backend/python-services/services/loc_service.py

"""
Location detection service.
Detects text regions and classifies them into 13 categories.

Uses DB_gc_loc architecture: ResNet-18 backbone + SegDetector decoder.
Supports both TorchScript (.pt) models and state_dict checkpoints.
"""

import os
import math
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import cv2
import logging
from collections import OrderedDict

from predictors.torchscript_predictor import resolve_model_path
from common.image_utils import array_from_image_stream


# RGB mean for normalization (from original implementation)
RGB_MEAN = np.array([122.67891434, 116.66876762, 104.00698793])

# Text type categories
TYPE_NAMES = [
	'Title',			# 0
	'Author',			# 1
	'TextualMark',		# 2
	'TempoNumeral',		# 3
	'MeasureNumber',	# 4
	'Times',			# 5
	'Chord',			# 6
	'PageMargin',		# 7
	'Instrument',		# 8
	'Other',			# 9
	'Lyric',			# 10
	'Alter1',			# 11
	'Alter2',			# 12
]


# ===================== ResNet-18 Backbone =====================

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


class _BasicBlock(nn.Module):
	expansion = 1

	def __init__(self, inplanes, planes, stride=1, downsample=None):
		super().__init__()
		self.conv1 = _conv3x3(inplanes, planes, stride)
		self.bn1 = nn.BatchNorm2d(planes)
		self.relu = nn.ReLU(inplace=True)
		self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, padding=1, bias=False)
		self.bn2 = nn.BatchNorm2d(planes)
		self.downsample = downsample

	def forward(self, x):
		residual = x
		out = self.relu(self.bn1(self.conv1(x)))
		out = self.bn2(self.conv2(out))
		if self.downsample is not None:
			residual = self.downsample(x)
		out += residual
		return self.relu(out)


class _ResNet(nn.Module):
	def __init__(self, block, layers):
		super().__init__()
		self.inplanes = 64
		self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=False)
		self.bn1 = nn.BatchNorm2d(64)
		self.relu = nn.ReLU(inplace=True)
		self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
		self.layer1 = self._make_layer(block, 64, layers[0])
		self.layer2 = self._make_layer(block, 128, layers[1], stride=2)
		self.layer3 = self._make_layer(block, 256, layers[2], stride=2)
		self.layer4 = self._make_layer(block, 512, layers[3], stride=2)
		# Not used in forward but exist in original model (needed for weight loading)
		self.avgpool = nn.AvgPool2d(7, stride=1)
		self.fc = nn.Linear(512, 1000)
		self.smooth = nn.Conv2d(2048, 256, kernel_size=1, stride=1, padding=1)

	def _make_layer(self, block, planes, blocks, stride=1):
		downsample = None
		if stride != 1 or self.inplanes != planes * block.expansion:
			downsample = nn.Sequential(
				nn.Conv2d(self.inplanes, planes * block.expansion,
						  kernel_size=1, stride=stride, bias=False),
				nn.BatchNorm2d(planes * block.expansion),
			)
		layers = [block(self.inplanes, planes, stride, downsample)]
		self.inplanes = planes * block.expansion
		for _ in range(1, blocks):
			layers.append(block(self.inplanes, planes))
		return nn.Sequential(*layers)

	def forward(self, x):
		x = self.maxpool(self.relu(self.bn1(self.conv1(x))))
		x2 = self.layer1(x)
		x3 = self.layer2(x2)
		x4 = self.layer3(x3)
		x5 = self.layer4(x4)
		return x2, x3, x4, x5


# ===================== SegDetector Decoder =====================

class _SegDetector(nn.Module):
	def __init__(self, n_cls=13, in_channels=(64, 128, 256, 512),
				 inner_channels=256, k=50, bias=False, adaptive=True):
		super().__init__()
		self.k = k
		self.up5 = nn.Upsample(scale_factor=2, mode='nearest')
		self.up4 = nn.Upsample(scale_factor=2, mode='nearest')
		self.up3 = nn.Upsample(scale_factor=2, mode='nearest')

		self.in5 = nn.Conv2d(in_channels[-1], inner_channels, 1, bias=bias)
		self.in4 = nn.Conv2d(in_channels[-2], inner_channels, 1, bias=bias)
		self.in3 = nn.Conv2d(in_channels[-3], inner_channels, 1, bias=bias)
		self.in2 = nn.Conv2d(in_channels[-4], inner_channels, 1, bias=bias)

		self.out5 = nn.Sequential(
			nn.Conv2d(inner_channels, inner_channels // 4, 3, padding=1, bias=bias),
			nn.Upsample(scale_factor=8, mode='nearest'))
		self.out4 = nn.Sequential(
			nn.Conv2d(inner_channels, inner_channels // 4, 3, padding=1, bias=bias),
			nn.Upsample(scale_factor=4, mode='nearest'))
		self.out3 = nn.Sequential(
			nn.Conv2d(inner_channels, inner_channels // 4, 3, padding=1, bias=bias),
			nn.Upsample(scale_factor=2, mode='nearest'))
		self.out2 = nn.Conv2d(inner_channels, inner_channels // 4, 3, padding=1, bias=bias)

		self.binarize = nn.Sequential(
			nn.Conv2d(inner_channels, inner_channels // 4, 3, padding=1, bias=bias),
			nn.BatchNorm2d(inner_channels // 4),
			nn.ReLU(inplace=True),
			nn.ConvTranspose2d(inner_channels // 4, inner_channels // 4, 2, 2),
			nn.BatchNorm2d(inner_channels // 4),
			nn.ReLU(inplace=True),
			nn.ConvTranspose2d(inner_channels // 4, 1, 2, 2),
			nn.Sigmoid())

		self.mulclass = nn.Sequential(
			nn.Conv2d(inner_channels, inner_channels // 4, 3, padding=1, bias=bias),
			nn.BatchNorm2d(inner_channels // 4),
			nn.ReLU(inplace=True),
			nn.ConvTranspose2d(inner_channels // 4, inner_channels // 4, 2, 2),
			nn.BatchNorm2d(inner_channels // 4),
			nn.ReLU(inplace=True),
			nn.ConvTranspose2d(inner_channels // 4, n_cls, 2, 2))

		if adaptive:
			self.thresh = nn.Sequential(
				nn.Conv2d(inner_channels, inner_channels // 4, 3, padding=1, bias=bias),
				nn.BatchNorm2d(inner_channels // 4),
				nn.ReLU(inplace=True),
				nn.ConvTranspose2d(inner_channels // 4, inner_channels // 4, 2, 2),
				nn.BatchNorm2d(inner_channels // 4),
				nn.ReLU(inplace=True),
				nn.ConvTranspose2d(inner_channels // 4, 1, 2, 2),
				nn.Sigmoid())

	def forward(self, features):
		c2, c3, c4, c5 = features
		in5 = self.in5(c5)
		in4 = self.in4(c4)
		in3 = self.in3(c3)
		in2 = self.in2(c2)

		out4 = self.up5(in5) + in4
		out3 = self.up4(out4) + in3
		out2 = self.up3(out3) + in2

		p5 = self.out5(in5)
		p4 = self.out4(out4)
		p3 = self.out3(out3)
		p2 = self.out2(out2)

		fuse = torch.cat((p5, p4, p3, p2), 1)
		binary = self.binarize(fuse)
		mcls = self.mulclass(fuse)
		return binary, mcls


class _LocModel(nn.Module):
	"""ResNet-18 + SegDetector combined model for text localization."""

	def __init__(self):
		super().__init__()
		self.backbone = _ResNet(_BasicBlock, [2, 2, 2, 2])
		self.decoder = _SegDetector(n_cls=13, adaptive=True, k=50)

	def forward(self, x):
		return self.decoder(self.backbone(x))


def _load_loc_model(model_path, device):
	"""Load loc model, handling both TorchScript and state_dict formats.

	First tries TorchScript loading; on failure, builds architecture and loads state_dict.
	"""
	resolved = resolve_model_path(model_path)

	# Try TorchScript first
	try:
		model = torch.jit.load(resolved, map_location=device)
		model.eval()
		logging.info('LocService: TorchScript model loaded: %s', resolved)
		return model
	except Exception as e:
		logging.info('LocService: not TorchScript (%s), trying state_dict...', str(e)[:60])

	# Fall back to state_dict loading with embedded architecture
	checkpoint = torch.load(resolved, map_location=device, weights_only=False)

	if isinstance(checkpoint, dict) and 'model' in checkpoint:
		state_dict = checkpoint['model']
	elif isinstance(checkpoint, OrderedDict):
		state_dict = checkpoint
	else:
		state_dict = checkpoint

	# Strip 'model.module.' prefix (SegDetectorModel → DataParallel wrapping)
	new_state_dict = OrderedDict()
	for key, value in state_dict.items():
		new_key = key
		if new_key.startswith('model.module.'):
			new_key = new_key[len('model.module.'):]
		elif new_key.startswith('module.'):
			new_key = new_key[len('module.'):]
		new_state_dict[new_key] = value

	model = _LocModel()
	model.load_state_dict(new_state_dict, strict=False)
	model.eval()
	model.to(device)

	# Log key loading stats
	model_keys = set(model.state_dict().keys())
	loaded_keys = set(new_state_dict.keys())
	matched = model_keys & loaded_keys
	logging.info('LocService: state_dict loaded: %s (%d/%d keys matched)',
				 resolved, len(matched), len(model_keys))

	return model


class LocService:
	"""
	Location detection service.

	Uses DB_gc_loc architecture (ResNet-18 + SegDetector).
	Supports both TorchScript models and state_dict checkpoints.
	"""

	def __init__(self, model_path, device='cuda', image_short_side=736,
				 box_thresh=0.01, class_num=13, **kwargs):
		self.device = device
		self.model = _load_loc_model(model_path, device)
		self.image_short_side = image_short_side
		self.box_thresh = box_thresh
		self.class_num = class_num

	def resize_image(self, img):
		"""Resize image keeping aspect ratio, with short side = image_short_side."""
		height, width = img.shape[:2]
		if height < width:
			new_height = self.image_short_side
			new_width = int(math.ceil(new_height / height * width / 32) * 32)
		else:
			new_width = self.image_short_side
			new_height = int(math.ceil(new_width / width * height / 32) * 32)
		return cv2.resize(img, (new_width, new_height))

	def preprocess(self, image):
		"""Preprocess image for model input."""
		img = image.astype('float32')
		original_shape = img.shape[:2]
		img = self.resize_image(img)
		img -= RGB_MEAN
		img /= 255.
		img = torch.from_numpy(img).permute(2, 0, 1).float().unsqueeze(0)
		return img.to(self.device), original_shape

	def represent_boxes(self, pred, out_class, original_shape, resized_shape):
		"""Post-process model output to extract bounding boxes."""
		pred_np = pred.cpu().numpy()[0, 0]
		class_np = out_class.cpu().numpy()[0, 0]

		binary = (pred_np > self.box_thresh).astype(np.uint8) * 255
		contours, _ = cv2.findContours(binary, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)

		boxes = []
		h_scale = original_shape[0] / resized_shape[0]
		w_scale = original_shape[1] / resized_shape[1]

		for contour in contours:
			if len(contour) < 4:
				continue

			rect = cv2.minAreaRect(contour)
			box_points = cv2.boxPoints(rect)
			box_points = np.int0(box_points)

			mask = np.zeros(pred_np.shape, dtype=np.uint8)
			cv2.drawContours(mask, [contour], -1, 1, -1)
			class_region = class_np * mask
			if mask.sum() > 0:
				box_class = int(np.argmax(np.bincount(class_region[mask > 0].astype(int))))
			else:
				box_class = 0

			score_region = pred_np * mask
			score = score_region.sum() / max(mask.sum(), 1)

			scaled_points = box_points.astype(float)
			scaled_points[:, 0] *= w_scale
			scaled_points[:, 1] *= h_scale

			boxes.append({
				'x0': float(scaled_points[0, 0]),
				'y0': float(scaled_points[0, 1]),
				'x1': float(scaled_points[1, 0]),
				'y1': float(scaled_points[1, 1]),
				'x2': float(scaled_points[2, 0]),
				'y2': float(scaled_points[2, 1]),
				'x3': float(scaled_points[3, 0]),
				'y3': float(scaled_points[3, 1]),
				'score': float(score),
				'class': box_class,
			})

		return boxes

	def predict(self, buffers, **kwargs):
		"""Detect text regions in images."""
		for buffer in buffers:
			image = array_from_image_stream(buffer)
			if image is None:
				yield []
				continue

			img_tensor, original_shape = self.preprocess(image)
			resized_shape = (img_tensor.shape[2], img_tensor.shape[3])

			with torch.no_grad():
				output = self.model(img_tensor)

				if isinstance(output, tuple) and len(output) == 2:
					pred, mcls = output
				else:
					pred = output
					mcls = torch.zeros_like(pred)

				b, c, h, w = mcls.shape
				out_class = mcls.permute(0, 2, 3, 1).reshape(-1, self.class_num)
				out_class = F.softmax(out_class, -1)
				out_class = out_class.max(1)[1].reshape(b, h, w).unsqueeze(1)

				boxes = self.represent_boxes(pred, out_class, original_shape, resized_shape)

				valid_boxes = []
				for box in boxes:
					if not (box['x0'] == box['x1'] and box['x2'] == box['x1']):
						valid_boxes.append(box)

				yield valid_boxes