Update minist unsuper.

unsuper/Readme.md

@@ -2,4 +2,8 @@
   1. 重复的权重，虽然权重看起来都一样，但是有稍微的不同，不是完全一样
   2. 3x3太小了导致了样本的信噪比太低，大部分的样本切出来都是0
 2. 5x5的时候会有网格状重复
-3. 7x7的时候边框区域问题
+3. 7x7的时候边框区域问题
+
+1. 输入的信噪比
+2. loss函数的设计
+3. grad信息的应用

unsuper/minist.py

@@ -30,7 +30,7 @@ test_loader = torch.utils.data.DataLoader(test_dataset, batch_size=batch_size, s
 class ConvNet(nn.Module):
     def __init__(self):
         super(ConvNet, self).__init__()
-        self.conv1 = nn.Conv2d(1, 8, 3, 1, 0)
+        self.conv1 = nn.Conv2d(1, 8, 5, 1, 0)
         self.pool = nn.MaxPool2d(2, 2)
         self.conv2 = nn.Conv2d(8, 1, 5, 1, 0)
         self.fc1 = nn.Linear(1 * 4 * 4, 10)
@@ -105,27 +105,11 @@ model = ConvNet().to(device)
 model.train()
 
 # Train the model unsuper
-epochs = 1
+epochs = 2
 n_total_steps = len(train_loader)
 for epoch in range(epochs):
     for i, (images, labels) in enumerate(train_loader):
         images = images.to(device)
-
-        # # images = images[:,:,12:15,12:15]
-        # kernel_size = 3
-        # mean_filter = torch.ones((1, 1, kernel_size, kernel_size), device=device) / (kernel_size * kernel_size)
-        # images = F.conv2d(images, mean_filter, padding=1)
-        # images = F.conv2d(images, mean_filter, padding=1)
-        # images = F.conv2d(images, mean_filter, padding=1)
-        # # images = F.conv2d(images, mean_filter, padding=1)
-        # # images = F.conv2d(images, mean_filter, padding=1)
-
-        # images = torch.rand(3, 3).to(device=device)
-        # # images[1, 1] = images[1, 1] * 2
-        # # images[0, 0] = images[1, 1] * 2
-        # # images[2, 2] = images[1, 1] * 2
-        # images = images.view(1, 1, 3, 3)
-
         outputs = model.forward_unsuper(images)
 
         outputs = outputs.permute(0, 2, 3, 1)  # 64 8 24 24 -> 64 24 24 8
@@ -137,9 +121,10 @@ for epoch in range(epochs):
         max = torch.expand_copy(max.reshape(-1, 1), sample.shape)
 
         all = range(0, sample.shape[0])
-        ratio_max = abs / mean
+        # ratio_max = abs / mean
-        ratio_nor = (max - abs) / max
+        # ratio_nor = (max - abs) / max
-        ratio_nor[all, max_index] = ratio_max[all, max_index].clone()
+        ratio_nor = torch.pow(abs / mean, 4)
+        # ratio_nor[all, max_index] = ratio_max[all, max_index].clone()
         ratio_nor = torch.where(torch.isnan(ratio_nor), 1.0, ratio_nor)
         label = sample * ratio_nor
 
@@ -147,7 +132,7 @@ for epoch in range(epochs):
         model.conv1.weight.grad = None
         loss.backward()
 
-        model.conv1.weight.data = model.conv1.weight.data - model.conv1.weight.grad * 100
+        model.conv1.weight.data = model.conv1.weight.data - model.conv1.weight.grad * 10000
 
         if (i + 1) % 100 == 0:
             print(f"Epoch [{epoch+1}/{epochs}], Step [{i+1}/{n_total_steps}], Loss: {loss.item():.8f}")
@@ -158,6 +143,8 @@ show.DumpTensorToImage(g.view(-1, g.shape[2], g.shape[3]).cpu(), "conv1_weight_g
 w = model.conv1.weight.data
 show.DumpTensorToImage(w.view(-1, w.shape[2], w.shape[3]), "conv1_weight_update.png")
 
+# model.conv1.weight.data = torch.rand(model.conv1.weight.data.shape, device=device)
+
 # loader = torch.utils.data.DataLoader(test_dataset, batch_size=1, shuffle=False)
 # images, labels = next(iter(loader))
 # images = images.to(device)