Skip to content

Add object detection, tracking, mobile notifications, and search to any security camera.

kinda works

roryclear88232199f23e
diff --git a/experimental/rtsp_yolov8.py b/experimental/rtsp_yolov8.py
index 648173a..edc53c9 100644
--- a/experimental/rtsp_yolov8.py
+++ b/experimental/rtsp_yolov8.py
@@ -1,26 +1,372 @@
+from tinygrad.nn import Conv2d, BatchNorm2d
+from tinygrad.tensor import Tensor
+from tinygrad import TinyJit
+from tinygrad.device import is_dtype_supported
+from tinygrad import dtypes
+import numpy as np
+from itertools import chain
+from pathlib import Path
+import cv2
+from collections import defaultdict
+import time, sys
+from tinygrad.helpers import fetch
+from tinygrad.nn.state import safe_load, load_state_dict
+import json
+
+#Model architecture from https://github.com/ultralytics/ultralytics/issues/189
+#The upsampling class has been taken from this pull request https://github.com/tinygrad/tinygrad/pull/784 by dc-dc-dc. Now 2(?) models use upsampling. (retinet and this)
+
+#Pre processing image functions.
+def compute_transform(image, new_shape=(640, 640), auto=False, scaleFill=False, scaleup=True, stride=32) -> Tensor:
+ shape = image.shape[:2] # current shape [height, width]
+ new_shape = (new_shape, new_shape) if isinstance(new_shape, int) else new_shape
+ r = min(new_shape[0] / shape[0], new_shape[1] / shape[1])
+ r = min(r, 1.0) if not scaleup else r
+ new_unpad = (int(round(shape[1] * r)), int(round(shape[0] * r)))
+ dw, dh = new_shape[1] - new_unpad[0], new_shape[0] - new_unpad[1]
+ dw, dh = (np.mod(dw, stride), np.mod(dh, stride)) if auto else (0.0, 0.0)
+ new_unpad = (new_shape[1], new_shape[0]) if scaleFill else new_unpad
+ dw /= 2
+ dh /= 2
+ image = cv2.resize(image, new_unpad, interpolation=cv2.INTER_LINEAR) if shape[::-1] != new_unpad else image
+ top, bottom = int(round(dh - 0.1)), int(round(dh + 0.1))
+ left, right = int(round(dw - 0.1)), int(round(dw + 0.1))
+ image = cv2.copyMakeBorder(image, top, bottom, left, right, cv2.BORDER_CONSTANT, value=(114, 114, 114))
+ return Tensor(image)
+
+def preprocess(im, imgsz=640, model_stride=32, model_pt=True): return compute_transform(im, new_shape=imgsz, auto=True, stride=model_stride)
+
+# utility functions for forward pass.
+def dist2bbox(distance, anchor_points, xywh=True, dim=-1):
+ lt, rb = distance.chunk(2, dim)
+ x1y1 = anchor_points - lt
+ x2y2 = anchor_points + rb
+ if xywh:
+ c_xy = (x1y1 + x2y2) / 2
+ wh = x2y2 - x1y1
+ return c_xy.cat(wh, dim=1)
+ return x1y1.cat(x2y2, dim=1)
+
+def make_anchors(feats, strides, grid_cell_offset=0.5):
+ anchor_points, stride_tensor = [], []
+ assert feats is not None
+ for i, stride in enumerate(strides):
+ _, _, h, w = feats[i].shape
+ sx = Tensor.arange(w) + grid_cell_offset
+ sy = Tensor.arange(h) + grid_cell_offset
+
+ # this is np.meshgrid but in tinygrad
+ sx = sx.reshape(1, -1).repeat([h, 1]).reshape(-1)
+ sy = sy.reshape(-1, 1).repeat([1, w]).reshape(-1)
+
+ anchor_points.append(Tensor.stack(sx, sy, dim=-1).reshape(-1, 2))
+ stride_tensor.append(Tensor.full((h * w), stride))
+ anchor_points = anchor_points[0].cat(anchor_points[1], anchor_points[2])
+ stride_tensor = stride_tensor[0].cat(stride_tensor[1], stride_tensor[2]).unsqueeze(1)
+ return anchor_points, stride_tensor
+
+# this function is from the original implementation
+def autopad(k, p=None, d=1): # kernel, padding, dilation
+ if d > 1:
+ k = d * (k - 1) + 1 if isinstance(k, int) else [d * (x - 1) + 1 for x in k] # actual kernel-size
+ if p is None:
+ p = k // 2 if isinstance(k, int) else [x // 2 for x in k] # auto-pad
+ return p
+
+def clip_boxes(boxes, shape):
+ boxes[..., [0, 2]] = np.clip(boxes[..., [0, 2]], 0, shape[1]) # x1, x2
+ boxes[..., [1, 3]] = np.clip(boxes[..., [1, 3]], 0, shape[0]) # y1, y2
+ return boxes
+
+def scale_boxes(img1_shape, predictions, img0_shape, ratio_pad=None):
+ gain = ratio_pad if ratio_pad else min(img1_shape[0] / img0_shape[0], img1_shape[1] / img0_shape[1])
+ pad = ((img1_shape[1] - img0_shape[1] * gain) / 2, (img1_shape[0] - img0_shape[0] * gain) / 2)
+ for pred in predictions:
+ boxes_np = pred[:4].numpy() if isinstance(pred[:4], Tensor) else pred[:4]
+ boxes_np[..., [0, 2]] -= pad[0]
+ boxes_np[..., [1, 3]] -= pad[1]
+ boxes_np[..., :4] /= gain
+ boxes_np = clip_boxes(boxes_np, img0_shape)
+ pred[:4] = boxes_np
+ return predictions
+
+def get_variant_multiples(variant):
+ return {'n':(0.33, 0.25, 2.0), 's':(0.33, 0.50, 2.0), 'm':(0.67, 0.75, 1.5), 'l':(1.0, 1.0, 1.0), 'x':(1, 1.25, 1.0) }.get(variant, None)
+
+def label_predictions(all_predictions):
+ class_index_count = defaultdict(int)
+ for pred in all_predictions:
+ class_id = int(pred[-1])
+ if pred[-2] != 0: class_index_count[class_id] += 1
+
+ return dict(class_index_count)
+
+#this is taken from https://github.com/tinygrad/tinygrad/pull/784/files by dc-dc-dc (Now 2 models use upsampling)
+class Upsample:
+ def __init__(self, scale_factor:int, mode: str = "nearest") -> None:
+ assert mode == "nearest" # only mode supported for now
+ self.mode = mode
+ self.scale_factor = scale_factor
+
+ def __call__(self, x: Tensor) -> Tensor:
+ assert len(x.shape) > 2 and len(x.shape) <= 5
+ (b, c), _lens = x.shape[:2], len(x.shape[2:])
+ tmp = x.reshape([b, c, -1] + [1] * _lens) * Tensor.ones(*[1, 1, 1] + [self.scale_factor] * _lens)
+ return tmp.reshape(list(x.shape) + [self.scale_factor] * _lens).permute([0, 1] + list(chain.from_iterable([[y+2, y+2+_lens] for y in range(_lens)]))).reshape([b, c] + [x * self.scale_factor for x in x.shape[2:]])
+
+class Conv_Block:
+ def __init__(self, c1, c2, kernel_size=1, stride=1, groups=1, dilation=1, padding=None):
+ self.conv = Conv2d(c1,c2, kernel_size, stride, padding=autopad(kernel_size, padding, dilation), bias=False, groups=groups, dilation=dilation)
+ self.bn = BatchNorm2d(c2, eps=0.001)
+
+ def __call__(self, x):
+ return self.bn(self.conv(x)).silu()
+
+class Bottleneck:
+ def __init__(self, c1, c2 , shortcut: bool, g=1, kernels: list = (3,3), channel_factor=0.5):
+ c_ = int(c2 * channel_factor)
+ self.cv1 = Conv_Block(c1, c_, kernel_size=kernels[0], stride=1, padding=None)
+ self.cv2 = Conv_Block(c_, c2, kernel_size=kernels[1], stride=1, padding=None, groups=g)
+ self.residual = c1 == c2 and shortcut
+
+ def __call__(self, x):
+ return x + self.cv2(self.cv1(x)) if self.residual else self.cv2(self.cv1(x))
+
+class C2f:
+ def __init__(self, c1, c2, n=1, shortcut=False, g=1, e=0.5):
+ self.c = int(c2 * e)
+ self.cv1 = Conv_Block(c1, 2 * self.c, 1,)
+ self.cv2 = Conv_Block((2 + n) * self.c, c2, 1)
+ self.bottleneck = [Bottleneck(self.c, self.c, shortcut, g, kernels=[(3, 3), (3, 3)], channel_factor=1.0) for _ in range(n)]
+
+ def __call__(self, x):
+ y= list(self.cv1(x).chunk(2, 1))
+ y.extend(m(y[-1]) for m in self.bottleneck)
+ z = y[0]
+ for i in y[1:]: z = z.cat(i, dim=1)
+ return self.cv2(z)
+
+class SPPF:
+ def __init__(self, c1, c2, k=5):
+ c_ = c1 // 2 # hidden channels
+ self.cv1 = Conv_Block(c1, c_, 1, 1, padding=None)
+ self.cv2 = Conv_Block(c_ * 4, c2, 1, 1, padding=None)
+
+ # TODO: this pads with 0s, whereas torch function pads with -infinity. This results in a < 2% difference in prediction which does not make a difference visually.
+ self.maxpool = lambda x : x.pad((k // 2, k // 2, k // 2, k // 2)).max_pool2d(kernel_size=k, stride=1)
+
+ def __call__(self, x):
+ x = self.cv1(x)
+ x2 = self.maxpool(x)
+ x3 = self.maxpool(x2)
+ x4 = self.maxpool(x3)
+ return self.cv2(x.cat(x2, x3, x4, dim=1))
+
+class DFL:
+ def __init__(self, c1=16):
+ self.conv = Conv2d(c1, 1, 1, bias=False)
+ x = Tensor.arange(c1)
+ self.conv.weight.replace(x.reshape(1, c1, 1, 1))
+ self.c1 = c1
+
+ def __call__(self, x):
+ b, c, a = x.shape # batch, channels, anchors
+ return self.conv(x.reshape(b, 4, self.c1, a).transpose(2, 1).softmax(1)).reshape(b, 4, a)
+
+#backbone
+class Darknet:
+ def __init__(self, w, r, d):
+ self.b1 = [Conv_Block(c1=3, c2= int(64*w), kernel_size=3, stride=2, padding=1), Conv_Block(int(64*w), int(128*w), kernel_size=3, stride=2, padding=1)]
+ self.b2 = [C2f(c1=int(128*w), c2=int(128*w), n=round(3*d), shortcut=True), Conv_Block(int(128*w), int(256*w), 3, 2, 1), C2f(int(256*w), int(256*w), round(6*d), True)]
+ self.b3 = [Conv_Block(int(256*w), int(512*w), kernel_size=3, stride=2, padding=1), C2f(int(512*w), int(512*w), round(6*d), True)]
+ self.b4 = [Conv_Block(int(512*w), int(512*w*r), kernel_size=3, stride=2, padding=1), C2f(int(512*w*r), int(512*w*r), round(3*d), True)]
+ self.b5 = [SPPF(int(512*w*r), int(512*w*r), 5)]
+
+ def return_modules(self):
+ return [*self.b1, *self.b2, *self.b3, *self.b4, *self.b5]
+
+ def __call__(self, x):
+ x1 = x.sequential(self.b1)
+ x2 = x1.sequential(self.b2)
+ x3 = x2.sequential(self.b3)
+ x4 = x3.sequential(self.b4)
+ x5 = x4.sequential(self.b5)
+ return (x2, x3, x5)
+
+#yolo fpn (neck)
+class Yolov8NECK:
+ def __init__(self, w, r, d): #width_multiple, ratio_multiple, depth_multiple
+ self.up = Upsample(2, mode='nearest')
+ self.n1 = C2f(c1=int(512*w*(1+r)), c2=int(512*w), n=round(3*d), shortcut=False)
+ self.n2 = C2f(c1=int(768*w), c2=int(256*w), n=round(3*d), shortcut=False)
+ self.n3 = Conv_Block(c1=int(256*w), c2=int(256*w), kernel_size=3, stride=2, padding=1)
+ self.n4 = C2f(c1=int(768*w), c2=int(512*w), n=round(3*d), shortcut=False)
+ self.n5 = Conv_Block(c1=int(512* w), c2=int(512 * w), kernel_size=3, stride=2, padding=1)
+ self.n6 = C2f(c1=int(512*w*(1+r)), c2=int(512*w*r), n=round(3*d), shortcut=False)
+
+ def return_modules(self):
+ return [self.n1, self.n2, self.n3, self.n4, self.n5, self.n6]
+
+ def __call__(self, p3, p4, p5):
+ x = self.n1(self.up(p5).cat(p4, dim=1))
+ head_1 = self.n2(self.up(x).cat(p3, dim=1))
+ head_2 = self.n4(self.n3(head_1).cat(x, dim=1))
+ head_3 = self.n6(self.n5(head_2).cat(p5, dim=1))
+ return [head_1, head_2, head_3]
+
+#task specific head.
+class DetectionHead:
+ def __init__(self, nc=80, filters=()):
+ self.ch = 16
+ self.nc = nc # number of classes
+ self.nl = len(filters)
+ self.no = nc + self.ch * 4 #
+ self.stride = [8, 16, 32]
+ c1 = max(filters[0], self.nc)
+ c2 = max((filters[0] // 4, self.ch * 4))
+ self.dfl = DFL(self.ch)
+ self.cv3 = [[Conv_Block(x, c1, 3), Conv_Block(c1, c1, 3), Conv2d(c1, self.nc, 1)] for x in filters]
+ self.cv2 = [[Conv_Block(x, c2, 3), Conv_Block(c2, c2, 3), Conv2d(c2, 4 * self.ch, 1)] for x in filters]
+
+ def __call__(self, x):
+ for i in range(self.nl):
+ x[i] = (x[i].sequential(self.cv2[i]).cat(x[i].sequential(self.cv3[i]), dim=1))
+ self.anchors, self.strides = (x.transpose(0, 1) for x in make_anchors(x, self.stride, 0.5))
+ y = [(i.reshape(x[0].shape[0], self.no, -1)) for i in x]
+ x_cat = y[0].cat(y[1], y[2], dim=2)
+ box, cls = x_cat[:, :self.ch * 4], x_cat[:, self.ch * 4:]
+ dbox = dist2bbox(self.dfl(box), self.anchors.unsqueeze(0), xywh=True, dim=1) * self.strides
+ z = dbox.cat(cls.sigmoid(), dim=1)
+ return z
+
+class YOLOv8:
+ def __init__(self, w, r, d, num_classes): #width_multiple, ratio_multiple, depth_multiple
+ self.net = Darknet(w, r, d)
+ self.fpn = Yolov8NECK(w, r, d)
+ self.head = DetectionHead(num_classes, filters=(int(256*w), int(512*w), int(512*w*r)))
+
+ def __call__(self, x):
+ x = self.net(x)
+ x = self.fpn(*x)
+ x = self.head(x)
+ # TODO: postprocess needs to be in the model to be compiled to webgpu
+ return postprocess(x)
+
+ def return_all_trainable_modules(self):
+ backbone_modules = [*range(10)]
+ yolov8neck_modules = [12, 15, 16, 18, 19, 21]
+ yolov8_head_weights = [(22, self.head)]
+ return [*zip(backbone_modules, self.net.return_modules()), *zip(yolov8neck_modules, self.fpn.return_modules()), *yolov8_head_weights]
+
+def convert_f16_safetensor_to_f32(input_file: Path, output_file: Path):
+ with open(input_file, 'rb') as f:
+ metadata_length = int.from_bytes(f.read(8), 'little')
+ metadata = json.loads(f.read(metadata_length).decode())
+ float32_values = np.fromfile(f, dtype=np.float16).astype(np.float32)
+
+ for v in metadata.values():
+ if v["dtype"] == "F16": v.update({"dtype": "F32", "data_offsets": [offset * 2 for offset in v["data_offsets"]]})
+
+ with open(output_file, 'wb') as f:
+ new_metadata_bytes = json.dumps(metadata).encode()
+ f.write(len(new_metadata_bytes).to_bytes(8, 'little'))
+ f.write(new_metadata_bytes)
+ float32_values.tofile(f)
+
+def compute_iou_matrix(boxes):
+ x1, y1, x2, y2 = boxes[:, 0], boxes[:, 1], boxes[:, 2], boxes[:, 3]
+ areas = (x2 - x1) * (y2 - y1)
+ x1 = Tensor.maximum(x1[:, None], x1[None, :])
+ y1 = Tensor.maximum(y1[:, None], y1[None, :])
+ x2 = Tensor.minimum(x2[:, None], x2[None, :])
+ y2 = Tensor.minimum(y2[:, None], y2[None, :])
+ w = Tensor.maximum(Tensor(0), x2 - x1)
+ h = Tensor.maximum(Tensor(0), y2 - y1)
+ intersection = w * h
+ union = areas[:, None] + areas[None, :] - intersection
+ return intersection / union
+
+def postprocess(output, max_det=300, conf_threshold=0.25, iou_threshold=0.45):
+ xc, yc, w, h, class_scores = output[0][0], output[0][1], output[0][2], output[0][3], output[0][4:]
+ class_ids = Tensor.argmax(class_scores, axis=0)
+ probs = Tensor.max(class_scores, axis=0)
+ probs = Tensor.where(probs >= conf_threshold, probs, 0)
+ x1 = xc - w / 2
+ y1 = yc - h / 2
+ x2 = xc + w / 2
+ y2 = yc + h / 2
+ boxes = Tensor.stack(x1, y1, x2, y2, probs, class_ids, dim=1)
+ order = Tensor.topk(probs, max_det)[1]
+ boxes = boxes[order]
+ iou = compute_iou_matrix(boxes[:, :4])
+ iou = Tensor.triu(iou, diagonal=1)
+ same_class_mask = boxes[:, -1][:, None] == boxes[:, -1][None, :]
+ high_iou_mask = (iou > iou_threshold) & same_class_mask
+ no_overlap_mask = high_iou_mask.sum(axis=0) == 0
+ boxes = boxes * no_overlap_mask.unsqueeze(-1)
+ return boxes
+
+def get_weights_location(yolo_variant: str) -> Path:
+ weights_location = Path(__file__).parents[1] / "weights" / f'yolov8{yolo_variant}.safetensors'
+ fetch(f'https://gitlab.com/r3sist/yolov8_weights/-/raw/master/yolov8{yolo_variant}.safetensors', weights_location)
+ #f16
+ return weights_location.with_name(f"{weights_location.stem}.safetensors")
+ f32_weights = weights_location.with_name(f"{weights_location.stem}_f32.safetensors")
+ if not f32_weights.exists(): convert_f16_safetensor_to_f32(weights_location, f32_weights)
+ return f32_weights
+
+import struct
+from time import monotonic
+from uvicorn import Server, Config
+import asyncio
+from starlette.types import Receive, Send, Scope
+from starlette.responses import Response
+
+@TinyJit
+def do_inf(im):
+ im = im.unsqueeze(0)
+ im = im[..., ::-1].permute(0, 3, 1, 2)
+ im = im / 255.0
+ predictions = yolo_infer(im)
+ return predictions
+
+# Model initialization (same as before)
+yolo_variant = sys.argv[1] if len(sys.argv) >= 2 else (print("No variant given, so choosing 'n' as the default. Yolov8 has different variants, you can choose from ['n', 's', 'm', 'l', 'x']") or 'n')
+depth, width, ratio = get_variant_multiples(yolo_variant)
+yolo_infer = YOLOv8(w=width, r=ratio, d=depth, num_classes=80)
+state_dict = safe_load(get_weights_location(yolo_variant))
+load_state_dict(yolo_infer, state_dict)
+
import cv2
import time
# Replace with your RTSP stream URL
rtsp_url = ""
+
+def clear_buffer(cap, frames_to_discard=5):
+ for _ in range(frames_to_discard):
+ cap.grab() # Grabs the frame but does not decode
+ time.sleep(0.03) # slight delay to allow new frames to queue
+
# Open stream
cap = cv2.VideoCapture(rtsp_url)
if not cap.isOpened():
print("Failed to open RTSP stream.")
else:
- # --- Capture 10 frames, 1 per second ---
- for i in range(10):
+ for i in range(1000):
+ clear_buffer(cap) # discard old buffered frames
ret, frame = cap.read()
if ret:
- filename = f"frame_{i}.jpg"
- cv2.imwrite(filename, frame)
- print(type(frame))
- print(f"Saved {filename}")
+ #filename = f"frame_{i}.jpg"
+ #cv2.imwrite(filename, frame)
+ print(i)
+ frame = preprocess(frame)
+ preds = do_inf(frame)
+ print(preds.numpy())
else:
print(f"Failed to read frame {i}")
- time.sleep(1)
# --- Setup video writer for last 10 seconds ---
@@ -49,4 +395,4 @@ else:
# --- Cleanup ---
out.release()
- cap.release()
+ cap.release()
\ No newline at end of file