You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

162 line
7.0KB

  1. # Auto-anchor utils
  2. import random
  3. import numpy as np
  4. import torch
  5. import yaml
  6. from tqdm import tqdm
  7. from utils.general import colorstr
  8. def check_anchor_order(m):
  9. # Check anchor order against stride order for YOLOv5 Detect() module m, and correct if necessary
  10. a = m.anchor_grid.prod(-1).view(-1) # anchor area
  11. da = a[-1] - a[0] # delta a
  12. ds = m.stride[-1] - m.stride[0] # delta s
  13. if da.sign() != ds.sign(): # same order
  14. print('Reversing anchor order')
  15. m.anchors[:] = m.anchors.flip(0)
  16. m.anchor_grid[:] = m.anchor_grid.flip(0)
  17. def check_anchors(dataset, model, thr=4.0, imgsz=640):
  18. # Check anchor fit to data, recompute if necessary
  19. prefix = colorstr('autoanchor: ')
  20. print(f'\n{prefix}Analyzing anchors... ', end='')
  21. m = model.module.model[-1] if hasattr(model, 'module') else model.model[-1] # Detect()
  22. shapes = imgsz * dataset.shapes / dataset.shapes.max(1, keepdims=True)
  23. scale = np.random.uniform(0.9, 1.1, size=(shapes.shape[0], 1)) # augment scale
  24. wh = torch.tensor(np.concatenate([l[:, 3:5] * s for s, l in zip(shapes * scale, dataset.labels)])).float() # wh
  25. def metric(k): # compute metric
  26. r = wh[:, None] / k[None]
  27. x = torch.min(r, 1. / r).min(2)[0] # ratio metric
  28. best = x.max(1)[0] # best_x
  29. aat = (x > 1. / thr).float().sum(1).mean() # anchors above threshold
  30. bpr = (best > 1. / thr).float().mean() # best possible recall
  31. return bpr, aat
  32. anchors = m.anchor_grid.clone().cpu().view(-1, 2) # current anchors
  33. bpr, aat = metric(anchors)
  34. print(f'anchors/target = {aat:.2f}, Best Possible Recall (BPR) = {bpr:.4f}', end='')
  35. if bpr < 0.98: # threshold to recompute
  36. print('. Attempting to improve anchors, please wait...')
  37. na = m.anchor_grid.numel() // 2 # number of anchors
  38. try:
  39. anchors = kmean_anchors(dataset, n=na, img_size=imgsz, thr=thr, gen=1000, verbose=False)
  40. except Exception as e:
  41. print(f'{prefix}ERROR: {e}')
  42. new_bpr = metric(anchors)[0]
  43. if new_bpr > bpr: # replace anchors
  44. anchors = torch.tensor(anchors, device=m.anchors.device).type_as(m.anchors)
  45. m.anchor_grid[:] = anchors.clone().view_as(m.anchor_grid) # for inference
  46. m.anchors[:] = anchors.clone().view_as(m.anchors) / m.stride.to(m.anchors.device).view(-1, 1, 1) # loss
  47. check_anchor_order(m)
  48. print(f'{prefix}New anchors saved to model. Update model *.yaml to use these anchors in the future.')
  49. else:
  50. print(f'{prefix}Original anchors better than new anchors. Proceeding with original anchors.')
  51. print('') # newline
  52. def kmean_anchors(dataset='./data/coco128.yaml', n=9, img_size=640, thr=4.0, gen=1000, verbose=True):
  53. """ Creates kmeans-evolved anchors from training dataset
  54. Arguments:
  55. dataset: path to data.yaml, or a loaded dataset
  56. n: number of anchors
  57. img_size: image size used for training
  58. thr: anchor-label wh ratio threshold hyperparameter hyp['anchor_t'] used for training, default=4.0
  59. gen: generations to evolve anchors using genetic algorithm
  60. verbose: print all results
  61. Return:
  62. k: kmeans evolved anchors
  63. Usage:
  64. from utils.autoanchor import *; _ = kmean_anchors()
  65. """
  66. from scipy.cluster.vq import kmeans
  67. thr = 1. / thr
  68. prefix = colorstr('autoanchor: ')
  69. def metric(k, wh): # compute metrics
  70. r = wh[:, None] / k[None]
  71. x = torch.min(r, 1. / r).min(2)[0] # ratio metric
  72. # x = wh_iou(wh, torch.tensor(k)) # iou metric
  73. return x, x.max(1)[0] # x, best_x
  74. def anchor_fitness(k): # mutation fitness
  75. _, best = metric(torch.tensor(k, dtype=torch.float32), wh)
  76. return (best * (best > thr).float()).mean() # fitness
  77. def print_results(k):
  78. k = k[np.argsort(k.prod(1))] # sort small to large
  79. x, best = metric(k, wh0)
  80. bpr, aat = (best > thr).float().mean(), (x > thr).float().mean() * n # best possible recall, anch > thr
  81. print(f'{prefix}thr={thr:.2f}: {bpr:.4f} best possible recall, {aat:.2f} anchors past thr')
  82. print(f'{prefix}n={n}, img_size={img_size}, metric_all={x.mean():.3f}/{best.mean():.3f}-mean/best, '
  83. f'past_thr={x[x > thr].mean():.3f}-mean: ', end='')
  84. for i, x in enumerate(k):
  85. print('%i,%i' % (round(x[0]), round(x[1])), end=', ' if i < len(k) - 1 else '\n') # use in *.cfg
  86. return k
  87. if isinstance(dataset, str): # *.yaml file
  88. with open(dataset, encoding='ascii', errors='ignore') as f:
  89. data_dict = yaml.safe_load(f) # model dict
  90. from utils.datasets import LoadImagesAndLabels
  91. dataset = LoadImagesAndLabels(data_dict['train'], augment=True, rect=True)
  92. # Get label wh
  93. shapes = img_size * dataset.shapes / dataset.shapes.max(1, keepdims=True)
  94. wh0 = np.concatenate([l[:, 3:5] * s for s, l in zip(shapes, dataset.labels)]) # wh
  95. # Filter
  96. i = (wh0 < 3.0).any(1).sum()
  97. if i:
  98. print(f'{prefix}WARNING: Extremely small objects found. {i} of {len(wh0)} labels are < 3 pixels in size.')
  99. wh = wh0[(wh0 >= 2.0).any(1)] # filter > 2 pixels
  100. # wh = wh * (np.random.rand(wh.shape[0], 1) * 0.9 + 0.1) # multiply by random scale 0-1
  101. # Kmeans calculation
  102. print(f'{prefix}Running kmeans for {n} anchors on {len(wh)} points...')
  103. s = wh.std(0) # sigmas for whitening
  104. k, dist = kmeans(wh / s, n, iter=30) # points, mean distance
  105. assert len(k) == n, print(f'{prefix}ERROR: scipy.cluster.vq.kmeans requested {n} points but returned only {len(k)}')
  106. k *= s
  107. wh = torch.tensor(wh, dtype=torch.float32) # filtered
  108. wh0 = torch.tensor(wh0, dtype=torch.float32) # unfiltered
  109. k = print_results(k)
  110. # Plot
  111. # k, d = [None] * 20, [None] * 20
  112. # for i in tqdm(range(1, 21)):
  113. # k[i-1], d[i-1] = kmeans(wh / s, i) # points, mean distance
  114. # fig, ax = plt.subplots(1, 2, figsize=(14, 7), tight_layout=True)
  115. # ax = ax.ravel()
  116. # ax[0].plot(np.arange(1, 21), np.array(d) ** 2, marker='.')
  117. # fig, ax = plt.subplots(1, 2, figsize=(14, 7)) # plot wh
  118. # ax[0].hist(wh[wh[:, 0]<100, 0],400)
  119. # ax[1].hist(wh[wh[:, 1]<100, 1],400)
  120. # fig.savefig('wh.png', dpi=200)
  121. # Evolve
  122. npr = np.random
  123. f, sh, mp, s = anchor_fitness(k), k.shape, 0.9, 0.1 # fitness, generations, mutation prob, sigma
  124. pbar = tqdm(range(gen), desc=f'{prefix}Evolving anchors with Genetic Algorithm:') # progress bar
  125. for _ in pbar:
  126. v = np.ones(sh)
  127. while (v == 1).all(): # mutate until a change occurs (prevent duplicates)
  128. v = ((npr.random(sh) < mp) * random.random() * npr.randn(*sh) * s + 1).clip(0.3, 3.0)
  129. kg = (k.copy() * v).clip(min=2.0)
  130. fg = anchor_fitness(kg)
  131. if fg > f:
  132. f, k = fg, kg.copy()
  133. pbar.desc = f'{prefix}Evolving anchors with Genetic Algorithm: fitness = {f:.4f}'
  134. if verbose:
  135. print_results(k)
  136. return print_results(k)