Upload 2 files

approximator.py

@@ -0,0 +1,271 @@
+import torch
+import numpy as np
+from torch import nn, einsum
+from torch.nn import functional as F
+from einops.layers.torch import Rearrange
+from einops import rearrange, reduce
+from math import ceil
+
+
+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)
+
+
+ # helper functions
+
+def exists(val):
+    return val is not None
+
+def moore_penrose_iter_pinv(x, iters = 6):
+    device = x.device
+
+    abs_x = torch.abs(x)
+    col = abs_x.sum(dim = -1)
+    row = abs_x.sum(dim = -2)
+    z = rearrange(x, '... i j -> ... j i') / (torch.max(col) * torch.max(row))
+
+    I = torch.eye(x.shape[-1], device = device)
+    I = rearrange(I, 'i j -> () i j')
+
+    for _ in range(iters):
+        xz = x @ z
+        z = 0.25 * z @ (13 * I - (xz @ (15 * I - (xz @ (7 * I - xz)))))
+
+    return z
+
+# main attention class
+
+class NystromAttention(nn.Module):
+    def __init__(
+        self,
+        dim,
+        dim_head = 64,
+        heads = 8,
+        num_landmarks = 256,
+        pinv_iterations = 6,
+        residual = True,
+        residual_conv_kernel = 33,
+        eps = 1e-8,
+        dropout = 0.
+    ):
+        super().__init__()
+        self.eps = eps
+        inner_dim = heads * dim_head
+
+        self.num_landmarks = num_landmarks
+        self.pinv_iterations = pinv_iterations
+
+        self.heads = heads
+        self.scale = dim_head ** -0.5
+        self.to_qkv = nn.Linear(dim, inner_dim * 3, bias = False)
+
+        self.to_out = nn.Sequential(
+            nn.Linear(inner_dim, dim),
+            nn.Dropout(dropout)
+        )
+
+        self.residual = residual
+        if residual:
+            kernel_size = residual_conv_kernel
+            padding = residual_conv_kernel // 2
+            self.res_conv = nn.Conv2d(heads, heads, (kernel_size, 1), padding = (padding, 0), groups = heads, bias = False)
+
+    def forward(self, x, mask = None, return_attn = False):
+        b, n, _, h, m, iters, eps = *x.shape, self.heads, self.num_landmarks, self.pinv_iterations, self.eps
+
+        # pad so that sequence can be evenly divided into m landmarks
+
+        remainder = n % m
+        if remainder > 0:
+            padding = m - (n % m)
+            x = F.pad(x, (0, 0, padding, 0), value = 0)
+
+            if exists(mask):
+                mask = F.pad(mask, (padding, 0), value = False)
+
+        # derive query, keys, values
+
+        q, k, v = self.to_qkv(x).chunk(3, dim = -1)
+        q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h = h), (q, k, v))
+
+        # set masked positions to 0 in queries, keys, values
+
+        if exists(mask):
+            mask = rearrange(mask, 'b n -> b () n')
+            q, k, v = map(lambda t: t * mask[..., None], (q, k, v))
+
+        q = q * self.scale
+
+        # generate landmarks by sum reduction, and then calculate mean using the mask
+
+        l = ceil(n / m)
+        landmark_einops_eq = '... (n l) d -> ... n d'
+        q_landmarks = reduce(q, landmark_einops_eq, 'sum', l = l)
+        k_landmarks = reduce(k, landmark_einops_eq, 'sum', l = l)
+
+        # calculate landmark mask, and also get sum of non-masked elements in preparation for masked mean
+
+        divisor = l
+        if exists(mask):
+            mask_landmarks_sum = reduce(mask, '... (n l) -> ... n', 'sum', l = l)
+            divisor = mask_landmarks_sum[..., None] + eps
+            mask_landmarks = mask_landmarks_sum > 0
+
+        # masked mean (if mask exists)
+
+        q_landmarks /= divisor
+        k_landmarks /= divisor
+
+        # similarities
+
+        einops_eq = '... i d, ... j d -> ... i j'
+        sim1 = einsum(einops_eq, q, k_landmarks)
+        sim2 = einsum(einops_eq, q_landmarks, k_landmarks)
+        sim3 = einsum(einops_eq, q_landmarks, k)
+
+        # masking
+
+        if exists(mask):
+            mask_value = -torch.finfo(q.dtype).max
+            sim1.masked_fill_(~(mask[..., None] * mask_landmarks[..., None, :]), mask_value)
+            sim2.masked_fill_(~(mask_landmarks[..., None] * mask_landmarks[..., None, :]), mask_value)
+            sim3.masked_fill_(~(mask_landmarks[..., None] * mask[..., None, :]), mask_value)
+
+        # eq (15) in the paper and aggregate values
+
+        attn1, attn2, attn3 = map(lambda t: t.softmax(dim = -1), (sim1, sim2, sim3))
+        attn2_inv = moore_penrose_iter_pinv(attn2, iters)
+
+        out = (attn1 @ attn2_inv) @ (attn3 @ v)
+
+        # add depth-wise conv residual of values
+
+        if self.residual:
+            out += self.res_conv(v)
+
+        # merge and combine heads
+
+        out = rearrange(out, 'b h n d -> b n (h d)', h = h)
+        out = self.to_out(out)
+        out = out[:, -n:]
+
+        if return_attn:
+            attn = attn1 @ attn2_inv @ attn3
+            return out, attn
+
+        return out
+
+
+
+
+
+
+
+class NystromBlock(nn.Module):
+    def __init__(self,dim,dim_ffn, dropout):
+        super().__init__()     
+        self.Nystrom = NystromAttention(
+        dim,
+        dim_head = 64,
+        heads = 4,
+        num_landmarks = 32,
+        pinv_iterations = 3,
+        residual = True,
+        residual_conv_kernel = 33,
+        eps = 1e-8,
+        dropout = dropout)
+        self.norm = nn.LayerNorm(dim)       
+        
+        self.ffn = FeedForward(dim,dim_ffn,dropout)
+
+    def forward(self, x):
+        res = x  
+        x = self.norm(x) 
+        x = self.Nystrom(x)
+        x = res + x
+        res = x
+        x = self.norm(x)
+        x = self.ffn(x)
+        out = x + res
+        return out       
+      
+
+
+
+class ApproximatorGatingUnit(nn.Module):
+    def __init__(self,d_model,d_ffn,dropout):
+        super().__init__()
+        #self.proj = nn.Linear(d_model, d_model)    
+        self.Approx_1 = NystromBlock(d_model,d_ffn,dropout)
+        self.Approx_2 = NystromBlock(d_model,d_ffn,dropout)    
+
+	
+       
+
+    def forward(self, x):
+        u, v = x, x 
+        u = self.Approx_1(u)  
+        v = self.Approx_2(v)
+        out = u * v
+        return out
+
+
+class ApproximatorBlock(nn.Module):
+    def __init__(self, d_model, d_ffn,dropout):
+        super().__init__()
+       
+        self.norm = nn.LayerNorm(d_model)       
+        self.agu = ApproximatorGatingUnit(d_model,d_ffn,dropout)
+        self.ffn = FeedForward(d_model,d_ffn,dropout)
+    def forward(self, x):
+        residual = x
+        x = self.norm(x)
+        x = self.agu(x)   
+        x = x + residual      
+        residual = x
+        x = self.norm(x)
+        x = self.ffn(x)
+        out = x + residual
+        return out
+
+
+
+
+
+
+
+
+
+class Approximator(nn.Module):
+    def __init__(self, d_model, d_ffn, num_layers,dropout):
+        super().__init__()
+        
+        self.model = nn.Sequential(
+            
+            *[ApproximatorBlock(d_model,d_ffn,dropout) for _ in range(num_layers)],
+            
+            
+        )
+
+    def forward(self, x):
+        
+        x = self.model(x)
+        
+        return x
+
+
+
+
+
+
+

train.py

@@ -0,0 +1,195 @@
+#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 approximator import Approximator
+
+# 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 ApproximatorImageClassification(Approximator):
+    def __init__(
+        self,
+        image_size=32,
+        patch_size=4,
+        in_channels=3,
+        num_classes=10,
+        d_model=256,
+        d_ffn=512,
+        num_layers=4,
+        dropout=0.5
+    ):
+        num_patches = check_sizes(image_size, patch_size)
+        super().__init__(d_model, d_ffn, num_layers,dropout)
+        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 = ApproximatorImageClassification().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 = "/home/abdullah/Desktop/Proposals_experiments/Approximator/Experiments_cifar10/logs_approximator/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 = "/home/abdullah/Desktop/Proposals_experiments/Approximator/Experiments_cifar10/weights_approximator"
+model_name = "ApproximatorImageClassification_cifar10"
+torch.save(model.state_dict(), f"{path}/{model_name}.pth")
+print(f"Saved Model State to {path}/{model_name}.pth ")
+