From 39dcf0d561bb66e42835a48b9430c8c1450c2a3a Mon Sep 17 00:00:00 2001 From: Glenn Jocher Date: Sat, 24 Aug 2019 16:43:43 +0200 Subject: [PATCH] removed xy/wh loss reporting --- test.py | 6 +++--- train.py | 42 +++++++++++------------------------------- utils/gcp.sh | 2 +- utils/utils.py | 22 +++++++++++----------- 4 files changed, 26 insertions(+), 46 deletions(-) diff --git a/test.py b/test.py index a16e57c3..bf063102 100644 --- a/test.py +++ b/test.py @@ -55,7 +55,7 @@ def test(cfg, seen = 0 model.eval() coco91class = coco80_to_coco91_class() - s = ('%30s' + '%10s' * 6) % ('Class', 'Images', 'Targets', 'P', 'R', 'mAP', 'F1') + s = ('%20s' + '%10s' * 6) % ('Class', 'Images', 'Targets', 'P', 'R', 'mAP', 'F1') p, r, f1, mp, mr, map, mf1 = 0., 0., 0., 0., 0., 0., 0. loss = torch.zeros(3) jdict, stats, ap, ap_class = [], [], [], [] @@ -73,7 +73,7 @@ def test(cfg, # Compute loss if hasattr(model, 'hyp'): # if model has loss hyperparameters - loss += compute_loss(train_out, targets, model)[1][[0, 2, 3]].cpu() # GIoU, obj, cls + loss += compute_loss(train_out, targets, model)[1][:3].cpu() # GIoU, obj, cls # Run NMS output = non_max_suppression(inf_out, conf_thres=conf_thres, nms_thres=nms_thres) @@ -155,7 +155,7 @@ def test(cfg, nt = torch.zeros(1) # Print results - pf = '%30s' + '%10.3g' * 6 # print format + pf = '%20s' + '%10.3g' * 6 # print format print(pf % ('all', seen, nt.sum(), mp, mr, map, mf1)) # Print results per class diff --git a/train.py b/train.py index e444f05a..55b48a57 100644 --- a/train.py +++ b/train.py @@ -16,28 +16,8 @@ try: # Mixed precision training https://github.com/NVIDIA/apex except: mixed_precision = False # not installed -# 320 --epochs 1 -# 0.109 0.297 0.150 0.126 7.04 1.666 4.062 0.1845 42.6 3.34 12.61 8.338 0.2705 0.001 -4 0.9 0.0005 a 320 giou + best_anchor False -# 0.223 0.218 0.138 0.189 9.28 1.153 4.376 0.08263 24.28 3.05 20.93 2.842 0.2759 0.001357 -5.036 0.9158 0.0005722 b mAP/F1 - 50/50 weighting -# 0.231 0.215 0.135 0.191 9.51 1.432 3.007 0.06082 24.87 3.477 24.13 2.802 0.3436 0.001127 -5.036 0.9232 0.0005874 c -# 0.246 0.194 0.128 0.192 8.12 1.101 3.954 0.0817 22.83 3.967 19.83 1.779 0.3352 0.000895 -5.036 0.9238 0.0007973 d -# 0.187 0.237 0.144 0.186 14.6 1.607 4.202 0.09439 39.27 3.726 31.26 2.634 0.273 0.001542 -5.036 0.8364 0.0008393 e -# 0.250 0.217 0.136 0.195 3.3 1.2 2 0.604 15.7 3.67 20 1.36 0.194 0.00128 -4 0.95 0.000201 0.8 0.388 1.2 0.119 0.0589 0.401 f -# 0.269 0.225 0.149 0.218 6.71 1.13 5.25 0.246 22.4 3.64 17.8 1.31 0.256 0.00146 -4 0.936 0.00042 0.123 0.18 1.81 0.0987 0.0788 0.441 g -# 0.179 0.274 0.165 0.187 7.95 1.22 7.62 0.224 17 5.71 17.7 3.28 0.295 0.00136 -4 0.875 0.000319 0.131 0.208 2.14 0.14 0.0773 0.228 h -# 0.296 0.228 0.152 0.220 5.18 1.43 4.27 0.265 11.7 4.81 11.5 1.56 0.281 0.0013 -4 0.944 0.000427 0.0599 0.142 1.03 0.0552 0.0555 0.434 i - -# 320 --epochs 2 -# 0.242 0.296 0.196 0.231 5.67 0.8541 4.286 0.1539 21.61 1.957 22.9 2.894 0.3689 0.001844 -4 0.913 0.000467 # ha 0.417 mAP @ epoch 100 -# 0.298 0.244 0.167 0.247 4.99 0.8896 4.067 0.1694 21.41 2.033 25.61 1.783 0.4115 0.00128 -4 0.950 0.000377 # hb -# 0.268 0.268 0.178 0.240 4.36 1.104 5.596 0.2087 14.47 2.599 16.27 2.406 0.4114 0.001585 -4 0.950 0.000524 # hc -# 0.161 0.327 0.190 0.193 7.82 1.153 4.062 0.1845 24.28 3.05 20.93 2.842 0.2759 0.001357 -4 0.916 0.000572 # hd 0.438 mAP @ epoch 100 - # Hyperparameters (j-series, 50.5 mAP yolov3-320) evolved by @ktian08 https://github.com/ultralytics/yolov3/issues/310 -# Transfer learning edge layers: 0.1 lr0, 0.9 momentum hyp = {'giou': 1.582, # giou loss gain - 'xy': 4.688, # xy loss gain - 'wh': 0.1857, # wh loss gain 'cls': 27.76, # cls loss gain (CE=~1.0, uCE=~20) 'cls_pw': 1.446, # cls BCELoss positive_weight 'obj': 21.35, # obj loss gain (*=80 for uBCE with 80 classes) @@ -125,10 +105,11 @@ def train(): # possible weights are 'yolov3.weights', 'yolov3-tiny.conv.15', 'darknet53.conv.74' etc. cutoff = load_darknet_weights(model, weights) - if opt.transfer: # transfer learning + if opt.transfer: # transfer learning edge (yolo) layers nf = int(model.module_defs[model.yolo_layers[0] - 1]['filters']) # yolo layer size (i.e. 255) - for x in optimizer.param_groups: # lower parameter count can handle more aggressive training hyps + for x in optimizer.param_groups: + # lower param count allows more aggressive training settings: ~0.1 lr0, ~0.9 momentum x['lr'] *= 10 x['momentum'] *= 0.9 @@ -197,13 +178,12 @@ def train(): model_info(model, report='summary') # 'full' or 'summary' nb = len(dataloader) maps = np.zeros(nc) # mAP per class - results = (0, 0, 0, 0, 0, 0, 0) # P, R, mAP, F1, test_loss + results = (0, 0, 0, 0, 0, 0, 0, 0) # 'P', 'R', 'mAP', 'F1', 'val GIoU', 'val Objectness', 'val Classification' t0 = time.time() for epoch in range(start_epoch, epochs): # epoch ------------------------------------------------------------------ model.train() - print(('\n' + '%10s' * 9) % - ('Epoch', 'gpu_mem', 'GIoU/xy', 'wh', 'obj', 'cls', 'total', 'targets', 'img_size')) + print(('\n' + '%10s' * 8) % ('Epoch', 'gpu_mem', 'GIoU', 'obj', 'cls', 'total', 'targets', 'img_size')) # Update scheduler if epoch > 0: @@ -222,7 +202,7 @@ def train(): image_weights = labels_to_image_weights(dataset.labels, nc=nc, class_weights=w) dataset.indices = random.choices(range(dataset.n), weights=image_weights, k=dataset.n) # rand weighted idx - mloss = torch.zeros(5).to(device) # mean losses + mloss = torch.zeros(4).to(device) # mean losses pbar = tqdm(enumerate(dataloader), total=nb) # progress bar for i, (imgs, targets, paths, _) in pbar: # batch ------------------------------------------------------------- ni = i + nb * epoch # number integrated batches (since train start) @@ -280,7 +260,7 @@ def train(): # Print batch results mloss = (mloss * i + loss_items) / (i + 1) # update mean losses mem = torch.cuda.memory_cached() / 1E9 if torch.cuda.is_available() else 0 # (GB) - s = ('%10s' * 2 + '%10.3g' * 7) % ( + s = ('%10s' * 2 + '%10.3g' * 6) % ( '%g/%g' % (epoch, epochs - 1), '%.3gG' % mem, *mloss, len(targets), img_size) pbar.set_description(s) @@ -298,13 +278,13 @@ def train(): # Write epoch results with open('results.txt', 'a') as file: - file.write(s + '%11.3g' * 7 % results + '\n') # P, R, mAP, F1, test_losses=(GIoU, obj, cls) + file.write(s + '%10.3g' * 7 % results + '\n') # P, R, mAP, F1, test_losses=(GIoU, obj, cls) # Write Tensorboard results if tb_writer: - x = list(mloss[:5]) + list(results[:7]) - titles = ['GIoU/XY', 'Width/Height', 'Objectness', 'Classification', 'Train loss', 'Precision', 'Recall', - 'mAP', 'F1', 'val GIoU', 'val Objectness', 'val Classification'] + x = list(mloss) + list(results) + titles = ['GIoU', 'Objectness', 'Classification', 'Train loss', + 'Precision', 'Recall', 'mAP', 'F1', 'val GIoU', 'val Objectness', 'val Classification'] for xi, title in zip(x, titles): tb_writer.add_scalar(title, xi, epoch) diff --git a/utils/gcp.sh b/utils/gcp.sh index ae3d19ad..2cc323f4 100755 --- a/utils/gcp.sh +++ b/utils/gcp.sh @@ -16,7 +16,7 @@ sudo shutdown # Re-clone rm -rf yolov3 # Warning: remove existing -git clone https://github.com/ultralytics/yolov3 # master +git clone https://github.com/ultralytics/yolov3 && cd yolov3 # master # git clone -b test --depth 1 https://github.com/ultralytics/yolov3 test # branch # Train diff --git a/utils/utils.py b/utils/utils.py index a067f7a1..b6977ee2 100755 --- a/utils/utils.py +++ b/utils/utils.py @@ -374,7 +374,7 @@ def compute_loss(p, targets, model): # predictions, targets, model lobj *= k * h['obj'] lcls *= k * h['cls'] loss = lbox + lobj + lcls - return loss, torch.cat((lbox, ft([0]), lobj, lcls, loss)).detach() + return loss, torch.cat((lbox, lobj, lcls, loss)).detach() def build_targets(model, targets): @@ -661,9 +661,9 @@ def kmeans_targets(path='../coco/trainvalno5k.txt', n=9, img_size=416): # from def print_mutation(hyp, results, bucket=''): # Print mutation results to evolve.txt (for use with train.py --evolve) - a = '%11s' * len(hyp) % tuple(hyp.keys()) # hyperparam keys - b = '%11.3g' * len(hyp) % tuple(hyp.values()) # hyperparam values - c = '%11.3g' * len(results) % results # results (P, R, mAP, F1, test_loss) + a = '%10s' * len(hyp) % tuple(hyp.keys()) # hyperparam keys + b = '%10.3g' * len(hyp) % tuple(hyp.values()) # hyperparam values + c = '%10.3g' * len(results) % results # results (P, R, mAP, F1, test_loss) print('\n%s\n%s\nEvolved fitness: %s\n' % (a, b, c)) if bucket: @@ -672,7 +672,7 @@ def print_mutation(hyp, results, bucket=''): with open('evolve.txt', 'a') as f: # append result f.write(c + b + '\n') x = np.unique(np.loadtxt('evolve.txt', ndmin=2), axis=0) # load unique rows - np.savetxt('evolve.txt', x[np.argsort(-fitness(x))], '%11.3g') # save sort by fitness + np.savetxt('evolve.txt', x[np.argsort(-fitness(x))], '%10.3g') # save sort by fitness if bucket: os.system('gsutil cp evolve.txt gs://%s' % bucket) # upload evolve.txt @@ -680,7 +680,7 @@ def print_mutation(hyp, results, bucket=''): def fitness(x): # Returns fitness (for use with results.txt or evolve.txt) - return 0.50 * x[:, 2] + 0.50 * x[:, 3] # fitness = 0.9 * mAP + 0.1 * F1 + return 0.50 * x[:, 2] + 0.50 * x[:, 3] # fitness = 0.5 * mAP + 0.5 * F1 # Plotting functions --------------------------------------------------------------------------------------------------- @@ -803,10 +803,10 @@ def plot_results(start=0, stop=0): # from utils.utils import *; plot_results() # Plot training results files 'results*.txt' fig, ax = plt.subplots(2, 5, figsize=(14, 7)) ax = ax.ravel() - s = ['GIoU', 'Confidence', 'Classification', 'Precision', 'Recall', - 'val GIoU', 'val Confidence', 'val Classification', 'mAP', 'F1'] + s = ['GIoU', 'Objectness', 'Classification', 'Precision', 'Recall', + 'val GIoU', 'val Objectness', 'val Classification', 'mAP', 'F1'] for f in sorted(glob.glob('results*.txt') + glob.glob('../../Downloads/results*.txt')): - results = np.loadtxt(f, usecols=[2, 4, 5, 9, 10, 13, 14, 15, 11, 12], ndmin=2).T + results = np.loadtxt(f, usecols=[2, 3, 4, 8, 9, 12, 13, 14, 10, 11], ndmin=2).T n = results.shape[1] # number of rows x = range(start, min(stop, n) if stop else n) for i in range(10): @@ -826,9 +826,9 @@ def plot_results(start=0, stop=0): # from utils.utils import *; plot_results() def plot_results_overlay(start=0, stop=0): # from utils.utils import *; plot_results_overlay() # Plot training results files 'results*.txt', overlaying train and val losses s = ['train', 'train', 'train', 'Precision', 'mAP', 'val', 'val', 'val', 'Recall', 'F1'] # legends - t = ['GIoU', 'Confidence', 'Classification', 'P-R', 'mAP-F1'] # titles + t = ['GIoU', 'Objectness', 'Classification', 'P-R', 'mAP-F1'] # titles for f in sorted(glob.glob('results*.txt') + glob.glob('../../Downloads/results*.txt')): - results = np.loadtxt(f, usecols=[2, 4, 5, 9, 11, 13, 14, 15, 10, 12], ndmin=2).T + results = np.loadtxt(f, usecols=[2, 3, 4, 8, 9, 12, 13, 14, 10, 11], ndmin=2).T n = results.shape[1] # number of rows x = range(start, min(stop, n) if stop else n) fig, ax = plt.subplots(1, 5, figsize=(14, 3.5))