Closed initze closed 2 months ago
The new datacube version training has bad performance
Training looks alright, though not great IoU and F1 = 0
# Model Specification model: # Model Architecture. Available: # Unet, UnetPlusPlus, Unet3Plus, MAnet, Linknet, FPN, PSPNet, DeepLabV3, DeepLabV3Plus, PAN] architecture: UnetPlusPlus # Model Encoder. Examples: # resnet18, resnet34, resnet50, resnet101, resnet152 # Check https://github.com/qubvel/segmentation_models.pytorch#encoders for the full list of available encoders encoder: resnet34 # Encoder weights to use (if transfer learning is desired) # `imagenet` is available for all encoders, some of them have more options available # `random` initializes the weights randomly # Check https://github.com/qubvel/segmentation_models.pytorch#encoders for the # full list of weights available for each encoder encoder_weights: random # Loss Function to use. Available: # JaccardLoss, DiceLoss, FocalLoss, LovaszLoss, SoftBCEWithLogitsLoss loss_function: SoftBCEWithLogitsLoss # Data Configuration data_threads: 8 # Number of threads for data loading, must be 0 on Windows tile_size: 1024 # tile size in pixels # Sampling mode to Use. Available: # deterministic, random, targets_only sampling_mode: deterministic data_sources: # Enabled input features - PlanetScope - TCVIS #- RelativeElevation - AbsoluteElevation - Slope datasets: train: augment: true augment_types: - HorizontalFlip - VerticalFlip - Blur - RandomRotate90 - RandomBrightnessContrast - MultiplicativeNoise - Cutout shuffle: true scenes: - 4688749_0470419_2021-07-12_105c - 4691214_0470419_2021-07-14_2455 - 4704585_1170810_2021-07-18_2427 - 4705397_1170810_2021-07-18_2423 - 5811537_0870113_2022-07-25_2276 - 4691214_0470419_2021-07-14_2455 - 4705397_1170810_2021-07-18_2423 - 5811537_0870113_2022-07-25_2276 - 4713120_5272315_2021-07-21_2463 - 4770635_4572716_2021-08-06_2416 - 3649691_4572917_2020-08-14_1039 - 4706735_4472312_2021-07-19_1034 - 4709512_4372018_2021-07-20_2427 - 4751812_3973218_2021-08-01_1026 - 4694518_0870514_2021-07-15_1064 - '20200722_081437_1032' - '20200722_081438_1032' - 20180702_025359_0f31 - 20180702_025400_0f31 - 20180702_025401_0f31 - '20180703_192559_1006' - '20180703_192600_1006' - '20180704_075856_1044' - '20180704_075857_1044' - '20180708_075756_1011' - '20180708_075757_1011' - 20180710_025407_101b - 20180710_025408_101b - 20180716_025342_0f35 - 20180716_025343_0f35 - 20180716_025344_0f35 - 20180722_025446_102e - 20180729_193040_0f43 - 20180729_193041_0f43 - '20180730_025452_1012' - '20180801_075816_1021' - '20180801_075817_1021' - 20180825_164603_0f2a - 20180825_164604_0f2a - '20180829_025607_1034' - '20180829_025608_1034' - '20180829_025609_1034' - 20180921_203252_101f - 20180923_080226_0f34 - 20190606_030140_0f3f - 20190606_030141_0f3f - 20190607_001259_0f44 - 20190607_001300_0f44 - '20190607_204203_1044' - '20190607_204204_1044' - '20190607_204205_1044' - 20190610_201139_0f15 - 20190610_201140_0f15 - '20190618_013705_23_1062' - '20190618_013707_30_1062' - 20190619_203634_0f4e - 20190619_203635_0f4e - 20190620_201330_100a - 20190620_201331_100a - 20190620_201332_100a - 20190621_204122_0f12 - 20190621_204124_0f12 - '20190622_164800_1048' - '20190622_164801_1048' - '20190622_164802_1048' - '20190628_030539_1027' - '20190628_030541_1027' - '20190628_030542_1027' - 20190628_064356_41_106a - 20190628_064358_48_106a - 20190702_050924_0f49 - 20190702_050925_0f49 - '20190702_064217_29_1065' - '20190703_201901_1006' - '20190703_201902_1006' - '20190704_172438_93_1067' - '20190704_172440_99_1067' - 20190708_201059_101b - 20190708_201100_101b - '20190713_184804_64_1067' - '20190713_184806_71_1067' - '20190714_194358_1035' - '20190714_194359_1035' - '20190714_194400_1035' - 20190715_000856_0f2a - 20190715_000857_0f2a - 20190716_014045_22_106a - '20190719_030812_1006' - '20190719_030813_1006' - '20190719_030814_1006' - '20190722_030702_1039' - '20190722_030703_1039' - 20190723_080935_103b - 20190723_080936_103b - 20190723_080937_103b - 20190723_080938_103b - 20190726_193719_0f17 - 20190726_193720_0f17 - 20190726_193721_0f17 - '20190727_030648_1040' - '20190727_030649_1040' - '20190727_030650_1040' - 20190727_194259_100c - 20190727_194300_100c - '20190727_202025_1034' - '20190727_202026_1034' - '20190731_000801_1053' - '20190731_000802_1_1053' - '20190801_080518_1014' - '20190801_080519_1014' - 20190803_000717_0f49 - 20190803_000718_0f49 - '20190803_164435_1048' - '20190803_164436_1048' - '20190803_164437_1048' - '20190803_164438_1048' - '20190803_164439_1048' - '20190808_201538_1018' - '20190808_201539_1018' - 20190813_082915_68_105c - 20190813_082918_17_105c - 20190814_080303_100c - 20190814_080304_100c - '20190815_030625_1001' - '20190815_030626_1001' - 20190815_080515_0f4e - 20190815_080516_0f4e - '20190819_080913_1012' - '20190819_080914_1012' - '20190819_080915_1012' - '20190819_080917_1012' - 20190820_164142_0f21 - 20190820_164143_0f21 - 20190820_164144_0f21 - 20190822_081235_101f - 20190822_081236_101f - 20190822_081237_101f - '20190829_030342_1006' - '20190829_030343_1006' - 20190831_030654_0f35 - 20190831_030655_0f35 - 20190902_203942_103b - 20190902_203943_103b - 20190903_075606_0e0f - 20190903_075607_0e0f - 20190904_193046_0e2f - 20190904_193047_0e2f - 20190904_193048_0e2f - '20190905_081012_1004' - '20190905_081013_1004' - '20190905_081014_1004' - '20190905_195022_1032' - '20190905_195023_1032' - 20190907_081306_100c - 20190907_081307_100c - 20190907_081308_100c - 20190907_081309_100c - '20190907_195118_1034' - '20190907_195119_1034' - 20190907_203545_84_105c - '20190913_030436_1005' - '20190913_030437_1005' - '20190916_030410_1010' - '20190916_030411_1010' - '20190921_080320_1011' - '20190921_080321_1011' - '20190928_201420_1035' - '20190928_201421_1035' - '20190928_201422_1035' - 5773266_0971513_2022-07-08_2439 - 5761052_1070519_2022-07-04_241f - 5796176_4372315_2022-07-19_242d - 4737567_5268618_2021-07-28_2465 - 5811537_0870514_2022-07-25_2276 - 5773288_1071219_2022-07-09_248b - 5853699_4973918_2022-08-13_245c - 5851529_0870615_2022-08-12_2403 - 5761052_1171112_2022-07-04_241f - 5828620_4271614_2022-08-02_2495 - 4740689_5173016_2021-07-29_242a - 5803306_3971013_2022-07-22_248c val: augment: false shuffle: false scenes: - '20200814_044131_1026' - '20200814_044132_1026' - '20200814_044133_1026' - 5761052_1171112_2022-07-04_241f test: augment: false shuffle: false scenes: null # Training Parameters batch_size: 18 learning_rate: 0.001 # Learning rate scheduler. Available: # ExponentialLR, StepLR (https://pytorch.org/docs/stable/optim.html) # if no lr_step_size given then lr_step_size=10, gamma=0.1 for StepLR and gamma=0.9 for ExponentialLR learning_rate_scheduler: ExponentialLR lr_step_size: 10 lr_gamma: 0.9 # Training Schedule schedule: - phase: Training epochs: 50 steps: - train_on: train - validate_on: val
Seems to be better now.
deprecated
The new datacube version training has bad performance
Training looks alright, though not great IoU and F1 = 0
Config