New ONNX ops and testsHEADmaster

New ops: Slice, Squeeze, Unsqueeze
New tests based on papers:
 - Wide-to-Deep, Deep-to-Wide Transformation
 - Pruining of stably inactive (always negative) and active (always
   positive) ReLUs

5 files changed, 290 insertions, 3 deletions

diff --git a/examples/mnist/mnist.py b/examples/mnist/mnist.py
index 0a81878..a1706be 100644
--- a/examples/mnist/mnist.py
+++ b/examples/mnist/mnist.py
@@ -24,7 +24,7 @@ def train_model(name: str, dim):
     optimizer = torch.optim.Adam(net.parameters(), lr=0.5e-4)
 
     print(f"Training {name} ({dim} neurons)...")
-    for epoch in range(100):
+    for epoch in range(10):
         global loss
         for data in trainloader:
             inputs, targets = data
@@ -33,8 +33,7 @@ def train_model(name: str, dim):
             loss = loss_fn(outputs, targets)
             loss.backward()
             optimizer.step()
-        if (epoch + 1) % 10 == 0:
-            print(f"  Epoch {epoch+1}, Loss: {loss.item():.4f}")
+        print(f"  Epoch {epoch+1}, Loss: {loss.item():.4f}")
     return net
 
 if __name__ == "__main__":
diff --git a/examples/mnist/mnist_deep_to_wide.py b/examples/mnist/mnist_deep_to_wide.py
new file mode 100644
index 0000000..6c4dde9
--- /dev/null
+++ b/examples/mnist/mnist_deep_to_wide.py
@@ -0,0 +1,85 @@
+import torch, torch.nn as nn
+from torchvision.datasets import MNIST
+from torch.utils.data import DataLoader
+from torchvision import transforms
+
+class Deep_MNIST_Block(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.fc1 = nn.Linear(784, 1)
+        self.fc2 = nn.Linear(1, 10, bias=False)
+        
+    def forward(self, x_s):
+        x, s = x_s
+        z = torch.relu(self.fc1(x))
+        ds = self.fc2(z)
+        return (x, s + ds)
+
+class Deep_MNIST_MLP(nn.Module):
+    def __init__(self, num_blocks):
+        super().__init__()
+        self.flatten = nn.Flatten()
+        self.blocks = nn.Sequential(*[Deep_MNIST_Block() for _ in range(num_blocks)])
+        self.final_bias = nn.Parameter(torch.zeros(10))
+
+    def forward(self, x):
+        x = self.flatten(x)
+        s = torch.zeros(x.shape[0], 10, device=x.device)
+        _, final_s = self.blocks((x, s))
+        return final_s + self.final_bias
+
+class Wide_MNIST_MLP(nn.Module):
+    def __init__(self, deep_net):
+        super().__init__()
+        self.flatten = nn.Flatten()
+        num_neurons = len(deep_net.blocks)
+        self.layers = nn.Sequential(
+            nn.Linear(784, num_neurons),
+            nn.ReLU(),
+            nn.Linear(num_neurons, 10),
+        )
+        with torch.no_grad():
+            w1_all = []
+            b1_all = []
+            w2_all = []
+            
+            for block in deep_net.blocks:
+                w1_all.append(block.fc1.weight.data)
+                b1_all.append(block.fc1.bias.data)
+                w2_all.append(block.fc2.weight.data)
+                
+            self.layers[0].weight.copy_(torch.cat(w1_all, dim=0)) # pyright: ignore
+            self.layers[0].bias.copy_(torch.cat(b1_all, dim=0)) # pyright: ignore
+            self.layers[2].weight.copy_(torch.cat(w2_all, dim=1)) # pyright: ignore
+            self.layers[2].bias.copy_(deep_net.final_bias.data) # pyright: ignore
+
+    def forward(self, x):
+        x = self.flatten(x)
+        return self.layers(x)
+
+train_dataset = MNIST('./', download=True, transform=transforms.ToTensor(), train=True)
+trainloader = DataLoader(train_dataset, batch_size=128, shuffle=True)
+
+def train_deep_model(name: str, num_blocks):
+    net = Deep_MNIST_MLP(num_blocks=num_blocks)
+    loss_fn = nn.CrossEntropyLoss()
+    optimizer = torch.optim.Adam(net.parameters(), lr=1e-3)
+
+    print(f"Training {name} ({num_blocks} blocks)...")
+    for epoch in range(10):
+        global loss
+        for inputs, targets in trainloader:
+            optimizer.zero_grad()
+            outputs = net(inputs)
+            loss = loss_fn(outputs, targets)
+            loss.backward()
+            optimizer.step()
+        print(f"  Epoch {epoch+1}, Loss: {loss.item():.4f}")
+    return net
+
+if __name__ == "__main__":
+    torch_net_a = train_deep_model("Deep Network", 8).eval()
+    torch_net_b = Wide_MNIST_MLP(torch_net_a).eval()
+    
+    torch.onnx.export(torch_net_a, (torch.randn(1, 28, 28),), "mnist_a.onnx")
+    torch.onnx.export(torch_net_b, (torch.randn(1, 28, 28),), "mnist_b.onnx")
diff --git a/examples/mnist/mnist_stably_active.py b/examples/mnist/mnist_stably_active.py
new file mode 100644
index 0000000..267d682
--- /dev/null
+++ b/examples/mnist/mnist_stably_active.py
@@ -0,0 +1,69 @@
+import torch, torch.nn as nn
+from torchvision.datasets import MNIST
+from torch.utils.data import DataLoader
+from torchvision import transforms
+
+class MNIST_MLP(nn.Module):
+    def __init__(self, hidden_dim):
+        super().__init__()
+        self.flatten = nn.Flatten()
+        self.layers = nn.Sequential(
+            nn.Linear(784, hidden_dim),
+            nn.ReLU(),
+            nn.Linear(hidden_dim, 10),
+        )
+    def forward(self, x):
+        x = self.flatten(x)
+        return self.layers(x)
+
+class MNIST_Linear(nn.Module):
+    def __init__(self, weight, bias):
+        super().__init__()
+        self.flatten = nn.Flatten()
+        self.fc = nn.Linear(784, 10)
+        self.fc.weight.data = weight
+        self.fc.bias.data = bias
+    def forward(self, x):
+        x = self.flatten(x)
+        return self.fc(x)
+
+train_dataset = MNIST('./', download=True, transform=transforms.ToTensor(), train=True)
+trainloader = DataLoader(train_dataset, batch_size=128, shuffle=True)
+
+def train_model(name: str, dim):
+    net = MNIST_MLP(hidden_dim=dim)
+    loss_fn = nn.CrossEntropyLoss()
+    optimizer = torch.optim.Adam(net.parameters(), lr=0.5e-4)
+
+    print(f"Training {name} ({dim} neurons)...")
+    for epoch in range(10):
+        global loss
+        for data in trainloader:
+            inputs, targets = data
+            optimizer.zero_grad()
+            outputs = net(inputs)
+            loss = loss_fn(outputs, targets)
+            loss.backward()
+            optimizer.step()
+        print(f"  Epoch {epoch+1}, Loss: {loss.item():.4f}")
+    return net
+
+if __name__ == "__main__":
+    torch_net_a = train_model("Base Network", 6).eval()
+
+    with torch.no_grad():
+        torch_net_a.layers[0].weight.fill_(0.01) # pyright: ignore
+        torch_net_a.layers[0].bias.fill_(5.0) # pyright: ignore
+
+        W1 = torch_net_a.layers[0].weight.data
+        b1 = torch_net_a.layers[0].bias.data
+        W2 = torch_net_a.layers[2].weight.data
+        b2 = torch_net_a.layers[2].bias.data
+
+        W_collapsed = torch.matmul(W2, W1) # pyright: ignore
+        b_collapsed = torch.matmul(W2, b1) + b2 # pyright: ignore
+
+        torch_net_b = MNIST_Linear(W_collapsed, b_collapsed).eval()
+
+    torch.onnx.export(torch_net_a, (torch.randn(1, 28, 28),), "mnist_a.onnx")
+    torch.onnx.export(torch_net_b, (torch.randn(1, 28, 28),), "mnist_b.onnx")
diff --git a/examples/mnist/mnist_stably_inactive.py b/examples/mnist/mnist_stably_inactive.py
new file mode 100644
index 0000000..ad81461
--- /dev/null
+++ b/examples/mnist/mnist_stably_inactive.py
@@ -0,0 +1,52 @@
+import torch, torch.nn as nn
+from torchvision.datasets import MNIST
+from torch.utils.data import DataLoader
+from torchvision import transforms
+
+class MNIST_MLP(nn.Module):
+    def __init__(self, hidden_dim):
+        super().__init__()
+        self.layers = nn.Sequential(
+            nn.Flatten(),
+            nn.Linear(28 * 28, hidden_dim),
+            nn.ReLU(),
+            nn.Linear(hidden_dim, 10),
+        )
+    def forward(self, x):
+        return self.layers(x)
+
+train_dataset = MNIST('./', download=True, transform=transforms.ToTensor(), train=True)
+trainloader = DataLoader(train_dataset, batch_size=128, shuffle=True)
+
+def train_model(name: str, dim):
+    net = MNIST_MLP(hidden_dim=dim)
+    loss_fn = nn.CrossEntropyLoss()
+    optimizer = torch.optim.Adam(net.parameters(), lr=0.5e-4)
+
+    print(f"Training {name} ({dim} neurons)...")
+    for epoch in range(10):
+        global loss
+        for data in trainloader:
+            inputs, targets = data
+            optimizer.zero_grad()
+            outputs = net(inputs)
+            loss = loss_fn(outputs, targets)
+            loss.backward()
+            optimizer.step()
+        print(f"  Epoch {epoch+1}, Loss: {loss.item():.4f}")
+    return net
+
+if __name__ == "__main__":
+    torch_net_a = train_model("Base Network", 6).eval()
+    
+    with torch.no_grad():
+        torch_net_a.layers[1].weight[5] = -1.0 # pyright: ignore
+        torch_net_a.layers[1].bias[5] = -1.0 # pyright: ignore
+        
+        torch_net_b = MNIST_MLP(6).eval()
+        torch_net_b.load_state_dict(torch_net_a.state_dict())
+        
+        torch_net_b.layers[3].weight[:, 5] = 0.0 # pyright: ignore
+        
+    torch.onnx.export(torch_net_a, (torch.randn(1, 28, 28),), "mnist_a.onnx")
+    torch.onnx.export(torch_net_b, (torch.randn(1, 28, 28),), "mnist_b.onnx")
diff --git a/examples/mnist/mnist_wide_to_deep.py b/examples/mnist/mnist_wide_to_deep.py
new file mode 100644
index 0000000..64f4ec7
--- /dev/null
+++ b/examples/mnist/mnist_wide_to_deep.py
@@ -0,0 +1,82 @@
+import torch, torch.nn as nn
+from torchvision.datasets import MNIST
+from torch.utils.data import DataLoader
+from torchvision import transforms
+
+class Wide_MNIST_MLP(nn.Module):
+    def __init__(self, hidden_dim):
+        super().__init__()
+        self.flatten = nn.Flatten()
+        self.layers = nn.Sequential(
+            nn.Linear(28 * 28, hidden_dim),
+            nn.ReLU(),
+            nn.Linear(hidden_dim, 10),
+        )
+    def forward(self, x):
+        x = self.flatten(x)
+        return self.layers(x)
+
+class Deep_MNIST_Block(nn.Module):
+    def __init__(self, w1, b1, w2):
+        super().__init__()
+        self.fc1 = nn.Linear(784, 1)
+        self.fc1.weight.data = w1.clone().unsqueeze(0)
+        self.fc1.bias.data = b1.clone().unsqueeze(0)
+        self.fc2 = nn.Linear(1, 10, bias=False)
+        self.fc2.weight.data = w2.clone().unsqueeze(1)
+        
+    def forward(self, x_s):
+        x, s = x_s
+        z = torch.relu(self.fc1(x))
+        ds = self.fc2(z)
+        return (x, s + ds)
+
+class Deep_MNIST_MLP(nn.Module):
+    def __init__(self, wide_net):
+        super().__init__()
+        self.flatten = nn.Flatten()
+        
+        w1 = wide_net.layers[0].weight.data
+        b1 = wide_net.layers[0].bias.data
+        w2 = wide_net.layers[2].weight.data
+        b2 = wide_net.layers[2].bias.data
+        num_neurons = w1.shape[0]
+        
+        self.blocks = nn.Sequential(*[
+            Deep_MNIST_Block(w1[j], b1[j], w2[:, j]) for j in range(num_neurons)
+        ])
+        self.final_bias = nn.Parameter(b2.clone())
+
+    def forward(self, x):
+        x = self.flatten(x)
+        s = torch.zeros(x.shape[0], 10, device=x.device)
+        _, final_s = self.blocks((x, s))
+        return final_s + self.final_bias
+
+train_dataset = MNIST('./', download=True, transform=transforms.ToTensor(), train=True)
+trainloader = DataLoader(train_dataset, batch_size=128, shuffle=True)
+
+def train_model(name: str, dim):
+    net = Wide_MNIST_MLP(hidden_dim=dim)
+    loss_fn = nn.CrossEntropyLoss()
+    optimizer = torch.optim.Adam(net.parameters(), lr=0.5e-4)
+
+    print(f"Training {name} ({dim} neurons)...")
+    for epoch in range(10):
+        global loss
+        for data in trainloader:
+            inputs, targets = data
+            optimizer.zero_grad()
+            outputs = net(inputs)
+            loss = loss_fn(outputs, targets)
+            loss.backward()
+            optimizer.step()
+        print(f"  Epoch {epoch+1}, Loss: {loss.item():.4f}")
+    return net
+
+if __name__ == "__main__":
+    torch_net_a = train_model("Wide Network", 8).eval()
+    torch_net_b = Deep_MNIST_MLP(torch_net_a).eval()
+
+    torch.onnx.export(torch_net_a, (torch.randn(1, 28, 28),), "mnist_a.onnx")
+    torch.onnx.export(torch_net_b, (torch.randn(1, 28, 28),), "mnist_b.onnx")