deepvisiontools 1.0.2__tar.gz → 1.0.3__tar.gz

{deepvisiontools-1.0.2 → deepvisiontools-1.0.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: deepvisiontools
-Version: 1.0.2
+Version: 1.0.3
 Summary: deepvisiontools is a library built on torch for training and inference processes for deep learning computer vision tasks.
 Author-email: Lucas Bernigaud Samatan <lucas.bernigaud-samatan@inrae.fr>, Jordan Bernigaud Samatan <jordan.bernigaud-samatan@inrae.fr>
 Maintainer-email: Jordan Bernigaud Samatan <jordan.bernigaud-samatan@inrae.fr>, Lucas Bernigaud Samatan <lucas.bernigaud-samatan@inrae.fr>
@@ -733,6 +733,7 @@ Dynamic: license-file
 
 ### patchs
 
+- *1.0.3: fixed ddp desynchronization during validation, made ddp more robust, added ddp timeout control*
 - *1.0.2: added option to concat datasets for deepvisiondataset (DeepVisionDataset.concat)*
 - *1.0.1: fixed inconsistency for yolo wrappers due to modified behaviour in ultralytics between 8.3 and 8.4 versions. Fixed and compatible with both.*
 

{deepvisiontools-1.0.2 → deepvisiontools-1.0.3}/README.md

@@ -16,6 +16,7 @@
 
 ### patchs
 
+- *1.0.3: fixed ddp desynchronization during validation, made ddp more robust, added ddp timeout control*
 - *1.0.2: added option to concat datasets for deepvisiondataset (DeepVisionDataset.concat)*
 - *1.0.1: fixed inconsistency for yolo wrappers due to modified behaviour in ultralytics between 8.3 and 8.4 versions. Fixed and compatible with both.*
 

{deepvisiontools-1.0.2 → deepvisiontools-1.0.3}/src/deepvisiontools/__init__.py

@@ -1,4 +1,4 @@
-__version__ = "1.0.2"
+__version__ = "1.0.3"
 
 from .config import current, set_current, use_config
 from .datasets import DeepVisionDataset, DeepVisionLoader

{deepvisiontools-1.0.2 → deepvisiontools-1.0.3}/src/deepvisiontools/models/base/basemodel.py

@@ -243,6 +243,62 @@ class DeepVisionModel(torch.nn.Module):
 
     # ----------------------------- core steps -----------------------------
 
+    def forward(
+        self,
+        images: torch.Tensor,
+        targets: BatchData | None = None,
+        *,
+        extra_args: dict[str, Any] | None = None,
+        mode: Literal["train", "eval", "valid", "predict"] = "train",
+    ) -> ModelOutput:
+        """
+        DDP-compatible forward dispatcher.
+
+        `DeepVisionModel` keeps the public high-level façade methods
+        `train_step`, `eval_step`, and `predict`. This `forward` method is the
+        module-level entrypoint used by wrappers such as
+        `torch.nn.parallel.DistributedDataParallel`, and dispatches to the same
+        façade methods without changing their behavior.
+
+        Args:
+            images (`torch.Tensor`): Image batch.
+            targets (`BatchData | None`): Optional target batch data.
+            extra_args (`dict[str, Any] | None`): Optional backend-specific
+                keyword payload forwarded to the selected façade method.
+            mode (`Literal["train", "eval", "valid", "predict"]`): Execution
+                path. `"train"` calls `train_step`, `"eval"` and `"valid"` call
+                `eval_step`, and `"predict"` calls `predict`.
+
+        Returns:
+            `ModelOutput`: Standard DeepVisionTools model output.
+
+        ???+ note "DDP note"
+            ```
+            In DDP, the Trainer must call the DDP wrapper itself, not
+            ddp.module.train_step(...). Calling the wrapper triggers PyTorch's
+            distributed forward/backward bookkeeping, while this dispatcher keeps
+            the DeepVisionTools façade API unchanged.
+            ```
+        """
+        if mode == "train":
+            return self.train_step(images, targets, extra_args=extra_args)
+
+        if mode in {"eval", "valid"}:
+            return self.eval_step(images, targets, extra_args=extra_args)
+
+        if mode == "predict":
+            if targets is not None:
+                raise ValueError(
+                    "DeepVisionModel.forward(mode='predict') expects targets=None."
+                )
+            return self.predict(images, extra_args=extra_args)
+
+        raise ValueError(
+            "Unsupported DeepVisionModel.forward mode. "
+            "Expected one of {'train', 'eval', 'valid', 'predict'}, "
+            f"got {mode!r}."
+        )
+
     def train_step(
         self,
         images: torch.Tensor,

{deepvisiontools-1.0.2 → deepvisiontools-1.0.3}/src/deepvisiontools/train/trainconfig.py

@@ -3,6 +3,7 @@ from __future__ import annotations
 import copy
 import datetime as _dt
 import importlib
+import math
 import platform
 import sys
 from collections.abc import Callable, Sequence
@@ -371,6 +372,9 @@ class DevicesConfig:
         ```
         - device=None + strategy="auto": choose "cuda" if available, else "cpu" with warning.
         - strategy="auto" matches trainer logic (DDP if >1 GPU).
+        - ddp_timeout_seconds controls the distributed process-group timeout for
+          DDP collectives. If None, DDPStrategy reads DVT_DDP_TIMEOUT_SECONDS
+          when present, otherwise it uses its internal default.
         ```
     """
 
@@ -381,6 +385,9 @@ class DevicesConfig:
     num_workers: int | None = None
     pin_memory: bool | None = None
 
+    # None -> DDPStrategy env/default policy
+    ddp_timeout_seconds: float | None = None
+
     def __post_init__(self) -> None:
         if self.device is not None:
             dev = str(self.device).strip().lower()
@@ -400,6 +407,15 @@ class DevicesConfig:
         if self.pin_memory is not None:
             self.pin_memory = bool(self.pin_memory)
 
+        if self.ddp_timeout_seconds is not None:
+            timeout = float(self.ddp_timeout_seconds)
+            if not math.isfinite(timeout) or timeout <= 0.0:
+                raise ValueError(
+                    "DevicesConfig.ddp_timeout_seconds must be a positive finite "
+                    f"number of seconds. Got {self.ddp_timeout_seconds!r}."
+                )
+            self.ddp_timeout_seconds = timeout
+
 
 @dataclass
 class WarmupConfig: