trying different approach

config.json

@@ -1,31 +1,25 @@
 {
-  "_name_or_path": "Maslionok/pipeline1",
+  "_name_or_path": "Maslionok/pipeline1/LID-40-3-2000000-1-4.bin",
   "architectures": [
-    "BertForTokenClassification"
+    "Pipeline_One"
   ],
-  "custom_pipelines": {
-      "language-detection": {
-          "impl": "impresso_langident_wrapper.Pipeline_One",
-          "pt": ["AutoModel"],
-          "tf": []
-      }
-  },
+  "model_type": "floret",
+  "num_labels": 40,
   "id2label": {
-      "0": "English",
-      "1": "German",
-      "2": "French"
+    "0": "English",
+    "1": "German",
+    "2": "French"
   },
   "label2id": {
-      "English": 0,
-      "German": 1,
-      "French": 2
+    "English": 0,
+    "German": 1,
+    "French": 2
   },
-  "model_type": "bert",
-  "vocab_size": 2000000,  
-  "embedding_dim": 300,   
-  "hash_count": 4,        
-  "minn": 3,             
-  "maxn": 6,             
-  "bucket": 2000000,      
-  "num_labels": 40
+  "custom_pipelines": {
+    "language-detection": {
+      "impl": "Pipeline_One",
+      "pt": [],
+      "tf": []
+    }
+  }
 }

impresso_langident_wrapper.py

@@ -1,29 +1,29 @@
 from transformers import Pipeline
 from transformers.pipelines import PIPELINE_REGISTRY
-
+import floret  # Ensure floret is imported
 
 
 class Pipeline_One(Pipeline):
     def _sanitize_parameters(self, **kwargs):
-        # Add any additional parameter handling if necessary
         return kwargs, {}, {}
 
     def preprocess(self, text, **kwargs):
         return text
 
     def _forward(self, inputs):
-        model_output = self.model.predict(inputs, k=1)
+        # Lazy load the model when first used
+        if not hasattr(self, "model"):
+            self.model = floret.load_model(self.model.config["Maslionok/pipeline1"])  # Load model dynamically
+
+        model_output = self.model.predict([inputs], k=1)
         return model_output
 
     def postprocess(self, outputs, **kwargs):
-        return outputs
-
+        return outputs[0]  # Return top prediction
 
-from transformers import AutoModelForSequenceClassification, TFAutoModelForSequenceClassification
 
+# Register the pipeline
 PIPELINE_REGISTRY.register_pipeline(
     "language-detection",
-    pipeline_class=Pipeline_One,
-    pt_model=AutoModelForSequenceClassification,
-    tf_model=TFAutoModelForSequenceClassification,
-)
+    pipeline_class=Pipeline_One
+)