trax
trax copied to clipboard
google
Trax — Deep Learning with Clear Code and Speed
### Description I use just some input activations, one SelfAttention layer and `n_heads=2`, but my code breaks. However, when I set `n_heads=1`, everything works fine. ### Environment information ``` OS:...
### Description ... ### Environment information tensorflow-gpu installed through anaconda, trax installed via pip. ``` OS: Ubuntu 20.04 $ pip freeze | grep trax trax==1.3.9 $ pip freeze | grep...
### Description I was using the pre-defined model `trax.models.rnn.LSTMSeq2SeqAttn()` to train on a NLP dataset, which is transformed to vectors. The training was okay and I have the saved model...
In Keras we can easily implement our regularizers such as : tf.keras.layers.Dense(3, kernel_regularizer='l2'). But I could not find such an attribute or function to do a similar task in Trax....
### Description LSHSelfAttention only works with different lengths, after initiated, if the parameter use_reference_code is set to "True". So I cant use the LSHSelfAttention in a Reformer Model with BucketByLength....
### Description Hello, I am having issues with the tensorflow imports while using trax, I've never experienced the issue before and not sure how to solve it. ### Environment information...
I've been fumbling around the whole day trying to get Trax to use GPU but FAILED.. is there some kind of clear guidelines on how to get Trax to see...
### Description Hello. I am taking Coursera NLP attention model course, with Trax package. During the course assignment, the task is to train model and evaluate the model performance. Course...
### Description PretrainedBERT model fails on its initialization. ### Environment information ``` OS: macOS Big Sur 11.4 $ pip freeze | grep trax trax==1.3.9 $ pip freeze | grep tensor...
When I read docs here https://trax-ml.readthedocs.io/en/latest/trax.models.html#trax.models.reformer.reformer.EncoderDecoderBlock I see a name `EncoderDecoderBlock`, but later a description of it: > Reversible transformer decoder layer. Is it a typo?