commit files to HF hub

README.md

@@ -11,11 +11,11 @@ model-index:
   results:
   - task:
       name: Image Classification
-      type: image-classification
+      type: pair-classification
     metrics:
       - name: Accuracy
         type: accuracy
-        value: 0.375
+        value: 0.3928571343421936
 ---
 
 # krenzcolor_chkpt_classifier

config.json

@@ -4,6 +4,17 @@
     "ViTForImageClassification"
   ],
   "attention_probs_dropout_prob": 0.0,
+  "custom_pipelines": {
+    "pair-classification": {
+      "impl": "pair_classification.PairClassificationPipeline",
+      "pt": [
+        "ViTForImageClassification"
+      ],
+      "tf": [
+        "TFViTForImageClassification"
+      ]
+    }
+  },
   "encoder_stride": 16,
   "hidden_act": "gelu",
   "hidden_dropout_prob": 0.0,

pair_classification.py

@@ -0,0 +1,29 @@
+from torchvision import transforms
+from transformers import ImageClassificationPipeline
+import torch
+
+
+class PairClassificationPipeline(ImageClassificationPipeline):
+    pipe_to_tensor = transforms.ToTensor()
+    pipe_to_pil = transforms.ToPILImage()
+
+    def preprocess(self, image):       
+        left_image, right_image = self.horizontal_split_image(image)
+        model_inputs = self.extract_split_feature(left_image, right_image)
+        # model_inputs = super().preprocess(image)
+        # print(model_inputs['pixel_values'].shape)
+        return model_inputs
+
+    def horizontal_split_image(self, image):
+        # image = image.resize((448,224))
+        w, h = image.size
+        half_w = w//2
+        left_image = image.crop([0,0,half_w,h])
+        right_image = image.crop([half_w,0,2*half_w,h])
+        return left_image, right_image
+    
+    def extract_split_feature(self, left_image, right_image):
+        model_inputs = self.feature_extractor(images=left_image, return_tensors=self.framework)
+        right_inputs = self.feature_extractor(images=right_image, return_tensors=self.framework)
+        model_inputs['pixel_values'] = torch.cat([model_inputs['pixel_values'],right_inputs['pixel_values']], dim=1)
+        return model_inputs

pytorch_model.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5c850dd17c2fb514bd822185575af28eb23a38a41eeff521fca00ad8d9f43efc
+oid sha256:b5233e44bcfb172d9200f5f0c63a00ac4411a1783bf6dd42b48dce384a468da1
 size 345635761

runs/events.out.tfevents.1666962937.sa103.22908.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a878e370777ce85d35b80fa02f4354af89b306a44916d4b592109528e5dad9ed
+size 1939