strhub/models/abinet/model.py

import torch
import torch.nn as nn


class Model(nn.Module):

    def __init__(self, dataset_max_length: int, null_label: int):
        super().__init__()
        self.max_length = dataset_max_length + 1  # additional stop token
        self.null_label = null_label

    def _get_length(self, logit, dim=-1):
        """ Greed decoder to obtain length from logit"""
        out = (logit.argmax(dim=-1) == self.null_label)
        abn = out.any(dim)
        out = ((out.cumsum(dim) == 1) & out).max(dim)[1]
        out = out + 1  # additional end token
        out = torch.where(abn, out, out.new_tensor(logit.shape[1], device=out.device))
        return out

    @staticmethod
    def _get_padding_mask(length, max_length):
        length = length.unsqueeze(-1)
        grid = torch.arange(0, max_length, device=length.device).unsqueeze(0)
        return grid >= length

    @staticmethod
    def _get_location_mask(sz, device=None):
        mask = torch.eye(sz, device=device)
        mask = mask.float().masked_fill(mask == 1, float('-inf'))
        return mask