ajusta verificação de conteúdo

app.py

@@ -1,66 +1,129 @@
 import gradio as gr
 from transformers import pipeline
 import re
+import torch
+import timm
+from PIL import Image
+import numpy as np
+
+try:
+    eva02_model = timm.create_model('hf_hub:SmilingWolf/wd-eva02-large-tagger-v3', pretrained=True)
+    eva02_model.eval()
+    
+    import requests
+    tags_response = requests.get('https://huggingface.co/SmilingWolf/wd-eva02-large-tagger-v3/resolve/main/selected_tags.csv')
+    tags_lines = tags_response.text.strip().split('\n')
+    eva02_tags = [line.split(',')[1] for line in tags_lines[1:]]  # Skip header
+    
+    print(f"Modelo EVA02 carregado com {len(eva02_tags)} tags")
+except Exception as e:
+    print(f"Erro ao carregar EVA02: {e}")
+    eva02_model = None
+    eva02_tags = []
 
 content_model = pipeline("image-classification", model="facebook/convnext-base-224")
 nsfw_model = pipeline("image-classification", model="Falconsai/nsfw_image_detection")
 
+def analyze_with_eva02(image):
+    if eva02_model is None:
+        return [], []
+    
+    image_tensor = torch.from_numpy(np.array(image)).permute(2, 0, 1).float() / 255.0
+    image_tensor = torch.nn.functional.interpolate(
+        image_tensor.unsqueeze(0), 
+        size=(448, 448), 
+        mode='bilinear', 
+        align_corners=False
+    )
+    
+    with torch.no_grad():
+        features = eva02_model(image_tensor)
+        probs = torch.sigmoid(features[0])
+    
+    detected_tags = []
+    tag_scores = []
+    
+    for i, (tag, prob) in enumerate(zip(eva02_tags, probs)):
+        if prob > 0.5:
+            detected_tags.append(tag)
+            tag_scores.append(float(prob))
+    
+    return detected_tags, tag_scores
+
 def analyze_image(image):
-    content_preds = content_model(image)
-    top_content = max(content_preds, key=lambda x: x["score"])
-    
-    nsfw_preds = nsfw_model(image)
-    top_nsfw = max(nsfw_preds, key=lambda x: x["score"])
-    
-    all_labels = []
-    for pred in content_preds:
-        all_labels.append(pred["label"].lower())
-    
-    for pred in nsfw_preds:
-        all_labels.append(pred["label"].lower())
-    
-    combined_labels = " ".join(all_labels)
+    if eva02_model is not None:
+        eva02_tags_detected, eva02_scores = analyze_with_eva02(image)
+        combined_labels = " ".join(eva02_tags_detected).lower()
+        print(f"EVA02 detectou: {eva02_tags_detected}")
+    else:
+        content_preds = content_model(image)
+        top_content = max(content_preds, key=lambda x: x["score"])
+        
+        nsfw_preds = nsfw_model(image)
+        top_nsfw = max(nsfw_preds, key=lambda x: x["score"])
+        
+        all_labels = []
+        for pred in content_preds:
+            all_labels.append(pred["label"].lower())
+        
+        for pred in nsfw_preds:
+            all_labels.append(pred["label"].lower())
+        
+        combined_labels = " ".join(all_labels)
     
     human_keywords = [
         "human", "person", "people", "man", "woman", "child", "baby", "boy", "girl",
-        "face", "portrait", "selfie", "crowd", "family", "couple", "teenager"
+        "face", "portrait", "selfie", "crowd", "family", "couple", "teenager",
+        "1boy", "1girl", "2boys", "2girls", "multiple boys", "multiple girls",
+        "male", "female", "adult", "teen", "kid", "toddler", "infant"
     ]
     is_human = any(keyword in combined_labels for keyword in human_keywords)
     
     dog_keywords = [
         "dog", "puppy", "retriever", "labrador", "golden", "beagle", "bulldog",
         "poodle", "german shepherd", "chihuahua", "terrier", "hound", "mastiff",
-        "canine", "pet", "animal", "malamute", "malemute", "alaskan"
+        "canine", "pet", "animal", "malamute", "malemute", "alaskan",
+        "1dog", "2dogs", "multiple dogs", "doggy", "doggie", "pup"
     ]
     is_dog = any(keyword in combined_labels for keyword in dog_keywords)
     
     violence_keywords = [
         "blood", "wound", "injury", "hurt", "pain", "fight", "violence", "weapon",
-        "knife", "gun", "attack", "aggression", "conflict", "battle", "war"
+        "knife", "gun", "attack", "aggression", "conflict", "battle", "war",
+        "bloody", "injured", "wounded", "bleeding", "scar", "bruise", "cut"
     ]
     
     suffering_keywords = [
         "sad", "crying", "tears", "depressed", "miserable", "suffering", "pain",
         "distress", "anguish", "grief", "mourning", "funeral", "death", "dead",
-        "dying", "illness", "sick", "injured", "abandoned", "neglected"
+        "dying", "illness", "sick", "injured", "abandoned", "neglected",
+        "crying", "tears", "sad", "depressed", "miserable", "grief", "mourning"
     ]
     
     abuse_keywords = [
         "abuse", "mistreatment", "cruelty", "torture", "beaten", "chained",
-        "caged", "starving", "malnourished", "neglected", "abandoned"
+        "caged", "starving", "malnourished", "neglected", "abandoned",
+        "chained", "caged", "starving", "malnourished", "abused", "mistreated"
     ]
     
     death_keywords = [
         "death", "dead", "dying", "corpse", "carcass", "deceased", "lifeless",
-        "motionless", "still", "rigid", "pale", "cold"
+        "motionless", "still", "rigid", "pale", "cold", "skull", "bones",
+        "grave", "tombstone", "funeral", "coffin", "burial"
     ]
     
     suspicious_keywords = [
         "lying", "laying", "ground", "floor", "side", "horizontal", "flat",
-        "unconscious", "sleeping", "resting", "still", "motionless", "quiet"
+        "unconscious", "sleeping", "resting", "still", "motionless", "quiet",
+        "on ground", "on floor", "lying down", "sleeping", "unconscious"
     ]
     
-    adult_content = top_nsfw["label"].lower() == "nsfw"
+    if eva02_model is not None:
+        adult_keywords = ["nsfw", "explicit", "nude", "naked", "sexual", "adult", "mature"]
+        adult_content = any(keyword in combined_labels for keyword in adult_keywords)
+    else:
+        adult_content = top_nsfw["label"].lower() == "nsfw"
+    
     violence = any(keyword in combined_labels for keyword in violence_keywords)
     suffering = any(keyword in combined_labels for keyword in suffering_keywords)
     abuse = any(keyword in combined_labels for keyword in abuse_keywords)

requirements.txt

@@ -1,3 +1,7 @@
 transformers
 torch
-gradio
+gradio
+timm
+requests
+pillow
+numpy