update example

660eae7f · Dmytro Mishkin · 3995a4a0 · 660eae7f
Commit 660eae7f authored 2 years ago by Dmytro Mishkin
--- a/assignment_0_3_correspondences_template/imagefiltering.ipynb
+++ b/assignment_0_3_correspondences_template/imagefiltering.ipynb
@@ -550,7 +550,7 @@
    "print (f'A = {A}')\n",
    "patches = extract_antializased_affine_patches(timg1,\n",
    "                                 A, \n",
-    "                                 torch.zeros(A.size(0)).long(),\n",
+    "                                 torch.zeros(A.size(0)).long().reshape(-1,1),\n",
    "                                 32, 1.0)\n",
    "\n",
    "print (f'patches.shape = {patches.shape}')\n",
@@ -591,7 +591,7 @@
    "print (f'A = {A}')\n",
    "patches = extract_antializased_affine_patches(timg1,\n",
    "                                 A, \n",
-    "                                 torch.zeros(A.size(0)).long(),\n",
+    "                                 torch.zeros(A.size(0)).long().reshape(-1,1),\n",
    "                                 32, 1.0)\n",
    "\n",
    "print (f'patches.shape = {patches.shape}')\n",

 %% Cell type:markdown id: tags:
  
 # Lab 0: introduction into image filtering using PyTorch
 This is a notebook, which could help you with testing first lab assignment.
 It contains utility functions for visualization, some test input for the functions you needs to implement,
 and the output of the reference solution for the same test input.
  
 template functions for the assignment contain a short description of what the function is supposed to do,
 and produce an incorrect output, which is nevertheless in proper format: type and shape.
  
 You are not allowed to use kornia or opencv or any other library functions, which are specifically designed
 to perform the operations requested in assignment
  
 %% Cell type:code id: tags:
  
 ``` python
 %load_ext autoreload
 %autoreload 2
 import matplotlib.pyplot as plt
 import numpy as np
 import torch
 import kornia
  
  
 def plot_torch(x, y, *kwargs):
    plt.plot(x.detach().cpu().numpy(), y.detach().cpu().numpy(), *kwargs)
    return
  
 def imshow_torch(tensor, *kwargs):
    plt.figure()
    plt.imshow(kornia.tensor_to_image(tensor), *kwargs)
    return
 inp = torch.linspace(-12, 12, 101)
 ```
  
 %% Cell type:code id: tags:
  
 ``` python
 ```
  
 %% Cell type:code id: tags:
  
 ``` python
 from imagefiltering import gaussian_deriv1d
 plot_torch(inp, gaussian_deriv1d(inp, 3.0), 'r-')
 ```
  
 %% Output
  

  
 %% Cell type:markdown id: tags:
  
 ## Reference example
  
 ```python
 from lab0_reference.imagefiltering import gaussian_deriv1d
 plot_torch(inp, gaussian_deriv1d(inp, 3.0), 'r-')
 ```
  
 ![image.png](attachment:image.png)
  
 %% Cell type:code id: tags:
  
 ``` python
 ```
  
 %% Cell type:markdown id: tags:
  
 ## Convolution vs correlation
  
 Despite the name, pytorch `convolution` operator does in fact **correlation**. The correlation result is convolution, rotated by 180 degree.
 To distingguish the difference, remember that **convolution copies the kernel for the impulse signal**.
  
  
 %% Cell type:code id: tags:
  
 ``` python
 from imagefiltering import filter2d #, dgauss, gaussfilter, gaussderiv, gaussderiv2
 from torch.nn.functional import conv2d
 inp = torch.zeros((1,1,32,32))
 inp[...,16,16] = 1.
 imshow_torch(inp)
  
 kernel = torch.ones(3,3)
  
 out = filter2d(inp, kernel)
 imshow_torch(out)
  
 ### Signal-copy illustration.
 inp2 = torch.zeros((1,1,5,5))
 inp2[...,2,2] = 1.
 kernel2 = torch.arange(1,10).reshape(3,3).float()
 out2 = filter2d(inp2, kernel2)
  
 out2_corr = conv2d(inp2, kernel2[None, None], stride=1, padding=1)
  
 print (f'inp2 = {inp2}, \n\n kernel2={kernel2},\n \n true conv output2=\n {out2}')
 print (f'pytorch convolution aka correlation result =\n {out2_corr}')
  
 ```
  
 %% Output
  
    inp2 = tensor([[[[0., 0., 0., 0., 0.],
              [0., 0., 0., 0., 0.],
              [0., 0., 1., 0., 0.],
              [0., 0., 0., 0., 0.],
              [0., 0., 0., 0., 0.]]]]),
    
     kernel2=tensor([[1., 2., 3.],
            [4., 5., 6.],
            [7., 8., 9.]]),
    
     true conv output2=
     tensor([[[[0., 0., 0., 0., 0.],
              [0., 0., 0., 0., 0.],
              [0., 0., 1., 0., 0.],
              [0., 0., 0., 0., 0.],
              [0., 0., 0., 0., 0.]]]])
    pytorch convolution aka correlation result =
     tensor([[[[0., 0., 0., 0., 0.],
              [0., 9., 8., 7., 0.],
              [0., 6., 5., 4., 0.],
              [0., 3., 2., 1., 0.],
              [0., 0., 0., 0., 0.]]]])
  

  

  
 %% Cell type:code id: tags:
  
 ``` python
 ```
  
 %% Cell type:code id: tags:
  
 ``` python
 ```
  
 %% Cell type:markdown id: tags:
  
 ## Reference example
 ```python
 from imagefiltering import filter2d
 inp = torch.zeros((1,1,32,32))
 inp[...,16,16] = 1.
 imshow_torch(inp)
  
 kernel = torch.ones(3,3)
  
 out = filter2d(inp, kernel)
 imshow_torch(out)
 ```
  
    inp2 = tensor([[[[0., 0., 0., 0., 0.],
              [0., 0., 0., 0., 0.],
              [0., 0., 1., 0., 0.],
              [0., 0., 0., 0., 0.],
              [0., 0., 0., 0., 0.]]]]),
  
     kernel2=tensor([[1., 2., 3.],
            [4., 5., 6.],
            [7., 8., 9.]]),
  
     true conv output2=
     tensor([[[[0., 0., 0., 0., 0.],
              [0., 1., 2., 3., 0.],
              [0., 4., 5., 6., 0.],
              [0., 7., 8., 9., 0.],
              [0., 0., 0., 0., 0.]]]])
    pytorch convolution aka correlation result =
     tensor([[[[0., 0., 0., 0., 0.],
              [0., 9., 8., 7., 0.],
              [0., 6., 5., 4., 0.],
              [0., 3., 2., 1., 0.],
              [0., 0., 0., 0., 0.]]]])
 ![image.png](attachment:image.png)
  
 %% Cell type:code id: tags:
  
 ``` python
 from imagefiltering import gaussian_filter2d
 inp = torch.zeros((1,1,32,32))
 inp[...,15,15] = 1.
 imshow_torch(inp)
  
 sigma = 3.0
 out = gaussian_filter2d(inp, sigma)
 imshow_torch(out)
 ```
  
 %% Output
  

  

  
 %% Cell type:markdown id: tags:
  
 ## Reference example
  
 ```python
 from lab0_reference.imagefiltering import gaussian_filter2d
 inp = torch.zeros((1,1,32,32))
 inp[...,15,15] = 1.
 imshow_torch(inp)
  
 sigma = 3.0
 out = gaussian_filter2d(inp, sigma)
 imshow_torch(out)
 ```
 ![image.png](attachment:image.png)
  
 %% Cell type:code id: tags:
  
 ``` python
 ```
  
 %% Cell type:code id: tags:
  
 ``` python
 def imshow_torch_channels(tensor, dim = 1, *kwargs):
    num_ch = tensor.size(dim)
    fig=plt.figure(figsize=(num_ch*5,5))
    tensor_splitted = torch.split(tensor, 1, dim=dim)
    for i in range(num_ch):
        fig.add_subplot(1, num_ch, i+1)
        plt.imshow(kornia.tensor_to_image(tensor_splitted[i].squeeze(dim)), *kwargs)
    return
  
 from imagefiltering import spatial_gradient_first_order
 inp = torch.zeros((1,1,32,32))
 inp[...,15,15] = 1.
 imshow_torch(inp)
  
 sigma = 3.0
 out2 = spatial_gradient_first_order(inp, sigma)
 print (out.shape)
 imshow_torch_channels(out2, 2)
 ```
  
 %% Output
  
    torch.Size([1, 1, 32, 32])
  

  

  
 %% Cell type:markdown id: tags:
  
 ## Reference example
  
  
 ```python
 from lab0_reference.imagefiltering import spatial_gradient_first_order
 inp = torch.zeros((1,1,32,32))
 inp[...,15,15] = 1.
 imshow_torch(inp)
  
 sigma = 3.0
 out = spatial_gradient_first_order(inp, sigma)
 print (out.shape)
 imshow_torch_channels(out, 2)
 ```
 ![image-2.png](attachment:image-2.png)
  
 %% Cell type:markdown id: tags:
  
 ## Converting (center, unit_x, unit_y) into affine transform A
 ![image.png](attachment:image.png)
  
 %% Cell type:code id: tags:
  
 ``` python
 from imagefiltering import affine
 inp = torch.tensor([[3, 3.]]), torch.tensor([[6, 3.]]), torch.tensor([[3, 6.]])
 A = affine(*inp)
 print (A)
 ```
  
 %% Output
  
    tensor([[[1., 0., 0.],
             [0., 1., 0.],
             [0., 0., 1.]]])
  
 %% Cell type:markdown id: tags:
  
 ## Reference example
  
 ```python
 from lab0_reference.imagefiltering import affine
 inp = 3, 3, 6, 3, 3, 6
 A = affine(*inp)
 print (A)
 ```
  
    tensor([[3., 0., 3.],
            [0., 3., 3.],
            [0., 0., 1.]])
  
 %% Cell type:markdown id: tags:
  
 ## Affine patch extraction
 The y-axis here is flipped, so for the both image and patch we have it pointing down.
 ![image.png](attachment:image.png)
  
 %% Cell type:code id: tags:
  
 ``` python
 from imagefiltering import extract_affine_patches, extract_antializased_affine_patches, affine
 import cv2
  
  
 def load_torch_image_rgb(fname):
    img1 = cv2.imread(fname)
    timg1 = kornia.image_to_tensor(img1, False).float()
    timg1 = kornia.color.bgr_to_rgb(timg1) / 255.
    return timg1
  
 timg1 = load_torch_image_rgb('graffiti.ppm')
  
 imshow_torch(timg1)
  
 patch_centers = torch.tensor([[300., 200.], [400., 300], [330, 200], [100,100]])
 patch_unitx = torch.tensor([[350., 210.], [450., 300], [380, 200], [150,150]])
 patch_unity = torch.tensor([[270., 150.], [400., 360], [330,250], [50, 150]])
  
 A = affine(patch_centers, patch_unitx, patch_unity)
 print (f'A = {A}')
 patches = extract_antializased_affine_patches(timg1,
                                 A,
-                                 torch.zeros(A.size(0)).long(),
+                                 torch.zeros(A.size(0)).long().reshape(-1,1),
                                 32, 1.0)
  
 print (f'patches.shape = {patches.shape}')
 imshow_torch_channels(patches, 0)
 ```
  
 %% Output
  
    A = tensor([[[1., 0., 0.],
             [0., 1., 0.],
             [0., 0., 1.]],
    
            [[1., 0., 0.],
             [0., 1., 0.],
             [0., 0., 1.]],
    
            [[1., 0., 0.],
             [0., 1., 0.],
             [0., 0., 1.]],
    
            [[1., 0., 0.],
             [0., 1., 0.],
             [0., 0., 1.]]])
    patches.shape = torch.Size([4, 3, 32, 32])
  

  

  
 %% Cell type:markdown id: tags:
  
 ## Reference example
  
 ```python
 from imagefiltering import extract_affine_patches, extract_antializased_affine_patches, affine
 import cv2
  
  
 def load_torch_image_rgb(fname):
    img1 = cv2.imread(fname)
    timg1 = kornia.image_to_tensor(img1, False).float()
    timg1 = kornia.color.bgr_to_rgb(timg1) / 255.
    return timg1
  
 timg1 = load_torch_image_rgb('graffiti.ppm')
  
 imshow_torch(timg1)
  
 patch_centers = torch.tensor([[300., 200.], [400., 300], [600,600], [100,100]])
 patch_unitx = torch.tensor([[350., 210.], [450., 300], [650,600], [150,150]])
 patch_unity = torch.tensor([[270., 150.], [400., 360], [600,650], [50, 150]])
  
 A = affine(patch_centers, patch_unitx, patch_unity)
 print (f'A = {A}')
 patches = extract_antializased_affine_patches(timg1,
                                 A,
-                                 torch.zeros(A.size(0)).long(),
+                                 torch.zeros(A.size(0)).long().reshape(-1,1),
                                 32, 1.0)
  
 print (f'patches.shape = {patches.shape}')
 imshow_torch_channels(patches, 0)
  
 ```
  
    A = tensor([[[ 50., -30., 300.],
             [ 10., -50., 200.],
             [  0.,   0.,   1.]],
  
            [[ 50.,   0., 400.],
             [  0.,  60., 300.],
             [  0.,   0.,   1.]],
  
            [[ 50.,   0., 330.],
             [  0.,  50., 200.],
             [  0.,   0.,   1.]],
  
            [[ 50., -50., 100.],
             [ 50.,  50., 100.],
             [  0.,   0.,   1.]]])
    patches.shape = torch.Size([4, 3, 32, 32])
 ![image-4.png](attachment:image-4.png)
  
 %% Cell type:code id: tags:
  
 ``` python
 ```
  
 %% Cell type:code id: tags:
  
 ``` python
 ```