Upload 8 files

contextualizer.py

@@ -0,0 +1,93 @@
+import torch
+from torch import nn
+
+
+
+
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, hidden_dim, dropout):
+        super().__init__()
+        self.net = nn.Sequential(
+            nn.Linear(dim, hidden_dim),
+            nn.GELU(),
+            nn.Dropout(dropout),
+            nn.Linear(hidden_dim, dim),
+            nn.Dropout(dropout)
+        )
+    def forward(self, x):
+        return self.net(x)
+
+
+   
+
+
+
+
+
+
+class ContextualizerBlock(nn.Module):
+    def __init__(self, d_model,d_ffn,dropout,num_tokens):
+        super().__init__()
+        
+        self.context_proj =  nn.Linear(d_model,d_model)
+        self.mlp = FeedForward(d_model,d_ffn,dropout)     
+        self.norm = nn.LayerNorm(d_model)       
+        self.upsample = nn.Upsample(scale_factor=num_tokens,mode='nearest')
+        self.downsample = nn.Upsample(scale_factor= 1/num_tokens, mode='nearest')
+    def forward(self, x):
+        res = x
+        x = self.norm(x)
+        
+        context = x
+        dim0 = context.shape[0]
+        dim1 = context.shape[1]
+        dim2 = context.shape[2]
+        context = context.reshape([dim0,1,dim1*dim2])
+        
+        context = self.downsample(context)
+        context = context.reshape([dim0,dim2])
+        context = self.context_proj(context)
+        
+        context = context.reshape([dim0,1,dim2])
+        context = self.upsample(context)
+        context = context.reshape([dim0,dim1,dim2]) 
+        x = context
+        x = x + res  
+        res = x
+        x = self.norm(x)
+        x = self.mlp(x)
+        out = x + res
+        return out 
+        return 
+    
+class Contextualizer(nn.Module):
+    def __init__(self, d_model, d_ffn, num_layers,dropout,num_tokens):
+        super().__init__()
+        
+        self.model = nn.Sequential(
+            
+            *[ContextualizerBlock(d_model,d_ffn,dropout,num_tokens) for _ in range(num_layers)],
+            
+            
+        )
+
+    def forward(self, x):
+        
+        x = self.model(x)
+        
+        return x
+       
+       
+       
+       
+       
+       
+       
+       
+
+
+
+
+
+

contextualizer_mlp.py

@@ -0,0 +1,93 @@
+import torch
+from torch import nn
+
+
+
+
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, hidden_dim, dropout):
+        super().__init__()
+        self.net = nn.Sequential(
+            nn.Linear(dim, hidden_dim),
+            nn.GELU(),
+            nn.Dropout(dropout),
+            nn.Linear(hidden_dim, dim),
+            nn.Dropout(dropout)
+        )
+    def forward(self, x):
+        return self.net(x)
+
+
+   
+
+
+
+
+
+
+class ContextualizerBlock(nn.Module):
+    def __init__(self, d_model,d_ffn,dropout,num_tokens):
+        super().__init__()
+        
+        self.context_mlp = FeedForward(d_model,d_ffn,dropout)
+        self.mlp = FeedForward(d_model,d_ffn,dropout)     
+        self.norm = nn.LayerNorm(d_model)       
+        self.upsample = nn.Upsample(scale_factor=num_tokens,mode='nearest')
+        self.downsample = nn.Upsample(scale_factor= 1/num_tokens, mode='nearest')
+    def forward(self, x):
+        res = x
+        x = self.norm(x)
+        
+        context = x
+        dim0 = context.shape[0]
+        dim1 = context.shape[1]
+        dim2 = context.shape[2]
+        context = context.reshape([dim0,1,dim1*dim2])
+        
+        context = self.downsample(context)
+        context = context.reshape([dim0,dim2])
+        context = self.context_mlp(context)
+        
+        context = context.reshape([dim0,1,dim2])
+        context = self.upsample(context)
+        context = context.reshape([dim0,dim1,dim2]) 
+        x = context
+        x = x + res  
+        res = x
+        x = self.norm(x)
+        x = self.mlp(x)
+        out = x + res
+        return out 
+        return 
+    
+class Contextualizer(nn.Module):
+    def __init__(self, d_model, d_ffn, num_layers,dropout,num_tokens):
+        super().__init__()
+        
+        self.model = nn.Sequential(
+            
+            *[ContextualizerBlock(d_model,d_ffn,dropout,num_tokens) for _ in range(num_layers)],
+            
+            
+        )
+
+    def forward(self, x):
+        
+        x = self.model(x)
+        
+        return x
+       
+       
+       
+       
+       
+       
+       
+       
+
+
+
+
+
+

contextualizer_mlp_nin.py

@@ -0,0 +1,126 @@
+import torch
+from torch import nn
+
+
+
+
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, hidden_dim, dropout):
+        super().__init__()
+        self.net = nn.Sequential(
+            nn.Linear(dim, hidden_dim),
+            nn.GELU(),
+            nn.Dropout(dropout),
+            nn.Linear(hidden_dim, dim),
+            nn.Dropout(dropout)
+        )
+    def forward(self, x):
+        return self.net(x)
+
+
+   
+
+
+
+
+
+
+class ContextualizerBlock(nn.Module):
+    def __init__(self, d_model,d_ffn,dropout,num_tokens):
+        super().__init__()
+        
+        self.context_mlp = FeedForward(d_model,d_ffn,dropout)
+        self.mlp = FeedForward(d_model,d_ffn,dropout)     
+        self.norm = nn.LayerNorm(d_model)       
+        self.upsample = nn.Upsample(scale_factor=num_tokens,mode='nearest')
+        self.downsample = nn.Upsample(scale_factor= 1/num_tokens, mode='nearest')
+    def forward(self, x):
+        res = x
+        x = self.norm(x)
+        
+        context = x
+        dim0 = context.shape[0]
+        dim1 = context.shape[1]
+        dim2 = context.shape[2]
+        context = context.reshape([dim0,1,dim1*dim2])
+        
+        context = self.downsample(context)
+        context = context.reshape([dim0,dim2])
+        context = self.context_mlp(context)
+        
+        context = context.reshape([dim0,1,dim2])
+        context = self.upsample(context)
+        context = context.reshape([dim0,dim1,dim2]) 
+        x = context
+        x = x + res  
+        res = x
+        x = self.norm(x)
+        x = self.mlp(x)
+        out = x + res
+        return out 
+        return
+ 
+ 
+class MixerGatingUnit(nn.Module):
+    def __init__(self,d_model,d_ffn,dropout,num_tokens):
+        super().__init__()     
+        self.Mixer = ContextualizerBlock(d_model,d_ffn,dropout,num_tokens)
+        self.proj = nn.Linear(d_model,d_model)
+
+    def forward(self, x):
+        u, v = x, x 
+        u = self.proj(u)  
+        v = self.Mixer(v)
+        out = u * v
+        return out
+
+
+class ContextualizerNiNBlock(nn.Module):
+    def __init__(self, d_model,d_ffn,dropout,num_tokens):
+        super().__init__()
+       
+        self.norm = nn.LayerNorm(d_model)       
+        self.mgu = MixerGatingUnit(d_model,d_ffn,dropout,num_tokens)
+        self.ffn = FeedForward(d_model,d_ffn,dropout)
+    def forward(self, x):
+        residual = x
+        x = self.norm(x)
+        x = self.mgu(x)   
+        x = x + residual      
+        residual = x
+        x = self.norm(x)
+        x = self.ffn(x)
+        out = x + residual
+        return out
+    
+class ContextualizerNiN(nn.Module):
+    def __init__(self, d_model, d_ffn, num_layers,dropout,num_tokens):
+        super().__init__()
+        
+        self.model = nn.Sequential(
+            
+            *[ContextualizerNiNBlock(d_model,d_ffn,dropout,num_tokens) for _ in range(num_layers)],
+            
+            
+        )
+
+    def forward(self, x):
+        
+        x = self.model(x)
+        
+        return x
+       
+       
+       
+       
+       
+       
+       
+       
+
+
+
+
+
+

contextualizer_nin.py

@@ -0,0 +1,126 @@
+import torch
+from torch import nn
+
+
+
+
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, hidden_dim, dropout):
+        super().__init__()
+        self.net = nn.Sequential(
+            nn.Linear(dim, hidden_dim),
+            nn.GELU(),
+            nn.Dropout(dropout),
+            nn.Linear(hidden_dim, dim),
+            nn.Dropout(dropout)
+        )
+    def forward(self, x):
+        return self.net(x)
+
+
+   
+
+
+
+
+
+
+class ContextualizerBlock(nn.Module):
+    def __init__(self, d_model,d_ffn,dropout,num_tokens):
+        super().__init__()
+        
+        self.context_proj = nn.Linear(d_model,d_model)
+        self.mlp = FeedForward(d_model,d_ffn,dropout)     
+        self.norm = nn.LayerNorm(d_model)       
+        self.upsample = nn.Upsample(scale_factor=num_tokens,mode='nearest')
+        self.downsample = nn.Upsample(scale_factor= 1/num_tokens, mode='nearest')
+    def forward(self, x):
+        res = x
+        x = self.norm(x)
+        
+        context = x
+        dim0 = context.shape[0]
+        dim1 = context.shape[1]
+        dim2 = context.shape[2]
+        context = context.reshape([dim0,1,dim1*dim2])
+        
+        context = self.downsample(context)
+        context = context.reshape([dim0,dim2])
+        context = self.context_proj(context)
+        
+        context = context.reshape([dim0,1,dim2])
+        context = self.upsample(context)
+        context = context.reshape([dim0,dim1,dim2]) 
+        x = context
+        x = x + res  
+        res = x
+        x = self.norm(x)
+        x = self.mlp(x)
+        out = x + res
+        return out 
+        return
+ 
+ 
+class MixerGatingUnit(nn.Module):
+    def __init__(self,d_model,d_ffn,dropout,num_tokens):
+        super().__init__()     
+        self.Mixer = ContextualizerBlock(d_model,d_ffn,dropout,num_tokens)
+        self.proj = nn.Linear(d_model,d_model)
+
+    def forward(self, x):
+        u, v = x, x 
+        u = self.proj(u)  
+        v = self.Mixer(v)
+        out = u * v
+        return out
+
+
+class ContextualizerNiNBlock(nn.Module):
+    def __init__(self, d_model,d_ffn,dropout,num_tokens):
+        super().__init__()
+       
+        self.norm = nn.LayerNorm(d_model)       
+        self.mgu = MixerGatingUnit(d_model,d_ffn,dropout,num_tokens)
+        self.ffn = FeedForward(d_model,d_ffn,dropout)
+    def forward(self, x):
+        residual = x
+        x = self.norm(x)
+        x = self.mgu(x)   
+        x = x + residual      
+        residual = x
+        x = self.norm(x)
+        x = self.ffn(x)
+        out = x + residual
+        return out
+    
+class ContextualizerNiN(nn.Module):
+    def __init__(self, d_model, d_ffn, num_layers,dropout,num_tokens):
+        super().__init__()
+        
+        self.model = nn.Sequential(
+            
+            *[ContextualizerNiNBlock(d_model,d_ffn,dropout,num_tokens) for _ in range(num_layers)],
+            
+            
+        )
+
+    def forward(self, x):
+        
+        x = self.model(x)
+        
+        return x
+       
+       
+       
+       
+       
+       
+       
+       
+
+
+
+
+
+

train.py

@@ -0,0 +1,194 @@
+#imports
+
+import os
+import csv
+import torch
+from torch import nn 
+from torch.utils.data import DataLoader 
+from torchvision import datasets 
+from torchvision.transforms import ToTensor, Normalize, RandomCrop, RandomHorizontalFlip, Compose 
+from contextualizer import Contextualizer
+
+# data transforms
+
+transform = Compose([
+RandomCrop(32, padding=4),
+RandomHorizontalFlip(), 
+ToTensor(),
+Normalize((0.5, 0.5,0.5),(0.5, 0.5,0.5))
+
+])
+
+training_data = datasets.CIFAR10(
+                                       root='data',
+                                       train=True,
+                                       download=True,
+                                       transform=transform 
+                                       )
+
+test_data = datasets.CIFAR10(
+                                       root='data',
+                                       train=False,
+                                       download=True,
+                                       transform=transform 
+                                       )                                       
+# create dataloaders  
+                                     
+batch_size = 128
+
+train_dataloader = DataLoader(training_data, batch_size=batch_size,shuffle=True)
+test_dataloader = DataLoader(test_data, batch_size=batch_size)
+
+
+for X, y in test_dataloader:
+    print(f"Shape of X [N,C,H,W]:{X.shape}")
+    print(f"Shape of y:{y.shape}{y.dtype}")
+    break
+
+# size checking for loading images
+def check_sizes(image_size, patch_size):
+    sqrt_num_patches, remainder = divmod(image_size, patch_size)
+    assert remainder == 0, "`image_size` must be divisibe by `patch_size`"
+    num_patches = sqrt_num_patches ** 2
+    return num_patches
+
+
+
+# create model
+# Get cpu or gpu device for training.
+device = "cuda" if torch.cuda.is_available() else "cpu"
+
+print(f"using {device} device") 
+
+# model definition
+
+class ContextualizerImageClassification(Contextualizer):
+    def __init__(
+        self,
+        image_size=32,
+        patch_size=4,
+        in_channels=3,
+        num_classes=10,
+        d_ffn=512,
+        d_model = 256,
+        num_tokens = 64,
+        num_layers=4,
+        dropout=0.5
+    ):
+        num_patches = check_sizes(image_size, patch_size)
+        super().__init__(d_model,d_ffn,num_layers,dropout, num_tokens)
+        self.patcher = nn.Conv2d(
+            in_channels, d_model, kernel_size=patch_size, stride=patch_size
+        )
+        self.classifier = nn.Linear(d_model, num_classes)
+
+    def forward(self, x):
+        
+        patches = self.patcher(x)
+        batch_size, num_channels, _, _ = patches.shape
+        patches = patches.permute(0, 2, 3, 1)
+        patches = patches.view(batch_size, -1, num_channels)
+        embedding = self.model(patches)
+        embedding = embedding.mean(dim=1) # global average pooling
+        out = self.classifier(embedding)
+        return out
+
+model = ContextualizerImageClassification().to(device)
+print(model)
+
+# Optimizer
+
+loss_fn = nn.CrossEntropyLoss()
+optimizer = torch.optim.Adam(model.parameters(),lr=1e-3)
+
+
+# Training Loop
+
+def train(dataloader, model, loss_fn, optimizer):
+    size = len(dataloader.dataset)
+    num_batches = len(dataloader)
+    model.train()
+    train_loss = 0
+    correct = 0
+    for batch, (X,y) in enumerate(dataloader):
+        X, y = X.to(device), y.to(device)
+       
+        #compute prediction error
+        pred = model(X)
+        loss = loss_fn(pred,y)
+        
+        # backpropagation
+        optimizer.zero_grad()
+        loss.backward()
+        optimizer.step()
+        train_loss += loss.item()
+        _, labels = torch.max(pred.data, 1)
+        correct += labels.eq(y.data).type(torch.float).sum()
+
+        
+
+
+        if batch % 100 == 0:
+            loss, current = loss.item(), batch * len(X)
+            print(f"loss: {loss:>7f}   [{current:>5d}/{size:>5d}]")
+
+    train_loss /= num_batches
+    train_accuracy = 100. * correct.item() / size
+    print(train_accuracy)
+    return train_loss,train_accuracy 
+
+
+
+# Test loop
+
+def test(dataloader, model, loss_fn):
+    size = len(dataloader.dataset)            
+    num_batches = len(dataloader)
+    model.eval()
+    test_loss = 0
+    correct = 0
+    with torch.no_grad():
+        for X,y in dataloader:
+            X,y = X.to(device), y.to(device)
+            pred = model(X)
+            test_loss += loss_fn(pred, y).item()
+            correct += (pred.argmax(1) == y).type(torch.float).sum().item()
+    test_loss /= num_batches
+    correct /= size
+    print(f"Test Error: \n Accuracy: {(100*correct):>0.1f}%, Avg loss: {test_loss:>8f} \n")  
+    test_accuracy = 100*correct      
+    return test_loss, test_accuracy
+
+
+
+# apply train and test
+
+logname = "/PATH/Contextualizer/Experiments_cifar10/logs_contextualizer/logs_cifar10.csv"
+if not os.path.exists(logname):
+  with open(logname, 'w') as logfile:
+    logwriter = csv.writer(logfile, delimiter=',')
+    logwriter.writerow(['epoch', 'train loss', 'train acc',
+                        'test loss', 'test acc'])
+
+
+epochs = 100
+for epoch in range(epochs):
+    print(f"Epoch {epoch+1}\n-----------------------------------")
+    train_loss, train_acc = train(train_dataloader, model, loss_fn, optimizer)
+    # learning rate scheduler
+    #if scheduler is not None:
+    #    scheduler.step()
+    test_loss, test_acc = test(test_dataloader, model, loss_fn)
+    with open(logname, 'a') as logfile:
+        logwriter = csv.writer(logfile, delimiter=',')
+        logwriter.writerow([epoch+1, train_loss, train_acc,
+                            test_loss, test_acc])
+print("Done!")
+
+# saving trained model
+
+path = "/PATH/Contextualizer/Experiments_cifar10/weights_contextualizer"
+model_name = "ContextualizerImageClassification_cifar10"
+torch.save(model.state_dict(), f"{path}/{model_name}.pth")
+print(f"Saved Model State to {path}/{model_name}.pth ")
+

train_mlp.py

@@ -0,0 +1,194 @@
+#imports
+
+import os
+import csv
+import torch
+from torch import nn 
+from torch.utils.data import DataLoader 
+from torchvision import datasets 
+from torchvision.transforms import ToTensor, Normalize, RandomCrop, RandomHorizontalFlip, Compose 
+from contextualizer_mlp import Contextualizer
+
+# data transforms
+
+transform = Compose([
+RandomCrop(32, padding=4),
+RandomHorizontalFlip(), 
+ToTensor(),
+Normalize((0.5, 0.5,0.5),(0.5, 0.5,0.5))
+
+])
+
+training_data = datasets.CIFAR10(
+                                       root='data',
+                                       train=True,
+                                       download=True,
+                                       transform=transform 
+                                       )
+
+test_data = datasets.CIFAR10(
+                                       root='data',
+                                       train=False,
+                                       download=True,
+                                       transform=transform 
+                                       )                                       
+# create dataloaders  
+                                     
+batch_size = 128
+
+train_dataloader = DataLoader(training_data, batch_size=batch_size,shuffle=True)
+test_dataloader = DataLoader(test_data, batch_size=batch_size)
+
+
+for X, y in test_dataloader:
+    print(f"Shape of X [N,C,H,W]:{X.shape}")
+    print(f"Shape of y:{y.shape}{y.dtype}")
+    break
+
+# size checking for loading images
+def check_sizes(image_size, patch_size):
+    sqrt_num_patches, remainder = divmod(image_size, patch_size)
+    assert remainder == 0, "`image_size` must be divisibe by `patch_size`"
+    num_patches = sqrt_num_patches ** 2
+    return num_patches
+
+
+
+# create model
+# Get cpu or gpu device for training.
+device = "cuda" if torch.cuda.is_available() else "cpu"
+
+print(f"using {device} device") 
+
+# model definition
+
+class ContextualizerImageClassification(Contextualizer):
+    def __init__(
+        self,
+        image_size=32,
+        patch_size=4,
+        in_channels=3,
+        num_classes=10,
+        d_ffn=512,
+        d_model = 256,
+        num_tokens = 64,
+        num_layers=4,
+        dropout=0.5
+    ):
+        num_patches = check_sizes(image_size, patch_size)
+        super().__init__(d_model,d_ffn,num_layers,dropout, num_tokens)
+        self.patcher = nn.Conv2d(
+            in_channels, d_model, kernel_size=patch_size, stride=patch_size
+        )
+        self.classifier = nn.Linear(d_model, num_classes)
+
+    def forward(self, x):
+        
+        patches = self.patcher(x)
+        batch_size, num_channels, _, _ = patches.shape
+        patches = patches.permute(0, 2, 3, 1)
+        patches = patches.view(batch_size, -1, num_channels)
+        embedding = self.model(patches)
+        embedding = embedding.mean(dim=1) # global average pooling
+        out = self.classifier(embedding)
+        return out
+
+model = ContextualizerImageClassification().to(device)
+print(model)
+
+# Optimizer
+
+loss_fn = nn.CrossEntropyLoss()
+optimizer = torch.optim.Adam(model.parameters(),lr=1e-3)
+
+
+# Training Loop
+
+def train(dataloader, model, loss_fn, optimizer):
+    size = len(dataloader.dataset)
+    num_batches = len(dataloader)
+    model.train()
+    train_loss = 0
+    correct = 0
+    for batch, (X,y) in enumerate(dataloader):
+        X, y = X.to(device), y.to(device)
+       
+        #compute prediction error
+        pred = model(X)
+        loss = loss_fn(pred,y)
+        
+        # backpropagation
+        optimizer.zero_grad()
+        loss.backward()
+        optimizer.step()
+        train_loss += loss.item()
+        _, labels = torch.max(pred.data, 1)
+        correct += labels.eq(y.data).type(torch.float).sum()
+
+        
+
+
+        if batch % 100 == 0:
+            loss, current = loss.item(), batch * len(X)
+            print(f"loss: {loss:>7f}   [{current:>5d}/{size:>5d}]")
+
+    train_loss /= num_batches
+    train_accuracy = 100. * correct.item() / size
+    print(train_accuracy)
+    return train_loss,train_accuracy 
+
+
+
+# Test loop
+
+def test(dataloader, model, loss_fn):
+    size = len(dataloader.dataset)            
+    num_batches = len(dataloader)
+    model.eval()
+    test_loss = 0
+    correct = 0
+    with torch.no_grad():
+        for X,y in dataloader:
+            X,y = X.to(device), y.to(device)
+            pred = model(X)
+            test_loss += loss_fn(pred, y).item()
+            correct += (pred.argmax(1) == y).type(torch.float).sum().item()
+    test_loss /= num_batches
+    correct /= size
+    print(f"Test Error: \n Accuracy: {(100*correct):>0.1f}%, Avg loss: {test_loss:>8f} \n")  
+    test_accuracy = 100*correct      
+    return test_loss, test_accuracy
+
+
+
+# apply train and test
+
+logname = "/PATH/Contextualizer_mlp/Experiments_cifar10/logs_contextualizer/logs_cifar10.csv"
+if not os.path.exists(logname):
+  with open(logname, 'w') as logfile:
+    logwriter = csv.writer(logfile, delimiter=',')
+    logwriter.writerow(['epoch', 'train loss', 'train acc',
+                        'test loss', 'test acc'])
+
+
+epochs = 100
+for epoch in range(epochs):
+    print(f"Epoch {epoch+1}\n-----------------------------------")
+    train_loss, train_acc = train(train_dataloader, model, loss_fn, optimizer)
+    # learning rate scheduler
+    #if scheduler is not None:
+    #    scheduler.step()
+    test_loss, test_acc = test(test_dataloader, model, loss_fn)
+    with open(logname, 'a') as logfile:
+        logwriter = csv.writer(logfile, delimiter=',')
+        logwriter.writerow([epoch+1, train_loss, train_acc,
+                            test_loss, test_acc])
+print("Done!")
+
+# saving trained model
+
+path = "/PATH/Contextualizer_mlp/Experiments_cifar10/weights_contextualizer"
+model_name = "ContextualizerImageClassification_cifar10"
+torch.save(model.state_dict(), f"{path}/{model_name}.pth")
+print(f"Saved Model State to {path}/{model_name}.pth ")
+

train_mlp_nin.py

@@ -0,0 +1,194 @@
+#imports
+
+import os
+import csv
+import torch
+from torch import nn 
+from torch.utils.data import DataLoader 
+from torchvision import datasets 
+from torchvision.transforms import ToTensor, Normalize, RandomCrop, RandomHorizontalFlip, Compose 
+from contextualizer_mlp_nin import ContextualizerNiN
+
+# data transforms
+
+transform = Compose([
+RandomCrop(32, padding=4),
+RandomHorizontalFlip(), 
+ToTensor(),
+Normalize((0.5, 0.5,0.5),(0.5, 0.5,0.5))
+
+])
+
+training_data = datasets.CIFAR10(
+                                       root='data',
+                                       train=True,
+                                       download=True,
+                                       transform=transform 
+                                       )
+
+test_data = datasets.CIFAR10(
+                                       root='data',
+                                       train=False,
+                                       download=True,
+                                       transform=transform 
+                                       )                                       
+# create dataloaders  
+                                     
+batch_size = 128
+
+train_dataloader = DataLoader(training_data, batch_size=batch_size,shuffle=True)
+test_dataloader = DataLoader(test_data, batch_size=batch_size)
+
+
+for X, y in test_dataloader:
+    print(f"Shape of X [N,C,H,W]:{X.shape}")
+    print(f"Shape of y:{y.shape}{y.dtype}")
+    break
+
+# size checking for loading images
+def check_sizes(image_size, patch_size):
+    sqrt_num_patches, remainder = divmod(image_size, patch_size)
+    assert remainder == 0, "`image_size` must be divisibe by `patch_size`"
+    num_patches = sqrt_num_patches ** 2
+    return num_patches
+
+
+
+# create model
+# Get cpu or gpu device for training.
+device = "cuda" if torch.cuda.is_available() else "cpu"
+
+print(f"using {device} device") 
+
+# model definition
+
+class ContextualizerNiNImageClassification(ContextualizerNiN):
+    def __init__(
+        self,
+        image_size=32,
+        patch_size=4,
+        in_channels=3,
+        num_classes=10,
+        d_ffn=512,
+        d_model = 256,
+        num_tokens = 64,
+        num_layers=4,
+        dropout=0.5
+    ):
+        num_patches = check_sizes(image_size, patch_size)
+        super().__init__(d_model,d_ffn,num_layers,dropout, num_tokens)
+        self.patcher = nn.Conv2d(
+            in_channels, d_model, kernel_size=patch_size, stride=patch_size
+        )
+        self.classifier = nn.Linear(d_model, num_classes)
+
+    def forward(self, x):
+        
+        patches = self.patcher(x)
+        batch_size, num_channels, _, _ = patches.shape
+        patches = patches.permute(0, 2, 3, 1)
+        patches = patches.view(batch_size, -1, num_channels)
+        embedding = self.model(patches)
+        embedding = embedding.mean(dim=1) # global average pooling
+        out = self.classifier(embedding)
+        return out
+
+model = ContextualizerNiNImageClassification().to(device)
+print(model)
+
+# Optimizer
+
+loss_fn = nn.CrossEntropyLoss()
+optimizer = torch.optim.Adam(model.parameters(),lr=1e-3)
+
+
+# Training Loop
+
+def train(dataloader, model, loss_fn, optimizer):
+    size = len(dataloader.dataset)
+    num_batches = len(dataloader)
+    model.train()
+    train_loss = 0
+    correct = 0
+    for batch, (X,y) in enumerate(dataloader):
+        X, y = X.to(device), y.to(device)
+       
+        #compute prediction error
+        pred = model(X)
+        loss = loss_fn(pred,y)
+        
+        # backpropagation
+        optimizer.zero_grad()
+        loss.backward()
+        optimizer.step()
+        train_loss += loss.item()
+        _, labels = torch.max(pred.data, 1)
+        correct += labels.eq(y.data).type(torch.float).sum()
+
+        
+
+
+        if batch % 100 == 0:
+            loss, current = loss.item(), batch * len(X)
+            print(f"loss: {loss:>7f}   [{current:>5d}/{size:>5d}]")
+
+    train_loss /= num_batches
+    train_accuracy = 100. * correct.item() / size
+    print(train_accuracy)
+    return train_loss,train_accuracy 
+
+
+
+# Test loop
+
+def test(dataloader, model, loss_fn):
+    size = len(dataloader.dataset)            
+    num_batches = len(dataloader)
+    model.eval()
+    test_loss = 0
+    correct = 0
+    with torch.no_grad():
+        for X,y in dataloader:
+            X,y = X.to(device), y.to(device)
+            pred = model(X)
+            test_loss += loss_fn(pred, y).item()
+            correct += (pred.argmax(1) == y).type(torch.float).sum().item()
+    test_loss /= num_batches
+    correct /= size
+    print(f"Test Error: \n Accuracy: {(100*correct):>0.1f}%, Avg loss: {test_loss:>8f} \n")  
+    test_accuracy = 100*correct      
+    return test_loss, test_accuracy
+
+
+
+# apply train and test
+
+logname = "/PATH/Contextualizer_mlp_NiN/Experiments_cifar10/logs_contextualizer/logs_cifar10.csv"
+if not os.path.exists(logname):
+  with open(logname, 'w') as logfile:
+    logwriter = csv.writer(logfile, delimiter=',')
+    logwriter.writerow(['epoch', 'train loss', 'train acc',
+                        'test loss', 'test acc'])
+
+
+epochs = 100
+for epoch in range(epochs):
+    print(f"Epoch {epoch+1}\n-----------------------------------")
+    train_loss, train_acc = train(train_dataloader, model, loss_fn, optimizer)
+    # learning rate scheduler
+    #if scheduler is not None:
+    #    scheduler.step()
+    test_loss, test_acc = test(test_dataloader, model, loss_fn)
+    with open(logname, 'a') as logfile:
+        logwriter = csv.writer(logfile, delimiter=',')
+        logwriter.writerow([epoch+1, train_loss, train_acc,
+                            test_loss, test_acc])
+print("Done!")
+
+# saving trained model
+
+path = "/PATH/Contextualizer_mlp_NiN/Experiments_cifar10/weights_contextualizer"
+model_name = "ContextualizerMLPNiNImageClassification_cifar10"
+torch.save(model.state_dict(), f"{path}/{model_name}.pth")
+print(f"Saved Model State to {path}/{model_name}.pth ")
+

train_nin.py

@@ -0,0 +1,194 @@
+#imports
+
+import os
+import csv
+import torch
+from torch import nn 
+from torch.utils.data import DataLoader 
+from torchvision import datasets 
+from torchvision.transforms import ToTensor, Normalize, RandomCrop, RandomHorizontalFlip, Compose 
+from contextualizer_nin import ContextualizerNiN
+
+# data transforms
+
+transform = Compose([
+RandomCrop(32, padding=4),
+RandomHorizontalFlip(), 
+ToTensor(),
+Normalize((0.5, 0.5,0.5),(0.5, 0.5,0.5))
+
+])
+
+training_data = datasets.CIFAR10(
+                                       root='data',
+                                       train=True,
+                                       download=True,
+                                       transform=transform 
+                                       )
+
+test_data = datasets.CIFAR10(
+                                       root='data',
+                                       train=False,
+                                       download=True,
+                                       transform=transform 
+                                       )                                       
+# create dataloaders  
+                                     
+batch_size = 128
+
+train_dataloader = DataLoader(training_data, batch_size=batch_size,shuffle=True)
+test_dataloader = DataLoader(test_data, batch_size=batch_size)
+
+
+for X, y in test_dataloader:
+    print(f"Shape of X [N,C,H,W]:{X.shape}")
+    print(f"Shape of y:{y.shape}{y.dtype}")
+    break
+
+# size checking for loading images
+def check_sizes(image_size, patch_size):
+    sqrt_num_patches, remainder = divmod(image_size, patch_size)
+    assert remainder == 0, "`image_size` must be divisibe by `patch_size`"
+    num_patches = sqrt_num_patches ** 2
+    return num_patches
+
+
+
+# create model
+# Get cpu or gpu device for training.
+device = "cuda" if torch.cuda.is_available() else "cpu"
+
+print(f"using {device} device") 
+
+# model definition
+
+class ContextualizerNiNImageClassification(ContextualizerNiN):
+    def __init__(
+        self,
+        image_size=32,
+        patch_size=4,
+        in_channels=3,
+        num_classes=10,
+        d_ffn=512,
+        d_model = 256,
+        num_tokens = 64,
+        num_layers=4,
+        dropout=0.5
+    ):
+        num_patches = check_sizes(image_size, patch_size)
+        super().__init__(d_model,d_ffn,num_layers,dropout, num_tokens)
+        self.patcher = nn.Conv2d(
+            in_channels, d_model, kernel_size=patch_size, stride=patch_size
+        )
+        self.classifier = nn.Linear(d_model, num_classes)
+
+    def forward(self, x):
+        
+        patches = self.patcher(x)
+        batch_size, num_channels, _, _ = patches.shape
+        patches = patches.permute(0, 2, 3, 1)
+        patches = patches.view(batch_size, -1, num_channels)
+        embedding = self.model(patches)
+        embedding = embedding.mean(dim=1) # global average pooling
+        out = self.classifier(embedding)
+        return out
+
+model = ContextualizerNiNImageClassification().to(device)
+print(model)
+
+# Optimizer
+
+loss_fn = nn.CrossEntropyLoss()
+optimizer = torch.optim.Adam(model.parameters(),lr=1e-3)
+
+
+# Training Loop
+
+def train(dataloader, model, loss_fn, optimizer):
+    size = len(dataloader.dataset)
+    num_batches = len(dataloader)
+    model.train()
+    train_loss = 0
+    correct = 0
+    for batch, (X,y) in enumerate(dataloader):
+        X, y = X.to(device), y.to(device)
+       
+        #compute prediction error
+        pred = model(X)
+        loss = loss_fn(pred,y)
+        
+        # backpropagation
+        optimizer.zero_grad()
+        loss.backward()
+        optimizer.step()
+        train_loss += loss.item()
+        _, labels = torch.max(pred.data, 1)
+        correct += labels.eq(y.data).type(torch.float).sum()
+
+        
+
+
+        if batch % 100 == 0:
+            loss, current = loss.item(), batch * len(X)
+            print(f"loss: {loss:>7f}   [{current:>5d}/{size:>5d}]")
+
+    train_loss /= num_batches
+    train_accuracy = 100. * correct.item() / size
+    print(train_accuracy)
+    return train_loss,train_accuracy 
+
+
+
+# Test loop
+
+def test(dataloader, model, loss_fn):
+    size = len(dataloader.dataset)            
+    num_batches = len(dataloader)
+    model.eval()
+    test_loss = 0
+    correct = 0
+    with torch.no_grad():
+        for X,y in dataloader:
+            X,y = X.to(device), y.to(device)
+            pred = model(X)
+            test_loss += loss_fn(pred, y).item()
+            correct += (pred.argmax(1) == y).type(torch.float).sum().item()
+    test_loss /= num_batches
+    correct /= size
+    print(f"Test Error: \n Accuracy: {(100*correct):>0.1f}%, Avg loss: {test_loss:>8f} \n")  
+    test_accuracy = 100*correct      
+    return test_loss, test_accuracy
+
+
+
+# apply train and test
+
+logname = "/PATH/Contextualizer_NiN/Experiments_cifar10/logs_contextualizer/logs_cifar10.csv"
+if not os.path.exists(logname):
+  with open(logname, 'w') as logfile:
+    logwriter = csv.writer(logfile, delimiter=',')
+    logwriter.writerow(['epoch', 'train loss', 'train acc',
+                        'test loss', 'test acc'])
+
+
+epochs = 100
+for epoch in range(epochs):
+    print(f"Epoch {epoch+1}\n-----------------------------------")
+    train_loss, train_acc = train(train_dataloader, model, loss_fn, optimizer)
+    # learning rate scheduler
+    #if scheduler is not None:
+    #    scheduler.step()
+    test_loss, test_acc = test(test_dataloader, model, loss_fn)
+    with open(logname, 'a') as logfile:
+        logwriter = csv.writer(logfile, delimiter=',')
+        logwriter.writerow([epoch+1, train_loss, train_acc,
+                            test_loss, test_acc])
+print("Done!")
+
+# saving trained model
+
+path = "/PATH/Contextualizer_NiN/Experiments_cifar10/weights_contextualizer"
+model_name = "ContextualizerNiNImageClassification_cifar10"
+torch.save(model.state_dict(), f"{path}/{model_name}.pth")
+print(f"Saved Model State to {path}/{model_name}.pth ")
+