# -*- coding: utf-8 -*- """Face_Segmentation_Project.ipynb Automatically generated by Colab. Original file is located at https://colab.research.google.com/drive/1sLYDzM6pK7GBZW2JH7xMoVa0kOYYS9B5 ## STEP 1 — Google Drive Mount Karo """ from google.colab import drive drive.mount('/content/drive') """## ✅ STEP 2 — Set Path """ # Pehle path define karo (Apne folder ke hisaab se check kar lena) TRAIN_NPY = "/content/drive/MyDrive/Face_Segmentation_Project/data/raw/Part 1- Train data - images.npy" import numpy as np print("Loading .npy file... (thoda time lagega 1.26GB hai)") data = np.load(TRAIN_NPY, allow_pickle=True) print("\n--- DATA INFO ---") print("Type :", type(data)) print("Shape :", data.shape) print("Dtype :", data.dtype) # Agar object/dictionary ho if data.dtype == object: print("\nObject type hai, keys dekho:") try: d = data.item() print("Keys:", d.keys()) except: print("Array of objects - length:", len(data)) print("First element type:", type(data[0])) print("First element shape:", data[0].shape if hasattr(data[0], 'shape') else "no shape") else: print("\nMin value :", data.min()) print("Max value :", data.max()) print("Total images:", data.shape[0]) # STEP 3 — Data Visualization (Andar kya hai dekho) import matplotlib.pyplot as plt # Pehla sample nikaalte hain sample = data[0] image_data = sample[0] # Pehla part image hona chahiye mask_data = sample[1] # Dusra part mask ya annotations print(f"Image shape: {image_data.shape}") print(f"Mask/Metadata type: {type(mask_data)}") # Ek bar visualize karke dekhte hain plt.imshow(image_data) plt.title("Sample Image") plt.show() # ✅ STEP 4 — Mask Check Karo (Andar kya hai?) # List ka content check karte hain print("Mask List Content:") print(mask_data) # Agar list ke pehle element mein 'notes' ya 'points' jaisi koi key hai toh: if len(mask_data) > 0: print("\nFirst element in list:", mask_data[0]) # 🛠️ STEP 5 — Coordinates se Binary Mask Banana import cv2 import numpy as np def create_mask(image_shape, mask_list): # Ek khali black image banao (0s) mask = np.zeros(image_shape[:2], dtype=np.uint8) h, w = image_shape[:2] for item in mask_list: # Normalized coordinates ko pixel values mein badlo p1 = item['points'][0] p2 = item['points'][1] x1, y1 = int(p1['x'] * w), int(p1['y'] * h) x2, y2 = int(p2['x'] * w), int(p2['y'] * h) # Is area ko white (255) kar do cv2.rectangle(mask, (x1, y1), (x2, y2), 255, -1) return mask # Test karte hain test_mask = create_mask(image_data.shape, mask_data) # Visualize Image + Mask plt.figure(figsize=(10, 5)) plt.subplot(1, 2, 1) plt.imshow(image_data); plt.title("Original Image") plt.subplot(1, 2, 2) plt.imshow(test_mask, cmap='gray'); plt.title("Generated Mask") plt.show() # ✅ STEP 6 — Data Preprocessing for U-Net import cv2 import numpy as np from tqdm import tqdm IMG_HEIGHT = 256 IMG_WIDTH = 256 X = [] y = [] print("Processing 409 images and masks...") for i in tqdm(range(len(data))): img_array = data[i][0] mask_list = data[i][1] img_resized = cv2.resize(img_array, (IMG_WIDTH, IMG_HEIGHT)) if len(img_resized.shape) == 2: img_resized = cv2.cvtColor(img_resized, cv2.COLOR_GRAY2RGB) elif img_resized.shape[2] == 4: img_resized = cv2.cvtColor(img_resized, cv2.COLOR_RGBA2RGB) m_array = create_mask(img_array.shape, mask_list) mask_resized = cv2.resize(m_array, (IMG_WIDTH, IMG_HEIGHT)) img_norm = img_resized / 255.0 mask_norm = mask_resized / 255.0 # ✅ Original image add karo X.append(img_norm) y.append(mask_norm) # ✅ Augmentation 1 — Horizontal Flip X.append(cv2.flip(img_norm, 1)) y.append(cv2.flip(mask_norm, 1)) # ✅ Augmentation 2 — Rotation 15 degree M = cv2.getRotationMatrix2D((128, 128), 15, 1.0) X.append(cv2.warpAffine(img_norm, M, (256, 256))) y.append(cv2.warpAffine(mask_norm, M, (256, 256))) # ✅ Augmentation 3 — Brightness Change bright = np.clip(img_norm * 1.2, 0, 1) X.append(bright) y.append(mask_norm) X = np.array(X, dtype=np.float32) y = np.array(y, dtype=np.float32) y = np.expand_dims(y, axis=-1) print("\n✅ Preprocessing Complete!") print(f"Images Shape: {X.shape}") print(f"Masks Shape : {y.shape}") # 🚀 STEP 7 — Model Building: U-Net with MobileNetV2 from sklearn.model_selection import train_test_split X_train, X_temp, y_train, y_temp = train_test_split( X, y, test_size=0.2, random_state=42 ) X_val, X_test, y_val, y_test = train_test_split( X_temp, y_temp, test_size=0.5, random_state=42 ) print(f"Training data : {X_train.shape}") print(f"Validation data : {X_val.shape}") print(f"Test data : {X_test.shape}") from tensorflow.keras.applications import MobileNetV2 from tensorflow.keras.layers import Input, Conv2D, UpSampling2D, Concatenate, BatchNormalization, Activation, Dropout from tensorflow.keras.models import Model import tensorflow as tf def build_unet(input_shape): inputs = Input(input_shape) base_model = MobileNetV2(input_tensor=inputs, weights="imagenet", include_top=False) # ✅ FIX 1: Encoder freeze karo base_model.trainable = False for layer in base_model.layers[-30:]: layer.trainable = True # Skip connections s1 = inputs s2 = base_model.get_layer("block_1_expand_relu").output s3 = base_model.get_layer("block_3_expand_relu").output s4 = base_model.get_layer("block_6_expand_relu").output bridge = base_model.get_layer("block_13_expand_relu").output # ✅ FIX 2: Dropout add kiya har decoder block mein u1 = UpSampling2D((2, 2))(bridge) u1 = Concatenate()([u1, s4]) u1 = Conv2D(256, 3, padding="same", kernel_initializer="he_normal")(u1) u1 = BatchNormalization()(u1) u1 = Activation("relu")(u1) u1 = Dropout(0.3)(u1) # ✅ u2 = UpSampling2D((2, 2))(u1) u2 = Concatenate()([u2, s3]) u2 = Conv2D(128, 3, padding="same", kernel_initializer="he_normal")(u2) u2 = BatchNormalization()(u2) u2 = Activation("relu")(u2) u2 = Dropout(0.3)(u2) # ✅ u3 = UpSampling2D((2, 2))(u2) u3 = Concatenate()([u3, s2]) u3 = Conv2D(64, 3, padding="same", kernel_initializer="he_normal")(u3) u3 = BatchNormalization()(u3) u3 = Activation("relu")(u3) u3 = Dropout(0.2)(u3) # ✅ u4 = UpSampling2D((2, 2))(u3) u4 = Concatenate()([u4, s1]) u4 = Conv2D(32, 3, padding="same", kernel_initializer="he_normal")(u4) u4 = BatchNormalization()(u4) u4 = Activation("relu")(u4) outputs = Conv2D(1, 1, activation="sigmoid")(u4) return Model(inputs, outputs) model = build_unet((256, 256, 3)) # ✅ FIX 3: Combined Loss def dice_coefficient(y_true, y_pred): y_true_f = tf.cast(tf.reshape(y_true, [-1]), tf.float32) y_pred_f = tf.cast(tf.reshape(y_pred, [-1]), tf.float32) intersection = tf.reduce_sum(y_true_f * y_pred_f) return (2. * intersection + 1.0) / (tf.reduce_sum(y_true_f) + tf.reduce_sum(y_pred_f) + 1.0) def dice_loss(y_true, y_pred): return 1 - dice_coefficient(y_true, y_pred) def combined_loss(y_true, y_pred): return dice_loss(y_true, y_pred) + tf.keras.losses.binary_crossentropy(y_true, y_pred) model.compile( optimizer=tf.keras.optimizers.Adam(learning_rate=1e-3), loss=combined_loss, # ✅ Combined loss metrics=[dice_coefficient, "accuracy"] ) print("✅ Model Compiled Successfully!") # Training se PEHLE yeh check karo: print(f"X_train shape: {X_train.shape}") print(f"X_val shape : {X_val.shape}") print(f"X_test shape : {X_test.shape}") # 🚀 STEP 9 — Model Training (Fit) from tensorflow.keras.callbacks import ModelCheckpoint, ReduceLROnPlateau, EarlyStopping callbacks = [ ModelCheckpoint("best_face_model.keras", monitor='val_dice_coefficient', mode='max', save_best_only=True, verbose=1), ReduceLROnPlateau(monitor='val_loss', factor=0.2, patience=5, min_lr=1e-7, verbose=1), EarlyStopping(monitor='val_loss', patience=12, restore_best_weights=True, verbose=1) ] print("🚀 Training shuru ho rahi hai!") history = model.fit( X_train, y_train, validation_data=(X_val, y_val), epochs=50, batch_size=8, callbacks=callbacks, shuffle=True ) print("\n✅ Training Complete!") # Step - Evaluation Metrix from sklearn.metrics import f1_score import time import numpy as np def evaluate_model(model, X_test, y_test): y_pred = model.predict(X_test) y_pred_binary = (y_pred > 0.5).astype(np.float32) # 1. Dice Coefficient intersection = np.sum(y_test * y_pred_binary) dice = (2. * intersection + 1.0) / ( np.sum(y_test) + np.sum(y_pred_binary) + 1.0 ) # 2. IoU Score union = np.sum(y_test) + np.sum(y_pred_binary) - intersection iou = intersection / union # 3. F1 Score f1 = f1_score( y_test.flatten().astype(int), y_pred_binary.flatten().astype(int) ) # 4. ✅ Inference Speed — FIXED VERSION model.predict(X_test[:1], verbose=0) # Warmup 1 model.predict(X_test[:1], verbose=0) # Warmup 2 times = [] for _ in range(5): start = time.time() model.predict(X_test[:1], verbose=0) times.append((time.time() - start) * 1000) speed = np.mean(times) print("="*45) print(f"Dice Coefficient : {dice:.4f} (Target >0.92)") print(f"IoU Score : {iou:.4f} (Target >0.88)") print(f"F1 Score : {f1:.4f} (Target >0.90)") print(f"Inference Speed : {speed:.1f}ms (Target <100ms)") print("="*45) # Dobara chalao evaluate_model(model, X_test, y_test) # STEP 11 — Training Curves Plot import matplotlib.pyplot as plt plt.figure(figsize=(12, 4)) plt.subplot(1, 2, 1) plt.plot(history.history['dice_coefficient'], label='Train Dice') plt.plot(history.history['val_dice_coefficient'], label='Val Dice') plt.title('Dice Coefficient Progress') plt.xlabel('Epoch') plt.legend() plt.subplot(1, 2, 2) plt.plot(history.history['loss'], label='Train Loss') plt.plot(history.history['val_loss'], label='Val Loss') plt.title('Loss Progress') plt.xlabel('Epoch') plt.legend() plt.tight_layout() plt.show() import matplotlib.pyplot as plt def visualize_prediction(index): img = X_val[index] gt_mask = y_val[index] # Model se prediction lo pred_mask = model.predict(np.expand_dims(img, axis=0))[0] # Thresholding (0.5 se upar white, niche black) pred_mask_binary = (pred_mask > 0.5).astype(np.uint8) plt.figure(figsize=(12, 4)) plt.subplot(1, 3, 1) plt.imshow(img); plt.title("Original Image") plt.subplot(1, 3, 2) plt.imshow(gt_mask.squeeze(), cmap='gray'); plt.title("Ground Truth Mask") plt.subplot(1, 3, 3) plt.imshow(pred_mask_binary.squeeze(), cmap='gray'); plt.title("Predicted Mask") plt.show() # Pehli 3 validation images check karte hain for i in [0, 5, 10]: visualize_prediction(i) # 🚀 STEP 12 — Improving Accuracy with Hugging Face (SegFormer) !pip install -q transformers datasets from transformers import SegformerImageProcessor, SegformerForSemanticSegmentation import torch from PIL import Image # 1. Pre-trained Model Load Karein device = "cuda" if torch.cuda.is_available() else "cpu" # Processor images ko transform karega, model segmentation karega processor = SegformerImageProcessor.from_pretrained("nvidia/mit-b0") hf_model = SegformerForSemanticSegmentation.from_pretrained("nvidia/mit-b0", num_labels=1, ignore_mismatched_sizes=True).to(device) print("✅ Hugging Face SegFormer Model Loaded!") # 🛠️ STEP 12 — SegFormer Dataset Class import torch from torch.utils.data import Dataset, DataLoader import numpy as np class FaceDataset(Dataset): def __init__(self, images, masks, processor): self.images = images self.masks = masks self.processor = processor def __len__(self): return len(self.images) def __getitem__(self, idx): # Image aur Mask uthayein # (X_train already normalized hai, hume original format chahiye processor ke liye) image = (self.images[idx] * 255).astype(np.uint8) mask = self.masks[idx].squeeze().astype(np.longlong) # SegFormer expects long for masks # Processor se transform karein encoded_inputs = self.processor(image, mask, return_tensors="pt") # Squeeze batch dimension jo processor add karta hai for k, v in encoded_inputs.items(): encoded_inputs[k] = v.squeeze(0) return encoded_inputs # 1. Dataset Objects Banayein train_dataset = FaceDataset(X_train, y_train, processor) val_dataset = FaceDataset(X_val, y_val, processor) # 2. DataLoaders Banayein train_dataloader = DataLoader(train_dataset, batch_size=4, shuffle=True) val_dataloader = DataLoader(val_dataset, batch_size=4) print(f"✅ DataLoaders Ready! Training samples: {len(train_dataset)}") # 🚀 STEP 13 — Fine-tuning SegFormer (Training Loop) from tqdm.notebook import tqdm from torch.optim import AdamW # 1. Optimizer setup optimizer = AdamW(hf_model.parameters(), lr=5e-5) # 2. Training Loop num_epochs = 10 # SegFormer jaldi seekh jata hai hf_model.train() print("🚀 SegFormer Training shuru ho rahi hai...") for epoch in range(num_epochs): train_loss = 0 # Tqdm progress bar ke liye pbar = tqdm(train_dataloader, desc=f"Epoch {epoch+1}") for batch in pbar: # Data ko GPU/CPU par bhejein pixel_values = batch["pixel_values"].to(device) labels = batch["labels"].to(device) # Forward pass outputs = hf_model(pixel_values=pixel_values, labels=labels) loss = outputs.loss # Backward pass loss.backward() optimizer.step() optimizer.zero_grad() train_loss += loss.item() pbar.set_postfix({'loss': loss.item()}) avg_train_loss = train_loss / len(train_dataloader) print(f"✅ Epoch {epoch+1} Complete | Average Loss: {avg_train_loss:.4f}") print("\n🎉 SegFormer Fine-tuning Finished!") # 🎨 STEP 14 — SegFormer Prediction & Visualization import torch.nn as nn def visualize_segformer_prediction(index): # 1. Validation set se data lein inputs = val_dataset[index] pixel_values = inputs["pixel_values"].unsqueeze(0).to(device) gt_mask = inputs["labels"].numpy() # 2. Prediction karein hf_model.eval() with torch.no_grad(): outputs = hf_model(pixel_values=pixel_values) logits = outputs.logits # [1, num_labels, height/4, width/4] # 3. Logits ko original image size (256x256) par upsample karein upsampled_logits = nn.functional.interpolate( logits, size=(256, 256), mode='bilinear', align_corners=False ) # 4. Binary mask banayein (Squeeze karke threshold lagayein) pred_mask = torch.sigmoid(upsampled_logits).cpu().numpy().squeeze() pred_mask_binary = (pred_mask > 0.5).astype(np.uint8) # 5. Plotting plt.figure(figsize=(12, 4)) # Original Image (Pixel values ko denormalize karke dikhayenge) display_img = pixel_values.cpu().squeeze().permute(1, 2, 0).numpy() display_img = (display_img - display_img.min()) / (display_img.max() - display_img.min()) plt.subplot(1, 3, 1) plt.imshow(display_img) plt.title("Original Image") plt.subplot(1, 3, 2) plt.imshow(gt_mask, cmap='gray') plt.title("Ground Truth Mask") plt.subplot(1, 3, 3) plt.imshow(pred_mask_binary, cmap='gray') plt.title("SegFormer Prediction") plt.show() # Pehle 3 validation images par results dekhein for i in [0, 5, 10]: visualize_segformer_prediction(i) # 💾 STEP 15 — Model aur Processor ko Save Karein import os # 1. Ek folder banate hain model ke liye save_directory = "my_face_segformer_model" os.makedirs(save_directory, exist_ok=True) # 2. Model ke weights aur config save karein hf_model.save_pretrained(save_directory) # 3. Processor (jo image resize aur normalize karta hai) bhi save karein processor.save_pretrained(save_directory) print(f"✅ Model successfully saved in: {save_directory}") # --- Google Colab Users ke liye --- # Agar aap chahte hain ki ye hamesha ke liye safe rahe, toh ise Zip karke download kar lein !zip -r face_model.zip my_face_segformer_model/ print("📦 Zip file ban gayi hai! Ab aap ise left side ke 'Files' tab se download kar sakte hain.") from google.colab import drive import shutil import os # 1. Drive mount karein drive.mount('/content/drive') # 2. Aapka bataya hua path define karein # Hum Face_Segmentation_Project ke andar 'saved_model' naam ka folder banayenge base_path = '/content/drive/MyDrive/Face_Segmentation_Project' model_save_path = os.path.join(base_path, 'trained_segformer_model') # 3. Folder banayein agar nahi hai toh os.makedirs(model_save_path, exist_ok=True) # 4. Colab ki temporary files ko Drive par copy karein source_folder = '/content/my_face_segformer_model' # Purani files clear karke fresh copy karein if os.path.exists(model_save_path): shutil.rmtree(model_save_path) shutil.copytree(source_folder, model_save_path) print(f"✅ Kaam ho gaya! Aapka model yahan save hai:") print(f"📂 {model_save_path}") import os # Actor folders banao actors = [ "Robert_Downey_Jr", "Scarlett_Johansson", "Chris_Evans", "Tom_Cruise", "Leonardo_DiCaprio", "Brad_Pitt", "Chris Hemsworth" ] for actor in actors: os.makedirs(f"/content/actor_db/{actor}", exist_ok=True) print("✅ Actor folders ban gaye!") print(os.listdir("/content/actor_db")) import requests import os actor_images = { "Robert_Downey_Jr": [ "https://wallpapers.com/images/hd/focused-photography-robert-downey-jr-pjs7jatnx0yfofsc.jpg", "https://static1.srcdn.com/wordpress/wp-content/uploads/2024/07/instar53643496.jpg" ], "Scarlett_Johansson": [ "https://tse2.mm.bing.net/th/id/OIP.OwNfXHWEAynGbQwhmUNQWwHaKk?rs=1&pid=ImgDetMain&o=7&rm=3", "https://m.media-amazon.com/images/M/MV5BMjAyNjE1NjgyMF5BMl5BanBnXkFtZTcwMjg0ODY1NA@@._V1_.jpg" ], "Chris_Evans": [ "https://cdn.britannica.com/28/215028-050-94E9EA1E/American-actor-Chris-Evans-2019.jpg", "https://tse4.mm.bing.net/th/id/OIP._aupdpOey-23R2KHLQd0OgHaLH?rs=1&pid=ImgDetMain&o=7&rm=3" ], "Tom_Cruise": [ "https://image.tmdb.org/t/p/original/8qBylBsQf4llkGrWR3qAsOtOU8O.jpg", "https://m.media-amazon.com/images/M/MV5BMmU1YWU1NmMtMjAyMi00MjFiLWFmZmUtOTc1ZjI5ODkxYmQyXkEyXkFqcGc@._V1_FMjpg_UX1000_.jpg" ], "Leonardo_DiCaprio": [ "https://image.tmdb.org/t/p/original/aLUFp0zWpLVyIOgY0scIpuuKZLE.jpg", "https://wallpapers.com/images/hd/leonardo-dicaprio-actor-nl2l96xppg1rj9hz.jpg" ], "Brad_Pitt": [ "https://fr.web.img2.acsta.net/pictures/20/02/10/10/37/1374948.jpg", "https://www.gratistodo.com/wp-content/uploads/2016/09/Brad-Pitt-7.jpg" ], "Chris Hemsworth": [ "https://cdn.britannica.com/92/215392-050-96A4BC1D/Australian-actor-Chris-Hemsworth-2019.jpg", "https://static1.cbrimages.com/wordpress/wp-content/uploads/2023/06/chris-hemsworth-thor-love-and-thunder-mcu.jpg" ] } success = 0 fail = 0 for actor_name, urls in actor_images.items(): for i, url in enumerate(urls): try: response = requests.get(url, timeout=15) if response.status_code == 200: path = f"/content/actor_db/{actor_name}/{i}.jpg" with open(path, 'wb') as f: f.write(response.content) print(f"✅ {actor_name} — photo {i+1} downloaded") success += 1 else: print(f"❌ {actor_name} — photo {i+1} failed ({response.status_code})") fail += 1 except Exception as e: print(f"❌ {actor_name} — photo {i+1} error: {e}") fail += 1 print(f"\n✅ Success: {success} | ❌ Failed: {fail}") # Verify karo print("\n📁 Database Status:") for actor in os.listdir("/content/actor_db"): photos = os.listdir(f"/content/actor_db/{actor}") print(f" {actor}: {len(photos)} photos") !pip install deepface -q !pip install tf-keras -q print("✅ DeepFace installed!") # Test Karo from deepface import DeepFace import cv2 import numpy as np # Test — ek image pe try karo test_img = "/content/actor_db/Robert_Downey_Jr/0.jpg" result = DeepFace.analyze( img_path=test_img, actions=['age', 'gender'], enforce_detection=False ) print("✅ DeepFace working!") print(f"Age: {result[0]['age']}") print(f"Gender: {result[0]['dominant_gender']}") # Step 1 — Sahi versions install karo !pip install tf-keras==2.19.0 -q !pip install keras==2.19.0 -q print("✅ Done!") import keras import tensorflow as tf def dice_coefficient(y_true, y_pred): y_true_f = tf.cast(tf.reshape(y_true,[-1]),tf.float32) y_pred_f = tf.cast(tf.reshape(y_pred,[-1]),tf.float32) i = tf.reduce_sum(y_true_f*y_pred_f) return (2.*i+1.)/(tf.reduce_sum(y_true_f)+tf.reduce_sum(y_pred_f)+1.) def combined_loss(y_true, y_pred): return 1-dice_coefficient(y_true,y_pred)+tf.keras.losses.binary_crossentropy(y_true,y_pred) # Pehle load karo model = keras.models.load_model( "/content/best_face_model.keras", custom_objects={ 'dice_coefficient': dice_coefficient, 'combined_loss': combined_loss } ) # Weights alag save karo model.save_weights("/content/model_weights.weights.h5") print("✅ Weights saved!") # Commented out IPython magic to ensure Python compatibility. # %%writefile /content/app.py # import os # os.environ['CUDA_VISIBLE_DEVICES'] = '-1' # os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3' # # import streamlit as st # import numpy as np # import cv2 # from PIL import Image # import tensorflow as tf # import time # import io # # st.set_page_config(page_title="Scene Cast AI", page_icon="🎬", layout="wide") # # st.markdown("""""", unsafe_allow_html=True) # # # ── Model Architecture Rebuild + Weights Load ── # from tensorflow.keras.applications import MobileNetV2 # from tensorflow.keras.layers import Input, Conv2D, UpSampling2D, Concatenate, BatchNormalization, Activation, Dropout # from tensorflow.keras.models import Model # # def dice_coefficient(y_true, y_pred): # y_true_f = tf.cast(tf.reshape(y_true,[-1]),tf.float32) # y_pred_f = tf.cast(tf.reshape(y_pred,[-1]),tf.float32) # i = tf.reduce_sum(y_true_f*y_pred_f) # return (2.*i+1.)/(tf.reduce_sum(y_true_f)+tf.reduce_sum(y_pred_f)+1.) # # def combined_loss(y_true, y_pred): # return 1-dice_coefficient(y_true,y_pred)+tf.keras.losses.binary_crossentropy(y_true,y_pred) # # def build_unet(input_shape=(256,256,3)): # inputs = Input(input_shape) # base = MobileNetV2(input_tensor=inputs, weights=None, include_top=False) # base.trainable = False # s1=inputs # s2=base.get_layer("block_1_expand_relu").output # s3=base.get_layer("block_3_expand_relu").output # s4=base.get_layer("block_6_expand_relu").output # bridge=base.get_layer("block_13_expand_relu").output # u1=UpSampling2D((2,2))(bridge) # u1=Concatenate()([u1,s4]) # u1=Conv2D(256,3,padding="same",kernel_initializer="he_normal")(u1) # u1=BatchNormalization()(u1); u1=Activation("relu")(u1); u1=Dropout(0.3)(u1) # u2=UpSampling2D((2,2))(u1) # u2=Concatenate()([u2,s3]) # u2=Conv2D(128,3,padding="same",kernel_initializer="he_normal")(u2) # u2=BatchNormalization()(u2); u2=Activation("relu")(u2); u2=Dropout(0.3)(u2) # u3=UpSampling2D((2,2))(u2) # u3=Concatenate()([u3,s2]) # u3=Conv2D(64,3,padding="same",kernel_initializer="he_normal")(u3) # u3=BatchNormalization()(u3); u3=Activation("relu")(u3); u3=Dropout(0.2)(u3) # u4=UpSampling2D((2,2))(u3) # u4=Concatenate()([u4,s1]) # u4=Conv2D(32,3,padding="same",kernel_initializer="he_normal")(u4) # u4=BatchNormalization()(u4); u4=Activation("relu")(u4) # outputs=Conv2D(1,1,activation="sigmoid")(u4) # return Model(inputs,outputs) # # @st.cache_resource(show_spinner=False) # def load_model(): # with tf.device('/CPU:0'): # m = build_unet() # m.load_weights("/content/model_weights.weights.h5", by_name=True, skip_mismatch=True) # m.compile( # optimizer=tf.keras.optimizers.Adam(1e-4), # loss=combined_loss, # metrics=[dice_coefficient] # ) # return m # # def create_overlay(image, mask): # img=cv2.resize(np.array(image),(256,256)) # cm=np.zeros_like(img); cm[mask>127]=[245,197,24] # ov=cv2.addWeighted(img,0.55,cm,0.45,0) # ctrs,_=cv2.findContours(mask,cv2.RETR_EXTERNAL,cv2.CHAIN_APPROX_SIMPLE) # cv2.drawContours(ov,ctrs,-1,(245,197,24),2) # return ov # # def face_count(mask): # ctrs,_=cv2.findContours(mask,cv2.RETR_EXTERNAL,cv2.CHAIN_APPROX_SIMPLE) # return len([c for c in ctrs if cv2.contourArea(c)>256*256*0.005]) # # def img_bytes(arr): # buf=io.BytesIO(); Image.fromarray(arr).save(buf,format="PNG"); return buf.getvalue() # # def get_face_crops(image,mask): # img=cv2.resize(np.array(image),(256,256)) # ctrs,_=cv2.findContours(mask,cv2.RETR_EXTERNAL,cv2.CHAIN_APPROX_SIMPLE) # crops=[] # for c in ctrs: # if cv2.contourArea(c)>256*256*0.005: # x,y,w,h=cv2.boundingRect(c) # crop=img[max(0,y-10):min(255,y+h+10),max(0,x-10):min(255,x+w+10)] # if crop.size>0: crops.append(crop) # return crops # # def identify_actors(crops): # results=[]; db="/content/actor_db" # if not os.path.exists(db): return results # has_photos=any(len(os.listdir(f"{db}/{a}"))>0 for a in os.listdir(db) if os.path.isdir(f"{db}/{a}")) # if not has_photos: return results # try: # from deepface import DeepFace # for crop in crops: # try: # dfs=DeepFace.find(img_path=crop,db_path=db,model_name="VGG-Face",enforce_detection=False,silent=True) # if len(dfs)>0 and len(dfs[0])>0: # best=dfs[0].iloc[0] # name=best['identity'].split('/')[-2].replace('_',' ') # dist=best.get('distance',1) # conf=max(0,(1-dist)*100) # results.append({"name":name,"confidence":conf}) # else: # results.append({"name":"Unknown Actor","confidence":0}) # except: # results.append({"name":"Unknown Actor","confidence":0}) # except ImportError: # pass # return results # # with st.sidebar: # st.markdown("""
#
🎬
#
Scene Cast AI
#
Face Segmentation Engine
#
""",unsafe_allow_html=True) # st.markdown("**⚙️ SETTINGS**") # threshold=st.slider("Confidence Threshold",0.1,0.9,0.5,0.05) # show_ov=st.checkbox("Face Overlay",value=True) # show_mask=st.checkbox("Binary Mask",value=True) # show_conf=st.checkbox("Confidence Map",value=False) # show_actor=st.checkbox("🎭 Actor Recognition",value=True) # st.markdown("---") # st.markdown("""**📊 MODEL INFO** #
# 🏗️ Architecture: U-Net
# 🔧 Encoder: MobileNetV2
# 📦 Input: 256×256
# 📈 Val Dice: 0.8742 #
""",unsafe_allow_html=True) # # st.markdown("""
#
🎬 Powered by Deep Learning
#

Scene Cast AI

#

Real-time face segmentation + actor identification.

#
#
""",unsafe_allow_html=True) # # with st.spinner("🔄 Loading AI Model..."): # try: # model=load_model() # st.success("✅ Model loaded!", icon="🤖") # except Exception as e: # st.error(f"❌ Error: {e}") # st.stop() # # st.markdown('
📁 Upload Movie Scene
',unsafe_allow_html=True) # f=st.file_uploader("Upload",type=["jpg","jpeg","png","webp"],label_visibility="collapsed") # # if f: # # 🔥 FIX: Streamlit upload → OpenCV decode # file_bytes = np.asarray(bytearray(f.read()), dtype=np.uint8) # img = cv2.imdecode(file_bytes, cv2.IMREAD_COLOR) # # if img is None: # st.error("❌ Image read nahi ho pa rahi. Dusri image try karo.") # st.stop() # # img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) # image = Image.fromarray(img) # # # 👇 same code as before # ir = cv2.resize(np.array(image),(256,256))/255.0 # ii = np.expand_dims(ir,0).astype(np.float32) # # with st.spinner("🧠 Analyzing..."): # with tf.device('/CPU:0'): # t0 = time.time() # pred = model.predict(ii,verbose=0)[0] # spd = (time.time()-t0)*1000 # # mask = (pred.squeeze()>threshold).astype(np.uint8)*255 # fc = face_count(mask) # cov = (np.sum(mask>0)/(256*256))*100 # # st.markdown('
🎯 Detection Results
',unsafe_allow_html=True) # # ncols = 1+show_ov+show_mask+show_conf # dcols = st.columns(ncols) # with dcols[0]: # st.markdown('
🎬 Original Frame
',unsafe_allow_html=True) # st.image(image,use_container_width=True) # st.markdown(f'
📐 {image.width}×{image.height}px | 👥 ~{fc} face(s)
',unsafe_allow_html=True) # idx=1 # if show_ov: # ov=create_overlay(image,mask) # with dcols[idx]: # st.markdown('
✨ Face Overlay
',unsafe_allow_html=True) # st.image(ov,use_container_width=True) # st.markdown(f'
🟡 Gold=face | {cov:.1f}%
',unsafe_allow_html=True) # idx+=1 # if show_mask: # with dcols[idx]: # st.markdown('
🔲 Binary Mask
',unsafe_allow_html=True) # st.image(mask,use_container_width=True) # st.markdown(f'
⬜ White=face | {threshold:.2f}
',unsafe_allow_html=True) # idx+=1 # if show_conf: # cm=cv2.applyColorMap((pred.squeeze()*255).astype(np.uint8),cv2.COLORMAP_INFERNO) # cm=cv2.cvtColor(cm,cv2.COLOR_BGR2RGB) # with dcols[idx]: # st.markdown('
🌡️ Confidence
',unsafe_allow_html=True) # st.image(cm,use_container_width=True) # st.markdown('
🔴 Hot=face
',unsafe_allow_html=True) # if show_actor: # st.markdown('
🎭 Actor Recognition
',unsafe_allow_html=True) # with st.spinner("🔍 Identifying..."): # crops=get_face_crops(image,mask); actors=identify_actors(crops) # if actors: # acols=st.columns(min(len(actors),4)) # for i,a in enumerate(actors): # cc="#4ade80" if a['confidence']>60 else "#fb923c" # with acols[i%4]: # st.markdown(f"""
🎭
#
{a['name']}
#
Confidence: {a['confidence']:.1f}%
#
""",unsafe_allow_html=True) # else: # st.info("🔍 Actor DB mein photos upload karo!",icon="ℹ️") # st.markdown('
📊 Performance
',unsafe_allow_html=True) # sc="fast" if spd<100 else "warn" # st.markdown(f"""
#
🎯
0.8845
Dice
>0.92
#
📐
0.7929
IoU
>0.88
#
⚖️
0.8849
F1
>0.90
#
⚡
{spd:.0f}ms
Speed
<100ms
#
""",unsafe_allow_html=True) # c1,c2,c3=st.columns(3) # for col,icon,val,lbl,clr in [(c1,"👥",str(fc),"Faces","#f5c518"),(c2,"📊",f"{cov:.1f}%","Coverage","#4cc9f0"),(c3,"🎚️",f"{threshold:.2f}","Threshold","#4ade80")]: # with col: # st.markdown(f"""
#
{icon}
#
{val}
#
{lbl}
#
""",unsafe_allow_html=True) # st.markdown('
📥 Export
',unsafe_allow_html=True) # d1,d2,d3=st.columns(3) # with d1: # st.download_button("⬇️ Mask",img_bytes(mask),"mask.png","image/png",use_container_width=True) # with d2: # ov2=create_overlay(image,mask) # st.download_button("⬇️ Overlay",img_bytes(ov2),"overlay.png","image/png",use_container_width=True) # with d3: # log=f"File:{f.name}\nFaces:{fc}\nSpeed:{spd:.1f}ms\nDice:0.8845\nIoU:0.7929" # st.download_button("⬇️ Log",log,"log.txt","text/plain",use_container_width=True) # else: # st.markdown("""
#
🎬
#
Ready to Detect & Identify
#
Upload movie screenshot → AI detects + identifies actors!
#
""",unsafe_allow_html=True) !sed -i 's/use_container_width=True/use_column_width=True/g' /content/app.py # Yeh cell chalao — warning fix hogi !sed -i 's/use_column_width=True/use_container_width=True/g' /content/app.py print("✅ Fixed!") !pip install streamlit pyngrok -q # Ek Cell Mein Sab !fuser -k 8501/tcp !pkill -f streamlit import time time.sleep(2) import subprocess, threading def run(): subprocess.Popen([ 'streamlit', 'run', '/content/app.py', '--server.port=8501', '--server.headless=true', '--server.enableCORS=false', '--server.enableXsrfProtection=false' ]) threading.Thread(target=run).start() time.sleep(15) from google.colab.output import eval_js url = eval_js("google.colab.kernel.proxyPort(8501)") print(f"✅ URL: {url}")