Fix HSC zeropoint decoding inverse

aion/codecs/preprocessing/image.py

@@ -102,6 +102,6 @@ def forward(self, image, survey):
         return image
 
     def backward(self, image, survey):
-        zpscale = self.reverse_zeropoint(27.0) if survey == "HSC" else 1.0
+        zpscale = self.convert_zeropoint(27.0) if survey == "HSC" else 1.0
         image *= zpscale
         return image

tests/codecs/test_image_codec.py

@@ -3,9 +3,68 @@
 
 from aion.codecs import ImageCodec
 from aion.codecs.config import HF_REPO_ID
+from aion.codecs.preprocessing.image import RescaleToLegacySurvey
 from aion.modalities import Image
 
 
+def test_hsc_rescaling_uses_legacy_survey_zeropoint():
+    rescaler = RescaleToLegacySurvey()
+    image = torch.tensor([[[[-2.5, 0.0], [1.0, 12.5]]]])
+    scale = rescaler.convert_zeropoint(27.0)
+
+    rescaled = rescaler.forward(image.clone(), "HSC")
+
+    assert torch.allclose(rescaled, image / scale)
+
+
+def test_hsc_rescaling_round_trip():
+    rescaler = RescaleToLegacySurvey()
+    image = torch.tensor([[[[-2.5, 0.0], [1.0, 12.5]]]])
+
+    rescaled = rescaler.forward(image.clone(), "HSC")
+    restored = rescaler.backward(rescaled, "HSC")
+
+    assert torch.allclose(restored, image)
+
+
+def test_non_hsc_rescaling_is_unchanged():
+    rescaler = RescaleToLegacySurvey()
+    image = torch.tensor([[[[-2.5, 0.0], [1.0, 12.5]]]])
+
+    rescaled = rescaler.forward(image.clone(), "DES")
+    restored = rescaler.backward(rescaled, "DES")
+
+    assert torch.equal(rescaled, image)
+    assert torch.equal(restored, image)
+
+
+def test_hsc_reversible_preprocessing_round_trip():
+    codec = ImageCodec(
+        quantizer_levels=[1] * 5,
+        hidden_dims=8,
+        multisurvey_projection_dims=12,
+        n_compressions=2,
+        num_consecutive=1,
+        embedding_dim=5,
+    )
+    bands = ["HSC-G", "HSC-R", "HSC-I", "HSC-Z", "HSC-Y"]
+    image = torch.linspace(-10.0, 10.0, 5 * 96 * 96).reshape(1, 5, 96, 96)
+
+    # Crop and clamp are lossless for this already-cropped, in-range input.
+    processed = codec.center_crop(image.clone())
+    processed = codec.clamp(processed, bands)
+    processed = codec.rescaler.forward(processed, codec._get_survey(bands))
+    processed = codec._range_compress(processed)
+    processed, channel_mask = codec.image_padder.forward(processed, bands)
+
+    restored = codec._reverse_range_compress(processed)
+    restored = codec.image_padder.backward(restored, bands)
+    restored = codec.rescaler.backward(restored, codec._get_survey(bands))
+
+    assert channel_mask.sum().item() == len(bands)
+    assert torch.allclose(restored, image, rtol=1e-5, atol=1e-6)
+
+
 @pytest.mark.parametrize("embedding_dim", [5, 10])
 @pytest.mark.parametrize("multisurvey_projection_dims", [12, 24])
 @pytest.mark.parametrize("hidden_dims", [8, 16])