fix output data type of image classification

src/transformers/pipelines/image_classification.py

@@ -23,6 +23,8 @@
     from ..models.auto.modeling_tf_auto import TF_MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING_NAMES
 
 if is_torch_available():
+    import torch
+
     from ..models.auto.modeling_auto import MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING_NAMES
 
 logger = logging.get_logger(__name__)
@@ -180,7 +182,10 @@ def postprocess(self, model_outputs, function_to_apply=None, top_k=5):
             top_k = self.model.config.num_labels
 
         outputs = model_outputs["logits"][0]
-        outputs = outputs.numpy()
+        if self.framework == "pt" and outputs.dtype in (torch.bfloat16, torch.float16):
+            outputs = outputs.to(torch.float32).numpy()
+        else:
+            outputs = outputs.numpy()
 
         if function_to_apply == ClassificationFunction.SIGMOID:
             scores = sigmoid(outputs)

tests/pipelines/test_pipelines_image_classification.py

@@ -18,6 +18,7 @@
     MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING,
     TF_MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING,
     PreTrainedTokenizerBase,
+    is_torch_available,
     is_vision_available,
 )
 from transformers.pipelines import ImageClassificationPipeline, pipeline
@@ -34,6 +35,9 @@
 from .test_pipelines_common import ANY
 
 
+if is_torch_available():
+    import torch
+
 if is_vision_available():
     from PIL import Image
 else:
@@ -177,6 +181,30 @@ def test_custom_tokenizer(self):
 
         self.assertIs(image_classifier.tokenizer, tokenizer)
 
+    @require_torch
+    def test_torch_float16_pipeline(self):
+        image_classifier = pipeline(
+            "image-classification", model="hf-internal-testing/tiny-random-vit", torch_dtype=torch.float16
+        )
+        outputs = image_classifier("http://images.cocodataset.org/val2017/000000039769.jpg")
+
+        self.assertEqual(
+            nested_simplify(outputs, decimals=3),
+            [{"label": "LABEL_1", "score": 0.574}, {"label": "LABEL_0", "score": 0.426}],
+        )
+
+    @require_torch
+    def test_torch_bfloat16_pipeline(self):
+        image_classifier = pipeline(
+            "image-classification", model="hf-internal-testing/tiny-random-vit", torch_dtype=torch.bfloat16
+        )
+        outputs = image_classifier("http://images.cocodataset.org/val2017/000000039769.jpg")
+
+        self.assertEqual(
+            nested_simplify(outputs, decimals=3),
+            [{"label": "LABEL_1", "score": 0.574}, {"label": "LABEL_0", "score": 0.426}],
+        )
+
     @slow
     @require_torch
     def test_perceiver(self):