Upload app.py with huggingface_hub

app.py

@@ -1,234 +1,2057 @@
-# ============================================================================
-# app.py - الواجهة التفاعلية لمشروع Interfuser
-# ============================================================================
-# هذا الملف مسؤول فقط عن بناء وتشغيل واجهة المستخدم باستخدام Gradio.
-# يعتمد على:
-# - model_utils.py: لإدارة وتحميل النماذج.
-# - simulation_modules.py: لمعالجة المخرجات والتحكم والعرض.
-# ============================================================================
-
-import os
+import traceback
 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
+import copy
+from typing import Optional, Tuple, Union, List
+from torch import Tensor
+from dataclasses import dataclass
 import numpy as np
-import cv2
+from timm.models.resnet import resnet50d,resnet101d, resnet26d, resnet18d
+from torch.utils.data import DataLoader, Dataset
+from collections import deque, OrderedDict
+import os
 import json
-import logging
-import traceback
+import cv2
+from pathlib import Path
+from torchvision import transforms
+from torch.utils.data import random_split
+from timm.models.registry import register_model
+import gradio as gr
+import zipfile
+import tempfile
+import shutil
+import tarfile
+import gdown
+import time 
+from huggingface_hub import hf_hub_download # طريقة أفضل للتنزيل من Hub
+import requests # <-- هذا هو السطر الذي يجب إضافته
 from PIL import Image
 
-# مكتبة الواجهة الرسومية
-import gradio as gr
 
-# مكتبات معالجة الصور
-from torchvision import transforms
+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__()
+        assert isinstance(backbone, nn.Module)
+        img_size = to_2tuple(img_size)
+        patch_size = to_2tuple(patch_size)
+        self.img_size = img_size
+        self.patch_size = 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]  # last feature if backbone outputs list/tuple of features
+                feature_size = o.shape[-2:]
+                feature_dim = o.shape[1]
+                backbone.train(training)
+        else:
+            feature_size = to_2tuple(feature_size)
+            if hasattr(self.backbone, "feature_info"):
+                feature_dim = self.backbone.feature_info.channels()[-1]
+            else:
+                feature_dim = self.backbone.num_features
 
-# --- الجزء الأول: استيراد الوحدات الخاصة بنا ---
-try:
-    from model_utils import get_available_models, load_model_by_name, get_current_model
-    from simulation_modules import (
-        DisplayInterface,
-        InterfuserController,
-        ControllerConfig,
-        render_waypoints,
-        render_self_car,
-        render
-    )
-except ImportError as e:
-    print(f"خطأ في الاستيراد: تأكد من وجود ملفات model_utils.py و simulation_modules.py. الخطأ: {e}")
-    exit()
+        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]  # last feature if backbone outputs list/tuple of features
+        x = self.proj(x)
+        global_x = torch.mean(x, [2, 3], keepdim=False)[:, :, None]
+        return x, global_x
+
+
+class PositionEmbeddingSine(nn.Module):
+    """
+    This is a more standard version of the position embedding, very similar to the one
+    used by the Attention is all you need paper, generalized to work on images.
+    """
+
+    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
+
+
+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):
+        # Output:
+        #   (N, C*2) x_0 y_0 ...
+
+        if self.data_format == "NHWC":
+            feature = (
+                feature.transpose(1, 3)
+                .tranpose(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(
+            torch.autograd.Variable(self.pos_x) * weight, dim=1, keepdim=True
+        )
+        expected_y = torch.sum(
+            torch.autograd.Variable(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
+
+
+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]
+        mask = mask.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1).repeat(1, 1, 1, 100, 100)
+        velocity = measurements[:, 6:7].unsqueeze(-1).unsqueeze(-1)
+        velocity = velocity.repeat(1, 32, 2, 2)
+
+        n, d, c = x.shape
+        x = x.transpose(1, 2)
+        x = x.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 = []
+        for i in range(6):
+            xt = self.tconv4_list[i](x)
+            xs.append(xt)
+        xs = torch.stack(xs, dim=1)
+        x = torch.sum(xs * mask, dim=1)
+        x = self.spatial_softmax(x)
+        return 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):
+        # input shape: n 10 embed_dim
+        bs, n, dim = x.shape
+        x = x + self.rank_embed
+        x = x.reshape(-1, dim)
+
+        mask = measurements[:, :6]
+        mask = torch.unsqueeze(mask, -1).repeat(n, 1, 2)
+
+        rs = []
+        for i in range(6):
+            res = self.head_fc1_list[i](x)
+            res = self.head_relu(res)
+            res = self.head_fc2_list[i](res)
+            rs.append(res)
+        rs = torch.stack(rs, 1)
+        x = torch.sum(rs * mask, dim=1)
+
+        x = x.view(bs, n, 2)
+        if self.cumsum:
+            x = torch.cumsum(x, 1)
+        return x
+
+
+class GRUWaypointsPredictor(nn.Module):
+    def __init__(self, input_dim, waypoints=10):
+        super().__init__()
+        # self.gru = torch.nn.GRUCell(input_size=input_dim, hidden_size=64)
+        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 = output.reshape(bs * self.waypoints, -1)
+        output = self.decoder(output).reshape(bs, self.waypoints, 2)
+        output = torch.cumsum(output, 1)
+        return output
+
+class GRUWaypointsPredictorWithCommand(nn.Module):
+    def __init__(self, input_dim, waypoints=10):
+        super().__init__()
+        # self.gru = torch.nn.GRUCell(input_size=input_dim, hidden_size=64)
+        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]
+        mask = mask.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 = output.reshape(bs * self.waypoints, -1)
+            output = self.decoders[i](output).reshape(bs, self.waypoints, 2)
+            output = torch.cumsum(output, 1)
+            outputs.append(output)
+        outputs = torch.stack(outputs, 1)
+        output = torch.sum(outputs * mask, dim=1)
+        return output
+
+
+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)
+
+        if self.return_intermediate:
+            return torch.stack(intermediate)
+
+        return 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)
+        # Implementation of Feedforward model
+        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)
+        src = self.norm2(src)
+        return 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))))
+        src = src + self.dropout2(src2)
+        return src
+
+    def forward(
+        self,
+        src,
+        src_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+        pos: Optional[Tensor] = None,
+    ):
+        if self.normalize_before:
+            return self.forward_pre(src, src_mask, src_key_padding_mask, pos)
+        return 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)
+        # Implementation of Feedforward model
+        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)
+        tgt = self.norm3(tgt)
+        return 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))))
+        tgt = tgt + self.dropout3(tgt2)
+        return tgt
+
+    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,
+    ):
+        if self.normalize_before:
+            return self.forward_pre(
+                tgt,
+                memory,
+                tgt_mask,
+                memory_mask,
+                tgt_key_padding_mask,
+                memory_key_padding_mask,
+                pos,
+                query_pos,
+            )
+        return self.forward_post(
+            tgt,
+            memory,
+            tgt_mask,
+            memory_mask,
+            tgt_key_padding_mask,
+            memory_key_padding_mask,
+            pos,
+            query_pos,
+        )
+
+
+class Interfuser(nn.Module):
+    def __init__(
+        self,
+        img_size=224,
+        multi_view_img_size=112,
+        patch_size=8,
+        in_chans=3,
+        embed_dim=768,
+        enc_depth=6,
+        dec_depth=6,
+        dim_feedforward=2048,
+        normalize_before=False,
+        rgb_backbone_name="r26",
+        lidar_backbone_name="r26",
+        num_heads=8,
+        norm_layer=None,
+        dropout=0.1,
+        end2end=False,
+        direct_concat=True,
+        separate_view_attention=False,
+        separate_all_attention=False,
+        act_layer=None,
+        weight_init="",
+        freeze_num=-1,
+        with_lidar=False,
+        with_right_left_sensors=True,
+        with_center_sensor=False,
+        traffic_pred_head_type="det",
+        waypoints_pred_head="heatmap",
+        reverse_pos=True,
+        use_different_backbone=False,
+        use_view_embed=True,
+        use_mmad_pretrain=None,
+    ):
+        super().__init__()
+        self.traffic_pred_head_type = traffic_pred_head_type
+        self.num_features = (
+            self.embed_dim
+        ) = embed_dim  # num_features for consistency with other models
+        norm_layer = norm_layer or partial(nn.LayerNorm, eps=1e-6)
+        act_layer = act_layer or nn.GELU
+
+        self.reverse_pos = reverse_pos
+        self.waypoints_pred_head = waypoints_pred_head
+        self.with_lidar = with_lidar
+        self.with_right_left_sensors = with_right_left_sensors
+        self.with_center_sensor = with_center_sensor
+
+        self.direct_concat = direct_concat
+        self.separate_view_attention = separate_view_attention
+        self.separate_all_attention = separate_all_attention
+        self.end2end = end2end
+        self.use_view_embed = use_view_embed
+
+        if self.direct_concat:
+            in_chans = in_chans * 4
+            self.with_center_sensor = False
+            self.with_right_left_sensors = False
+
+        if self.separate_view_attention:
+            self.attn_mask = build_attn_mask("seperate_view")
+        elif self.separate_all_attention:
+            self.attn_mask = build_attn_mask("seperate_all")
+        else:
+            self.attn_mask = None
+
+        if use_different_backbone:
+            if rgb_backbone_name == "r50":
+                self.rgb_backbone = resnet50d(
+                    pretrained=True,
+                    in_chans=in_chans,
+                    features_only=True,
+                    out_indices=[4],
+                )
+            elif rgb_backbone_name == "r26":
+                self.rgb_backbone = resnet26d(
+                    pretrained=True,
+                    in_chans=in_chans,
+                    features_only=True,
+                    out_indices=[4],
+                )
+            elif rgb_backbone_name == "r18":
+                self.rgb_backbone = resnet18d(
+                    pretrained=True,
+                    in_chans=in_chans,
+                    features_only=True,
+                    out_indices=[4],
+                )
+            if lidar_backbone_name == "r50":
+                self.lidar_backbone = resnet50d(
+                    pretrained=False,
+                    in_chans=in_chans,
+                    features_only=True,
+                    out_indices=[4],
+                )
+            elif lidar_backbone_name == "r26":
+                self.lidar_backbone = resnet26d(
+                    pretrained=False,
+                    in_chans=in_chans,
+                    features_only=True,
+                    out_indices=[4],
+                )
+            elif lidar_backbone_name == "r18":
+                self.lidar_backbone = resnet18d(
+                    pretrained=False, in_chans=3, features_only=True, out_indices=[4]
+                )
+            rgb_embed_layer = partial(HybridEmbed, backbone=self.rgb_backbone)
+            lidar_embed_layer = partial(HybridEmbed, backbone=self.lidar_backbone)
+
+            if use_mmad_pretrain:
+                params = torch.load(use_mmad_pretrain)["state_dict"]
+                updated_params = OrderedDict()
+                for key in params:
+                    if "backbone" in key:
+                        updated_params[key.replace("backbone.", "")] = params[key]
+                self.rgb_backbone.load_state_dict(updated_params)
+
+            self.rgb_patch_embed = rgb_embed_layer(
+                img_size=img_size,
+                patch_size=patch_size,
+                in_chans=in_chans,
+                embed_dim=embed_dim,
+            )
+            self.lidar_patch_embed = lidar_embed_layer(
+                img_size=img_size,
+                patch_size=patch_size,
+                in_chans=3,
+                embed_dim=embed_dim,
+            )
+        else:
+            if rgb_backbone_name == "r50":
+                self.rgb_backbone = resnet50d(
+                    pretrained=True, in_chans=3, features_only=True, out_indices=[4]
+                )
+            elif rgb_backbone_name == "r101":
+                self.rgb_backbone = resnet101d(
+                    pretrained=True, in_chans=3, features_only=True, out_indices=[4]
+                )
+            elif rgb_backbone_name == "r26":
+                self.rgb_backbone = resnet26d(
+                    pretrained=True, in_chans=3, features_only=True, out_indices=[4]
+                )
+            elif rgb_backbone_name == "r18":
+                self.rgb_backbone = resnet18d(
+                    pretrained=True, in_chans=3, features_only=True, out_indices=[4]
+                )
+            embed_layer = partial(HybridEmbed, backbone=self.rgb_backbone)
+
+            self.rgb_patch_embed = embed_layer(
+                img_size=img_size,
+                patch_size=patch_size,
+                in_chans=in_chans,
+                embed_dim=embed_dim,
+            )
+            self.lidar_patch_embed = embed_layer(
+                img_size=img_size,
+                patch_size=patch_size,
+                in_chans=in_chans,
+                embed_dim=embed_dim,
+            )
+
+        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))
+
+        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)
+        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)
+
+        if self.traffic_pred_head_type == "det":
+            self.traffic_pred_head = nn.Sequential(
+                *[
+                    nn.Linear(embed_dim + 32, 64),
+                    nn.ReLU(),
+                    nn.Linear(64, 7),
+                    nn.Sigmoid(),
+                ]
+            )
+        elif self.traffic_pred_head_type == "seg":
+            self.traffic_pred_head = nn.Sequential(
+                *[nn.Linear(embed_dim, 64), nn.ReLU(), nn.Linear(64, 1), nn.Sigmoid()]
+            )
+
+        self.position_encoding = PositionEmbeddingSine(embed_dim // 2, normalize=True)
+
+        encoder_layer = TransformerEncoderLayer(
+            embed_dim, num_heads, dim_feedforward, dropout, act_layer, normalize_before
+        )
+        self.encoder = TransformerEncoder(encoder_layer, enc_depth, None)
+
+        decoder_layer = TransformerDecoderLayer(
+            embed_dim, num_heads, dim_feedforward, dropout, act_layer, normalize_before
+        )
+        decoder_norm = nn.LayerNorm(embed_dim)
+        self.decoder = TransformerDecoder(
+            decoder_layer, 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)
+
+    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 view processing
+            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 view processing
+            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 view processing
+            (
+                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])
+
+        features = torch.cat(features, 0)
+        return features
+
+    def forward(self, x):
+        front_image = x["rgb"]
+        left_image = x["rgb_left"]
+        right_image = x["rgb_right"]
+        front_center_image = x["rgb_center"]
+        measurements = x["measurements"]
+        target_point = x["target_point"]
+        lidar = x["lidar"]
+
+        if self.direct_concat:
+            img_size = front_image.shape[-1]
+            left_image = torch.nn.functional.interpolate(
+                left_image, size=(img_size, img_size)
+            )
+            right_image = torch.nn.functional.interpolate(
+                right_image, size=(img_size, img_size)
+            )
+            front_center_image = torch.nn.functional.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)
+            )
+            tgt = tgt.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]
+
+        hs = hs.permute(1, 0, 2)  # Batchsize ,  N, C
+        if self.end2end:
+            waypoints = self.waypoints_generator(hs, target_point)
+            return waypoints
+
+        if self.waypoints_pred_head != "heatmap":
+            traffic_feature = hs[:, :400]
+            is_junction_feature = hs[:, 400]
+            traffic_light_state_feature = hs[:, 400]
+            stop_sign_feature = hs[:, 400]
+            waypoints_feature = hs[:, 401:411]
+        else:
+            traffic_feature = hs[:, :400]
+            is_junction_feature = hs[:, 400]
+            traffic_light_state_feature = hs[:, 400]
+            stop_sign_feature = hs[:, 400]
+            waypoints_feature = hs[:, 401:405]
+
+        if self.waypoints_pred_head == "heatmap":
+            waypoints = self.waypoints_generator(waypoints_feature, measurements)
+        elif self.waypoints_pred_head == "gru":
+            waypoints = self.waypoints_generator(waypoints_feature, target_point)
+        elif self.waypoints_pred_head == "gru-command":
+            waypoints = self.waypoints_generator(waypoints_feature, target_point, measurements)
+        elif self.waypoints_pred_head == "linear":
+            waypoints = self.waypoints_generator(waypoints_feature, measurements)
+        elif self.waypoints_pred_head == "linear-sum":
+            waypoints = self.waypoints_generator(waypoints_feature, measurements)
+
+        is_junction = self.junction_pred_head(is_junction_feature)
+        traffic_light_state = self.traffic_light_pred_head(traffic_light_state_feature)
+        stop_sign = self.stop_sign_head(stop_sign_feature)
+
+        velocity = measurements[:, 6:7].unsqueeze(-1)
+        velocity = velocity.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
+
+# ================== 3. فئة التتبع ==================
+class TrackedObject:
+    def __init__(self):
+        self.last_step = 0
+        self.last_pos = [0, 0]
+        # استخدام deque يعطينا كفاءة أفضل ويحدد حجمًا أقصى للذاكرة
+        self.historical_pos = deque(maxlen=10) 
+        self.historical_steps = deque(maxlen=10)
+        self.historical_features = deque(maxlen=10)
+
+    # هذه هي الدالة المفق��دة التي يجب إضافتها
+    def update(self, step, object_info):
+        """
+        تحديث حالة الكائن بالبيانات الجديدة من الإطار الحالي.
+        """
+        self.last_step = step
+        self.last_pos = object_info[:2]
+        
+        # إضافة البيانات الجديدة إلى السجل التاريخي
+        self.historical_pos.append(self.last_pos)
+        self.historical_steps.append(step)
+        
+        # التأكد من وجود ميزات إضافية قبل إضافتها
+        if len(object_info) > 2:
+            self.historical_features.append(object_info[2])
+
+class Tracker:
+    def __init__(self, frequency=10):
+        self.tracks = []
+        self.alive_ids = []
+        self.frequency = frequency
+
+    def update_and_predict(self, det_data, pos, theta, frame_num):
+        det_data_weighted = det_data * reweight_array
+        detected_objects = find_peak_box(det_data_weighted)
+        objects_info = []
+        R = np.array([[np.cos(-theta), -np.sin(-theta)], [np.sin(-theta), np.cos(-theta)]])
+
+        for obj in detected_objects:
+            i, j = obj['coords']
+            obj_data = obj['raw_data']
+
+            center_y, center_x = convert_grid_to_xy(i, j)
+            center_x += obj_data[1]
+            center_y += obj_data[2]
+
+            loc = R.T.dot(np.array([center_x, center_y]))
+            objects_info.append([loc[0] + pos[0], loc[1] + pos[1], obj_data[1:]])  # [x, y, features...]
+
+        updates_ids = self._update(objects_info, frame_num)
+        speed_results, heading_results = self._predict(updates_ids)
+
+        for k, poi in enumerate(updates_ids):
+            i, j = poi
+            if heading_results[k] is not None:
+                factor = MERGE_PERCENT * 0.1
+                det_data[i, j, 3] = heading_results[k] * factor + det_data[i, j, 3] * (1 - factor)
+            if speed_results[k] is not None:
+                factor = MERGE_PERCENT * 0.1
+                det_data[i, j, 6] = speed_results[k] * factor + det_data[i, j, 6] * (1 - factor)
+        return det_data
+
+    def _update(self, objects_info, step):
+        latest_ids = []
+        if len(self.tracks) == 0:
+            for object_info in objects_info:
+                to = TrackedObject()
+                to.update(step, object_info)
+                self.tracks.append(to)
+                latest_ids.append(len(self.tracks) - 1)
+        else:
+            matched_ids = set()
+            for idx, object_info in enumerate(objects_info):
+                min_id, min_error = -1, float('inf')
+                pos_x, pos_y = object_info[:2]
+                for _id in self.alive_ids:
+                    if _id in matched_ids:
+                        continue
+                    track_pos = self.tracks[_id].last_pos
+                    distance = np.sqrt((track_pos[0] - pos_x)**2 + (track_pos[1] - pos_y)**2)
+                    if distance < 2.0 and distance < min_error:
+                        min_error = distance
+                        min_id = _id
+                if min_id != -1:
+                    self.tracks[min_id].update(step, objects_info[idx])
+                    latest_ids.append(min_id)
+                    matched_ids.add(min_id)
+                else:
+                    to = TrackedObject()
+                    to.update(step, object_info)
+                    self.tracks.append(to)
+                    latest_ids.append(len(self.tracks) - 1)
+        self.alive_ids = [i for i, track in enumerate(self.tracks) if track.last_step > step - 6]
+        return latest_ids
+
+    def _match(self, objects_info):
+        results = []
+        matched_ids = set()
+        for object_info in objects_info:
+            min_id, min_error = -1, float('inf')
+            pos_x, pos_y = object_info[:2]
+            for _id in self.alive_ids:
+                if _id in matched_ids:
+                    continue
+                track_pos = self.tracks[_id].last_pos
+                distance = np.sqrt((track_pos[0] - pos_x)**2 + (track_pos[1] - pos_y)**2)
+                if distance < min_error:
+                    min_error = distance
+                    min_id = _id
+            results.append(min_id)
+            if min_id != -1:
+                matched_ids.add(min_id)
+        return results
+
+    def _predict(self, updates_ids):
+        speed_results, heading_results = [], []
+        for each_id in updates_ids:
+            to = self.tracks[each_id]
+            avg_speed, avg_heading = [], []
+            for feature in to.historical_features:
+                avg_speed.append(feature[2])
+                avg_heading.append(feature[:2])
+            if len(avg_speed) < 2:
+                speed_results.append(None)
+                heading_results.append(None)
+                continue
+            avg_speed = np.mean(avg_speed)
+            avg_heading = np.mean(np.stack(avg_heading), axis=0)
+            yaw_angle = get_yaw_angle(avg_heading)
+            heading_results.append((4 - yaw_angle / np.pi) % 2)
+            speed_results.append(avg_speed)
+        return speed_results, heading_results
 
-# --- الجزء الثاني: الإعدادات والثوابت ---
-logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')
 
-SAMPLE_DATA_DIR = "sample_data"
-DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 
-# تعريف التحويلات (يجب أن تطابق ما تم استخدامه أثناء التدريب)
-RGB_TRANSFORM = transforms.Compose([
+
+
+# ================== 0. تعريف PID Controller ==================
+class PIDController:
+    def __init__(self, K_P=1.0, K_I=0.0, K_D=0.0, n=20):
+        self._K_P = K_P
+        self._K_I = K_I
+        self._K_D = K_D
+        self._window = deque([0 for _ in range(n)], maxlen=n)
+        self._max = 0.0
+        self._min = 0.0
+
+    def step(self, error):
+        self._window.append(error)
+        self._max = max(self._max, abs(error))
+        self._min = -abs(self._max)
+
+        if len(self._window) >= 2:
+            integral = np.mean(self._window)
+            derivative = self._window[-1] - self._window[-2]
+        else:
+            integral = 0.0
+            derivative = 0.0
+
+        return self._K_P * error + self._K_I * integral + self._K_D * derivative
+# ================== 4. فئة المتحكم ==================
+class InterfuserController:
+    def __init__(self, config):
+        self.turn_controller = PIDController(
+            K_P=config.turn_KP,
+            K_I=config.turn_KI,
+            K_D=config.turn_KD,
+            n=config.turn_n,
+        )
+        self.speed_controller = PIDController(
+            K_P=config.speed_KP,
+            K_I=config.speed_KI,
+            K_D=config.speed_KD,
+            n=config.speed_n,
+        )
+        self.config = config
+        self.collision_buffer = np.array(config.collision_buffer)
+        self.detect_threshold = config.detect_threshold
+        self.stop_steps = 0
+        self.forced_forward_steps = 0
+        self.red_light_steps = 0
+        self.block_red_light = 0
+        self.in_stop_sign_effect = False
+        self.block_stop_sign_distance = 0
+        self.stop_sign_timer = 0
+        self.stop_sign_trigger_times = 0
+
+    def run_step(
+        self, speed, waypoints, junction, traffic_light_state, stop_sign, meta_data
+    ):
+        # --- تحديث حالة التوقف ---
+        if speed < 0.2:
+            self.stop_steps += 1
+        else:
+            self.stop_steps = max(0, self.stop_steps - 10)
+
+        if speed < 0.06 and self.in_stop_sign_effect:
+            self.in_stop_sign_effect = False
+
+        if junction < 0.3:
+            self.stop_sign_trigger_times = 0
+
+        if traffic_light_state > 0.7:
+            self.red_light_steps += 1
+        else:
+            self.red_light_steps = 0
+
+        if self.red_light_steps > 1000:
+            self.block_red_light = 80
+            self.red_light_steps = 0
+
+        if self.block_red_light > 0:
+            self.block_red_light -= 1
+            traffic_light_state = 0.01
+
+        if stop_sign < 0.6 and self.block_stop_sign_distance < 0.1:
+            self.in_stop_sign_effect = True
+            self.block_stop_sign_distance = 2.0
+            self.stop_sign_trigger_times = 3
+
+        self.block_stop_sign_distance = max(
+            0, self.block_stop_sign_distance - 0.05 * speed
+        )
+
+        if self.block_stop_sign_distance < 0.1:
+            if self.stop_sign_trigger_times > 0:
+                self.block_stop_sign_distance = 2.0
+                self.stop_sign_trigger_times -= 1
+                self.in_stop_sign_effect = True
+
+        # --- حساب زاوية الانعطاف ---
+        aim = (waypoints[1] + waypoints[0]) / 2.0
+        angle = np.degrees(np.pi / 2 - np.arctan2(aim[1], aim[0])) / 90
+        if speed < 0.01:
+            angle = 0
+        steer = self.turn_controller.step(angle)
+        steer = np.clip(steer, -1.0, 1.0)
+
+        brake = False
+        throttle = 0.0
+        desired_speed = 0.0
+
+        downsampled_waypoints = downsample_waypoints(waypoints)
+
+        d_0 = get_max_safe_distance(
+            meta_data,
+            downsampled_waypoints,
+            t=0,
+            collision_buffer=self.collision_buffer,
+            threshold=self.detect_threshold,
+        )
+        d_05 = get_max_safe_distance(
+            meta_data,
+            downsampled_waypoints,
+            t=0.5,
+            collision_buffer=self.collision_buffer,
+            threshold=self.detect_threshold,
+        )
+        d_075 = get_max_safe_distance(
+            meta_data,
+            downsampled_waypoints,
+            t=0.75,
+            collision_buffer=self.collision_buffer,
+            threshold=self.detect_threshold,
+        )
+        d_1 = get_max_safe_distance(
+            meta_data,
+            downsampled_waypoints,
+            t=1,
+            collision_buffer=self.collision_buffer,
+            threshold=self.detect_threshold,
+        )
+        d_15 = get_max_safe_distance(
+            meta_data,
+            downsampled_waypoints,
+            t=1.5,
+            collision_buffer=self.collision_buffer,
+            threshold=self.detect_threshold,
+        )
+        d_2 = get_max_safe_distance(
+            meta_data,
+            downsampled_waypoints,
+            t=2,
+            collision_buffer=self.collision_buffer,
+            threshold=self.detect_threshold,
+        )
+
+        d_05 = min(d_0, d_05, d_075)
+        d_1 = min(d_05, d_075, d_15, d_2)
+
+        safe_dis = min(d_05, d_1)
+        d_0 = max(0, d_0 - 2.0)
+        d_05 = max(0, d_05 - 2.0)
+        d_1 = max(0, d_1 - 2.0)
+
+        # --- تفعيل الفرملة فقط إذا كانت الإشارة حمراء أو هناك علامة Stop ---
+        if traffic_light_state > 0.5:
+            brake = True
+            desired_speed = 0.0
+        elif stop_sign > 0.6 and traffic_light_state <= 0.5:
+            if self.stop_sign_timer < 20:
+                brake = True
+                desired_speed = 0.0
+                self.stop_sign_timer += 1
+            else:
+                brake = False
+                desired_speed = max(0, min(self.config.max_speed, speed + 0.2))
+        else:
+            brake = False
+            desired_speed = max(0, min(self.config.max_speed, speed + 0.2))
+
+        delta = np.clip(desired_speed - speed, 0.0, self.config.clip_delta)
+        throttle = self.speed_controller.step(delta)
+        throttle = np.clip(throttle, 0.0, self.config.max_throttle)
+
+        # --- إذا كانت السرعة أعلى من 1.1 مرة السرعة المستهدفة، نفرم ---
+        if speed > desired_speed * self.config.brake_ratio:
+            brake = True
+
+        # --- إعداد معلومات التشخيص ---
+        meta_info_1 = f"speed: {speed:.2f}, target_speed: {desired_speed:.2f}"
+        meta_info_2 = f"on_road_prob: {junction:.2f}, red_light_prob: {traffic_light_state:.2f}, stop_sign_prob: {1 - stop_sign:.2f}"
+        meta_info_3 = f"stop_steps: {self.stop_steps}, block_stop_sign_distance: {self.block_stop_sign_distance:.1f}"
+
+        # --- حالة خاصة بعد فترة طويلة من التوقف ---
+        if self.stop_steps > 1200:
+            self.forced_forward_steps = 12
+            self.stop_steps = 0
+        if self.forced_forward_steps > 0:
+            throttle = 0.8
+            brake = False
+            self.forced_forward_steps -= 1
+        if self.in_stop_sign_effect:
+            throttle = 0
+            brake = True
+
+        return steer, throttle, brake, (meta_info_1, meta_info_2, meta_info_3, safe_dis)
+
+
+class ControllerConfig:
+    turn_KP, turn_KI, turn_KD, turn_n = 1.0, 0.1, 0.1, 20
+    speed_KP, speed_KI, speed_KD, speed_n = 0.5, 0.05, 0.1, 20
+    max_speed, max_throttle, clip_delta = 6.0, 0.75, 0.25
+    collision_buffer, detect_threshold = [0.0, 0.0], 0.04
+    brake_speed, brake_ratio = 0.4, 1.1
+
+
+# ================== 5. واجهة العرض ==================
+class DisplayInterface:
+    def __init__(self, width=1200, height=600):
+        self._width = width
+        self._height = height
+
+    def run_interface(self, data):
+        dashboard = np.zeros((self._height, self._width, 3), dtype=np.uint8)
+        font = cv2.FONT_HERSHEY_SIMPLEX
+        dashboard[:, :800] = cv2.resize(data.get('camera_view'), (800, 600))
+        dashboard[:400, 800:1200] = cv2.resize(data['map_t0'], (400, 400))
+        dashboard[400:600, 800:1000] = cv2.resize(data['map_t1'], (200, 200))
+        dashboard[400:600, 1000:1200] = cv2.resize(data['map_t2'], (200, 200))
+
+        # خطوط فصل
+        cv2.line(dashboard, (800, 0), (800, 600), (255, 255, 255), 2)
+        cv2.line(dashboard, (800, 400), (1200, 400), (255, 255, 255), 2)
+        cv2.line(dashboard, (1000, 400), (1000, 600), (255, 255, 255), 2)
+
+        y_pos = 40
+        for key, text in data['text_info'].items():
+            cv2.putText(dashboard, text, (820, y_pos), font, 0.6, (255, 255, 255), 1)
+            y_pos += 30
+
+        y_pos += 10
+        for t, counts in data['object_counts'].items():
+            count_str = f"{t}: C={counts['car']} B={counts['bike']} P={counts['pedestrian']}"
+            cv2.putText(dashboard, count_str, (820, y_pos), font, 0.5, (255, 255, 255), 1)
+            y_pos += 20
+
+        cv2.putText(dashboard, "t0", (1160, 30), font, 0.8, (0, 255, 255), 2)
+        cv2.putText(dashboard, "t1", (960, 430), font, 0.8, (0, 255, 255), 2)
+        cv2.putText(dashboard, "t2", (1160, 430), font, 0.8, (0, 255, 255), 2)
+
+        return dashboard
+
+# --- تحديد التحوّلات ---
+transform = transforms.Compose([
+    # الخطوة الأولى الآن هي تغيير الحجم مباشرة
     transforms.Resize((224, 224)),
     transforms.ToTensor(),
-    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
+    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
 ])
 
-LIDAR_TRANSFORM = transforms.Compose([
-    transforms.Resize((224, 224)),
-    transforms.ToTensor()
+lidar_transform = transforms.Compose([
+    # الخطوة الأولى الآن هي تغيير الحجم مباشرة
+    transforms.Resize((112, 112)),
+    transforms.ToTensor(),
+    transforms.Normalize(mean=[0.5], std=[0.5]),
 ])
 
-# --- الجزء الثالث: الكائنات العامة للواجهة ---
-logging.info("تهيئة مكونات الواجهة...")
-GLOBAL_DISPLAY_INTERFACE = DisplayInterface()
-controller_config = ControllerConfig()
-GLOBAL_CONTROLLER = InterfuserController(controller_config)
-logging.info(f"تم تهيئة مكونات الواجهة. الجهاز المستخدم: {DEVICE}")
+class LMDriveDataset(Dataset):
+    def __init__(self, data_dir, transform=None, lidar_transform=None):
+        self.data_dir = Path(data_dir)
+        self.transform = transform
+        self.lidar_transform = lidar_transform
+        self.samples = []
 
-# --- الجزء الرابع: دوال المعالجة الرئيسية ---
+        measurement_dir = self.data_dir / "measurements"
+        image_dir = self.data_dir / "rgb_full"
 
-def process_and_run_inference(
-    model: torch.nn.Module,
-    paths: dict
-):
-    """
-    محرك المعالجة: يعالج بيانات إطار واحد بناءً على المسارات ويشغل النموذج.
-    """
-    try:
-        rgb_path = paths['rgb']
-        if not rgb_path or not os.path.exists(rgb_path): 
-            raise FileNotFoundError("ملف الصورة الأمامية (RGB) غير موجود.")
+        measurement_files = sorted([f for f in os.listdir(measurement_dir) if f.endswith(".json")])
+        image_files = sorted([f for f in os.listdir(image_dir) if f.endswith(".jpg")])
 
-        rgb_image = Image.open(rgb_path).convert("RGB")
-        
-        def open_optional_image(path, default_img):
-            return Image.open(path).convert("RGB") if path and os.path.exists(path) else default_img
+        num_samples = min(len(measurement_files), len(image_files))
+
+        for i in range(num_samples):
+            frame_id = i
+            measurement_path = str(measurement_dir / f"{frame_id:04d}.json")
+            image_name = f"{frame_id:04d}.jpg"
+            image_path = str(image_dir / image_name)
+
+            if not os.path.exists(measurement_path) or not os.path.exists(image_path):
+                continue
+
+            with open(measurement_path, "r") as f:
+                measurements_data = json.load(f)
+
+            self.samples.append({
+                "image_path": image_path,
+                "measurement_path": measurement_path,
+                "frame_id": frame_id,
+                "measurements": measurements_data
+            })
+
+    def __len__(self):
+        return len(self.samples)
 
-        rgb_left_image = open_optional_image(paths.get('left'), rgb_image)
-        rgb_right_image = open_optional_image(paths.get('right'), rgb_image)
-        rgb_center_image = open_optional_image(paths.get('center'), rgb_image)
+    def __getitem__(self, idx):
+        sample = self.samples[idx]
 
-        lidar_path = paths.get('lidar')
-        if lidar_path and os.path.exists(lidar_path):
-            lidar_array = np.load(lidar_path)
-            max_val = lidar_array.max()
-            if max_val > 0: lidar_array = (lidar_array / max_val) * 255.0
-            lidar_image = Image.fromarray(lidar_array.astype(np.uint8)).convert('RGB')
+        # قراءة الصورة الكاملة (2400x800)
+        full_image = cv2.imread(sample["image_path"])
+        if full_image is None:
+            raise ValueError(f"Failed to load image: {sample['image_path']}")
+        full_image = cv2.cvtColor(full_image, cv2.COLOR_BGR2RGB)
+
+        # تقسيم الصورة إلى أجزاء (كل جزء 600x800)
+        front_image = full_image[:600, :800]      # الجزء الأول
+        left_image = full_image[600:1200, :800]   # الجزء الثاني
+        right_image = full_image[1200:1800, :800] # الجزء الثالث
+        center_image = full_image[1800:2400, :800]# الجزء الرابع
+
+        # تطبيق التحويل على كل صورة
+        front_image_tensor = self.transform(front_image)
+        left_image_tensor = self.transform(left_image)
+        right_image_tensor = self.transform(right_image)
+        center_image_tensor = self.transform(center_image)
+
+        # تحميل الليدار
+        lidar_path = str(self.data_dir / "lidar" / f"{sample['frame_id']:04d}.png")
+        lidar = cv2.imread(lidar_path)
+
+        if lidar is None:
+            lidar = np.zeros((112, 112, 3), dtype=np.uint8)  # مكان فارغ
         else:
-            lidar_image = Image.fromarray(np.zeros((224, 224, 3), dtype=np.uint8))
+            if len(lidar.shape) == 2:
+                lidar = cv2.cvtColor(lidar, cv2.COLOR_GRAY2BGR)
+            lidar = cv2.cvtColor(lidar, cv2.COLOR_BGR2RGB)
 
-        target_point_list = json.load(open(paths['target_point']))['value']
-        
-        inputs = {
-            'rgb': RGB_TRANSFORM(rgb_image).unsqueeze(0).to(DEVICE), 
-            'rgb_left': RGB_TRANSFORM(rgb_left_image).unsqueeze(0).to(DEVICE), 
-            'rgb_right': RGB_TRANSFORM(rgb_right_image).unsqueeze(0).to(DEVICE),
-            'rgb_center': RGB_TRANSFORM(rgb_center_image).unsqueeze(0).to(DEVICE), 
-            'lidar': LIDAR_TRANSFORM(lidar_image).unsqueeze(0).to(DEVICE),
-            'measurements': torch.tensor([json.load(open(paths['measurements']))['values']], dtype=torch.float32).to(DEVICE),
-            'target_point': torch.tensor([target_point_list], dtype=torch.float32).to(DEVICE)
+        lidar_tensor = self.lidar_transform(lidar)
+
+        # استخراج القياسات
+        measurements_data = sample["measurements"]
+
+        x = measurements_data.get("x", 0.0)
+        y = measurements_data.get("y", 0.0)
+        theta = measurements_data.get("theta", 0.0)
+        speed = measurements_data.get("speed", 0.0)
+        steer = measurements_data.get("steer", 0.0)
+        throttle = measurements_data.get("throttle", 0.0)
+        brake = int(measurements_data.get("brake", False))
+        command = measurements_data.get("command", 0)
+        is_junction = int(measurements_data.get("is_junction", False))
+        should_brake = int(measurements_data.get("should_brake", 0))
+        x_command = measurements_data.get("x_command", 0.0)
+        y_command = measurements_data.get("y_command", 0.0)
+
+        target_point = torch.tensor([x_command, y_command], dtype=torch.float32)
+
+        measurements = torch.tensor(
+            [x, y, theta, speed, steer, throttle, brake, command, is_junction, should_brake],
+            dtype=torch.float32
+        )
+
+        return {
+            "rgb": front_image_tensor,
+            "rgb_left": left_image_tensor,
+            "rgb_right": right_image_tensor,
+            "rgb_center": center_image_tensor,
+            "lidar": lidar_tensor,
+            "measurements": measurements,
+            "target_point": target_point
         }
 
-        with torch.no_grad():
-            outputs = model(inputs)
-        
-        return outputs
 
-    except Exception as e:
-        logging.error(traceback.format_exc())
-        raise gr.Error(f"حدث خطأ أثناء معالجة البيانات: {e}")
 
 
-def run_simulation(scenario_name: str):
+SAVE_VIDEO = True
+FPS = 10
+WAYPOINT_SCALE_FACTOR = 5.0
+T1_FUTURE_TIME = 1.0
+T2_FUTURE_TIME = 2.0
+TRACKER_FREQUENCY = 10
+MERGE_PERCENT = 0.4
+PIXELS_PER_METER = 8
+MAX_DISTANCE = 32
+IMG_SIZE = MAX_DISTANCE * PIXELS_PER_METER * 2
+EGO_CAR_X = IMG_SIZE // 2
+EGO_CAR_Y = IMG_SIZE - (4.0 * PIXELS_PER_METER)
+reweight_array = np.ones((20, 20, 7))
+last_valid_waypoints = None
+last_valid_theta = 0.0
+
+def to_2tuple(x):
+    if isinstance(x, tuple): return x
+    return (x, x)
+
+
+def _get_clones(module, N):
+    return nn.ModuleList([copy.deepcopy(module) for i in range(N)])
+
+
+def _get_activation_fn(activation):
+    """Return an activation function given a string"""
+    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):
+    mask = torch.ones((151, 151), dtype=torch.bool).cuda()
+    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
+
+
+def get_yaw_angle(forward_vector):
+    forward_vector = forward_vector / np.linalg.norm(forward_vector)
+    yaw = math.atan2(forward_vector[1], forward_vector[0])
+    return yaw
+
+
+@register_model
+def interfuser_baseline(**kwargs):
+    model = Interfuser(
+        enc_depth=6,
+        dec_depth=6,
+        embed_dim=256,
+        rgb_backbone_name="r50",
+        lidar_backbone_name="r18",
+        waypoints_pred_head="gru",
+        use_different_backbone=True,
+    )
+    # model.save_pretrained("/content/t")
+    return model
+
+def ensure_rgb(image):
+    """تحويل الصورة إلى RGB إذا كانت grayscale."""
+    if len(image.shape) == 2 or image.shape[2] == 1:
+        return cv2.cvtColor(image, cv2.COLOR_GRAY2BGR)
+    return image
+def process_camera_image(tensor_image):
+    """تحويل صورة الكاميرا من Tensor إلى NumPy Array."""
+    image_np = tensor_image.permute(1, 2, 0).cpu().numpy()
+    image_np = (image_np * np.array([0.229, 0.224, 0.225])) + np.array([0.485, 0.456, 0.406])
+    image_np = np.clip(image_np, 0, 1)
+    return (image_np * 255).astype(np.uint8)[:, :, ::-1]  # BGR
+
+
+def convert_grid_to_xy(i, j):
+    """تحويل الشبكة إلى إحداثيات x, y."""
+    return (j - 9.5) * 1.6, (19.5 - i) * 1.6
+
+
+def add_rect(img, loc, ori, box, value, color):
     """
-    المنسق الرئيسي: يجهز المدخلات، يستدعي محرك المعالجة، ثم يعرض النتائج.
+    إضافة مستطيل إلى الخريطة.
     """
-    logging.info(f"بدء المحاكاة للسيناريو: '{scenario_name}'")
-    
-    model = get_current_model()
-    if model is None:
-        raise gr.Error("لا يوجد نموذج محمل. يرجى اختيار نموذج من القائمة أولاً.")
-
-    scenario_path = os.path.join(SAMPLE_DATA_DIR, scenario_name)
-    paths = {
-        'rgb': os.path.join(scenario_path, 'rgb.png'),
-        'left': os.path.join(scenario_path, 'rgb_left.png'),
-        'right': os.path.join(scenario_path, 'rgb_right.png'),
-        'center': os.path.join(scenario_path, 'rgb_center.png'),
-        'lidar': os.path.join(scenario_path, 'lidar.npy'),
-        'measurements': os.path.join(scenario_path, 'measurements.json'),
-        'target_point': os.path.join(scenario_path, 'target_point.json')
-    }
-
-    traffic, waypoints, is_junction, traffic_light, stop_sign, _ = process_and_run_inference(model, paths)
-    
-    pred_wp = waypoints[0].cpu().numpy()
-    pred_traffic_map = traffic[0].sigmoid().cpu().numpy().reshape(20, 20, 7)
-    current_speed = json.load(open(paths['measurements'])).get('values', [0]*7)[3]
-
-    steer, throttle, brake, metadata = GLOBAL_CONTROLLER.run_step(
-        current_speed=current_speed, waypoints=pred_wp,
-        junction=torch.sigmoid(is_junction)[0,1].item(),
-        traffic_light_state=torch.sigmoid(traffic_light)[0,0].item(),
-        stop_sign=torch.sigmoid(stop_sign)[0,1].item(),
-        meta_data={}
+    center_x = int(loc[0] * PIXELS_PER_METER + MAX_DISTANCE * PIXELS_PER_METER)
+    center_y = int(loc[1] * PIXELS_PER_METER + MAX_DISTANCE * PIXELS_PER_METER)
+
+    size_px = (
+        int(box[0] * PIXELS_PER_METER),
+        int(box[1] * PIXELS_PER_METER)
     )
-    
-    traffic_render, _ = render(pred_traffic_map)
-    waypoints_render = render_waypoints(pred_wp)
-    combined_map = cv2.addWeighted(traffic_render, 1.0, waypoints_render, 1.0, 0.0)
-    final_map = render_self_car(combined_map)
-
-    display_data = {
-        'camera_view': np.array(Image.open(paths['rgb'])),
-        'map_t0': final_map, 'map_t1': np.zeros_like(final_map), 'map_t2': np.zeros_like(final_map),
-        'text_info': {
-            "Controller": metadata,
-            "Brake": f"Brake Activated: {'YES' if brake else 'NO'}"
-        }
-    }
-    
-    dashboard = GLOBAL_DISPLAY_INTERFACE.run_interface(display_data)
-    logging.info("انتهت المحاكاة بنجاح.")
-    
-    return dashboard, metadata
+
+    angle_deg = -np.degrees(math.atan2(ori[1], ori[0]))
+
+    box_points = cv2.boxPoints(((center_x, center_y), size_px, angle_deg))
+    box_points = np.int32(box_points)
+
+    adjusted_color = [int(x * value) for x in color]
+    cv2.fillConvexPoly(img, box_points, adjusted_color)
+    return img
+
+def find_peak_box(data):
+    """
+    اكتشاف القمم في البيانات وتصنيفها.
+    """
+    det_data = np.zeros((22, 22, 7))
+    det_data[1:21, 1:21] = data
+    detected_objects = []
+
+    for i in range(1, 21):
+        for j in range(1, 21):
+            if det_data[i, j, 0] > 0.6 and (
+                det_data[i, j, 0] > det_data[i, j - 1, 0]
+                and det_data[i, j, 0] > det_data[i, j + 1, 0]
+                and det_data[i, j, 0] > det_data[i - 1, j, 0]
+                and det_data[i, j, 0] > det_data[i + 1, j, 0]
+            ):
+                length = det_data[i, j, 4]
+                width = det_data[i, j, 5]
+                confidence = det_data[i, j, 0]
+
+                obj_type = 'unknown'
+                if length > 4.0:
+                    obj_type = 'car'
+                elif length / width > 1.5:
+                    obj_type = 'bike'
+                else:
+                    obj_type = 'pedestrian'
+
+                detected_objects.append({
+                    'coords': (i - 1, j - 1),
+                    'type': obj_type,
+                    'confidence': confidence,
+                    'raw_data': det_data[i, j]
+                })
+
+    return detected_objects
+
+
+def render(det_data, t=0):
+    """
+    رسم كائنات الكشف على الخريطة BEV.
+    """
+    CLASS_COLORS = {'car': (0, 0, 255), 'bike': (0, 255, 0), 'pedestrian': (255, 0, 0), 'unknown': (128, 128, 128)}
+    det_weighted = det_data * reweight_array
+    detected_objects = find_peak_box(det_weighted)
+    counts = {cls: 0 for cls in CLASS_COLORS.keys()}
+    [counts.update({obj['type']: counts.get(obj['type'], 0) + 1}) for obj in detected_objects]
+    img = np.zeros((IMG_SIZE, IMG_SIZE, 3), np.uint8)
+
+    for obj in detected_objects:
+        i, j = obj['coords']
+        obj_data = obj['raw_data']
+        speed = obj_data[6]
+        center_x, center_y = convert_grid_to_xy(i, j)
+        theta = obj_data[3] * np.pi
+        ori = np.array([math.cos(theta), math.sin(theta)])
+        loc_x = center_x + obj_data[1] + t * speed * ori[0]
+        loc_y = center_y + obj_data[2] - t * speed * ori[1]
+        box = np.array([obj_data[4], obj_data[5]])
+        if obj['type'] == 'pedestrian':
+            box *= 1.5
+        add_rect(
+            img,
+            loc=np.array([loc_x, loc_y]),
+            ori=ori,
+            box=box,
+            value=obj['confidence'],
+            color=CLASS_COLORS[obj['type']]
+        )
+    return img, counts
+
+
+def render_self_car(loc, ori, box, pixels_per_meter=PIXELS_PER_METER):
+    """
+    رسم السيارة الذاتية على الخريطة BEV.
+    Args:
+        loc: موقع السيارة [x, y] في النظام العالمي.
+        ori: اتجاه السيارة [cos(theta), sin(theta)].
+        box: أبعاد السيارة [طول, عرض].
+        pixels_per_meter: عدد البكسلات لكل متر.
+    Returns:
+        self_car_map: خريطة السيارة ذاتية القيادة (RGB - 3 قنوات).
+    """
+    img = np.zeros((IMG_SIZE, IMG_SIZE, 3), np.uint8)
+    center_x = int(loc[0] * pixels_per_meter + MAX_DISTANCE * pixels_per_meter)
+    center_y = int(loc[1] * pixels_per_meter + MAX_DISTANCE * pixels_per_meter)
+    size_px = (
+        int(box[0] * pixels_per_meter),
+        int(box[1] * pixels_per_meter)
+    )
+    angle_deg = -np.degrees(math.atan2(ori[1], ori[0]))
+    box_points = cv2.boxPoints(((center_x, center_y), size_px, angle_deg))
+    box_points = np.int32(box_points)
+    ego_color = (0, 255, 255)  # أصفر
+    cv2.fillConvexPoly(img, box_points, ego_color)
+    return img  # ← نرجع الصورة بأكملها وليس جزءًا منها
+
+def render_waypoints(waypoints, pixels_per_meter=PIXELS_PER_METER):
+    global last_valid_waypoints
+    img = np.zeros((IMG_SIZE, IMG_SIZE, 3), np.uint8)
+    current_waypoints = waypoints
+    if waypoints is not None and len(waypoints) > 2:
+        last_valid_waypoints = waypoints
+    else:
+        current_waypoints = last_valid_waypoints
+    if current_waypoints is None:
+        return img
+    origin_x, origin_y = EGO_CAR_X, EGO_CAR_Y
+    for i, point in enumerate(current_waypoints):
+        px = int(origin_x + point[1] * pixels_per_meter)
+        py = int(origin_y - point[0] * pixels_per_meter)
+        color = (0, 0, 255) if i == len(current_waypoints) - 1 else (0, 255, 0)
+        cv2.circle(img, (px, py), 4, color, -1)
+    return img
+
+
+def collision_detections(map1, map2, threshold=0.04):
+    """
+    تحقق من وجود تداخل بين خريطة البيئة ونموذج السيارة.
+    """
+    print("map1 shape:", map1.shape)
+    print("map2 shape:", map2.shape)
+
+    # تحويل map2 إلى grayscale إذا كانت تحتوي على 3 قنوات (RGB)
+    if len(map2.shape) == 3 and map2.shape[2] == 3:
+        map2 = cv2.cvtColor(map2, cv2.COLOR_BGR2GRAY)
+
+    # التأكد من أن map1 و map2 لها نفس الأبعاد
+    assert map1.shape == map2.shape
+
+    overlap_map = (map1 > 0.01) & (map2 > 0.01)
+    ratio = float(np.sum(overlap_map)) / np.sum(map2 > 0)
+    return ratio < threshold
+
+def get_max_safe_distance(meta_data, downsampled_waypoints, t, collision_buffer, threshold):
+    """
+    حساب أقصى مسافة آمنة قبل حدوث تصادم.
+    """
+    surround_map = meta_data.reshape(20, 20, 7)[..., :3][..., 0]
+    if np.sum(surround_map) < 1:
+        return np.linalg.norm(downsampled_waypoints[-3])
+    hero_bounding_box = np.array([2.45, 1.0]) + collision_buffer
+    safe_distance = 0.0
+    for i in range(len(downsampled_waypoints) - 2):
+        aim = (downsampled_waypoints[i + 1] + downsampled_waypoints[i + 2]) / 2.0
+        loc = downsampled_waypoints[i]
+        ori = aim - loc
+        self_car_map = render_self_car(loc=loc, ori=ori, box=hero_bounding_box, pixels_per_meter=PIXELS_PER_METER)
+        # تصغير الخريطة والتحويل إلى grayscale
+        self_car_map_resized = cv2.resize(self_car_map, (20, 20))
+        self_car_map_gray = cv2.cvtColor(self_car_map_resized, cv2.COLOR_BGR2GRAY)
+        if not collision_detections(surround_map, self_car_map_gray, threshold):
+            break
+        safe_distance = max(safe_distance, np.linalg.norm(loc))
+    return safe_distance
+
+def downsample_waypoints(waypoints, precision=0.2):
+    """
+    تقليل عدد نقاط المسار.
+    """
+    downsampled_waypoints = []
+    last_waypoint = np.array([0.0, 0.0])
+    for i in range(len(waypoints)):
+        now_waypoint = waypoints[i]
+        dis = np.linalg.norm(now_waypoint - last_waypoint)
+        if dis > precision:
+            interval = int(dis / precision)
+            move_vector = (now_waypoint - last_waypoint) / (interval + 1)
+            for j in range(interval):
+                downsampled_waypoints.append(last_waypoint + move_vector * (j + 1))
+        downsampled_waypoints.append(now_waypoint)
+        last_waypoint = now_waypoint
+    return downsampled_waypoints
+
 
 
-# --- الجزء الخامس: بناء و��جهة Gradio ---
-with gr.Blocks(title="Interfuser Demo", theme=gr.themes.Soft()) as demo:
-    gr.Markdown("# 🚀 Interfuser: واجهة القيادة التفاعلية")
-    gr.Markdown("اختر النموذج والسيناريو، ثم اضغط على 'تشغيل' للمقارنة بين أداء النماذج المختلفة.")
 
+# ==============================================================================
+#           Gradio Application Logic
+# ==============================================================================
+
+# --- Load the Model (do this once globally) ---
+print("Loading the Interfuser model...")
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+model = interfuser_baseline() # This function needs to be defined from your script
+# Ensure the model file is in the same directory in your Hugging Face Space
+model_path = "interfuser_best_model.pth"
+if not os.path.exists(model_path):
+    raise FileNotFoundError(f"Model file not found at {model_path}. Please upload it to the Space.")
+
+state_dic = torch.load(model_path, map_location=device, weights_only=True)
+model.load_state_dict(state_dic)
+model.to(device)
+model.eval()
+print("Model loaded successfully.")
+
+def run_single_frame(
+    rgb_image_path: str,
+    rgb_left_image_path: str,
+    rgb_right_image_path: str,
+    rgb_center_image_path: str,
+    lidar_image_path: str,
+    measurements_path: str,
+    target_point_list: list 
+):
+    """
+    تعالج إطارًا واحدًا من البيانات، وتُنشئ لوحة تحكم مرئية كاملة،
+    وتُرجع كلاً من الصورة والبيانات المهيكلة.
+    """
     try:
-        available_models = get_available_models()
-        available_scenarios = [d for d in os.listdir(SAMPLE_DATA_DIR) if os.path.isdir(os.path.join(SAMPLE_DATA_DIR, d))]
-    except FileNotFoundError as e:
-        gr.Error(f"خطأ في الإعداد: {e}. هل قمت بإنشاء مجلدي 'model' و 'sample_data'؟")
-        available_models, available_scenarios = [], []
-
-    with gr.Row():
-        with gr.Column(scale=1, min_width=300):
-            model_selector = gr.Dropdown(
-                choices=available_models, label="1. اختر النموذج",
-                value=available_models[0] if available_models else None,
-                interactive=True
-            )
-            model_load_status = gr.Textbox(label="حالة تحميل النموذج", interactive=False)
-            
-            scenario_selector = gr.Dropdown(
-                choices=available_scenarios, label="2. اختر سيناريو القيادة",
-                value=available_scenarios[0] if available_scenarios else None)
-            
-            run_button = gr.Button("▶️ تشغيل المحاكاة", variant="primary")
-            
-            controller_output = gr.Textbox(label="بيانات متحكم القيادة", interactive=False)
-
-        with gr.Column(scale=3):
-            dashboard_output = gr.Image(label="لوحة المعلومات الحية (Dashboard)", interactive=False)
-
-    # --- ربط الأحداث ---
+        # ==========================================================
+        #           1. قراءة ومعالجة المدخلات من المسارات
+        # ==========================================================
+        if not rgb_image_path:
+          raise gr.Error("الرجاء توفير مسار الصورة الأمامية (RGB).")        
+        # --- أ. قراءة الصور ---
+        # <<< تصحيح: استخدام .name لقراءة الملفات >>>
+        rgb_image_pil = Image.open(rgb_image_path.name)
     
-    # 1. عند تغيير النموذج في القائمة، قم بتحميله
-    model_selector.change(
-        fn=load_model_by_name,
-        inputs=[model_selector],
-        outputs=[model_load_status] # عرض حالة التحميل للمستخدم
-    )
+        # إذا كانت مسارات الصور الأخرى غير متوفرة، استخدم الصورة الأمامية
+        # نتحقق من وجود الكائن نفسه ثم نستخدم .name
+        rgb_left_pil = Image.open(rgb_left_image_path.name) if rgb_left_image_path else rgb_image_pil
+        rgb_right_pil = Image.open(rgb_right_image_path.name) if rgb_right_image_path else rgb_image_pil
+        rgb_center_pil = Image.open(rgb_center_image_path.name) if rgb_center_image_path else rgb_image_pil
+    
+        # --- ب. قراءة ومعالجة الليدار ---
+        if lidar_image_path:
+            # <<< تصحيح: استخدام .name لقراءة ملف .npy >>>
+            lidar_array = np.load(lidar_image_path.name)
+            if lidar_array.max() > 0:
+                lidar_array = (lidar_array / lidar_array.max()) * 255.0
+            lidar_pil = Image.fromarray(lidar_array.astype(np.uint8))
+            lidar_image_pil = lidar_pil.convert('RGB') if lidar_pil.mode != 'RGB' else lidar_pil
+        else:
+            lidar_image_pil = Image.fromarray(np.zeros((112, 112, 3), dtype=np.uint8))
     
-    # 2. عند الضغط على زر التشغيل، قم بتشغيل المحاكاة
-    if available_models and available_scenarios:
-        run_button.click(
-            fn=run_simulation,
-            inputs=[scenario_selector],
-            outputs=[dashboard_output, controller_output]
+        # --- ج. تحويل الصور إلى تنسورات ---
+        rgb_tensor = transform(rgb_image_pil).unsqueeze(0).to(device)
+        rgb_left_tensor = transform(rgb_left_pil).unsqueeze(0).to(device)
+        rgb_right_tensor = transform(rgb_right_pil).unsqueeze(0).to(device)
+        rgb_center_tensor = transform(rgb_center_pil).unsqueeze(0).to(device)
+        lidar_tensor = lidar_transform(lidar_image_pil).unsqueeze(0).to(device)
+        
+        # --- د. قراءة البيانات الرقمية ---
+        # <<< تصحيح: استخدام .name لقراءة ملف JSON >>>
+        with open(measurements_path.name, 'r') as f:
+            measurements_dict = json.load(f)
+
+        measurements_values = [
+            measurements_dict.get('x', 0.0), measurements_dict.get('y', 0.0),
+            measurements_dict.get('theta', 0.0), measurements_dict.get('speed', 5.0),
+            measurements_dict.get('steer', 0.0), measurements_dict.get('throttle', 0.0),
+            measurements_dict.get('brake', 0.0), measurements_dict.get('command', 2.0),
+            measurements_dict.get('is_junction', 0.0), measurements_dict.get('should_brake', 0.0)
+        ]
+        measurements_tensor = torch.tensor([measurements_values], dtype=torch.float32).to(device)
+        target_point_tensor = torch.tensor([target_point_list], dtype=torch.float32).to(device)
+
+        inputs = {
+            'rgb': rgb_tensor, 'rgb_left': rgb_left_tensor, 'rgb_right': rgb_right_tensor,
+            'rgb_center': rgb_center_tensor, 'lidar': lidar_tensor,
+            'measurements': measurements_tensor, 'target_point': target_point_tensor
+        }
+
+        # ==========================================================
+        #           2. تشغيل النموذج والمعالجات اللاحقة
+        # ==========================================================
+        with torch.no_grad():
+            outputs = model(inputs)
+            traffic, waypoints, is_junction, traffic_light, stop_sign, _ = outputs
+
+        measurements_np = measurements_tensor[0].cpu().numpy()
+        pos, theta, speed = measurements_np[:2], measurements_np[2], measurements_np[3]
+        
+        traffic_np = traffic[0].detach().cpu().numpy().reshape(20, 20, -1)
+        waypoints_np = waypoints[0].detach().cpu().numpy() * WAYPOINT_SCALE_FACTOR
+
+        tracker = Tracker()
+        updated_traffic = tracker.update_and_predict(traffic_np.copy(), pos, theta, frame_num=0)
+        
+        controller = InterfuserController(ControllerConfig())
+        steer, throttle, brake, metadata_tuple = controller.run_step(
+            speed=speed, waypoints=waypoints_np, junction=is_junction.sigmoid()[0, 1].item(),
+            traffic_light_state=traffic_light.sigmoid()[0, 0].item(),
+            stop_sign=stop_sign.sigmoid()[0, 1].item(), meta_data=updated_traffic
         )
-    else:
-        gr.Warning("لا يمكن العثور على نماذج أو سيناريوهات. يرجى التأكد من إعداد المجلدات بشكل صحيح.")
+
+        # ==========================================================
+        #           3. إنشاء التصور المرئي (Dashboard)
+        # ==========================================================
+        map_t0, counts_t0 = render(updated_traffic, t=0)
+        map_t1, counts_t1 = render(updated_traffic, t=T1_FUTURE_TIME)
+        map_t2, counts_t2 = render(updated_traffic, t=T2_FUTURE_TIME)
         
-    # تحميل النموذج الافتراضي عند بدء تشغيل الواجهة
-    demo.load(
-        fn=load_model_by_name,
-        inputs=[model_selector],
-        outputs=[model_load_status]
-    )
+        wp_map = render_waypoints(waypoints_np)
+        self_car_map = render_self_car(loc=np.array([0,0]), ori=[math.cos(0), math.sin(0)], box=[4.0, 2.0])
+        
+        map_t0 = cv2.add(cv2.add(map_t0, wp_map), self_car_map)
+        map_t0 = cv2.resize(map_t0, (400, 400))
+        map_t1 = cv2.add(ensure_rgb(map_t1), ensure_rgb(self_car_map)); map_t1 = cv2.resize(map_t1, (200, 200))
+        map_t2 = cv2.add(ensure_rgb(map_t2), ensure_rgb(self_car_map)); map_t2 = cv2.resize(map_t2, (200, 200))
+        
+        display = DisplayInterface()
+        light_state = "Red" if traffic_light.sigmoid()[0,0].item() > 0.5 else "Green"
+        stop_sign_state = "Yes" if stop_sign.sigmoid()[0,1].item() > 0.5 else "No"
+        
+        interface_data = {
+            'camera_view': np.array(rgb_image_pil),
+            'map_t0': map_t0, 'map_t1': map_t1, 'map_t2': map_t2,
+            'text_info': { 
+                'Frame': 'API Frame', 'Control': f"S:{steer:.2f} T:{throttle:.2f} B:{int(brake)}", 
+                'Light': f"L: {light_state}", 'Stop': f"St: {stop_sign_state}" 
+            },
+            'object_counts': {'t0': counts_t0, 't1': counts_t1, 't2': counts_t2}
+        }
+        
+        dashboard_image = display.run_interface(interface_data)
+
+        # ==========================================================
+        #           4. تجهيز وإرجاع المخرجات النهائية
+        # ==========================================================
+        result_dict = {
+            "predicted_waypoints": waypoints_np.tolist(),
+            "control_commands": {"steer": steer, "throttle": throttle, "brake": bool(brake)},
+            "perception": {"traffic_light_status": light_state, "stop_sign_detected": (stop_sign_state == "Yes"), "is_at_junction_prob": round(is_junction.sigmoid()[0,1].item(), 3)},
+            "metadata": {"speed_info": metadata_tuple[0], "perception_info": metadata_tuple[1], "stop_info": metadata_tuple[2], "safe_distance": metadata_tuple[3]}
+        }
+        
+        # تحويل صورة numpy إلى PIL Image قبل إرجاعها
+        return Image.fromarray(dashboard_image), result_dict
+
+    except Exception as e:
+        print(traceback.format_exc())
+        raise gr.Error(f"Error processing single frame: {e}")
+
+with gr.Blocks() as demo:
+    gr.Markdown("# 🚗 محاكاة القيادة الذاتية باستخدام Interfuser")
+    
+    with gr.Tabs():
+               with gr.TabItem("نقطة نهاية API (إطار واحد)", id=1):
+                    gr.Markdown("### اختبار النموذج بإدخال مباشر (Single Frame Inference)")
+                    gr.Markdown("هذه الواجهة مخصصة للمطورين. قم برفع الملفات المطلوبة لتشغيل النموذج على إطار واحد.")
+                    
+                    with gr.Row():
+                        with gr.Column(scale=1):
+                            gr.Markdown("#### ملفات الصور")
+                            api_rgb_image_path = gr.File(label="RGB (Front) File (.jpg, .png)")
+                            api_rgb_left_image_path = gr.File(label="RGB (Left) File (Optional)")
+                            api_rgb_right_image_path = gr.File(label="RGB (Right) File (Optional)")
+                            api_rgb_center_image_path = gr.File(label="RGB (Center) File (Optional)")
+                            api_lidar_image_path = gr.File(label="LiDAR File (.npy, Optional)")
+                        
+                        with gr.Column(scale=2):
+                            gr.Markdown("#### ملفات ومحتويات البيانات")
+                            api_measurements_path = gr.File(label="Measurements File (.json)")
+                            api_target_point_list = gr.JSON(label="Target Point (List [x, y])", value=[0.0, 100.0])
+                            api_output_image = gr.Image(label="Dashboard Result", type="pil")
+                            api_output_json = gr.JSON(label="نتائج النموذج (JSON)")
+                            gr.Markdown("---")
+                            api_run_button = gr.Button("🚀 تشغيل إطار واحد", variant="primary")
+                    gr.Markdown("---")
+                    gr.Markdown("#### المخرجات")
+                    
+                    # <<< بداية التعديل: إضافة مكون لعرض الصورة الناتجة >>>
+                    with gr.Row():
+                        # سيتم عرض لوحة التحكم هنا
+                        api_output_image = gr.Image(label="Dashboard Result", type="pil")
+                        # سيتم عرض بيانات JSON هنا
+                        api_output_json = gr.JSON(label="نتائج النموذج (JSON)")
+        
+       api_run_button.click(
+            fn=run_single_frame,
+            inputs=[
+                api_rgb_image_path, 
+                api_rgb_left_image_path, 
+                api_rgb_right_image_path, 
+                api_rgb_center_image_path, 
+                api_lidar_image_path,
+                api_measurements_path, 
+                api_target_point_list
+            ],
+            # الآن نربط المخرجين اللذين تُرجعهما الدالة بالمكونين الصحيحين
+            outputs=[api_output_image, api_output_json],
+            api_name="run_single_frame"
+        )
 
-# --- الجزء السادس: إطلاق الواجهة ---
+# ==============================================================================
+#           7. تشغيل التطبيق
+# ==============================================================================
 if __name__ == "__main__":
-    demo.launch(debug=True)
+    demo.queue().launch(debug=True)