-
I am working on streaming, long-form decoding on TED-Lium. I have a causal Zipformer encoder, and have adapted the `streaming_decode.py` script to decode full recordings in a streaming fashion and als…
-
https://huggingface.co/blog/how-to-generate#beam-search
Original paragraph:
At time step 1, besides the most likely hypothesis \text{"The", "woman"}"The", "woman", beam search also keeps track of …
-
Compared to OpenNMT, why do we need this [block](https://github.com/IBM/pytorch-seq2seq/blob/master/seq2seq/models/TopKDecoder.py#L257) which handles the dropped sequences that see EOS earlier. (This …
-
Are there any plans to support online inference with beam search?
Usecase:
Audio comes in n milisecond chunks and gets processed by an accoustic model that outputs a log_prob frame for each audio …
-
### What needs to happen?
It seems that one of the python libraries Beam uses, js2py, has a vulnerability:
[CVE-2024-28397](https://nvd.nist.gov/vuln/detail/CVE-2024-28397)
[Exposure](https://g…
-
Hi, great work on the beam search implementation. I was going through your code to understand your implementation and noticed that you a) Don't terminate a beam when the EOS token is predicted by the …
-
Need sampling with temperature for inference.
At each search step, sample beam_size tokens from the token distribution instead of taking the top k.
-
I was wondering if you plan to implement beam search for generation of translation candidates, as described here:
![image](https://cloud.githubusercontent.com/assets/1579822/24578279/0ff538be-16e6-11…
-
@patrickvonplaten
I was recently tinkering around the generation strategies for a decoder, and came across this [blog post](https://huggingface.co/blog/how-to-generate) on decoding methods. It was …
-
Really nice work! Have you found the reason why beam search fail in dgl transformer?