Readding the validation of the minimal cnn input size (#5345) (#5346)

Author: vincentpierre

ml-agents/mlagents/trainers/tests/torch/test_utils.py

@@ -36,6 +36,29 @@ def test_min_visual_size():
             enc.forward(vis_input)
 
 
+@pytest.mark.parametrize(
+    "encoder_type",
+    [
+        EncoderType.SIMPLE,
+        EncoderType.NATURE_CNN,
+        EncoderType.SIMPLE,
+        EncoderType.MATCH3,
+    ],
+)
+def test_invalid_visual_input_size(encoder_type):
+    with pytest.raises(UnityTrainerException):
+        obs_spec = create_observation_specs_with_shapes(
+            [
+                (
+                    ModelUtils.MIN_RESOLUTION_FOR_ENCODER[encoder_type] - 1,
+                    ModelUtils.MIN_RESOLUTION_FOR_ENCODER[encoder_type],
+                    1,
+                )
+            ]
+        )
+        ModelUtils.create_input_processors(obs_spec, 20, encoder_type, 20, False)
+
+
 @pytest.mark.parametrize("num_visual", [0, 1, 2])
 @pytest.mark.parametrize("num_vector", [0, 1, 2])
 @pytest.mark.parametrize("normalize", [True, False])

ml-agents/mlagents/trainers/torch/utils.py

@@ -159,6 +159,9 @@ def get_encoder_for_obs(
         # VISUAL
         if dim_prop in ModelUtils.VALID_VISUAL_PROP:
             visual_encoder_class = ModelUtils.get_encoder_for_type(vis_encode_type)
+            ModelUtils._check_resolution_for_encoder(
+                shape[0], shape[1], vis_encode_type
+            )
             return (visual_encoder_class(shape[0], shape[1], shape[2], h_size), h_size)
         # VECTOR
         if dim_prop in ModelUtils.VALID_VECTOR_PROP: