def test_anchor_generator_with_tuples(): from mmdet.core.anchor import build_anchor_generator if torch.cuda.is_available(): device = 'cuda' else: device = 'cpu' anchor_generator_cfg = dict(type='SSDAnchorGenerator', scale_major=False, input_size=300, basesize_ratio_range=(0.15, 0.9), strides=[8, 16, 32, 64, 100, 300], ratios=[[2], [2, 3], [2, 3], [2, 3], [2], [2]]) featmap_sizes = [(38, 38), (19, 19), (10, 10), (5, 5), (3, 3), (1, 1)] anchor_generator = build_anchor_generator(anchor_generator_cfg) anchors = anchor_generator.grid_anchors(featmap_sizes, device) anchor_generator_cfg_tuples = dict(type='SSDAnchorGenerator', scale_major=False, input_size=300, basesize_ratio_range=(0.15, 0.9), strides=[(8, 8), (16, 16), (32, 32), (64, 64), (100, 100), (300, 300)], ratios=[[2], [2, 3], [2, 3], [2, 3], [2], [2]]) anchor_generator_tuples = build_anchor_generator( anchor_generator_cfg_tuples) anchors_tuples = anchor_generator_tuples.grid_anchors( featmap_sizes, device) for anchor, anchor_tuples in zip(anchors, anchors_tuples): assert torch.equal(anchor, anchor_tuples)
def test_standard_anchor_generator(): from mmdet.core.anchor import build_anchor_generator anchor_generator_cfg = dict(type='AnchorGenerator', scales=[8], ratios=[0.5, 1.0, 2.0], strides=[4, 8]) anchor_generator = build_anchor_generator(anchor_generator_cfg) assert anchor_generator is not None
def __init__(self, anchor_generator, in_channels, kernel_size=3, norm_cfg=dict(type='BN'), weighted_sum=False, bbox_coder=dict( type='DeltaXYWHBBoxCoder', target_means=[0., 0., 0., 0.], target_stds=[1., 1., 1., 1.]), loss_cls=dict( type='CrossEntropyLoss', reduction='sum', loss_weight=1.0), loss_bbox=dict( type='L1Loss', reduction='sum', loss_weight=1.2), train_cfg=None, test_cfg=None, *args, **kwargs): super(SiameseRPNHead, self).__init__(*args, **kwargs) self.anchor_generator = build_anchor_generator(anchor_generator) self.bbox_coder = build_bbox_coder(bbox_coder) self.train_cfg = train_cfg self.test_cfg = test_cfg self.assigner = build_assigner(self.train_cfg.assigner) self.sampler = build_sampler(self.train_cfg.sampler) self.cls_heads = nn.ModuleList() self.reg_heads = nn.ModuleList() for i in range(len(in_channels)): self.cls_heads.append( CorrelationHead(in_channels[i], in_channels[i], 2 * self.anchor_generator.num_base_anchors[0], kernel_size, norm_cfg)) self.reg_heads.append( CorrelationHead(in_channels[i], in_channels[i], 4 * self.anchor_generator.num_base_anchors[0], kernel_size, norm_cfg)) self.weighted_sum = weighted_sum if self.weighted_sum: self.cls_weight = nn.Parameter(torch.ones(len(in_channels))) self.reg_weight = nn.Parameter(torch.ones(len(in_channels))) self.loss_cls = build_loss(loss_cls) self.loss_bbox = build_loss(loss_bbox)
def test_yolo_anchor_generator(): from mmdet.core.anchor import build_anchor_generator if torch.cuda.is_available(): device = 'cuda' else: device = 'cpu' anchor_generator_cfg = dict(type='YOLOAnchorGenerator', strides=[32, 16, 8], base_sizes=[ [(116, 90), (156, 198), (373, 326)], [(30, 61), (62, 45), (59, 119)], [(10, 13), (16, 30), (33, 23)], ]) featmap_sizes = [(14, 18), (28, 36), (56, 72)] anchor_generator = build_anchor_generator(anchor_generator_cfg) # check base anchors expected_base_anchors = [ torch.Tensor([[-42.0000, -29.0000, 74.0000, 61.0000], [-62.0000, -83.0000, 94.0000, 115.0000], [-170.5000, -147.0000, 202.5000, 179.0000]]), torch.Tensor([[-7.0000, -22.5000, 23.0000, 38.5000], [-23.0000, -14.5000, 39.0000, 30.5000], [-21.5000, -51.5000, 37.5000, 67.5000]]), torch.Tensor([[-1.0000, -2.5000, 9.0000, 10.5000], [-4.0000, -11.0000, 12.0000, 19.0000], [-12.5000, -7.5000, 20.5000, 15.5000]]) ] base_anchors = anchor_generator.base_anchors for i, base_anchor in enumerate(base_anchors): assert base_anchor.allclose(expected_base_anchors[i]) # check number of base anchors for each level assert anchor_generator.num_base_anchors == [3, 3, 3] # check anchor generation anchors = anchor_generator.grid_anchors(featmap_sizes, device) assert len(anchors) == 3
def main(): anchor_generator_cfg = dict(type="AnchorGenerator", scales=[8], ratios=[0.5, 1.0, 2.0], strides=[4, 8, 16, 32, 64]) anchor_generator: AnchorGenerator = build_anchor_generator( anchor_generator_cfg) multi_level_anchors = anchor_generator.grid_anchors( featmap_sizes=[ torch.Size([256, 256]), torch.Size([128, 128]), torch.Size([64, 64]), torch.Size([32, 32]), torch.Size([16, 16]), ], device="cpu", ) anchors = torch.cat(multi_level_anchors).numpy() widths = anchors[:, 2] - anchors[:, 0] heights = anchors[:, 3] - anchors[:, 1] data = np.stack([heights, widths], axis=1) clusters = kmeans(data, k=50) print(f"aspect rations: {clusters[: 0] / clusters[: 1]}") print(f"sizes: {np.sqrt(clusters[: 0] * clusters[: 1])}")
def test_ssd_anchor_generator(): from mmdet.core.anchor import build_anchor_generator if torch.cuda.is_available(): device = 'cuda' else: device = 'cpu' anchor_generator_cfg = dict(type='SSDAnchorGenerator', scale_major=False, input_size=300, basesize_ratio_range=(0.15, 0.9), strides=[8, 16, 32, 64, 100, 300], ratios=[[2], [2, 3], [2, 3], [2, 3], [2], [2]]) featmap_sizes = [(38, 38), (19, 19), (10, 10), (5, 5), (3, 3), (1, 1)] anchor_generator = build_anchor_generator(anchor_generator_cfg) # check base anchors expected_base_anchors = [ torch.Tensor([[-6.5000, -6.5000, 14.5000, 14.5000], [-11.3704, -11.3704, 19.3704, 19.3704], [-10.8492, -3.4246, 18.8492, 11.4246], [-3.4246, -10.8492, 11.4246, 18.8492]]), torch.Tensor([[-14.5000, -14.5000, 30.5000, 30.5000], [-25.3729, -25.3729, 41.3729, 41.3729], [-23.8198, -7.9099, 39.8198, 23.9099], [-7.9099, -23.8198, 23.9099, 39.8198], [-30.9711, -4.9904, 46.9711, 20.9904], [-4.9904, -30.9711, 20.9904, 46.9711]]), torch.Tensor([[-33.5000, -33.5000, 65.5000, 65.5000], [-45.5366, -45.5366, 77.5366, 77.5366], [-54.0036, -19.0018, 86.0036, 51.0018], [-19.0018, -54.0036, 51.0018, 86.0036], [-69.7365, -12.5788, 101.7365, 44.5788], [-12.5788, -69.7365, 44.5788, 101.7365]]), torch.Tensor([[-44.5000, -44.5000, 108.5000, 108.5000], [-56.9817, -56.9817, 120.9817, 120.9817], [-76.1873, -22.0937, 140.1873, 86.0937], [-22.0937, -76.1873, 86.0937, 140.1873], [-100.5019, -12.1673, 164.5019, 76.1673], [-12.1673, -100.5019, 76.1673, 164.5019]]), torch.Tensor([[-53.5000, -53.5000, 153.5000, 153.5000], [-66.2185, -66.2185, 166.2185, 166.2185], [-96.3711, -23.1855, 196.3711, 123.1855], [-23.1855, -96.3711, 123.1855, 196.3711]]), torch.Tensor([[19.5000, 19.5000, 280.5000, 280.5000], [6.6342, 6.6342, 293.3658, 293.3658], [-34.5549, 57.7226, 334.5549, 242.2774], [57.7226, -34.5549, 242.2774, 334.5549]]), ] base_anchors = anchor_generator.base_anchors for i, base_anchor in enumerate(base_anchors): assert base_anchor.allclose(expected_base_anchors[i]) # check valid flags expected_valid_pixels = [5776, 2166, 600, 150, 36, 4] multi_level_valid_flags = anchor_generator.valid_flags( featmap_sizes, (300, 300), device) for i, single_level_valid_flag in enumerate(multi_level_valid_flags): assert single_level_valid_flag.sum() == expected_valid_pixels[i] # check number of base anchors for each level assert anchor_generator.num_base_anchors == [4, 6, 6, 6, 4, 4] # check anchor generation anchors = anchor_generator.grid_anchors(featmap_sizes, device) assert len(anchors) == 6
def test_ssd_anchor_generator(): from mmdet.core.anchor import build_anchor_generator if torch.cuda.is_available(): device = 'cuda' else: device = 'cpu' # min_sizes max_sizes must set at the same time with pytest.raises(AssertionError): anchor_generator_cfg = dict(type='SSDAnchorGenerator', scale_major=False, min_sizes=[48, 100, 150, 202, 253, 300], max_sizes=None, strides=[8, 16, 32, 64, 100, 300], ratios=[[2], [2, 3], [2, 3], [2, 3], [2], [2]]) build_anchor_generator(anchor_generator_cfg) # length of min_sizes max_sizes must be the same with pytest.raises(AssertionError): anchor_generator_cfg = dict(type='SSDAnchorGenerator', scale_major=False, min_sizes=[48, 100, 150, 202, 253, 300], max_sizes=[100, 150, 202, 253], strides=[8, 16, 32, 64, 100, 300], ratios=[[2], [2, 3], [2, 3], [2, 3], [2], [2]]) build_anchor_generator(anchor_generator_cfg) # test setting anchor size manually anchor_generator_cfg = dict(type='SSDAnchorGenerator', scale_major=False, min_sizes=[48, 100, 150, 202, 253, 304], max_sizes=[100, 150, 202, 253, 304, 320], strides=[16, 32, 64, 107, 160, 320], ratios=[[2, 3], [2, 3], [2, 3], [2, 3], [2, 3], [2, 3]]) featmap_sizes = [(38, 38), (19, 19), (10, 10), (5, 5), (3, 3), (1, 1)] anchor_generator = build_anchor_generator(anchor_generator_cfg) expected_base_anchors = [ torch.Tensor([[-16.0000, -16.0000, 32.0000, 32.0000], [-26.6410, -26.6410, 42.6410, 42.6410], [-25.9411, -8.9706, 41.9411, 24.9706], [-8.9706, -25.9411, 24.9706, 41.9411], [-33.5692, -5.8564, 49.5692, 21.8564], [-5.8564, -33.5692, 21.8564, 49.5692]]), torch.Tensor([[-34.0000, -34.0000, 66.0000, 66.0000], [-45.2372, -45.2372, 77.2372, 77.2372], [-54.7107, -19.3553, 86.7107, 51.3553], [-19.3553, -54.7107, 51.3553, 86.7107], [-70.6025, -12.8675, 102.6025, 44.8675], [-12.8675, -70.6025, 44.8675, 102.6025]]), torch.Tensor([[-43.0000, -43.0000, 107.0000, 107.0000], [-55.0345, -55.0345, 119.0345, 119.0345], [-74.0660, -21.0330, 138.0660, 85.0330], [-21.0330, -74.0660, 85.0330, 138.0660], [-97.9038, -11.3013, 161.9038, 75.3013], [-11.3013, -97.9038, 75.3013, 161.9038]]), torch.Tensor([[-47.5000, -47.5000, 154.5000, 154.5000], [-59.5332, -59.5332, 166.5332, 166.5332], [-89.3356, -17.9178, 196.3356, 124.9178], [-17.9178, -89.3356, 124.9178, 196.3356], [-121.4371, -4.8124, 228.4371, 111.8124], [-4.8124, -121.4371, 111.8124, 228.4371]]), torch.Tensor([[-46.5000, -46.5000, 206.5000, 206.5000], [-58.6651, -58.6651, 218.6651, 218.6651], [-98.8980, -9.4490, 258.8980, 169.4490], [-9.4490, -98.8980, 169.4490, 258.8980], [-139.1044, 6.9652, 299.1044, 153.0348], [6.9652, -139.1044, 153.0348, 299.1044]]), torch.Tensor([[8.0000, 8.0000, 312.0000, 312.0000], [4.0513, 4.0513, 315.9487, 315.9487], [-54.9605, 52.5198, 374.9604, 267.4802], [52.5198, -54.9605, 267.4802, 374.9604], [-103.2717, 72.2428, 423.2717, 247.7572], [72.2428, -103.2717, 247.7572, 423.2717]]) ] base_anchors = anchor_generator.base_anchors for i, base_anchor in enumerate(base_anchors): assert base_anchor.allclose(expected_base_anchors[i]) # check valid flags expected_valid_pixels = [2400, 600, 150, 54, 24, 6] multi_level_valid_flags = anchor_generator.valid_flags( featmap_sizes, (320, 320), device) for i, single_level_valid_flag in enumerate(multi_level_valid_flags): assert single_level_valid_flag.sum() == expected_valid_pixels[i] # check number of base anchors for each level assert anchor_generator.num_base_anchors == [6, 6, 6, 6, 6, 6] # check anchor generation anchors = anchor_generator.grid_anchors(featmap_sizes, device) assert len(anchors) == 6 # test vgg ssd anchor setting anchor_generator_cfg = dict(type='SSDAnchorGenerator', scale_major=False, input_size=300, basesize_ratio_range=(0.15, 0.9), strides=[8, 16, 32, 64, 100, 300], ratios=[[2], [2, 3], [2, 3], [2, 3], [2], [2]]) featmap_sizes = [(38, 38), (19, 19), (10, 10), (5, 5), (3, 3), (1, 1)] anchor_generator = build_anchor_generator(anchor_generator_cfg) # check base anchors expected_base_anchors = [ torch.Tensor([[-6.5000, -6.5000, 14.5000, 14.5000], [-11.3704, -11.3704, 19.3704, 19.3704], [-10.8492, -3.4246, 18.8492, 11.4246], [-3.4246, -10.8492, 11.4246, 18.8492]]), torch.Tensor([[-14.5000, -14.5000, 30.5000, 30.5000], [-25.3729, -25.3729, 41.3729, 41.3729], [-23.8198, -7.9099, 39.8198, 23.9099], [-7.9099, -23.8198, 23.9099, 39.8198], [-30.9711, -4.9904, 46.9711, 20.9904], [-4.9904, -30.9711, 20.9904, 46.9711]]), torch.Tensor([[-33.5000, -33.5000, 65.5000, 65.5000], [-45.5366, -45.5366, 77.5366, 77.5366], [-54.0036, -19.0018, 86.0036, 51.0018], [-19.0018, -54.0036, 51.0018, 86.0036], [-69.7365, -12.5788, 101.7365, 44.5788], [-12.5788, -69.7365, 44.5788, 101.7365]]), torch.Tensor([[-44.5000, -44.5000, 108.5000, 108.5000], [-56.9817, -56.9817, 120.9817, 120.9817], [-76.1873, -22.0937, 140.1873, 86.0937], [-22.0937, -76.1873, 86.0937, 140.1873], [-100.5019, -12.1673, 164.5019, 76.1673], [-12.1673, -100.5019, 76.1673, 164.5019]]), torch.Tensor([[-53.5000, -53.5000, 153.5000, 153.5000], [-66.2185, -66.2185, 166.2185, 166.2185], [-96.3711, -23.1855, 196.3711, 123.1855], [-23.1855, -96.3711, 123.1855, 196.3711]]), torch.Tensor([[19.5000, 19.5000, 280.5000, 280.5000], [6.6342, 6.6342, 293.3658, 293.3658], [-34.5549, 57.7226, 334.5549, 242.2774], [57.7226, -34.5549, 242.2774, 334.5549]]), ] base_anchors = anchor_generator.base_anchors for i, base_anchor in enumerate(base_anchors): assert base_anchor.allclose(expected_base_anchors[i]) # check valid flags expected_valid_pixels = [5776, 2166, 600, 150, 36, 4] multi_level_valid_flags = anchor_generator.valid_flags( featmap_sizes, (300, 300), device) for i, single_level_valid_flag in enumerate(multi_level_valid_flags): assert single_level_valid_flag.sum() == expected_valid_pixels[i] # check number of base anchors for each level assert anchor_generator.num_base_anchors == [4, 6, 6, 6, 4, 4] # check anchor generation anchors = anchor_generator.grid_anchors(featmap_sizes, device) assert len(anchors) == 6