Fangjun Kuang

I just created a colab notebook (see https://colab.research.google.com/drive/1iyc_q8aHuKd-RZxtYv9EqfyjB2QZDSOx?usp=sharing) to verify the idea. The following code you posted: ``` graph, _ = self.graph_compiler.compile(texts, self.P, replicate_den=True) T = x.size()[1] tot_scores = []...

For the following `dense_fsa_vec` (you can find the code for all the below comments in the above colab notebook)  with the following decoding...

> In other words, would the lats.get_tot_scores lead to the wrong number in some concept for P(W|O_{1:t}) if t is not equal to T? Maybe @danpovey has more to say...

What's the distribution of your utterance length?

There is a similar issue at https://stackoverflow.com/questions/74516660/ios-app-crashes-with-required-condition-is-false-isformatsamplerateandchannelc Could you have a look?

have you tried the ios-swiftui example?

By the way, this repo is for sherpa-onnx. Does it work with sherpa-onnx?

This PR is not for merge. It is useful for visualizing the node gradient in the lattice during training.

Note: The above plots are from the first batch at the very beginning of the training, i.e., the model weights are randomly initialized and no backward pass has been performed...

For better comparison, the plots between the model with randomly initialized weights and the pre-trained model are listed as follows: | Randomly initialized | Pre-trained| |---------------------|-------------| |  |...