mxl1990 / tsne-pytorch Goto Github PK

Hello

I have an array of Tensors which represents words in certain language.
Is it possible to run tsne on this array? thanks

Hello, my test code is as follow:

X1 = torch.randn([10, 2048]).cuda()
X_emb = TorchTSNE(n_components=2, perplexity=100., n_iter=1000).fit_transform(X1)

Sometimes, it will occur '{IndexError}index 2048 is out of bounds for dimension 1 with size 2048', I find the reason is that as 121 line in the tsne_pytorch.py, there may be out of bounds:

119 for i in range(d):
120     if l[i, 1] != 0:
121         M[:, i+1] = M[:, i]
122         i += 1

Concretely, i may equal to d-1 in 121 line which causes M[:,i+1] will be out of bounds(we know M is dxd).

I wonder if the modified code as follow is right?

for i in range(d-1):
    if l[i, 1] != 0:
        M[:, i+1] = M[:, i]
        i += 1

Thank you for sharing the code. Is max_iter = 1000 fixed, or set it according to own data set?

Here is an example:

from tsne_torch import TorchTSNE as TSNE
# from sklearn.manifold import TSNE
import torch
import torchvision
import torchvision.transforms as transforms
import matplotlib.pyplot as plt

dataset = torchvision.datasets.MNIST(root="path of the dataset",
                                     train=False,
                                     download=False,
                                     transform=transforms.ToTensor())
x = dataset.data
n = x.shape[0]
x = x.reshape(n, -1)
n0 = int(0.01 * n)
x = torch.tensor(x[:n0], dtype=torch.float32)

X_tsne = TSNE(n_components=2, perplexity=10, n_iter=250,
              verbose=True).fit_transform(x)
print(X_tsne)

This gives a series of nans. However, if we use sklearn instead (as commented in the original code), this gives some meaningful results.

Could you please figure out why this is the case?

Hello,
As i understand , this line deals with imaginary eigenvalues, and creates real eigenvectors in M. according to the torch documentation for eig, the suggestion is to perform linear combinations of the ith and i+1th initial eigenvectors to get the real eigenvectors. however in the code there seems to just be a reassignment, which will still leave M[:,i] complex if it was originally so ( because then M[:,i+1] would have been complex as well)
can you please verify this line is correct?

Hello，I have a question when I run (l, M) = torch.eig(torch.mm(X.t(), X), True) ,sometimes the program is stuck，Is this a torch version issue？