Update mnist unsuper learning.

unsuper/minist.py

@@ -16,10 +16,8 @@ torch.cuda.manual_seed_all(seed)
 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 # device = torch.device("mps")
 
-# Hyper-parameters
 num_epochs = 1
 batch_size = 64
-learning_rate = 0.2
 
 transform = transforms.Compose([transforms.ToTensor()])
 
@@ -36,93 +34,132 @@ class ConvNet(nn.Module):
         self.pool = nn.MaxPool2d(2, 2)
         self.conv2 = nn.Conv2d(8, 1, 5, 1, 0)
         self.fc1 = nn.Linear(1 * 4 * 4, 10)
-        # self.fc2 = nn.Linear(120, 84)
-        # self.fc3 = nn.Linear(84, 10)
 
     def forward(self, x):
-        x = self.pool(F.relu(self.conv1(x)))
+        x = self.pool(self.conv1(x))
-        x = self.pool(F.relu(self.conv2(x)))
+        x = self.pool(self.conv2(x))
         x = x.view(x.shape[0], -1)
-        # x = F.relu(self.fc1(x))
-        # x = F.relu(self.fc2(x))
-        # x = self.fc3(x)
-
         x = self.fc1(x)
         return x
 
-    def printFector(self, x, label):
+    def forward_unsuper(self, x):
-        show.DumpTensorToImage(x.view(-1, x.shape[2], x.shape[3]), "input_image.png", Contrast=[0, 1.0])
+        x = self.pool(self.conv1(x))
+        return x
+
+    def forward_finetune(self, x):
+        x = self.pool(self.conv1(x))
+        x = self.pool(self.conv2(x))
+        x = x.view(x.shape[0], -1)
+        x = self.fc1(x)
+        return x
+
+    def printFector(self, x, label, dir=""):
+        show.DumpTensorToImage(x.view(-1, x.shape[2], x.shape[3]), dir + "/input_image.png", Contrast=[0, 1.0])
         # show.DumpTensorToLog(x, "input_image.log")
         x = self.conv1(x)
         w = self.conv1.weight
-        show.DumpTensorToImage(w.view(-1, w.shape[2], w.shape[3]), "conv1_weight.png", Contrast=[-1.0, 1.0])
+        show.DumpTensorToImage(w.view(-1, w.shape[2], w.shape[3]), dir + "/conv1_weight.png", Contrast=[-1.0, 1.0])
         # show.DumpTensorToLog(w, "conv1_weight.log")
 
-        show.DumpTensorToImage(x.view(-1, x.shape[2], x.shape[3]), "conv1_output.png", Contrast=[-1.0, 1.0])
+        show.DumpTensorToImage(x.view(-1, x.shape[2], x.shape[3]), dir + "/conv1_output.png", Contrast=[-1.0, 1.0])
         # show.DumpTensorToLog(x, "conv1_output.png")
 
         x = self.pool(F.relu(x))
         x = self.conv2(x)
         w = self.conv2.weight
-        show.DumpTensorToImage(w.view(-1, w.shape[2], w.shape[3]).cpu(), "conv2_weight.png", Contrast=[-1.0, 1.0])
+        show.DumpTensorToImage(
+            w.view(-1, w.shape[2], w.shape[3]).cpu(), dir + "/conv2_weight.png", Contrast=[-1.0, 1.0]
+        )
 
-        show.DumpTensorToImage(x.view(-1, x.shape[2], x.shape[3]).cpu(), "conv2_output.png", Contrast=[-1.0, 1.0])
+        show.DumpTensorToImage(
+            x.view(-1, x.shape[2], x.shape[3]).cpu(), dir + "/conv2_output.png", Contrast=[-1.0, 1.0]
+        )
         x = self.pool(F.relu(x))
-        show.DumpTensorToImage(x.view(-1, x.shape[2], x.shape[3]).cpu(), "pool_output.png", Contrast=[-1.0, 1.0])
+        show.DumpTensorToImage(x.view(-1, x.shape[2], x.shape[3]).cpu(), dir + "/pool_output.png", Contrast=[-1.0, 1.0])
         pool_shape = x.shape
         x = x.view(x.shape[0], -1)
         x = self.fc1(x)
         show.DumpTensorToImage(
-            self.fc1.weight.view(-1, pool_shape[2], pool_shape[3]), "fc_weight.png", Contrast=[-1.0, 1.0]
+            self.fc1.weight.view(-1, pool_shape[2], pool_shape[3]), dir + "/fc_weight.png", Contrast=[-1.0, 1.0]
         )
-        show.DumpTensorToImage(x.view(-1).cpu(), "fc_output.png")
+        show.DumpTensorToImage(x.view(-1).cpu(), dir + "/fc_output.png")
 
         criterion = nn.CrossEntropyLoss()
         loss = criterion(x, label)
-        optimizer.zero_grad()
         loss.backward()
 
+        if self.conv1.weight.requires_grad:
             w = self.conv1.weight.grad
-        show.DumpTensorToImage(w.view(-1, w.shape[2], w.shape[3]).cpu(), "conv1_weight_grad.png")
+            show.DumpTensorToImage(w.view(-1, w.shape[2], w.shape[3]).cpu(), dir + "/conv1_weight_grad.png")
+        if self.conv2.weight.requires_grad:
             w = self.conv2.weight.grad
-        show.DumpTensorToImage(w.view(-1, w.shape[2], w.shape[3]), "conv2_weight_grad.png")
+            show.DumpTensorToImage(w.view(-1, w.shape[2], w.shape[3]), dir + "/conv2_weight_grad.png")
-        show.DumpTensorToImage(self.fc1.weight.grad.view(-1, pool_shape[2], pool_shape[3]), "fc_weight_grad.png")
+        if self.fc1.weight.requires_grad:
+            show.DumpTensorToImage(
+                self.fc1.weight.grad.view(-1, pool_shape[2], pool_shape[3]), dir + "/fc_weight_grad.png"
+            )
 
 
 model = ConvNet().to(device)
+model.train()
 
-criterion = nn.CrossEntropyLoss()
+# Train the model unsuper
-optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
+epochs = 10
-
+model.conv1.weight.requires_grad = True
+model.conv2.weight.requires_grad = False
+model.fc1.weight.requires_grad = False
+optimizer_unsuper = torch.optim.SGD(model.parameters(), lr=0.1)
+n_total_steps = len(train_loader)
+for epoch in range(epochs):
+    for i, (images, labels) in enumerate(train_loader):
+        images = images.to(device)
+        outputs = model.forward_unsuper(images)
+        sample = outputs.view(outputs.shape[0], -1)
+        sample_mean = torch.mean(sample, dim=1, keepdim=True)
+        diff_mean = torch.mean(torch.abs(sample - sample_mean), dim=1, keepdim=True)
+        diff_ratio = (sample - sample_mean) / diff_mean
+        diff_ratio_mean = torch.mean(diff_ratio * diff_ratio, dim=1)
+        label = diff_ratio_mean * 0.5
+        loss = F.l1_loss(diff_ratio_mean, label)
+        optimizer_unsuper.zero_grad()
+        loss.backward()
+        optimizer_unsuper.step()
+        if (i + 1) % 100 == 0:
+            print(f"Epoch [{epoch+1}/{num_epochs}], Step [{i+1}/{n_total_steps}], Loss: {loss.item():.8f}")
 
 # Train the model
+model.conv1.weight.requires_grad = False
+model.conv2.weight.requires_grad = True
+model.fc1.weight.requires_grad = True
+criterion = nn.CrossEntropyLoss()
+optimizer = torch.optim.SGD(filter(lambda p: p.requires_grad, model.parameters()), lr=0.6)
 n_total_steps = len(train_loader)
 for epoch in range(num_epochs):
     for i, (images, labels) in enumerate(train_loader):
         images = images.to(device)
         labels = labels.to(device)
-
+        outputs = model.forward_finetune(images)
-        # Forward pass
-        outputs = model(images)
         loss = criterion(outputs, labels)
-
-        # Backward and optimize
         optimizer.zero_grad()
         loss.backward()
         optimizer.step()
-
         if (i + 1) % 100 == 0:
             print(f"Epoch [{epoch+1}/{num_epochs}], Step [{i+1}/{n_total_steps}], Loss: {loss.item():.4f}")
 
-test_loader = torch.utils.data.DataLoader(test_dataset, batch_size=1, shuffle=False)
-for images, labels in test_loader:
-    images = images.to(device)
-    labels = labels.to(device)
-    model.printFector(images, labels)
-    break
-
 print("Finished Training")
 
+test_loader = torch.utils.data.DataLoader(test_dataset, batch_size=1, shuffle=False)
+test_loader = iter(test_loader)
+images, labels = next(test_loader)
+images = images.to(device)
+labels = labels.to(device)
+model.printFector(images, labels, "dump1")
+
+images, labels = next(test_loader)
+images = images.to(device)
+labels = labels.to(device)
+model.printFector(images, labels, "dump2")
+
 # Test the model
 with torch.no_grad():
     n_correct = 0