Python Pytorch学习之图像检索实践

class TripletData(Dataset):
    def __init__(self, path, transforms, split="train"):
 
        self.path = path
        self.split = split    # train or valid
        self.cats = 102       # number of categories
        self.transforms = transforms
 
        
    def __getitem__(self, idx):
 
        # our positive class for the triplet
        idx = str(idx%self.cats + 1)
 
        # choosing our pair of positive images (im1, im2)
        positives = os.listdir(os.path.join(self.path, idx))
        im1, im2 = random.sample(positives, 2)
 
        # choosing a negative class and negative image (im3)
        negative_cats = [str(x+1) for x in range(self.cats)]
        negative_cats.remove(idx)
        negative_cat = str(random.choice(negative_cats))
        negatives = os.listdir(os.path.join(self.path, negative_cat))
 
        im3 = random.choice(negatives)
 
        im1,im2,im3 = os.path.join(self.path, idx, im1), os.path.join(self.path, idx, im2), os.path.join(self.path, negative_cat, im3)
 
        im1 = self.transforms(Image.open(im1))
 
        im2 = self.transforms(Image.open(im2))
 
        im3 = self.transforms(Image.open(im3))
 
        return [im1, im2, im3]
 
    
    # we'll put some value that we want since there can be far too many triplets possible
    # multiples of the number of images/ number of categories is a good choice
    def __len__(self):
        return self.cats*8
# Transforms
train_transforms = transforms.Compose([
    transforms.Resize((224,224)),
    transforms.RandomHorizontalFlip(),
    transforms.ToTensor(),
    transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
])
val_transforms = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
])
# Datasets and Dataloaders
train_data = TripletData(PATH_TRAIN, train_transforms)
val_data = TripletData(PATH_VALID, val_transforms)
train_loader = torch.utils.data.DataLoader(dataset = train_data, batch_size=32, shuffle=True, num_workers=4)
val_loader = torch.utils.data.DataLoader(dataset = val_data, batch_size=32, shuffle=False, num_workers=4)

class TripletLoss(nn.Module):
    def __init__(self, margin=1.0):
        
        super(TripletLoss, self).__init__()
        self.margin = margin
        
        
    def calc_euclidean(self, x1, x2):
        return (x1 - x2).pow(2).sum(1)
    
    
    # Distances in embedding space is calculated in euclidean
    def forward(self, anchor, positive, negative):
        
        distance_positive = self.calc_euclidean(anchor, positive)
        
        distance_negative = self.calc_euclidean(anchor, negative)
        
        losses = torch.relu(distance_positive - distance_negative + self.margin)
        
        return losses.mean()
      
 
device = 'cuda'
 
# Our base model
model = models.resnet18().cuda()
optimizer = optim.Adam(model.parameters(), lr=0.001)
triplet_loss = TripletLoss()
 
# Training
for epoch in range(epochs):
    
    model.train()
    epoch_loss = 0.0
    
    for data in tqdm(train_loader):
        
        optimizer.zero_grad()
        x1,x2,x3 = data
        e1 = model(x1.to(device))
        e2 = model(x2.to(device))
        e3 = model(x3.to(device)) 
        
        loss = triplet_loss(e1,e2,e3)
        epoch_loss += loss
        loss.backward()
        optimizer.step()
        
    print("Train Loss: {}".format(epoch_loss.item()))
 
    
    
class TripletLoss(nn.Module):
    def __init__(self, margin=1.0):
        
        super(TripletLoss, self).__init__()
        self.margin = margin
        
        
    def calc_euclidean(self, x1, x2):
        return (x1 - x2).pow(2).sum(1)
    
    
    # Distances in embedding space is calculated in euclidean
    def forward(self, anchor, positive, negative):
        
        distance_positive = self.calc_euclidean(anchor, positive)
        
        distance_negative = self.calc_euclidean(anchor, negative)
        
        losses = torch.relu(distance_positive - distance_negative + self.margin)
        
        return losses.mean()
      
 
device = 'cuda'
 
 
# Our base model
model = models.resnet18().cuda()
optimizer = optim.Adam(model.parameters(), lr=0.001)
triplet_loss = TripletLoss()
 
 
# Training
for epoch in range(epochs):
    model.train()
    epoch_loss = 0.0
    for data in tqdm(train_loader):
 
        optimizer.zero_grad()
        
        x1,x2,x3 = data
        
        e1 = model(x1.to(device))
        e2 = model(x2.to(device))
        e3 = model(x3.to(device)) 
        
        loss = triplet_loss(e1,e2,e3)
        epoch_loss += loss
        loss.backward()
        optimizer.step()
        
    print("Train Loss: {}".format(epoch_loss.item()))

#!pip install faiss-gpu
import faiss                            
faiss_index = faiss.IndexFlatL2(1000)   # build the index
 
# storing the image representations
im_indices = []
 
with torch.no_grad():
    for f in glob.glob(os.path.join(PATH_TRAIN, '*/*')):
        
        im = Image.open(f)
        im = im.resize((224,224))
        im = torch.tensor([val_transforms(im).numpy()]).cuda()
    
        preds = model(im)
        preds = np.array([preds[0].cpu().numpy()])
        faiss_index.add(preds) #add the representation to index
        im_indices.append(f)   #store the image name to find it later on
 
        
# Retrieval with a query image
with torch.no_grad():
    for f in os.listdir(PATH_TEST):
        
        # query/test image
        im = Image.open(os.path.join(PATH_TEST,f))
        im = im.resize((224,224))
        im = torch.tensor([val_transforms(im).numpy()]).cuda()
    
        test_embed = model(im).cpu().numpy()
        
        _, I = faiss_index.search(test_embed, 5)
        print("Retrieved Image: {}".format(im_indices[I[0][0]]))