benathi / multisense-prob-fasttext Goto Github PK

Hello,

I am interested in saving the variance in the unimodal case. So I implemented the negativeSampling in the case where multi is false and var is true.

For some reason the variance does not seem to change a lot, it stays around the initial value. I tried to use the variance without the setting it to the log, but then the loss blows up. Is this the reason you store the log?

Also, it seems to me that when var is false, you never change its value. I get that in the spherical case, it simplifies the energy (equation 5), but shouldn't you still update the variance (just a scalar, not a vector anymore).

Many thanks

since learning rate [0.1] for fasttext, and 1e-5 here

Hi,

I try to download the pretrained embeddings for the model but both 7z and zip link failed.
here is the error message

wget https://s3.amazonaws.com/probabilistic-ft-multisense/multift-english/mv-wacky_e10_d300_vs2e-4_lr1e-5_mar1.zip -P modelfiles/
--2018-07-23 19:46:29-- https://s3.amazonaws.com/probabilistic-ft-multisense/multift-english/mv-wacky_e10_d300_vs2e-4_lr1e-5_mar1.zip
Resolving s3.amazonaws.com (s3.amazonaws.com)... 52.216.129.173
Connecting to s3.amazonaws.com (s3.amazonaws.com)|52.216.129.173|:443... connected.
HTTP request sent, awaiting response... 403 Forbidden
2018-07-23 19:46:30 ERROR 403: Forbidden.

Would you mind solving this link issue?

Best,
Guoyin

Is PTF embedding better than word2vec embedding on text classification and sentence similarity task?
Thanks!

Hi,

I am a fresher of word2vec research. When I copied this code to my own jupyter, I found there was the error existed as below:

NameError Traceback (most recent call last)
in
1 # caching
----> 2 if self.cache and self.bucket != 0:
3 print ("Cache subword")
4 self.subword_emb = self.cache_subword_rep(basename=basename)
5 if multi:

NameError: name 'self' is not defined

I tried to find a method to address this issue, but did not make any sense. Just wandering whether you have got the same experience? I am looking forward to your reply. Thank you.

Kind Regards,
Siyue

Where can I download the original paper? THANKS

Hi, I trained the model on 50k files with same arguments given in train_text8_multi.sh , which took around 7 to 8 hours for training. Now I am trying to load it in python, I used below code:

import multift
ft = multift.MultiFastText(basename="modelfiles/multi_text8_e10_d300_vs2e-4_lr1e-5_margin1", multi=True)

The size of modelfiles folder is around 5GB.
But now it is giving me error:
EmptyDataError: No columns to parse from file