imageAI.py

try:
  import google.colab
  IN_COLAB = True
  from google.colab import drive,files
  from google.colab import output
  drive.mount('/gdrive')
  Gbase="/gdrive/MyDrive/generate/"
  cache_dir="/gdrive/MyDrive/hf/"
  import sys
  sys.path.append(Gbase)
except:
  IN_COLAB = False
  Gbase="./"
  cache_dir="./hf/"


import cv2,os
import numpy as np
import random,string
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.utils.data import Dataset, DataLoader

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"Using device: {device}")

IMAGE_SIZE = 64
NUM_SAMPLES = 1000
BATCH_SIZE = 4
EPOCHS = 500
LEARNING_RATE = 0.001


def generate_sample(num_shapes=1):
    image = np.zeros((IMAGE_SIZE, IMAGE_SIZE), dtype=np.uint8)
    instructions = []
    
    #num_shapes = random.randint(1, 3)
    for _ in range(num_shapes):
        shape = random.choice(['line', 'rectangle', 'circle', 'ellipse', 'polygon'])
        color = random.randint(0, 255)
        thickness = random.randint(1, 3)
        
        if shape == 'line':
            start_point = (random.randint(0, IMAGE_SIZE), random.randint(0, IMAGE_SIZE))
            end_point = (random.randint(0, IMAGE_SIZE), random.randint(0, IMAGE_SIZE))
            cv2.line(image, start_point, end_point, color, thickness)
            instructions.append(f"cv2.line(image, {start_point}, {end_point}, {color}, {thickness})")
        
        elif shape == 'rectangle':
            start_point = (random.randint(0, IMAGE_SIZE - 10), random.randint(0, IMAGE_SIZE - 10))
            end_point = (start_point[0] + random.randint(10, IMAGE_SIZE - start_point[0]),
                         start_point[1] + random.randint(10, IMAGE_SIZE - start_point[1]))
            cv2.rectangle(image, start_point, end_point, color, thickness)
            instructions.append(f"cv2.rectangle(image, {start_point}, {end_point}, {color}, {thickness})")
        
        elif shape == 'circle':
            center = (random.randint(10, IMAGE_SIZE - 10), random.randint(10, IMAGE_SIZE - 10))
            radius = random.randint(5, min(center[0], center[1], IMAGE_SIZE - center[0], IMAGE_SIZE - center[1]))
            cv2.circle(image, center, radius, color, thickness)
            instructions.append(f"cv2.circle(image, {center}, {radius}, {color}, {thickness})")
        
        elif shape == 'ellipse':
            center = (random.randint(10, IMAGE_SIZE - 10), random.randint(10, IMAGE_SIZE - 10))
            axes = (random.randint(5, 30), random.randint(5, 30))
            angle = random.randint(0, 360)
            cv2.ellipse(image, center, axes, angle, 0, 360, color, thickness)
            instructions.append(f"cv2.ellipse(image, {center}, {axes}, {angle}, 0, 360, {color}, {thickness})")
        
        elif shape == 'polygon':
            num_points = random.randint(3, 6)
            points = np.array([(random.randint(0, IMAGE_SIZE), random.randint(0, IMAGE_SIZE)) for _ in range(num_points)], np.int32)
            points = points.reshape((-1, 1, 2))
            cv2.polylines(image, [points], True, color, thickness)
            instructions.append(f"cv2.polylines(image, [{points.tolist()}], True, {color}, {thickness})")
        
    return {'image': image, 'instructions': instructions}

def generate_dataset(NUM_SAMPLES=NUM_SAMPLES,maxNumShape=3):
    dataset = []
    for _ in range(NUM_SAMPLES):
        num_shapes = random.randint(1, maxNumShape)
        sample = generate_sample(num_shapes=num_shapes)
        dataset.append(sample)
    return dataset

class ImageDataset(Dataset):
    def __init__(self, dataset):
        self.dataset = dataset
    
    def __len__(self):
        return len(self.dataset)
    
    def __getitem__(self, idx):
        sample = self.dataset[idx]
        image = torch.FloatTensor(sample['image']).unsqueeze(0) / 255.0
        return image, len(sample['instructions'])

class SimpleModel(nn.Module):
    def __init__(self, path=None):
        super(SimpleModel, self).__init__()
        self.conv1 = nn.Conv2d(1, 32, 3, padding=1)
        self.bn1 = nn.BatchNorm2d(32)
        self.conv2 = nn.Conv2d(32, 64, 3, padding=1)
        self.bn2 = nn.BatchNorm2d(64)
        self.conv3 = nn.Conv2d(64, 128, 3, padding=1)
        self.bn3 = nn.BatchNorm2d(128)
        self.pool = nn.MaxPool2d(2, 2)
        self.fc1 = nn.Linear(128 * 8 * 8, 512)
        self.fc2 = nn.Linear(512, 128)
        self.fc3 = nn.Linear(128, 1)
        self.dropout = nn.Dropout(0.5)
        
        if path and os.path.exists(path):
            self.load_state_dict(torch.load(path, map_location=device))
    
    def forward(self, x):
        x = self.pool(F.leaky_relu(self.bn1(self.conv1(x))))
        x = self.pool(F.leaky_relu(self.bn2(self.conv2(x))))
        x = self.pool(F.leaky_relu(self.bn3(self.conv3(x))))
        x = x.view(-1, 128 * 8 * 8)
        x = F.leaky_relu(self.fc1(x))
        x = self.dropout(x)
        x = F.leaky_relu(self.fc2(x))
        x = self.dropout(x)
        x = self.fc3(x)
        return x

    def predict(self, image):
        self.eval()
        with torch.no_grad():
            if isinstance(image, str) and os.path.isfile(image):
                # 如果輸入是圖片文件路徑
                img = cv2.imread(image, cv2.IMREAD_GRAYSCALE)
                img = cv2.resize(img, (IMAGE_SIZE, IMAGE_SIZE))
            elif isinstance(image, np.ndarray):
                # 如果輸入是 numpy 數組
                if image.ndim == 3:
                    img = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
                else:
                    img = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
                img = cv2.resize(img, (IMAGE_SIZE, IMAGE_SIZE))
            else:
                raise ValueError("Input should be an image file path or a numpy array")
            
            img_tensor = torch.FloatTensor(img).unsqueeze(0).unsqueeze(0) / 255.0
            img_tensor = img_tensor.to(device)
            output = self(img_tensor).item()
            
            # 將輸出四捨五入到最接近的整數
            num_instructions = round(output)
            
            # 生成相應數量的繪圖指令
            instructions = []
            for _ in range(num_instructions):
                shape = random.choice(['line', 'rectangle', 'circle', 'ellipse', 'polygon'])
                if shape == 'line':
                    instructions.append(f"cv2.line(image, {(random.randint(0, IMAGE_SIZE), random.randint(0, IMAGE_SIZE))}, {(random.randint(0, IMAGE_SIZE), random.randint(0, IMAGE_SIZE))}, {random.randint(0, 255)}, {random.randint(1, 3)})")
                elif shape == 'rectangle':
                    instructions.append(f"cv2.rectangle(image, {(random.randint(0, IMAGE_SIZE-10), random.randint(0, IMAGE_SIZE-10))}, {(random.randint(10, IMAGE_SIZE), random.randint(10, IMAGE_SIZE))}, {random.randint(0, 255)}, {random.randint(1, 3)})")
                elif shape == 'circle':
                    instructions.append(f"cv2.circle(image, {(random.randint(10, IMAGE_SIZE-10), random.randint(10, IMAGE_SIZE-10))}, {random.randint(5, 30)}, {random.randint(0, 255)}, {random.randint(1, 3)})")
                elif shape == 'ellipse':
                    instructions.append(f"cv2.ellipse(image, {(random.randint(10, IMAGE_SIZE-10), random.randint(10, IMAGE_SIZE-10))}, {(random.randint(5, 30), random.randint(5, 30))}, {random.randint(0, 360)}, 0, 360, {random.randint(0, 255)}, {random.randint(1, 3)})")
                elif shape == 'polygon':
                    num_points = random.randint(3, 6)
                    points = [(random.randint(0, IMAGE_SIZE), random.randint(0, IMAGE_SIZE)) for _ in range(num_points)]
                    instructions.append(f"cv2.polylines(image, [np.array({points})], True, {random.randint(0, 255)}, {random.randint(1, 3)})")
              
            
            return instructions

def train(model, train_loader, optimizer, criterion):
    model.train()
    total_loss = 0
    for batch_idx, (data, target) in enumerate(train_loader):
        data, target = data.to(device), target.float().to(device)
        optimizer.zero_grad()
        output = model(data).squeeze()
        loss = criterion(output, target)
        loss.backward()
        optimizer.step()
        total_loss += loss.item()
        if batch_idx % 100 == 0:
            print(f'Train Batch {batch_idx}/{len(train_loader)} Loss: {loss.item():.6f}')
    return total_loss / len(train_loader)

def test(model, test_loader, criterion, print_predictions=False):
    model.eval()
    test_loss = 0
    all_predictions = []
    all_targets = []
    with torch.no_grad():
        for data, target in test_loader:
            data, target = data.to(device), target.float().to(device)
            output = model(data).squeeze()
            test_loss += criterion(output, target).item()
            all_predictions.extend(output.cpu().numpy())
            all_targets.extend(target.cpu().numpy())
    
    test_loss /= len(test_loader)
    print(f'Test set: Average loss: {test_loss:.4f}')
    
    if print_predictions:
        print("Sample predictions:")
        for pred, targ in zip(all_predictions[:10], all_targets[:10]):
            print(f"Prediction: {pred:.2f}, Target: {targ:.2f}")
    
    return test_loss, all_predictions, all_targets

def train1(NUM_SAMPLES=NUM_SAMPLES, maxNumShape=1, EPOCHS=EPOCHS):
    model = SimpleModel(path=os.path.join(Gbase, 'best_model.pth')).to(device)
    optimizer = optim.Adam(model.parameters(), lr=LEARNING_RATE)
    
    optimizer_path = os.path.join(Gbase, 'optimizer.pth')
    if os.path.exists(optimizer_path):
        print("Loading optimizer state...")
        optimizer.load_state_dict(torch.load(optimizer_path, map_location=device))
    
    criterion = nn.MSELoss()
    
    seed = 618 * 382 * 33
    random.seed(seed)
    np.random.seed(seed)
    torch.manual_seed(seed)
    if torch.cuda.is_available():
        torch.cuda.manual_seed(seed)

    dataset = generate_dataset(NUM_SAMPLES=NUM_SAMPLES, maxNumShape=maxNumShape)
    train_size = int(0.8 * len(dataset))
    train_dataset = ImageDataset(dataset[:train_size])
    test_dataset = ImageDataset(dataset[train_size:])

    train_loader = DataLoader(train_dataset, batch_size=BATCH_SIZE, shuffle=True)
    test_loader = DataLoader(test_dataset, batch_size=BATCH_SIZE, shuffle=True)

    best_loss = float('inf')

    for epoch in range(EPOCHS):
        print(f'Epoch {epoch+1}/{EPOCHS}')
        train_loss = train(model, train_loader, optimizer, criterion)
        test_loss, predictions, targets = test(model, test_loader, criterion, print_predictions=True)
        
        if test_loss < best_loss:
            best_loss = test_loss
            torch.save(model.state_dict(), os.path.join(Gbase, 'best_model.pth'))
            torch.save(optimizer.state_dict(), os.path.join(Gbase, 'optimizer.pth'))
            print(f"New best model saved with test loss: {best_loss:.4f}")




def main():
    # Set random seed
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    model = SimpleModel(path=Gbase+ 'best_model.pth').to(device)
    optimizer = optim.Adam(model.parameters(), lr=LEARNING_RATE)
    if os.path.exists(Gbase+'optimizer.pth'):
        print("Loading optimizer state...")
        optimizer.load_state_dict(torch.load('optimizer.pth'))
    criterion = nn.MSELoss()
    test_image =Gbase+"image.jpg"
    # np.random.randint(0, 256, (IMAGE_SIZE, IMAGE_SIZE), dtype=np.uint8)
    instructions = model.predict(test_image)
    print("Generated instructions:")
    for instruction in instructions:
        print(instruction)
    # 檢查是否存在已保存的優化器狀態
   
    #return 
    seed = 618 * 382 * 33
    random.seed(seed)
    np.random.seed(seed)
    torch.manual_seed(seed)

    # Generate dataset
    dataset = generate_dataset()

    # Split dataset into train and test
    train_size = int(0.8 * len(dataset))
    train_dataset = ImageDataset(dataset[:train_size])
    test_dataset = ImageDataset(dataset[train_size:])

    train_loader = DataLoader(train_dataset, batch_size=BATCH_SIZE, shuffle=True)
    test_loader = DataLoader(test_dataset, batch_size=BATCH_SIZE, shuffle=True)

    

    best_loss = float('inf')

    for epoch in range(EPOCHS):
        print(f'Epoch {epoch+1}/{EPOCHS}')
        train_loss = train(model, train_loader, optimizer, criterion, device)
        test_loss, predictions, targets = test(model, test_loader, criterion, device, print_predictions=True)
        
        if test_loss < best_loss:
            best_loss = test_loss
            torch.save(model.state_dict(),Gbase+ 'best_model.pth')
            torch.save(optimizer.state_dict(),Gbase+ 'optimizer.pth')
            print(f"New best model saved with test loss: {best_loss:.4f}")

    # 測試 predict 方法
    

if __name__ == "__main__":
    train1(NUM_SAMPLES=1000 ,maxNumShape=1, EPOCHS=100)
    train1(NUM_SAMPLES=1000 ,maxNumShape=1, EPOCHS=100)
    train1(NUM_SAMPLES=1000 ,maxNumShape=1, EPOCHS=100)
    train1(NUM_SAMPLES=10000 ,maxNumShape=2, EPOCHS=10)
    train1(NUM_SAMPLES=10000 ,maxNumShape=3,  EPOCHS=10)
    train1(NUM_SAMPLES=100000 ,maxNumShape=5, EPOCHS=10)
    train1(NUM_SAMPLES=100000 ,maxNumShape=5, EPOCHS=10)
    train1(NUM_SAMPLES=100000 ,maxNumShape=5, EPOCHS=10)
    while True:
        train1(NUM_SAMPLES=100000 ,maxNumShape=8, EPOCHS=10)