Add single-node DDP support, distributed utils, samplers, and gradient accumulation

CHANGELOG.md

@@ -1,5 +1,13 @@
 # Changelog
 
+## 1.0.6
+- FIX: DDP training stability and config handling (single init, rank/world-size propagation, robust setup/token flow)
+- NEW: tqdm progress bars for training/validation and interactive embedding extraction
+- FIX: interactive plots alignment bugs and compute_embeddings device argument handling
+- FIX: Streamlit compatibility updates for model explorer (`use_container_width`, watcher workaround)
+- FIX: safer core behavior (no-grad embedding extraction, CPU fallbacks in train/SWA/LR finder, clearer missing global config error)
+- IMPROVED: metric selection logic with configurable `target_metric_mode` (`auto` / `max` / `min`)
+
 ## 1.0.5
 - FIX: class order in model explorer was broken!
 - FIX: interactive plot was looking for a missing argument 

README.md

@@ -32,7 +32,10 @@ Read the paper: [https://onlinelibrary.wiley.com/doi/10.1111/ele.14495](https://
 [>> Comprehensive help files <<](help)
 
 1\. Install BioEncoder (into a virtual environment with pytorch/CUDA): 
+
+Install `torch` / `torchvision` first for your platform (CPU or CUDA), then install BioEncoder:
 ````
+pip install torch torchvision
 pip install bioencoder
 ````
 

bioencoder/__init__.py

@@ -2,15 +2,49 @@
 from .core import utils
 # from .scripts import *
 
-from .scripts.archive import archive
-from .scripts.configure import configure
-from .scripts.split_dataset import split_dataset
-from .scripts.train import train
-from .scripts.swa import swa
-from .scripts.lr_finder import lr_finder
-from .scripts.interactive_plots import interactive_plots
-from .scripts.inference import inference
-from .scripts.model_explorer_wrapper import model_explorer_wrapper as model_explorer
+def archive(*args, **kwargs):
+    from .scripts.archive import archive as _archive
+    return _archive(*args, **kwargs)
+
+
+def configure(*args, **kwargs):
+    from .scripts.configure import configure as _configure
+    return _configure(*args, **kwargs)
+
+
+def split_dataset(*args, **kwargs):
+    from .scripts.split_dataset import split_dataset as _split_dataset
+    return _split_dataset(*args, **kwargs)
+
+
+def train(*args, **kwargs):
+    from .scripts.train import train as _train
+    return _train(*args, **kwargs)
+
+
+def swa(*args, **kwargs):
+    from .scripts.swa import swa as _swa
+    return _swa(*args, **kwargs)
+
+
+def lr_finder(*args, **kwargs):
+    from .scripts.lr_finder import lr_finder as _lr_finder
+    return _lr_finder(*args, **kwargs)
+
+
+def interactive_plots(*args, **kwargs):
+    from .scripts.interactive_plots import interactive_plots as _interactive_plots
+    return _interactive_plots(*args, **kwargs)
+
+
+def inference(*args, **kwargs):
+    from .scripts.inference import inference as _inference
+    return _inference(*args, **kwargs)
+
+
+def model_explorer(*args, **kwargs):
+    from .scripts.model_explorer_wrapper import model_explorer_wrapper as _model_explorer
+    return _model_explorer(*args, **kwargs)
 
 from importlib.metadata import version
 __version__ = version("bioencoder")

bioencoder/core/losses.py

@@ -173,7 +173,7 @@ def forward(self, pred:torch.Tensor, target:torch.Tensor):
         with torch.no_grad():
             true_dist = torch.zeros_like(pred)
             true_dist.fill_(self.smoothing / (self.cls - 1))
-            true_dist.scatter_(1, target.unsqueeze(1), self.confidence)
+            true_dist.scatter_(class_dim, target.unsqueeze(class_dim), self.confidence)
         return torch.mean(torch.sum(-true_dist * pred, dim=self.dim))