Upload 2 files

litetensormapper.py

@@ -0,0 +1,134 @@
+import torch
+from torch import nn, Tensor
+
+
+
+
+class VecDyT(nn.Module):
+    def __init__(self, input_shape):
+
+        super().__init__()
+
+        self.alpha = nn.Parameter(torch.randn(input_shape))
+
+    def forward(self, x):
+        x = torch.tanh(self.alpha * x)
+        return x
+
+
+class VecDyGeluSine(nn.Module):
+    def __init__(self, input_shape):
+
+        super().__init__()
+
+        self.alpha = nn.Parameter(torch.randn(input_shape))
+        self.beta = nn.Parameter(torch.randn(input_shape))
+        self.gamma = nn.Parameter(torch.randn(1))
+        self.etta = nn.Parameter(torch.randn(1))
+        self.gelu = nn.GELU()
+
+    def forward(self, x):
+
+
+
+        x = self.gamma * self.gelu(self.alpha * x) + self.etta * torch.sin(self.beta * x)
+
+        return x
+
+
+
+
+
+class TTT(nn.Module):
+    def __init__(self, dim: int):
+
+        super(TTT, self).__init__()
+
+
+        self.mapping = nn.Linear(dim,dim,bias=False)
+
+
+
+    def forward(self, in_seq: Tensor) -> Tensor:
+
+
+        outs = []
+
+        for seq in range(in_seq.size(1)):
+
+            state = in_seq[:,seq,:]
+            train_view = state + torch.randn_like(state)
+            label_view = state
+            loss = nn.functional.mse_loss(self.mapping(train_view), label_view)
+            grads = torch.autograd.grad(
+                loss, self.mapping.parameters(),create_graph=True)
+            with torch.no_grad():
+                for param, grad in zip(self.mapping.parameters(), grads):
+
+                    param -= 0.01 * grad
+
+            readout = self.mapping(in_seq[:,seq,:]).detach()
+            outs.append(readout)
+        out = torch.stack(outs, dim=1)
+
+        return out
+
+class FFUnit(nn.Module):
+    def __init__(self,dim):
+
+        super().__init__()
+
+        self.proj =  nn.Linear(dim,dim,bias=False)
+        self.modulate = VecDyGeluSine(dim)
+
+
+    def forward(self, x):
+
+        u, v = x, x
+
+        u = self.modulate(u)
+        v = self.proj(v)
+        g = u * v
+
+        return g
+
+class LiteTensorMapperBlock(nn.Module):
+    def __init__(self, dim, num_patch):
+
+        super().__init__()
+
+        self.norm_1 =  VecDyT(dim)
+        self.norm_2 =  VecDyT(dim)
+        self.memory = TTT(dim)
+        self.feedforward = FFUnit(dim)
+
+
+    def forward(self, x):
+
+
+        memorypath, FFpath = x, x
+
+        memorypath = self.norm_1(memorypath)
+
+        memorypath = self.memory(memorypath)
+
+        FFpath = self.norm_2(FFpath)
+
+        FFpath = self.feedforward(FFpath)
+
+        x = memorypath + FFpath
+
+        return x
+
+
+class LiteTensorMapper(nn.Module):
+    def __init__(self, d_model,num_patch, num_layers):
+        super().__init__()
+
+        self.model = nn.Sequential(
+            *[LiteTensorMapperBlock(d_model,num_patch) for _ in range(num_layers)]
+        )
+
+    def forward(self, x):
+
+        return self.model(x)

train.py

@@ -0,0 +1,186 @@
+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
+
+
+
+
+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
+                                       )
+
+
+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
+
+
+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
+
+
+
+
+device = "cuda" if torch.cuda.is_available() else "cpu"
+
+print(f"using {device} device")
+
+
+
+class TensorMapperImageClassification(LiteTensorMapper):
+    def __init__(
+        self,
+        image_size=32,
+        patch_size=4,
+        in_channels=3,
+        num_classes=10,
+        d_model = 256,
+        num_layers=4,
+
+
+    ):
+        num_patches = check_sizes(image_size, patch_size)
+        super().__init__(d_model, num_patches,num_layers)
+        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)
+        out = self.classifier(embedding)
+        return out
+
+model = TensorMapperImageClassification().to(device)
+print(model)
+
+
+
+loss_fn = nn.CrossEntropyLoss()
+optimizer = torch.optim.Adam(model.parameters(),lr=1e-3)
+
+
+
+
+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)
+
+
+        pred = model(X)
+        loss = loss_fn(pred,y)
+
+
+        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
+
+
+
+
+
+def test(dataloader, model, loss_fn):
+    size = len(dataloader.dataset)
+    num_batches = len(dataloader)
+    model.eval()
+    test_loss = 0
+    correct = 0
+
+    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
+
+
+
+
+
+logname = "/content/sample_data/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)
+    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!")
+
+
+
+path = "/content/sample_data/"
+model_name = "LiteTensorMapperImageClassification_cifar10"
+torch.save(model.state_dict(), f"{path}/{model_name}.pth")
+print(f"Saved Model State to {path}/{model_name}.pth ")