diff --git a/unsuper/Readme.md b/unsuper/Readme.md index cb7a678..bc8fdab 100644 --- a/unsuper/Readme.md +++ b/unsuper/Readme.md @@ -1,21 +1,9 @@ +## 方向 1. 输入的信噪比 2. loss函数的设计 3. grad信息的应用 - -abs.cpu().detach().numpy() -array([[8.1206687e-02, 2.2388995e-05, 3.7080176e-02, 5.7033218e-02, - 1.7404296e-03, 7.6270252e-02, 5.9453689e-02, 4.0801242e-05]], - dtype=float32) -ratio_nor.cpu().detach().numpy() -array([[4.3121886e+00, 3.2778070e-07, 8.9907879e-01, 2.1270120e+00, - 1.9807382e-03, 3.8038602e+00, 2.3113825e+00, 1.0885816e-06]], - dtype=float32) - -都比较差的时候区分不开 - - ## 发现的问题 1. 3x3的时候会有重复 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 @@ -31,10 +19,13 @@ array([[4.3121886e+00, 3.2778070e-07, 8.9907879e-01, 2.1270120e+00, 2. 网格状，就是为了尽量降低最终输出的绝对值，降低loss 1. 形成4个一组，一共2组，的形式，交替成为最大值，把另外一组的输出降低，最后都输出最低的绝对值 2. 需要类似batchnormal的方式对各个conv核心之间进行归一化 - 3. 采用label不改变原来sample output的 abs均值的限制 - 1. 生成2个分别交替充当最大值的极端（只有2个像素不是0）的卷积核，其他的卷积核都是输出接近0的网格 - 4. 多个卷积核之间需要有差异，同一个卷积核的不同样本输入也要有差异，卷积核的分布要有要求 - 5. 每个卷积核尽量平摊权重到所有像素，而不是集中一个像素？提高鲁棒性？ + 3. 采用label不改变原来sample output的 abs均值的限制，也就是不改变所有卷积核输出的总能量 + 1. 生成2个分别交替充当最大值的极端（只有2个像素不是0）的卷积核， + 2. 其他的几个卷积核都是输出接近0的网格 + 3. 为什么生成了极端的卷积核？为什么还是有相同的卷积核？是不是数据集导致了这样的结果？ + 4. 多个卷积核之间需要有差异，同一个卷积核的不同样本输入也要有差异，卷积核的分布要有要求 + 5. 每个卷积核尽量平摊权重到所有像素，而不是集中一个像素？提高鲁棒性？ + 6. 采用自动的ratio增益控制之后，好像没有重复了卷积核了，但是还有极端的2个像素有效的卷积核 ## 可能的策略 1. 每个卷积核的改变权重(grad)能量守恒 diff --git a/unsuper/minist.py b/unsuper/minist.py index 858c29c..cdef917 100644 --- a/unsuper/minist.py +++ b/unsuper/minist.py @@ -117,7 +117,7 @@ for epoch in range(epochs): images = images.to(device) # images = torch.ones((1, 1, 5, 5), device=device) - # type = random.randint(0, 7) + # type = random.randint(0, 3) # if type == 0: # rand = random.randint(0, 4) # images[:, :, rand, :] = images[:, :, rand, :] * 0.5 @@ -131,8 +131,8 @@ for epoch in range(epochs): # images[:, :, 3, 3] = images[:, :, 3, 3] * 0.5 # images[:, :, 4, 4] = images[:, :, 4, 4] * 0.5 # if type == 3: - # randx = random.randint(0, 2) - # randy = random.randint(0, 2) + # randx = random.randint(1, 3) + # randy = random.randint(1, 3) # images[:, :, randx, randy] = images[:, :, randx, randy] * 0.5 # images[:, :, randx, randy + 1] = images[:, :, randx, randy + 1] * 0.5 # images[:, :, randx, randy - 1] = images[:, :, randx, randy - 1] * 0.5 @@ -143,7 +143,6 @@ for epoch in range(epochs): outputs = outputs.permute(0, 2, 3, 1) # 64 8 24 24 -> 64 24 24 8 sample = outputs.reshape(-1, outputs.shape[3]) # -> 36864 8 - # sample = outputs.reshape(-1, 8,24*24) # -> 36864 8 # sample = torch.mean(sample,dim=2) # -> 36864 8 @@ -152,7 +151,15 @@ for epoch in range(epochs): mean = torch.mean(abs, dim=1) mean = torch.expand_copy(mean.reshape(-1, 1), abs.shape) max = torch.expand_copy(max.reshape(-1, 1), abs.shape) - ratio = torch.pow(abs / mean, 2) + + e = torch.sum(torch.pow(abs - mean, 2), dim=1) + e = torch.expand_copy(e.reshape(-1, 1), abs.shape) + e = 1 / e + e = torch.where(torch.isinf(e), 1.0, e) + e = torch.pow(e, 0.5) + + ratio = abs / mean * e + # ratio = torch.pow(abs / mean, e ) ratio = torch.where(torch.isnan(ratio), 0.0, ratio) label = ratio * abs @@ -160,15 +167,13 @@ for epoch in range(epochs): label = label - label_mean + mean sample = torch.abs(sample) - sample_nz = sample[abs > 0] - label_nz = label[abs > 0] - loss = F.l1_loss(sample_nz, label_nz) + loss = F.l1_loss(sample[abs > 0], label[abs > 0]) model.conv1.weight.grad = None loss.backward() # if epoch >= (epochs - 1): # continue - model.conv1.weight.data = model.conv1.weight.data - model.conv1.weight.grad * 0.1 + model.conv1.weight.data = model.conv1.weight.data - model.conv1.weight.grad * 0.01 model.conv1.weight.data = model.normal_conv1_weight() if (i + 1) % 100 == 0: