc]`):
|
|
210
|
+
`dvc metrics show` and `dvc plots show` then compare the runs.
|
|
211
|
+
|
|
212
|
+
</details>
|
|
213
|
+
|
|
214
|
+
<details>
|
|
215
|
+
<summary>Running with Docker</summary>
|
|
216
|
+
|
|
217
|
+
From a repository checkout, [`Dockerfile`](https://github.com/marcus-vreis/VisionForge/blob/main/Dockerfile) and
|
|
218
|
+
[`docker-compose.yml`](https://github.com/marcus-vreis/VisionForge/blob/main/docker-compose.yml) build the React SPA in Stage 1 and bake Python 3.13,
|
|
219
|
+
PyTorch (`CUDA_TAG=cu128` by default, or `VARIANT=cpu` on `ubuntu:22.04`) and
|
|
220
|
+
VisionForge into Stage 2:
|
|
221
|
+
|
|
222
|
+
```bash
|
|
223
|
+
docker compose up # NVIDIA GPU (requires driver + nvidia-container-toolkit)
|
|
224
|
+
docker compose --profile cpu up # CPU-only image
|
|
225
|
+
```
|
|
226
|
+
|
|
227
|
+
`./datasets` is mounted read-only (`/work/datasets:ro`) and `./outputs`,
|
|
228
|
+
`./user_models` and `./user_tasks` are mounted under `/work`. Inside the
|
|
229
|
+
container (`VISIONFORGE_CONTAINER=1`, `uid 1000`), type the mounted path (such
|
|
230
|
+
as `/work/datasets/my_dataset`) in the dataset field. See
|
|
231
|
+
[`docs/dev/DOCKER.md`](https://github.com/marcus-vreis/VisionForge/blob/main/docs/dev/DOCKER.md).
|
|
232
|
+
|
|
233
|
+
</details>
|
|
234
|
+
|
|
235
|
+
## First run
|
|
236
|
+
|
|
237
|
+
Point the dataset picker at a folder, pick a task tab, press *Train*. If you
|
|
238
|
+
have no dataset yet, the **⤓ datasets** button downloads one (CIFAR-10, for
|
|
239
|
+
example) already in the layout the task expects.
|
|
240
|
+
|
|
241
|
+
The interface is available in English and Portuguese (PT/EN switch in the header),
|
|
242
|
+
with a light, dark or automatic theme.
|
|
243
|
+
|
|
244
|
+
Every task is configured in the same six steps — Data, Strategy, Model,
|
|
245
|
+
Training, Image and augmentation, and Review. *Train* unlocks on the Review
|
|
246
|
+
step, and the experiment name is suggested from the configuration.
|
|
247
|
+
|
|
248
|
+
In the **Datasets** panel, the **Explore** tab (**next release**, on
|
|
249
|
+
`development`) inspects classification (`ImageFolder`) and detection (YOLO)
|
|
250
|
+
datasets before training: class and split counts, a paginated thumbnail gallery
|
|
251
|
+
(with bounding boxes drawn for detection), and a background integrity scan for
|
|
252
|
+
unreadable files and exact SHA-256 duplicates across splits. From the same tab,
|
|
253
|
+
you can incorporate new images, merge two datasets, resplit with the test set
|
|
254
|
+
locked by default, or convert between YOLO and COCO — always writing a new
|
|
255
|
+
sibling folder with `visionforge_dataset.json` and leaving the original dataset
|
|
256
|
+
untouched.
|
|
257
|
+
|
|
258
|
+
While a job runs, the **Results** tab plots loss and task metric curves epoch by
|
|
259
|
+
epoch, shows the learning rate and the best epoch so far, renders the validation
|
|
260
|
+
confusion matrix and optional live Grad-CAM (`training.live_gradcam_every`) for
|
|
261
|
+
classification, and lists completed runs during sweeps, comparisons,
|
|
262
|
+
cross-validation and replicates. Beside it, the **Code** tab (**next release**,
|
|
263
|
+
on `development`) reads the real training, validation and evaluation functions
|
|
264
|
+
from the installed package via `inspect` — for classification, regression,
|
|
265
|
+
segmentation, anomaly detection and `torchvision` detection — and highlights the
|
|
266
|
+
branch your configuration takes
|
|
267
|
+
([`docs/gui/codigo.md`](https://github.com/marcus-vreis/VisionForge/blob/main/docs/gui/codigo.md)).
|
|
268
|
+
|
|
269
|
+
To train classification on **Google Colab**, open the `colab` button in the
|
|
270
|
+
header: each profile logs in with its Google account via `google-colab-cli`
|
|
271
|
+
(installed into an isolated venv under `~/.visionforge/colab-cli`) and can
|
|
272
|
+
queue classification runs on a Colab GPU (`T4`, `L4`, `A100`, `H100`, `G4`) in
|
|
273
|
+
parallel with local training. The VM fetches the dataset from a public URL,
|
|
274
|
+
`torchvision`, Hugging Face, or — in the **next release** — a folder inside
|
|
275
|
+
Google Drive for desktop; epoch events stream live to the interface, the
|
|
276
|
+
finished run returns as a verified experiment bundle into History, and the VM
|
|
277
|
+
is released when the job ends, fails or is stopped
|
|
278
|
+
([`docs/gui/colab.md`](https://github.com/marcus-vreis/VisionForge/blob/main/docs/gui/colab.md)).
|
|
279
|
+
|
|
280
|
+
The step-by-step walkthrough — dataset download, first run, confidence
|
|
281
|
+
intervals, re-running from YAML — is in
|
|
282
|
+
[`docs/QUICKSTART.md`](https://github.com/marcus-vreis/VisionForge/blob/main/docs/QUICKSTART.md).
|
|
283
|
+
|
|
284
|
+
For automation, export the configuration as YAML from the Review step (or copy
|
|
285
|
+
one of the templates in
|
|
286
|
+
[`configs/`](https://github.com/marcus-vreis/VisionForge/tree/main/configs) and
|
|
287
|
+
point `data.base_dir` at your dataset) and run it from the CLI (in the **next
|
|
288
|
+
release**, a `strategy:` + `config:` YAML also runs `replicates`,
|
|
289
|
+
`replicated_comparison`, `sweep` and `cv` from the command line):
|
|
290
|
+
|
|
291
|
+
```bash
|
|
292
|
+
visionforge run experiment.yaml
|
|
293
|
+
```
|
|
294
|
+
|
|
295
|
+
## The five tasks
|
|
296
|
+
|
|
297
|
+
| Task | Models | Main metrics |
|
|
298
|
+
|---|---|---|
|
|
299
|
+
| **Classification** | ResNet, EfficientNet, ViT, Swin, ConvNeXt, VGG, AlexNet, timm | Accuracy, F1, AUC-ROC, confusion matrix |
|
|
300
|
+
| **Object detection** | YOLOv8–v26, RT-DETR, Faster R-CNN, SSD, RetinaNet | mAP@50, mAP@50-95 |
|
|
301
|
+
| **Image regression** | CNN backbones + linear head (CSV manifests) | MSE, RMSE, MAE, R² |
|
|
302
|
+
| **Semantic segmentation** | DeepLabV3, FCN, LR-ASPP, U-Net | mean IoU, Dice, pixel accuracy |
|
|
303
|
+
| **Anomaly detection** | Convolutional autoencoder, PatchCore | image AUROC, F1, threshold |
|
|
304
|
+
|
|
305
|
+
Every panel has the same shape: experiment name, strategy, model, training,
|
|
306
|
+
dataset, preprocessing. Learning what one task looks like teaches you all five.
|
|
307
|
+
|
|
308
|
+
Need something else? You can [add your own models and whole task
|
|
309
|
+
families](https://github.com/marcus-vreis/VisionForge/tree/main/docs/custom)
|
|
310
|
+
without touching the package.
|
|
311
|
+
|
|
312
|
+
## Results you can defend
|
|
313
|
+
|
|
314
|
+

|
|
315
|
+
|
|
316
|
+
- **Replicates** — the same config over N seeds, reported as
|
|
317
|
+
`mean ± 95% CI` instead of one lucky number.
|
|
318
|
+
- **K-fold cross-validation** — per-fold metrics plus mean ± std, with
|
|
319
|
+
fold-safe transforms. Each fold keeps the epoch that did best on its own fold,
|
|
320
|
+
and that same fold scores it: the mean is optimistic (there is no nested
|
|
321
|
+
split), and the LaTeX caption says so.
|
|
322
|
+
- **Sweeps** — grid, random or Optuna over any config field.
|
|
323
|
+
- **Paired significance tests** — compare configurations over the *same* seeds,
|
|
324
|
+
with a paired t or Wilcoxon test and the interval of the difference that
|
|
325
|
+
inverts it (t or Hodges-Lehmann), Cohen's `d_z` and
|
|
326
|
+
Holm-Bonferroni correction. It refuses to compare runs whose seeds do not
|
|
327
|
+
line up, and says when the seed count makes significance unreachable — so
|
|
328
|
+
"not significant" is never mistaken for "no effect".
|
|
329
|
+
- **Full provenance** — every run writes a `run.json` with the config, the
|
|
330
|
+
seed, the per-epoch history, the environment (torch, CUDA, GPU model) and a
|
|
331
|
+
dataset fingerprint, so "same data" is a checkable claim.
|
|
332
|
+
- **Paper-ready output** — every report is also written as a `booktabs` LaTeX
|
|
333
|
+
table, with notes stating what each interval covers.
|
|
334
|
+
In the **next release**, figures are written as PNG at 300 dpi by default; set
|
|
335
|
+
`output.figure_formats: [png, pdf, svg]` for vector files, and replicate sets
|
|
336
|
+
also leave a `replicates_strip` figure beside the table.
|
|
337
|
+
- **Data audits and verifiable bundles** — audit splits for unreadable images,
|
|
338
|
+
exact SHA-256 duplicates across splits and group leakage before training, and
|
|
339
|
+
export or verify a self-contained experiment archive with a SHA-256 manifest
|
|
340
|
+
without extracting it
|
|
341
|
+
([`docs/REPRODUCIBILITY.md`](https://github.com/marcus-vreis/VisionForge/blob/main/docs/REPRODUCIBILITY.md)):
|
|
342
|
+
|
|
343
|
+
```bash
|
|
344
|
+
visionforge audit-data datasets/my_dataset --groups groups.csv --strict-groups --json audit.json
|
|
345
|
+
visionforge bundle outputs/models/my_run evidence.zip
|
|
346
|
+
visionforge verify-bundle evidence.zip
|
|
347
|
+
```
|
|
348
|
+
|
|
349
|
+
- **Learning rate per model** (classification) — with
|
|
350
|
+
`model_comparison.lr_policy: suggested` each architecture trains at its own
|
|
351
|
+
suggested LR, and `grid_search.lr_mode: suggested_multiplier` reads the LR
|
|
352
|
+
axis as a multiplier of it. The effective LR of every trial is recorded in
|
|
353
|
+
`run.json`, in the reports and in Review. The default does not change, and
|
|
354
|
+
configs written before still load (`schema_version` 2).
|
|
355
|
+
|
|
356
|
+
After training: run history with config diffs, per-checkpoint testing on new
|
|
357
|
+
data, batch prediction to CSV, Grad-CAM, ONNX export, TensorBoard.
|
|
358
|
+
|
|
359
|
+
## Your workspace
|
|
360
|
+
|
|
361
|
+
VisionForge reads your own code from folders **next to wherever you run it** —
|
|
362
|
+
no repository needed, nothing to edit inside the package:
|
|
363
|
+
|
|
364
|
+
```
|
|
365
|
+
my-research/ ← run `visionforge gui` from here
|
|
366
|
+
├── user_models/ ← your architectures
|
|
367
|
+
├── user_tasks/ ← your task families
|
|
368
|
+
├── datasets/ ← whatever you point the picker at
|
|
369
|
+
└── outputs/ ← runs, checkpoints, reports, run.json
|
|
370
|
+
```
|
|
371
|
+
|
|
372
|
+
Run it from a different folder and it looks there instead — so keep one folder
|
|
373
|
+
per project. `visionforge doctor` always prints the paths it resolved. To
|
|
374
|
+
scaffold a custom task under `user_tasks/` (with `--package` for a package
|
|
375
|
+
directory or `--force` to overwrite):
|
|
376
|
+
|
|
377
|
+
```bash
|
|
378
|
+
visionforge new-task my_task
|
|
379
|
+
```
|
|
380
|
+
|
|
381
|
+
To keep everything on this computer, start with `visionforge gui --offline`
|
|
382
|
+
(or set `VISIONFORGE_OFFLINE=1`): any connection to another host fails, weight
|
|
383
|
+
downloads and the update check included.
|
|
384
|
+
|
|
385
|
+
## Checking the install
|
|
386
|
+
|
|
387
|
+
```bash
|
|
388
|
+
visionforge doctor # environment: driver, torch, workspace
|
|
389
|
+
visionforge selftest --quick # trains every task on synthetic data, ~15s
|
|
390
|
+
```
|
|
391
|
+
|
|
392
|
+
`selftest` starts the real API and trains through the same endpoints the
|
|
393
|
+
browser uses. It checks that the pipeline works, not that a model is good — one
|
|
394
|
+
epoch on synthetic data says nothing about accuracy.
|
|
395
|
+
|
|
396
|
+
## Documentation
|
|
397
|
+
|
|
398
|
+
- [`docs/INSTALL.md`](https://github.com/marcus-vreis/VisionForge/blob/main/docs/INSTALL.md) — install, step by step, per system (Portuguese)
|
|
399
|
+
- [`docs/QUICKSTART.md`](https://github.com/marcus-vreis/VisionForge/blob/main/docs/QUICKSTART.md) — first run, start to finish
|
|
400
|
+
- [`docs/DATASETS.md`](https://github.com/marcus-vreis/VisionForge/blob/main/docs/DATASETS.md) — dataset layouts and the download providers
|
|
401
|
+
- [`docs/guias/`](https://github.com/marcus-vreis/VisionForge/blob/main/docs/guias/comecando-do-zero.md) — three entry guides, one per reader: starting from zero, for people who already train models, and a tour of what VisionForge does (Portuguese)
|
|
402
|
+
- [`docs/gui/README.md`](https://github.com/marcus-vreis/VisionForge/blob/main/docs/gui/README.md) — index of the GUI guides, with screenshots: first training run, strategies, what to do after a run (Portuguese)
|
|
403
|
+
- [`docs/gui/codigo.md`](https://github.com/marcus-vreis/VisionForge/blob/main/docs/gui/codigo.md) — the Code tab and the functions inspected per task (Portuguese)
|
|
404
|
+
- [`docs/gui/colab.md`](https://github.com/marcus-vreis/VisionForge/blob/main/docs/gui/colab.md) — classification training on Google Colab and Google Drive datasets (Portuguese)
|
|
405
|
+
- [`docs/gui/tokens.md`](https://github.com/marcus-vreis/VisionForge/blob/main/docs/gui/tokens.md) — where tokens and keys are stored and how to remove them (Portuguese)
|
|
406
|
+
- [`docs/REPRODUCIBILITY.md`](https://github.com/marcus-vreis/VisionForge/blob/main/docs/REPRODUCIBILITY.md) — data audits, grouped CV, resume guarantees and experiment bundles (Portuguese)
|
|
407
|
+
- [`docs/custom/`](https://github.com/marcus-vreis/VisionForge/tree/main/docs/custom) — your own models and task families
|
|
408
|
+
- [`docs/ROADMAP.md`](https://github.com/marcus-vreis/VisionForge/blob/main/docs/ROADMAP.md) — what is on `development`, what is planned next and criteria for 1.0
|
|
409
|
+
- [`CHANGELOG.md`](https://github.com/marcus-vreis/VisionForge/blob/main/CHANGELOG.md) — what shipped in each release
|
|
410
|
+
|
|
411
|
+
Working on VisionForge itself?
|
|
412
|
+
[`CONTRIBUTING.md`](https://github.com/marcus-vreis/VisionForge/blob/main/CONTRIBUTING.md)
|
|
413
|
+
has the dev setup, and
|
|
414
|
+
[`docs/dev/`](https://github.com/marcus-vreis/VisionForge/tree/main/docs/dev)
|
|
415
|
+
has the architecture, the decision log and the validation record.
|
|
416
|
+
|
|
417
|
+
## Status
|
|
418
|
+
|
|
419
|
+
Under active development and usable for real work. Below 1.0 the config schema
|
|
420
|
+
and the HTTP API may change between minor releases — configs carry a
|
|
421
|
+
`schema_version` and are migrated on load, so an exported YAML keeps working.
|
|
422
|
+
|
|
423
|
+
Worth knowing before you start:
|
|
424
|
+
|
|
425
|
+
- **One local training at a time** (a Colab job can run alongside it in its own
|
|
426
|
+
queue), and submissions queue and start on their own. A job already training
|
|
427
|
+
can be stopped from the training screen (■ Stop): it ends at its next stop
|
|
428
|
+
point (the epoch, in a single run) and keeps what it saved. A custom task
|
|
429
|
+
that owns its loop cannot be stopped.
|
|
430
|
+
- **Built-in dataset download formats classification and detection.**
|
|
431
|
+
`torchvision` and Hugging Face deliver classification (`ImageFolder`),
|
|
432
|
+
Roboflow delivers detection (YOLO layout) and classification, and Kaggle
|
|
433
|
+
downloads any archive as published; regression, segmentation and anomaly need
|
|
434
|
+
a dataset already in their layout.
|
|
435
|
+
- **No K-fold for detection or anomaly**, by design: Ultralytics owns its
|
|
436
|
+
training loop, and an unsupervised fold without anomalies measures nothing.
|
|
437
|
+
|
|
438
|
+
Found something?
|
|
439
|
+
[Open an issue](https://github.com/marcus-vreis/VisionForge/issues/new/choose).
|
|
440
|
+
|
|
441
|
+
## Citing
|
|
442
|
+
|
|
443
|
+
If VisionForge is useful in your research, please cite it — see
|
|
444
|
+
[`CITATION.cff`](https://github.com/marcus-vreis/VisionForge/blob/main/CITATION.cff).
|
|
445
|
+
Every release is archived on Zenodo under
|
|
446
|
+
[doi:10.5281/zenodo.23425765](https://doi.org/10.5281/zenodo.23425765), which
|
|
447
|
+
always points to the latest version; each version also has its own DOI there.
|
|
448
|
+
|
|
449
|
+
## License
|
|
450
|
+
|
|
451
|
+
[MIT](https://github.com/marcus-vreis/VisionForge/blob/main/LICENSE)
|
|
452
|
+
|
|
453
|
+
## Third-party licenses
|
|
454
|
+
|
|
455
|
+
VisionForge is MIT, but some of what it installs is not.
|
|
456
|
+
[`THIRD_PARTY_NOTICES.md`](https://github.com/marcus-vreis/VisionForge/blob/main/THIRD_PARTY_NOTICES.md) lists them; the main one:
|
|
457
|
+
|
|
458
|
+
- **Ultralytics (AGPL-3.0)** is the YOLO / RT-DETR detection backend, and it
|
|
459
|
+
is installed only with the `detection` extra
|
|
460
|
+
(`pip install "visionforge-studio[detection]"`). If you distribute a program
|
|
461
|
+
built on it, or let users reach a modified version over a network, the
|
|
462
|
+
AGPL-3.0 asks you to publish that program's source under the same license;
|
|
463
|
+
Ultralytics offers an Enterprise License otherwise. The `torchvision`
|
|
464
|
+
detection backend, marked "no AGPL" in the form, does not use it.
|
|
465
|
+
- Smaller cases: `pi-heif` (bundles LGPL-3.0 libraries), `text-unidecode`
|
|
466
|
+
(Artistic or GPL-2.0+), `certifi` and `tqdm` (MPL-2.0).
|
|
467
|
+
- The GUI bundles three fonts under the SIL Open Font License 1.1.
|
|
468
|
+
|
|
469
|
+
## Network and privacy
|
|
470
|
+
|
|
471
|
+
VisionForge has no telemetry of its own. It goes online only for:
|
|
472
|
+
|
|
473
|
+
| When | Where | Turn it off |
|
|
474
|
+
|---|---|---|
|
|
475
|
+
| The GUI opens: update check | `pypi.org` (short timeout, cached, silent offline) | `VISIONFORGE_NO_UPDATE_CHECK=1` |
|
|
476
|
+
| A model with pretrained weights, first use | `download.pytorch.org`, `huggingface.co` (timm), GitHub releases (Ultralytics) | turn `pretrained` off |
|
|
477
|
+
| You download a dataset | Roboflow, Kaggle, Hugging Face | don't use the download |
|
|
478
|
+
| You export a run | MLflow server you set, Weights & Biases, DVC remote, Hugging Face Hub, GitHub, Zenodo | don't export |
|
|
479
|
+
| You connect Google Colab | PyPI (`google-colab-cli`), Google | don't connect |
|
|
480
|
+
| `visionforge doctor --fix` | `download.pytorch.org` | run `doctor` without `--fix` |
|
|
481
|
+
| `VISIONFORGE_DOCS=1` (API docs pages) | `cdn.jsdelivr.net`, Google Fonts | leave it unset |
|
|
482
|
+
|
|
483
|
+
Some bundled SDKs report usage on their own. Before importing them,
|
|
484
|
+
VisionForge sets `HF_HUB_DISABLE_TELEMETRY=1`, `DO_NOT_TRACK=1`,
|
|
485
|
+
`MLFLOW_DISABLE_TELEMETRY=true` and `DVC_NO_ANALYTICS=1` (only when you have
|
|
486
|
+
not set them) and switches off Ultralytics' usage events inside its own
|
|
487
|
+
process, without touching your Ultralytics settings file.
|
|
488
|
+
|
|
489
|
+
---
|
|
490
|
+
|
|
491
|
+
# Português (pt-BR)
|
|
492
|
+
|
|
493
|
+
**Treine, compare e defenda modelos de visão computacional na sua própria
|
|
494
|
+
GPU.** Cinco tipos de tarefa numa interface local-first, sem notebooks — roda na sua máquina; nada sai dela sem você pedir (datasets, Colab e exportadores são opt-in, com prévia local).
|
|
495
|
+
|
|
496
|
+
## Instalação
|
|
497
|
+
|
|
498
|
+
Você precisa de **Python 3.12+** e uns cinco minutos. Nada além disso — o
|
|
499
|
+
pacote publicado já vem com a interface pronta.
|
|
500
|
+
|
|
501
|
+
Um passo a passo por sistema (Windows, Linux, macOS), incluindo como instalar o
|
|
502
|
+
próprio Python e os erros mais comuns, está em [`docs/INSTALL.md`](https://github.com/marcus-vreis/VisionForge/blob/main/docs/INSTALL.md).
|
|
503
|
+
No Mac: só Apple Silicon (o PyTorch não tem versões para Mac Intel desde a 2.3),
|
|
504
|
+
o treino roda na CPU e o `doctor --fix` não é necessário.
|
|
505
|
+
|
|
506
|
+
```bash
|
|
507
|
+
mkdir minha-pesquisa && cd minha-pesquisa
|
|
508
|
+
python -m venv .venv
|
|
509
|
+
.venv\Scripts\activate
|
|
510
|
+
```
|
|
511
|
+
|
|
512
|
+
No Linux ou no macOS a última linha vira `source .venv/bin/activate`.
|
|
513
|
+
|
|
514
|
+
```bash
|
|
515
|
+
pip install visionforge-studio
|
|
516
|
+
```
|
|
517
|
+
|
|
518
|
+
Agora deixe o VisionForge olhar a sua máquina e instalar o PyTorch certo — ele
|
|
519
|
+
lê o driver da sua GPU, diz qual wheel você precisa e instala depois que você
|
|
520
|
+
responder `y`:
|
|
521
|
+
|
|
522
|
+
```bash
|
|
523
|
+
visionforge doctor --fix
|
|
524
|
+
```
|
|
525
|
+
|
|
526
|
+
A instalação acaba aqui. Para abrir a interface:
|
|
527
|
+
|
|
528
|
+
```bash
|
|
529
|
+
visionforge gui
|
|
530
|
+
```
|
|
531
|
+
|
|
532
|
+
Ela abre em <http://127.0.0.1:8000>. A primeira tela pergunta o seu nome e
|
|
533
|
+
oferece um guia rápido da interface.
|
|
534
|
+
|
|
535
|
+
<details>
|
|
536
|
+
<summary>Por que o PyTorch é instalado à parte</summary>
|
|
537
|
+
|
|
538
|
+
Porque a build dele precisa combinar com o seu hardware, e nenhum resolvedor de
|
|
539
|
+
dependências consegue escolher entre a wheel de CPU e a de CUDA por você — no
|
|
540
|
+
Windows o PyPI só tem a de CPU, então é ela que a instalação normal traz. O
|
|
541
|
+
`doctor --fix` troca pela build que o seu driver pede.
|
|
542
|
+
|
|
543
|
+
Já sabe que quer CUDA? Instale o torch **antes**, pelo índice do PyTorch, e o
|
|
544
|
+
pacote depois — assim a build de CPU nem é baixada:
|
|
545
|
+
|
|
546
|
+
```bash
|
|
547
|
+
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu128
|
|
548
|
+
pip install visionforge-studio
|
|
549
|
+
```
|
|
550
|
+
|
|
551
|
+
Disponíveis no lugar de `cu128`: `cu130` (driver 580+) · `cu126` · `cpu`. As builds
|
|
552
|
+
legadas `cu118` · `cu121` · `cu124` param no torch 2.7 e ficam sem as correções de
|
|
553
|
+
segurança (ADR-198); use só com um driver antigo demais para o CUDA 12.6.
|
|
554
|
+
O `cu128` é o mais abrangente — vai de Turing (sm_75) até Blackwell, e é o único
|
|
555
|
+
que roda numa RTX série 50.
|
|
556
|
+
|
|
557
|
+
Para atualizar depois, `pip install --upgrade visionforge-studio` — ele mantém o
|
|
558
|
+
torch que você já tem.
|
|
559
|
+
|
|
560
|
+
**No PyPI a distribuição chama `visionforge-studio`** — o nome `visionforge`
|
|
561
|
+
puro pertence a um projeto sem relação com este. O nome de import, o comando e
|
|
562
|
+
o projeto continuam sendo `visionforge`.
|
|
563
|
+
|
|
564
|
+
</details>
|
|
565
|
+
|
|
566
|
+
<details>
|
|
567
|
+
<summary>O que já vem junto</summary>
|
|
568
|
+
|
|
569
|
+
Tudo que as cinco tarefas precisam está naquela instalação: o backend de
|
|
570
|
+
detecção torchvision, os backbones extras do `timm`, as buscas guiadas por
|
|
571
|
+
Optuna, os escalares do TensorBoard e os downloaders de dataset do Roboflow,
|
|
572
|
+
Kaggle e Hugging Face. Eram extras opcionais; a separação custava mais do que
|
|
573
|
+
economizava, e eles somam uns 83 MB contra os gigabytes que o PyTorch já baixa.
|
|
574
|
+
A exceção é o backend YOLO / RT-DETR, o Ultralytics, com licença AGPL-3.0:
|
|
575
|
+
`pip install "visionforge-studio[detection]"` o adiciona.
|
|
576
|
+
|
|
577
|
+
Escolha a versão do PyTorch compatível com o seu hardware. As integrações
|
|
578
|
+
opcionais de acompanhamento e exportação abaixo estão na branch `development`
|
|
579
|
+
(**próxima release**): para acompanhar
|
|
580
|
+
treinos de qualquer tarefa localmente, instale `visionforge-studio[mlflow]` e ative
|
|
581
|
+
MLflow no seletor de perfis. Na pasta em que abriu a GUI, execute
|
|
582
|
+
`mlflow ui --backend-store-uri sqlite:///outputs/mlflow/mlflow.db` para ver
|
|
583
|
+
os parâmetros, as métricas por época e o `run.json` final. Se a exportação
|
|
584
|
+
falhar, abra a execução no Histórico para tentar novamente. Para acompanhar
|
|
585
|
+
no Weights & Biases, instale `visionforge-studio[wandb]`, ligue o W&B no
|
|
586
|
+
seletor de perfis e salve lá uma chave de API; os treinos vão para um projeto
|
|
587
|
+
privado. Para enviar o checkpoint, o model card e as métricas de um treino a um
|
|
588
|
+
repositório privado do Hugging Face Hub, instale `visionforge-studio[hub]`,
|
|
589
|
+
salve um token de escrita no perfil e confira a prévia local no Histórico antes
|
|
590
|
+
de enviar
|
|
591
|
+
([`docs/gui/hugging-face-hub.md`](https://github.com/marcus-vreis/VisionForge/blob/main/docs/gui/hugging-face-hub.md)).
|
|
592
|
+
Para citar um treino por DOI, abra-o no Histórico e crie um rascunho
|
|
593
|
+
no Zenodo: o bundle sobe sem os caminhos desta máquina, e você preenche
|
|
594
|
+
autores e licença e publica no próprio Zenodo. Para guardar os treinos junto
|
|
595
|
+
do código, envie o relatório a um repositório privado do GitHub: uma pasta
|
|
596
|
+
por treino, sem os pesos. Para versionar os pesos junto do git, abra o
|
|
597
|
+
treino no Histórico e versione com DVC num repositório deste computador
|
|
598
|
+
(`visionforge-studio[dvc]`); depois, `dvc metrics show` e `dvc plots show`
|
|
599
|
+
comparam os treinos.
|
|
600
|
+
|
|
601
|
+
</details>
|
|
602
|
+
|
|
603
|
+
<details>
|
|
604
|
+
<summary>Rodando com Docker</summary>
|
|
605
|
+
|
|
606
|
+
A partir de um clone do repositório, o [`Dockerfile`](https://github.com/marcus-vreis/VisionForge/blob/main/Dockerfile) e o
|
|
607
|
+
[`docker-compose.yml`](https://github.com/marcus-vreis/VisionForge/blob/main/docker-compose.yml) compilam a SPA React no primeiro estágio e empacotam
|
|
608
|
+
Python 3.13, PyTorch (`CUDA_TAG=cu128` por padrão, ou `VARIANT=cpu` sobre
|
|
609
|
+
`ubuntu:22.04`) e o VisionForge no segundo estágio:
|
|
610
|
+
|
|
611
|
+
```bash
|
|
612
|
+
docker compose up # GPU NVIDIA (exige driver + nvidia-container-toolkit)
|
|
613
|
+
docker compose --profile cpu up # imagem só CPU
|
|
614
|
+
```
|
|
615
|
+
|
|
616
|
+
`./datasets` entra como somente leitura (`/work/datasets:ro`) e `./outputs`,
|
|
617
|
+
`./user_models` e `./user_tasks` são montadas em `/work`. Dentro do container
|
|
618
|
+
(`VISIONFORGE_CONTAINER=1`, `uid 1000`), digite o caminho montado (como
|
|
619
|
+
`/work/datasets/meu_dataset`) no campo de dataset. Veja
|
|
620
|
+
[`docs/dev/DOCKER.md`](https://github.com/marcus-vreis/VisionForge/blob/main/docs/dev/DOCKER.md).
|
|
621
|
+
|
|
622
|
+
</details>
|
|
623
|
+
|
|
624
|
+
## Primeiro treino
|
|
625
|
+
|
|
626
|
+
Aponte o seletor de dataset para uma pasta, escolha a aba da tarefa e clique em
|
|
627
|
+
*Treinar*. Se você ainda não tem dataset, o botão **⤓ datasets** baixa um
|
|
628
|
+
(CIFAR-10, por exemplo) já no formato que a tarefa espera.
|
|
629
|
+
|
|
630
|
+
A interface está em português e inglês (seletor PT/EN no cabeçalho), com tema
|
|
631
|
+
claro, escuro ou automático.
|
|
632
|
+
|
|
633
|
+
Toda tarefa é configurada nas mesmas seis etapas — Dados, Estratégia, Modelo,
|
|
634
|
+
Treinamento, Imagem e aumento, e Revisão. O botão *Treinar* destrava na Revisão
|
|
635
|
+
e o nome do experimento vem sugerido a partir da configuração.
|
|
636
|
+
|
|
637
|
+
No painel **Datasets**, a aba **Explorar** (**próxima release**, em
|
|
638
|
+
`development`) inspeciona datasets de classificação (`ImageFolder`) e detecção
|
|
639
|
+
(YOLO) antes do treino: contagem por classe e divisão, galeria paginada de
|
|
640
|
+
miniaturas (com caixas desenhadas na detecção) e varredura de integridade em
|
|
641
|
+
segundo plano para arquivos ilegíveis e cópias exatas por SHA-256 entre
|
|
642
|
+
divisões. Na mesma aba dá para incorporar novas imagens, juntar dois datasets,
|
|
643
|
+
refazer a divisão mantendo o conjunto de teste travado por padrão ou converter
|
|
644
|
+
entre YOLO e COCO — sempre gravando uma pasta irmã nova com
|
|
645
|
+
`visionforge_dataset.json`, sem alterar o dataset original.
|
|
646
|
+
|
|
647
|
+
Enquanto o treino roda, a aba **Resultados** desenha as curvas de loss e da
|
|
648
|
+
métrica da tarefa época a época, mostra o learning rate e a melhor época até ali,
|
|
649
|
+
exibe a matriz de confusão da validação e o Grad-CAM ao vivo opcional
|
|
650
|
+
(`training.live_gradcam_every`) na classificação, e lista as execuções já
|
|
651
|
+
concluídas em buscas, comparações, validação cruzada e réplicas. Ao lado dela, a
|
|
652
|
+
aba **Código** (**próxima release**, em `development`) lê do pacote instalado,
|
|
653
|
+
com `inspect`, as funções reais de treino, validação e avaliação — em
|
|
654
|
+
classificação, regressão, segmentação, detecção de anomalias e detecção
|
|
655
|
+
`torchvision` — e marca o caminho que a sua configuração segue
|
|
656
|
+
([`docs/gui/codigo.md`](https://github.com/marcus-vreis/VisionForge/blob/main/docs/gui/codigo.md)).
|
|
657
|
+
|
|
658
|
+
Para treinar classificação no **Google Colab**, abra o botão `colab` no
|
|
659
|
+
cabeçalho: cada perfil entra com a sua conta Google pela `google-colab-cli`
|
|
660
|
+
(instalada num ambiente isolado em `~/.visionforge/colab-cli`) e enfileira
|
|
661
|
+
treinos de classificação numa GPU do Colab (`T4`, `L4`, `A100`, `H100`, `G4`)
|
|
662
|
+
em paralelo com a fila local. A máquina virtual busca o dataset de uma URL
|
|
663
|
+
pública, do `torchvision`, do Hugging Face ou — na **próxima release** — de uma
|
|
664
|
+
pasta dentro do Google Drive para desktop; os eventos de cada época chegam ao
|
|
665
|
+
vivo na interface, a execução concluída volta como bundle verificado para o
|
|
666
|
+
Histórico e a máquina é liberada quando o trabalho termina, falha ou é parado
|
|
667
|
+
([`docs/gui/colab.md`](https://github.com/marcus-vreis/VisionForge/blob/main/docs/gui/colab.md)).
|
|
668
|
+
|
|
669
|
+
O passo a passo completo — baixar dataset, primeiro treino, intervalos de
|
|
670
|
+
confiança, repetir a partir do YAML — está em
|
|
671
|
+
[`docs/QUICKSTART.md`](https://github.com/marcus-vreis/VisionForge/blob/main/docs/QUICKSTART.md).
|
|
672
|
+
|
|
673
|
+
Para automação, exporte a configuração como YAML na etapa de Revisão (ou copie
|
|
674
|
+
um dos modelos em
|
|
675
|
+
[`configs/`](https://github.com/marcus-vreis/VisionForge/tree/main/configs) e
|
|
676
|
+
aponte `data.base_dir` para o seu dataset) e rode pelo CLI (na **próxima
|
|
677
|
+
release**, um YAML com `strategy:` + `config:` também executa `replicates`,
|
|
678
|
+
`replicated_comparison`, `sweep` e `cv` pela linha de comando):
|
|
679
|
+
|
|
680
|
+
```bash
|
|
681
|
+
visionforge run experiment.yaml
|
|
682
|
+
```
|
|
683
|
+
|
|
684
|
+
## As cinco tarefas
|
|
685
|
+
|
|
686
|
+
| Tarefa | Modelos | Métricas principais |
|
|
687
|
+
|---|---|---|
|
|
688
|
+
| **Classificação** | ResNet, EfficientNet, ViT, Swin, ConvNeXt, VGG, AlexNet, timm | Acurácia, F1, AUC-ROC, matriz de confusão |
|
|
689
|
+
| **Detecção de objetos** | YOLOv8–v26, RT-DETR, Faster R-CNN, SSD, RetinaNet | mAP@50, mAP@50-95 |
|
|
690
|
+
| **Regressão em imagens** | Backbones CNN + cabeça linear (manifesto CSV) | MSE, RMSE, MAE, R² |
|
|
691
|
+
| **Segmentação semântica** | DeepLabV3, FCN, LR-ASPP, U-Net | IoU médio, Dice, acurácia por pixel |
|
|
692
|
+
| **Detecção de anomalias** | Autoencoder convolucional, PatchCore | AUROC por imagem, F1, limiar |
|
|
693
|
+
|
|
694
|
+
Todos os painéis têm a mesma forma: nome do experimento, estratégia, modelo,
|
|
695
|
+
treinamento, dataset, pré-processamento. Aprender uma tarefa ensina as cinco.
|
|
696
|
+
|
|
697
|
+
Precisa de outra coisa? Dá para [adicionar seus próprios modelos e até tarefas
|
|
698
|
+
inteiras](https://github.com/marcus-vreis/VisionForge/tree/main/docs/custom)
|
|
699
|
+
sem tocar no pacote.
|
|
700
|
+
|
|
701
|
+
## Resultados que você consegue defender
|
|
702
|
+
|
|
703
|
+
- **Réplicas** — o mesmo config em N sementes, reportado como
|
|
704
|
+
`média ± IC 95%` em vez de um número de sorte.
|
|
705
|
+
- **Validação cruzada K-fold** — métrica por fold mais média ± desvio, com
|
|
706
|
+
transformações que não vazam entre folds. Cada fold guarda a época que foi
|
|
707
|
+
melhor no próprio fold, e esse mesmo fold a avalia: a média é otimista (não há
|
|
708
|
+
divisão aninhada), e a legenda do LaTeX diz isso.
|
|
709
|
+
- **Buscas** — grid, aleatória ou Optuna sobre qualquer campo do config.
|
|
710
|
+
- **Testes de significância pareados** — compare configurações nas *mesmas*
|
|
711
|
+
sementes, com teste t pareado ou Wilcoxon e o IC da diferença que inverte
|
|
712
|
+
esse teste (t ou Hodges-Lehmann), `d_z` de Cohen e
|
|
713
|
+
correção de Holm-Bonferroni. Ele se recusa a comparar execuções cujas
|
|
714
|
+
sementes não batem, e avisa quando o número de sementes torna a significância
|
|
715
|
+
inalcançável — para que "não significativo" nunca seja lido como "sem
|
|
716
|
+
efeito".
|
|
717
|
+
- **Procedência completa** — toda execução escreve um `run.json` com o config,
|
|
718
|
+
a semente, o histórico por época, o ambiente (torch, CUDA, modelo da GPU) e
|
|
719
|
+
uma impressão digital do dataset, para que "mesmos dados" seja uma afirmação
|
|
720
|
+
verificável.
|
|
721
|
+
- **Saída pronta para artigo** — todo relatório sai também como tabela LaTeX
|
|
722
|
+
`booktabs`, com notas dizendo o que cada intervalo cobre. Na **próxima
|
|
723
|
+
release**, as figuras saem em PNG a 300 dpi por padrão; defina
|
|
724
|
+
`output.figure_formats: [png, pdf, svg]` para arquivos vetoriais, e conjuntos
|
|
725
|
+
de réplicas também deixam uma figura `replicates_strip` ao lado da tabela.
|
|
726
|
+
- **Auditoria de dados e pacotes verificáveis** — audite as divisões atrás de
|
|
727
|
+
imagens ilegíveis, cópias exatas por SHA-256 entre splits e vazamento de
|
|
728
|
+
grupos antes de treinar, e exporte ou verifique um pacote do experimento com
|
|
729
|
+
manifesto SHA-256 sem precisar extraí-lo
|
|
730
|
+
([`docs/REPRODUCIBILITY.md`](https://github.com/marcus-vreis/VisionForge/blob/main/docs/REPRODUCIBILITY.md)):
|
|
731
|
+
|
|
732
|
+
```bash
|
|
733
|
+
visionforge audit-data datasets/meu_dataset --groups groups.csv --strict-groups --json auditoria.json
|
|
734
|
+
visionforge bundle outputs/models/meu_run evidencia.zip
|
|
735
|
+
visionforge verify-bundle evidencia.zip
|
|
736
|
+
```
|
|
737
|
+
|
|
738
|
+
- **Learning rate por modelo** (classificação) — com
|
|
739
|
+
`model_comparison.lr_policy: suggested` cada arquitetura treina com o seu LR
|
|
740
|
+
sugerido, e `grid_search.lr_mode: suggested_multiplier` lê o eixo de LR como
|
|
741
|
+
multiplicador dele. O LR efetivo de cada trial fica no `run.json`, nos
|
|
742
|
+
relatórios e na Revisão. O padrão não muda, e configs antigas carregam iguais
|
|
743
|
+
(`schema_version` 2).
|
|
744
|
+
|
|
745
|
+
Depois do treino: histórico com diff de config, teste de checkpoints em dados
|
|
746
|
+
novos, predição em lote para CSV, Grad-CAM, exportação ONNX, TensorBoard.
|
|
747
|
+
|
|
748
|
+
## Sua pasta de trabalho
|
|
749
|
+
|
|
750
|
+
O VisionForge lê o seu código de pastas **ao lado de onde você o executa** —
|
|
751
|
+
sem repositório, sem editar nada dentro do pacote:
|
|
752
|
+
|
|
753
|
+
```
|
|
754
|
+
minha-pesquisa/ ← rode `visionforge gui` daqui
|
|
755
|
+
├── user_models/ ← suas arquiteturas
|
|
756
|
+
├── user_tasks/ ← suas tarefas
|
|
757
|
+
├── datasets/ ← o que você apontar no seletor
|
|
758
|
+
└── outputs/ ← execuções, checkpoints, relatórios, run.json
|
|
759
|
+
```
|
|
760
|
+
|
|
761
|
+
Se rodar de outra pasta, ele procura lá — então mantenha uma pasta por projeto.
|
|
762
|
+
O `visionforge doctor` sempre imprime os caminhos que resolveu. Para criar o
|
|
763
|
+
esqueleto de uma tarefa própria em `user_tasks/` (com `--package` para criar um
|
|
764
|
+
pacote ou `--force` para sobrescrever):
|
|
765
|
+
|
|
766
|
+
```bash
|
|
767
|
+
visionforge new-task minha_tarefa
|
|
768
|
+
```
|
|
769
|
+
|
|
770
|
+
Para manter tudo neste computador, use `visionforge gui --offline` (ou
|
|
771
|
+
`VISIONFORGE_OFFLINE=1`): qualquer conexão com outro host falha, inclusive o
|
|
772
|
+
download de pesos e a busca por atualização.
|
|
773
|
+
|
|
774
|
+
## Conferindo a instalação
|
|
775
|
+
|
|
776
|
+
```bash
|
|
777
|
+
visionforge doctor # ambiente: driver, torch, pastas
|
|
778
|
+
visionforge selftest --quick # treina todas as tarefas em dados sintéticos, ~15s
|
|
779
|
+
```
|
|
780
|
+
|
|
781
|
+
O `selftest` sobe a API de verdade e treina pelos mesmos endpoints que o
|
|
782
|
+
navegador usa. Ele confere que o caminho funciona, não que o modelo é bom — uma
|
|
783
|
+
época em dado sintético não diz nada sobre acurácia.
|
|
784
|
+
|
|
785
|
+
## Documentação
|
|
786
|
+
|
|
787
|
+
- [`docs/INSTALL.md`](https://github.com/marcus-vreis/VisionForge/blob/main/docs/INSTALL.md) — instalação passo a passo, por sistema
|
|
788
|
+
- [`docs/QUICKSTART.md`](https://github.com/marcus-vreis/VisionForge/blob/main/docs/QUICKSTART.md) — o primeiro treino, do início ao fim
|
|
789
|
+
- [`docs/DATASETS.md`](https://github.com/marcus-vreis/VisionForge/blob/main/docs/DATASETS.md) — formatos de dataset e os provedores de download
|
|
790
|
+
- [`docs/guias/`](https://github.com/marcus-vreis/VisionForge/blob/main/docs/guias/comecando-do-zero.md) — três guias de entrada, um por leitor: começando do zero, para quem já treina modelos e o que o VisionForge faz
|
|
791
|
+
- [`docs/gui/README.md`](https://github.com/marcus-vreis/VisionForge/blob/main/docs/gui/README.md) — índice dos guias da interface, com prints: primeiro treino, estratégias e o que fazer depois de um treino
|
|
792
|
+
- [`docs/gui/codigo.md`](https://github.com/marcus-vreis/VisionForge/blob/main/docs/gui/codigo.md) — a aba Código e as funções inspecionadas em cada tarefa
|
|
793
|
+
- [`docs/gui/colab.md`](https://github.com/marcus-vreis/VisionForge/blob/main/docs/gui/colab.md) — treino de classificação no Google Colab e datasets no Google Drive
|
|
794
|
+
- [`docs/gui/tokens.md`](https://github.com/marcus-vreis/VisionForge/blob/main/docs/gui/tokens.md) — onde ficam os tokens e chaves e como removê-los
|
|
795
|
+
- [`docs/REPRODUCIBILITY.md`](https://github.com/marcus-vreis/VisionForge/blob/main/docs/REPRODUCIBILITY.md) — auditoria dos dados, CV por grupos, limites da retomada e exportação de registros
|
|
796
|
+
- [`docs/custom/`](https://github.com/marcus-vreis/VisionForge/tree/main/docs/custom) — seus próprios modelos e tarefas
|
|
797
|
+
- [`docs/ROADMAP.md`](https://github.com/marcus-vreis/VisionForge/blob/main/docs/ROADMAP.md) — o que está em `development`, próximos passos e critérios para a 1.0
|
|
798
|
+
- [`CHANGELOG.md`](https://github.com/marcus-vreis/VisionForge/blob/main/CHANGELOG.md) — o que entrou em cada versão
|
|
799
|
+
|
|
800
|
+
Trabalhando no próprio VisionForge? O
|
|
801
|
+
[`CONTRIBUTING.md`](https://github.com/marcus-vreis/VisionForge/blob/main/CONTRIBUTING.md)
|
|
802
|
+
traz o ambiente de desenvolvimento, e
|
|
803
|
+
[`docs/dev/`](https://github.com/marcus-vreis/VisionForge/tree/main/docs/dev)
|
|
804
|
+
reúne a arquitetura, o registro de decisões e a validação.
|
|
805
|
+
|
|
806
|
+
## Situação atual
|
|
807
|
+
|
|
808
|
+
Em desenvolvimento ativo e já utilizável para trabalho de verdade. Abaixo da
|
|
809
|
+
1.0 o schema de config e a API HTTP podem mudar entre versões menores — os
|
|
810
|
+
configs carregam `schema_version` e são migrados ao abrir, então um YAML
|
|
811
|
+
exportado antes continua funcionando.
|
|
812
|
+
|
|
813
|
+
Vale saber antes de começar:
|
|
814
|
+
|
|
815
|
+
- **Um treino local por vez** (um treino no Colab pode rodar em paralelo, na
|
|
816
|
+
fila da própria conta), e os envios entram numa fila e começam sozinhos. Um
|
|
817
|
+
treino já em andamento pode ser parado pela tela de treino (■ Parar): ele
|
|
818
|
+
encerra no próximo ponto de parada (a época, num treino simples) e mantém o que
|
|
819
|
+
salvou. Uma tarefa custom com laço próprio não pode ser parada.
|
|
820
|
+
- **O download integrado formata classificação e detecção.** `torchvision` e
|
|
821
|
+
Hugging Face entregam classificação (`ImageFolder`), o Roboflow entrega
|
|
822
|
+
detecção (layout YOLO) e classificação, e o Kaggle baixa o arquivo como o
|
|
823
|
+
autor publicou; regressão, segmentação e anomalia precisam de um dataset já
|
|
824
|
+
no formato delas.
|
|
825
|
+
- **Sem K-fold em detecção e anomalia**, de propósito: o Ultralytics é dono do
|
|
826
|
+
próprio laço de treino, e um fold de validação sem anomalias não mede nada.
|
|
827
|
+
|
|
828
|
+
Achou algum problema?
|
|
829
|
+
[Abra uma issue](https://github.com/marcus-vreis/VisionForge/issues/new/choose).
|
|
830
|
+
|
|
831
|
+
## Como citar
|
|
832
|
+
|
|
833
|
+
Se o VisionForge for útil na sua pesquisa, cite o projeto — veja o
|
|
834
|
+
[`CITATION.cff`](https://github.com/marcus-vreis/VisionForge/blob/main/CITATION.cff).
|
|
835
|
+
Toda release é arquivada no Zenodo em
|
|
836
|
+
[doi:10.5281/zenodo.23425765](https://doi.org/10.5281/zenodo.23425765), que
|
|
837
|
+
aponta sempre para a versão mais recente; cada versão também tem o seu próprio
|
|
838
|
+
DOI lá.
|
|
839
|
+
|
|
840
|
+
## Licença
|
|
841
|
+
|
|
842
|
+
[MIT](https://github.com/marcus-vreis/VisionForge/blob/main/LICENSE)
|
|
843
|
+
|
|
844
|
+
## Licenças de terceiros
|
|
845
|
+
|
|
846
|
+
O VisionForge é MIT, mas parte do que ele instala não é.
|
|
847
|
+
[`THIRD_PARTY_NOTICES.md`](https://github.com/marcus-vreis/VisionForge/blob/main/THIRD_PARTY_NOTICES.md) lista esses pacotes; o principal:
|
|
848
|
+
|
|
849
|
+
- **Ultralytics (AGPL-3.0)** é o backend de detecção YOLO / RT-DETR, e só é
|
|
850
|
+
instalado com o extra `detection`
|
|
851
|
+
(`pip install "visionforge-studio[detection]"`). Quem distribui um programa
|
|
852
|
+
feito sobre ele, ou deixa usuários acessarem uma versão modificada pela rede,
|
|
853
|
+
precisa publicar o código desse programa sob a mesma licença; fora disso, a
|
|
854
|
+
Ultralytics oferece uma Enterprise License. O backend de detecção
|
|
855
|
+
`torchvision`, marcado "sem AGPL" no formulário, não usa o Ultralytics.
|
|
856
|
+
- Casos menores: `pi-heif` (traz bibliotecas LGPL-3.0), `text-unidecode`
|
|
857
|
+
(Artistic ou GPL-2.0+), `certifi` e `tqdm` (MPL-2.0).
|
|
858
|
+
- A interface traz três fontes sob a SIL Open Font License 1.1.
|
|
859
|
+
|
|
860
|
+
## Rede e privacidade
|
|
861
|
+
|
|
862
|
+
O VisionForge não tem telemetria própria. Ele só acessa a internet em:
|
|
863
|
+
|
|
864
|
+
| Quando | Para onde | Como desligar |
|
|
865
|
+
|---|---|---|
|
|
866
|
+
| A interface abre: checagem de atualização | `pypi.org` (timeout curto, cache, silenciosa offline) | `VISIONFORGE_NO_UPDATE_CHECK=1` |
|
|
867
|
+
| Modelo com pesos pré-treinados, primeiro uso | `download.pytorch.org`, `huggingface.co` (timm), releases do GitHub (Ultralytics) | desligue `pretrained` |
|
|
868
|
+
| Você baixa um dataset | Roboflow, Kaggle, Hugging Face | não use o download |
|
|
869
|
+
| Você exporta um run | servidor MLflow que você configurou, Weights & Biases, remoto do DVC, Hugging Face Hub, GitHub, Zenodo | não exporte |
|
|
870
|
+
| Você conecta o Google Colab | PyPI (`google-colab-cli`), Google | não conecte |
|
|
871
|
+
| `visionforge doctor --fix` | `download.pytorch.org` | rode o `doctor` sem `--fix` |
|
|
872
|
+
| `VISIONFORGE_DOCS=1` (páginas da documentação da API) | `cdn.jsdelivr.net`, Google Fonts | deixe sem definir |
|
|
873
|
+
|
|
874
|
+
Alguns SDKs embutidos reportam uso por conta própria. Antes de importá-los, o
|
|
875
|
+
VisionForge define `HF_HUB_DISABLE_TELEMETRY=1`, `DO_NOT_TRACK=1`,
|
|
876
|
+
`MLFLOW_DISABLE_TELEMETRY=true` e `DVC_NO_ANALYTICS=1` (só quando você não
|
|
877
|
+
definiu) e desliga os eventos de uso do Ultralytics dentro do próprio processo,
|
|
878
|
+
sem mexer no arquivo de configurações do Ultralytics.
|