Merge pull request #29 from hmorimitsu/docfixes

Docfixes

Author: hmorimitsu

.github/workflows/pytest.yml

@@ -17,8 +17,8 @@ jobs:
       fail-fast: false
       matrix:
         python-version: [3.6, 3.9]
-        torch: [1.7.1, 1.9.0]
-        pytorch-lightning: [1.1.8, 1.4.0]
+        torch: [1.7.1, 1.10.0]
+        pytorch-lightning: [1.1.8, 1.5.2]
 
     steps:
       - uses: actions/checkout@v2

docs/source/starting/inference.rst

@@ -71,7 +71,7 @@ The code below shows a way to do this:
     predictions = model(inputs)
 
     # Remove extra padding that may have been added to the inputs
-    predictions = io_adapter.unpad(predictions)
+    predictions = io_adapter.unpad_and_unscale(predictions)
 
     # The output is a dict with possibly several keys,
     # but it should always store the optical flow prediction in a key called 'flows'.
@@ -93,4 +93,12 @@ The code below shows a way to do this:
     cv.imshow('image1', images[0])
     cv.imshow('image2', images[1])
     cv.imshow('flow', flow_bgr_npy)
-    cv.waitKey()
+    cv.waitKey()
+
+Inference on batches of images
+==============================
+
+For simplicity, the base PTLFlow scripts do not provide a direct way to do inference on batches.
+However, it should be easy to extend the base scripts to your use case.
+One example of a workaround to work with batches can be found in
+`[this GitHub issue] <https://github.com/hmorimitsu/ptlflow/issues/28>`__.

ptlflow/models/base_model/base_model.py

@@ -338,7 +338,7 @@ def configure_optimizers(self) -> Dict[str, Any]:
 
         self.train_dataloader()  # Just to initialize dataloader variables
 
-        if self.args.max_steps is None:
+        if self.args.max_steps is None or self.args.max_steps <= 0:
             if self.args.max_epochs is None:
                 self.args.max_epochs = 10
                 logging.warning('--max_epochs is not set. It will be set to %d.', self.args.max_epochs)