How does the view method work in PyTorch

Question

I am confused about the method view   in the following code snippet   class Net nn Module       def   init   self           super Net  self    init             self conv1   nn Conv2d 3  6  5          self pool    nn MaxPool2d 2 2          self conv2   nn Conv2d 6  16  5          self fc1     nn Linear 16 5 5  120          self fc2     nn Linear 120  84          self fc3     nn Linear 84  10       def forward self  x           x   self pool F relu self conv1 x            x   self pool F relu self conv2 x            x   x view -1  16 5 5          x   F relu self fc1 x           x   F relu self fc2 x           x   self fc3 x          return x  net   Net     My confusion is regarding the following line   x   x view -1  16 5 5    What does tensor view   function do  I have seen its usage in many places  but I can t understand how it interprets its parameters    What happens if I give negative values as parameters to the view   function  For example  what happens if I call  tensor variable view 1  1  -1    Can anyone explain the main principle of view   function with some examples

User · Answer

Let s try to understand view by the following examples       a torch range 1 16   print a       tensor   1    2    3    4    5    6    7    8    9   10   11   12   13   14               15   16     print a view -1 2        tensor    1    2                  3    4                  5    6                  7    8                  9   10                 11   12                 13   14                 15   16      print a view 2 -1 4      3d tensor      tensor     1    2    3    4                   5    6    7    8                     9   10   11   12                  13   14   15   16      print a view 2 -1 2        tensor     1    2                   3    4                   5    6                   7    8                     9   10                  11   12                  13   14                  15   16       print a view 4 -1 2        tensor     1    2                   3    4                     5    6                   7    8                     9   10                  11   12                    13   14                  15   16        -1 as an argument value is an easy way to compute the value of say x provided we know values of y  z or the other way round in case of 3d and for 2d again an easy way to compute the value of say x provided we know values of y or vice versa

User · Answer

I figured it out that x view -1  16   5   5  is equivalent to x flatten 1   where the parameter 1 indicates the flatten process starts from the 1st dimension not flattening the  sample  dimension  As you can see  the latter usage is semantically more clear and easier to use  so I prefer flatten

User · Answer

Let s do some examples  from simpler to more difficult    The view method returns a tensor with the same data as the self tensor  which means that the returned tensor has the same number of elements   but with a different shape   For example   a   torch arange 1  17     a s shape is  16    a view 4  4    output below   1   2   3   4   5   6   7   8   9  10  11  12  13  14  15  16  torch FloatTensor of size 4x4   a view 2  2  4    output below  0          1   2   3   4 5   6   7   8   1           9  10  11  12 13  14  15  16  torch FloatTensor of size 2x2x4   Assuming that -1 is not one of the parameters  when you multiply them together  the result must be equal to the number of elements in the tensor   If you do  a view 3  3   it will raise a RuntimeError because shape  3 x 3  is invalid for input with 16 elements  In other words  3 x 3 does not equal 16 but 9  You can use -1 as one of the parameters that you pass to the function  but only once  All that happens is that the method will do the math for you on how to fill that dimension   For example a view 2  -1  4  is equivalent to a view 2  2  4    16    2 x 4    2  Notice that the returned tensor shares the same data  If you make a change in the  view  you are changing the original tensor s data   b   a view 4  4  b 0  2    2 a 2     3 0 False  Now  for a more complex use case   The documentation says that each new view dimension must either be a subspace of an original dimension  or only span d  d   1       d   k that satisfy the following contiguity-like condition that for all i   0       k - 1  stride i    stride i   1  x size i   1   Otherwise  contiguous   needs to be called before the tensor can be viewed   For example   a   torch rand 5  4  3  2    size  5  4  3  2  a t   a permute 0  2  3  1    size  5  3  2  4     The commented line below will raise a RuntimeError  because one dimension   spans across two contiguous subspaces   a t view -1  4     instead do  a t contiguous   view -1  4     To see why the first one does not work and the second does    compare a stride   and a t stride   a stride      24  6  2  1  a t stride      24  2  1  6    Notice that for a t  stride 0     stride 1  x size 1  since 24    2 x 3

User · Answer

I really liked  Jadiel de Armas examples   I would like to add a small insight to how elements are ordered for  view        For a Tensor with shape  a b c   the order of it s elements are determined by a numbering system  where the first digit has a numbers  second digit has b numbers and third digit has c numbers  The mapping of the elements in the new Tensor returned by  view      preserves this order of the original Tensor

User · Answer

weights reshape a  b  will return a new tensor with the same data as weights with size  a  b  as in it copies the data to another part of memory   weights resize  a  b  returns the same tensor with a different shape  However  if the new shape results in fewer elements than the original tensor  some elements will be removed from the tensor  but not from memory   If the new shape results in more elements than the original tensor  new elements will be uninitialized in memory   weights view a  b  will return a new tensor with the same data as weights with size  a  b

User · Answer

What is the meaning of parameter -1    You can read -1 as dynamic number of parameters or  anything   Because of that there can be only one parameter -1 in view     If you ask x view -1 1  this will output tensor shape  anything  1  depending on the number of elements in x  For example   import torch x   torch tensor  1  2  3  4   print x x shape  print        print x view -1 1   x view -1 1  shape  print x view 1 -1   x view 1 -1  shape    Will output   tensor  1  2  3  4   torch Size  4       tensor   1            2            3            4    torch Size  4  1   tensor   1  2  3  4    torch Size  1  4

User · Answer

torch Tensor view    Simply put  torch Tensor view   which is inspired by numpy ndarray reshape   or numpy reshape    creates a new view of the tensor  as long as the new shape is compatible with the shape of the original tensor    Let s understand this in detail using a concrete example   In  43   t   torch arange 18    In  44   t  Out 44    tensor   0   1   2   3   4   5   6   7   8   9  10  11  12  13  14  15  16  17     With this tensor t of shape  18    new views can only be created for the following shapes    1  18  or equivalently   1  -1  or  -1  18      2  9   or equivalently   2  -1  or  -1  9      3  6   or equivalently   3  -1  or  -1  6      6  3   or equivalently   6  -1  or  -1  3      9  2   or equivalently   9  -1  or  -1  2      18  1  or equivalently   18  -1  or  -1  1     As we can already observe from the above shape tuples  the multiplication of the elements of the shape tuple  e g  2 9  3 6 etc   must always be equal to the total number of elements in the original tensor  18 in our example     Another thing to observe is that we used a -1 in one of the places in each of the shape tuples  By using a -1  we are being lazy in doing the computation ourselves and rather delegate the task to PyTorch to do calculation of that value for the shape when it creates the new view  One important thing to note is that we can only use a single -1 in the shape tuple  The remaining values should be explicitly supplied by us  Else PyTorch will complain by throwing a RuntimeError      RuntimeError  only one dimension can be inferred   So  with all of the above mentioned shapes  PyTorch will always return a new view of the original tensor t  This basically means that it just changes the stride information of the tensor for each of the new views that are requested   Below are some examples illustrating how the strides of the tensors are changed with each new view     stride of our original tensor  t  In  53   t stride    Out 53    1     Now  we will see the strides for the new views     shape  1  18  In  54   t1   t view 1  -1    stride tensor  t1  with shape  1  18  In  55   t1 stride    Out 55    18  1     shape  2  9  In  56   t2   t view 2  -1    stride of tensor  t2  with shape  2  9  In  57   t2 stride          Out 57    9  1     shape  3  6  In  59   t3   t view 3  -1     stride of tensor  t3  with shape  3  6  In  60   t3 stride    Out 60    6  1     shape  6  3  In  62   t4   t view 6 -1    stride of tensor  t4  with shape  6  3  In  63   t4 stride    Out 63    3  1     shape  9  2  In  65   t5   t view 9  -1     stride of tensor  t5  with shape  9  2  In  66   t5 stride   Out 66    2  1     shape  18  1  In  68   t6   t view 18  -1    stride of tensor  t6  with shape  18  1  In  69   t6 stride   Out 69    1  1    So that s the magic of the view   function  It just changes the strides of the  original  tensor for each of the new views  as long as the shape of the new view is compatible with the original shape   Another interesting thing one might observe from the strides tuples is that the value of the element in the 0th position is equal to the value of the element in the 1st position of the shape tuple   In  74   t3 shape  Out 74   torch Size  3  6                             In  75   t3 stride        Out 75    6  1                                        This is because   In  76   t3  Out 76    tensor    0   1   2   3   4   5             6   7   8   9  10  11            12  13  14  15  16  17      the stride  6  1  says that to go from one element to the next element along the 0th dimension  we have to jump or take 6 steps   i e  to go from 0 to 6  one has to take 6 steps   But to go from one element to the next element in the 1st dimension  we just need only one step  for e g  to go from 2 to 3    Thus  the strides information is at the heart of how the elements are accessed from memory for performing the computation     torch reshape    This function would return a view and is exactly the same as using torch Tensor view   as long as the new shape is compatible with the shape of the original tensor  Otherwise  it will return a copy    However  the notes of torch reshape   warns that       contiguous inputs and inputs with compatible strides can be reshaped without copying  but one should not depend on the copying vs  viewing behavior

User · Answer

A tensor in pytorch is a view of an underlying contiguous block of numbers in memory  known as a storage   pytorch can achieve fast operations by modifying the shape parameters of a view of a storage without changing the underlying memory allocations themselves  Hence multiple different tensors may reference the same underlying storage object   view is a way of specifying a change of shape on an existing tensor

User · Answer

The view function is meant to reshape the tensor    Say you have a tensor  import torch a   torch range 1  16    a is a tensor that has 16 elements from 1 to 16 included   If you want to reshape this tensor to make it a 4 x 4 tensor then you can use   a   a view 4  4    Now a will be a 4 x 4 tensor  Note that after the reshape the total number of elements need to remain the same  Reshaping the tensor a to a 3 x 5 tensor would not be appropriate   What is the meaning of parameter -1   If there is any situation that you don t know how many rows you want but are sure of the number of columns  then you can specify this with a -1   Note that you can extend this to tensors with more dimensions  Only one of the axis value can be -1   This is a way of telling the library   give me a tensor that has these many columns and you compute the appropriate number of rows that is necessary to make this happen    This can be seen in the neural network code that you have given above  After the line x   self pool F relu self conv2 x    in the forward function  you will have a 16 depth feature map  You have to flatten this to give it to the fully connected layer  So you tell pytorch to reshape the tensor you obtained to have specific number of columns and tell it to decide the number of rows by itself   Drawing a similarity between numpy and pytorch  view is similar to numpy s reshape function

[python] How does the "view" method work in PyTorch?

Examples related to python

Examples related to memory

Examples related to pytorch

Examples related to torch

Examples related to tensor