learning-loop-node 0.27.1__tar.gz → 0.27.2__tar.gz

{learning_loop_node-0.27.1 → learning_loop_node-0.27.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: learning_loop_node
-Version: 0.27.1
+Version: 0.27.2
 Summary: Python Library for Nodes which connect to the Zauberzeug Learning Loop
 Author-email: Zauberzeug GmbH <info@zauberzeug.com>
 License: MIT
@@ -61,6 +61,7 @@ You can configure connection to our Learning Loop by specifying the following en
 | MIN_UNCERTAIN_THRESHOLD  | -            | smallest confidence (float) at which auto-upload will happen | Detector (opt.)           | 0.3          |
 | MAX_UNCERTAIN_THRESHOLD  | -            | largest confidence (float) at which auto-upload will happen  | Detector (opt.)           | 0.6          |
 | MAX_OUTBOX_SIZE          | -            | max. number of non-priority images queued for upload         | Detector (opt.)           | 1000         |
+| CONF_THRESHOLD           | -            | minimum confidence of a detection, overriding the model's (`--conf-threshold`, detectors built with `node_parser(conf_threshold=True)`) | Detector (opt.) | model's, 0.2 |
 | EXCLUSIVE_MODEL_BUILD    | -            | Reject detections during update to save VRAM (set to 1)      | Detector (opt.)           | 0            |
 | INFERENCE_BATCH_SIZE     | -            | Batch size of trainer when calculating detections            | Trainer (opt.)            | 10           |
 | VRAM_LIMIT_GB            | -            | GPU memory (GB) a training may use; the batch size is probed against it (`--vram-limit-gb`, trainers built with `node_parser(vram_limit=True)`) | Trainer (opt.) | 0 (whole card) |
@@ -349,4 +350,4 @@ After a short time the converted model should be available as well.
 - Nodes add properties to `model.json`, which contains all the information which are needed by subsequent nodes. These are typically the properties:
   - `resolution`: resolution in which the model expects images (as `int`, since the resolution is mostly square - later, `resolution_x`` resolution_y` would also be conceivable or `resolutions` to give a list of possible resolutions)
   - `categories`: list of categories with name, id, (later also type), in the order in which they are used by the model -- this is neccessary to be robust about renamings
-  - `conf_threshold` (optional): the minimum confidence the trainer chose for this model's detections
+  - `conf_threshold` (optional): the minimum confidence the trainer chose for this model's detections; a detector filters at it unless its `CONF_THRESHOLD` is set, and at 0.2 for a model without one

{learning_loop_node-0.27.1 → learning_loop_node-0.27.2}/README.md

@@ -28,6 +28,7 @@ You can configure connection to our Learning Loop by specifying the following en
 | MIN_UNCERTAIN_THRESHOLD  | -            | smallest confidence (float) at which auto-upload will happen | Detector (opt.)           | 0.3          |
 | MAX_UNCERTAIN_THRESHOLD  | -            | largest confidence (float) at which auto-upload will happen  | Detector (opt.)           | 0.6          |
 | MAX_OUTBOX_SIZE          | -            | max. number of non-priority images queued for upload         | Detector (opt.)           | 1000         |
+| CONF_THRESHOLD           | -            | minimum confidence of a detection, overriding the model's (`--conf-threshold`, detectors built with `node_parser(conf_threshold=True)`) | Detector (opt.) | model's, 0.2 |
 | EXCLUSIVE_MODEL_BUILD    | -            | Reject detections during update to save VRAM (set to 1)      | Detector (opt.)           | 0            |
 | INFERENCE_BATCH_SIZE     | -            | Batch size of trainer when calculating detections            | Trainer (opt.)            | 10           |
 | VRAM_LIMIT_GB            | -            | GPU memory (GB) a training may use; the batch size is probed against it (`--vram-limit-gb`, trainers built with `node_parser(vram_limit=True)`) | Trainer (opt.) | 0 (whole card) |
@@ -316,4 +317,4 @@ After a short time the converted model should be available as well.
 - Nodes add properties to `model.json`, which contains all the information which are needed by subsequent nodes. These are typically the properties:
   - `resolution`: resolution in which the model expects images (as `int`, since the resolution is mostly square - later, `resolution_x`` resolution_y` would also be conceivable or `resolutions` to give a list of possible resolutions)
   - `categories`: list of categories with name, id, (later also type), in the order in which they are used by the model -- this is neccessary to be robust about renamings
-  - `conf_threshold` (optional): the minimum confidence the trainer chose for this model's detections
+  - `conf_threshold` (optional): the minimum confidence the trainer chose for this model's detections; a detector filters at it unless its `CONF_THRESHOLD` is set, and at 0.2 for a model without one

{learning_loop_node-0.27.1 → learning_loop_node-0.27.2}/learning_loop_node/data_classes/general.py

@@ -12,6 +12,8 @@ from ..enums import CategoryType
 
 KWONLY_SLOTS = {'kw_only': True, 'slots': True} if sys.version_info >= (3, 10) else {}
 
+DEFAULT_CONF_THRESHOLD = 0.2
+
 
 @dataclass(**KWONLY_SLOTS)
 class Category():
@@ -59,8 +61,9 @@ class ModelInformation():
         default=None, metadata={"description": "The path of the parent directory of the model in the file system."})
     model_size: Optional[str] = field(default=None, metadata={
                                       "description": "The size of the model (i.e. the specification or variant of the model architecture)."})
-    conf_threshold: Optional[float] = field(default=None, metadata={
-        "description": "The minimum confidence the trainer chose for the detections of this model, if it chose one."})
+    conf_threshold: float = field(default=DEFAULT_CONF_THRESHOLD, metadata={
+        "description": "The minimum confidence for the detections of this model: Loaded from the model if available. "
+                       "A detector node's own setting overrides it."})
 
     @property
     def context(self):
@@ -80,6 +83,8 @@ class ModelInformation():
             except Exception as exc:
                 raise Exception(f"could not read model information from file '{model_info_file_path}'") from exc
             try:
+                # a null is a value the writer did not determine, so it reads as if the key were absent
+                content = {key: value for key, value in content.items() if value is not None}
                 model_information = from_dict(data_class=ModelInformation, data=content)
                 model_information.model_root_path = model_root_path
             except Exception as exc:

{learning_loop_node-0.27.1 → learning_loop_node-0.27.2}/learning_loop_node/detector/detector_logic.py

@@ -10,6 +10,7 @@ class DetectorLogicFactory(Protocol):
 
     The factory controls how the detector is constructed — implementations
     can build synchronously or offload heavy work to a thread pool.
+    `model_info.conf_threshold` is the confidence the detector filters at.
     """
     @property
     def model_format(self) -> str: ...

{learning_loop_node-0.27.1 → learning_loop_node-0.27.2}/learning_loop_node/detector/detector_node.py

@@ -7,7 +7,7 @@ import shutil
 import subprocess
 import sys
 from collections.abc import Sequence
-from dataclasses import asdict, dataclass
+from dataclasses import asdict, dataclass, replace
 from datetime import datetime
 from typing import Dict, List, Literal, Optional, Union
 
@@ -50,9 +50,11 @@ from .rest import upload as rest_upload
 class DetectorNode(Node):
 
     def __init__(self, name: str, detector_factory: DetectorLogicFactory,
-                 uuid: Optional[str] = None, use_backdoor_controls: bool = False) -> None:
+                 uuid: Optional[str] = None, use_backdoor_controls: bool = False,
+                 conf_threshold: Optional[float] = None) -> None:
         super().__init__(name, uuid=uuid, node_type='detector', needs_login=False, needs_sio=False)
         self._detector_factory = detector_factory
+        self._conf_threshold = conf_threshold
         self.node_version: Optional[str] = os.environ.get('NODE_VERSION')
         self._detector: _DetectorState = _Initializing()
         self._exclusive_model_build: bool = os.environ.get('EXCLUSIVE_MODEL_BUILD', '0').lower() in ('1', 'true')
@@ -547,6 +549,10 @@ class DetectorNode(Node):
         if model_info is None:
             logging.warning('No model.json found in %s', model_dir)
             raise Exception('model.json not found')
+        if self._conf_threshold is not None:
+            model_info = replace(model_info, conf_threshold=self._conf_threshold)
+        self.log.info('Set confidence threshold for model %s to %s%s', model_info.version,
+                      model_info.conf_threshold, ' (set on the node)' if self._conf_threshold is not None else '')
 
         if self._exclusive_model_build and isinstance(self._detector, _ActiveDetector):
             async with self.detection_lock:

{learning_loop_node-0.27.1 → learning_loop_node-0.27.2}/learning_loop_node/helpers/entrypoint.py

@@ -12,13 +12,14 @@ from argparse import Action, Namespace
 import configargparse
 import uvicorn
 
+from ..data_classes.general import DEFAULT_CONF_THRESHOLD
 from ..trainer.batch_size import VRAM_LIMIT_GB_FLAG, VRAM_LIMIT_GB_HELP
 
 logger = logging.getLogger(__name__)
 
 
 def node_parser(*, description: str, legacy_env_prefix: str = '',
-                vram_limit: bool = False) -> configargparse.ArgumentParser:
+                vram_limit: bool = False, conf_threshold: bool = False) -> configargparse.ArgumentParser:
     """Build the parser for a node, pre-loaded with the settings every node has.
 
     :param legacy_env_prefix: A prefix an earlier version of this node required, e.g.
@@ -26,6 +27,9 @@ def node_parser(*, description: str, legacy_env_prefix: str = '',
         current name (``MY_DETECTOR_NODE_HOST``) and the prefix on the flag it was originally
         applied to (``MY_DETECTOR_HOST``). Leave empty for a node that never used one.
     :param vram_limit: Add :data:`~learning_loop_node.trainer.batch_size.VRAM_LIMIT_GB_FLAG`.
+    :param conf_threshold: Add ``--conf-threshold`` for the ``conf_threshold`` of a
+        :class:`~learning_loop_node.detector.detector_node.DetectorNode`. It has no default:
+        one would override the threshold the trainer chose for the model.
     """
     parser = _NodeArgumentParser(description=description, legacy_env_prefix=legacy_env_prefix)
     parser.add_argument('--host', default='0.0.0.0', env_var='NODE_HOST',
@@ -33,6 +37,10 @@ def node_parser(*, description: str, legacy_env_prefix: str = '',
     parser.add_argument('--port', type=int, default=80, env_var='NODE_PORT', help='Port to bind to')
     if vram_limit:
         parser.add_argument(VRAM_LIMIT_GB_FLAG, type=float, default=0, help=VRAM_LIMIT_GB_HELP)
+    if conf_threshold:
+        parser.add_argument('--conf-threshold', type=float, default=None,
+                            help='Minimum confidence for detections (default: the one the trainer chose for the '
+                                 f'model, else {DEFAULT_CONF_THRESHOLD})')
     return parser
 
 

{learning_loop_node-0.27.1 → learning_loop_node-0.27.2}/learning_loop_node/trainer/trainer_logic_generic.py

@@ -498,10 +498,11 @@ class TrainerLogicGeneric(ABC):
             raise CriticalError('The model is uploaded, but the loop never answered with its id') from e
 
     def _dump_model_json(self) -> str:
-        """Dumps the categories and the trainer's model metadata to a json file and returns the path to the file.
-        """
-        content = {'categories': [asdict(c) for c in self._training.categories],
-                   **self._get_model_metadata()} if self._training else None
+        """Dump the categories and the trainer's model metadata to a json file and return its path."""
+        content = None
+        if self._training:
+            content = {'categories': [asdict(c) for c in self._training.categories],
+                       **{key: value for key, value in self._get_model_metadata().items() if value is not None}}
         json_path = '/tmp/model.json'
         with open(json_path, 'w') as f:
             json.dump(content, f)
@@ -620,9 +621,11 @@ class TrainerLogicGeneric(ABC):
         raise NotImplementedError
 
     def _get_model_metadata(self) -> dict[str, Any]:
-        """Called when the files of `_get_latest_model_files` are uploaded.
-        Returns values the trainer determined for this model, merged into its model.json.
+        """Return the values the trainer determined for this model, to be merged into its model.json.
+
+        Called when the files of `_get_latest_model_files` are uploaded.
         The keys must be fields of `ModelInformation` (e.g. `conf_threshold`) so detectors can read them.
+        A value of `None` is left out of the file.
         """
         return {}
 

{learning_loop_node-0.27.1 → learning_loop_node-0.27.2}/learning_loop_node.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: learning_loop_node
-Version: 0.27.1
+Version: 0.27.2
 Summary: Python Library for Nodes which connect to the Zauberzeug Learning Loop
 Author-email: Zauberzeug GmbH <info@zauberzeug.com>
 License: MIT
@@ -61,6 +61,7 @@ You can configure connection to our Learning Loop by specifying the following en
 | MIN_UNCERTAIN_THRESHOLD  | -            | smallest confidence (float) at which auto-upload will happen | Detector (opt.)           | 0.3          |
 | MAX_UNCERTAIN_THRESHOLD  | -            | largest confidence (float) at which auto-upload will happen  | Detector (opt.)           | 0.6          |
 | MAX_OUTBOX_SIZE          | -            | max. number of non-priority images queued for upload         | Detector (opt.)           | 1000         |
+| CONF_THRESHOLD           | -            | minimum confidence of a detection, overriding the model's (`--conf-threshold`, detectors built with `node_parser(conf_threshold=True)`) | Detector (opt.) | model's, 0.2 |
 | EXCLUSIVE_MODEL_BUILD    | -            | Reject detections during update to save VRAM (set to 1)      | Detector (opt.)           | 0            |
 | INFERENCE_BATCH_SIZE     | -            | Batch size of trainer when calculating detections            | Trainer (opt.)            | 10           |
 | VRAM_LIMIT_GB            | -            | GPU memory (GB) a training may use; the batch size is probed against it (`--vram-limit-gb`, trainers built with `node_parser(vram_limit=True)`) | Trainer (opt.) | 0 (whole card) |
@@ -349,4 +350,4 @@ After a short time the converted model should be available as well.
 - Nodes add properties to `model.json`, which contains all the information which are needed by subsequent nodes. These are typically the properties:
   - `resolution`: resolution in which the model expects images (as `int`, since the resolution is mostly square - later, `resolution_x`` resolution_y` would also be conceivable or `resolutions` to give a list of possible resolutions)
   - `categories`: list of categories with name, id, (later also type), in the order in which they are used by the model -- this is neccessary to be robust about renamings
-  - `conf_threshold` (optional): the minimum confidence the trainer chose for this model's detections
+  - `conf_threshold` (optional): the minimum confidence the trainer chose for this model's detections; a detector filters at it unless its `CONF_THRESHOLD` is set, and at 0.2 for a model without one

{learning_loop_node-0.27.1 → learning_loop_node-0.27.2}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "learning_loop_node"
-version = "0.27.1"
+version = "0.27.2"
 description = "Python Library for Nodes which connect to the Zauberzeug Learning Loop"
 authors = [{ name = "Zauberzeug GmbH", email = "info@zauberzeug.com" }]
 license = { text = "MIT" }