major(feat): add support for Gradio API model and enhance ONNX model handling

- Introduced new model (model_8) using Gradio API for inference.
- Added preprocessing and postprocessing functions for Gradio API integration.
- Enhanced ONNX model handling with improved logging and error management.
- Updated softmax function to return Python floats for better compatibility.
- Added new model configuration files for model_8 and updated existing configurations.

app_optimized.py

@@ -12,6 +12,8 @@ import json
 from huggingface_hub import CommitScheduler, hf_hub_download, snapshot_download
 from dotenv import load_dotenv
 import concurrent.futures
+import ast
+import torch
 
 from utils.utils import softmax, augment_image
 from forensics.gradient import gradient_processing
@@ -27,7 +29,6 @@ from utils.registry import register_model, MODEL_REGISTRY, ModelEntry
 from agents.ensemble_weights import ModelWeightManager
 from transformers import pipeline, AutoImageProcessor, SwinForImageClassification, Swinv2ForImageClassification, AutoFeatureExtractor, AutoModelForImageClassification
 from torchvision import transforms
-import torch
 
 logging.basicConfig(level=logging.INFO)
 logger = logging.getLogger(__name__)
@@ -73,7 +74,8 @@ MODEL_PATHS = {
     "model_4": "cmckinle/sdxl-flux-detector_v1.1",
     "model_5": "LPX55/detection-model-5-ONNX",
     "model_6": "LPX55/detection-model-6-ONNX",
-    "model_7": "LPX55/detection-model-7-ONNX"
+    "model_7": "LPX55/detection-model-7-ONNX",
+    "model_8": "aiwithoutborders-xyz/CommunityForensics-DeepfakeDet-ViT"
 }
 
 CLASS_NAMES = {
@@ -84,6 +86,7 @@ CLASS_NAMES = {
     "model_5": ['Realism', 'Deepfake'],
     "model_6": ['ai_gen', 'human'],
     "model_7": ['Fake', 'Real'],
+    "model_8": ['Fake', 'Real'],
 }
 
 def preprocess_resize_256(image):
@@ -98,7 +101,7 @@ def preprocess_resize_224(image):
 
 def postprocess_pipeline(prediction, class_names):
     # Assumes HuggingFace pipeline output
-    return {pred['label']: pred['score'] for pred in prediction}
+    return {pred['label']: float(pred['score']) for pred in prediction}
 
 def postprocess_logits(outputs, class_names):
     # Assumes model output with logits
@@ -106,6 +109,83 @@ def postprocess_logits(outputs, class_names):
     probabilities = softmax(logits)
     return {class_names[i]: probabilities[i] for i in range(len(class_names))}
 
+def postprocess_binary_output(output, class_names):
+    # output can be a dictionary {"probabilities": numpy_array} or directly a numpy_array
+    probabilities_array = None
+    if isinstance(output, dict) and "probabilities" in output:
+        probabilities_array = output["probabilities"]
+    elif isinstance(output, np.ndarray):
+        probabilities_array = output
+    else:
+        logger.warning(f"Unexpected output type for binary post-processing: {type(output)}. Expected dict with 'probabilities' or numpy.ndarray.")
+        return {class_names[0]: 0.0, class_names[1]: 1.0}
+
+    logger.info(f"Debug: Probabilities array entering postprocess_binary_output: {probabilities_array}, type: {type(probabilities_array)}, shape: {probabilities_array.shape}")
+
+    if probabilities_array is None:
+        logger.warning("Probabilities array is None after extracting from output. Returning default scores.")
+        return {class_names[0]: 0.0, class_names[1]: 1.0}
+
+    if probabilities_array.size == 1:
+        fake_prob = float(probabilities_array.item())
+    elif probabilities_array.size == 2:
+        fake_prob = float(probabilities_array[0])
+    else:
+        logger.warning(f"Unexpected probabilities array shape for binary post-processing: {probabilities_array.shape}. Expected size 1 or 2.")
+        return {class_names[0]: 0.0, class_names[1]: 1.0}
+
+    real_prob = 1.0 - fake_prob # Ensure Fake and Real sum to 1
+    return {class_names[0]: fake_prob, class_names[1]: real_prob}
+
+# New function to infer using Gradio API for model_8
+def infer_gradio_api(image_path):
+    client = Client("aiwithoutborders-xyz/OpenSight-Community-Forensics-Preview")
+    result_str = client.predict(
+        input_image=handle_file(image_path),
+        api_name="/simple_predict"
+    )
+    logger.info(f"Debug: Raw result_str from Gradio API (model_8): {result_str}, type: {type(result_str)}")
+    try:
+        # Safely evaluate the string as a Python literal
+        result_dict = ast.literal_eval(result_str)
+        fake_probability = result_dict.get('Fake Probability', 0.0)
+        logger.info(f"Debug: Parsed result_dict: {result_dict}, Extracted fake_probability: {fake_probability}")
+        return {"probabilities": np.array([fake_probability])} # Return as a numpy array with one element
+    except Exception as e:
+        logger.error(f"Error parsing Gradio API output: {e}. Raw output: {result_str}")
+        return {"probabilities": np.array([0.0])}
+
+# New preprocess function for Gradio API
+def preprocess_gradio_api(image: Image.Image):
+    # The Gradio API expects a file path, so we need to save the PIL Image to a temporary file.
+    temp_file_path = "./temp_gradio_input.png"
+    image.save(temp_file_path)
+    return temp_file_path
+
+# New postprocess function for Gradio API (adapting postprocess_binary_output)
+def postprocess_gradio_api(gradio_output, class_names):
+    # gradio_output is expected to be a dictionary like {"probabilities": np.array([fake_prob])}
+    probabilities_array = None
+    if isinstance(gradio_output, dict) and "probabilities" in gradio_output:
+        probabilities_array = gradio_output["probabilities"]
+    elif isinstance(gradio_output, np.ndarray):
+        probabilities_array = gradio_output
+    else:
+        logger.warning(f"Unexpected output type for Gradio API post-processing: {type(gradio_output)}. Expected dict with 'probabilities' or numpy.ndarray.")
+        return {class_names[0]: 0.0, class_names[1]: 1.0}
+
+    logger.info(f"Debug: Probabilities array entering postprocess_gradio_api: {probabilities_array}, type: {type(probabilities_array)}, shape: {probabilities_array.shape}")
+
+    if probabilities_array is None or probabilities_array.size == 0:
+        logger.warning("Probabilities array is None or empty after extracting from Gradio API output. Returning default scores.")
+        return {class_names[0]: 0.0, class_names[1]: 1.0}
+
+    # It should always be a single element array for fake probability
+    fake_prob = float(probabilities_array.item())
+    real_prob = 1.0 - fake_prob
+
+    return {class_names[0]: fake_prob, class_names[1]: real_prob}
+
 def register_model_with_metadata(model_id, model, preprocess, postprocess, class_names, display_name, contributor, model_path, architecture=None, dataset=None):
     entry = ModelEntry(model, preprocess, postprocess, class_names, display_name=display_name, contributor=contributor, model_path=model_path, architecture=architecture, dataset=dataset)
     MODEL_REGISTRY[model_id] = entry
@@ -157,12 +237,15 @@ def preprocess_onnx_input(image: Image.Image, preprocessor_config: dict):
         image = image.convert('RGB')
     
     # Get image size and normalization values from preprocessor_config or use defaults
-    image_size = preprocessor_config.get('size', {'height': 224, 'width': 224})
+    # Use 'size' for initial resize and 'crop_size' for center cropping
+    initial_resize_size = preprocessor_config.get('size', {'height': 224, 'width': 224})
+    crop_size = preprocessor_config.get('crop_size', initial_resize_size['height'])
     mean = preprocessor_config.get('image_mean', [0.485, 0.456, 0.406])
     std = preprocessor_config.get('image_std', [0.229, 0.224, 0.225])
 
     transform = transforms.Compose([
-        transforms.Resize((image_size['height'], image_size['width'])),
+        transforms.Resize((initial_resize_size['height'], initial_resize_size['width'])),
+        transforms.CenterCrop(crop_size), # Apply center crop
         transforms.ToTensor(),
         transforms.Normalize(mean=mean, std=std),
     ])
@@ -170,16 +253,28 @@ def preprocess_onnx_input(image: Image.Image, preprocessor_config: dict):
     # ONNX expects numpy array with batch dimension (1, C, H, W)
     return input_tensor.unsqueeze(0).cpu().numpy()
 
-def infer_onnx_model(hf_model_id, preprocessed_image_np):
+def infer_onnx_model(hf_model_id, preprocessed_image_np, model_config: dict):
     try:
         ort_session, _, _ = get_onnx_model_from_cache(hf_model_id)
         
+        # Debug: Print expected input shape from ONNX model
+        for input_meta in ort_session.get_inputs():
+            logger.info(f"Debug: ONNX model expected input name: {input_meta.name}, shape: {input_meta.shape}, type: {input_meta.type}")
+
+        logger.info(f"Debug: preprocessed_image_np shape: {preprocessed_image_np.shape}")
         ort_inputs = {ort_session.get_inputs()[0].name: preprocessed_image_np}
         ort_outputs = ort_session.run(None, ort_inputs)
         
-        # Assuming the output is logits, apply softmax to get probabilities
         logits = ort_outputs[0]
-        probabilities = softmax(logits[0]) # Remove batch dim, apply softmax
+        logger.info(f"Debug: logits type: {type(logits)}, shape: {logits.shape}")
+        # If the model outputs a single logit (e.g., shape (1,)), use .item() to convert to scalar
+        # Otherwise, assume it's a batch of logits (e.g., shape (1, num_classes)) and take the first element (batch dim)
+        # The num_classes in config.json can be misleading; rely on actual output shape.
+        
+        # Apply softmax to the logits to get probabilities for the classes
+        # The softmax function in utils/utils.py now ensures a list of floats
+        probabilities = softmax(logits[0]) # Assuming logits[0] is the relevant output for a single prediction
+
         return {"logits": logits, "probabilities": probabilities}
 
     except Exception as e:
@@ -188,15 +283,32 @@ def infer_onnx_model(hf_model_id, preprocessed_image_np):
         return {"logits": np.array([]), "probabilities": np.array([])}
 
 def postprocess_onnx_output(onnx_output, model_config):
-    # Get class names from model_config, or use a default if not found
-    class_names = model_config.get('id2label', {0: 'Fake', 1: 'Real'}) # Default to Fake/Real if not found
-    class_names = [class_names[i] for i in sorted(class_names.keys())]
+    # Get class names from model_config
+    # Prioritize id2label, then check num_classes, otherwise default
+    class_names_map = model_config.get('id2label')
+    if class_names_map:
+        class_names = [class_names_map[k] for k in sorted(class_names_map.keys())]
+    elif model_config.get('num_classes') == 1: # Handle models that output a single value (e.g., probability of 'Fake')
+        class_names = ['Fake', 'Real'] # Assume first class is 'Fake' and second 'Real'
+    else:
+        class_names = {0: 'Fake', 1: 'Real'} # Default to Fake/Real if not found or not 1 class
+        class_names = [class_names[i] for i in sorted(class_names.keys())]
 
     probabilities = onnx_output.get("probabilities")
-    if probabilities is not None and len(probabilities) == len(class_names):
-        return {class_names[i]: probabilities[i] for i in range(len(class_names))}
+    
+    if probabilities is not None:
+        if model_config.get('num_classes') == 1 and len(probabilities) == 2: # Special handling for single output models
+            # The single output is the probability of the 'Fake' class
+            fake_prob = float(probabilities[0])
+            real_prob = float(probabilities[1])
+            return {class_names[0]: fake_prob, class_names[1]: real_prob}
+        elif len(probabilities) == len(class_names):
+            return {class_names[i]: float(probabilities[i]) for i in range(len(class_names))}
+        else:
+            logger.warning("ONNX post-processing: Probabilities length mismatch with class names.")
+            return {name: 0.0 for name in class_names}
     else:
-        logger.warning("ONNX post-processing failed or class names mismatch.")
+        logger.warning("ONNX post-processing failed: 'probabilities' key not found in output.")
         return {name: 0.0 for name in class_names}
 
 # Register the ONNX quantized model
@@ -217,7 +329,8 @@ class ONNXModelWrapper:
 
     def __call__(self, image_np):
         self.load() # Ensure model is loaded on first call
-        return infer_onnx_model(self.hf_model_id, image_np)
+        # Pass model_config to infer_onnx_model
+        return infer_onnx_model(self.hf_model_id, image_np, self._model_config)
 
     def preprocess(self, image: Image.Image):
         self.load()
@@ -229,105 +342,133 @@ class ONNXModelWrapper:
 
 # Consolidate all model loading and registration
 for model_key, hf_model_path in MODEL_PATHS.items():
-    display_name = model_key.replace("model_", "").upper()
+    logger.debug(f"Attempting to register model: {model_key} with path: {hf_model_path}")
+    model_num = model_key.replace("model_", "").upper()
     contributor = "Unknown"
     architecture = "Unknown"
     dataset = "TBA"
 
-    # Attempt to derive contributor, architecture, dataset based on model_key
-    if model_key == "model_1":
-        contributor = "haywoodsloan"
-        architecture = "SwinV2"
-        dataset = "DeepFakeDetection"
-    elif model_key == "model_2":
-        contributor = "Heem2"
-        architecture = "ViT"
-        dataset = "DeepFakeDetection"
-    elif model_key == "model_3":
-        contributor = "Organika"
-        architecture = "VIT"
-        dataset = "SDXL"
-    elif model_key == "model_4":
-        contributor = "cmckinle"
-        architecture = "VIT"
-        dataset = "SDXL, FLUX"
-    elif model_key == "model_5":
-        contributor = "prithivMLmods"
-        architecture = "VIT"
-    elif model_key == "model_6":
-        contributor = "ideepankarsharma2003"
-        architecture = "SWINv1"
-        dataset = "SDXL, Midjourney"
-    elif model_key == "model_7":
-        contributor = "date3k2"
-        architecture = "VIT"
-
     current_class_names = CLASS_NAMES.get(model_key, [])
 
+    # Logic for ONNX models (1, 2, 3, 5, 6, 7)
     if "ONNX" in hf_model_path:
+        logger.debug(f"Model {model_key} identified as ONNX.")
         logger.info(f"Registering ONNX model: {model_key} from {hf_model_path}")
         onnx_wrapper_instance = ONNXModelWrapper(hf_model_path)
+
+        # Attempt to derive contributor, architecture, dataset based on model_key
+        if model_key == "model_1":
+            contributor = "haywoodsloan"
+            architecture = "SwinV2"
+            dataset = "DeepFakeDetection"
+        elif model_key == "model_2":
+            contributor = "Heem2"
+            architecture = "ViT"
+            dataset = "DeepFakeDetection"
+        elif model_key == "model_3":
+            contributor = "Organika"
+            architecture = "VIT"
+            dataset = "SDXL"
+        elif model_key == "model_5":
+            contributor = "prithivMLmods"
+            architecture = "VIT"
+        elif model_key == "model_6":
+            contributor = "ideepankarsharma2003"
+            architecture = "SWINv1"
+            dataset = "SDXL, Midjourney"
+        elif model_key == "model_7":
+            contributor = "date3k2"
+            architecture = "VIT"
+
+        display_name_parts = [model_num]
+        if architecture and architecture not in ["Unknown"]:
+            display_name_parts.append(architecture)
+        if dataset and dataset not in ["TBA"]:
+            display_name_parts.append(dataset)
+        display_name = "-".join(display_name_parts)
+        display_name += "_ONNX" # Always append _ONNX for ONNX models
+
         register_model_with_metadata(
-            model_key,
-            onnx_wrapper_instance, # The callable wrapper for the ONNX model
-            onnx_wrapper_instance.preprocess, 
-            onnx_wrapper_instance.postprocess, 
-            current_class_names, # Initial class names; will be overridden by model_config if available
-            display_name=display_name + ("_ONNX" if "ONNX" not in display_name else ""),
+            model_id=model_key,
+            model=onnx_wrapper_instance, # The callable wrapper for the ONNX model
+            preprocess=onnx_wrapper_instance.preprocess,
+            postprocess=onnx_wrapper_instance.postprocess,
+            class_names=current_class_names, # Initial class names; will be overridden by model_config if available
+            display_name=display_name,
             contributor=contributor,
             model_path=hf_model_path,
             architecture=architecture,
             dataset=dataset
         )
-    else:
+    # Logic for Gradio API model (model_8)
+    elif model_key == "model_8":
+        logger.debug(f"Model {model_key} identified as Gradio API.")
+        logger.info(f"Registering Gradio API model: {model_key} from {hf_model_path}")
+        contributor = "aiwithoutborders-xyz"
+        architecture = "ViT"
+        dataset = "DeepfakeDetection"
+
+        display_name_parts = [model_num]
+        if architecture and architecture not in ["Unknown"]:
+            display_name_parts.append(architecture)
+        if dataset and dataset not in ["TBA"]:
+            display_name_parts.append(dataset)
+        display_name = "-".join(display_name_parts)
+
+        register_model_with_metadata(
+            model_id=model_key,
+            model=infer_gradio_api,
+            preprocess=preprocess_gradio_api,
+            postprocess=postprocess_gradio_api,
+            class_names=current_class_names,
+            display_name=display_name,
+            contributor=contributor,
+            model_path=hf_model_path,
+            architecture=architecture,
+            dataset=dataset
+        )
+    # Logic for PyTorch/Hugging Face pipeline models (currently only model_4)
+    elif model_key == "model_4": # Explicitly handle model_4
+        logger.debug(f"Model {model_key} identified as PyTorch/HuggingFace pipeline.")
         logger.info(f"Registering HuggingFace pipeline/AutoModel: {model_key} from {hf_model_path}")
-        model_instance = None
-        preprocess_func = None
-        postprocess_func = postprocess_pipeline # Default for pipelines
-        current_processor = None
-
-        if model_key == "model_5": 
-            model_instance = pipeline("image-classification", model=hf_model_path, device=device)
-            preprocess_func = preprocess_resize_224
-        elif model_key in ["model_4"]:
-            current_processor = AutoFeatureExtractor.from_pretrained(hf_model_path, device=device)
-            model_instance = AutoModelForImageClassification.from_pretrained(hf_model_path).to(device)
-
-            preprocess_func = preprocess_resize_256 # Using the shared preprocess_resize_256
-            postprocess_func = postprocess_logits # Using the shared postprocess_logits
-
-            def custom_infer(image, processor_local=current_processor, model_local=model_instance):
-                inputs = processor_local(image, return_tensors="pt").to(device)
-                with torch.no_grad():
-                    outputs = model_local(**inputs)
-                return outputs
-            model_instance = custom_infer
-
-        elif model_key in ["model_6", "model_7"]:
-            current_processor = AutoImageProcessor.from_pretrained(hf_model_path, use_fast=True)
-            if model_key == "model_6":
-                model_instance = SwinForImageClassification.from_pretrained(hf_model_path).to(device)
-            else: # model_7
-                model_instance = AutoModelForImageClassification.from_pretrained(hf_model_path).to(device)
-
-            model_instance = pipeline(model=model_instance, task="image-classification", image_processor=current_processor, device=device)
-            preprocess_func = preprocess_resize_224
-
-        if model_instance and preprocess_func:
-            register_model_with_metadata(
-                model_id=model_key,
-                model=model_instance,
-                preprocess=preprocess_func,
-                postprocess=postprocess_func,
-                class_names=current_class_names,
-                display_name=display_name,
-                contributor=contributor,
-                model_path=hf_model_path,
-                architecture=architecture,
-                dataset=dataset
-            )
-        else:
-            logger.warning(f"Could not automatically load and register model: {model_key} from {hf_model_path}")
+        contributor = "cmckinle"
+        architecture = "VIT"
+        dataset = "SDXL, FLUX"
+
+        display_name_parts = [model_num]
+        if architecture and architecture not in ["Unknown"]:
+            display_name_parts.append(architecture)
+        if dataset and dataset not in ["TBA"]:
+            display_name_parts.append(dataset)
+        display_name = "-".join(display_name_parts)
+
+        current_processor = AutoFeatureExtractor.from_pretrained(hf_model_path, device=device)
+        model_instance = AutoModelForImageClassification.from_pretrained(hf_model_path).to(device)
+
+        preprocess_func = preprocess_resize_256
+        postprocess_func = postprocess_logits
+
+        def custom_infer(image, processor_local=current_processor, model_local=model_instance):
+            inputs = processor_local(image, return_tensors="pt").to(device)
+            with torch.no_grad():
+                outputs = model_local(**inputs)
+            return outputs
+        model_instance = custom_infer
+
+        register_model_with_metadata(
+            model_id=model_key,
+            model=model_instance,
+            preprocess=preprocess_func,
+            postprocess=postprocess_func,
+            class_names=current_class_names,
+            display_name=display_name,
+            contributor=contributor,
+            model_path=hf_model_path,
+            architecture=architecture,
+            dataset=dataset
+        )
+    else: # Fallback for any unhandled models (shouldn't happen if MODEL_PATHS is fully covered)
+        logger.warning(f"Could not automatically load and register model: {model_key} from {hf_model_path}. No matching registration logic found.")
 
 
 def infer(image: Image.Image, model_id: str, confidence_threshold: float = 0.75) -> dict:
@@ -346,8 +487,14 @@ def infer(image: Image.Image, model_id: str, confidence_threshold: float = 0.75)
     try:
         result = entry.model(img)
         scores = entry.postprocess(result, entry.class_names)
-        ai_score = float(scores.get(entry.class_names[0], 0.0))
-        real_score = float(scores.get(entry.class_names[1], 0.0))
+        
+        def _to_float_scalar(value):
+            if isinstance(value, np.ndarray):
+                return float(value.item()) # Convert numpy array scalar to Python float
+            return float(value) # Already a Python scalar or convertible type
+
+        ai_score = _to_float_scalar(scores.get(entry.class_names[0], 0.0))
+        real_score = _to_float_scalar(scores.get(entry.class_names[1], 0.0))
         label = "AI" if ai_score >= confidence_threshold else ("REAL" if real_score >= confidence_threshold else "UNCERTAIN")
         return {
             "Model": entry.display_name,
@@ -445,20 +592,30 @@ def full_prediction(img, confidence_threshold, rotate_degrees, noise_level, shar
         model_start = time.time()
         result = infer(img_pil, model_id, confidence_threshold)
         model_end = time.time()
+
+        # Helper to ensure values are Python floats, handling numpy scalars
+        def _ensure_float_scalar(value):
+            if isinstance(value, np.ndarray):
+                return float(value.item()) # Convert numpy array scalar to Python float
+            return float(value) # Already a Python scalar or convertible type
+
+        ai_score_val = _ensure_float_scalar(result.get("AI Score", 0.0))
+        real_score_val = _ensure_float_val = _ensure_float_scalar(result.get("Real Score", 0.0))
+
         monitor_agent.monitor_prediction(
             model_id,
             result["Label"],
-            max(result.get("AI Score", 0.0), result.get("Real Score", 0.0)),
+            max(ai_score_val, real_score_val),
             model_end - model_start
         )
         model_predictions_raw[model_id] = result
-        confidence_scores[model_id] = max(result.get("AI Score", 0.0), result.get("Real Score", 0.0))
+        confidence_scores[model_id] = max(ai_score_val, real_score_val)
         results.append(result)
         table_rows.append([
             result.get("Model", ""),
             result.get("Contributor", ""),
-            round(result.get("AI Score", 0.0), 3) if result.get("AI Score") is not None else 0.0,
-            round(result.get("Real Score", 0.0), 3) if result.get("Real Score") is not None else 0.0,
+            round(ai_score_val, 5),
+            round(real_score_val, 5),
             result.get("Label", "Error")
         ])
         # Yield partial results: only update the table, others are None
@@ -817,7 +974,7 @@ demo = gr.TabbedInterface(
 
 )
 footerMD = """
-### ⚠️ ENSEMBLE TEAM IN TRAINING ⚠️ \n\n
+## ⚠️ ENSEMBLE TEAM IN TRAINING ⚠️ \n\n
 
 1. **DISCLAIMER: METADATA AS WELL AS MEDIA SUBMITTED TO THIS SPACE MAY BE VIEWED AND SELECTED FOR FUTURE DATASETS, PLEASE DO NOT SUBMIT PERSONAL CONTENT. FOR UNTRACKED, PRIVATE USE OF THE MODELS YOU MAY STILL USE [THE ORIGINAL SPACE HERE](https://huggingface.co/spaces/aiwithoutborders-xyz/OpenSight-Deepfake-Detection-Models-Playground), SOTA MODEL INCLUDED.**
 2. **UPDATE 6-13-25**: APOLOGIES FOR THE CONFUSION, WE ARE WORKING TO REVERT THE ORIGINAL REPO BACK TO ITS NON-DATA COLLECTION STATE -- ONLY THE "SIMPLE PREDICTION" ENDPOINT IS CURRENTLY 100% PRIVATE. PLEASE STAY TUNED AS WE FIGURE OUT A SOLUTION FOR THE ENSEMBLE + AGENT TEAM ENDPOINT. IT CAN GET RESOURCE INTENSIVE TO RUN A FULL PREDICTION. ALTERNATIVELY, WE **ENCOURAGE** ANYONE TO FORK AND CONTRIBUTE TO THE PROJECT.

temp_model_config/config.json

@@ -0,0 +1,28 @@
+{
+  "_attn_implementation_autoset": true,
+  "_name_or_path": "aiwithoutborders-xyz/CommunityForensics-DeepfakeDet-ViT",
+  "architectures": [
+    "ViTForImageClassification"
+  ],
+  "attention_probs_dropout_prob": 0.0,
+  "encoder_stride": 16,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.0,
+  "hidden_size": 384,
+  "image_size": 384,
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "layer_norm_eps": 1e-06,
+  "mlp_ratio": 4,
+  "model_type": "vit",
+  "num_attention_heads": 12,
+  "num_channels": 3,
+  "num_classes": 1,
+  "num_heads": 6,
+  "num_hidden_layers": 12,
+  "num_layers": 12,
+  "patch_size": 16,
+  "qkv_bias": true,
+  "torch_dtype": "float32",
+  "transformers_version": "4.49.0"
+}

temp_original_vit_config/config.json

@@ -0,0 +1,26 @@
+{
+  "architectures": [
+    "ViTForImageClassification"
+  ],
+  "attention_probs_dropout_prob": 0.0,
+  "encoder_stride": 16,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.0,
+  "hidden_size": 384,
+  "image_size": 384,
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "layer_norm_eps": 1e-06,
+  "mlp_ratio": 4,
+  "model_type": "vit",
+  "num_attention_heads": 12,
+  "num_channels": 3,
+  "num_classes": 1,
+  "num_heads": 6,
+  "num_hidden_layers": 12,
+  "num_layers": 12,
+  "patch_size": 16,
+  "qkv_bias": true,
+  "torch_dtype": "float32",
+  "transformers_version": "4.50.0.dev0"
+}

utils/utils.py

@@ -5,7 +5,8 @@ from torchvision import transforms
 
 def softmax(vector):
     e = np.exp(vector - np.max(vector))  # for numerical stability
-    return e / e.sum()
+    probabilities = e / e.sum()
+    return [float(p.item()) for p in probabilities] # Convert numpy array elements to Python floats using .item()
 
 def augment_image(img_pil, methods, rotate_degrees=0, noise_level=0, sharpen_strength=1):
     for method in methods: