Serious bug with type casting using ch.Tensor

[markroxor]

* **I'm submitting a ...** - [x] bug report - [ ] feature request

• How do we replicate the issue?

  
  ############
  import torch as ch
  from sklearn.decomposition import PCA
  import numpy as np
  Xtrain = np.zeros([n, n_dim])
  if n_dim==3:
  m = 5
  c = 2
  d = 5
  for i in range(n):
      Xtrain[i,0] = i
      Xtrain[i,1] = m*i + c
  #         X[i,1] = np.sqrt(i)
      Xtrain[i,2] = (d - 2*Xtrain[i,0] - 3*Xtrain[i,1])/2

Xtrain = ch.Tensor(Xtrain)

Xtrain = ch.from_numpy(Xtrain)

Xtrain = Xtrain.numpy()

pca = PCA(n_components=3) pca.fit(Xtrain) U, S, VT = np.linalg.svd(Xtrain - Xtrain.mean(0)) print(VT)

[[ 0.10088665 0.50443327 -0.85753656] [ 0.82430398 0.44025326 0.35594945] [ 0.55708601 -0.74278135 -0.37139068]]

print(pca.components_)

[[-1.0088666e-01 -5.0443327e-01 8.5753661e-01] [-5.8938619e-02 8.6345071e-01 5.0097811e-01] [-9.9315059e-01 -7.4505806e-09 -1.1684125e-01]]

• Expected behavior (i.e. solution)

  
  import torch as ch
  from sklearn.decomposition import PCA
  import numpy as np
  ############
  Xtrain = np.zeros([n, n_dim])
  if n_dim==3:
  m = 5
  c = 2
  d = 5
  for i in range(n):
      Xtrain[i,0] = i
      Xtrain[i,1] = m*i + c
  #         X[i,1] = np.sqrt(i)
      Xtrain[i,2] = (d - 2*Xtrain[i,0] - 3*Xtrain[i,1])/2

Xtrain = ch.Tensor(Xtrain)

Xtrain = ch.from_numpy(Xtrain) Xtrain = Xtrain.numpy()

pca = PCA(n_components=3) pca.fit(Xtrain) U, S, VT = np.linalg.svd(Xtrain - Xtrain.mean(0)) print(VT)

[[ 0.10088666 0.5044333 -0.85753655] [ 0.824304 0.44025326 0.35594946] [ 0.557086 -0.74278134 -0.37139067]] print(pca.components_) [[-0.10088665 -0.50443327 0.85753656] [-0.82430398 -0.44025326 -0.35594945] [-0.55708601 0.74278135 0.37139068]]

• Please tell us about your environment:

  • Description:    Ubuntu 18.04 LTS
    Release:    18.04
    Codename:   bionic

  • python 2.7

[markroxor]

import torch as ch
from sklearn.decomposition import PCA
import numpy as np
n = 5
n_dim = 3

X = ch.zeros(n, n_dim)
m = 5
c = 2
d = 5
for i in range(n):
    X[i,0] = i
    X[i,1] = m*i + c
    X[i,2] = (d - 2*X[i,0] - 3*X[i,1])/2

X = X.numpy()

Xtrain = np.zeros([n, n_dim])
m = 5
c = 2
d = 5
for i in range(n):
    Xtrain[i,0] = i
    Xtrain[i,1] = m*i + c
    Xtrain[i,2] = (d - 2*Xtrain[i,0] - 3*Xtrain[i,1])/2

print(X, type(X))
>>> (array([[  0. ,   2. ,  -0.5],
       [  1. ,   7. ,  -9. ],
       [  2. ,  12. , -17.5],
       [  3. ,  17. , -26. ],
       [  4. ,  22. , -34.5]], dtype=float32), <type 'numpy.ndarray'>)
print(Xtrain, type(Xtrain))
>>> (array([[  0. ,   2. ,  -0.5],
       [  1. ,   7. ,  -9. ],
       [  2. ,  12. , -17.5],
       [  3. ,  17. , -26. ],
       [  4. ,  22. , -34.5]]), <type 'numpy.ndarray'>)
print(np.allclose(Xtrain, X))
>>> True
pca_x = PCA(n_components=3)
pca_x.fit(X)

pca_xtrain = PCA(n_components=3)
pca_xtrain.fit(Xtrain)

print(pca_x.components_)
>>> [[-1.0088666e-01 -5.0443327e-01  8.5753661e-01]
 [-5.8938619e-02  8.6345071e-01  5.0097811e-01]
 [-9.9315059e-01 -7.4505806e-09 -1.1684125e-01]]
print(pca_xtrain.components_)
>>> [[-0.10088665 -0.50443327  0.85753656]
 [-0.82430398 -0.44025326 -0.35594945]
 [-0.55708601  0.74278135  0.37139068]]

[markroxor]

possibly because the dtype is float64 and float32

[markroxor]

opened #7251 at pytorch.

[markroxor]

I think something like X = ch.zeros((n, n_dim), , dtype=torch.float64) should solve my issue. Unfortunately I couldn't find it in the documentation as it is still a unstable build feature.

[markroxor]

Fix available in pytorch v0.4.