home/cnn_num_layer_test.md

@@ -6,6 +6,10 @@ I compare which model of the CNN is better using either of 3 or 4.
 - process of the regression
 The score is negative in the testing set (seems not good)
 
+
+![Screenshot_2021-03-26_hist_overlap_test_ori_png__PNG_Image__640___480_pixels_](uploads/ad3eab4e59d1aa47563414a49e92bb5f/Screenshot_2021-03-26_hist_overlap_test_ori_png__PNG_Image__640___480_pixels_.png)
+
+
 ```
 0 0.710105836391449 0.5692983269691467 0.18870583176612854 0.7803471684455872 0.7499625086784363
 1 0.6804407835006714 0.587370753288269 -0.1154908537864685 0.8309608101844788 0.7888878583908081
@@ -49,97 +53,3 @@ The score is negative in the testing set (seems not good)
 
 ## Model
 
-```
-class DeepPixel(torch.nn.Module):
-
-    def __init__(self, in_channels, kernel_size=(7,7), last_kernel=(22,22), num_layer=4):
-        super().__init__()
-
-
-        # self.input_conv = torch.nn.Sequential(
-        #     torch.nn.Conv2d(in_channels, in_channels, kernel_size=7, stride=1, padding=3),
-        #     torch.nn.BatchNorm2d(in_channels),
-        #     torch.nn.Tanh()
-        # )
-
-        # check the input layer with different settings
-        self.input_conv = torch.nn.Sequential(
-            torch.nn.Conv2d(in_channels, in_channels, kernel_size=7, stride=1, padding=3),
-            torch.nn.BatchNorm2d(in_channels),
-            torch.nn.ReLU()
-        )
-
-
-
-        self.downsampler = torch.nn.Sequential()
-        self.downsampler.add_module('CONV_1', torch.nn.Conv2d(in_channels, 8, kernel_size=kernel_size, stride=2, padding=3))
-        self.downsampler.add_module('BN_1', torch.nn.BatchNorm2d(8))
-        self.downsampler.add_module('TANH_1',  torch.nn.ReLU())
-        self.downsampler.add_module('CONV_2',  torch.nn.Conv2d(8, 16, kernel_size=kernel_size, stride=2, padding=3))
-        self.downsampler.add_module('BN_2', torch.nn.BatchNorm2d(16))
-        self.downsampler.add_module('TANH_2',  torch.nn.ReLU())
-        self.downsampler.add_module('CONV_3',  torch.nn.Conv2d(16, 32, kernel_size=kernel_size, stride=2, padding=3))
-        self.downsampler.add_module('BN_3', torch.nn.BatchNorm2d(32))
-        self.downsampler.add_module('TANH_3',  torch.nn.ReLU())
-        if num_layer == 4:
-            self.downsampler.add_module('CONV_4',  torch.nn.Conv2d(32, 64, kernel_size=kernel_size, stride=2, padding=3))
-            self.downsampler.add_module('BN_4', torch.nn.BatchNorm2d(64))
-            self.downsampler.add_module('TANH_4',  torch.nn.ReLU())
-        else:
-            pass
-
-        self.upsampler = torch.nn.Sequential()
-        if num_layer == 4:
-            self.upsampler.add_module(
-                'CONVTRANS_1',  torch.nn.ConvTranspose2d(64, 32, kernel_size=kernel_size, stride=2, padding=3, output_padding=1))
-            self.upsampler.add_module('BN_1', torch.nn.BatchNorm2d(32))
-            self.upsampler.add_module('TANH_1',  torch.nn.ReLU())
-        else:
-            pass
-        self.upsampler.add_module(
-            'CONVTRANS_2',  torch.nn.ConvTranspose2d(32, 16, kernel_size=kernel_size, stride=2, padding=3, output_padding=1))
-        self.upsampler.add_module('BN_2', torch.nn.BatchNorm2d(16))
-        self.upsampler.add_module('TANH_2',  torch.nn.ReLU())
-        self.upsampler.add_module(
-            'CONVTRANS_3', torch.nn.ConvTranspose2d(16, 8, kernel_size=kernel_size, stride=2, padding=3, output_padding=1))
-        self.upsampler.add_module('BN_3', torch.nn.BatchNorm2d(8))
-        self.upsampler.add_module('TANH_3',  torch.nn.ReLU())
-        self.upsampler.add_module(
-            'CONVTRANS_4',  torch.nn.ConvTranspose2d(8, in_channels, kernel_size=kernel_size, stride=2, padding=3, output_padding=1))
-        self.upsampler.add_module('BN_4', torch.nn.BatchNorm2d(in_channels))
-        self.upsampler.add_module('TANH_4',  torch.nn.ReLU())
-
-        self.output_conv = torch.nn.Conv2d(in_channels, 1, kernel_size=last_kernel, stride=1, padding=3) # for stft
-
-        # self.output_conv2 = torch.nn.Conv2d(1, 1, kernel_size=(7, 7), stride=1, padding=3)  # for stft
-
-        #self.output_conv = torch.nn.Conv2d(in_channels, 1, kernel_size=(19, 23), stride=1, padding=3)  # for qtrans
-        #self.output_conv = torch.nn.Conv2d(in_channels, 1, kernel_size=(19, 23), stride=1, padding=3)  # for qtrans magnitude
-
-        self.positive = torch.nn.ReLU()
-
-
-
-    def forward(self, x):
-        #print('input conv')
-        x = self.input_conv(x)
-        #print(x.shape)
-        #print('downsampler')
-        x = self.downsampler(x)
-        #print(x.shape)
-        #print('upsampler')
-        x = self.upsampler(x)
-        #print(x.shape)
-        #print('outut conv')
-        x = self.output_conv(x)
-        #print(x.shape)
-        # x = self.positive(x) ####-----------------
-        #x = self.output_conv2(x)
-        #x = self.positive(x)
-        # print(x.shape)
-        #x = x.view(x.shape[0], -1)
-        return x
-
-
-```
-