Spaces:
Runtime error
Runtime error
File size: 19,330 Bytes
7a4b92f 5ab0373 7a4b92f fbb3995 7a4b92f 5ab0373 7a4b92f 2a572c2 7a4b92f 889bde6 7a4b92f 5ab0373 7a4b92f fbb3995 e8f4d7e ddf2165 7a4b92f 2a572c2 7a4b92f 058f18b 2a572c2 7a4b92f 058f18b fbb3995 1a4c886 2a572c2 cef04ce 058f18b 1a4c886 7e4e0ac fbb3995 058f18b 1a4c886 7e4e0ac fbb3995 058f18b 7e4e0ac 2a572c2 7e4e0ac 7a4b92f 058f18b 2a572c2 7a4b92f 2a572c2 7e4e0ac 058f18b 2a572c2 7e4e0ac 058f18b 2a572c2 7e4e0ac 058f18b 7e4e0ac 058f18b 7e4e0ac 2a572c2 7e4e0ac 7a4b92f 058f18b 7a4b92f 5ab0373 7a4b92f c9d11b2 7a4b92f 5ab0373 fbb3995 7a4b92f c9d11b2 7a4b92f 5ab0373 7a4b92f 5ab0373 fbb3995 7a4b92f 5ab0373 cb90e6f 24cdfbf fbb3995 5ab0373 fbb3995 5ab0373 1a4c886 5ab0373 fbb3995 5ab0373 fbb3995 5ab0373 1a4c886 7a4b92f fbb3995 7a4b92f fbb3995 7a4b92f fbb3995 7a4b92f fbb3995 7a4b92f 2482ba4 7a4b92f 058f18b 5b63380 fbb3995 058f18b fbb3995 72d14ec 5b63380 72d14ec fbb3995 f0af381 fbb3995 72d14ec 2ceaa4e f0af381 2ceaa4e f0af381 2ceaa4e 01fd3a6 2ceaa4e a2da5d3 fbb3995 711b619 058f18b fbb3995 01fd3a6 5ede26e 8ad7701 fbb3995 5b63380 fbb3995 058f18b ddf2165 5b63380 1ae5c71 5b63380 1ae5c71 5b63380 fb7d89a 5b63380 1ae5c71 fb7d89a 1ae5c71 fa2b33e 889bde6 5165dd5 5b63380 7d9063f 5b63380 ddf2165 5b63380 6697470 7a4b92f 058f18b 7a4b92f 5ab0373 1a4c886 5ab0373 1a4c886 5ab0373 058f18b 5ab0373 4a7fc58 5ab0373 1ae5c71 1a4c886 1ae5c71 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 |
import project_path
import torch
from tqdm import tqdm
from functools import partial
import numpy as np
import json
import time
from unittest.mock import patch
import math
# assumes yolov5 on sys.path
from lib.yolov5.models.experimental import attempt_load
from lib.yolov5.utils.torch_utils import select_device
from lib.yolov5.utils.general import clip_boxes, scale_boxes, xywh2xyxy
from lib.yolov5.utils.metrics import box_iou
import torch
import torchvision
from InferenceConfig import InferenceConfig, TrackerType
from lib.fish_eye.tracker import Tracker
from lib.fish_eye.associative import Associate
### Configuration options
WEIGHTS = 'models/v5m_896_300best.pt'
# will need to configure these based on GPU hardware
BATCH_SIZE = 32
CONF_THRES = 0.05 # detection
NMS_IOU = 0.2 # NMS IOU
MAX_AGE = 14 # time until missing fish get's new id
MIN_HITS = 16 # minimum number of frames with a specific fish for it to count
MIN_LENGTH = 0.3 # minimum fish length, in meters
IOU_THRES = 0.01 # IOU threshold for tracking
MIN_TRAVEL = -1 # Minimum distance a track has to travel
###
def norm(bbox, w, h):
"""
Normalize a bounding box.
Args:
bbox: list of length 4. Can be [x,y,w,h] or [x0,y0,x1,y1]
w: image width
h: image height
"""
bb = bbox.copy()
bb[0] /= w
bb[1] /= h
bb[2] /= w
bb[3] /= h
return bb
def do_full_inference(dataloader, image_meter_width, image_meter_height, gp=None, config=InferenceConfig()):
# Set up model
model, device = setup_model(config.weights)
# Detect boxes in frames
inference, image_shapes, width, height = do_detection(dataloader, model, device, gp=gp)
inference = filter_detection_size(inference, image_meter_width, width, 1.5)
if config.associative_tracker == TrackerType.BYTETRACK:
# Find low confidence detections
low_outputs = do_suppression(inference, conf_thres=config.byte_low_conf, iou_thres=config.nms_iou, gp=gp)
low_preds, real_width, real_height = format_predictions(image_shapes, low_outputs, width, height, gp=gp)
# Find high confidence detections
high_outputs = do_suppression(inference, conf_thres=config.byte_high_conf, iou_thres=config.nms_iou, gp=gp)
high_preds, real_width, real_height = format_predictions(image_shapes, high_outputs, width, height, gp=gp)
# Perform associative tracking (ByteTrack)
results = do_associative_tracking(
low_preds, high_preds, image_meter_width, image_meter_height,
reverse=False, min_length=config.min_length, min_travel=config.min_travel,
max_age=config.max_age, min_hits=config.min_hits,
gp=gp)
else:
# Find confident detections
outputs = do_suppression(inference, conf_thres=config.conf_thresh, iou_thres=config.nms_iou, gp=gp)
if config.associative_tracker == TrackerType.CONF_BOOST:
# Boost confidence based on found confident detections
do_confidence_boost(inference, outputs, boost_power=config.boost_power, boost_decay=config.boost_decay, gp=gp)
# Find confident detections from boosted list
outputs = do_suppression(inference, conf_thres=config.conf_thresh, iou_thres=config.nms_iou, gp=gp)
# Format confident detections
all_preds, real_width, real_height = format_predictions(image_shapes, outputs, width, height, gp=gp)
# Perform SORT tracking
results = do_tracking(
all_preds, image_meter_width, image_meter_height,
min_length=config.min_length, min_travel=config.min_travel,
max_age=config.max_age, min_hits=config.min_hits,
gp=gp)
return results
def setup_model(weights_fp=WEIGHTS, imgsz=896, batch_size=32):
if torch.cuda.is_available():
device = select_device('0', batch_size=batch_size)
else:
print("CUDA not available. Using CPU inference.")
device = select_device('cpu', batch_size=batch_size)
# Setup model for inference
model = attempt_load(weights_fp, device=device)
half = device.type != 'cpu' # half precision only supported on CUDA
if half:
model.half()
model.eval()
# Create dataloader for batched inference
img = torch.zeros((1, 3, imgsz, imgsz), device=device)
_ = model(img.half() if half else img) if device.type != 'cpu' else None # run once
return model, device
def do_detection(dataloader, model, device, gp=None, batch_size=BATCH_SIZE, verbose=True):
"""
Args:
frames_dir: a directory containing frames to be evaluated
image_meter_width: the width of each image, in meters (used for fish length calculation)
gp: a callback function which takes as input 1 parameter, (int) percent complete
prep_for_marking: re-index fish for manual marking output
"""
if (gp): gp(0, "Detection...")
inference = []
image_shapes = []
# Run detection
with tqdm(total=len(dataloader)*batch_size, desc="Running detection", ncols=0, disable=not verbose) as pbar:
for batch_i, (img, _, shapes) in enumerate(dataloader):
if gp: gp(batch_i / len(dataloader), pbar.__str__())
img = img.to(device, non_blocking=True)
img = img.half() if device.type != 'cpu' else img.float() # uint8 to fp16/32
img /= 255.0 # 0 - 255 to 0.0 - 1.0
size = tuple(img.shape)
nb, _, height, width = size # batch size, channels, height, width
# Run model & NMS
with torch.no_grad():
inf_out, _ = model(img, augment=False)
# Save shapes for resizing to original shape
batch_shape = []
for si, pred in enumerate(inf_out):
batch_shape.append((img[si].shape[1:], shapes[si]))
image_shapes.append(batch_shape)
inference.append(inf_out)
pbar.update(1*batch_size)
return inference, image_shapes, width, height
def do_suppression(inference, gp=None, batch_size=BATCH_SIZE, conf_thres=CONF_THRES, iou_thres=NMS_IOU, verbose=True):
"""
Args:
frames_dir: a directory containing frames to be evaluated
image_meter_width: the width of each image, in meters (used for fish length calculation)
gp: a callback function which takes as input 1 parameter, (int) percent complete
prep_for_marking: re-index fish for manual marking output
"""
if (gp): gp(0, "Suppression...")
# keep predictions to feed them ordered into the Tracker
# TODO: how to deal with large files?
outputs = []
with tqdm(total=len(inference)*batch_size, desc="Running suppression", ncols=0, disable=not verbose) as pbar:
for batch_i, inf_out in enumerate(inference):
if gp: gp(batch_i / len(inference), pbar.__str__())
with torch.no_grad():
output = non_max_suppression(inf_out, conf_thres=conf_thres, iou_thres=iou_thres)
outputs.append(output)
pbar.update(1*batch_size)
return outputs
def format_predictions(image_shapes, outputs, width, height, gp=None, batch_size=BATCH_SIZE, verbose=True):
"""
Args:
frames_dir: a directory containing frames to be evaluated
image_meter_width: the width of each image, in meters (used for fish length calculation)
gp: a callback function which takes as input 1 parameter, (int) percent complete
prep_for_marking: re-index fish for manual marking output
"""
if (gp): gp(0, "Formatting...")
# keep predictions to feed them ordered into the Tracker
# TODO: how to deal with large files?
all_preds = {}
with tqdm(total=len(image_shapes)*batch_size, desc="Running formatting", ncols=0, disable=not verbose) as pbar:
for batch_i, batch in enumerate(outputs):
if gp: gp(batch_i / len(image_shapes), pbar.__str__())
batch_shapes = image_shapes[batch_i]
# Format results
for si, pred in enumerate(batch):
(image_shape, original_shape) = batch_shapes[si]
# Clip boxes to image bounds and resize to input shape
clip_boxes(pred, (height, width))
box = pred[:, :4].clone() # xyxy
confs = pred[:, 4].clone().tolist()
scale_boxes(image_shape, box, original_shape[0], original_shape[1]) # to original shape
# get boxes into tracker input format - normalized xyxy with confidence score
# confidence score currently not used by tracker; set to 1.0
boxes = None
if box.shape[0]:
real_width = original_shape[0][1]
real_height = original_shape[0][0]
do_norm = partial(norm, w=original_shape[0][1], h=original_shape[0][0])
normed = list((map(do_norm, box[:, :4].tolist())))
boxes = np.stack([ [*bb, conf] for bb, conf in zip(normed, confs) ])
frame_num = (batch_i, si)
all_preds[frame_num] = boxes
pbar.update(1*batch_size)
return all_preds, real_width, real_height
# ---------------------------------------- TRACKING ------------------------------------------
def do_tracking(all_preds, image_meter_width, image_meter_height, gp=None, max_age=MAX_AGE, iou_thres=IOU_THRES, min_hits=MIN_HITS, min_length=MIN_LENGTH, min_travel=MIN_TRAVEL, verbose=True):
"""
Perform SORT tracking based on formatted detections
"""
if (gp): gp(0, "Tracking...")
# Initialize tracker
clip_info = {
'start_frame': 0,
'end_frame': len(all_preds),
'image_meter_width': image_meter_width,
'image_meter_height': image_meter_height
}
tracker = Tracker(clip_info, args={ 'max_age': max_age, 'min_hits': 0, 'iou_threshold': iou_thres}, min_hits=min_hits)
# Run tracking
with tqdm(total=len(all_preds), desc="Running tracking", ncols=0, disable=not verbose) as pbar:
for i, key in enumerate(sorted(all_preds.keys())):
if gp: gp(i / len(all_preds), pbar.__str__())
boxes = all_preds[key]
if boxes is not None:
tracker.update(boxes)
else:
tracker.update()
pbar.update(1)
json_data = tracker.finalize(min_length=min_length, min_travel=min_travel)
return json_data
def do_confidence_boost(inference, safe_preds, gp=None, batch_size=BATCH_SIZE, boost_power=1, boost_decay=1, verbose=True):
"""
Takes in the full YOLO detections 'inference' and formatted non-max suppressed detections 'safe_preds'
and boosts the confidence of detections around identified fish that are close in space in neighbouring frames.
"""
if (gp): gp(0, "Confidence Boost...")
# keep predictions to feed them ordered into the Tracker
# TODO: how to deal with large files?
boost_cutoff = 0.01
boost_range = math.floor(math.sqrt(1/boost_decay * math.log(boost_power / boost_cutoff)))
boost_scale = boost_power * math.exp(-boost_decay)
with tqdm(total=len(inference), desc="Running confidence boost", ncols=0, disable=not verbose) as pbar:
for batch_i in range(len(inference)):
if gp: gp(batch_i / len(inference), pbar.__str__())
safe = safe_preds[batch_i]
infer = inference[batch_i]
for i in range(len(safe)):
safe_frame = safe[i]
if len(safe_frame) == 0:
continue
next_batch = inference[batch_i + 1] if batch_i+1 < len(inference) else None
prev_batch = inference[batch_i - 1] if batch_i-1 >= 0 else None
for dt in range(-boost_range, boost_range+1):
if dt == 0: continue
idx = i+dt
temp_frame = None
if idx >= 0 and idx < len(infer):
temp_frame = infer[idx]
elif idx < 0 and prev_batch is not None and -idx >= len(prev_batch):
temp_frame = prev_batch[idx]
elif idx >= len(infer) and next_batch is not None and idx - len(infer) < len(next_batch):
temp_frame = next_batch[idx - len(infer)]
if temp_frame is not None:
boost_frame(safe_frame, temp_frame, dt, power=boost_scale, decay=boost_decay)
pbar.update(1*batch_size)
def boost_frame(safe_frame, base_frame, dt, power=1, decay=1):
"""
Boosts confidence of base_frame based on confidence in safe_frame, iou, and the time difference between frames.
"""
safe_boxes = safe_frame[:, :4]
boxes = xywh2xyxy(base_frame[:, :4]) # center_x, center_y, width, height) to (x1, y1, x2, y2)≈
# If running on CPU, you have to convert to double for the .prod() function in box_iou for some reason?
if torch.cuda.is_available():
ious = box_iou(boxes, safe_boxes)
else:
ious = box_iou(boxes.double(), safe_boxes).float()
score = torch.matmul(ious, safe_frame[:, 4])
# score = iou(safe_box, base_box) * confidence(safe_box)
base_frame[:, 4] *= 1 + power*(score)*math.exp(-decay*(dt*dt-1))
return base_frame
# ByteTrack
def do_associative_tracking(low_preds, high_preds, image_meter_width, image_meter_height, reverse=False, gp=None, max_age=MAX_AGE, iou_thres=IOU_THRES, min_hits=MIN_HITS, min_length=MIN_LENGTH, min_travel=MIN_TRAVEL, verbose=True):
if (gp): gp(0, "Tracking...")
# Initialize tracker
clip_info = {
'start_frame': 0,
'end_frame': len(low_preds),
'image_meter_width': image_meter_width,
'image_meter_height': image_meter_height
}
tracker = Tracker(clip_info, algorithm=Associate, reverse=reverse, args={ 'max_age': max_age, 'min_hits': 0, 'iou_threshold': iou_thres}, min_hits=min_hits)
# Run tracking
with tqdm(total=len(low_preds), desc="Running tracking", ncols=0, disable=not verbose) as pbar:
for i, key in enumerate(sorted(low_preds.keys(), reverse=reverse)):
if gp: gp(i / len(low_preds), pbar.__str__())
low_boxes = low_preds[key]
high_boxes = high_preds[key]
boxes = (low_boxes, high_boxes)
if low_boxes is not None and high_boxes is not None:
tracker.update(boxes)
else:
tracker.update((np.empty((0, 5)), np.empty((0, 5))))
pbar.update(1)
json_data = tracker.finalize(min_length=min_length, min_travel=min_travel)
return json_data
@patch('json.encoder.c_make_encoder', None)
def json_dump_round_float(some_object, out_path, num_digits=4):
"""Write a json file to disk with a specified level of precision.
See: https://gist.github.com/Sukonnik-Illia/ed9b2bec1821cad437d1b8adb17406a3
"""
# saving original method
of = json.encoder._make_iterencode
def inner(*args, **kwargs):
args = list(args)
# fifth argument is float formater which will we replace
fmt_str = '{:.' + str(num_digits) + 'f}'
args[4] = lambda o: fmt_str.format(o)
return of(*args, **kwargs)
with patch('json.encoder._make_iterencode', wraps=inner):
return json.dump(some_object, open(out_path, 'w'), indent=2)
def filter_detection_size(inference, image_meter_width, width, max_length):
outputs = []
for prediction in inference:
pix2width = image_meter_width/width
print("pix2w", pix2width)
width = prediction[..., 2]*pix2width
wc = width < max_length
bs = prediction.shape[0] # batches
output = [torch.zeros((0, 6), device=prediction.device)] * bs
for xi, x in enumerate(prediction):
x = x[wc[xi]] # confidence
output[xi] = x
outputs.append(output)
return outputs
def non_max_suppression(
prediction,
conf_thres=0.25,
iou_thres=0.45,
max_det=300
):
"""Non-Maximum Suppression (NMS) on inference results to reject overlapping detections
NOTE: SIMPLIFIED FOR SINGLE CLASS DETECTION
Returns:
list of detections, on (n,6) tensor per image [xyxy, conf, cls]
"""
# Checks
assert 0 <= conf_thres <= 1, f'Invalid Confidence threshold {conf_thres}, valid values are between 0.0 and 1.0'
assert 0 <= iou_thres <= 1, f'Invalid IoU {iou_thres}, valid values are between 0.0 and 1.0'
if isinstance(prediction, (list, tuple)): # YOLOv5 model in validation model, output = (inference_out, loss_out)
prediction = prediction[0] # select only inference output
device = prediction.device
mps = 'mps' in device.type # Apple MPS
if mps: # MPS not fully supported yet, convert tensors to CPU before NMS
prediction = prediction.cpu()
bs = prediction.shape[0] # batch size
xc = prediction[..., 4] > conf_thres # candidates
# Settings
# min_wh = 2 # (pixels) minimum box width and height
max_nms = 30000 # maximum number of boxes into torchvision.ops.nms()
redundant = True # require redundant detections
merge = False # use merge-NMS
output = [torch.zeros((0, 6), device=prediction.device)] * bs
for xi, x in enumerate(prediction): # image index, image inference
# Keep boxes that pass confidence threshold
x = x[xc[xi]] # confidence
# If none remain process next image
if not x.shape[0]:
continue
# Compute conf
x[:, 5:] *= x[:, 4:5] # conf = obj_conf * cls_conf
# Box/Mask
box = xywh2xyxy(x[:, :4]) # center_x, center_y, width, height) to (x1, y1, x2, y2)
mask = x[:, 6:] # zero columns if no masks
# Detections matrix nx6 (xyxy, conf, cls)
conf, j = x[:, 5:6].max(1, keepdim=True)
x = torch.cat((box, conf, j.float(), mask), 1)[conf.view(-1) > conf_thres]
# Check shape
n = x.shape[0] # number of boxes
if not n: # no boxes
continue
x = x[x[:, 4].argsort(descending=True)[:max_nms]] # sort by confidence and remove excess boxes
# Batched NMS
boxes = x[:, :4] # boxes (offset by class), scores
scores = x[:, 4]
i = torchvision.ops.nms(boxes, scores, iou_thres) # NMS
i = i[:max_det] # limit detections
if merge and (1 < n < 3E3): # Merge NMS (boxes merged using weighted mean)
# update boxes as boxes(i,4) = weights(i,n) * boxes(n,4)
iou = box_iou(boxes[i], boxes) > iou_thres # iou matrix
weights = iou * scores[None] # box weights
x[i, :4] = torch.mm(weights, x[:, :4]).float() / weights.sum(1, keepdim=True) # merged boxes
if redundant:
i = i[iou.sum(1) > 1] # require redundancy
output[xi] = x[i]
if mps:
output[xi] = output[xi].to(device)
logging = False
return output
|