configs updated

WangLibo1995 · May 8, 2024 · 0e7ed6f · 0e7ed6f
1 parent 637bd22
commit 0e7ed6f
Show file tree

Hide file tree

Showing 11 changed files with 226 additions and 31 deletions.
diff --git a/config/loveda/dcswin.py b/config/loveda/dcswin.py
@@ -2,8 +2,8 @@
 from geoseg.losses import *
 from geoseg.datasets.loveda_dataset import *
 from geoseg.models.DCSwin import dcswin_small
-from catalyst.contrib.nn import Lookahead
-from catalyst import utils
+from tools.utils import Lookahead
+from tools.utils import process_model_params
 
 # training hparam
 max_epoch = 30
@@ -63,7 +63,7 @@
 
 # define the optimizer
 layerwise_params = {"backbone.*": dict(lr=backbone_lr, weight_decay=backbone_weight_decay)}  # 0.1xlr for backbone
-net_params = utils.process_model_params(net, layerwise_params=layerwise_params)
+net_params = process_model_params(net, layerwise_params=layerwise_params)
 base_optimizer = torch.optim.AdamW(net_params, lr=lr, weight_decay=weight_decay)
 optimizer = Lookahead(base_optimizer)
 lr_scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=max_epoch, eta_min=1e-6)

diff --git a/config/loveda/unetformer.py b/config/loveda/unetformer.py
@@ -2,8 +2,8 @@
 from geoseg.losses import *
 from geoseg.datasets.loveda_dataset import *
 from geoseg.models.UNetFormer import UNetFormer
-from catalyst.contrib.nn import Lookahead
-from catalyst import utils
+from tools.utils import Lookahead
+from tools.utils import process_model_params
 
 # training hparam
 max_epoch = 30
@@ -79,7 +79,7 @@ def train_aug(img, mask):
 
 # define the optimizer
 layerwise_params = {"backbone.*": dict(lr=backbone_lr, weight_decay=backbone_weight_decay)}
-net_params = utils.process_model_params(net, layerwise_params=layerwise_params)
+net_params = process_model_params(net, layerwise_params=layerwise_params)
 base_optimizer = torch.optim.AdamW(net_params, lr=lr, weight_decay=weight_decay)
 optimizer = Lookahead(base_optimizer)
 lr_scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=max_epoch, eta_min=1e-6)

diff --git a/config/potsdam/dcswin.py b/config/potsdam/dcswin.py
@@ -2,8 +2,8 @@
 from geoseg.losses import *
 from geoseg.datasets.potsdam_dataset import *
 from geoseg.models.DCSwin import dcswin_small
-from catalyst.contrib.nn import Lookahead
-from catalyst import utils
+from tools.utils import Lookahead
+from tools.utils import process_model_params
 from timm.scheduler.poly_lr import PolyLRScheduler
 
 # training hparam
@@ -96,7 +96,7 @@ def val_aug(img, mask):
 
 # define the optimizer
 layerwise_params = {"backbone.*": dict(lr=backbone_lr, weight_decay=backbone_weight_decay)}
-net_params = utils.process_model_params(net, layerwise_params=layerwise_params)
+net_params = process_model_params(net, layerwise_params=layerwise_params)
 base_optimizer = torch.optim.AdamW(net_params, lr=lr, weight_decay=weight_decay)
 optimizer = Lookahead(base_optimizer)
 lr_scheduler = torch.optim.lr_scheduler.CosineAnnealingWarmRestarts(optimizer, T_0=10, T_mult=2)
diff --git a/config/potsdam/ftunetformer.py b/config/potsdam/ftunetformer.py
@@ -2,8 +2,8 @@
 from geoseg.losses import *
 from geoseg.datasets.potsdam_dataset import *
 from geoseg.models.FTUNetFormer import ft_unetformer
-from catalyst.contrib.nn import Lookahead
-from catalyst import utils
+from tools.utils import Lookahead
+from tools.utils import process_model_params
 
 # training hparam
 max_epoch = 45
@@ -64,7 +64,7 @@
 
 # define the optimizer
 layerwise_params = {"backbone.*": dict(lr=backbone_lr, weight_decay=backbone_weight_decay)}
-net_params = utils.process_model_params(net, layerwise_params=layerwise_params)
+net_params = process_model_params(net, layerwise_params=layerwise_params)
 base_optimizer = torch.optim.AdamW(net_params, lr=lr, weight_decay=weight_decay)
 optimizer = Lookahead(base_optimizer)
 lr_scheduler = torch.optim.lr_scheduler.CosineAnnealingWarmRestarts(optimizer, T_0=15, T_mult=2)
diff --git a/config/potsdam/unetformer.py b/config/potsdam/unetformer.py
@@ -2,8 +2,8 @@
 from geoseg.losses import *
 from geoseg.datasets.potsdam_dataset import *
 from geoseg.models.UNetFormer import UNetFormer
-from catalyst.contrib.nn import Lookahead
-from catalyst import utils
+from tools.utils import Lookahead
+from tools.utils import process_model_params
 
 # training hparam
 max_epoch = 45
@@ -17,8 +17,6 @@
 num_classes = len(CLASSES)
 classes = CLASSES
 
-test_time_aug = 'd4'
-output_mask_dir, output_mask_rgb_dir = None, None
 weights_name = "unetformer-r18-768crop-ms-e45"
 weights_path = "model_weights/potsdam/{}".format(weights_name)
 test_weights_name = "unetformer-r18-768crop-ms-e45"
@@ -64,7 +62,7 @@
 
 # define the optimizer
 layerwise_params = {"backbone.*": dict(lr=backbone_lr, weight_decay=backbone_weight_decay)}
-net_params = utils.process_model_params(net, layerwise_params=layerwise_params)
+net_params = process_model_params(net, layerwise_params=layerwise_params)
 base_optimizer = torch.optim.AdamW(net_params, lr=lr, weight_decay=weight_decay)
 optimizer = Lookahead(base_optimizer)
 lr_scheduler = torch.optim.lr_scheduler.CosineAnnealingWarmRestarts(optimizer, T_0=15, T_mult=2)

diff --git a/config/uavid/unetformer.py b/config/uavid/unetformer.py
@@ -6,8 +6,8 @@
 from geoseg.losses import *
 from geoseg.datasets.uavid_dataset import *
 from geoseg.models.UNetFormer import UNetFormer
-from catalyst.contrib.nn import Lookahead
-from catalyst import utils
+from tools.utils import Lookahead
+from tools.utils import process_model_params
 
 # training hparam
 max_epoch = 40
@@ -66,7 +66,7 @@
 
 # define the optimizer
 layerwise_params = {"backbone.*": dict(lr=backbone_lr, weight_decay=backbone_weight_decay)}
-net_params = utils.process_model_params(net, layerwise_params=layerwise_params)
+net_params = process_model_params(net, layerwise_params=layerwise_params)
 base_optimizer = torch.optim.AdamW(net_params, lr=lr, weight_decay=weight_decay)
 optimizer = Lookahead(base_optimizer)
 lr_scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=max_epoch)

diff --git a/config/vaihingen/dcswin.py b/config/vaihingen/dcswin.py
@@ -2,8 +2,8 @@
 from geoseg.losses import *
 from geoseg.datasets.vaihingen_dataset import *
 from geoseg.models.DCSwin import dcswin_small
-from catalyst.contrib.nn import Lookahead
-from catalyst import utils
+from tools.utils import Lookahead
+from tools.utils import process_model_params
 
 # training hparam
 max_epoch = 70
@@ -96,7 +96,7 @@ def val_aug(img, mask):
 
 # define the optimizer
 layerwise_params = {"backbone.*": dict(lr=backbone_lr, weight_decay=backbone_weight_decay)}
-net_params = utils.process_model_params(net, layerwise_params=layerwise_params)
+net_params = process_model_params(net, layerwise_params=layerwise_params)
 base_optimizer = torch.optim.AdamW(net_params, lr=lr, weight_decay=weight_decay)
 optimizer = Lookahead(base_optimizer)
 lr_scheduler = torch.optim.lr_scheduler.CosineAnnealingWarmRestarts(optimizer, T_0=10, T_mult=2)

diff --git a/config/vaihingen/ftunetformer.py b/config/vaihingen/ftunetformer.py
@@ -2,8 +2,8 @@
 from geoseg.losses import *
 from geoseg.datasets.vaihingen_dataset import *
 from geoseg.models.FTUNetFormer import ft_unetformer
-from catalyst.contrib.nn import Lookahead
-from catalyst import utils
+from tools.utils import Lookahead
+from tools.utils import process_model_params
 
 # training hparam
 max_epoch = 45
@@ -64,7 +64,7 @@
 
 # define the optimizer
 layerwise_params = {"backbone.*": dict(lr=backbone_lr, weight_decay=backbone_weight_decay)}
-net_params = utils.process_model_params(net, layerwise_params=layerwise_params)
+net_params = process_model_params(net, layerwise_params=layerwise_params)
 base_optimizer = torch.optim.AdamW(net_params, lr=lr, weight_decay=weight_decay)
 optimizer = Lookahead(base_optimizer)
 lr_scheduler = torch.optim.lr_scheduler.CosineAnnealingWarmRestarts(optimizer, T_0=15, T_mult=2)

diff --git a/config/vaihingen/unetformer.py b/config/vaihingen/unetformer.py
@@ -2,8 +2,8 @@
 from geoseg.losses import *
 from geoseg.datasets.vaihingen_dataset import *
 from geoseg.models.UNetFormer import UNetFormer
-from catalyst.contrib.nn import Lookahead
-from catalyst import utils
+from tools.utils import Lookahead
+from tools.utils import process_model_params
 
 # training hparam
 max_epoch = 105
@@ -62,7 +62,7 @@
 
 # define the optimizer
 layerwise_params = {"backbone.*": dict(lr=backbone_lr, weight_decay=backbone_weight_decay)}
-net_params = utils.process_model_params(net, layerwise_params=layerwise_params)
+net_params = process_model_params(net, layerwise_params=layerwise_params)
 base_optimizer = torch.optim.AdamW(net_params, lr=lr, weight_decay=weight_decay)
 optimizer = Lookahead(base_optimizer)
 lr_scheduler = torch.optim.lr_scheduler.CosineAnnealingWarmRestarts(optimizer, T_0=15, T_mult=2)

diff --git a/requirements.txt b/requirements.txt
@@ -1,5 +1,4 @@
 timm
-catalyst==20.09
 lightning
 albumentations
 ttach
@@ -9,4 +8,6 @@ opencv-python
 scipy
 matplotlib
 einops
-addict
+addict
+ftfy
+regex