pokouqiege/models/line_detect/heads/arc/arc_heads.py

import torch
from torch import nn

class ArcHeads(nn.Sequential):
    def __init__(self, in_channels, layers):
        d = []
        next_feature = in_channels
        for out_channels in layers:
            d.append(nn.Conv2d(next_feature, out_channels, 3, stride=1, padding=1))
            d.append(nn.ReLU(inplace=True))
            next_feature = out_channels
        super().__init__(*d)
        for m in self.children():
            if isinstance(m, nn.Conv2d):
                nn.init.kaiming_normal_(m.weight, mode="fan_out", nonlinearity="relu")
                nn.init.constant_(m.bias, 0)


class ArcPredictor(nn.Module):
    def __init__(self, in_channels, out_channels=1 ):
        super().__init__()
        input_features = in_channels
        deconv_kernel = 4
        self.kps_score_lowres = nn.ConvTranspose2d(
            input_features,
            out_channels,
            deconv_kernel,
            stride=2,
            padding=deconv_kernel // 2 - 1,
        )
        nn.init.kaiming_normal_(self.kps_score_lowres.weight, mode="fan_out", nonlinearity="relu")
        nn.init.constant_(self.kps_score_lowres.bias, 0)
        self.up_scale = 2
        self.out_channels = out_channels

    def forward(self, x):
        print(f'before kps_score_lowres x:{x.shape}')
        x = self.kps_score_lowres(x)
        print(f'kps_score_lowres x:{x.shape}')
        return x
        # return torch.nn.functional.interpolate(
        #     x, scale_factor=float(self.up_scale), mode="bilinear", align_corners=False, recompute_scale_factor=False
        # )




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


class ArcEquationHead(nn.Module):
    """
    Input:
        feature_logits : [N, 1, H, W]

    Output:
        arc_params : [N, 7]
            0: center x (cx)
            1: center y (cy)
            2: long axis length (a)
            3: short axis length (b)
            4: ellipse angle (theta)
            5: auxiliary x coordinate
            6: auxiliary y coordinate
    """

    def __init__(self, num_outputs=9, hidden=512):
        super().__init__()

        # Use GAP to remove spatial dependency
        self.gap = nn.AdaptiveAvgPool2d((1, 1))

        # Final MLP that maps pooled feature ¡ú arc parameters
        self.mlp = nn.Sequential(
            nn.Linear(1, hidden),
            nn.ReLU(inplace=True),
            nn.Linear(hidden, num_outputs)
        )

    def forward(self, feature_logits):
        """
        feature_logits: [N, 1, H, W]
        """
        N, _, H, W = feature_logits.shape

        # --------------------------------------------
        # Global average pooling
        # Input  : [N, 1, H, W]
        # Output : [N, 1]
        # --------------------------------------------
        x = self.gap(feature_logits)
        x = x.view(N, -1)

        # Predict raw parameters
        arc_params = self.mlp(x)   # [N, 7]

        # --------------------------------------------
        # Parameter constraints
        # --------------------------------------------
        # H=1500
        # W = 2000

        # Ellipse center
        arc_params[..., 0] = torch.sigmoid(arc_params[..., 0]) * W   # cx in image width range
        arc_params[..., 1] = torch.sigmoid(arc_params[..., 1]) * H   # cy in image height range


        # Axes lengths must be positive
        arc_params[..., 2] = torch.sigmoid(arc_params[..., 2]) * W  # cx in image width range
        arc_params[..., 3] = torch.sigmoid(arc_params[..., 3]) * W  # cy in image height range
        # arc_params[..., 2] = F.relu(arc_params[..., 2]) + 1e-6       # a > 0
        # arc_params[..., 3] = F.relu(arc_params[..., 3]) + 1e-6       # b > 0

        # Angle between 0~2¦Ð
        arc_params[..., 4] = torch.sigmoid(arc_params[..., 4]) * (2 * 3.1415926535)

        # ------------------------------------------------
        # Last two values are auxiliary points
        # Now mapped to the same spatial range as image
        # ------------------------------------------------
        arc_params[..., 5] = torch.sigmoid(arc_params[..., 5]) * W   # x auxiliary
        arc_params[..., 6] = torch.sigmoid(arc_params[..., 6]) * H   # y auxiliary
        arc_params[..., 7] = torch.sigmoid(arc_params[..., 7]) * W  # x auxiliary
        arc_params[..., 8] = torch.sigmoid(arc_params[..., 8]) * H  # y auxiliary

        print(f'arc_params in head:{arc_params}')

        return arc_params



# class ArcEquationHead(nn.Module):
#     """
#     Input:
#         feature_logits : [N, 1, H, W]  # N:Ô²»¡Êý£¬H/W:ÌØÕ÷ͼ³ß´ç£¨¶ÔӦԭʼͼÏñ¿Õ¼äλÖã©
#     Output:
#         arc_params : [N, 9]  # [cx, cy, a, b, theta, x1, y1, x2, y2]
#     """
#
#     def __init__(self, num_outputs=9, hidden=1024, feat_size=(672, 672)):
#         super().__init__()
#         self.feat_H, self.feat_W = feat_size  # ÌØÕ÷ͼµÄ¹Ì¶¨³ß´ç£¨ÐèÓëfeature_logitsÒ»Ö£©
#
#         self.flatten = nn.Flatten()
#         self.input_dim = self.feat_H * self.feat_W  # ÊäÈëά¶È£ºH*W£¨¶ø·Ç1£©
#
#         self.mlp = nn.Sequential(
#             nn.Linear(self.input_dim, hidden),
#             nn.ReLU(inplace=True),
#             nn.Dropout(0.2),  # ·ÀÖ¹¹ýÄâºÏ
#             nn.Linear(hidden, hidden // 2),
#             nn.ReLU(inplace=True),
#             nn.Dropout(0.1),
#             nn.Linear(hidden // 2, num_outputs)
#         )
#
#         self._init_weights()
#
#     def _init_weights(self):
#         for m in self.mlp.modules():
#             if isinstance(m, nn.Linear):
#                 nn.init.xavier_uniform_(m.weight)  # ¾ùÔȳõʼ»¯£¬±ÜÃâÊä³ö¼¯ÖÐ
#                 if m.bias is not None:
#                     nn.init.zeros_(m.bias)
#
#     def forward(self, feature_logits):
#         N, C, H, W = feature_logits.shape
#         assert H == self.feat_H and W == self.feat_W, "ÌØÕ÷ͼ³ß´çÐèÓë³õʼ»¯Ê±µÄfeat_sizeÒ»ÖÂ"
#
#         # 1. Flatten¿Õ¼äÌØÕ÷£º[N,1,H,W] ¡ú [N, H*W]£¨±£Áôÿ¸öÏñËصĿռäÐÅÏ¢£©
#         x = self.flatten(feature_logits)  # [N, H*W]
#
#         # 2. MLPÔ¤²âԭʼ²ÎÊý
#         arc_params = self.mlp(x)  # [N,9]
#
#         # 3. ÓÅ»¯²ÎÊýÔ¼Êø£¨±ÜÃâÖÐÐÄ/Ö᳤Òì³££©
#         # ÍÖÔ²ÖÐÐÄ£ºÓ³Éäµ½ÌØÕ÷ͼ³ß´ç£¨ÈôÌØÕ÷ͼÊÇԭʼͼÏñϲÉÑù£¬Ðè³ËÒÔËõ·ÅÒò×Ó£©
#         arc_params[..., 0] = torch.sigmoid(arc_params[..., 0]) * W  # cx
#         arc_params[..., 1] = torch.sigmoid(arc_params[..., 1]) * H  # cy
#
#         arc_params[..., 2] = torch.sigmoid(arc_params[..., 2]) * W  # cx in image width range
#         arc_params[..., 3] = torch.sigmoid(arc_params[..., 3]) * H  # cy in image height range
#         # arc_params[..., 2] = F.relu(arc_params[..., 2]) + 1e-6       # a > 0
#         # arc_params[..., 3] = F.relu(arc_params[..., 3]) + 1e-6       # b > 0
#
#         # Angle between 0~2¦Ð
#         arc_params[..., 4] = torch.sigmoid(arc_params[..., 4]) * (2 * 3.1415926535)
#
#         # ------------------------------------------------
#         # Last two values are auxiliary points
#         # Now mapped to the same spatial range as image
#         # ------------------------------------------------
#         arc_params[..., 5] = torch.sigmoid(arc_params[..., 5]) * W   # x auxiliary
#         arc_params[..., 6] = torch.sigmoid(arc_params[..., 6]) * H   # y auxiliary
#         arc_params[..., 7] = torch.sigmoid(arc_params[..., 7]) * W  # x auxiliary
#         arc_params[..., 8] = torch.sigmoid(arc_params[..., 8]) * H  # y auxiliary
#         print(f'arc_params in head:{arc_params}')
#         return arc_params