herandy / depict Goto Github PK

Hello! I attempted to run your code on another dataset, but I'm not sure how to adapt the parameters for the new dataset. Could you please tell me how to configure your code and which parameters to change? Thank you!

Hi,
When I try to train DEPICT I receive the following error in random epochs. I have received it in epochs 165 400 2000. The error is
TypeError: ('An update must have the same type as the original shared variable (shared_var=<TensorType(float32, matrix)>, shared_var.type=TensorType(float32, matrix), update_val=Elemwise{add,no_inplace}.0, update_val.type=TensorType(float64, matrix)).', 'If the difference is related to the broadcast pattern, you can call the tensor.unbroadcast(var, axis_to_unbroadcast[, ...]) function to remove broadcastable dimensions.')
It occurs in this block of code:
updates = lasagne.updates.adam(
loss, params2, learning_rate=learning_rate)
train_fn = theano.function([input_var, target_var],
[loss, loss_recons, loss_clus], updates=updates)

Did you face it ? Any ideas on how can I bypass it? The versions of Theano and Lasagne are as instructed in the guidelines.

In the paper, q is updated according to the commented line. What is the reason behind the change in the code?

y_prob = y_prob ** 2 / cluster_frequency
# y_prob = y_prob / np.sqrt(cluster_frequency)

Thanks.

Hi, do you have TensorFlow codes for this paper?

It seems that the loss function in your code is different from that in your paper. I've changed into the loss you mentioned in your paper, it didn't work very well. Could you give me some advice?

If the u is not uniform prior for the empirical label distribution, is the eq.(7)(8) still right?

will you please tell me how to get datasets

Hi, do you have Pytorch codes for this paper?

Hi, I can't understand your derivation from fomular (6) to formular (7) by setting the gradient to zero. And it changes to another form in the algorithm 1 of the paper. In fact when I compute that I can't get the same solution. Could you explain this?

Hello!

First of all thanks for sharing this work, it's amazing!

I have a bit of a peculiar question I guess. Do you think DEPICT could easily be adapted to fit to text classification. I'm really not trained in the field so my simple mind only tells me change Conv layers to 1D and adjust their options and you're good to go, but of course it has to be a whole lot more than that.

Was wondering if you had ever thought of this application and if you think it would work. If not the specific model, then a similar approach to text classification. Seems to fit in the efforts of classifying documents that can take one out of n possible sentiment target values.

Would love to hear your thoughts!

Kind regards,
Theodore.

hello! i've attempted to run your code on the YTF dataset, but the clustering accuracy i can achieved is 56%, i wonder i missed some important process in the code running, could you show your running detail or give me some advice, thank you!