MobileNets

README.md

@@ -38,9 +38,9 @@ Models from this repo are faster to train (**single card trainable**) and often
 | semantic segmentation | ResNet-101 | [DeeplabV3](https://arxiv.org/abs/1706.05587) |
 | semantic segmentation | - | [ENet](https://arxiv.org/abs/1606.02147) |
 | semantic segmentation | - | [ERFNet](https://ieeexplore.ieee.org/abstract/document/8063438/) |
-| lane detection | ENet, ERFNet, VGG16, ResNets (18, 34, 50, 101), RepVGGs (A0, A1, B0, B1g2, B2), Swin (Tiny) | Baseline |
+| lane detection | ENet, ERFNet, VGG16, ResNets (18, 34, 50, 101), MobileNets (V2, V3-Large), RepVGGs (A0, A1, B0, B1g2, B2), Swin (Tiny) | Baseline |
 | lane detection | ERFNet, VGG16, ResNets (18, 34, 50, 101), RepVGGs (A1) | [SCNN](https://arxiv.org/abs/1712.06080) |
-| lane detection | ResNets (18, 34, 50, 101) | [RESA](https://arxiv.org/abs/2008.13719) |
+| lane detection | ResNets (18, 34, 50, 101), MobileNets (V2, V3-Large) | [RESA](https://arxiv.org/abs/2008.13719) |
 | lane detection | ERFNet, ENet | [SAD](https://arxiv.org/abs/1908.00821) ([*Postponed*](https://github.com/voldemortX/pytorch-auto-drive/wiki/Notes)) |
 | lane detection | ERFNet | [PRNet](http://www.ecva.net/papers/eccv_2020/papers_ECCV/papers/123630698.pdf) (*In progress*) |
 | lane detection | ResNets (18, 34, 50, 101), ResNet18-reduced | [LSTR](https://arxiv.org/abs/2011.04233) |

configs/lane_detection/baseline/mobilenetv2_culane.py

@@ -5,7 +5,7 @@
 
 # Optimization pipeline
 from configs.lane_detection.common.optims.segloss_5class import loss
-from configs.lane_detection.common.optims.sgd02 import optimizer
+from configs.lane_detection.common.optims.sgd009 import optimizer
 from configs.lane_detection.common.optims.ep12_poly_warmup200 import lr_scheduler
 
 

configs/lane_detection/baseline/mobilenetv2_tusimple.py

@@ -5,7 +5,7 @@
 
 # Optimization pipeline
 from configs.lane_detection.common.optims.segloss_7class import loss
-from configs.lane_detection.common.optims.sgd02 import optimizer
+from configs.lane_detection.common.optims.sgd009 import optimizer
 from configs.lane_detection.common.optims.ep50_poly_warmup200 import lr_scheduler
 
 

configs/lane_detection/baseline/mobilenetv3-large_culane.py

@@ -5,7 +5,7 @@
 
 # Optimization pipeline
 from configs.lane_detection.common.optims.segloss_5class import loss
-from configs.lane_detection.common.optims.sgd02 import optimizer
+from configs.lane_detection.common.optims.sgd009 import optimizer
 from configs.lane_detection.common.optims.ep12_poly_warmup200 import lr_scheduler
 
 

configs/lane_detection/baseline/mobilenetv3-large_tusimple.py

@@ -0,0 +1,86 @@
+# Data pipeline
+from configs.lane_detection.common.datasets.tusimple_seg import dataset
+from configs.lane_detection.common.datasets.train_level0_360 import train_augmentation
+from configs.lane_detection.common.datasets.test_360 import test_augmentation
+
+# Optimization pipeline
+from configs.lane_detection.common.optims.segloss_7class import loss
+from configs.lane_detection.common.optims.sgd009 import optimizer
+from configs.lane_detection.common.optims.ep50_poly_warmup200 import lr_scheduler
+
+
+train = dict(
+    exp_name='mobilenetv3-large_baseline_tusimple',
+    workers=10,
+    batch_size=20,
+    checkpoint=None,
+    # Device args
+    world_size=0,
+    dist_url='env://',
+    device='cuda',
+
+    val_num_steps=0,  # Seg IoU validation (mostly useless)
+    save_dir='./checkpoints',
+
+    input_size=(360, 640),
+    original_size=(720, 1280),
+    num_classes=7,
+    num_epochs=50,
+    collate_fn=None,  # 'dict_collate_fn' for LSTR
+    seg=True  # Seg-based method or not
+)
+
+test = dict(
+    exp_name='mobilenetv3-large_baseline_tusimple',
+    workers=10,
+    batch_size=80,
+    checkpoint='./checkpoints/mobilenetv3-large_baseline_tusimple/model.pt',
+    # Device args
+    device='cuda',
+
+    save_dir='./checkpoints',
+
+    seg=True,
+    gap=10,
+    ppl=56,
+    thresh=0.3,
+    collate_fn=None,  # 'dict_collate_fn' for LSTR
+    input_size=(360, 640),
+    original_size=(720, 1280),
+    max_lane=5,
+    dataset_name='tusimple'
+)
+
+model = dict(
+    name='DeepLabV1Lane',
+    backbone_cfg=dict(
+        # MobileNetV3-Large 1.0
+        # Manually download https://download.pytorch.org/models/mobilenet_v3_large-8738ca79.pth,
+        # bug in torch 1.6
+        name='MobileNetV3Encoder',
+        pretrained='mobilenet_v3_large-8738ca79.pth',
+        arch='large',
+        reduction_factor=1,
+
+        # OS-16 (DeepLab style)
+        strides=(1, 2, 1, 2, 1, 1, 2, 1, 1, 1, 1, 1, 1, 1, 1),
+        dilations=(1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 2),
+        out_indices=(16, )
+    ),
+    classifier_cfg=dict(
+        name='DeepLabV1Head',
+        in_channels=128,
+        num_classes=7,
+        dilation=1
+    ),
+    reducer_cfg=dict(
+        name='RESAReducer',
+        in_channels=960,
+        reduce=128
+    ),
+    lane_classifier_cfg=dict(
+        name='SimpleLaneExist',
+        num_output=7 - 1,
+        flattened_size=1540,
+    )
+)

configs/lane_detection/baseline/swin-tiny_culane.py

@@ -25,7 +25,7 @@
     dist_url='tcp://localhost:12345',
     device='cuda',
     val_num_steps=0,  # Seg IoU validation (mostly useless)
-    save_dir='./checkpoints'
+    save_dir='./checkpoints',
 
     input_size=(288, 800),
     original_size=(590, 1640),
@@ -42,7 +42,7 @@
     checkpoint='./checkpoints/swin-tiny_baseline_culane/model.pt',
     # Device args
     device='cuda',
-    save_dir='./checkpoints'
+    save_dir='./checkpoints',
 
     seg=True,
     gap=20,

configs/lane_detection/common/optims/sgd009.py

@@ -0,0 +1,7 @@
+optimizer = dict(
+    name='torch_optimizer',
+    torch_optim_class='SGD',
+    lr=0.09,
+    momentum=0.9,
+    weight_decay=1e-4
+)

configs/lane_detection/resa/mobilenetv2_culane.py

@@ -0,0 +1,89 @@
+# Data pipeline
+from configs.lane_detection.common.datasets.culane_seg import dataset
+from configs.lane_detection.common.datasets.train_level0_288 import train_augmentation
+from configs.lane_detection.common.datasets.test_288 import test_augmentation
+
+# Optimization pipeline
+from configs.lane_detection.common.optims.segloss_5class import loss
+from configs.lane_detection.common.optims.sgd006 import optimizer
+from configs.lane_detection.common.optims.ep12_poly_warmup500 import lr_scheduler
+
+
+train = dict(
+    exp_name='mobilenetv2_resa_culane',
+    workers=10,
+    batch_size=20,
+    checkpoint=None,
+    # Device args
+    world_size=0,
+    dist_url='env://',
+    device='cuda',
+
+    val_num_steps=0,  # Seg IoU validation (mostly useless)
+    save_dir='./checkpoints',
+
+    input_size=(288, 800),
+    original_size=(590, 1640),
+    num_classes=5,
+    num_epochs=12,
+    collate_fn=None,  # 'dict_collate_fn' for LSTR
+    seg=True  # Seg-based method or not
+)
+
+test = dict(
+    exp_name='mobilenetv2_resa_culane',
+    workers=10,
+    batch_size=80,
+    checkpoint='./checkpoints/mobilenetv2_resa_culane/model.pt',
+    # Device args
+    device='cuda',
+
+    save_dir='./checkpoints',
+
+    seg=True,
+    gap=20,
+    ppl=18,
+    thresh=0.3,
+    collate_fn=None,  # 'dict_collate_fn' for LSTR
+    input_size=(288, 800),
+    original_size=(590, 1640),
+    max_lane=4,
+    dataset_name='culane'
+)
+
+model = dict(
+    name='RESA_Net',
+    backbone_cfg=dict(
+        name='MobileNetV2Encoder',
+        pretrained='https://download.pytorch.org/models/mobilenet_v2-b0353104.pth',
+        widen_factor=1,
+
+        # OS-16 (DeepLab style)
+        strides=(1, 2, 2, 2, 1, 1, 1),
+        dilations=(1, 1, 1, 1, 1, 2, 2),
+        out_indices=(6, )
+    ),
+    reducer_cfg=dict(
+        name='RESAReducer',
+        in_channels=320,
+        reduce=128
+    ),
+    spatial_conv_cfg=dict(
+        name='RESA',
+        num_channels=128,
+        iteration=4,
+        alpha=2.0
+    ),
+    classifier_cfg=dict(
+        name='BUSD',
+        in_channels=128,
+        num_classes=5
+    ),
+    lane_classifier_cfg=dict(
+        name='EDLaneExist',
+        num_output=5 - 1,
+        flattened_size=1125,
+        dropout=0.1,
+        pool='avg'
+    )
+)

configs/lane_detection/resa/mobilenetv2_tusimple.py

@@ -0,0 +1,89 @@
+# Data pipeline
+from configs.lane_detection.common.datasets.tusimple_seg import dataset
+from configs.lane_detection.common.datasets.train_level0_360 import train_augmentation
+from configs.lane_detection.common.datasets.test_360 import test_augmentation
+
+# Optimization pipeline
+from configs.lane_detection.common.optims.segloss_7class import loss
+from configs.lane_detection.common.optims.sgd006 import optimizer
+from configs.lane_detection.common.optims.ep50_poly_warmup200 import lr_scheduler
+
+
+train = dict(
+    exp_name='mobilenetv2_resa_tusimple',
+    workers=10,
+    batch_size=20,
+    checkpoint=None,
+    # Device args
+    world_size=0,
+    dist_url='env://',
+    device='cuda',
+
+    val_num_steps=0,  # Seg IoU validation (mostly useless)
+    save_dir='./checkpoints',
+
+    input_size=(360, 640),
+    original_size=(720, 1280),
+    num_classes=7,
+    num_epochs=50,
+    collate_fn=None,  # 'dict_collate_fn' for LSTR
+    seg=True  # Seg-based method or not
+)
+
+test = dict(
+    exp_name='mobilenetv2_resa_tusimple',
+    workers=10,
+    batch_size=80,
+    checkpoint='./checkpoints/mobilenetv2_resa_tusimple/model.pt',
+    # Device args
+    device='cuda',
+
+    save_dir='./checkpoints',
+
+    seg=True,
+    gap=10,
+    ppl=56,
+    thresh=0.3,
+    collate_fn=None,  # 'dict_collate_fn' for LSTR
+    input_size=(360, 640),
+    original_size=(720, 1280),
+    max_lane=5,
+    dataset_name='tusimple'
+)
+
+model = dict(
+    name='RESA_Net',
+    backbone_cfg=dict(
+        name='MobileNetV2Encoder',
+        pretrained='https://download.pytorch.org/models/mobilenet_v2-b0353104.pth',
+        widen_factor=1,
+
+        # OS-16 (DeepLab style)
+        strides=(1, 2, 2, 2, 1, 1, 1),
+        dilations=(1, 1, 1, 1, 1, 2, 2),
+        out_indices=(6, )
+    ),
+    reducer_cfg=dict(
+        name='RESAReducer',
+        in_channels=320,
+        reduce=128
+    ),
+    spatial_conv_cfg=dict(
+        name='RESA',
+        num_channels=128,
+        iteration=4,
+        alpha=2.0
+    ),
+    classifier_cfg=dict(
+        name='BUSD',
+        in_channels=128,
+        num_classes=7
+    ),
+    lane_classifier_cfg=dict(
+        name='EDLaneExist',
+        num_output=7 - 1,
+        flattened_size=1100,
+        dropout=0.1,
+        pool='avg'
+    )
+)