manga_translator/ocr/model_32px.py

import math
from typing import List
from collections import defaultdict
import os
import shutil
import cv2
import numpy as np
import einops

import torch
import torch.nn as nn
import torch.nn.functional as F

from .common import OfflineOCR
from ..utils import TextBlock, Quadrilateral, chunks
from ..utils.bubble import is_ignore

class Model32pxOCR(OfflineOCR):
    _MODEL_MAPPING = {
        'model': {
            'url': 'https://github.com/zyddnys/manga-image-translator/releases/download/beta-0.3/ocr.zip',
            'hash': '47405638b96fa2540a5ee841a4cd792f25062c09d9458a973362d40785f95d7a',
            'archive': {
                'ocr.ckpt': '.',
                'alphabet-all-v5.txt': '.',
            },
        },
    }

    def __init__(self, *args, **kwargs):
        os.makedirs(self.model_dir, exist_ok=True)
        if os.path.exists('ocr.ckpt'):
            shutil.move('ocr.ckpt', self._get_file_path('ocr.ckpt'))
        if os.path.exists('alphabet-all-v5.txt'):
            shutil.move('alphabet-all-v5.txt', self._get_file_path('alphabet-all-v5.txt'))
        super().__init__(*args, **kwargs)

    async def _load(self, device: str):
        with open(self._get_file_path('alphabet-all-v5.txt'), 'r', encoding = 'utf-8') as fp:
            dictionary = [s[:-1] for s in fp.readlines()]

        self.model = OCR(dictionary, 768)
        sd = torch.load(self._get_file_path('ocr.ckpt'), map_location = 'cpu')
        self.model.load_state_dict(sd['model'] if 'model' in sd else sd)
        self.model.eval()
        self.device = device
        if (device == 'cuda' or device == 'mps'):
            self.use_gpu = True
        else:
            self.use_gpu = False
        if self.use_gpu:
            self.model = self.model.to(device)


    async def _unload(self):
        del self.model

    async def _infer(self, image: np.ndarray, textlines: List[Quadrilateral], args: dict, verbose: bool = False) -> List[TextBlock]:
        text_height = 32
        max_chunk_size = 16
        ignore_bubble = args.get('ignore_bubble', 0)

        quadrilaterals = list(self._generate_text_direction(textlines))
        region_imgs = [q.get_transformed_region(image, d, text_height) for q, d in quadrilaterals]
        out_regions = []

        perm = range(len(region_imgs))
        is_quadrilaterals = False
        if len(quadrilaterals) > 0 and isinstance(quadrilaterals[0][0], Quadrilateral):
            perm = sorted(range(len(region_imgs)), key = lambda x: region_imgs[x].shape[1])
            is_quadrilaterals = True

        ix = 0
        for indices in chunks(perm, max_chunk_size):
            N = len(indices)
            widths = [region_imgs[i].shape[1] for i in indices]
            max_width = 4 * (max(widths) + 7) // 4
            region = np.zeros((N, text_height, max_width, 3), dtype = np.uint8)
            for i, idx in enumerate(indices):
                W = region_imgs[idx].shape[1]
                tmp = region_imgs[idx]
                # Determine whether to skip the text block, and return True to skip.
                if ignore_bubble >=1 and ignore_bubble <=50 and is_ignore(region_imgs[idx],ignore_bubble):
                    ix+=1
                    continue
                region[i, :, : W, :]=tmp
                if verbose:
                    os.makedirs('result/ocrs/', exist_ok=True)
                    if quadrilaterals[idx][1] == 'v':
                        cv2.imwrite(f'result/ocrs/{ix}.png', cv2.rotate(cv2.cvtColor(region[i, :, :, :], cv2.COLOR_RGB2BGR), cv2.ROTATE_90_CLOCKWISE))
                    else:
                        cv2.imwrite(f'result/ocrs/{ix}.png', cv2.cvtColor(region[i, :, :, :], cv2.COLOR_RGB2BGR))
                ix += 1
            image_tensor = (torch.from_numpy(region).float() - 127.5) / 127.5
            image_tensor = einops.rearrange(image_tensor, 'N H W C -> N C H W')
            if self.use_gpu:
                image_tensor = image_tensor.to(self.device)
            with torch.no_grad():
                ret = self.model.infer_beam_batch(image_tensor, widths, beams_k = 5, max_seq_length = 255)
            for i, (pred_chars_index, prob, fr, fg, fb, br, bg, bb) in enumerate(ret):
                if prob < 0.7:
                    continue
                fr = (torch.clip(fr.view(-1), 0, 1).mean() * 255).long().item()
                fg = (torch.clip(fg.view(-1), 0, 1).mean() * 255).long().item()
                fb = (torch.clip(fb.view(-1), 0, 1).mean() * 255).long().item()
                br = (torch.clip(br.view(-1), 0, 1).mean() * 255).long().item()
                bg = (torch.clip(bg.view(-1), 0, 1).mean() * 255).long().item()
                bb = (torch.clip(bb.view(-1), 0, 1).mean() * 255).long().item()
                seq = []
                for chid in pred_chars_index:
                    ch = self.model.dictionary[chid]
                    if ch == '<S>':
                        continue
                    if ch == '</S>':
                        break
                    if ch == '<SP>':
                        ch = ' '
                    seq.append(ch)
                txt = ''.join(seq)
                self.logger.info(f'prob: {prob} {txt} fg: ({fr}, {fg}, {fb}) bg: ({br}, {bg}, {bb})')
                cur_region = quadrilaterals[indices[i]][0]
                if isinstance(cur_region, Quadrilateral):
                    cur_region.text = txt
                    cur_region.prob = prob
                    cur_region.fg_r = fr
                    cur_region.fg_g = fg
                    cur_region.fg_b = fb
                    cur_region.bg_r = br
                    cur_region.bg_g = bg
                    cur_region.bg_b = bb
                else:
                    cur_region.text.append(txt)
                    cur_region.update_font_colors(np.array([fr, fg, fb]), np.array([br, bg, bb]))

                out_regions.append(cur_region)

        if is_quadrilaterals:
            return out_regions
        return textlines


class ResNet(nn.Module):

    def __init__(self, input_channel, output_channel, block, layers):
        super(ResNet, self).__init__()

        self.output_channel_block = [int(output_channel / 4), int(output_channel / 2), output_channel, output_channel]

        self.inplanes = int(output_channel / 8)
        self.conv0_1 = nn.Conv2d(input_channel, int(output_channel / 8),
                                 kernel_size=3, stride=1, padding=1, bias=False)
        self.bn0_1 = nn.BatchNorm2d(int(output_channel / 8))
        self.conv0_2 = nn.Conv2d(int(output_channel / 8), self.inplanes,
                                 kernel_size=3, stride=1, padding=1, bias=False)

        self.maxpool1 = nn.AvgPool2d(kernel_size=2, stride=2, padding=0)
        self.layer1 = self._make_layer(block, self.output_channel_block[0], layers[0])
        self.bn1 = nn.BatchNorm2d(self.output_channel_block[0])
        self.conv1 = nn.Conv2d(self.output_channel_block[0], self.output_channel_block[
                               0], kernel_size=3, stride=1, padding=1, bias=False)

        self.maxpool2 = nn.AvgPool2d(kernel_size=2, stride=2, padding=0)
        self.layer2 = self._make_layer(block, self.output_channel_block[1], layers[1], stride=1)
        self.bn2 = nn.BatchNorm2d(self.output_channel_block[1])
        self.conv2 = nn.Conv2d(self.output_channel_block[1], self.output_channel_block[
                               1], kernel_size=3, stride=1, padding=1, bias=False)

        self.maxpool3 = nn.AvgPool2d(kernel_size=2, stride=(2, 1), padding=(0, 1))
        self.layer3 = self._make_layer(block, self.output_channel_block[2], layers[2], stride=1)
        self.bn3 = nn.BatchNorm2d(self.output_channel_block[2])
        self.conv3 = nn.Conv2d(self.output_channel_block[2], self.output_channel_block[
                               2], kernel_size=3, stride=1, padding=1, bias=False)

        self.layer4 = self._make_layer(block, self.output_channel_block[3], layers[3], stride=1)
        self.bn4_1 = nn.BatchNorm2d(self.output_channel_block[3])
        self.conv4_1 = nn.Conv2d(self.output_channel_block[3], self.output_channel_block[
                                 3], kernel_size=2, stride=(2, 1), padding=(0, 1), bias=False)
        self.bn4_2 = nn.BatchNorm2d(self.output_channel_block[3])
        self.conv4_2 = nn.Conv2d(self.output_channel_block[3], self.output_channel_block[
                                 3], kernel_size=2, stride=1, padding=0, bias=False)
        self.bn4_3 = nn.BatchNorm2d(self.output_channel_block[3])

    def _make_layer(self, block, planes, blocks, stride=1):
        downsample = None
        if stride != 1 or self.inplanes != planes * block.expansion:
            downsample = nn.Sequential(
                nn.BatchNorm2d(self.inplanes),
                nn.Conv2d(self.inplanes, planes * block.expansion,
                          kernel_size=1, stride=stride, bias=False),
            )

        layers = []
        layers.append(block(self.inplanes, planes, stride, downsample))
        self.inplanes = planes * block.expansion
        for i in range(1, blocks):
            layers.append(block(self.inplanes, planes))

        return nn.Sequential(*layers)

    def forward(self, x):
        x = self.conv0_1(x)
        x = self.bn0_1(x)
        x = F.relu(x)
        x = self.conv0_2(x)

        x = self.maxpool1(x)
        x = self.layer1(x)
        x = self.bn1(x)
        x = F.relu(x)
        x = self.conv1(x)

        x = self.maxpool2(x)
        x = self.layer2(x)
        x = self.bn2(x)
        x = F.relu(x)
        x = self.conv2(x)

        x = self.maxpool3(x)
        x = self.layer3(x)
        x = self.bn3(x)
        x = F.relu(x)
        x = self.conv3(x)

        x = self.layer4(x)
        x = self.bn4_1(x)
        x = F.relu(x)
        x = self.conv4_1(x)
        x = self.bn4_2(x)
        x = F.relu(x)
        x = self.conv4_2(x)
        x = self.bn4_3(x)

        return x

class BasicBlock(nn.Module):
    expansion = 1

    def __init__(self, inplanes, planes, stride=1, downsample=None):
        super(BasicBlock, self).__init__()
        self.bn1 = nn.BatchNorm2d(inplanes)
        self.conv1 = self._conv3x3(inplanes, planes)
        self.bn2 = nn.BatchNorm2d(planes)
        self.conv2 = self._conv3x3(planes, planes)
        self.downsample = downsample
        self.stride = stride

    def _conv3x3(self, in_planes, out_planes, stride=1):
        "3x3 convolution with padding"
        return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride,
                         padding=1, bias=False)

    def forward(self, x):
        residual = x

        out = self.bn1(x)
        out = F.relu(out)
        out = self.conv1(out)

        out = self.bn2(out)
        out = F.relu(out)
        out = self.conv2(out)

        if self.downsample is not None:
            residual = self.downsample(residual)

        return out + residual

def conv3x3(in_planes, out_planes, stride=1, groups=1, dilation=1):
    """3x3 convolution with padding"""
    return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride,
                     padding=dilation, groups=groups, bias=False, dilation=dilation)


def conv1x1(in_planes, out_planes, stride=1):
    """1x1 convolution"""
    return nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, bias=False)

class ResNet_FeatureExtractor(nn.Module):
    """ FeatureExtractor of FAN (http://openaccess.thecvf.com/content_ICCV_2017/papers/Cheng_Focusing_Attention_Towards_ICCV_2017_paper.pdf) """

    def __init__(self, input_channel, output_channel=128):
        super(ResNet_FeatureExtractor, self).__init__()
        self.ConvNet = ResNet(input_channel, output_channel, BasicBlock, [3, 6, 7, 5])

    def forward(self, input):
        return self.ConvNet(input)

class PositionalEncoding(nn.Module):
    def __init__(self, d_model, dropout=0.1, max_len=5000):
        super(PositionalEncoding, self).__init__()
        self.dropout = nn.Dropout(p=dropout)

        pe = torch.zeros(max_len, d_model)
        position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1)
        div_term = torch.exp(torch.arange(0, d_model, 2).float() * (-math.log(10000.0) / d_model))
        pe[:, 0::2] = torch.sin(position * div_term)
        pe[:, 1::2] = torch.cos(position * div_term)
        pe = pe.unsqueeze(0).transpose(0, 1)
        self.register_buffer('pe', pe)

    def forward(self, x, offset = 0):
        x = x + self.pe[offset: offset + x.size(0), :]
        return x#self.dropout(x)

def generate_square_subsequent_mask(sz):
    mask = (torch.triu(torch.ones(sz, sz)) == 1).transpose(0, 1)
    mask = mask.float().masked_fill(mask == 0, float('-inf')).masked_fill(mask == 1, float(0.0))
    return mask

class AddCoords(nn.Module):

    def __init__(self, with_r=False):
        super().__init__()
        self.with_r = with_r

    def forward(self, input_tensor):
        """
        Args:
            input_tensor: shape(batch, channel, x_dim, y_dim)
        """
        batch_size, _, x_dim, y_dim = input_tensor.size()

        xx_channel = torch.arange(x_dim).repeat(1, y_dim, 1)
        yy_channel = torch.arange(y_dim).repeat(1, x_dim, 1).transpose(1, 2)

        xx_channel = xx_channel.float() / (x_dim - 1)
        yy_channel = yy_channel.float() / (y_dim - 1)

        xx_channel = xx_channel * 2 - 1
        yy_channel = yy_channel * 2 - 1

        xx_channel = xx_channel.repeat(batch_size, 1, 1, 1).transpose(2, 3)
        yy_channel = yy_channel.repeat(batch_size, 1, 1, 1).transpose(2, 3)

        ret = torch.cat([
            input_tensor,
            xx_channel.type_as(input_tensor),
            yy_channel.type_as(input_tensor)], dim=1)

        if self.with_r:
            rr = torch.sqrt(torch.pow(xx_channel.type_as(input_tensor) - 0.5, 2) + torch.pow(yy_channel.type_as(input_tensor) - 0.5, 2))
            ret = torch.cat([ret, rr], dim=1)

        return ret

class Beam:
    def __init__(self, char_seq = [], logprobs = []):
        # L
        if isinstance(char_seq, list):
            self.chars = torch.tensor(char_seq, dtype=torch.long)
            self.logprobs = torch.tensor(logprobs, dtype=torch.float32)
        else:
            self.chars = char_seq.clone()
            self.logprobs = logprobs.clone()

    def avg_logprob(self):
        return self.logprobs.mean().item()

    def sort_key(self):
        return -self.avg_logprob()

    def seq_end(self, end_tok):
        return self.chars.view(-1)[-1] == end_tok

    def extend(self, idx, logprob):
        return Beam(
            torch.cat([self.chars, idx.unsqueeze(0)], dim = -1),
            torch.cat([self.logprobs, logprob.unsqueeze(0)], dim = -1),
        )

DECODE_BLOCK_LENGTH = 8

class Hypothesis:
    def __init__(self, device, start_tok: int, end_tok: int, padding_tok: int, memory_idx: int, num_layers: int, embd_dim: int):
        self.device = device
        self.start_tok = start_tok
        self.end_tok = end_tok
        self.padding_tok = padding_tok
        self.memory_idx = memory_idx
        self.embd_size = embd_dim
        self.num_layers = num_layers
        # L, 1, E
        self.cached_activations = [torch.zeros(0, 1, self.embd_size).to(self.device)] * (num_layers + 1)
        self.out_idx = torch.LongTensor([start_tok]).to(self.device)
        self.out_logprobs = torch.FloatTensor([0]).to(self.device)
        self.length = 0

    def seq_end(self):
        return self.out_idx.view(-1)[-1] == self.end_tok

    def logprob(self):
        return self.out_logprobs.mean().item()

    def sort_key(self):
        return -self.logprob()

    def prob(self):
        return self.out_logprobs.mean().exp().item()

    def __len__(self):
        return self.length

    def extend(self, idx, logprob):
        ret = Hypothesis(self.device, self.start_tok, self.end_tok, self.padding_tok, self.memory_idx, self.num_layers, self.embd_size)
        ret.cached_activations = [item.clone() for item in self.cached_activations]
        ret.length = self.length + 1
        ret.out_idx = torch.cat([self.out_idx, torch.LongTensor([idx]).to(self.device)], dim = 0)
        ret.out_logprobs = torch.cat([self.out_logprobs, torch.FloatTensor([logprob]).to(self.device)], dim = 0)
        return ret

    def output(self):
        return self.cached_activations[-1]

def next_token_batch(
    hyps: List[Hypothesis],
    memory: torch.Tensor, # S, K, E
    memory_mask: torch.BoolTensor,
    decoders: nn.TransformerDecoder,
    pe: PositionalEncoding,
    embd: nn.Embedding
    ):
    layer: nn.TransformerDecoderLayer
    N = len(hyps)

    # N
    last_toks = torch.stack([item.out_idx[-1] for item in hyps], dim = 0)
    # 1, N, E
    tgt: torch.FloatTensor = pe(embd(last_toks).unsqueeze_(0), offset = len(hyps[0]))

    # # L, N
    # out_idxs = torch.stack([item.out_idx for item in hyps], dim = 0).permute(1, 0)
    # # L, N, E
    # tgt2: torch.FloatTensor = pe(embd(out_idxs))
    # # 1, N, E
    # tgt_v2 = tgt2[-1, :, :].unsqueeze_(0)
    # print(((tgt_v1 - tgt_v2) ** 2).sum())

    # tgt = tgt_v2

    # S, N, E
    memory = torch.stack([memory[:, idx, :] for idx in [item.memory_idx for item in hyps]], dim = 1)
    for l, layer in enumerate(decoders.layers):
        # TODO: keys and values are recomputed every time
        # L - 1, N, E
        combined_activations = torch.cat([item.cached_activations[l] for item in hyps], dim = 1)
        # L, N, E
        combined_activations = torch.cat([combined_activations, tgt], dim = 0)
        for i in range(N):
            hyps[i].cached_activations[l] = combined_activations[:, i: i + 1, :]
        tgt2 = layer.self_attn(tgt, combined_activations, combined_activations)[0]
        tgt = tgt + layer.dropout1(tgt2)
        tgt = layer.norm1(tgt)
        tgt2 = layer.multihead_attn(tgt, memory, memory, key_padding_mask = memory_mask)[0]
        tgt = tgt + layer.dropout2(tgt2)
        tgt = layer.norm2(tgt)
        tgt2 = layer.linear2(layer.dropout(layer.activation(layer.linear1(tgt))))
        tgt = tgt + layer.dropout3(tgt2)
        # 1, N, E
        tgt = layer.norm3(tgt)
    #print(tgt[0, 0, 0])
    for i in range(N):
        hyps[i].cached_activations[decoders.num_layers] = torch.cat([hyps[i].cached_activations[decoders.num_layers], tgt[:, i: i + 1, :]], dim = 0)
    # N, E
    return tgt.squeeze_(0)

class OCR(nn.Module):
    def __init__(self, dictionary, max_len):
        super(OCR, self).__init__()
        self.max_len = max_len
        self.dictionary = dictionary
        self.dict_size = len(dictionary)
        self.backbone = ResNet_FeatureExtractor(3, 320)
        encoder = nn.TransformerEncoderLayer(320, 4, dropout = 0.0)
        decoder = nn.TransformerDecoderLayer(320, 4, dropout = 0.0)
        self.encoders = nn.TransformerEncoder(encoder, 3)
        self.decoders = nn.TransformerDecoder(decoder, 2)
        self.pe = PositionalEncoding(320, max_len = max_len)
        self.embd = nn.Embedding(self.dict_size, 320)
        self.pred1 = nn.Sequential(nn.Linear(320, 320), nn.ReLU(), nn.Dropout(0.1))
        self.pred = nn.Linear(320, self.dict_size)
        self.pred.weight = self.embd.weight
        self.color_pred1 = nn.Sequential(nn.Linear(320, 64), nn.ReLU())
        self.fg_r_pred = nn.Linear(64, 1)
        self.fg_g_pred = nn.Linear(64, 1)
        self.fg_b_pred = nn.Linear(64, 1)
        self.bg_r_pred = nn.Linear(64, 1)
        self.bg_g_pred = nn.Linear(64, 1)
        self.bg_b_pred = nn.Linear(64, 1)

    def forward(self,
        img: torch.FloatTensor,
        char_idx: torch.LongTensor,
        mask: torch.BoolTensor,
        source_mask: torch.BoolTensor
        ):
        feats = self.backbone(img)
        feats = torch.einsum('n e h s -> s n e', feats)
        feats = self.pe(feats)
        memory = self.encoders(feats, src_key_padding_mask = source_mask)
        N, L = char_idx.shape
        char_embd = self.embd(char_idx)
        char_embd = torch.einsum('n t e -> t n e', char_embd)
        char_embd = self.pe(char_embd)
        casual_mask = generate_square_subsequent_mask(L).to(img.device)
        decoded = self.decoders(char_embd, memory, tgt_mask = casual_mask, tgt_key_padding_mask = mask, memory_key_padding_mask = source_mask)
        decoded = decoded.permute(1, 0, 2)
        pred_char_logits = self.pred(self.pred1(decoded))
        color_feats = self.color_pred1(decoded)
        return pred_char_logits, \
            self.fg_r_pred(color_feats), \
            self.fg_g_pred(color_feats), \
            self.fg_b_pred(color_feats), \
            self.bg_r_pred(color_feats), \
            self.bg_g_pred(color_feats), \
            self.bg_b_pred(color_feats)

    def infer_beam_batch(self, img: torch.FloatTensor, img_widths: List[int], beams_k: int = 5, start_tok = 1, end_tok = 2, pad_tok = 0, max_finished_hypos: int = 2, max_seq_length = 384):
        N, C, H, W = img.shape
        assert H == 32 and C == 3
        feats = self.backbone(img)
        feats = torch.einsum('n e h s -> s n e', feats)
        valid_feats_length = [(x + 3) // 4 + 2 for x in img_widths]
        input_mask = torch.zeros(N, feats.size(0), dtype = torch.bool).to(img.device)
        for i, l in enumerate(valid_feats_length):
            input_mask[i, l:] = True
        feats = self.pe(feats)
        memory = self.encoders(feats, src_key_padding_mask = input_mask)
        hypos = [Hypothesis(img.device, start_tok, end_tok, pad_tok, i, self.decoders.num_layers, 320) for i in range(N)]
        # N, E
        decoded = next_token_batch(hypos, memory, input_mask, self.decoders, self.pe, self.embd)
        # N, n_chars
        pred_char_logprob = self.pred(self.pred1(decoded)).log_softmax(-1)
        # N, k
        pred_chars_values, pred_chars_index = torch.topk(pred_char_logprob, beams_k, dim = 1)
        new_hypos = []
        finished_hypos = defaultdict(list)
        for i in range(N):
            for k in range(beams_k):
                new_hypos.append(hypos[i].extend(pred_chars_index[i, k], pred_chars_values[i, k]))
        hypos = new_hypos
        for _ in range(max_seq_length):
            # N * k, E
            decoded = next_token_batch(hypos, memory, torch.stack([input_mask[hyp.memory_idx] for hyp in hypos]) , self.decoders, self.pe, self.embd)
            # N * k, n_chars
            pred_char_logprob = self.pred(self.pred1(decoded)).log_softmax(-1)
            # N * k, k
            pred_chars_values, pred_chars_index = torch.topk(pred_char_logprob, beams_k, dim = 1)
            hypos_per_sample = defaultdict(list)
            h: Hypothesis
            for i, h in enumerate(hypos):
                for k in range(beams_k):
                    hypos_per_sample[h.memory_idx].append(h.extend(pred_chars_index[i, k], pred_chars_values[i, k]))
            hypos = []
            # hypos_per_sample now contains N * k^2 hypos
            for i in hypos_per_sample.keys():
                cur_hypos: List[Hypothesis] = hypos_per_sample[i]
                cur_hypos = sorted(cur_hypos, key = lambda a: a.sort_key())[: beams_k + 1]
                #print(cur_hypos[0].out_idx[-1])
                to_added_hypos = []
                sample_done = False
                for h in cur_hypos:
                    if h.seq_end():
                        finished_hypos[i].append(h)
                        if len(finished_hypos[i]) >= max_finished_hypos:
                            sample_done = True
                            break
                    else:
                        if len(to_added_hypos) < beams_k:
                            to_added_hypos.append(h)
                if not sample_done:
                    hypos.extend(to_added_hypos)
            if len(hypos) == 0:
                break
        # add remaining hypos to finished
        for i in range(N):
            if i not in finished_hypos:
                cur_hypos: List[Hypothesis] = hypos_per_sample[i]
                cur_hypo = sorted(cur_hypos, key = lambda a: a.sort_key())[0]
                finished_hypos[i].append(cur_hypo)
        assert len(finished_hypos) == N
        result = []
        for i in range(N):
            cur_hypos = finished_hypos[i]
            cur_hypo = sorted(cur_hypos, key = lambda a: a.sort_key())[0]
            decoded = cur_hypo.output()
            color_feats = self.color_pred1(decoded)
            fg_r, fg_g, fg_b, bg_r, bg_g, bg_b = self.fg_r_pred(color_feats), \
                self.fg_g_pred(color_feats), \
                self.fg_b_pred(color_feats), \
                self.bg_r_pred(color_feats), \
                self.bg_g_pred(color_feats), \
                self.bg_b_pred(color_feats)
            result.append((cur_hypo.out_idx, cur_hypo.prob(), fg_r, fg_g, fg_b, bg_r, bg_g, bg_b))
        return result

    def infer_beam(self, img: torch.FloatTensor, beams_k: int = 5, start_tok = 1, end_tok = 2, pad_tok = 0, max_seq_length = 384):
        N, C, H, W = img.shape
        assert H == 32 and N == 1 and C == 3
        feats = self.backbone(img)
        feats = torch.einsum('n e h s -> s n e', feats)
        feats = self.pe(feats)
        memory = self.encoders(feats)
        def run(tokens, add_start_tok = True, char_only = True):
            if add_start_tok:
                if isinstance(tokens, list):
                    # N(=1), L
                    tokens = torch.tensor([start_tok] + tokens, dtype = torch.long, device = img.device).unsqueeze_(0)
                else:
                    # N, L
                    tokens = torch.cat([torch.tensor([start_tok], dtype = torch.long, device = img.device), tokens], dim = -1).unsqueeze_(0)
            N, L = tokens.shape
            embd = self.embd(tokens)
            embd = torch.einsum('n t e -> t n e', embd)
            embd = self.pe(embd)
            casual_mask = generate_square_subsequent_mask(L).to(img.device)
            decoded = self.decoders(embd, memory, tgt_mask = casual_mask)
            decoded = decoded.permute(1, 0, 2)
            pred_char_logprob = self.pred(self.pred1(decoded)).log_softmax(-1)
            if char_only:
                return pred_char_logprob
            else:
                color_feats = self.color_pred1(decoded)
                return pred_char_logprob, \
                    self.fg_r_pred(color_feats), \
                    self.fg_g_pred(color_feats), \
                    self.fg_b_pred(color_feats), \
                    self.bg_r_pred(color_feats), \
                    self.bg_g_pred(color_feats), \
                    self.bg_b_pred(color_feats)
        # N, L, embd_size
        initial_char_logprob = run([])
        # N, L
        initial_pred_chars_values, initial_pred_chars_index = torch.topk(initial_char_logprob, beams_k, dim = 2)
        # beams_k, L
        initial_pred_chars_values = initial_pred_chars_values.squeeze(0).permute(1, 0)
        initial_pred_chars_index = initial_pred_chars_index.squeeze(0).permute(1, 0)
        beams = sorted([Beam(tok, logprob) for tok, logprob in zip(initial_pred_chars_index, initial_pred_chars_values)], key = lambda a: a.sort_key())
        for _ in range(max_seq_length):
            new_beams = []
            all_ended = True
            for beam in beams:
                if not beam.seq_end(end_tok):
                    logprobs = run(beam.chars)
                    pred_chars_values, pred_chars_index = torch.topk(logprobs, beams_k, dim = 2)
                    # beams_k, L
                    pred_chars_values = pred_chars_values.squeeze(0).permute(1, 0)
                    pred_chars_index = pred_chars_index.squeeze(0).permute(1, 0)
                    #print(pred_chars_index.view(-1)[-1])
                    new_beams.extend([beam.extend(tok[-1], logprob[-1]) for tok, logprob in zip(pred_chars_index, pred_chars_values)])
                    #new_beams.extend([Beam(tok, logprob) for tok, logprob in zip(pred_chars_index, pred_chars_values)]) # extend other top k
                    all_ended = False
                else:
                    new_beams.append(beam) # seq ended, add back to queue
            beams = sorted(new_beams, key = lambda a: a.sort_key())[: beams_k] # keep top k
            #print(beams[0].chars)
            if all_ended:
                break
        final_tokens = beams[0].chars[:-1]
        #print(beams[0].logprobs.mean().exp())
        return run(final_tokens, char_only = False), beams[0].logprobs.mean().exp().item()

def test():
    with open('../SynthText/alphabet-all-v2.txt', 'r') as fp:
        dictionary = [s[:-1] for s in fp.readlines()]
    img = torch.randn(4, 3, 32, 1224)
    idx = torch.zeros(4, 32).long()
    mask = torch.zeros(4, 32).bool()
    model = ResNet_FeatureExtractor(3, 256)
    out = model(img)

def test_inference():
    with torch.no_grad():
        with open('../SynthText/alphabet-all-v3.txt', 'r') as fp:
            dictionary = [s[:-1] for s in fp.readlines()]
        img = torch.zeros(1, 3, 32, 128)
        model = OCR(dictionary, 32)
        m = torch.load("ocr_ar_v2-3-test.ckpt", map_location='cpu')
        model.load_state_dict(m['model'])
        model.eval()
        (char_probs, _, _, _, _, _, _, _), _ = model.infer_beam(img, max_seq_length = 20)
        _, pred_chars_index = char_probs.max(2)
        pred_chars_index = pred_chars_index.squeeze_(0)
        seq = []
        for chid in pred_chars_index:
            ch = dictionary[chid]
            if ch == '<SP>':
                ch == ' '
            seq.append(ch)
        print(''.join(seq))

if __name__ == "__main__":
    test()