Add model architecture code

modeling_interfuser.py

@@ -0,0 +1,579 @@
+# modeling_interfuser.py
+
+import torch
+from torch import nn
+import torch.nn.functional as F
+from transformers import PreTrainedModel, PretrainedConfig
+from transformers.utils.generic import ModelOutput
+from functools import partial
+import math
+from collections import OrderedDict
+import copy
+from typing import Optional, List, Tuple, Union
+from torch import Tensor
+from dataclasses import dataclass
+import numpy as np
+
+# ==============================================================================
+# ملاحظة: هذا الملف يحتوي على كل التعريفات اللازمة لتشغيل النموذج.
+# ==============================================================================
+
+# --- الكلاسات الوهمية للـ Backbones ---
+# في الاستخدام الحقيقي، يجب استبدالها بالشبكات الحقيقية من مكتبة مثل timm.
+class DummyResNet(nn.Module):
+    def __init__(self, name="r26", **kwargs):
+        super().__init__()
+        out_channels = 512 if name == "r18" else 2048
+        self.features = nn.Sequential(
+            nn.Conv2d(kwargs.get('in_chans', 3), out_channels, kernel_size=7, stride=2, padding=3),
+            nn.AdaptiveAvgPool2d((1, 1))
+        )
+        self.num_features = out_channels
+    def forward(self, x):
+        return [self.features(x)]
+
+def resnet18d(**kwargs): return DummyResNet(name="r18", **kwargs)
+def resnet26d(**kwargs): return DummyResNet(name="r26", **kwargs)
+def resnet50d(**kwargs): return DummyResNet(name="r50", **kwargs)
+def to_2tuple(x): return (x, x) if not isinstance(x, tuple) else x
+
+# --- جميع الكلاسات المساعدة ---
+# (HybridEmbed, PositionEmbeddingSine, TransformerEncoder, SpatialSoftmax, etc.)
+# تم نسخها بالكامل هنا.
+class HybridEmbed(nn.Module):
+    def __init__(self, backbone, img_size=224, patch_size=1, feature_size=None, in_chans=3, embed_dim=768):
+        super().__init__()
+        self.img_size = to_2tuple(img_size)
+        self.patch_size = to_2tuple(patch_size)
+        self.backbone = backbone
+        if feature_size is None:
+            with torch.no_grad():
+                training = backbone.training
+                if training: backbone.eval()
+                o = self.backbone(torch.zeros(1, in_chans, img_size[0], img_size[1]))
+                if isinstance(o, (list, tuple)): o = o[-1]
+                feature_dim = o.shape[1]
+                backbone.train(training)
+        else:
+            feature_dim = self.backbone.num_features
+        self.proj = nn.Conv2d(feature_dim, embed_dim, kernel_size=1, stride=1)
+    def forward(self, x):
+        x = self.backbone(x)
+        if isinstance(x, (list, tuple)): x = x[-1]
+        x = self.proj(x)
+        global_x = torch.mean(x, [2, 3], keepdim=False)[:, :, None]
+        return x, global_x
+
+# ... (يتم لصق بقية الكلاسات المساعدة هنا: PositionEmbeddingSine, Transformer*...)
+# (للاختصار، لن أعرضها كلها مرة أخرى، ولكن يجب أن تكون كلها في هذا الملف)
+class PositionEmbeddingSine(nn.Module):
+    def __init__( self, num_pos_feats=64, temperature=10000, normalize=False, scale=None):
+        super().__init__()
+        self.num_pos_feats = num_pos_feats
+        self.temperature = temperature
+        self.normalize = normalize
+        if scale is not None and normalize is False: raise ValueError("normalize should be True if scale is passed")
+        if scale is None: scale = 2 * math.pi
+        self.scale = scale
+    def forward(self, tensor):
+        x = tensor; bs, _, h, w = x.shape
+        not_mask = torch.ones((bs, h, w), device=x.device)
+        y_embed = not_mask.cumsum(1, dtype=torch.float32)
+        x_embed = not_mask.cumsum(2, dtype=torch.float32)
+        if self.normalize:
+            eps = 1e-6
+            y_embed = y_embed / (y_embed[:, -1:, :] + eps) * self.scale
+            x_embed = x_embed / (x_embed[:, :, -1:] + eps) * self.scale
+        dim_t = torch.arange(self.num_pos_feats, dtype=torch.float32, device=x.device)
+        dim_t = self.temperature ** (2 * (dim_t // 2) / self.num_pos_feats)
+        pos_x = x_embed[:, :, :, None] / dim_t; pos_y = y_embed[:, :, :, None] / dim_t
+        pos_x = torch.stack((pos_x[:, :, :, 0::2].sin(), pos_x[:, :, :, 1::2].cos()), dim=4).flatten(3)
+        pos_y = torch.stack((pos_y[:, :, :, 0::2].sin(), pos_y[:, :, :, 1::2].cos()), dim=4).flatten(3)
+        pos = torch.cat((pos_y, pos_x), dim=3).permute(0, 3, 1, 2)
+        return pos
+# (لصق باقي الكلاسات المساعدة هنا)
+class TransformerEncoder(nn.Module):
+    def __init__(self, encoder_layer, num_layers, norm=None):
+        super().__init__()
+        self.layers = _get_clones(encoder_layer, num_layers)
+        self.num_layers = num_layers
+        self.norm = norm
+
+    def forward(
+        self,
+        src,
+        mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+        pos: Optional[Tensor] = None,
+    ):
+        output = src
+        for layer in self.layers:
+            output = layer(
+                output,
+                src_mask=mask,
+                src_key_padding_mask=src_key_padding_mask,
+                pos=pos,
+            )
+        if self.norm is not None:
+            output = self.norm(output)
+        return output
+
+
+class SpatialSoftmax(nn.Module):
+    def __init__(self, height, width, channel, temperature=None, data_format="NCHW"):
+        super().__init__()
+        self.data_format = data_format
+        self.height = height
+        self.width = width
+        self.channel = channel
+        if temperature:
+            self.temperature = nn.Parameter(torch.ones(1) * temperature)
+        else:
+            self.temperature = 1.0
+        pos_x, pos_y = np.meshgrid(
+            np.linspace(-1.0, 1.0, self.height), np.linspace(-1.0, 1.0, self.width)
+        )
+        pos_x = torch.from_numpy(pos_x.reshape(self.height * self.width)).float()
+        pos_y = torch.from_numpy(pos_y.reshape(self.height * self.width)).float()
+        self.register_buffer("pos_x", pos_x)
+        self.register_buffer("pos_y", pos_y)
+
+    def forward(self, feature):
+        if self.data_format == "NHWC":
+            feature = (
+                feature.transpose(1, 3)
+                .transpose(2, 3)
+                .view(-1, self.height * self.width)
+            )
+        else:
+            feature = feature.view(-1, self.height * self.width)
+        weight = F.softmax(feature / self.temperature, dim=-1)
+        expected_x = torch.sum(self.pos_x * weight, dim=1, keepdim=True)
+        expected_y = torch.sum(self.pos_y * weight, dim=1, keepdim=True)
+        expected_xy = torch.cat([expected_x, expected_y], 1)
+        feature_keypoints = expected_xy.view(-1, self.channel, 2)
+        feature_keypoints[:, :, 1] = (feature_keypoints[:, :, 1] - 1) * 12
+        feature_keypoints[:, :, 0] = feature_keypoints[:, :, 0] * 12
+        return feature_keypoints
+
+# ... (بقية الكلاسات المساعدة مثل MultiPath_Generator, LinearWaypointsPredictor, etc. توضع هنا)
+class MultiPath_Generator(nn.Module):
+    def __init__(self, in_channel, embed_dim, out_channel):
+        super().__init__()
+        self.spatial_softmax = SpatialSoftmax(100, 100, out_channel)
+        self.tconv0 = nn.Sequential(nn.ConvTranspose2d(in_channel, 256, 4, 2, 1, bias=False), nn.BatchNorm2d(256), nn.ReLU(True))
+        self.tconv1 = nn.Sequential(nn.ConvTranspose2d(256, 256, 4, 2, 1, bias=False), nn.BatchNorm2d(256), nn.ReLU(True))
+        self.tconv2 = nn.Sequential(nn.ConvTranspose2d(256, 192, 4, 2, 1, bias=False), nn.BatchNorm2d(192), nn.ReLU(True))
+        self.tconv3 = nn.Sequential(nn.ConvTranspose2d(192, 64, 4, 2, 1, bias=False), nn.BatchNorm2d(64), nn.ReLU(True))
+        self.tconv4_list = torch.nn.ModuleList([nn.Sequential(nn.ConvTranspose2d(64, out_channel, 8, 2, 3, bias=False), nn.Tanh()) for _ in range(6)])
+        self.upsample = nn.Upsample(size=(50, 50), mode="bilinear")
+
+    def forward(self, x, measurements):
+        mask = measurements[:, :6].unsqueeze(-1).unsqueeze(-1).unsqueeze(-1).repeat(1, 1, 1, 100, 100)
+        velocity = measurements[:, 6:7].unsqueeze(-1).unsqueeze(-1).repeat(1, 32, 2, 2)
+        n, d, c = x.shape
+        x = x.transpose(1, 2).view(n, -1, 2, 2)
+        x = torch.cat([x, velocity], dim=1)
+        x = self.tconv0(x); x = self.tconv1(x); x = self.tconv2(x); x = self.tconv3(x)
+        x = self.upsample(x)
+        xs = torch.stack([self.tconv4_list[i](x) for i in range(6)], dim=1)
+        x = torch.sum(xs * mask, dim=1)
+        return self.spatial_softmax(x)
+
+class LinearWaypointsPredictor(nn.Module):
+    def __init__(self, input_dim, cumsum=True):
+        super().__init__()
+        self.cumsum = cumsum
+        self.rank_embed = nn.Parameter(torch.zeros(1, 10, input_dim))
+        self.head_fc1_list = nn.ModuleList([nn.Linear(input_dim, 64) for _ in range(6)])
+        self.head_relu = nn.ReLU(inplace=True)
+        self.head_fc2_list = nn.ModuleList([nn.Linear(64, 2) for _ in range(6)])
+
+    def forward(self, x, measurements):
+        bs, n, dim = x.shape
+        x = (x + self.rank_embed).reshape(-1, dim)
+        mask = measurements[:, :6].unsqueeze(-1).repeat(n, 1, 2)
+        rs = [self.head_fc2_list[i](self.head_relu(self.head_fc1_list[i](x))) for i in range(6)]
+        x = torch.sum(torch.stack(rs, 1) * mask, dim=1).view(bs, n, 2)
+        return torch.cumsum(x, 1) if self.cumsum else x
+
+class GRUWaypointsPredictor(nn.Module):
+    def __init__(self, input_dim, waypoints=10):
+        super().__init__()
+        self.gru = torch.nn.GRU(input_size=input_dim, hidden_size=64, batch_first=True)
+        self.encoder = nn.Linear(2, 64)
+        self.decoder = nn.Linear(64, 2)
+        self.waypoints = waypoints
+
+    def forward(self, x, target_point):
+        bs = x.shape[0]
+        z = self.encoder(target_point).unsqueeze(0)
+        output, _ = self.gru(x, z)
+        output = self.decoder(output.reshape(bs * self.waypoints, -1)).reshape(bs, self.waypoints, 2)
+        return torch.cumsum(output, 1)
+
+class GRUWaypointsPredictorWithCommand(nn.Module):
+    def __init__(self, input_dim, waypoints=10):
+        super().__init__()
+        self.grus = nn.ModuleList([torch.nn.GRU(input_size=input_dim, hidden_size=64, batch_first=True) for _ in range(6)])
+        self.encoder = nn.Linear(2, 64)
+        self.decoders = nn.ModuleList([nn.Linear(64, 2) for _ in range(6)])
+        self.waypoints = waypoints
+
+    def forward(self, x, target_point, measurements):
+        bs, n, dim = x.shape
+        mask = measurements[:, :6, None, None].repeat(1, 1, self.waypoints, 2)
+        z = self.encoder(target_point).unsqueeze(0)
+        outputs = []
+        for i in range(6):
+            output, _ = self.grus[i](x, z)
+            output = self.decoders[i](output.reshape(bs * self.waypoints, -1)).reshape(bs, self.waypoints, 2)
+            outputs.append(torch.cumsum(output, 1))
+        return torch.sum(torch.stack(outputs, 1) * mask, dim=1)
+
+class TransformerDecoder(nn.Module):
+    def __init__(self, decoder_layer, num_layers, norm=None, return_intermediate=False):
+        super().__init__()
+        self.layers = _get_clones(decoder_layer, num_layers)
+        self.num_layers = num_layers
+        self.norm = norm
+        self.return_intermediate = return_intermediate
+
+    def forward(self, tgt, memory, tgt_mask: Optional[Tensor] = None, memory_mask: Optional[Tensor] = None, tgt_key_padding_mask: Optional[Tensor] = None, memory_key_padding_mask: Optional[Tensor] = None, pos: Optional[Tensor] = None, query_pos: Optional[Tensor] = None):
+        output = tgt
+        intermediate = []
+        for layer in self.layers:
+            output = layer(output, memory, tgt_mask=tgt_mask, memory_mask=memory_mask, tgt_key_padding_mask=tgt_key_padding_mask, memory_key_padding_mask=memory_key_padding_mask, pos=pos, query_pos=query_pos)
+            if self.return_intermediate: intermediate.append(self.norm(output))
+        if self.norm is not None:
+            output = self.norm(output)
+            if self.return_intermediate: intermediate.pop(); intermediate.append(output)
+        return torch.stack(intermediate) if self.return_intermediate else output.unsqueeze(0)
+
+class TransformerEncoderLayer(nn.Module):
+    def __init__(self, d_model, nhead, dim_feedforward=2048, dropout=0.1, activation=nn.ReLU, normalize_before=False):
+        super().__init__()
+        self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=dropout)
+        self.linear1 = nn.Linear(d_model, dim_feedforward); self.dropout = nn.Dropout(dropout); self.linear2 = nn.Linear(dim_feedforward, d_model)
+        self.norm1 = nn.LayerNorm(d_model); self.norm2 = nn.LayerNorm(d_model)
+        self.dropout1 = nn.Dropout(dropout); self.dropout2 = nn.Dropout(dropout)
+        self.activation = activation(); self.normalize_before = normalize_before
+
+    def with_pos_embed(self, tensor, pos: Optional[Tensor]): return tensor if pos is None else tensor + pos
+    def forward_post(self, src, src_mask: Optional[Tensor] = None, src_key_padding_mask: Optional[Tensor] = None, pos: Optional[Tensor] = None):
+        q = k = self.with_pos_embed(src, pos)
+        src2 = self.self_attn(q, k, value=src, attn_mask=src_mask, key_padding_mask=src_key_padding_mask)[0]
+        src = src + self.dropout1(src2); src = self.norm1(src)
+        src2 = self.linear2(self.dropout(self.activation(self.linear1(src))))
+        src = src + self.dropout2(src2); return self.norm2(src)
+    def forward_pre(self, src, src_mask: Optional[Tensor] = None, src_key_padding_mask: Optional[Tensor] = None, pos: Optional[Tensor] = None):
+        src2 = self.norm1(src); q = k = self.with_pos_embed(src2, pos)
+        src2 = self.self_attn(q, k, value=src2, attn_mask=src_mask, key_padding_mask=src_key_padding_mask)[0]
+        src = src + self.dropout1(src2); src2 = self.norm2(src)
+        src2 = self.linear2(self.dropout(self.activation(self.linear1(src2))))
+        return src + self.dropout2(src2)
+    def forward(self, src, src_mask: Optional[Tensor] = None, src_key_padding_mask: Optional[Tensor] = None, pos: Optional[Tensor] = None):
+        return self.forward_pre(src, src_mask, src_key_padding_mask, pos) if self.normalize_before else self.forward_post(src, src_mask, src_key_padding_mask, pos)
+
+class TransformerDecoderLayer(nn.Module):
+    def __init__(self, d_model, nhead, dim_feedforward=2048, dropout=0.1, activation=nn.ReLU, normalize_before=False):
+        super().__init__()
+        self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=dropout)
+        self.multihead_attn = nn.MultiheadAttention(d_model, nhead, dropout=dropout)
+        self.linear1 = nn.Linear(d_model, dim_feedforward); self.dropout = nn.Dropout(dropout); self.linear2 = nn.Linear(dim_feedforward, d_model)
+        self.norm1 = nn.LayerNorm(d_model); self.norm2 = nn.LayerNorm(d_model); self.norm3 = nn.LayerNorm(d_model)
+        self.dropout1 = nn.Dropout(dropout); self.dropout2 = nn.Dropout(dropout); self.dropout3 = nn.Dropout(dropout)
+        self.activation = activation(); self.normalize_before = normalize_before
+
+    def with_pos_embed(self, tensor, pos: Optional[Tensor]): return tensor if pos is None else tensor + pos
+    def forward_post(self, tgt, memory, tgt_mask: Optional[Tensor] = None, memory_mask: Optional[Tensor] = None, tgt_key_padding_mask: Optional[Tensor] = None, memory_key_padding_mask: Optional[Tensor] = None, pos: Optional[Tensor] = None, query_pos: Optional[Tensor] = None):
+        q = k = self.with_pos_embed(tgt, query_pos)
+        tgt2 = self.self_attn(q, k, value=tgt, attn_mask=tgt_mask, key_padding_mask=tgt_key_padding_mask)[0]
+        tgt = tgt + self.dropout1(tgt2); tgt = self.norm1(tgt)
+        tgt2 = self.multihead_attn(query=self.with_pos_embed(tgt, query_pos), key=self.with_pos_embed(memory, pos), value=memory, attn_mask=memory_mask, key_padding_mask=memory_key_padding_mask)[0]
+        tgt = tgt + self.dropout2(tgt2); tgt = self.norm2(tgt)
+        tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt))))
+        tgt = tgt + self.dropout3(tgt2); return self.norm3(tgt)
+    def forward_pre(self, tgt, memory, tgt_mask: Optional[Tensor] = None, memory_mask: Optional[Tensor] = None, tgt_key_padding_mask: Optional[Tensor] = None, memory_key_padding_mask: Optional[Tensor] = None, pos: Optional[Tensor] = None, query_pos: Optional[Tensor] = None):
+        tgt2 = self.norm1(tgt); q = k = self.with_pos_embed(tgt2, query_pos)
+        tgt2 = self.self_attn(q, k, value=tgt2, attn_mask=tgt_mask, key_padding_mask=tgt_key_padding_mask)[0]
+        tgt = tgt + self.dropout1(tgt2); tgt2 = self.norm2(tgt)
+        tgt2 = self.multihead_attn(query=self.with_pos_embed(tgt2, query_pos), key=self.with_pos_embed(memory, pos), value=memory, attn_mask=memory_mask, key_padding_mask=memory_key_padding_mask)[0]
+        tgt = tgt + self.dropout2(tgt2); tgt2 = self.norm3(tgt)
+        tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt2))))
+        return tgt + self.dropout3(tgt2)
+    def forward(self, tgt, memory, tgt_mask: Optional[Tensor] = None, memory_mask: Optional[Tensor] = None, tgt_key_padding_mask: Optional[Tensor] = None, memory_key_padding_mask: Optional[Tensor] = None, pos: Optional[Tensor] = None, query_pos: Optional[Tensor] = None):
+        return self.forward_pre(tgt, memory, tgt_mask, memory_mask, tgt_key_padding_mask, memory_key_padding_mask, pos, query_pos) if self.normalize_before else self.forward_post(tgt, memory, tgt_mask, memory_mask, tgt_key_padding_mask, memory_key_padding_mask, pos, query_pos)
+
+def _get_clones(module, N): return nn.ModuleList([copy.deepcopy(module) for i in range(N)])
+def _get_activation_fn(activation):
+    if activation == "relu": return F.relu
+    if activation == "gelu": return F.gelu
+    if activation == "glu": return F.glu
+    raise RuntimeError(f"activation should be relu/gelu, not {activation}.")
+
+def build_attn_mask(mask_type, device):
+    mask = torch.ones((151, 151), dtype=torch.bool, device=device)
+    if mask_type == "seperate_all":
+        mask[:50, :50] = False; mask[50:67, 50:67] = False; mask[67:84, 67:84] = False
+        mask[84:101, 84:101] = False; mask[101:151, 101:151] = False
+    elif mask_type == "seperate_view":
+        mask[:50, :50] = False; mask[50:67, 50:67] = False; mask[67:84, 67:84] = False
+        mask[84:101, 84:101] = False; mask[101:151, :] = False; mask[:, 101:151] = False
+    return mask
+
+# --- تعريف فئة الإعدادات (Config) ---
+class InterfuserConfig(PretrainedConfig):
+    model_type = "interfuser"
+    def __init__(self, img_size=224, embed_dim=256, enc_depth=6, dec_depth=6, num_heads=8, rgb_backbone_name="r26", lidar_backbone_name="r18", use_different_backbone=True, waypoints_pred_head="gru", **kwargs):
+        super().__init__(**kwargs)
+        self.img_size = img_size
+        self.embed_dim = embed_dim
+        self.enc_depth = enc_depth
+        self.dec_depth = dec_depth
+        self.num_heads = num_heads
+        self.rgb_backbone_name = rgb_backbone_name
+        self.lidar_backbone_name = lidar_backbone_name
+        self.use_different_backbone = use_different_backbone
+        self.waypoints_pred_head = waypoints_pred_head
+        # أضف أي إعدادات أخرى ضرورية هنا
+        self.patch_size=8; self.in_chans=3; self.dim_feedforward=2048; self.normalize_before=False; self.dropout=0.1; self.end2end=False; self.direct_concat=False; self.separate_view_attention=False; self.separate_all_attention=False; self.freeze_num=-1; self.with_lidar=True; self.with_right_left_sensors=True; self.with_center_sensor=True; self.traffic_pred_head_type="det"; self.reverse_pos=True; self.use_view_embed=True; self.use_mmad_pretrain=None
+
+# --- تعريف فئة مخرجات النموذج (ModelOutput) ---
+@dataclass
+class InterfuserOutput(ModelOutput):
+    waypoints: torch.FloatTensor = None
+    traffic_predictions: Optional[torch.FloatTensor] = None
+    is_junction: Optional[torch.FloatTensor] = None
+    traffic_light_state: Optional[torch.FloatTensor] = None
+    stop_sign: Optional[torch.FloatTensor] = None
+    traffic_features: Optional[torch.FloatTensor] = None
+
+# --- تعريف النموذج الأصلي (Interfuser) ---
+# (يجب لصق كلاس Interfuser بالكامل هنا)
+class Interfuser(nn.Module):
+    def __init__(self, config: InterfuserConfig):
+        super().__init__()
+        self.config = config
+        
+        # استخلاص المتغيرات من كائن الـ config
+        embed_dim = config.embed_dim
+        norm_layer = partial(nn.LayerNorm, eps=1e-6)
+        act_layer = nn.GELU
+
+        self.num_features = self.embed_dim = embed_dim
+        self.traffic_pred_head_type = config.traffic_pred_head_type
+        self.waypoints_pred_head = config.waypoints_pred_head
+        self.end2end = config.end2end
+        
+        # ... باقي متغيرات الـ init من الكود الأصلي
+        self.direct_concat = config.direct_concat
+        self.with_center_sensor = config.with_center_sensor
+        self.with_right_left_sensors = config.with_right_left_sensors
+        self.with_lidar = config.with_lidar
+        self.use_view_embed = config.use_view_embed
+        self.separate_view_attention = config.separate_view_attention
+        self.separate_all_attention = config.separate_all_attention
+        
+        if self.direct_concat:
+            in_chans = config.in_chans * 4
+            self.with_center_sensor = False
+            self.with_right_left_sensors = False
+        else:
+            in_chans = config.in_chans
+
+        if self.separate_view_attention:
+            self.attn_mask = build_attn_mask("seperate_view", device=self.device)
+        elif self.separate_all_attention:
+            self.attn_mask = build_attn_mask("seperate_all", device=self.device)
+        else:
+            self.attn_mask = None
+
+        # تعريف الـ backbones (استخدام DummyResNet كمثال)
+        # في الاستخدام الحقيقي، استبدل هذا بالتحميل الفعلي للشبكات
+        backbone_map = {"r50": resnet50d, "r26": resnet26d, "r18": resnet18d}
+        
+        # RGB Backbone
+        rgb_backbone_class = backbone_map.get(config.rgb_backbone_name, resnet26d)
+        self.rgb_backbone = rgb_backbone_class(pretrained=True, in_chans=in_chans, features_only=True, out_indices=[4])
+        
+        # Lidar Backbone
+        if config.use_different_backbone:
+            lidar_backbone_class = backbone_map.get(config.lidar_backbone_name, resnet26d)
+            self.lidar_backbone = lidar_backbone_class(pretrained=False, in_chans=3, features_only=True, out_indices=[4])
+        else:
+            self.lidar_backbone = self.rgb_backbone
+
+        rgb_embed_layer = partial(HybridEmbed, backbone=self.rgb_backbone)
+        lidar_embed_layer = partial(HybridEmbed, backbone=self.lidar_backbone)
+
+        self.rgb_patch_embed = rgb_embed_layer(
+            img_size=config.img_size, patch_size=config.patch_size, in_chans=in_chans, embed_dim=embed_dim
+        )
+        self.lidar_patch_embed = lidar_embed_layer(
+            img_size=config.img_size, patch_size=config.patch_size, in_chans=3, embed_dim=embed_dim
+        )
+
+        # ... باقي تعريفات الطبقات من الكود الأصلي
+        self.global_embed = nn.Parameter(torch.zeros(1, embed_dim, 5))
+        self.view_embed = nn.Parameter(torch.zeros(1, embed_dim, 5, 1))
+
+        if self.end2end:
+            self.query_pos_embed = nn.Parameter(torch.zeros(1, embed_dim, 4))
+            self.query_embed = nn.Parameter(torch.zeros(4, 1, embed_dim))
+        elif self.waypoints_pred_head == "heatmap":
+            self.query_pos_embed = nn.Parameter(torch.zeros(1, embed_dim, 5))
+            self.query_embed = nn.Parameter(torch.zeros(400 + 5, 1, embed_dim))
+        else:
+            self.query_pos_embed = nn.Parameter(torch.zeros(1, embed_dim, 11))
+            self.query_embed = nn.Parameter(torch.zeros(400 + 11, 1, embed_dim))
+
+        # Waypoints Generator
+        if self.end2end: self.waypoints_generator = GRUWaypointsPredictor(embed_dim, 4)
+        elif self.waypoints_pred_head == "heatmap": self.waypoints_generator = MultiPath_Generator(embed_dim + 32, embed_dim, 10)
+        elif self.waypoints_pred_head == "gru": self.waypoints_generator = GRUWaypointsPredictor(embed_dim)
+        elif self.waypoints_pred_head == "gru-command": self.waypoints_generator = GRUWaypointsPredictorWithCommand(embed_dim)
+        elif self.waypoints_pred_head == "linear": self.waypoints_generator = LinearWaypointsPredictor(embed_dim, cumsum=False)
+        elif self.waypoints_pred_head == "linear-sum": self.waypoints_generator = LinearWaypointsPredictor(embed_dim, cumsum=True)
+        
+        self.junction_pred_head = nn.Linear(embed_dim, 2)
+        self.traffic_light_pred_head = nn.Linear(embed_dim, 2)
+        self.stop_sign_head = nn.Linear(embed_dim, 2)
+        
+        self.traffic_pred_head = nn.Sequential(*[nn.Linear(embed_dim + 32, 64), nn.ReLU(), nn.Linear(64, 7), nn.Sigmoid()])
+        self.position_encoding = PositionEmbeddingSine(embed_dim // 2, normalize=True)
+        
+        encoder_layer = TransformerEncoderLayer(embed_dim, config.num_heads, config.dim_feedforward, config.dropout, act_layer, config.normalize_before)
+        self.encoder = TransformerEncoder(encoder_layer, config.enc_depth, None)
+        
+        decoder_layer = TransformerDecoderLayer(embed_dim, config.num_heads, config.dim_feedforward, config.dropout, act_layer, config.normalize_before)
+        decoder_norm = nn.LayerNorm(embed_dim)
+        self.decoder = TransformerDecoder(decoder_layer, config.dec_depth, decoder_norm, return_intermediate=False)
+        
+        self.reset_parameters()
+
+    def reset_parameters(self):
+        nn.init.uniform_(self.global_embed)
+        nn.init.uniform_(self.view_embed)
+        nn.init.uniform_(self.query_embed)
+        nn.init.uniform_(self.query_pos_embed)
+
+    # ... يجب نسخ دالتي forward_features و forward من الكود الأصلي هنا بالكامل
+    def forward_features(self, front_image, left_image, right_image, front_center_image, lidar, measurements):
+        features = []
+        # Front view processing
+        front_image_token, front_image_token_global = self.rgb_patch_embed(front_image)
+        if self.use_view_embed:
+            front_image_token = front_image_token + self.view_embed[:, :, 0:1, :] + self.position_encoding(front_image_token)
+        else:
+            front_image_token = front_image_token + self.position_encoding(front_image_token)
+        front_image_token = front_image_token.flatten(2).permute(2, 0, 1)
+        front_image_token_global = front_image_token_global + self.view_embed[:, :, 0, :] + self.global_embed[:, :, 0:1]
+        front_image_token_global = front_image_token_global.permute(2, 0, 1)
+        features.extend([front_image_token, front_image_token_global])
+
+        if self.with_right_left_sensors:
+            left_image_token, left_image_token_global = self.rgb_patch_embed(left_image)
+            if self.use_view_embed: left_image_token = left_image_token + self.view_embed[:, :, 1:2, :] + self.position_encoding(left_image_token)
+            else: left_image_token = left_image_token + self.position_encoding(left_image_token)
+            left_image_token = left_image_token.flatten(2).permute(2, 0, 1)
+            left_image_token_global = left_image_token_global + self.view_embed[:, :, 1, :] + self.global_embed[:, :, 1:2]
+            left_image_token_global = left_image_token_global.permute(2, 0, 1)
+
+            right_image_token, right_image_token_global = self.rgb_patch_embed(right_image)
+            if self.use_view_embed: right_image_token = right_image_token + self.view_embed[:, :, 2:3, :] + self.position_encoding(right_image_token)
+            else: right_image_token = right_image_token + self.position_encoding(right_image_token)
+            right_image_token = right_image_token.flatten(2).permute(2, 0, 1)
+            right_image_token_global = right_image_token_global + self.view_embed[:, :, 2, :] + self.global_embed[:, :, 2:3]
+            right_image_token_global = right_image_token_global.permute(2, 0, 1)
+            features.extend([left_image_token, left_image_token_global, right_image_token, right_image_token_global])
+
+        if self.with_center_sensor:
+            front_center_image_token, front_center_image_token_global = self.rgb_patch_embed(front_center_image)
+            if self.use_view_embed: front_center_image_token = front_center_image_token + self.view_embed[:, :, 3:4, :] + self.position_encoding(front_center_image_token)
+            else: front_center_image_token = front_center_image_token + self.position_encoding(front_center_image_token)
+            front_center_image_token = front_center_image_token.flatten(2).permute(2, 0, 1)
+            front_center_image_token_global = front_center_image_token_global + self.view_embed[:, :, 3, :] + self.global_embed[:, :, 3:4]
+            front_center_image_token_global = front_center_image_token_global.permute(2, 0, 1)
+            features.extend([front_center_image_token, front_center_image_token_global])
+
+        if self.with_lidar:
+            lidar_token, lidar_token_global = self.lidar_patch_embed(lidar)
+            if self.use_view_embed: lidar_token = lidar_token + self.view_embed[:, :, 4:5, :] + self.position_encoding(lidar_token)
+            else: lidar_token = lidar_token + self.position_encoding(lidar_token)
+            lidar_token = lidar_token.flatten(2).permute(2, 0, 1)
+            lidar_token_global = lidar_token_global + self.view_embed[:, :, 4, :] + self.global_embed[:, :, 4:5]
+            lidar_token_global = lidar_token_global.permute(2, 0, 1)
+            features.extend([lidar_token, lidar_token_global])
+        
+        return torch.cat(features, 0)
+
+    def forward(self, x):
+        front_image, left_image, right_image, front_center_image = x["rgb"], x["rgb_left"], x["rgb_right"], x["rgb_center"]
+        measurements, target_point, lidar = x["measurements"], x["target_point"], x["lidar"]
+
+        if self.direct_concat:
+            img_size = front_image.shape[-1]
+            left_image = F.interpolate(left_image, size=(img_size, img_size))
+            right_image = F.interpolate(right_image, size=(img_size, img_size))
+            front_center_image = F.interpolate(front_center_image, size=(img_size, img_size))
+            front_image = torch.cat([front_image, left_image, right_image, front_center_image], dim=1)
+        
+        features = self.forward_features(front_image, left_image, right_image, front_center_image, lidar, measurements)
+        bs = front_image.shape[0]
+
+        if self.end2end:
+            tgt = self.query_pos_embed.repeat(bs, 1, 1)
+        else:
+            tgt = self.position_encoding(torch.ones((bs, 1, 20, 20), device=x["rgb"].device)).flatten(2)
+            tgt = torch.cat([tgt, self.query_pos_embed.repeat(bs, 1, 1)], 2)
+        tgt = tgt.permute(2, 0, 1)
+
+        memory = self.encoder(features, mask=self.attn_mask)
+        hs = self.decoder(self.query_embed.repeat(1, bs, 1), memory, query_pos=tgt)[0].permute(1, 0, 2)
+
+        if self.end2end:
+            waypoints = self.waypoints_generator(hs, target_point)
+            return waypoints
+
+        if self.waypoints_pred_head != "heatmap":
+            traffic_feature, is_junction_feature, waypoints_feature = hs[:, :400], hs[:, 400], hs[:, 401:411]
+        else:
+            traffic_feature, is_junction_feature, waypoints_feature = hs[:, :400], hs[:, 400], hs[:, 401:405]
+
+        # Waypoints prediction
+        if self.waypoints_pred_head == "heatmap": waypoints = self.waypoints_generator(waypoints_feature, measurements)
+        elif self.waypoints_pred_head.startswith("gru"): waypoints = self.waypoints_generator(waypoints_feature, target_point, measurements) if "command" in self.waypoints_pred_head else self.waypoints_generator(waypoints_feature, target_point)
+        elif self.waypoints_pred_head.startswith("linear"): waypoints = self.waypoints_generator(waypoints_feature, measurements)
+            
+        is_junction = self.junction_pred_head(is_junction_feature)
+        traffic_light_state = self.traffic_light_pred_head(is_junction_feature) # Original code uses same feature
+        stop_sign = self.stop_sign_head(is_junction_feature) # Original code uses same feature
+
+        velocity = measurements[:, 6:7].unsqueeze(-1).repeat(1, 400, 32)
+        traffic_feature_with_vel = torch.cat([traffic_feature, velocity], dim=2)
+        traffic = self.traffic_pred_head(traffic_feature_with_vel)
+        
+        return traffic, waypoints, is_junction, traffic_light_state, stop_sign, traffic_feature
+
+
+# --- تعريف الغلاف الرئيسي (Wrapper) ---
+# هذا هو الكلاس الذي سيتم استدعاؤه بواسطة AutoModel
+class InterfuserForHuggingFace(PreTrainedModel):
+    config_class = InterfuserConfig
+    def __init__(self, config: InterfuserConfig):
+        super().__init__(config)
+        self.model = Interfuser(config) # سيتم بناء النموذج الأصلي هنا
+    def _init_weights(self, module):
+        if hasattr(module, 'reset_parameters'):
+            module.reset_parameters()
+    def forward(self, rgb: torch.FloatTensor, rgb_left: torch.FloatTensor, rgb_right: torch.FloatTensor, rgb_center: torch.FloatTensor, lidar: torch.FloatTensor, measurements: torch.FloatTensor, target_point: torch.FloatTensor, return_dict: Optional[bool] = None) -> Union[Tuple, InterfuserOutput]:
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        inputs = {"rgb": rgb, "rgb_left": rgb_left, "rgb_right": rgb_right, "rgb_center": rgb_center, "lidar": lidar, "measurements": measurements, "target_point": target_point}
+        outputs = self.model(inputs)
+        if self.config.end2end:
+            if not return_dict: return (outputs,)
+            return InterfuserOutput(waypoints=outputs)
+        traffic, waypoints, is_junction, traffic_light_state, stop_sign, traffic_feature = outputs
+        if not return_dict: return outputs
+        return InterfuserOutput(waypoints=waypoints, traffic_predictions=traffic, is_junction=is_junction, traffic_light_state=traffic_light_state, stop_sign=stop_sign, traffic_features=traffic_feature)