What's necessary in order to train dialogues with length longer than 1023? about ardm

[qywu]

You can set fixed window size and shift it during the generation. Thus, you can manage the length to be under 1024.

[dimeldo]

Thus, you can manage the length to be under 1024.

Longer than 1024, you mean?

Also, would you be able to support this in your code, if it's not too much of an effort? I'm lacking the abilities to implement it myself.

[qywu]

You have a sampler to sample dialogs sequentially. I have attached a version that implements random sampling. You can just modify it.

class DialogFragmentSampler:
    def __init__(self, max_len=1024):
        """Sample dialog fragments from a dialog
        """
        self.max_tokens_len = max_len

    def __call__(self, dialog):
        """dialog is a dict which has key "token_ids" and "text" with list of turns
        """
        dialog_fragment = {}

        lengths = np.array([len(item) for item in dialog['token_ids']])

        # if the entire dialog is smaller than the max len
        if lengths.sum() < self.max_tokens_len:
            return dialog

        cumsum_len = lengths.cumsum()
        reverse_cumsum_len = cumsum_len[-1] - cumsum_len

        # based on the reverse cumsum, we can have a range to select from
        start_turns = np.arange(
            len(reverse_cumsum_len))[reverse_cumsum_len > self.max_tokens_len]
        # remove odd numbers
        start_turns = [idx for idx in start_turns if idx % 2 == 0]
        # randomly choose one
        random_start_turn = random.choice(start_turns)
        new_cumsum_len = cumsum_len - cumsum_len[random_start_turn]

        # find the maximum end turn (only odd turn)
        for i in reversed(range(len(new_cumsum_len))):
            if i % 2 == 1 and new_cumsum_len[i] < self.max_tokens_len:
                random_end_turn = i
                break

        dialog_fragment["text"] = dialog['text'][random_start_turn:
                                                 random_end_turn + 1]
        dialog_fragment["token_ids"] = dialog['token_ids'][random_start_turn:
                                                           random_end_turn + 1]

        return dialog_fragment

[dimeldo]

Isn't sampling for testing? I want the model to be able to learn to represent dialogs longer than 1024, not just generating. In any case, it seems like I'm inadequate to comprehend your code and what you're saying.

[qywu]

The current model can't encode more than 1024. You can try to apply the idea to models like XL-Net, which might solve your problem.

[dimeldo]

Gotcha, thanks a lot.