Darwin Bautista

Are you using the pretrained weights as is? What's the difference between the NED of AR vs NAR? For the standard STR benchmarks, the difference between the two is within...

I would reckon that the difference in outputs is due to differences in the internal data representations. Can't really say, since I'm not familiar with the implementation of Torch in...

> I found the torch version of parseq which is able to convert onnx and tensorrt too > https://github.com/bharatsubedi/PARseq_torch The reference implementation **is** PyTorch. PyTorch-Lightning is only used for training...

@gioivuathoi is correct. The forward method of the PARSeq model is dynamic in the sense that the actual code path is dependent on the input, especially for the autoregressive decoding...

During training, all input tokens are used in the input context in order to take advantage of the parallel processing of Transformers. For testing, only `[B]` (the beginning-of-sequence token) is...

A dataset that small + training schedule that long would definitely result in overfitting. 1. Don't use the default hyperparameters. 2. Try decoding with `decode_ar=False` and `refine_iters=0`.

It should be possible, but you won't be able to use the built-in finetuning code since the output shape will change. The process should look something like this: 1. Append...

Comment [#1](https://code.google.com/p/ufw-frontends/issues/detail?id=8#c1) originally posted by baudm on 2010-08-31T22:43:17.000Z: This should be Enhancement not defect, i do not know how to edit this...

Comment [#2](https://code.google.com/p/ufw-frontends/issues/detail?id=8#c2) originally posted by baudm on 2010-09-01T02:14:50.000Z: Blocked on this bug: https://bugs.launchpad.net/ufw/+bug/544976

Comment [#3](https://code.google.com/p/ufw-frontends/issues/detail?id=8#c3) originally posted by baudm on 2010-09-01T02:18:55.000Z: <empty>