Comments (4)
I haven't been able to train my models yet just with normal transformers, using larger context lengths (my weird TTS + STT system). CTC loss isn't converging at all. So haven't attempted a proper run with RMT architecture in the STT model. But setting it up with RMT while debugging the other one. I will let you know if i find success. I'm worried that training 2 transformers in tandem simply doesn't work for reasons. Either because of stupidly slow convergence, too lower batch size, or other reasons... Don't know. I've been looking at shifted tokens, scale_norm and other tricks to help with convergence. But i'm not getting any luck. I'm tempted to try RWKV as they claim really fast convergence. Either way, I'm going to need something like RMT in the end so i can have a well defined streaming architecture on the STT side.
from recurrent-memory-transformer-pytorch.
@pfeatherstone ahh, yea, i can look into that
care to share what you are seeing on your dataset with this approach?
from recurrent-memory-transformer-pytorch.
oh got it, makes sense
from recurrent-memory-transformer-pytorch.
Gave this a go, it turns out that torch.jit.trace()
doesn't accept None
in example_inputs
. So we cannot trace with mems
not None and expect to work when None, or vice versa. My workaround is to pass mems=torch.zeros(B,num_memory_tokens,dim)
in the first pass. Which means you're attending to self.read_memory_emb
ONLY in the first pass. Don't know if that's allowed.
from recurrent-memory-transformer-pytorch.
Related Issues (19)
- causal mask assert hit HOT 25
- Question: configuring scaled_dot_product_attention
- Attend : check mask isn't already 4D HOT 1
- Question: how does memory replay backprogagation work with multiple models in series HOT 8
- Question: why do we need read_memory_emb HOT 6
- Question: masks HOT 3
- Question: Global write tokens or recurrent HOT 7
- Bug: resiDual implementation HOT 3
- Question: first read memories HOT 12
- flash attention, and a potentially better improvement HOT 5
- What is the reasoning for no dropout? HOT 2
- token_shift HOT 4
- bptt depth implementation? HOT 3
- have you had a chance to train it yet? HOT 2
- What happens if texts from the dataset don't have equal lengths HOT 4
- Is rmt compatible with pretrain models like LLaMA? HOT 6
- Question: How to set seq_len ? HOT 1
- Question: how to adapt this for CTC loss HOT 2
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
-
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
-
TensorFlow
An Open Source Machine Learning Framework for Everyone
-
Django
The Web framework for perfectionists with deadlines.
-
Laravel
A PHP framework for web artisans
-
D3
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
-
Recommend Topics
-
javascript
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
-
web
Some thing interesting about web. New door for the world.
-
server
A server is a program made to process requests and deliver data to clients.
-
Machine learning
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
-
Visualization
Some thing interesting about visualization, use data art
-
Game
Some thing interesting about game, make everyone happy.
Recommend Org
-
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
-
Microsoft
Open source projects and samples from Microsoft.
-
Google
Google ❤️ Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.
from recurrent-memory-transformer-pytorch.