Spaces:
Runtime error
Runtime error
Upload model_definition.py with huggingface_hub
Browse files- model_definition.py +1394 -0
model_definition.py
ADDED
@@ -0,0 +1,1394 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# model_definition.py
|
2 |
+
# ============================================================================
|
3 |
+
# الاستيرادات الأساسية
|
4 |
+
# ============================================================================
|
5 |
+
import torch
|
6 |
+
import torch.nn as nn
|
7 |
+
import torch.nn.functional as F
|
8 |
+
import torch.optim as optim
|
9 |
+
from torch.optim import AdamW
|
10 |
+
from torch.optim.lr_scheduler import OneCycleLR
|
11 |
+
from torch.utils.data import Dataset, DataLoader
|
12 |
+
from torchvision import transforms
|
13 |
+
from functools import partial
|
14 |
+
from typing import Optional, List
|
15 |
+
from torch import Tensor
|
16 |
+
import os
|
17 |
+
import json
|
18 |
+
import numpy as np
|
19 |
+
import cv2
|
20 |
+
from PIL import Image
|
21 |
+
import torch
|
22 |
+
import torch.nn as nn
|
23 |
+
import torch.nn.functional as F
|
24 |
+
from torchvision import transforms
|
25 |
+
from functools import partial
|
26 |
+
from collections import deque, OrderedDict
|
27 |
+
import math
|
28 |
+
from torch.nn import MultiheadAttention
|
29 |
+
from torch.nn import TransformerEncoder, TransformerEncoderLayer
|
30 |
+
from torch.nn import TransformerDecoder, TransformerDecoderLayer
|
31 |
+
from timm.models.resnet import resnet50d, resnet26d, resnet18d
|
32 |
+
try:
|
33 |
+
from timm.layers import trunc_normal_
|
34 |
+
except ImportError:
|
35 |
+
from timm.models.layers import trunc_normal_
|
36 |
+
from huggingface_hub import hf_hub_download, HfApi
|
37 |
+
from huggingface_hub.utils import HfFolder
|
38 |
+
|
39 |
+
# مكتبات إضافية
|
40 |
+
import os
|
41 |
+
import json
|
42 |
+
import logging
|
43 |
+
import math
|
44 |
+
import copy
|
45 |
+
from pathlib import Path
|
46 |
+
from collections import OrderedDict
|
47 |
+
|
48 |
+
# مكتبات معالجة البيانات
|
49 |
+
import numpy as np
|
50 |
+
import cv2
|
51 |
+
|
52 |
+
# مكتبات اختيارية (يمكن تعطيلها إذا لم تكن متوفرة)
|
53 |
+
|
54 |
+
try:
|
55 |
+
from tqdm import tqdm
|
56 |
+
except ImportError:
|
57 |
+
# إذا لم تكن tqdm متوفرة، استخدم دالة بديلة
|
58 |
+
def tqdm(iterable, *args, **kwargs):
|
59 |
+
return iterable
|
60 |
+
|
61 |
+
# ============================================================================
|
62 |
+
# دوال مساعدة
|
63 |
+
# ============================================================================
|
64 |
+
def to_2tuple(x):
|
65 |
+
"""تحويل قيمة إلى tuple من عنصرين"""
|
66 |
+
if isinstance(x, (list, tuple)):
|
67 |
+
return tuple(x)
|
68 |
+
return (x, x)
|
69 |
+
# ============================================================================
|
70 |
+
# ============================================================================
|
71 |
+
|
72 |
+
class HybridEmbed(nn.Module):
|
73 |
+
def __init__(
|
74 |
+
self,
|
75 |
+
backbone,
|
76 |
+
img_size=224,
|
77 |
+
patch_size=1,
|
78 |
+
feature_size=None,
|
79 |
+
in_chans=3,
|
80 |
+
embed_dim=768,
|
81 |
+
):
|
82 |
+
super().__init__()
|
83 |
+
assert isinstance(backbone, nn.Module)
|
84 |
+
img_size = to_2tuple(img_size)
|
85 |
+
patch_size = to_2tuple(patch_size)
|
86 |
+
self.img_size = img_size
|
87 |
+
self.patch_size = patch_size
|
88 |
+
self.backbone = backbone
|
89 |
+
if feature_size is None:
|
90 |
+
with torch.no_grad():
|
91 |
+
training = backbone.training
|
92 |
+
if training:
|
93 |
+
backbone.eval()
|
94 |
+
o = self.backbone(torch.zeros(1, in_chans, img_size[0], img_size[1]))
|
95 |
+
if isinstance(o, (list, tuple)):
|
96 |
+
o = o[-1] # last feature if backbone outputs list/tuple of features
|
97 |
+
feature_size = o.shape[-2:]
|
98 |
+
feature_dim = o.shape[1]
|
99 |
+
backbone.train(training)
|
100 |
+
else:
|
101 |
+
feature_size = to_2tuple(feature_size)
|
102 |
+
if hasattr(self.backbone, "feature_info"):
|
103 |
+
feature_dim = self.backbone.feature_info.channels()[-1]
|
104 |
+
else:
|
105 |
+
feature_dim = self.backbone.num_features
|
106 |
+
|
107 |
+
self.proj = nn.Conv2d(feature_dim, embed_dim, kernel_size=1, stride=1)
|
108 |
+
|
109 |
+
def forward(self, x):
|
110 |
+
x = self.backbone(x)
|
111 |
+
if isinstance(x, (list, tuple)):
|
112 |
+
x = x[-1] # last feature if backbone outputs list/tuple of features
|
113 |
+
x = self.proj(x)
|
114 |
+
global_x = torch.mean(x, [2, 3], keepdim=False)[:, :, None]
|
115 |
+
return x, global_x
|
116 |
+
|
117 |
+
|
118 |
+
class PositionEmbeddingSine(nn.Module):
|
119 |
+
"""
|
120 |
+
This is a more standard version of the position embedding, very similar to the one
|
121 |
+
used by the Attention is all you need paper, generalized to work on images.
|
122 |
+
"""
|
123 |
+
|
124 |
+
def __init__(
|
125 |
+
self, num_pos_feats=64, temperature=10000, normalize=False, scale=None
|
126 |
+
):
|
127 |
+
super().__init__()
|
128 |
+
self.num_pos_feats = num_pos_feats
|
129 |
+
self.temperature = temperature
|
130 |
+
self.normalize = normalize
|
131 |
+
if scale is not None and normalize is False:
|
132 |
+
raise ValueError("normalize should be True if scale is passed")
|
133 |
+
if scale is None:
|
134 |
+
scale = 2 * math.pi
|
135 |
+
self.scale = scale
|
136 |
+
|
137 |
+
def forward(self, tensor):
|
138 |
+
x = tensor
|
139 |
+
bs, _, h, w = x.shape
|
140 |
+
not_mask = torch.ones((bs, h, w), device=x.device)
|
141 |
+
y_embed = not_mask.cumsum(1, dtype=torch.float32)
|
142 |
+
x_embed = not_mask.cumsum(2, dtype=torch.float32)
|
143 |
+
if self.normalize:
|
144 |
+
eps = 1e-6
|
145 |
+
y_embed = y_embed / (y_embed[:, -1:, :] + eps) * self.scale
|
146 |
+
x_embed = x_embed / (x_embed[:, :, -1:] + eps) * self.scale
|
147 |
+
|
148 |
+
dim_t = torch.arange(self.num_pos_feats, dtype=torch.float32, device=x.device)
|
149 |
+
dim_t = self.temperature ** (2 * (dim_t // 2) / self.num_pos_feats)
|
150 |
+
|
151 |
+
pos_x = x_embed[:, :, :, None] / dim_t
|
152 |
+
pos_y = y_embed[:, :, :, None] / dim_t
|
153 |
+
pos_x = torch.stack(
|
154 |
+
(pos_x[:, :, :, 0::2].sin(), pos_x[:, :, :, 1::2].cos()), dim=4
|
155 |
+
).flatten(3)
|
156 |
+
pos_y = torch.stack(
|
157 |
+
(pos_y[:, :, :, 0::2].sin(), pos_y[:, :, :, 1::2].cos()), dim=4
|
158 |
+
).flatten(3)
|
159 |
+
pos = torch.cat((pos_y, pos_x), dim=3).permute(0, 3, 1, 2)
|
160 |
+
return pos
|
161 |
+
|
162 |
+
|
163 |
+
class TransformerEncoder(nn.Module):
|
164 |
+
def __init__(self, encoder_layer, num_layers, norm=None):
|
165 |
+
super().__init__()
|
166 |
+
self.layers = _get_clones(encoder_layer, num_layers)
|
167 |
+
self.num_layers = num_layers
|
168 |
+
self.norm = norm
|
169 |
+
|
170 |
+
def forward(
|
171 |
+
self,
|
172 |
+
src,
|
173 |
+
mask: Optional[Tensor] = None,
|
174 |
+
src_key_padding_mask: Optional[Tensor] = None,
|
175 |
+
pos: Optional[Tensor] = None,
|
176 |
+
):
|
177 |
+
output = src
|
178 |
+
|
179 |
+
for layer in self.layers:
|
180 |
+
output = layer(
|
181 |
+
output,
|
182 |
+
src_mask=mask,
|
183 |
+
src_key_padding_mask=src_key_padding_mask,
|
184 |
+
pos=pos,
|
185 |
+
)
|
186 |
+
|
187 |
+
if self.norm is not None:
|
188 |
+
output = self.norm(output)
|
189 |
+
|
190 |
+
return output
|
191 |
+
|
192 |
+
|
193 |
+
class SpatialSoftmax(nn.Module):
|
194 |
+
def __init__(self, height, width, channel, temperature=None, data_format="NCHW"):
|
195 |
+
super().__init__()
|
196 |
+
|
197 |
+
self.data_format = data_format
|
198 |
+
self.height = height
|
199 |
+
self.width = width
|
200 |
+
self.channel = channel
|
201 |
+
|
202 |
+
if temperature:
|
203 |
+
self.temperature = Parameter(torch.ones(1) * temperature)
|
204 |
+
else:
|
205 |
+
self.temperature = 1.0
|
206 |
+
|
207 |
+
pos_x, pos_y = np.meshgrid(
|
208 |
+
np.linspace(-1.0, 1.0, self.height), np.linspace(-1.0, 1.0, self.width)
|
209 |
+
)
|
210 |
+
pos_x = torch.from_numpy(pos_x.reshape(self.height * self.width)).float()
|
211 |
+
pos_y = torch.from_numpy(pos_y.reshape(self.height * self.width)).float()
|
212 |
+
self.register_buffer("pos_x", pos_x)
|
213 |
+
self.register_buffer("pos_y", pos_y)
|
214 |
+
|
215 |
+
def forward(self, feature):
|
216 |
+
# Output:
|
217 |
+
# (N, C*2) x_0 y_0 ...
|
218 |
+
|
219 |
+
if self.data_format == "NHWC":
|
220 |
+
feature = (
|
221 |
+
feature.transpose(1, 3)
|
222 |
+
.tranpose(2, 3)
|
223 |
+
.view(-1, self.height * self.width)
|
224 |
+
)
|
225 |
+
else:
|
226 |
+
feature = feature.view(-1, self.height * self.width)
|
227 |
+
|
228 |
+
weight = F.softmax(feature / self.temperature, dim=-1)
|
229 |
+
expected_x = torch.sum(
|
230 |
+
torch.autograd.Variable(self.pos_x) * weight, dim=1, keepdim=True
|
231 |
+
)
|
232 |
+
expected_y = torch.sum(
|
233 |
+
torch.autograd.Variable(self.pos_y) * weight, dim=1, keepdim=True
|
234 |
+
)
|
235 |
+
expected_xy = torch.cat([expected_x, expected_y], 1)
|
236 |
+
feature_keypoints = expected_xy.view(-1, self.channel, 2)
|
237 |
+
feature_keypoints[:, :, 1] = (feature_keypoints[:, :, 1] - 1) * 12
|
238 |
+
feature_keypoints[:, :, 0] = feature_keypoints[:, :, 0] * 12
|
239 |
+
return feature_keypoints
|
240 |
+
|
241 |
+
|
242 |
+
class MultiPath_Generator(nn.Module):
|
243 |
+
def __init__(self, in_channel, embed_dim, out_channel):
|
244 |
+
super().__init__()
|
245 |
+
self.spatial_softmax = SpatialSoftmax(100, 100, out_channel)
|
246 |
+
self.tconv0 = nn.Sequential(
|
247 |
+
nn.ConvTranspose2d(in_channel, 256, 4, 2, 1, bias=False),
|
248 |
+
nn.BatchNorm2d(256),
|
249 |
+
nn.ReLU(True),
|
250 |
+
)
|
251 |
+
self.tconv1 = nn.Sequential(
|
252 |
+
nn.ConvTranspose2d(256, 256, 4, 2, 1, bias=False),
|
253 |
+
nn.BatchNorm2d(256),
|
254 |
+
nn.ReLU(True),
|
255 |
+
)
|
256 |
+
self.tconv2 = nn.Sequential(
|
257 |
+
nn.ConvTranspose2d(256, 192, 4, 2, 1, bias=False),
|
258 |
+
nn.BatchNorm2d(192),
|
259 |
+
nn.ReLU(True),
|
260 |
+
)
|
261 |
+
self.tconv3 = nn.Sequential(
|
262 |
+
nn.ConvTranspose2d(192, 64, 4, 2, 1, bias=False),
|
263 |
+
nn.BatchNorm2d(64),
|
264 |
+
nn.ReLU(True),
|
265 |
+
)
|
266 |
+
self.tconv4_list = torch.nn.ModuleList(
|
267 |
+
[
|
268 |
+
nn.Sequential(
|
269 |
+
nn.ConvTranspose2d(64, out_channel, 8, 2, 3, bias=False),
|
270 |
+
nn.Tanh(),
|
271 |
+
)
|
272 |
+
for _ in range(6)
|
273 |
+
]
|
274 |
+
)
|
275 |
+
|
276 |
+
self.upsample = nn.Upsample(size=(50, 50), mode="bilinear")
|
277 |
+
|
278 |
+
def forward(self, x, measurements):
|
279 |
+
mask = measurements[:, :6]
|
280 |
+
mask = mask.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1).repeat(1, 1, 1, 100, 100)
|
281 |
+
velocity = measurements[:, 6:7].unsqueeze(-1).unsqueeze(-1)
|
282 |
+
velocity = velocity.repeat(1, 32, 2, 2)
|
283 |
+
|
284 |
+
n, d, c = x.shape
|
285 |
+
x = x.transpose(1, 2)
|
286 |
+
x = x.view(n, -1, 2, 2)
|
287 |
+
x = torch.cat([x, velocity], dim=1)
|
288 |
+
x = self.tconv0(x)
|
289 |
+
x = self.tconv1(x)
|
290 |
+
x = self.tconv2(x)
|
291 |
+
x = self.tconv3(x)
|
292 |
+
x = self.upsample(x)
|
293 |
+
xs = []
|
294 |
+
for i in range(6):
|
295 |
+
xt = self.tconv4_list[i](x)
|
296 |
+
xs.append(xt)
|
297 |
+
xs = torch.stack(xs, dim=1)
|
298 |
+
x = torch.sum(xs * mask, dim=1)
|
299 |
+
x = self.spatial_softmax(x)
|
300 |
+
return x
|
301 |
+
|
302 |
+
|
303 |
+
class LinearWaypointsPredictor(nn.Module):
|
304 |
+
def __init__(self, input_dim, cumsum=True):
|
305 |
+
super().__init__()
|
306 |
+
self.cumsum = cumsum
|
307 |
+
self.rank_embed = nn.Parameter(torch.zeros(1, 10, input_dim))
|
308 |
+
self.head_fc1_list = nn.ModuleList([nn.Linear(input_dim, 64) for _ in range(6)])
|
309 |
+
self.head_relu = nn.ReLU(inplace=True)
|
310 |
+
self.head_fc2_list = nn.ModuleList([nn.Linear(64, 2) for _ in range(6)])
|
311 |
+
|
312 |
+
def forward(self, x, measurements):
|
313 |
+
# input shape: n 10 embed_dim
|
314 |
+
bs, n, dim = x.shape
|
315 |
+
x = x + self.rank_embed
|
316 |
+
x = x.reshape(-1, dim)
|
317 |
+
|
318 |
+
mask = measurements[:, :6]
|
319 |
+
mask = torch.unsqueeze(mask, -1).repeat(n, 1, 2)
|
320 |
+
|
321 |
+
rs = []
|
322 |
+
for i in range(6):
|
323 |
+
res = self.head_fc1_list[i](x)
|
324 |
+
res = self.head_relu(res)
|
325 |
+
res = self.head_fc2_list[i](res)
|
326 |
+
rs.append(res)
|
327 |
+
rs = torch.stack(rs, 1)
|
328 |
+
x = torch.sum(rs * mask, dim=1)
|
329 |
+
|
330 |
+
x = x.view(bs, n, 2)
|
331 |
+
if self.cumsum:
|
332 |
+
x = torch.cumsum(x, 1)
|
333 |
+
return x
|
334 |
+
|
335 |
+
|
336 |
+
class GRUWaypointsPredictor(nn.Module):
|
337 |
+
def __init__(self, input_dim, waypoints=10):
|
338 |
+
super().__init__()
|
339 |
+
# self.gru = torch.nn.GRUCell(input_size=input_dim, hidden_size=64)
|
340 |
+
self.gru = torch.nn.GRU(input_size=input_dim, hidden_size=64, batch_first=True)
|
341 |
+
self.encoder = nn.Linear(2, 64)
|
342 |
+
self.decoder = nn.Linear(64, 2)
|
343 |
+
self.waypoints = waypoints
|
344 |
+
|
345 |
+
def forward(self, x, target_point):
|
346 |
+
bs = x.shape[0]
|
347 |
+
z = self.encoder(target_point).unsqueeze(0)
|
348 |
+
output, _ = self.gru(x, z)
|
349 |
+
output = output.reshape(bs * self.waypoints, -1)
|
350 |
+
output = self.decoder(output).reshape(bs, self.waypoints, 2)
|
351 |
+
output = torch.cumsum(output, 1)
|
352 |
+
return output
|
353 |
+
|
354 |
+
class GRUWaypointsPredictorWithCommand(nn.Module):
|
355 |
+
def __init__(self, input_dim, waypoints=10):
|
356 |
+
super().__init__()
|
357 |
+
# self.gru = torch.nn.GRUCell(input_size=input_dim, hidden_size=64)
|
358 |
+
self.grus = nn.ModuleList([torch.nn.GRU(input_size=input_dim, hidden_size=64, batch_first=True) for _ in range(6)])
|
359 |
+
self.encoder = nn.Linear(2, 64)
|
360 |
+
self.decoders = nn.ModuleList([nn.Linear(64, 2) for _ in range(6)])
|
361 |
+
self.waypoints = waypoints
|
362 |
+
|
363 |
+
def forward(self, x, target_point, measurements):
|
364 |
+
bs, n, dim = x.shape
|
365 |
+
mask = measurements[:, :6, None, None]
|
366 |
+
mask = mask.repeat(1, 1, self.waypoints, 2)
|
367 |
+
|
368 |
+
z = self.encoder(target_point).unsqueeze(0)
|
369 |
+
outputs = []
|
370 |
+
for i in range(6):
|
371 |
+
output, _ = self.grus[i](x, z)
|
372 |
+
output = output.reshape(bs * self.waypoints, -1)
|
373 |
+
output = self.decoders[i](output).reshape(bs, self.waypoints, 2)
|
374 |
+
output = torch.cumsum(output, 1)
|
375 |
+
outputs.append(output)
|
376 |
+
outputs = torch.stack(outputs, 1)
|
377 |
+
output = torch.sum(outputs * mask, dim=1)
|
378 |
+
return output
|
379 |
+
|
380 |
+
|
381 |
+
class TransformerDecoder(nn.Module):
|
382 |
+
def __init__(self, decoder_layer, num_layers, norm=None, return_intermediate=False):
|
383 |
+
super().__init__()
|
384 |
+
self.layers = _get_clones(decoder_layer, num_layers)
|
385 |
+
self.num_layers = num_layers
|
386 |
+
self.norm = norm
|
387 |
+
self.return_intermediate = return_intermediate
|
388 |
+
|
389 |
+
def forward(
|
390 |
+
self,
|
391 |
+
tgt,
|
392 |
+
memory,
|
393 |
+
tgt_mask: Optional[Tensor] = None,
|
394 |
+
memory_mask: Optional[Tensor] = None,
|
395 |
+
tgt_key_padding_mask: Optional[Tensor] = None,
|
396 |
+
memory_key_padding_mask: Optional[Tensor] = None,
|
397 |
+
pos: Optional[Tensor] = None,
|
398 |
+
query_pos: Optional[Tensor] = None,
|
399 |
+
):
|
400 |
+
output = tgt
|
401 |
+
|
402 |
+
intermediate = []
|
403 |
+
|
404 |
+
for layer in self.layers:
|
405 |
+
output = layer(
|
406 |
+
output,
|
407 |
+
memory,
|
408 |
+
tgt_mask=tgt_mask,
|
409 |
+
memory_mask=memory_mask,
|
410 |
+
tgt_key_padding_mask=tgt_key_padding_mask,
|
411 |
+
memory_key_padding_mask=memory_key_padding_mask,
|
412 |
+
pos=pos,
|
413 |
+
query_pos=query_pos,
|
414 |
+
)
|
415 |
+
if self.return_intermediate:
|
416 |
+
intermediate.append(self.norm(output))
|
417 |
+
|
418 |
+
if self.norm is not None:
|
419 |
+
output = self.norm(output)
|
420 |
+
if self.return_intermediate:
|
421 |
+
intermediate.pop()
|
422 |
+
intermediate.append(output)
|
423 |
+
|
424 |
+
if self.return_intermediate:
|
425 |
+
return torch.stack(intermediate)
|
426 |
+
|
427 |
+
return output.unsqueeze(0)
|
428 |
+
|
429 |
+
|
430 |
+
class TransformerEncoderLayer(nn.Module):
|
431 |
+
def __init__(
|
432 |
+
self,
|
433 |
+
d_model,
|
434 |
+
nhead,
|
435 |
+
dim_feedforward=2048,
|
436 |
+
dropout=0.1,
|
437 |
+
activation=nn.ReLU(),
|
438 |
+
normalize_before=False,
|
439 |
+
):
|
440 |
+
super().__init__()
|
441 |
+
self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=dropout)
|
442 |
+
# Implementation of Feedforward model
|
443 |
+
self.linear1 = nn.Linear(d_model, dim_feedforward)
|
444 |
+
self.dropout = nn.Dropout(dropout)
|
445 |
+
self.linear2 = nn.Linear(dim_feedforward, d_model)
|
446 |
+
|
447 |
+
self.norm1 = nn.LayerNorm(d_model)
|
448 |
+
self.norm2 = nn.LayerNorm(d_model)
|
449 |
+
self.dropout1 = nn.Dropout(dropout)
|
450 |
+
self.dropout2 = nn.Dropout(dropout)
|
451 |
+
|
452 |
+
self.activation = activation()
|
453 |
+
self.normalize_before = normalize_before
|
454 |
+
|
455 |
+
def with_pos_embed(self, tensor, pos: Optional[Tensor]):
|
456 |
+
return tensor if pos is None else tensor + pos
|
457 |
+
|
458 |
+
def forward_post(
|
459 |
+
self,
|
460 |
+
src,
|
461 |
+
src_mask: Optional[Tensor] = None,
|
462 |
+
src_key_padding_mask: Optional[Tensor] = None,
|
463 |
+
pos: Optional[Tensor] = None,
|
464 |
+
):
|
465 |
+
q = k = self.with_pos_embed(src, pos)
|
466 |
+
src2 = self.self_attn(
|
467 |
+
q, k, value=src, attn_mask=src_mask, key_padding_mask=src_key_padding_mask
|
468 |
+
)[0]
|
469 |
+
src = src + self.dropout1(src2)
|
470 |
+
src = self.norm1(src)
|
471 |
+
src2 = self.linear2(self.dropout(self.activation(self.linear1(src))))
|
472 |
+
src = src + self.dropout2(src2)
|
473 |
+
src = self.norm2(src)
|
474 |
+
return src
|
475 |
+
|
476 |
+
def forward_pre(
|
477 |
+
self,
|
478 |
+
src,
|
479 |
+
src_mask: Optional[Tensor] = None,
|
480 |
+
src_key_padding_mask: Optional[Tensor] = None,
|
481 |
+
pos: Optional[Tensor] = None,
|
482 |
+
):
|
483 |
+
src2 = self.norm1(src)
|
484 |
+
q = k = self.with_pos_embed(src2, pos)
|
485 |
+
src2 = self.self_attn(
|
486 |
+
q, k, value=src2, attn_mask=src_mask, key_padding_mask=src_key_padding_mask
|
487 |
+
)[0]
|
488 |
+
src = src + self.dropout1(src2)
|
489 |
+
src2 = self.norm2(src)
|
490 |
+
src2 = self.linear2(self.dropout(self.activation(self.linear1(src2))))
|
491 |
+
src = src + self.dropout2(src2)
|
492 |
+
return src
|
493 |
+
|
494 |
+
def forward(
|
495 |
+
self,
|
496 |
+
src,
|
497 |
+
src_mask: Optional[Tensor] = None,
|
498 |
+
src_key_padding_mask: Optional[Tensor] = None,
|
499 |
+
pos: Optional[Tensor] = None,
|
500 |
+
):
|
501 |
+
if self.normalize_before:
|
502 |
+
return self.forward_pre(src, src_mask, src_key_padding_mask, pos)
|
503 |
+
return self.forward_post(src, src_mask, src_key_padding_mask, pos)
|
504 |
+
|
505 |
+
|
506 |
+
class TransformerDecoderLayer(nn.Module):
|
507 |
+
def __init__(
|
508 |
+
self,
|
509 |
+
d_model,
|
510 |
+
nhead,
|
511 |
+
dim_feedforward=2048,
|
512 |
+
dropout=0.1,
|
513 |
+
activation=nn.ReLU(),
|
514 |
+
normalize_before=False,
|
515 |
+
):
|
516 |
+
super().__init__()
|
517 |
+
self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=dropout)
|
518 |
+
self.multihead_attn = nn.MultiheadAttention(d_model, nhead, dropout=dropout)
|
519 |
+
# Implementation of Feedforward model
|
520 |
+
self.linear1 = nn.Linear(d_model, dim_feedforward)
|
521 |
+
self.dropout = nn.Dropout(dropout)
|
522 |
+
self.linear2 = nn.Linear(dim_feedforward, d_model)
|
523 |
+
|
524 |
+
self.norm1 = nn.LayerNorm(d_model)
|
525 |
+
self.norm2 = nn.LayerNorm(d_model)
|
526 |
+
self.norm3 = nn.LayerNorm(d_model)
|
527 |
+
self.dropout1 = nn.Dropout(dropout)
|
528 |
+
self.dropout2 = nn.Dropout(dropout)
|
529 |
+
self.dropout3 = nn.Dropout(dropout)
|
530 |
+
|
531 |
+
self.activation = activation()
|
532 |
+
self.normalize_before = normalize_before
|
533 |
+
|
534 |
+
def with_pos_embed(self, tensor, pos: Optional[Tensor]):
|
535 |
+
return tensor if pos is None else tensor + pos
|
536 |
+
|
537 |
+
def forward_post(
|
538 |
+
self,
|
539 |
+
tgt,
|
540 |
+
memory,
|
541 |
+
tgt_mask: Optional[Tensor] = None,
|
542 |
+
memory_mask: Optional[Tensor] = None,
|
543 |
+
tgt_key_padding_mask: Optional[Tensor] = None,
|
544 |
+
memory_key_padding_mask: Optional[Tensor] = None,
|
545 |
+
pos: Optional[Tensor] = None,
|
546 |
+
query_pos: Optional[Tensor] = None,
|
547 |
+
):
|
548 |
+
q = k = self.with_pos_embed(tgt, query_pos)
|
549 |
+
tgt2 = self.self_attn(
|
550 |
+
q, k, value=tgt, attn_mask=tgt_mask, key_padding_mask=tgt_key_padding_mask
|
551 |
+
)[0]
|
552 |
+
tgt = tgt + self.dropout1(tgt2)
|
553 |
+
tgt = self.norm1(tgt)
|
554 |
+
tgt2 = self.multihead_attn(
|
555 |
+
query=self.with_pos_embed(tgt, query_pos),
|
556 |
+
key=self.with_pos_embed(memory, pos),
|
557 |
+
value=memory,
|
558 |
+
attn_mask=memory_mask,
|
559 |
+
key_padding_mask=memory_key_padding_mask,
|
560 |
+
)[0]
|
561 |
+
tgt = tgt + self.dropout2(tgt2)
|
562 |
+
tgt = self.norm2(tgt)
|
563 |
+
tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt))))
|
564 |
+
tgt = tgt + self.dropout3(tgt2)
|
565 |
+
tgt = self.norm3(tgt)
|
566 |
+
return tgt
|
567 |
+
|
568 |
+
def forward_pre(
|
569 |
+
self,
|
570 |
+
tgt,
|
571 |
+
memory,
|
572 |
+
tgt_mask: Optional[Tensor] = None,
|
573 |
+
memory_mask: Optional[Tensor] = None,
|
574 |
+
tgt_key_padding_mask: Optional[Tensor] = None,
|
575 |
+
memory_key_padding_mask: Optional[Tensor] = None,
|
576 |
+
pos: Optional[Tensor] = None,
|
577 |
+
query_pos: Optional[Tensor] = None,
|
578 |
+
):
|
579 |
+
tgt2 = self.norm1(tgt)
|
580 |
+
q = k = self.with_pos_embed(tgt2, query_pos)
|
581 |
+
tgt2 = self.self_attn(
|
582 |
+
q, k, value=tgt2, attn_mask=tgt_mask, key_padding_mask=tgt_key_padding_mask
|
583 |
+
)[0]
|
584 |
+
tgt = tgt + self.dropout1(tgt2)
|
585 |
+
tgt2 = self.norm2(tgt)
|
586 |
+
tgt2 = self.multihead_attn(
|
587 |
+
query=self.with_pos_embed(tgt2, query_pos),
|
588 |
+
key=self.with_pos_embed(memory, pos),
|
589 |
+
value=memory,
|
590 |
+
attn_mask=memory_mask,
|
591 |
+
key_padding_mask=memory_key_padding_mask,
|
592 |
+
)[0]
|
593 |
+
tgt = tgt + self.dropout2(tgt2)
|
594 |
+
tgt2 = self.norm3(tgt)
|
595 |
+
tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt2))))
|
596 |
+
tgt = tgt + self.dropout3(tgt2)
|
597 |
+
return tgt
|
598 |
+
|
599 |
+
def forward(
|
600 |
+
self,
|
601 |
+
tgt,
|
602 |
+
memory,
|
603 |
+
tgt_mask: Optional[Tensor] = None,
|
604 |
+
memory_mask: Optional[Tensor] = None,
|
605 |
+
tgt_key_padding_mask: Optional[Tensor] = None,
|
606 |
+
memory_key_padding_mask: Optional[Tensor] = None,
|
607 |
+
pos: Optional[Tensor] = None,
|
608 |
+
query_pos: Optional[Tensor] = None,
|
609 |
+
):
|
610 |
+
if self.normalize_before:
|
611 |
+
return self.forward_pre(
|
612 |
+
tgt,
|
613 |
+
memory,
|
614 |
+
tgt_mask,
|
615 |
+
memory_mask,
|
616 |
+
tgt_key_padding_mask,
|
617 |
+
memory_key_padding_mask,
|
618 |
+
pos,
|
619 |
+
query_pos,
|
620 |
+
)
|
621 |
+
return self.forward_post(
|
622 |
+
tgt,
|
623 |
+
memory,
|
624 |
+
tgt_mask,
|
625 |
+
memory_mask,
|
626 |
+
tgt_key_padding_mask,
|
627 |
+
memory_key_padding_mask,
|
628 |
+
pos,
|
629 |
+
query_pos,
|
630 |
+
)
|
631 |
+
|
632 |
+
|
633 |
+
def _get_clones(module, N):
|
634 |
+
return nn.ModuleList([copy.deepcopy(module) for i in range(N)])
|
635 |
+
|
636 |
+
|
637 |
+
def _get_activation_fn(activation):
|
638 |
+
"""Return an activation function given a string"""
|
639 |
+
if activation == "relu":
|
640 |
+
return F.relu
|
641 |
+
if activation == "gelu":
|
642 |
+
return F.gelu
|
643 |
+
if activation == "glu":
|
644 |
+
return F.glu
|
645 |
+
raise RuntimeError(f"activation should be relu/gelu, not {activation}.")
|
646 |
+
|
647 |
+
|
648 |
+
def build_attn_mask(mask_type):
|
649 |
+
mask = torch.ones((151, 151), dtype=torch.bool).cuda()
|
650 |
+
if mask_type == "seperate_all":
|
651 |
+
mask[:50, :50] = False
|
652 |
+
mask[50:67, 50:67] = False
|
653 |
+
mask[67:84, 67:84] = False
|
654 |
+
mask[84:101, 84:101] = False
|
655 |
+
mask[101:151, 101:151] = False
|
656 |
+
elif mask_type == "seperate_view":
|
657 |
+
mask[:50, :50] = False
|
658 |
+
mask[50:67, 50:67] = False
|
659 |
+
mask[67:84, 67:84] = False
|
660 |
+
mask[84:101, 84:101] = False
|
661 |
+
mask[101:151, :] = False
|
662 |
+
mask[:, 101:151] = False
|
663 |
+
return mask
|
664 |
+
# class InterfuserModel(nn.Module):
|
665 |
+
|
666 |
+
class InterfuserModel(nn.Module):
|
667 |
+
def __init__(
|
668 |
+
self,
|
669 |
+
img_size=224,
|
670 |
+
multi_view_img_size=112,
|
671 |
+
patch_size=8,
|
672 |
+
in_chans=3,
|
673 |
+
embed_dim=768,
|
674 |
+
enc_depth=6,
|
675 |
+
dec_depth=6,
|
676 |
+
dim_feedforward=2048,
|
677 |
+
normalize_before=False,
|
678 |
+
rgb_backbone_name="r50",
|
679 |
+
lidar_backbone_name="r50",
|
680 |
+
num_heads=8,
|
681 |
+
norm_layer=None,
|
682 |
+
dropout=0.1,
|
683 |
+
end2end=False,
|
684 |
+
direct_concat=False,
|
685 |
+
separate_view_attention=False,
|
686 |
+
separate_all_attention=False,
|
687 |
+
act_layer=None,
|
688 |
+
weight_init="",
|
689 |
+
freeze_num=-1,
|
690 |
+
with_lidar=False,
|
691 |
+
with_right_left_sensors=False,
|
692 |
+
with_center_sensor=False,
|
693 |
+
traffic_pred_head_type="det",
|
694 |
+
waypoints_pred_head="heatmap",
|
695 |
+
reverse_pos=True,
|
696 |
+
use_different_backbone=False,
|
697 |
+
use_view_embed=False,
|
698 |
+
use_mmad_pretrain=None,
|
699 |
+
):
|
700 |
+
super().__init__()
|
701 |
+
self.traffic_pred_head_type = traffic_pred_head_type
|
702 |
+
self.num_features = (
|
703 |
+
self.embed_dim
|
704 |
+
) = embed_dim # num_features for consistency with other models
|
705 |
+
norm_layer = norm_layer or partial(nn.LayerNorm, eps=1e-6)
|
706 |
+
act_layer = act_layer or nn.GELU
|
707 |
+
|
708 |
+
self.reverse_pos = reverse_pos
|
709 |
+
self.waypoints_pred_head = waypoints_pred_head
|
710 |
+
self.with_lidar = with_lidar
|
711 |
+
self.with_right_left_sensors = with_right_left_sensors
|
712 |
+
self.with_center_sensor = with_center_sensor
|
713 |
+
|
714 |
+
self.direct_concat = direct_concat
|
715 |
+
self.separate_view_attention = separate_view_attention
|
716 |
+
self.separate_all_attention = separate_all_attention
|
717 |
+
self.end2end = end2end
|
718 |
+
self.use_view_embed = use_view_embed
|
719 |
+
|
720 |
+
if self.direct_concat:
|
721 |
+
in_chans = in_chans * 4
|
722 |
+
self.with_center_sensor = False
|
723 |
+
self.with_right_left_sensors = False
|
724 |
+
|
725 |
+
if self.separate_view_attention:
|
726 |
+
self.attn_mask = build_attn_mask("seperate_view")
|
727 |
+
elif self.separate_all_attention:
|
728 |
+
self.attn_mask = build_attn_mask("seperate_all")
|
729 |
+
else:
|
730 |
+
self.attn_mask = None
|
731 |
+
|
732 |
+
if use_different_backbone:
|
733 |
+
if rgb_backbone_name == "r50":
|
734 |
+
self.rgb_backbone = resnet50d(
|
735 |
+
pretrained=True,
|
736 |
+
in_chans=in_chans,
|
737 |
+
features_only=True,
|
738 |
+
out_indices=[4],
|
739 |
+
)
|
740 |
+
elif rgb_backbone_name == "r26":
|
741 |
+
self.rgb_backbone = resnet26d(
|
742 |
+
pretrained=True,
|
743 |
+
in_chans=in_chans,
|
744 |
+
features_only=True,
|
745 |
+
out_indices=[4],
|
746 |
+
)
|
747 |
+
elif rgb_backbone_name == "r18":
|
748 |
+
self.rgb_backbone = resnet18d(
|
749 |
+
pretrained=True,
|
750 |
+
in_chans=in_chans,
|
751 |
+
features_only=True,
|
752 |
+
out_indices=[4],
|
753 |
+
)
|
754 |
+
if lidar_backbone_name == "r50":
|
755 |
+
self.lidar_backbone = resnet50d(
|
756 |
+
pretrained=False,
|
757 |
+
in_chans=in_chans,
|
758 |
+
features_only=True,
|
759 |
+
out_indices=[4],
|
760 |
+
)
|
761 |
+
elif lidar_backbone_name == "r26":
|
762 |
+
self.lidar_backbone = resnet26d(
|
763 |
+
pretrained=False,
|
764 |
+
in_chans=in_chans,
|
765 |
+
features_only=True,
|
766 |
+
out_indices=[4],
|
767 |
+
)
|
768 |
+
elif lidar_backbone_name == "r18":
|
769 |
+
self.lidar_backbone = resnet18d(
|
770 |
+
pretrained=False, in_chans=3, features_only=True, out_indices=[4]
|
771 |
+
)
|
772 |
+
rgb_embed_layer = partial(HybridEmbed, backbone=self.rgb_backbone)
|
773 |
+
lidar_embed_layer = partial(HybridEmbed, backbone=self.lidar_backbone)
|
774 |
+
|
775 |
+
if use_mmad_pretrain:
|
776 |
+
params = torch.load(use_mmad_pretrain)["state_dict"]
|
777 |
+
updated_params = OrderedDict()
|
778 |
+
for key in params:
|
779 |
+
if "backbone" in key:
|
780 |
+
updated_params[key.replace("backbone.", "")] = params[key]
|
781 |
+
self.rgb_backbone.load_state_dict(updated_params)
|
782 |
+
|
783 |
+
self.rgb_patch_embed = rgb_embed_layer(
|
784 |
+
img_size=img_size,
|
785 |
+
patch_size=patch_size,
|
786 |
+
in_chans=in_chans,
|
787 |
+
embed_dim=embed_dim,
|
788 |
+
)
|
789 |
+
self.lidar_patch_embed = lidar_embed_layer(
|
790 |
+
img_size=img_size,
|
791 |
+
patch_size=patch_size,
|
792 |
+
in_chans=3,
|
793 |
+
embed_dim=embed_dim,
|
794 |
+
)
|
795 |
+
else:
|
796 |
+
if rgb_backbone_name == "r50":
|
797 |
+
self.rgb_backbone = resnet50d(
|
798 |
+
pretrained=True, in_chans=3, features_only=True, out_indices=[4]
|
799 |
+
)
|
800 |
+
elif rgb_backbone_name == "r101":
|
801 |
+
self.rgb_backbone = resnet101d(
|
802 |
+
pretrained=True, in_chans=3, features_only=True, out_indices=[4]
|
803 |
+
)
|
804 |
+
elif rgb_backbone_name == "r26":
|
805 |
+
self.rgb_backbone = resnet26d(
|
806 |
+
pretrained=True, in_chans=3, features_only=True, out_indices=[4]
|
807 |
+
)
|
808 |
+
elif rgb_backbone_name == "r18":
|
809 |
+
self.rgb_backbone = resnet18d(
|
810 |
+
pretrained=True, in_chans=3, features_only=True, out_indices=[4]
|
811 |
+
)
|
812 |
+
embed_layer = partial(HybridEmbed, backbone=self.rgb_backbone)
|
813 |
+
|
814 |
+
self.rgb_patch_embed = embed_layer(
|
815 |
+
img_size=img_size,
|
816 |
+
patch_size=patch_size,
|
817 |
+
in_chans=in_chans,
|
818 |
+
embed_dim=embed_dim,
|
819 |
+
)
|
820 |
+
self.lidar_patch_embed = embed_layer(
|
821 |
+
img_size=img_size,
|
822 |
+
patch_size=patch_size,
|
823 |
+
in_chans=in_chans,
|
824 |
+
embed_dim=embed_dim,
|
825 |
+
)
|
826 |
+
|
827 |
+
self.global_embed = nn.Parameter(torch.zeros(1, embed_dim, 5))
|
828 |
+
self.view_embed = nn.Parameter(torch.zeros(1, embed_dim, 5, 1))
|
829 |
+
|
830 |
+
if self.end2end:
|
831 |
+
self.query_pos_embed = nn.Parameter(torch.zeros(1, embed_dim, 4))
|
832 |
+
self.query_embed = nn.Parameter(torch.zeros(4, 1, embed_dim))
|
833 |
+
elif self.waypoints_pred_head == "heatmap":
|
834 |
+
self.query_pos_embed = nn.Parameter(torch.zeros(1, embed_dim, 5))
|
835 |
+
self.query_embed = nn.Parameter(torch.zeros(400 + 5, 1, embed_dim))
|
836 |
+
else:
|
837 |
+
self.query_pos_embed = nn.Parameter(torch.zeros(1, embed_dim, 11))
|
838 |
+
self.query_embed = nn.Parameter(torch.zeros(400 + 11, 1, embed_dim))
|
839 |
+
|
840 |
+
if self.end2end:
|
841 |
+
self.waypoints_generator = GRUWaypointsPredictor(embed_dim, 4)
|
842 |
+
elif self.waypoints_pred_head == "heatmap":
|
843 |
+
self.waypoints_generator = MultiPath_Generator(
|
844 |
+
embed_dim + 32, embed_dim, 10
|
845 |
+
)
|
846 |
+
elif self.waypoints_pred_head == "gru":
|
847 |
+
self.waypoints_generator = GRUWaypointsPredictor(embed_dim)
|
848 |
+
elif self.waypoints_pred_head == "gru-command":
|
849 |
+
self.waypoints_generator = GRUWaypointsPredictorWithCommand(embed_dim)
|
850 |
+
elif self.waypoints_pred_head == "linear":
|
851 |
+
self.waypoints_generator = LinearWaypointsPredictor(embed_dim)
|
852 |
+
elif self.waypoints_pred_head == "linear-sum":
|
853 |
+
self.waypoints_generator = LinearWaypointsPredictor(embed_dim, cumsum=True)
|
854 |
+
|
855 |
+
self.junction_pred_head = nn.Linear(embed_dim, 2)
|
856 |
+
self.traffic_light_pred_head = nn.Linear(embed_dim, 2)
|
857 |
+
self.stop_sign_head = nn.Linear(embed_dim, 2)
|
858 |
+
|
859 |
+
if self.traffic_pred_head_type == "det":
|
860 |
+
self.traffic_pred_head = nn.Sequential(
|
861 |
+
*[
|
862 |
+
nn.Linear(embed_dim + 32, 64),
|
863 |
+
nn.ReLU(),
|
864 |
+
nn.Linear(64, 7),
|
865 |
+
# nn.Sigmoid(),
|
866 |
+
]
|
867 |
+
)
|
868 |
+
elif self.traffic_pred_head_type == "seg":
|
869 |
+
self.traffic_pred_head = nn.Sequential(
|
870 |
+
*[nn.Linear(embed_dim, 64), nn.ReLU(), nn.Linear(64, 1), nn.Sigmoid()]
|
871 |
+
)
|
872 |
+
|
873 |
+
self.position_encoding = PositionEmbeddingSine(embed_dim // 2, normalize=True)
|
874 |
+
|
875 |
+
encoder_layer = TransformerEncoderLayer(
|
876 |
+
embed_dim, num_heads, dim_feedforward, dropout, act_layer, normalize_before
|
877 |
+
)
|
878 |
+
self.encoder = TransformerEncoder(encoder_layer, enc_depth, None)
|
879 |
+
|
880 |
+
decoder_layer = TransformerDecoderLayer(
|
881 |
+
embed_dim, num_heads, dim_feedforward, dropout, act_layer, normalize_before
|
882 |
+
)
|
883 |
+
decoder_norm = nn.LayerNorm(embed_dim)
|
884 |
+
self.decoder = TransformerDecoder(
|
885 |
+
decoder_layer, dec_depth, decoder_norm, return_intermediate=False
|
886 |
+
)
|
887 |
+
self.reset_parameters()
|
888 |
+
|
889 |
+
def reset_parameters(self):
|
890 |
+
nn.init.uniform_(self.global_embed)
|
891 |
+
nn.init.uniform_(self.view_embed)
|
892 |
+
nn.init.uniform_(self.query_embed)
|
893 |
+
nn.init.uniform_(self.query_pos_embed)
|
894 |
+
|
895 |
+
def forward_features(
|
896 |
+
self,
|
897 |
+
front_image,
|
898 |
+
left_image,
|
899 |
+
right_image,
|
900 |
+
front_center_image,
|
901 |
+
lidar,
|
902 |
+
measurements,
|
903 |
+
):
|
904 |
+
features = []
|
905 |
+
|
906 |
+
# Front view processing
|
907 |
+
front_image_token, front_image_token_global = self.rgb_patch_embed(front_image)
|
908 |
+
if self.use_view_embed:
|
909 |
+
front_image_token = (
|
910 |
+
front_image_token
|
911 |
+
+ self.view_embed[:, :, 0:1, :]
|
912 |
+
+ self.position_encoding(front_image_token)
|
913 |
+
)
|
914 |
+
else:
|
915 |
+
front_image_token = front_image_token + self.position_encoding(
|
916 |
+
front_image_token
|
917 |
+
)
|
918 |
+
front_image_token = front_image_token.flatten(2).permute(2, 0, 1)
|
919 |
+
front_image_token_global = (
|
920 |
+
front_image_token_global
|
921 |
+
+ self.view_embed[:, :, 0, :]
|
922 |
+
+ self.global_embed[:, :, 0:1]
|
923 |
+
)
|
924 |
+
front_image_token_global = front_image_token_global.permute(2, 0, 1)
|
925 |
+
features.extend([front_image_token, front_image_token_global])
|
926 |
+
|
927 |
+
if self.with_right_left_sensors:
|
928 |
+
# Left view processing
|
929 |
+
left_image_token, left_image_token_global = self.rgb_patch_embed(left_image)
|
930 |
+
if self.use_view_embed:
|
931 |
+
left_image_token = (
|
932 |
+
left_image_token
|
933 |
+
+ self.view_embed[:, :, 1:2, :]
|
934 |
+
+ self.position_encoding(left_image_token)
|
935 |
+
)
|
936 |
+
else:
|
937 |
+
left_image_token = left_image_token + self.position_encoding(
|
938 |
+
left_image_token
|
939 |
+
)
|
940 |
+
left_image_token = left_image_token.flatten(2).permute(2, 0, 1)
|
941 |
+
left_image_token_global = (
|
942 |
+
left_image_token_global
|
943 |
+
+ self.view_embed[:, :, 1, :]
|
944 |
+
+ self.global_embed[:, :, 1:2]
|
945 |
+
)
|
946 |
+
left_image_token_global = left_image_token_global.permute(2, 0, 1)
|
947 |
+
|
948 |
+
# Right view processing
|
949 |
+
right_image_token, right_image_token_global = self.rgb_patch_embed(
|
950 |
+
right_image
|
951 |
+
)
|
952 |
+
if self.use_view_embed:
|
953 |
+
right_image_token = (
|
954 |
+
right_image_token
|
955 |
+
+ self.view_embed[:, :, 2:3, :]
|
956 |
+
+ self.position_encoding(right_image_token)
|
957 |
+
)
|
958 |
+
else:
|
959 |
+
right_image_token = right_image_token + self.position_encoding(
|
960 |
+
right_image_token
|
961 |
+
)
|
962 |
+
right_image_token = right_image_token.flatten(2).permute(2, 0, 1)
|
963 |
+
right_image_token_global = (
|
964 |
+
right_image_token_global
|
965 |
+
+ self.view_embed[:, :, 2, :]
|
966 |
+
+ self.global_embed[:, :, 2:3]
|
967 |
+
)
|
968 |
+
right_image_token_global = right_image_token_global.permute(2, 0, 1)
|
969 |
+
|
970 |
+
features.extend(
|
971 |
+
[
|
972 |
+
left_image_token,
|
973 |
+
left_image_token_global,
|
974 |
+
right_image_token,
|
975 |
+
right_image_token_global,
|
976 |
+
]
|
977 |
+
)
|
978 |
+
|
979 |
+
if self.with_center_sensor:
|
980 |
+
# Front center view processing
|
981 |
+
(
|
982 |
+
front_center_image_token,
|
983 |
+
front_center_image_token_global,
|
984 |
+
) = self.rgb_patch_embed(front_center_image)
|
985 |
+
if self.use_view_embed:
|
986 |
+
front_center_image_token = (
|
987 |
+
front_center_image_token
|
988 |
+
+ self.view_embed[:, :, 3:4, :]
|
989 |
+
+ self.position_encoding(front_center_image_token)
|
990 |
+
)
|
991 |
+
else:
|
992 |
+
front_center_image_token = (
|
993 |
+
front_center_image_token
|
994 |
+
+ self.position_encoding(front_center_image_token)
|
995 |
+
)
|
996 |
+
|
997 |
+
front_center_image_token = front_center_image_token.flatten(2).permute(
|
998 |
+
2, 0, 1
|
999 |
+
)
|
1000 |
+
front_center_image_token_global = (
|
1001 |
+
front_center_image_token_global
|
1002 |
+
+ self.view_embed[:, :, 3, :]
|
1003 |
+
+ self.global_embed[:, :, 3:4]
|
1004 |
+
)
|
1005 |
+
front_center_image_token_global = front_center_image_token_global.permute(
|
1006 |
+
2, 0, 1
|
1007 |
+
)
|
1008 |
+
features.extend([front_center_image_token, front_center_image_token_global])
|
1009 |
+
|
1010 |
+
if self.with_lidar:
|
1011 |
+
lidar_token, lidar_token_global = self.lidar_patch_embed(lidar)
|
1012 |
+
if self.use_view_embed:
|
1013 |
+
lidar_token = (
|
1014 |
+
lidar_token
|
1015 |
+
+ self.view_embed[:, :, 4:5, :]
|
1016 |
+
+ self.position_encoding(lidar_token)
|
1017 |
+
)
|
1018 |
+
else:
|
1019 |
+
lidar_token = lidar_token + self.position_encoding(lidar_token)
|
1020 |
+
lidar_token = lidar_token.flatten(2).permute(2, 0, 1)
|
1021 |
+
lidar_token_global = (
|
1022 |
+
lidar_token_global
|
1023 |
+
+ self.view_embed[:, :, 4, :]
|
1024 |
+
+ self.global_embed[:, :, 4:5]
|
1025 |
+
)
|
1026 |
+
lidar_token_global = lidar_token_global.permute(2, 0, 1)
|
1027 |
+
features.extend([lidar_token, lidar_token_global])
|
1028 |
+
|
1029 |
+
features = torch.cat(features, 0)
|
1030 |
+
return features
|
1031 |
+
|
1032 |
+
def forward(self, x):
|
1033 |
+
front_image = x["rgb"]
|
1034 |
+
left_image = x["rgb_left"]
|
1035 |
+
right_image = x["rgb_right"]
|
1036 |
+
front_center_image = x["rgb_center"]
|
1037 |
+
measurements = x["measurements"]
|
1038 |
+
target_point = x["target_point"]
|
1039 |
+
lidar = x["lidar"]
|
1040 |
+
|
1041 |
+
if self.direct_concat:
|
1042 |
+
img_size = front_image.shape[-1]
|
1043 |
+
left_image = torch.nn.functional.interpolate(
|
1044 |
+
left_image, size=(img_size, img_size)
|
1045 |
+
)
|
1046 |
+
right_image = torch.nn.functional.interpolate(
|
1047 |
+
right_image, size=(img_size, img_size)
|
1048 |
+
)
|
1049 |
+
front_center_image = torch.nn.functional.interpolate(
|
1050 |
+
front_center_image, size=(img_size, img_size)
|
1051 |
+
)
|
1052 |
+
front_image = torch.cat(
|
1053 |
+
[front_image, left_image, right_image, front_center_image], dim=1
|
1054 |
+
)
|
1055 |
+
features = self.forward_features(
|
1056 |
+
front_image,
|
1057 |
+
left_image,
|
1058 |
+
right_image,
|
1059 |
+
front_center_image,
|
1060 |
+
lidar,
|
1061 |
+
measurements,
|
1062 |
+
)
|
1063 |
+
|
1064 |
+
bs = front_image.shape[0]
|
1065 |
+
|
1066 |
+
if self.end2end:
|
1067 |
+
tgt = self.query_pos_embed.repeat(bs, 1, 1)
|
1068 |
+
else:
|
1069 |
+
tgt = self.position_encoding(
|
1070 |
+
torch.ones((bs, 1, 20, 20), device=x["rgb"].device)
|
1071 |
+
)
|
1072 |
+
tgt = tgt.flatten(2)
|
1073 |
+
tgt = torch.cat([tgt, self.query_pos_embed.repeat(bs, 1, 1)], 2)
|
1074 |
+
tgt = tgt.permute(2, 0, 1)
|
1075 |
+
|
1076 |
+
memory = self.encoder(features, mask=self.attn_mask)
|
1077 |
+
hs = self.decoder(self.query_embed.repeat(1, bs, 1), memory, query_pos=tgt)[0]
|
1078 |
+
|
1079 |
+
hs = hs.permute(1, 0, 2) # Batchsize , N, C
|
1080 |
+
if self.end2end:
|
1081 |
+
waypoints = self.waypoints_generator(hs, target_point)
|
1082 |
+
return waypoints
|
1083 |
+
|
1084 |
+
if self.waypoints_pred_head != "heatmap":
|
1085 |
+
traffic_feature = hs[:, :400]
|
1086 |
+
is_junction_feature = hs[:, 400]
|
1087 |
+
traffic_light_state_feature = hs[:, 400]
|
1088 |
+
stop_sign_feature = hs[:, 400]
|
1089 |
+
waypoints_feature = hs[:, 401:411]
|
1090 |
+
else:
|
1091 |
+
traffic_feature = hs[:, :400]
|
1092 |
+
is_junction_feature = hs[:, 400]
|
1093 |
+
traffic_light_state_feature = hs[:, 400]
|
1094 |
+
stop_sign_feature = hs[:, 400]
|
1095 |
+
waypoints_feature = hs[:, 401:405]
|
1096 |
+
|
1097 |
+
if self.waypoints_pred_head == "heatmap":
|
1098 |
+
waypoints = self.waypoints_generator(waypoints_feature, measurements)
|
1099 |
+
elif self.waypoints_pred_head == "gru":
|
1100 |
+
waypoints = self.waypoints_generator(waypoints_feature, target_point)
|
1101 |
+
elif self.waypoints_pred_head == "gru-command":
|
1102 |
+
waypoints = self.waypoints_generator(waypoints_feature, target_point, measurements)
|
1103 |
+
elif self.waypoints_pred_head == "linear":
|
1104 |
+
waypoints = self.waypoints_generator(waypoints_feature, measurements)
|
1105 |
+
elif self.waypoints_pred_head == "linear-sum":
|
1106 |
+
waypoints = self.waypoints_generator(waypoints_feature, measurements)
|
1107 |
+
|
1108 |
+
is_junction = self.junction_pred_head(is_junction_feature)
|
1109 |
+
traffic_light_state = self.traffic_light_pred_head(traffic_light_state_feature)
|
1110 |
+
stop_sign = self.stop_sign_head(stop_sign_feature)
|
1111 |
+
|
1112 |
+
velocity = measurements[:, 6:7].unsqueeze(-1)
|
1113 |
+
velocity = velocity.repeat(1, 400, 32)
|
1114 |
+
traffic_feature_with_vel = torch.cat([traffic_feature, velocity], dim=2)
|
1115 |
+
traffic = self.traffic_pred_head(traffic_feature_with_vel)
|
1116 |
+
return traffic, waypoints, is_junction, traffic_light_state, stop_sign, traffic_feature
|
1117 |
+
def load_pretrained(self, model_path, strict=False):
|
1118 |
+
"""
|
1119 |
+
تحميل الأوزان المدربة مسبقاً - نسخة محسنة
|
1120 |
+
|
1121 |
+
Args:
|
1122 |
+
model_path (str): مسار ملف الأوزان
|
1123 |
+
strict (bool): إذا كان True، يتطلب تطابق تام للمفاتيح
|
1124 |
+
"""
|
1125 |
+
if not model_path or not Path(model_path).exists():
|
1126 |
+
logging.warning(f"ملف الأوزان غير موجود: {model_path}")
|
1127 |
+
logging.info("سيتم استخدام أوزان عشوائية")
|
1128 |
+
return False
|
1129 |
+
|
1130 |
+
try:
|
1131 |
+
logging.info(f"محاولة تحميل الأوزان من: {model_path}")
|
1132 |
+
|
1133 |
+
# تحميل الملف مع معالجة أنواع مختلفة من ملفات الحفظ
|
1134 |
+
checkpoint = torch.load(model_path, map_location='cpu', weights_only=False)
|
1135 |
+
|
1136 |
+
# استخراج state_dict من أنواع مختلفة من ملفات الحفظ
|
1137 |
+
if isinstance(checkpoint, dict):
|
1138 |
+
if 'model_state_dict' in checkpoint:
|
1139 |
+
state_dict = checkpoint['model_state_dict']
|
1140 |
+
logging.info("تم العثور على 'model_state_dict' في الملف")
|
1141 |
+
elif 'state_dict' in checkpoint:
|
1142 |
+
state_dict = checkpoint['state_dict']
|
1143 |
+
logging.info("تم العثور على 'state_dict' في الملف")
|
1144 |
+
elif 'model' in checkpoint:
|
1145 |
+
state_dict = checkpoint['model']
|
1146 |
+
logging.info("تم العثور على 'model' في الملف")
|
1147 |
+
else:
|
1148 |
+
state_dict = checkpoint
|
1149 |
+
logging.info("استخدام الملف كـ state_dict مباشرة")
|
1150 |
+
else:
|
1151 |
+
state_dict = checkpoint
|
1152 |
+
logging.info("استخدام الملف كـ state_dict مباشرة")
|
1153 |
+
|
1154 |
+
# تنظيف أسماء المفاتيح (إزالة 'module.' إذا كانت موجودة)
|
1155 |
+
clean_state_dict = OrderedDict()
|
1156 |
+
for k, v in state_dict.items():
|
1157 |
+
# إزالة 'module.' من بداية اسم المفتاح إذا كان موجوداً
|
1158 |
+
clean_key = k[7:] if k.startswith('module.') else k
|
1159 |
+
clean_state_dict[clean_key] = v
|
1160 |
+
|
1161 |
+
# تحميل الأوزان
|
1162 |
+
missing_keys, unexpected_keys = self.load_state_dict(clean_state_dict, strict=strict)
|
1163 |
+
|
1164 |
+
# تقرير حالة التحميل
|
1165 |
+
if missing_keys:
|
1166 |
+
logging.warning(f"مفاتيح مفقودة ({len(missing_keys)}): {missing_keys[:5]}..." if len(missing_keys) > 5 else f"مفاتيح مفقودة: {missing_keys}")
|
1167 |
+
|
1168 |
+
if unexpected_keys:
|
1169 |
+
logging.warning(f"مفاتيح غير متوقعة ({len(unexpected_keys)}): {unexpected_keys[:5]}..." if len(unexpected_keys) > 5 else f"مفاتيح غير متوقعة: {unexpected_keys}")
|
1170 |
+
|
1171 |
+
if not missing_keys and not unexpected_keys:
|
1172 |
+
logging.info("✅ تم تحميل جميع الأوزان بنجاح تام")
|
1173 |
+
elif not strict:
|
1174 |
+
logging.info("✅ تم تحميل الأوزان بنجاح (مع تجاهل عدم التطابق)")
|
1175 |
+
|
1176 |
+
return True
|
1177 |
+
|
1178 |
+
except Exception as e:
|
1179 |
+
logging.error(f"❌ خطأ في تحميل الأوزان: {str(e)}")
|
1180 |
+
logging.info("سيتم استخدام أوزان عشوائية")
|
1181 |
+
return False
|
1182 |
+
|
1183 |
+
|
1184 |
+
# ============================================================================
|
1185 |
+
# دوال مساعدة لتحميل النموذج
|
1186 |
+
# ============================================================================
|
1187 |
+
# ==============================================================================
|
1188 |
+
# ملف: config_and_loader.py
|
1189 |
+
# هذا هو المصدر الوحيد للحقيقة لجميع الإعدادات وعملية تحميل النموذج.
|
1190 |
+
# ==============================================================================
|
1191 |
+
|
1192 |
+
|
1193 |
+
|
1194 |
+
# def get_master_config(model_path="model/best_model.pth"):
|
1195 |
+
# """
|
1196 |
+
# [النسخة الكاملة والنهائية]
|
1197 |
+
# ينشئ ويدمج كل الإعدادات المطلوبة للتطبيق (النموذج، المتتبع، المتحكم).
|
1198 |
+
# """
|
1199 |
+
# model_params = {
|
1200 |
+
# "img_size": 224, "embed_dim": 256, "enc_depth": 6, "dec_depth": 6,
|
1201 |
+
# "rgb_backbone_name": 'r50', "lidar_backbone_name": 'r18',
|
1202 |
+
# "waypoints_pred_head": 'gru', "use_different_backbone": True,
|
1203 |
+
# "with_lidar": False, "with_right_left_sensors": False,
|
1204 |
+
# "with_center_sensor": False, "multi_view_img_size": 112,
|
1205 |
+
# "patch_size": 8, "in_chans": 3, "dim_feedforward": 2048,
|
1206 |
+
# "normalize_before": False, "num_heads": 8, "dropout": 0.1,
|
1207 |
+
# "end2end": False, "direct_concat": False, "separate_view_attention": False,
|
1208 |
+
# "separate_all_attention": False, "freeze_num": -1,
|
1209 |
+
# "traffic_pred_head_type": "det", "reverse_pos": True,
|
1210 |
+
# "use_view_embed": False, "use_mmad_pretrain": None,
|
1211 |
+
# }
|
1212 |
+
|
1213 |
+
# grid_conf = {
|
1214 |
+
# 'h': 20, 'w': 20, 'x_res': 1.0, 'y_res': 1.0,
|
1215 |
+
# 'y_min': 0.0, 'y_max': 20.0, 'x_min': -10.0, 'x_max': 10.0,
|
1216 |
+
# }
|
1217 |
+
|
1218 |
+
# controller_params = {
|
1219 |
+
# 'turn_KP': 0.75, 'turn_KI': 0.05, 'turn_KD': 0.25, 'turn_n': 20,
|
1220 |
+
# 'speed_KP': 0.55, 'speed_KI': 0.05, 'speed_KD': 0.15, 'speed_n': 20,
|
1221 |
+
# 'max_speed': 8.0, 'max_throttle': 0.75, 'min_speed': 0.1,
|
1222 |
+
# 'brake_sensitivity': 0.3, 'light_threshold': 0.5, 'stop_threshold': 0.6,
|
1223 |
+
# 'stop_sign_duration': 20, 'max_stop_time': 250,
|
1224 |
+
# 'forced_move_duration': 20, 'forced_throttle': 0.5,
|
1225 |
+
# 'max_red_light_time': 150, 'red_light_block_duration': 80,
|
1226 |
+
# 'accel_rate': 0.1, 'decel_rate': 0.2, 'critical_distance': 4.0,
|
1227 |
+
# 'follow_distance': 10.0, 'speed_match_factor': 0.9,
|
1228 |
+
# 'tracker_match_thresh': 2.5, 'tracker_prune_age': 5,
|
1229 |
+
# 'follow_grace_period': 20
|
1230 |
+
# }
|
1231 |
+
|
1232 |
+
# master_config = {
|
1233 |
+
# 'model_params': model_params,
|
1234 |
+
# 'grid_conf': grid_conf,
|
1235 |
+
# 'controller_params': controller_params,
|
1236 |
+
# 'paths': {'pretrained_weights': model_path},
|
1237 |
+
# 'simulation': {'frequency': 10.0}
|
1238 |
+
# }
|
1239 |
+
|
1240 |
+
# return master_config
|
1241 |
+
|
1242 |
+
|
1243 |
+
# def load_and_prepare_model(device: torch.device) -> InterfuserModel:
|
1244 |
+
# """
|
1245 |
+
# [النسخة النهائية الصحيحة - تستقبل مدخلاً واحدًا فقط]
|
1246 |
+
# تستخدم دالة الإعدادات الرئيسية لإنشاء وتحميل النموذج.
|
1247 |
+
# """
|
1248 |
+
# try:
|
1249 |
+
# logging.info("Attempting to load model using master config...")
|
1250 |
+
# # 1. الحصول على كل الإعدادات من المصدر الوحيد للحقيقة
|
1251 |
+
# config = get_master_config()
|
1252 |
+
|
1253 |
+
# # 2. إنشاء النموذج باستخدام إعدادات النموذج فقط
|
1254 |
+
# model = InterfuserModel(**config['model_params']).to(device)
|
1255 |
+
# logging.info(f"Model instantiated on device: {device}")
|
1256 |
+
|
1257 |
+
# # 3. تحميل الأوزان باستخدام الدالة الداخلية للنموذج
|
1258 |
+
# checkpoint_path = config['paths']['pretrained_weights']
|
1259 |
+
# model.load_pretrained(checkpoint_path, strict=False)
|
1260 |
+
|
1261 |
+
# # 4. وضع النموذج في وضع التقييم
|
1262 |
+
# model.eval()
|
1263 |
+
# logging.info("✅ Model prepared and set to evaluation mode.")
|
1264 |
+
|
1265 |
+
# return model
|
1266 |
+
|
1267 |
+
# except Exception as e:
|
1268 |
+
# logging.error(f"❌ CRITICAL ERROR in load_and_prepare_model: {e}", exc_info=True)
|
1269 |
+
# raise
|
1270 |
+
|
1271 |
+
|
1272 |
+
|
1273 |
+
# ==============================================================================
|
1274 |
+
# الدالة الأولى: get_master_config
|
1275 |
+
# ==============================================================================
|
1276 |
+
|
1277 |
+
def get_master_config():
|
1278 |
+
"""
|
1279 |
+
[النسخة الاحترافية]
|
1280 |
+
يعيد قاموسًا شاملاً يحتوي على جميع إعدادات التطبيق الثابتة.
|
1281 |
+
هذه الدالة هي المصدر الوحيد للحقيقة للإعدادات.
|
1282 |
+
"""
|
1283 |
+
|
1284 |
+
# --- القسم 1: معلومات مستودع النموذج على Hugging Face Hub ---
|
1285 |
+
huggingface_repo = {
|
1286 |
+
'repo_id': "BaseerAI/Interfuser-Baseer-v1", # استبدله باسم مستودع النموذج الخاص بك
|
1287 |
+
'filename': "best_model.pth" # اسم ملف الأوزان داخل المستودع
|
1288 |
+
}
|
1289 |
+
|
1290 |
+
# --- القسم 2: إعدادات بنية نموذج Interfuser ---
|
1291 |
+
model_params = {
|
1292 |
+
"img_size": 224, "embed_dim": 256, "enc_depth": 6, "dec_depth": 6,
|
1293 |
+
"rgb_backbone_name": 'r50', "lidar_backbone_name": 'r18',
|
1294 |
+
"waypoints_pred_head": 'gru', "use_different_backbone": True,
|
1295 |
+
"with_lidar": False, "with_right_left_sensors": False,
|
1296 |
+
"with_center_sensor": False, "multi_view_img_size": 112,
|
1297 |
+
"patch_size": 8, "in_chans": 3, "dim_feedforward": 2048,
|
1298 |
+
"normalize_before": False, "num_heads": 8, "dropout": 0.1,
|
1299 |
+
"end2end": False, "direct_concat": False, "separate_view_attention": False,
|
1300 |
+
"separate_all_attention": False, "freeze_num": -1,
|
1301 |
+
"traffic_pred_head_type": "det", "reverse_pos": True,
|
1302 |
+
"use_view_embed": False, "use_mmad_pretrain": None,
|
1303 |
+
}
|
1304 |
+
|
1305 |
+
# --- القسم 3: إعدادات الشبكة ومنظور عين الطائر (BEV) ---
|
1306 |
+
grid_conf = {
|
1307 |
+
'h': 20, 'w': 20, 'x_res': 1.0, 'y_res': 1.0,
|
1308 |
+
'y_min': 0.0, 'y_max': 20.0, 'x_min': -10.0, 'x_max': 10.0,
|
1309 |
+
}
|
1310 |
+
|
1311 |
+
# --- القسم 4: إعدادات وحدة التحكم (Controller) والمتتبع (Tracker) ---
|
1312 |
+
controller_params = {
|
1313 |
+
'turn_KP': 0.75, 'turn_KI': 0.05, 'turn_KD': 0.25, 'turn_n': 20,
|
1314 |
+
'speed_KP': 0.55, 'speed_KI': 0.05, 'speed_KD': 0.15, 'speed_n': 20,
|
1315 |
+
'max_speed': 8.0, 'max_throttle': 0.75, 'min_speed': 0.1,
|
1316 |
+
'brake_sensitivity': 0.3, 'light_threshold': 0.5, 'stop_threshold': 0.6,
|
1317 |
+
'stop_sign_duration': 20, 'max_stop_time': 250,
|
1318 |
+
'forced_move_duration': 20, 'forced_throttle': 0.5,
|
1319 |
+
'max_red_light_time': 150, 'red_light_block_duration': 80,
|
1320 |
+
'accel_rate': 0.1, 'decel_rate': 0.2, 'critical_distance': 4.0,
|
1321 |
+
'follow_distance': 10.0, 'speed_match_factor': 0.9,
|
1322 |
+
'tracker_match_thresh': 2.5, 'tracker_prune_age': 5,
|
1323 |
+
'follow_grace_period': 20
|
1324 |
+
}
|
1325 |
+
|
1326 |
+
# --- القسم 5: تجميع كل شيء في قاموس رئيسي واحد ---
|
1327 |
+
master_config = {
|
1328 |
+
'huggingface_repo': huggingface_repo,
|
1329 |
+
'model_params': model_params,
|
1330 |
+
'grid_conf': grid_conf,
|
1331 |
+
'controller_params': controller_params,
|
1332 |
+
'simulation': {
|
1333 |
+
'frequency': 10.0
|
1334 |
+
}
|
1335 |
+
}
|
1336 |
+
|
1337 |
+
return master_config
|
1338 |
+
|
1339 |
+
|
1340 |
+
# ==============================================================================
|
1341 |
+
# الدالة الثانية: load_and_prepare_model
|
1342 |
+
# ==============================================================================
|
1343 |
+
|
1344 |
+
def load_and_prepare_model(device: torch.device) -> InterfuserModel:
|
1345 |
+
"""
|
1346 |
+
[النسخة الاحترافية]
|
1347 |
+
تستخدم الإعدادات الرئيسية من `get_master_config` لإنشاء وتحميل النموذج.
|
1348 |
+
تقوم بتحويل معرّف النموذج من Hugging Face Hub إلى مسار ملف حقيقي.
|
1349 |
+
|
1350 |
+
Args:
|
1351 |
+
device (torch.device): الجهاز المستهدف (CPU/GPU)
|
1352 |
+
|
1353 |
+
Returns:
|
1354 |
+
Interfuser: النموذج المحمل وجاهز للاستدلال.
|
1355 |
+
"""
|
1356 |
+
try:
|
1357 |
+
logging.info("Initializing model loading process...")
|
1358 |
+
|
1359 |
+
# 1. الحصول على جميع الإعدادات من المصدر الوحيد للحقيقة
|
1360 |
+
config = get_master_config()
|
1361 |
+
|
1362 |
+
# 2. تحميل ملف الأوزان من Hugging Face Hub
|
1363 |
+
repo_info = config['huggingface_repo']
|
1364 |
+
logging.info(f"Downloading model weights from repo: '{repo_info['repo_id']}'")
|
1365 |
+
|
1366 |
+
# استخدام token إذا كان المستودع خاصًا
|
1367 |
+
# token = HfFolder.get_token() # أو يمكن تمريره مباشرة
|
1368 |
+
actual_model_path = hf_hub_download(
|
1369 |
+
repo_id=repo_info['repo_id'],
|
1370 |
+
filename=repo_info['filename'],
|
1371 |
+
# token=token, # قم بإلغاء التعليق إذا كان المستودع خاصًا
|
1372 |
+
)
|
1373 |
+
logging.info(f"Model weights are available at local path: {actual_model_path}")
|
1374 |
+
|
1375 |
+
# 3. إنشاء نسخة من النموذج باستخدام الإعدادات الصحيحة
|
1376 |
+
logging.info("Instantiating model with specified parameters...")
|
1377 |
+
model = InterfuserModel(**config['model_params']).to(device)
|
1378 |
+
|
1379 |
+
# 4. تحميل الأوزان التي تم تنزيلها إلى النموذج
|
1380 |
+
# نستخدم الدالة المساعدة الموجودة داخل كلاس النموذج نفسه
|
1381 |
+
success = model.load_pretrained(actual_model_path, strict=False)
|
1382 |
+
if not success:
|
1383 |
+
logging.warning("⚠️ Model weights were not loaded successfully. The model will use random weights.")
|
1384 |
+
|
1385 |
+
# 5. وضع النموذج في وضع التقييم (خطوة حاسمة)
|
1386 |
+
model.eval()
|
1387 |
+
logging.info("✅ Model prepared and set to evaluation mode. Ready for inference.")
|
1388 |
+
|
1389 |
+
return model
|
1390 |
+
|
1391 |
+
except Exception as e:
|
1392 |
+
# تسجيل الخطأ بالتفصيل ثم إطلاقه مرة أخرى ليتم التعامل معه في مستوى أعلى
|
1393 |
+
logging.error(f"❌ CRITICAL ERROR during model initialization: {e}", exc_info=True)
|
1394 |
+
raise
|