Help with get_top_covariances Course 2 Week 2?

Course Q&A Generative Adversarial Networks (GANS) Build Better Generative Adversarial Networks
1

I’ve been trying to complete the get_top_convariances function but have been running into some issues and could use some help.

My approach to this function is to start by creating the covariance matrix using the previously defined helper function. See here:

  cov = covariance_matrix_from_examples(classification_changes)

Once the covariance matrix is created, I want to remove the part of the covariance matrix that is the target_index’s covariance with itself, which I do here:

  relevant_indices = remove_from_list(cov, cov[target_index][target_index])

Finally I want to sort the relevant indices to return the highest_covariances using the helper function:

highest_covariances = cov[get_top_magnitude_indices(relevant_indices)[:top_n]]

But this keeps failing the first assertion. Can someone provide some guidance?

2

Hmmmm, I just compared my code to your description. I did those operations in the reverse order:

First use the get_top_magnitude_indices helper function.

Then use remove_from_list on that result.

Also note that I think you are misinterpreting the arguments that get passed to remove_from_list: they are index values, right?

1 Like
3

Uff i’m very confused, 'm trying to resolve this part but I can’t find a solution.

All help function function correctly, the particular test was good, but the get_top_convariances function it’s wrong

class_chg=classification_changes[:,:,-1]
cov_matrix=covariance_matrix_from_examples(classification_changes[:,:,-1])
get_top_magn=get_top_magnitude_indices(covariance_matrix_from_examples(classification_changes[:,:,-1]))[:top_n]

but the result :

foo- torch.Size([60, 128, 4]) tensor([[[-3.4697e-01, -6.9493e+00, -3.6928e+00, -6.8480e-01],
[ 3.4767e+00, 5.7545e+00, -2.0498e-01, -7.8584e-01],
[-1.1608e+00, 2.1348e+00, 5.7659e+00, 3.6523e+00],
…,
[ 6.1659e-01, -3.0746e+00, -5.8732e+00, -8.3355e+00],
[-4.1844e+00, 1.5068e+00, 1.7872e+00, 1.9598e+00],
[-1.9612e+00, -1.8472e+00, -2.5687e+00, -3.9748e-01]],
…,
[ 5.8083e+00, 2.9954e-01, -2.0871e+00, -2.6318e+00],
[-3.9213e+00, 5.0618e+00, 2.6743e-01, 4.0105e+00],
[-1.3784e+00, -1.3492e+00, -6.8733e+00, 5.4454e+00]]])

class_chg- torch.Size([60, 128]) tensor([[-6.8480e-01, -7.8584e-01, 3.6523e+00, …, -8.3355e+00,
1.9598e+00, -3.9748e-01],
[ 2.8420e-01, 5.2391e+00, 2.2153e+00, …, 6.3707e+00,
1.1093e+00, 2.4502e+00],
[ 4.8043e-01, 2.1717e+00, 2.7532e-01, …, 7.1343e-01,
-6.4532e+00, -4.3415e+00],
…,
[ 8.8360e+00, 4.5181e+00, -6.0043e-01, …, 9.0040e-01,
-1.3661e+00, 2.5932e+00],
[-1.9945e+00, -2.6196e+00, -4.0520e-03, …, 1.3709e+00,
1.7533e+00, -5.4681e-01],
[-1.3839e+00, -1.8083e+00, 5.7028e+00, …, -2.6318e+00,
4.0105e+00, 5.4454e+00]])

cov_matrix- (128, 128) [[10.31490342 -0.13934841 0.66691834 … 1.45728203 -2.49601633
-1.4768491 ]
[-0.13934841 12.93335031 -1.24922513 … 0.65039519 -0.61105151
-0.54250509]
[ 0.66691834 -1.24922513 8.0322546 … -0.66695757 0.06201199
-2.56580555]
…
[ 1.45728203 0.65039519 -0.66695757 … 11.35423392 0.92914036
-1.6231485 ]
[-2.49601633 -0.61105151 0.06201199 … 0.92914036 12.8996609
-0.18807426]
[-1.4768491 -0.54250509 -2.56580555 … -1.6231485 -0.18807426
12.94835014]]

get_top_magn- (3, 128) [[118 101 19 6 95 108 87 102 80 31 105 126 74 53 94 110 115 24
30 50 13 58 51 121 63 9 12 107 117 120 85 123 48 119 1 42
66 39 103 82 15 109 96 44 34 76 14 71 27 23 124 64 98 92
28 62 4 84 11 52 79 73 36 3 83 25 32 104 22 47 10 86
67 17 81 18 68 8 90 0 112 49 113 125 5 99 29 56 65 61
59 20 122 100 38 88 7 75 16 54 37 26 35 114 40 77 43 93
111 55 57 97 2 46 91 72 21 45 70 89 33 116 78 41 106 60
69 127]
[ 61 101 5 98 2 116 109 118 10 3 67 58 40 45 21 92 127 53
64 85 33 27 77 51 26 47 89 44 105 18 41 114 34 124 1 13
25 52 71 115 28 119 48 9 73 63 120 74 49 87 88 16 59 43
36 125 4 95 31 38 113 54 7 24 96 60 83 99 62 65 102 42
23 68 22 103 11 81 55 97 100 104 106 117 76 17 46 66 108 57
112 90 8 121 20 56 82 91 39 79 6 107 111 14 29 50 94 15
123 12 93 69 84 110 70 78 0 19 35 75 32 37 72 80 122 30
86 126]
[100 81 26 55 109 89 102 4 54 11 112 111 80 32 110 44 61 21
121 96 23 70 105 20 119 7 31 108 65 68 113 57 47 10 15 1
2 38 82 60 49 18 48 122 75 58 93 19 88 115 35 6 126 87
123 73 117 5 118 37 120 56 77 95 46 67 101 85 72 78 76 39
52 45 30 107 9 62 17 42 124 41 12 51 90 0 27 84 28 71
16 106 104 33 127 22 14 98 64 25 91 13 50 94 86 97 29 79
99 43 66 36 53 74 59 63 34 24 69 116 114 83 92 40 8 103
3 125]]

4

Hi, Fernando.

I compared your code to mine and mine looks different. :nerd_face: Why are you “slicing off” the last index of the last dimension of classification_changes when you pass that to covariance_matrix_from_examples? I did not need to do that in my code and it still worked. Then you don’t use the cov_matrix that you got as a result of that step as the input to the next step. This is a serial process, right? The output of each step is the input to the next one …

Please have another look at the instructions with the above in mind and see if that helps.

Regards,
Paul

5

Oh, sorry, maybe my point about the “serial” nature is covered in your code, but you could do it more efficiently: why make the call to covariance_matrix_from_examples twice?

6

yes it is true that I can do the process in series, but I do it step by step to see the result of each one.

however I still don’t see the second optional hints: “The target feature should not be included in the outputs.”

On the other hand I still don’t see why it worked

    class_chg=classification_changes
    print("class_chg-",np.shape(class_chg),class_chg)
    
    cov_matrix=covariance_matrix_from_examples(class_chg)
    print("cov_matrix-",np.shape(cov_matrix),cov_matrix)   

    get_top_magn=get_top_magnitude_indices(np.abs(cov_matrix)) 
    print("get_top_magn-",np.shape(get_top_magn),get_top_magn)
    
    relevant_indices = get_top_magn[:top_n]
    print('relevant_indices-',np.shape(relevant_indices),relevant_indices)
    
    highest_covariances = remove_from_list(relevant_indices, target_index)
    print('highest_covariances-',np.shape(highest_covariances),highest_covariances)

and the results are not close to what is expected:

class_chg- torch.Size([60, 128, 4]) tensor([[[ -1.8057, -1.2741, -0.0903, 4.6592],
[ 4.8889, 4.7921, -6.8099, 1.7678],
[ -4.7037, 1.9235, 3.7709, 3.8320],
…,
[ -4.1178, -0.3834, -1.7291, 4.1331],
[ -0.0168, 4.2047, 3.7771, 3.2863],
[ -2.6518, -1.6972, -3.1701, 2.8028]],

    [[  4.8896,   1.7211,   0.6857,  -1.5728],
     [ -5.0437,  -3.3760,  -0.2583,   1.0699],
     [ -5.5135,  -0.9380,   2.8819,   2.6553],
      ...,
     [  6.5315,   0.6002,  -1.9520,  -3.0223],
     [ -0.9957,   4.2713,   6.1071,   2.2693],
     [ -3.5694,  -4.2024,   0.0451,  -1.7943]]])

cov_matrix- (4, 4) [[ 9.9329278 1.92701195 -0.43278569 -4.94921648]
[ 1.92701195 10.888818 4.90974552 3.96154653]
[-0.43278569 4.90974552 10.03044214 1.87909212]
[-4.94921648 3.96154653 1.87909212 10.76941422]]
get_top_magn- (4, 4) [[2 1 0 3]
[0 3 1 2]
[0 3 2 1]
[2 1 3 0]]
relevant_indices- (3, 4) [[2 1 0 3]
[0 3 1 2]
[0 3 2 1]]
highest_covariances- (2, 4) [[2 1 0 3]
[0 3 2 1]]

7

With this version of the code, I think the problem is in how you call get_top_magnitude_indices. You are calling it on the complete covariance matrix. I think the intent is to select the row of the matrix corresponding to the target_index. I also did not take the absolute value of the inputs there, but am not sure whether that actually does any harm or not.

I have some similar prints in my code, but I used different names for things. Here’s my output for the first case in that test cell:

covary.shape = (4, 4)
covary =
[[10.20160607  2.21143831 -0.35353162 -5.11246264]
 [ 2.21143831 10.8387457   4.76233288  3.78990483]
 [-0.35353162  4.76233288  9.8154528   1.83621288]
 [-5.11246264  3.78990483  1.83621288 11.01953345]]
values [ 2.21143831 10.8387457   4.76233288  3.78990483]
top_indices = [1 2 3 0]
all_indices.shape = (4,)
relevant_indices.shape = (3,)
highest_cov.shape = (3,)
relevant_indices [2 3 0]
highest_covariances [4.76233288 3.78990483 2.21143831]

To interpret that note that all_indices is my name for what you called get_top_magn and that my version produces a (4,) 1D tensor, whereas yours gives a (4,4) result.

1 Like