Python  - PyTorch - nn.Module                                     Home : www.sharetechnote.com

 

 

 

 

PyTorch - nn.conv2D

 

As the name implies, conv2D is the function to perform convolution to a 2D data (e.g, an image). If you are completely new to the concept of convolution and serious about understanding it from the very basic. I would suggest you to start with 1 D convolution in my note here.  If you are already familiar with the basic concept of convolution or not interested in the basic concept and want to directly jump into 2D convolution which is more relevant to 2D image processing, take a look at some examples in Conv 2D section in my visual note (www.slide4math.com). Here goes one of the examples.

 

Conv2D function can perform the convolution to a much more complicated data structure (like to a set of 2D data, not only to single 2D data) and with some additional options like padding and stride.

 

What I want to you do in this section is not about performing the math of the convolution but a data structure and calculation flow performed by the Conv2D function.

 

The Generic usage of Conv2D() function is as follows.

    torch.nn.Conv2d(in_channels = n, out_channels = m, kernel_size = k, stride = s, padding = p)

The operation performed by this function can be illustrated as show below. In most case, the a and b is set to be the same. the c and d is usally set to be same as well.

One important thing that you might have noticed from Conv2D() usage would be that we only set the dimension of the filter (kernel), does not specify the number of the filters nor the values within the filters. As hinted by the diagram shown below, the number of the filter is automatically determined by the number of input data and number of output data. The initial values of each filter is automatically assigned by Conv2D() function (usually random values are assigned).

 

 

 

Now I will show you various examples for various data/calculation flow and corresponding source Pytorch source code. Going through these examples, I hope you intuitively understand the usage of Conv2D() function.

 

 

 

Input layer 1, output layer 1, kernel size 3x3, stride 1, padding 0

 

 

 

    import torch

     

    input1 = torch.ones(1,1,6,6)

    print("input = ",input1)

    ==> input =  tensor([[[[1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.]]]])

     

     

    net = torch.nn.Conv2d(in_channels = 1, out_channels = 1, kernel_size = 3)

     

    print("net = ",net)

    ==> net =  Conv2d(1, 1, kernel_size=(3, 3), stride=(1, 1))

     

     

    print("Parameters = ",list(net.parameters()))

    ==> Parameters =  [Parameter containing:

                                          tensor([[[[ 0.0881, -0.1189, -0.0778],

                                                       [ 0.0953,  0.0934,  0.0858],

                                                       [-0.2734,  0.1937,  0.0823]]]], requires_grad=True),

                               Parameter containing:

                                          tensor([0.1680], requires_grad=True)]

     

     

    print("Weight = ",net.weight)

    print("bias = ",net.bias)

    ==> Weight =  Parameter containing:

                                 tensor([[[[ 0.0881, -0.1189, -0.0778],

                                              [ 0.0953,  0.0934,  0.0858],

                                              [-0.2734,  0.1937,  0.0823]]]], requires_grad=True)

           bias =  Parameter containing:

                                 tensor([0.1680], requires_grad=True)

     

     

    out = net(input1)

    print("output = ",out)

    ==> output =  tensor([[[[0.3365, 0.3365, 0.3365, 0.3365],

                                     [0.3365, 0.3365, 0.3365, 0.3365],

                                     [0.3365, 0.3365, 0.3365, 0.3365],

                                     [0.3365, 0.3365, 0.3365, 0.3365]]]], grad_fn=<ThnnConv2DBackward>)  

 

 

 

Input layer 1, output layer 1, kernel size 3x3, stride 2, padding 0

 

 

 

 

    import torch

     

    input1 = torch.ones(1,1,6,6)

    print("input = ",input1)

    ==> input =  tensor([[[[1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.]]]])

     

     

    net = torch.nn.Conv2d(in_channels = 1, out_channels = 1, kernel_size = 3, stride = 2)

     

    print("net = ",net)

    ==> net =  Conv2d(1, 1, kernel_size=(3, 3), stride=(2, 2))

     

     

    print("Parameters = ",list(net.parameters()))

    ==> Parameters =  [Parameter containing:

                                        tensor([[[[ 0.1907,  0.2069,  0.1676],

                                                     [ 0.0233, -0.2632,  0.1108],

                                                     [-0.0301,  0.0857,  0.2350]]]], requires_grad=True),

                               Parameter containing:

                                        tensor([-0.0093], requires_grad=True)]

     

     

    print("Weight = ",net.weight)

    print("bias = ",net.bias)

    ==> Weight =  Parameter containing:

                                 tensor([[[[ 0.1907,  0.2069,  0.1676],

                                              [ 0.0233, -0.2632,  0.1108],

                                              [-0.0301,  0.0857,  0.2350]]]], requires_grad=True)

           bias =  Parameter containing:

                                 tensor([-0.0093], requires_grad=True)

     

     

    out = net(input1)

    print("output = ",out)

    ==> output =  tensor([[[[0.7175, 0.7175],

                                     [0.7175, 0.7175]]]], grad_fn=<ThnnConv2DBackward>)  

 

 

 

Input layer 1, output layer 1, kernel size 3x3, stride 2, padding 1

 

 

 

 

    import torch

     

    input1 = torch.ones(1,1,6,6)

    print("input = ",input1)

    ==> input =  tensor([[[[1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.]]]])

     

     

    net = torch.nn.Conv2d(in_channels = 1, out_channels = 1, kernel_size = 3, stride = 2, padding = 1)

     

    print("net = ",net)

    ==> net =  Conv2d(1, 1, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))

     

     

    print("Parameters = ",list(net.parameters()))

    ==> Parameters =  [Parameter containing:

                                    tensor([[[[-0.2392, -0.0492, -0.0347],

                                                 [-0.3049, -0.1630,  0.1242],

                                                 [-0.2988, -0.3229,  0.1064]]]], requires_grad=True),

                               Parameter containing:

                                    tensor([-0.0967], requires_grad=True)]

     

     

    print("Weight = ",net.weight)

    print("bias = ",net.bias)

    ==> Weight =  Parameter containing:

                                  tensor([[[[-0.2392, -0.0492, -0.0347],

                                               [-0.3049, -0.1630,  0.1242],

                                               [-0.2988, -0.3229,  0.1064]]]], requires_grad=True)

           bias =  Parameter containing:

                                  tensor([-0.0967], requires_grad=True)

     

     

    out = net(input1)

    print("output = ",out)

    ==> output =  tensor([[[[-0.3520, -0.9556, -0.9556],

                                     [-0.4359, -1.2787, -1.2787],

                                     [-0.4359, -1.2787, -1.2787]]]], grad_fn=<ThnnConv2DBackward>)

 

 

 

Input layer 3, output layer 3, kernel size 3x3, stride 1, padding 0

 

 

 

 

    import torch

     

    input1 = torch.ones(1,3,6,6)

    print("input = ",input1)

    ==> input =  tensor([[[[1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.]],

     

                                  [[1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.]],

     

                                  [[1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.]]]])

     

     

    net = torch.nn.Conv2d(in_channels = 3, out_channels = 3, kernel_size = 3, stride = 1)

     

    print("net = ",net)

    ==> net =  Conv2d(3, 3, kernel_size=(3, 3), stride=(1, 1))

     

     

    print("Parameters = ",list(net.parameters()))

    ==> Parameters =  [Parameters =  [Parameter containing:

                                                       tensor([[[[-0.0803, -0.1713,  0.1441],

                                                                    [-0.1072, -0.1571, -0.0254],

                                                                    [-0.1230, -0.0584,  0.0903]],

     

                                                                   [[-0.1486,  0.1374, -0.1192],

                                                                    [ 0.0228,  0.0543, -0.1124],

                                                                    [-0.1045,  0.1804, -0.0721]],

     

                                                                   [[-0.0944,  0.1255,  0.0371],

                                                                    [-0.0887, -0.1317,  0.1474],

                                                                    [ 0.0237,  0.1048,  0.0840]]],

     

     

                                                                 [[[-0.1206, -0.1360,  0.0185],

                                                                    [-0.0684,  0.1522,  0.1608],

                                                                    [ 0.0973,  0.0807,  0.1193]],

     

                                                                   [[ 0.1036, -0.0177,  0.1745],

                                                                    [-0.1605,  0.0437, -0.1423],

                                                                    [-0.0322,  0.0826, -0.1443]],

     

                                                                   [[ 0.1145, -0.1378, -0.1148],

                                                                    [-0.0828, -0.1226, -0.0900],

                                                                    [ 0.1138,  0.1260,  0.0788]]],

     

     

                                                                 [[[-0.1372, -0.0510,  0.1307],

                                                                   [ 0.1600,  0.0902,  0.0489],

                                                                   [-0.0889,  0.1738, -0.0099]],

     

                                                                  [[-0.0494, -0.0856,  0.1392],

                                                                   [-0.1584,  0.0696, -0.1846],

                                                                   [-0.1266, -0.1801,  0.0202]],

     

                                                                  [[ 0.0151,  0.1716, -0.1645],

                                                                   [-0.0296,  0.1748, -0.0985],

                                                                   [-0.1260, -0.1463,  0.0970]]]], requires_grad=True),

                                Parameter containing:

                                         tensor([0.0566, 0.0760, 0.0430], requires_grad=True)]

     

     

    print("Weight = ",net.weight)

    print("bias = ",net.bias)

    ==> Weight =  Parameter containing:

                                                       tensor([[[[-0.0803, -0.1713,  0.1441],

                                                                    [-0.1072, -0.1571, -0.0254],

                                                                    [-0.1230, -0.0584,  0.0903]],

     

                                                                   [[-0.1486,  0.1374, -0.1192],

                                                                    [ 0.0228,  0.0543, -0.1124],

                                                                    [-0.1045,  0.1804, -0.0721]],

     

                                                                   [[-0.0944,  0.1255,  0.0371],

                                                                    [-0.0887, -0.1317,  0.1474],

                                                                    [ 0.0237,  0.1048,  0.0840]]],

     

     

                                                                 [[[-0.1206, -0.1360,  0.0185],

                                                                    [-0.0684,  0.1522,  0.1608],

                                                                    [ 0.0973,  0.0807,  0.1193]],

     

                                                                   [[ 0.1036, -0.0177,  0.1745],

                                                                    [-0.1605,  0.0437, -0.1423],

                                                                    [-0.0322,  0.0826, -0.1443]],

     

                                                                   [[ 0.1145, -0.1378, -0.1148],

                                                                    [-0.0828, -0.1226, -0.0900],

                                                                    [ 0.1138,  0.1260,  0.0788]]],

     

     

                                                                 [[[-0.1372, -0.0510,  0.1307],

                                                                   [ 0.1600,  0.0902,  0.0489],

                                                                   [-0.0889,  0.1738, -0.0099]],

     

                                                                  [[-0.0494, -0.0856,  0.1392],

                                                                   [-0.1584,  0.0696, -0.1846],

                                                                   [-0.1266, -0.1801,  0.0202]],

     

                                                                  [[ 0.0151,  0.1716, -0.1645],

                                                                   [-0.0296,  0.1748, -0.0985],

                                                                   [-0.1260, -0.1463,  0.0970]]]], requires_grad=True),

             bias =  Parameter containing:

                                         tensor([0.0566, 0.0760, 0.0430], requires_grad=True)]

     

     

    out = net(input1)

    print("output = ",out)

    ==> output =  tensor([[[[-0.3861, -0.3861, -0.3861, -0.3861],

                                     [-0.3861, -0.3861, -0.3861, -0.3861],

                                     [-0.3861, -0.3861, -0.3861, -0.3861],

                                     [-0.3861, -0.3861, -0.3861, -0.3861]],

     

                                    [[ 0.1724,  0.1724,  0.1724,  0.1724],

                                     [ 0.1724,  0.1724,  0.1724,  0.1724],

                                     [ 0.1724,  0.1724,  0.1724,  0.1724],

                                     [ 0.1724,  0.1724,  0.1724,  0.1724]],

     

                                    [[-0.3028, -0.3028, -0.3028, -0.3028],

                                     [-0.3028, -0.3028, -0.3028, -0.3028],

                                     [-0.3028, -0.3028, -0.3028, -0.3028],

                                     [-0.3028, -0.3028, -0.3028, -0.3028]]]],

                                    grad_fn=<ThnnConv2DBackward>)

 

 

 

Input layer 1, output layer 3, kernel size 3x3, stride 1, padding 0

 

 

 

    import torch

     

    input1 = torch.ones(1,1,6,6)

    print("input = ",input1)

    ==> input =  tensor([[[[1., 1., 1., 1., 1., 1.],

                                  [1., 1., 1., 1., 1., 1.],

                                  [1., 1., 1., 1., 1., 1.],

                                  [1., 1., 1., 1., 1., 1.],

                                  [1., 1., 1., 1., 1., 1.],

                                  [1., 1., 1., 1., 1., 1.]]]])

     

     

    net = torch.nn.Conv2d(in_channels = 1, out_channels = 3, kernel_size = 3, stride = 1)

     

    print("net = ",net)

    ==> net =  Conv2d(1, 3, kernel_size=(3, 3), stride=(1, 1))  

     

     

    print("Parameters = ",list(net.parameters()))

    ==> Parameters =  [Parameter containing:

                                      tensor([[[[-0.2515, -0.1372,  0.3204],

                                                   [-0.0997,  0.3118,  0.2565],

                                                   [ 0.0022,  0.0423, -0.0538]]],

     

     

                                                [[[-0.2974,  0.3017, -0.0427],

                                                   [-0.0471,  0.0506,  0.2380],

                                                   [-0.0384, -0.1994, -0.0341]]],

     

     

                                                [[[-0.1916, -0.1838,  0.1763],

                                                   [-0.3117,  0.1773, -0.0955],

                                                   [-0.1647,  0.1660,  0.1878]]]], requires_grad=True),

                                Parameter containing:

                                       tensor([-0.0194,  0.2818, -0.1526], requires_grad=True)]

     

     

    print("Weight = ",net.weight)

    print("bias = ",net.bias)

    ==> Weight =  Parameter containing:

                                  tensor([[[[-0.2515, -0.1372,  0.3204],

                                               [-0.0997,  0.3118,  0.2565],

                                               [ 0.0022,  0.0423, -0.0538]]],

     

     

                                             [[[-0.2974,  0.3017, -0.0427],

                                                [-0.0471,  0.0506,  0.2380],

                                                [-0.0384, -0.1994, -0.0341]]],

     

     

                                              [[[-0.1916, -0.1838,  0.1763],

                                                 [-0.3117,  0.1773, -0.0955],

                                                 [-0.1647,  0.1660,  0.1878]]]], requires_grad=True)

             bias =  Parameter containing:

                                     tensor([-0.0194,  0.2818, -0.1526], requires_grad=True)

     

     

    out = net(input1)

    print("output = ",out)

    ==> output =  tensor([[[[ 0.3715,  0.3715,  0.3715,  0.3715],

                                     [ 0.3715,  0.3715,  0.3715,  0.3715],

                                     [ 0.3715,  0.3715,  0.3715,  0.3715],

                                     [ 0.3715,  0.3715,  0.3715,  0.3715]],

     

                                    [[ 0.2131,  0.2131,  0.2131,  0.2131],

                                     [ 0.2131,  0.2131,  0.2131,  0.2131],

                                     [ 0.2131,  0.2131,  0.2131,  0.2131],

                                     [ 0.2131,  0.2131,  0.2131,  0.2131]],

     

                                    [[-0.3923, -0.3923, -0.3923, -0.3923],

                                     [-0.3923, -0.3923, -0.3923, -0.3923],

                                     [-0.3923, -0.3923, -0.3923, -0.3923],

                                     [-0.3923, -0.3923, -0.3923, -0.3923]]]],

                                     grad_fn=<ThnnConv2DBackward>)

 

 

 

Input layer 3, output layer 1, kernel size 3x3, stride 1, padding 0

 

 

 

 

    import torch

     

    input1 = torch.ones(1,3,6,6)

    print("input = ",input1)

    ==> input =  tensor([[[[1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.]],

     

                                  [[1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.]],

     

                                  [[1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.]]]])

     

     

     

    net = torch.nn.Conv2d(in_channels = 3, out_channels = 1, kernel_size = 3, stride = 1)

     

    print("net = ",net)

    ==> net =  Conv2d(3, 1, kernel_size=(3, 3), stride=(1, 1))  

     

     

    print("Parameters = ",list(net.parameters()))

    ==> Parameters =  [Parameter containing:

                                      tensor([[[[-0.0912, -0.1064, -0.1396],

                                                   [ 0.0072, -0.1196,  0.1672],

                                                   [-0.0177, -0.0890,  0.0466]],

     

                                                  [[ 0.0092,  0.1629, -0.1772],

                                                   [-0.1293, -0.1035,  0.0919],

                                                   [-0.1625, -0.0141,  0.0412]],

     

                                                  [[ 0.0712, -0.1737,  0.0552],

                                                   [ 0.0434,  0.1417, -0.1580],

                                                   [ 0.1569,  0.1290,  0.1705]]]], requires_grad=True),

                               Parameter containing:

                                       tensor([-0.0269], requires_grad=True)]

     

     

    print("Weight = ",net.weight)

    print("bias = ",net.bias)

    ==> Weight =  Parameter containing:

                                tensor([[[[-0.0912, -0.1064, -0.1396],

                                             [ 0.0072, -0.1196,  0.1672],

                                             [-0.0177, -0.0890,  0.0466]],

     

                                            [[ 0.0092,  0.1629, -0.1772],

                                             [-0.1293, -0.1035,  0.0919],

                                             [-0.1625, -0.0141,  0.0412]],

     

                                            [[ 0.0712, -0.1737,  0.0552],

                                             [ 0.0434,  0.1417, -0.1580],

                                             [ 0.1569,  0.1290,  0.1705]]]], requires_grad=True)

              bias =  Parameter containing:

                                 tensor([-0.0269], requires_grad=True)

     

     

    out = net(input1)

    print("output = ",out)

    ==> output =  tensor([[[[-0.2145, -0.2145, -0.2145, -0.2145],

                                     [-0.2145, -0.2145, -0.2145, -0.2145],

                                     [-0.2145, -0.2145, -0.2145, -0.2145],

                                     [-0.2145, -0.2145, -0.2145, -0.2145]]]],

                                     grad_fn=<ThnnConv2DBackward>)

 

 

 

Input layer 2, output layer 4, kernel size 3x3, stride 1, padding 0

 

 

 

 

    import torch

     

    input1 = torch.ones(1,2,6,6)

    print("input = ",input1)

    ==> input =  tensor([[[[1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.]],

     

                                  [[1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.],

                                   [1., 1., 1., 1., 1., 1.]]]])

     

     

     

    net = torch.nn.Conv2d(in_channels = 2, out_channels = 4, kernel_size = 3, stride = 1)

     

    print("net = ",net)

    ==> net =  Conv2d(2, 4, kernel_size=(3, 3), stride=(1, 1))  

     

     

    print("Parameters = ",list(net.parameters()))

    ==> Parameters =  [Parameter containing:

                                    tensor([[[[ 0.2179, -0.1332,  0.0899],

                                                 [-0.0964, -0.1653,  0.1551],

                                                 [-0.1422, -0.2231,  0.1075]],

     

                                                [[-0.1066, -0.1183,  0.0421],

                                                 [ 0.1305,  0.2184, -0.1633],

                                                 [-0.1271,  0.2284,  0.2235]]],

     

     

                                              [[[ 0.1970, -0.0710, -0.0590],

                                                [-0.1749,  0.0487,  0.1591],

                                                [-0.1202,  0.0690,  0.1691]],

     

                                               [[ 0.0198, -0.0896,  0.2124],

                                                [-0.0867, -0.0135, -0.1714],

                                                [-0.0533,  0.1503, -0.2194]]],

     

     

                                              [[[-0.1208, -0.1256, -0.0556],

                                                [ 0.0442,  0.2287, -0.2230],

                                                [-0.0903,  0.1225, -0.1689]],

     

                                               [[-0.1276, -0.2175,  0.1301],

                                                [-0.0630, -0.0887,  0.0780],

                                                [ 0.0101,  0.1145,  0.0791]]],

     

     

                                             [[[ 0.0135,  0.0196,  0.1861],

                                               [ 0.1848,  0.2056, -0.1715],

                                               [ 0.1584, -0.2033,  0.0102]],

     

                                              [[ 0.2010,  0.1128, -0.0148],

                                               [ 0.2009,  0.2132,  0.0760],

                                               [ 0.1043,  0.1950,  0.1885]]]], requires_grad=True),

          Parameter containing:

                          tensor([-0.2167, -0.0877,  0.1046,  0.1399], requires_grad=True)]

     

     

    print("Weight = ",net.weight)

    print("bias = ",net.bias)

    ==> Weight =  [Parameter containing:

                                    tensor([[[[ 0.2179, -0.1332,  0.0899],

                                                 [-0.0964, -0.1653,  0.1551],

                                                 [-0.1422, -0.2231,  0.1075]],

     

                                                [[-0.1066, -0.1183,  0.0421],

                                                 [ 0.1305,  0.2184, -0.1633],

                                                 [-0.1271,  0.2284,  0.2235]]],

     

     

                                              [[[ 0.1970, -0.0710, -0.0590],

                                                [-0.1749,  0.0487,  0.1591],

                                                [-0.1202,  0.0690,  0.1691]],

     

                                               [[ 0.0198, -0.0896,  0.2124],

                                                [-0.0867, -0.0135, -0.1714],

                                                [-0.0533,  0.1503, -0.2194]]],

     

     

                                              [[[-0.1208, -0.1256, -0.0556],

                                                [ 0.0442,  0.2287, -0.2230],

                                                [-0.0903,  0.1225, -0.1689]],

     

                                               [[-0.1276, -0.2175,  0.1301],

                                                [-0.0630, -0.0887,  0.0780],

                                                [ 0.0101,  0.1145,  0.0791]]],

     

     

                                             [[[ 0.0135,  0.0196,  0.1861],

                                               [ 0.1848,  0.2056, -0.1715],

                                               [ 0.1584, -0.2033,  0.0102]],

     

                                              [[ 0.2010,  0.1128, -0.0148],

                                               [ 0.2009,  0.2132,  0.0760],

                                               [ 0.1043,  0.1950,  0.1885]]]], requires_grad=True),

          bias = Parameter containing:

                          tensor([-0.2167, -0.0877,  0.1046,  0.1399], requires_grad=True)]

     

     

    out = net(input1)

    print("output = ",out)

    ==> output =  tensor([[[[-0.0788, -0.0788, -0.0788, -0.0788],

                                     [-0.0788, -0.0788, -0.0788, -0.0788],

                                     [-0.0788, -0.0788, -0.0788, -0.0788],

                                     [-0.0788, -0.0788, -0.0788, -0.0788]],

     

                                    [[-0.1214, -0.1214, -0.1214, -0.1214],

                                     [-0.1214, -0.1214, -0.1214, -0.1214],

                                     [-0.1214, -0.1214, -0.1214, -0.1214],

                                     [-0.1214, -0.1214, -0.1214, -0.1214]],

     

                                   [[-0.3692, -0.3692, -0.3692, -0.3692],

                                    [-0.3692, -0.3692, -0.3692, -0.3692],

                                    [-0.3692, -0.3692, -0.3692, -0.3692],

                                    [-0.3692, -0.3692, -0.3692, -0.3692]],

     

                                   [[ 1.8202,  1.8202,  1.8202,  1.8202],

                                    [ 1.8202,  1.8202,  1.8202,  1.8202],

                                    [ 1.8202,  1.8202,  1.8202,  1.8202],

                                    [ 1.8202,  1.8202,  1.8202,  1.8202]]]],

                                    grad_fn=<ThnnConv2DBackward>)

 

 

 

Input layer 3, output layer 1, kernel size 3x3, stride 1, padding 0 wit input image

 

    import torch

    import torchvision.transforms.functional as TF

    import PIL

    from PIL import Image

    from matplotlib import pyplot

    from numpy import asarray

     

     

    img = Image.open('temp/digit/0/digit0.png')

    input = TF.to_tensor(img);

    input = input.unsqueeze_(0);

     

    print("input.shape = ",input.shape)

    ==> input.shape =  torch.Size([1, 3, 10, 10])

     

    net = torch.nn.Conv2d(in_channels = 3, out_channels = 1, kernel_size = 3)

     

    print("net = ",net)

     ==> net =  Conv2d(3, 1, kernel_size=(3, 3), stride=(1, 1))

     

    print("Weight = ",net.weight)

    print("bias = ",net.bias)

    ==> Weight =  Parameter containing:

                                 tensor([[[[-0.0482,  0.1148, -0.1228],

                                              [-0.1687,  0.0758,  0.0497],

                                              [ 0.0313, -0.1646, -0.1486]],

     

                                            [[ 0.0562,  0.1516,  0.1300],

                                             [ 0.0879, -0.0339, -0.1876],

                                             [-0.0335,  0.1665,  0.0831]],

     

                                           [[-0.0947, -0.1731, -0.0320],

                                            [ 0.0810,  0.0433, -0.1137],

                                            [-0.0629, -0.1213, -0.0357]]]], requires_grad=True)

    bias =  Parameter containing:

                      tensor([0.1899], requires_grad=True)

     

     

    out = net(input)

    print("output = ",out)

    ==> output =  tensor([[[[-0.2825,  0.0074,  0.0993, -0.1404, -0.2211, -0.2096, -0.0612,  -0.2893],

                                     [-0.1458, -0.0563, -0.2432, -0.2315,  0.0343, -0.2591, -0.2193, -0.2048],

                                     [-0.1072, -0.0550, -0.3664, -0.1790, -0.1037, -0.0855, -0.3280, -0.1874],

                                     [-0.0702, -0.1030, -0.3483, -0.2005, -0.1603, -0.0146, -0.3481, -0.1741],

                                     [-0.1308, -0.0748, -0.2734, -0.1411, -0.0867, -0.0083, -0.3704, -0.1826],

                                     [-0.1727, -0.0352, -0.3164,  0.0981,  0.2282, -0.0866, -0.4669,-0.1779],

                                     [-0.2981,  0.0287, -0.1716, -0.1647, -0.2899, -0.2744, -0.2395,-0.2673],

                                     [-0.2801, -0.2910, -0.2061, -0.2739, -0.2440, -0.3237, -0.2540,-0.2801]]]],

                                    grad_fn=<ThnnConv2DBackward>)

 

 

 

Reference :

 

[1] Convolutional Neural Networks (CNNs / ConvNets)

[2] CONVOLUTIONAL NEURAL NETWORKS IN PYTORCH