modeling.py

"""Model definition. Required to load the weights."""

import numpy as np
import torch
import torch.nn as nn
from PIL import Image

HEIGHT = 48
DOWNSAMPLE = 4      # width is divided by 4; T = width // 4


class CRNN(nn.Module):
    def __init__(self, nclass, hidden=256, layers=2, dropout=0.1):
        super().__init__()

        def blk(i, o, pool):
            m = [nn.Conv2d(i, o, 3, 1, 1, bias=False), nn.BatchNorm2d(o),
                 nn.ReLU(inplace=True)]
            if pool:
                m.append(nn.MaxPool2d(pool, pool))
            return m

        self.cnn = nn.Sequential(
            *blk(1, 64, (2, 2)),
            *blk(64, 128, (2, 2)),
            *blk(128, 256, None),
            *blk(256, 256, (2, 1)),
            *blk(256, 512, None),
            *blk(512, 512, (2, 1)),
            nn.Conv2d(512, 512, (3, 1), 1, 0, bias=False),
            nn.BatchNorm2d(512), nn.ReLU(inplace=True),
        )
        self.rnn = nn.LSTM(512, hidden, num_layers=layers, bidirectional=True,
                           batch_first=True, dropout=dropout if layers > 1 else 0)
        self.head = nn.Linear(hidden * 2, nclass)

    def forward(self, x):
        f = self.cnn(x).squeeze(2).permute(0, 2, 1)
        f, _ = self.rnn(f)
        return self.head(f)


def preprocess(img, max_width=1200):
    """PIL Image -> tensor (1,1,48,W). Input is a crop of ONE text line."""
    img = img.convert("L")
    if img.height != HEIGHT:
        r = HEIGHT / img.height
        img = img.resize((max(8, int(img.width * r)), HEIGHT), Image.BILINEAR)
    if img.width > max_width:
        img = img.crop((0, 0, max_width, HEIGHT))
    x = (np.array(img, dtype=np.float32) / 255.0 - 0.5) / 0.5
    return torch.from_numpy(x)[None, None]


def ctc_greedy(logits, itos, blank=0):
    ids = logits.argmax(-1)[0].tolist()
    prev, out = -1, []
    for k in ids:
        if k != prev and k != blank:
            out.append(itos[k])
        prev = k
    return "".join(out)


def load(path="model.pt", device="cpu"):
    ck = torch.load(path, map_location=device)
    cfg = ck.get("cfg", {})
    itos = ck["itos"]
    m = CRNN(len(itos), cfg.get("rnn_hidden", 256), cfg.get("rnn_layers", 2), 0.0)
    m.load_state_dict(ck["model"])
    return m.eval().to(device), itos