first training

data/MNIST/raw/t10k-images-idx3-ubyte

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0fa7898d509279e482958e8ce81c8e77db3f2f8254e26661ceb7762c4d494ce7
+size 7840016

data/MNIST/raw/t10k-images-idx3-ubyte.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8d422c7b0a1c1c79245a5bcf07fe86e33eeafee792b84584aec276f5a2dbc4e6
+size 1648877

data/MNIST/raw/t10k-labels-idx1-ubyte

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ff7bcfd416de33731a308c3f266cc351222c34898ecbeaf847f06e48f7ec33f2
+size 10008

data/MNIST/raw/t10k-labels-idx1-ubyte.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f7ae60f92e00ec6debd23a6088c31dbd2371eca3ffa0defaefb259924204aec6
+size 4542

data/MNIST/raw/train-images-idx3-ubyte

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ba891046e6505d7aadcbbe25680a0738ad16aec93bde7f9b65e87a2fc25776db
+size 47040016

data/MNIST/raw/train-images-idx3-ubyte.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:440fcabf73cc546fa21475e81ea370265605f56be210a4024d2ca8f203523609
+size 9912422

data/MNIST/raw/train-labels-idx1-ubyte

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:65a50cbbf4e906d70832878ad85ccda5333a97f0f4c3dd2ef09a8a9eef7101c5
+size 60008

data/MNIST/raw/train-labels-idx1-ubyte.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3552534a0a558bbed6aed32b30c495cca23d567ec52cac8be1a0730e8010255c
+size 28881

model.py

@@ -3,29 +3,34 @@ import torch.nn as nn
 import torch.nn.functional as F
 
 class MovieposterNet(nn.Module):
-    def __init__(self):
+    def __init__(self, num_classes=10):
         super(MovieposterNet, self).__init__()
-        self.conv1 = nn.Conv2d(1,8,5)
-        self.conv2 = nn.Conv2d(8,16, 5)
-        self.pool = nn.MaxPool2d(2,2)
-        self.fc1 = nn.Linear(256,128)
-        self.fc2 = nn.Linear(128,64)
-        self.fc3 = nn.Linear(64,10)
+        # Entrée : 3 canaux (RGB), 8 filtres, noyau 5x5
+        self.conv1 = nn.Conv2d(3, 8, 5)
+        self.conv2 = nn.Conv2d(8, 16, 5)
+        self.pool = nn.MaxPool2d(2, 2)
+        
+        # Calcul des dimensions après convolutions et pooling :
+        # Input: (3, 224, 224)
+        # Conv1: 224 - 5 + 1 = 220 -> Pool: 110
+        # Conv2: 110 - 5 + 1 = 106 -> Pool: 53
+        # Taille aplatie : 16 * 53 * 53 = 44944
+        self.fc1 = nn.Linear(16 * 53 * 53, 128)
+        self.fc2 = nn.Linear(128, 64)
+        self.fc3 = nn.Linear(64, num_classes)
 
     def forward(self, x):
-        x = F.relu(self.conv1(x))       # First convolution followed by
-        x = self.pool(x)                # a relu activation and a max pooling#
-        x = F.relu(self.conv2(x))
-        x = self.pool(x)      
-        x=torch.flatten(x,1)
+        x = self.pool(F.relu(self.conv1(x)))
+        x = self.pool(F.relu(self.conv2(x)))
+        x = torch.flatten(x, 1) 
         x = F.relu(self.fc1(x))
         x = F.relu(self.fc2(x))
         x = self.fc3(x)
         return x
     
-    
     def get_features(self, x):
+        # Pour TensorBoard : extraction des caractéristiques avant les couches denses
         x = self.pool(F.relu(self.conv1(x)))
         x = self.pool(F.relu(self.conv2(x)))
-        x = x.view(-1, 16 * 4 * 4)
+        x = torch.flatten(x, 1)
         return x

runs/Movieposter/00003/default/metadata.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dde9687965705836c2844ef4737cdae31c7c5c22a19aecb5e204e64f68269044
+size 456

runs/Movieposter/00003/default/tensors.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0b65839a3628693becb0a65a290fb7ae1b291ba33c2e5995a36da7e7a37e3e15
+size 41848595

runs/Movieposter/events.out.tfevents.1772969768.DESKTOP-5TAQJI5.15764.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f9cedd73a112b3e6198fe35ca760beafb2f22fe1209aa0f5f767a6c1f0f0c940
+size 239

runs/Movieposter/events.out.tfevents.1772971673.DESKTOP-5TAQJI5.21220.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7ed38084b8611ca16178122c6ac4307cdd817d6e51ba9d48f1d5a420f44fcb12
+size 8611

runs/Movieposter/events.out.tfevents.1772972078.DESKTOP-5TAQJI5.8588.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0cb8d524bb7d65ce97c42cb1b4bc329da13496f779b0b83e128b008afec80c22
+size 3008529

runs/Movieposter/projector_config.pbtxt

@@ -0,0 +1,10 @@
+embeddings {
+  tensor_name: "default:00003"
+  metadata_path: "00003/default\\metadata.tsv"
+  sprite {
+    image_path: "00003/default\\sprite.png"
+    single_image_dim: 224
+    single_image_dim: 224
+  }
+  tensor_path: "00003/default\\tensors.tsv"
+}

train.py

@@ -0,0 +1,144 @@
+import argparse
+from statistics import mean
+
+import torch
+import torchvision
+import torchvision.transforms as transforms
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.optim as optim
+from tqdm import tqdm
+from torch.utils.tensorboard import SummaryWriter
+from torchvision import datasets
+from torch.utils.data import DataLoader, random_split
+
+from model import MovieposterNet
+
+ # setting device on GPU if available, else CPU
+device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+
+def train(net, optimizer, loader, writer,epochs=10):
+    criterion = nn.CrossEntropyLoss()
+    for epoch in range(epochs):
+        running_loss = []
+        t = tqdm(loader)
+        for x, y in t:
+            x, y = x.to(device), y.to(device)
+            outputs = net(x)
+            loss = criterion(outputs, y)
+            running_loss.append(loss.item())
+            optimizer.zero_grad()
+            loss.backward()
+            optimizer.step()
+            t.set_description(f'training loss: {mean(running_loss)}')
+        writer.add_scalar('training loss', mean(running_loss), epoch)
+
+
+def test(model, dataloader):
+    test_corrects = 0
+    total = 0
+    with torch.no_grad():
+        for x, y in dataloader:
+            x = x.to(device)
+            y = y.to(device)
+            y_hat = model(x).argmax(1)
+            test_corrects += y_hat.eq(y).sum().item()
+            total += y.size(0)
+    return test_corrects / total
+	
+if __name__=='__main__':
+
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument('--exp_name', type=str, default = 'Movieposter', help='experiment name')
+    parser.add_argument('--epochs', type=int, default = int(10), help='nb of epochs')
+    parser.add_argument('--batch_size', type=int, default = int(64), help='batch size')
+    parser.add_argument('--lr', type=float, default =  float(1e-3), help='learning rate')
+
+
+    args = parser.parse_args()
+    print(args.exp_name)
+    exp_name = args.exp_name
+    epochs = args.epochs
+    batch_size = args.batch_size
+    lr = args.lr
+
+    writer = SummaryWriter(f'runs/Movieposter')
+
+    # 1. Définition des transformations
+    # Les posters sont en couleur (3 canaux) et de tailles variées, contrairement à MNIST.
+    transform = transforms.Compose([
+        transforms.Resize((224, 224)), # Redimensionnement standard pour les modèles de vision
+        transforms.ToTensor(),
+        transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)) # Normalisation sur 3 canaux (RGB)
+    ])
+
+    # 2. Chargement du dataset complet
+    # Le chemin '../' permet de remonter d'un niveau par rapport au dossier 'projet_AIF'
+    data_dir = '../sorted_movie_posters_paligema'
+    full_dataset = datasets.ImageFolder(root=data_dir, transform=transform)
+
+    # 3. Division en train/test (ex: 80% train, 20% test)
+    train_size = int(0.8 * len(full_dataset))
+    test_size = len(full_dataset) - train_size
+    trainset, testset = random_split(full_dataset, [train_size, test_size])
+
+    # 4. Création des DataLoaders
+    trainloader = DataLoader(trainset, batch_size=batch_size, shuffle=True, num_workers=2)
+    testloader = DataLoader(testset, batch_size=batch_size, shuffle=False, num_workers=2)
+
+    # Accès aux classes (genres)
+    classes = full_dataset.classes
+    print(f"Classes détectées : {classes}")
+
+
+    net =MovieposterNet().to(device)
+
+    # setting net on device(GPU if available, else CPU)
+    net = net.to(device)
+    optimizer = optim.SGD(net.parameters(), lr=lr)
+
+    train(net, optimizer,trainloader, writer, epochs)
+    test_acc = test(net,testloader)
+    print(f'Test accuracy: {test_acc}')   
+
+    # 1. Gestion du dossier de sauvegarde des poids
+    import os
+    if not os.path.exists('weights'):
+        os.makedirs('weights')
+    
+    torch.save(net.state_dict(), 'weights/movieposter_net.pth')
+
+    # 2. Récupération d'un échantillon de données pour TensorBoard
+    # On utilise le loader pour obtenir des tenseurs déjà transformés
+    dataiter = iter(trainloader)
+    images, labels = next(dataiter) 
+    
+    # On limite à 64 images pour la visualisation et on envoie sur le device
+    images = images[:64].to(device)
+    labels = labels[:64].to(device)
+
+    # 3. Enregistrement du graphe du modèle
+    # Vérifiez que les dimensions d'entrée du modèle correspondent (ex: 3, 224, 224)
+    writer.add_graph(net, images)
+
+    # 4. Enregistrement d'une grille d'images
+    img_grid = torchvision.utils.make_grid(images)
+    writer.add_image('movieposter_samples', img_grid)
+
+    # 5. Projecteur d'embeddings
+    # get_features() doit être définie dans MovieposterNet pour retourner l'avant-dernière couche
+    with torch.no_grad():
+        try:
+            embeddings = net.get_features(images)
+            # Conversion des indices en noms de classes pour la lisibilité
+            metadata = [classes[l] for l in labels]
+            writer.add_embedding(embeddings,
+                                metadata=metadata,
+                                label_img=images, 
+                                global_step=epochs)
+        except AttributeError:
+            print("Erreur : La méthode get_features n'est pas définie dans MovieposterNet.")
+
+    # 6. Fermeture du SummaryWriter
+    writer.close()

weights/movieposter_net.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:24daa4b8600c1c61c2a05b129871ee80c38b00d582b81991611c2381f45d68c4
+size 23066885