Refactor scheduler configuration

configs/_base_/schedules/cifar10_bs128.py

@@ -1,5 +1,10 @@
 # optimizer
 optimizer = dict(type='SGD', lr=0.1, momentum=0.9, weight_decay=0.0001)
 # learning policy
-lr_config = dict(policy='step', step=[100, 150])
-runner = dict(type='EpochBasedRunner', max_epochs=200)
+param_scheduler = dict(
+    type='MultiStepLR', by_epoch=True, milestones=[100, 150], gamma=0.1)
+
+# train, val, test setting
+train_cfg = dict(by_epoch=True, max_epochs=200)
+val_cfg = dict(interval=1)  # validate every epoch
+test_cfg = dict()

configs/_base_/schedules/cub_bs64.py

@@ -2,12 +2,12 @@
 optimizer = dict(
     type='SGD', lr=0.01, momentum=0.9, weight_decay=0.0005, nesterov=True)
 # learning policy
-lr_config = dict(
-    policy='CosineAnnealing',
-    min_lr=0,
-    warmup='linear',
-    warmup_iters=5,
-    warmup_ratio=0.01,
-    warmup_by_epoch=True)
+param_scheduler = [
+    dict(type='LinearLR', start_factor=0.01, by_epoch=True, begin=0, end=5),
+    dict(type='CosineAnnealingLR', T_max=95, by_epoch=True, begin=5, end=100)
+]
 
-runner = dict(type='EpochBasedRunner', max_epochs=100)
+# train, val, test setting
+train_cfg = dict(by_epoch=True, max_epochs=100)
+val_cfg = dict(interval=1)  # validate every epoch
+test_cfg = dict()

configs/_base_/schedules/imagenet_bs1024_adamw_conformer.py

@@ -16,13 +16,23 @@ optimizer = dict(
     paramwise_cfg=paramwise_cfg)
 
 # learning policy
-lr_config = dict(
-    policy='CosineAnnealing',
-    by_epoch=False,
-    min_lr_ratio=1e-2,
-    warmup='linear',
-    warmup_ratio=1e-3,
-    warmup_iters=5 * 1252,
-    warmup_by_epoch=False)
+param_scheduler = [
+    dict(
+        type='LinearLR',
+        start_factor=1e-3,
+        by_epoch=False,
+        begin=0,
+        end=5 * 1252),
+    dict(
+        type='CosineAnnealingLR',
+        T_max=295,
+        eta_min=1e-2,
+        by_epoch=True,
+        begin=5,
+        end=300)
+]
 
-runner = dict(type='EpochBasedRunner', max_epochs=300)
+# train, val, test setting
+train_cfg = dict(by_epoch=True, max_epochs=300)
+val_cfg = dict(interval=1)  # validate every epoch
+test_cfg = dict()

configs/_base_/schedules/imagenet_bs1024_adamw_swin.py

@@ -17,13 +17,18 @@ optimizer = dict(
     paramwise_cfg=paramwise_cfg)
 
 # learning policy
-lr_config = dict(
-    policy='CosineAnnealing',
-    by_epoch=False,
-    min_lr_ratio=1e-2,
-    warmup='linear',
-    warmup_ratio=1e-3,
-    warmup_iters=20,
-    warmup_by_epoch=True)
+param_scheduler = [
+    dict(type='LinearLR', start_factor=1e-3, by_epoch=False, begin=0, end=20),
+    dict(
+        type='CosineAnnealingLR',
+        T_max=280,
+        eta_min=1e-2,
+        by_epoch=True,
+        begin=10,
+        end=300)
+]
 
-runner = dict(type='EpochBasedRunner', max_epochs=300)
+# train, val, test setting
+train_cfg = dict(by_epoch=True, max_epochs=300)
+val_cfg = dict(interval=1)  # validate every epoch
+test_cfg = dict()

configs/_base_/schedules/imagenet_bs1024_coslr.py

@@ -1,11 +1,12 @@
 # optimizer
 optimizer = dict(type='SGD', lr=0.8, momentum=0.9, weight_decay=5e-5)
 # learning policy
-lr_config = dict(
-    policy='CosineAnnealing',
-    min_lr=0,
-    warmup='linear',
-    warmup_iters=5,
-    warmup_ratio=0.1,
-    warmup_by_epoch=True)
-runner = dict(type='EpochBasedRunner', max_epochs=100)
+param_scheduler = [
+    dict(type='LinearLR', start_factor=0.1, by_epoch=True, begin=0, end=5),
+    dict(type='CosineAnnealingLR', T_max=95, by_epoch=True, begin=5, end=100)
+]
+
+# train, val, test setting
+train_cfg = dict(by_epoch=True, max_epochs=300)
+val_cfg = dict(interval=1)  # validate every epoch
+test_cfg = dict()

configs/_base_/schedules/imagenet_bs1024_linearlr_bn_nowd.py

@@ -6,11 +6,21 @@ optimizer = dict(
     weight_decay=0.00004,
     paramwise_cfg=dict(norm_decay_mult=0))
 # learning policy
-lr_config = dict(
-    policy='poly',
-    min_lr=0,
-    by_epoch=False,
-    warmup='constant',
-    warmup_iters=5000,
-)
-runner = dict(type='EpochBasedRunner', max_epochs=300)
+param_scheduler = [
+    dict(type='ConstantLR', factor=0.1, by_epoch=False, begin=0, end=5 * 1252),
+    dict(type='PolyLR', eta_min=0, by_epoch=True, begin=5, end=300)
+]
+
+# old learning policy
+# lr_config = dict(
+#     policy='poly',
+#     min_lr=0,
+#     by_epoch=False,
+#     warmup='constant',
+#     warmup_iters=5000,
+# )
+
+# train, val, test setting
+train_cfg = dict(by_epoch=True, max_epochs=300)
+val_cfg = dict(interval=1)  # validate every other epoch
+test_cfg = dict()

configs/_base_/schedules/imagenet_bs2048.py

@@ -3,10 +3,14 @@ optimizer = dict(
     type='SGD', lr=0.8, momentum=0.9, weight_decay=0.0001, nesterov=True)
 
 # learning policy
-lr_config = dict(
-    policy='step',
-    warmup='linear',
-    warmup_iters=2500,
-    warmup_ratio=0.25,
-    step=[30, 60, 90])
-runner = dict(type='EpochBasedRunner', max_epochs=100)
+param_scheduler = [
+    dict(
+        type='LinearLR', start_factor=0.25, by_epoch=False, begin=0, end=2500),
+    dict(
+        type='MultiStepLR', by_epoch=True, milestones=[30, 60, 90], gamma=0.1)
+]
+
+# train, val, test setting
+train_cfg = dict(by_epoch=True, max_epochs=100)
+val_cfg = dict(interval=1)  # validate every epoch
+test_cfg = dict()

configs/_base_/schedules/imagenet_bs2048_AdamW.py

@@ -10,10 +10,30 @@ paramwise_cfg = dict(
         '.backbone.pos_embed': dict(decay_mult=0.0)
     })
 # learning policy
-lr_config = dict(
-    policy='CosineAnnealing',
-    min_lr=0,
-    warmup='linear',
-    warmup_iters=10000,
-    warmup_ratio=1e-4)
-runner = dict(type='EpochBasedRunner', max_epochs=300)
+param_scheduler = [
+    dict(
+        type='LinearLR',
+        start_factor=1e-3,
+        by_epoch=False,
+        begin=0,
+        end=10 * 626),
+    dict(
+        type='CosineAnnealingLR',
+        T_max=290,
+        eta_min=1e-2,
+        by_epoch=True,
+        begin=10,
+        end=300)
+]
+# old learning policy
+# lr_config = dict(
+#     policy='CosineAnnealing',
+#     min_lr=0,
+#     warmup='linear',
+#     warmup_iters=10000,
+#     warmup_ratio=1e-4)
+
+# train, val, test setting
+train_cfg = dict(by_epoch=True, max_epochs=300)
+val_cfg = dict(interval=1)  # validate every epoch
+test_cfg = dict()

configs/_base_/schedules/imagenet_bs2048_coslr.py

@@ -2,10 +2,24 @@
 optimizer = dict(
     type='SGD', lr=0.8, momentum=0.9, weight_decay=0.0001, nesterov=True)
 # learning policy
-lr_config = dict(
-    policy='CosineAnnealing',
-    min_lr=0,
-    warmup='linear',
-    warmup_iters=2500,
-    warmup_ratio=0.25)
-runner = dict(type='EpochBasedRunner', max_epochs=100)
+param_scheduler = [
+    dict(
+        type='LinearLR',
+        start_factor=0.25,
+        by_epoch=False,
+        begin=0,
+        end=5 * 626),
+    dict(type='CosineAnnealingLR', T_max=95, by_epoch=True, begin=5, end=100)
+]
+# old learning policy
+# lr_config = dict(
+#     policy='CosineAnnealing',
+#     min_lr=0,
+#     warmup='linear',
+#     warmup_iters=2500,
+#     warmup_ratio=0.25)
+
+# train, val, test setting
+train_cfg = dict(by_epoch=True, max_epochs=100)
+val_cfg = dict(interval=1)  # validate every epoch
+test_cfg = dict()

configs/_base_/schedules/imagenet_bs2048_rsb.py

@@ -2,11 +2,23 @@
 optimizer = dict(type='Lamb', lr=0.005, weight_decay=0.02)
 
 # learning policy
-lr_config = dict(
-    policy='CosineAnnealing',
-    min_lr=1.0e-6,
-    warmup='linear',
-    # For ImageNet-1k, 626 iters per epoch, warmup 5 epochs.
-    warmup_iters=5 * 626,
-    warmup_ratio=0.0001)
-runner = dict(type='EpochBasedRunner', max_epochs=100)
+param_scheduler = [
+    dict(
+        type='LinearLR',
+        start_factor=0.0001,
+        by_epoch=False,
+        begin=0,
+        end=5 * 626),
+    dict(
+        type='CosineAnnealingLR',
+        T_max=95,
+        eta_min=1.0e-6,
+        by_epoch=True,
+        begin=5,
+        end=100)
+]
+
+# train, val, test setting
+train_cfg = dict(by_epoch=True, max_epochs=100)
+val_cfg = dict(interval=1)  # validate every epoch
+test_cfg = dict()

configs/_base_/schedules/imagenet_bs256.py

@@ -1,5 +1,10 @@
 # optimizer
 optimizer = dict(type='SGD', lr=0.1, momentum=0.9, weight_decay=0.0001)
 # learning policy
-lr_config = dict(policy='step', step=[30, 60, 90])
-runner = dict(type='EpochBasedRunner', max_epochs=100)
+param_scheduler = dict(
+    type='MultiStepLR', by_epoch=True, milestones=[30, 60, 90], gamma=0.1)
+
+# train, val, test setting
+train_cfg = dict(by_epoch=True, max_epochs=100)
+val_cfg = dict(interval=1)  # validate every epoch
+test_cfg = dict()

configs/_base_/schedules/imagenet_bs256_140e.py

@@ -1,6 +1,10 @@
 # optimizer
 optimizer = dict(type='SGD', lr=0.1, momentum=0.9, weight_decay=0.0001)
 # learning policy
-lr_config = dict(policy='step', step=[40, 80, 120])
+param_scheduler = dict(
+    type='MultiStepLR', by_epoch=True, milestones=[40, 80, 120], gamma=0.1)
 
-runner = dict(type='EpochBasedRunner', max_epochs=140)
+# train, val, test setting
+train_cfg = dict(by_epoch=True, max_epochs=140)
+val_cfg = dict(interval=1)  # validate every epoch
+test_cfg = dict()

configs/_base_/schedules/imagenet_bs256_200e_coslr_warmup.py

@@ -1,10 +1,14 @@
 # optimizer
 optimizer = dict(type='SGD', lr=0.1, momentum=0.9, weight_decay=0.0001)
 # learning policy
-lr_config = dict(
-    policy='CosineAnnealing',
-    min_lr=0,
-    warmup='linear',
-    warmup_iters=25025,
-    warmup_ratio=0.25)
-runner = dict(type='EpochBasedRunner', max_epochs=200)
+param_scheduler = [
+    dict(
+        type='LinearLR', start_factor=0.25, by_epoch=False, begin=0,
+        end=25025),
+    dict(type='CosineAnnealingLR', T_max=195, by_epoch=True, begin=5, end=200)
+]
+
+# train, val, test setting
+train_cfg = dict(by_epoch=True, max_epochs=200)
+val_cfg = dict(interval=1)  # validate every epoch
+test_cfg = dict()

configs/_base_/schedules/imagenet_bs256_coslr.py

@@ -1,6 +1,10 @@
 # optimizer
 optimizer = dict(type='SGD', lr=0.1, momentum=0.9, weight_decay=0.0001)
 # learning policy
-lr_config = dict(policy='CosineAnnealing', min_lr=0)
+param_scheduler = dict(
+    type='CosineAnnealingLR', T_max=100, by_epoch=True, begin=0, end=100)
 
-runner = dict(type='EpochBasedRunner', max_epochs=100)
+# train, val, test setting
+train_cfg = dict(by_epoch=True, max_epochs=100)
+val_cfg = dict(interval=1)  # validate every epoch
+test_cfg = dict()

configs/_base_/schedules/imagenet_bs256_epochstep.py

@@ -1,5 +1,9 @@
 # optimizer
 optimizer = dict(type='SGD', lr=0.045, momentum=0.9, weight_decay=0.00004)
 # learning policy
-lr_config = dict(policy='step', gamma=0.98, step=1)
-runner = dict(type='EpochBasedRunner', max_epochs=300)
+param_scheduler = dict(type='StepLR', by_epoch=True, step_size=1, gamma=0.98)
+
+# train, val, test setting
+train_cfg = dict(by_epoch=True, max_epochs=300)
+val_cfg = dict(interval=1)  # validate every epoch
+test_cfg = dict()

configs/_base_/schedules/imagenet_bs4096_AdamW.py

@@ -13,11 +13,26 @@ optimizer = dict(
 )
 
 # learning policy
-lr_config = dict(
-    policy='CosineAnnealing',
-    min_lr=0,
-    warmup='linear',
-    warmup_iters=10000,
-    warmup_ratio=1e-4,
-)
-runner = dict(type='EpochBasedRunner', max_epochs=300)
+param_scheduler = [
+    dict(
+        type='LinearLR',
+        start_factor=1e-4,
+        by_epoch=False,
+        begin=0,
+        end=30 * 308),
+    dict(
+        type='CosineAnnealingLR', T_max=270, by_epoch=True, begin=30, end=300)
+]
+# old learning policy
+# lr_config = dict(
+#     policy='CosineAnnealing',
+#     min_lr=0,
+#     warmup='linear',
+#     warmup_iters=10000,
+#     warmup_ratio=1e-4,
+# )
+
+# train, val, test setting
+train_cfg = dict(by_epoch=True, max_epochs=300)
+val_cfg = dict(interval=1)  # validate every epoch
+test_cfg = dict()

configs/convmixer/convmixer-1024-20_10xb64_in1k.py

@@ -9,4 +9,4 @@ default_hooks = dict(optimizer=dict(grad_clip=dict(max_norm=5.0)))
 
 optimizer = dict(lr=0.01)
 
-runner = dict(type='EpochBasedRunner', max_epochs=150)
+train_cfg = dict(by_epoch=True, max_epochs=150)

configs/convmixer/convmixer-1536-20_10xb64_in1k.py

@@ -9,4 +9,4 @@ default_hooks = dict(optimizer=dict(grad_clip=dict(max_norm=5.0)))
 
 optimizer = dict(lr=0.01)
 
-runner = dict(type='EpochBasedRunner', max_epochs=150)
+train_cfg = dict(by_epoch=True, max_epochs=150)

configs/convmixer/convmixer-768-32_10xb64_in1k.py

@@ -9,4 +9,4 @@ default_hooks = dict(optimizer=dict(grad_clip=dict(max_norm=5.0)))
 
 optimizer = dict(lr=0.01)
 
-runner = dict(type='EpochBasedRunner', max_epochs=300)
+train_cfg = dict(by_epoch=True, max_epochs=300)

configs/densenet/README.md

@@ -6,7 +6,7 @@
 
 ## Abstract
 
-Recent work has shown that convolutional networks can be substantially deeper, more accurate, and efficient to train if they contain shorter connections between layers close to the input and those close to the output. In this paper, we embrace this observation and introduce the Dense Convolutional Network (DenseNet), which connects each layer to every other layer in a feed-forward fashion. Whereas traditional convolutional networks with L layers have L connections - one between each layer and its subsequent layer - our network has L(L+1)/2 direct connections. For each layer, the feature-maps of all preceding layers are used as inputs, and its own feature-maps are used as inputs into all subsequent layers. DenseNets have several compelling advantages: they alleviate the vanishing-gradient problem, strengthen feature propagation, encourage feature reuse, and substantially reduce the number of parameters. We evaluate our proposed architecture on four highly competitive object recognition benchmark tasks (CIFAR-10, CIFAR-100, SVHN, and ImageNet). DenseNets obtain significant improvements over the state-of-the-art on most of them, whilst requiring less computation to achieve high performance.
+Recent work has shown that convolutional networks can be substantially deeper, more accurate, and efficient to train if they contain shorter connections between layers close to the input and those close to the output. In this paper, we embrace this observation and introduce the Dense Convolutional Network (DenseNet), which connects each layer to every layer in a feed-forward fashion. Whereas traditional convolutional networks with L layers have L connections - one between each layer and its subsequent layer - our network has L(L+1)/2 direct connections. For each layer, the feature-maps of all preceding layers are used as inputs, and its own feature-maps are used as inputs into all subsequent layers. DenseNets have several compelling advantages: they alleviate the vanishing-gradient problem, strengthen feature propagation, encourage feature reuse, and substantially reduce the number of parameters. We evaluate our proposed architecture on four highly competitive object recognition benchmark tasks (CIFAR-10, CIFAR-100, SVHN, and ImageNet). DenseNets obtain significant improvements over the state-of-the-art on most of them, whilst requiring less computation to achieve high performance.
 
 <div align=center>
 <img src="https://user-images.githubusercontent.com/42952108/162675098-9a670883-b13a-4a5a-a9c9-06c39c616a0a.png" width="100%"/>

configs/densenet/densenet121_4xb256_in1k.py

@@ -7,4 +7,4 @@ _base_ = [
 
 data = dict(samples_per_gpu=256)
 
-runner = dict(type='EpochBasedRunner', max_epochs=90)
+train_cfg = dict(by_epoch=True, max_epochs=90)

configs/densenet/densenet161_4xb256_in1k.py

@@ -7,4 +7,4 @@ _base_ = [
 
 data = dict(samples_per_gpu=256)
 
-runner = dict(type='EpochBasedRunner', max_epochs=90)
+train_cfg = dict(by_epoch=True, max_epochs=90)

configs/densenet/densenet169_4xb256_in1k.py

@@ -7,4 +7,4 @@ _base_ = [
 
 data = dict(samples_per_gpu=256)
 
-runner = dict(type='EpochBasedRunner', max_epochs=90)
+train_cfg = dict(by_epoch=True, max_epochs=90)

configs/densenet/densenet201_4xb256_in1k.py

@@ -7,4 +7,4 @@ _base_ = [
 
 data = dict(samples_per_gpu=256)
 
-runner = dict(type='EpochBasedRunner', max_epochs=90)
+train_cfg = dict(by_epoch=True, max_epochs=90)

configs/lenet/lenet5_mnist.py

@@ -38,7 +38,8 @@ evaluation = dict(
 optimizer = dict(type='SGD', lr=0.01, momentum=0.9, weight_decay=0.0001)
 optimizer_config = dict(grad_clip=None)
 # learning policy
-lr_config = dict(policy='step', step=[15])
+param_scheduler = dict(
+    type='MultiStepLR', by_epoch=True, milestones=[15], gamma=0.1)
 # checkpoint saving
 checkpoint_config = dict(interval=1)
 # yapf:disable
@@ -49,8 +50,13 @@ log_config = dict(
         # dict(type='TensorboardLoggerHook')
     ])
 # yapf:enable
+
+# train, val, test setting
+train_cfg = dict(by_epoch=True, max_epochs=5)
+val_cfg = dict(interval=1)  # validate every epoch
+test_cfg = dict()
+
 # runtime settings
-runner = dict(type='EpochBasedRunner', max_epochs=5)
 dist_params = dict(backend='nccl')
 log_level = 'INFO'
 work_dir = './work_dirs/mnist/'

configs/mobilenet_v3/mobilenet-v3-large_8xb32_in1k.py

@@ -154,5 +154,8 @@ optimizer = dict(
     weight_decay=1e-5)
 optimizer_config = dict(grad_clip=None)
 # learning policy
-lr_config = dict(policy='step', step=2, gamma=0.973, by_epoch=True)
-runner = dict(type='EpochBasedRunner', max_epochs=600)
+param_scheduler = dict(type='StepLR', by_epoch=True, step_size=2, gamma=0.973)
+# train, val, test setting
+train_cfg = dict(by_epoch=True, max_epochs=600)
+val_cfg = dict(interval=1)  # validate every epoch
+test_cfg = dict()

configs/mobilenet_v3/mobilenet-v3-small_8xb16_cifar10.py

@@ -4,5 +4,6 @@ _base_ = [
     '../_base_/schedules/cifar10_bs128.py', '../_base_/default_runtime.py'
 ]
 
-lr_config = dict(policy='step', step=[120, 170])
-runner = dict(type='EpochBasedRunner', max_epochs=200)
+param_scheduler = dict(
+    type='MultiStepLR', by_epoch=True, milestones=[120, 170], gamma=0.1)
+train_cfg = dict(by_epoch=True, max_epochs=200)

configs/mobilenet_v3/mobilenet-v3-small_8xb32_in1k.py

@@ -154,5 +154,8 @@ optimizer = dict(
     weight_decay=1e-5)
 optimizer_config = dict(grad_clip=None)
 # learning policy
-lr_config = dict(policy='step', step=2, gamma=0.973, by_epoch=True)
-runner = dict(type='EpochBasedRunner', max_epochs=600)
+param_scheduler = dict(type='StepLR', by_epoch=True, step_size=2, gamma=0.973)
+# train, val, test setting
+train_cfg = dict(by_epoch=True, max_epochs=600)
+val_cfg = dict(interval=1)  # validate every epoch
+test_cfg = dict()

configs/resnest/resnest101_32xb64_in1k.py

@@ -171,11 +171,11 @@ optimizer = dict(
 optimizer_config = dict(grad_clip=None)
 
 # learning policy
-lr_config = dict(
-    policy='CosineAnnealing',
-    min_lr=0,
-    warmup='linear',
-    warmup_iters=5,
-    warmup_ratio=1e-6,
-    warmup_by_epoch=True)
-runner = dict(type='EpochBasedRunner', max_epochs=270)
+param_scheduler = [
+    dict(type='LinearLR', start_factor=1e-6, by_epoch=True, begin=0, end=5),
+    dict(type='CosineAnnealingLR', T_max=265, by_epoch=True, begin=5, end=270)
+]
+# train, val, test setting
+train_cfg = dict(by_epoch=True, max_epochs=270)
+val_cfg = dict(interval=1)  # validate every epoch
+test_cfg = dict()

configs/resnest/resnest200_64xb32_in1k.py

@@ -171,11 +171,12 @@ optimizer = dict(
 optimizer_config = dict(grad_clip=None)
 
 # learning policy
-lr_config = dict(
-    policy='CosineAnnealing',
-    min_lr=0,
-    warmup='linear',
-    warmup_iters=5,
-    warmup_ratio=1e-6,
-    warmup_by_epoch=True)
-runner = dict(type='EpochBasedRunner', max_epochs=270)
+param_scheduler = [
+    dict(type='LinearLR', start_factor=1e-6, by_epoch=True, begin=0, end=5),
+    dict(type='CosineAnnealingLR', T_max=265, by_epoch=True, begin=5, end=270)
+]
+
+# train, val, test setting
+train_cfg = dict(by_epoch=True, max_epochs=270)
+val_cfg = dict(interval=1)  # validate every epoch
+test_cfg = dict()

configs/resnest/resnest269_64xb32_in1k.py

@@ -171,11 +171,11 @@ optimizer = dict(
 optimizer_config = dict(grad_clip=None)
 
 # learning policy
-lr_config = dict(
-    policy='CosineAnnealing',
-    min_lr=0,
-    warmup='linear',
-    warmup_iters=5,
-    warmup_ratio=1e-6,
-    warmup_by_epoch=True)
-runner = dict(type='EpochBasedRunner', max_epochs=270)
+param_scheduler = [
+    dict(type='LinearLR', start_factor=1e-6, by_epoch=True, begin=0, end=5),
+    dict(type='CosineAnnealingLR', T_max=265, by_epoch=True, begin=5, end=270)
+]
+# train, val, test setting
+train_cfg = dict(by_epoch=True, max_epochs=270)
+val_cfg = dict(interval=1)  # validate every epoch
+test_cfg = dict()

configs/resnest/resnest50_32xb64_in1k.py

@@ -171,11 +171,11 @@ optimizer = dict(
 optimizer_config = dict(grad_clip=None)
 
 # learning policy
-lr_config = dict(
-    policy='CosineAnnealing',
-    min_lr=0,
-    warmup='linear',
-    warmup_iters=5,
-    warmup_ratio=1e-6,
-    warmup_by_epoch=True)
-runner = dict(type='EpochBasedRunner', max_epochs=270)
+param_scheduler = [
+    dict(type='LinearLR', start_factor=1e-6, by_epoch=True, begin=0, end=5),
+    dict(type='CosineAnnealingLR', T_max=265, by_epoch=True, begin=5, end=270)
+]
+# train, val, test setting
+train_cfg = dict(by_epoch=True, max_epochs=270)
+val_cfg = dict(interval=1)  # validate every epoch
+test_cfg = dict()

configs/resnet/resnet50_8xb128_coslr-90e_in21k.py

@@ -8,4 +8,4 @@ _base_ = [
 model = dict(head=dict(num_classes=21843))
 
 # runtime settings
-runner = dict(type='EpochBasedRunner', max_epochs=90)
+train_cfg = dict(by_epoch=True, max_epochs=90)

configs/resnet/resnet50_8xb16_cifar100.py

@@ -7,4 +7,5 @@ _base_ = [
 model = dict(head=dict(num_classes=100))
 
 optimizer = dict(type='SGD', lr=0.1, momentum=0.9, weight_decay=0.0005)
-lr_config = dict(policy='step', step=[60, 120, 160], gamma=0.2)
+param_scheduler = dict(
+    type='MultiStepLR', by_epoch=False, milestones=[60, 120, 160], gamma=0.2)

configs/t2t_vit/t2t-vit-t-14_8xb64_in1k.py

@@ -19,17 +19,20 @@ optimizer = dict(
 optimizer_config = dict(grad_clip=None)
 
 # learning policy
-# FIXME: lr in the first 300 epochs conforms to the CosineAnnealing and
-# the lr in the last 10 epoch equals to min_lr
-lr_config = dict(
-    policy='CosineAnnealingCooldown',
-    min_lr=1e-5,
-    cool_down_time=10,
-    cool_down_ratio=0.1,
-    by_epoch=True,
-    warmup_by_epoch=True,
-    warmup='linear',
-    warmup_iters=10,
-    warmup_ratio=1e-6)
+param_scheduler = [
+    dict(type='LinearLR', start_factor=1e-6, by_epoch=True, begin=0, end=10),
+    dict(
+        type='CosineAnnealingLR',
+        T_max=290,
+        eta_min=1e-5,
+        by_epoch=True,
+        begin=10,
+        end=300),
+    dict(type='ConstantLR', factor=0.1, by_epoch=True, begin=300, end=310),
+]
 custom_hooks = [dict(type='EMAHook', momentum=4e-5, priority='ABOVE_NORMAL')]
-runner = dict(type='EpochBasedRunner', max_epochs=310)
+
+# train, val, test setting
+train_cfg = dict(by_epoch=True, max_epochs=310)
+val_cfg = dict(interval=1)  # validate every epoch
+test_cfg = dict()

configs/t2t_vit/t2t-vit-t-19_8xb64_in1k.py

@@ -19,17 +19,20 @@ optimizer = dict(
 optimizer_config = dict(grad_clip=None)
 
 # learning policy
-# FIXME: lr in the first 300 epochs conforms to the CosineAnnealing and
-# the lr in the last 10 epoch equals to min_lr
-lr_config = dict(
-    policy='CosineAnnealingCooldown',
-    min_lr=1e-5,
-    cool_down_time=10,
-    cool_down_ratio=0.1,
-    by_epoch=True,
-    warmup_by_epoch=True,
-    warmup='linear',
-    warmup_iters=10,
-    warmup_ratio=1e-6)
+param_scheduler = [
+    dict(type='LinearLR', start_factor=1e-6, by_epoch=True, begin=0, end=10),
+    dict(
+        type='CosineAnnealingLR',
+        T_max=290,
+        eta_min=1e-5,
+        by_epoch=True,
+        begin=10,
+        end=300),
+    dict(type='ConstantLR', factor=0.1, by_epoch=True, begin=300, end=310),
+]
 custom_hooks = [dict(type='EMAHook', momentum=4e-5, priority='ABOVE_NORMAL')]
-runner = dict(type='EpochBasedRunner', max_epochs=310)
+
+# train, val, test setting
+train_cfg = dict(by_epoch=True, max_epochs=310)
+val_cfg = dict(interval=1)  # validate every epoch
+test_cfg = dict()

configs/t2t_vit/t2t-vit-t-24_8xb64_in1k.py

@@ -19,17 +19,20 @@ optimizer = dict(
 optimizer_config = dict(grad_clip=None)
 
 # learning policy
-# FIXME: lr in the first 300 epochs conforms to the CosineAnnealing and
-# the lr in the last 10 epoch equals to min_lr
-lr_config = dict(
-    policy='CosineAnnealingCooldown',
-    min_lr=1e-5,
-    cool_down_time=10,
-    cool_down_ratio=0.1,
-    by_epoch=True,
-    warmup_by_epoch=True,
-    warmup='linear',
-    warmup_iters=10,
-    warmup_ratio=1e-6)
+param_scheduler = [
+    dict(type='LinearLR', start_factor=1e-6, by_epoch=True, begin=0, end=10),
+    dict(
+        type='CosineAnnealingLR',
+        T_max=290,
+        eta_min=1e-5,
+        by_epoch=True,
+        begin=10,
+        end=300),
+    dict(type='ConstantLR', factor=0.1, by_epoch=True, begin=300, end=310),
+]
+
 custom_hooks = [dict(type='EMAHook', momentum=4e-5, priority='ABOVE_NORMAL')]
-runner = dict(type='EpochBasedRunner', max_epochs=310)
+# train, val, test setting
+train_cfg = dict(by_epoch=True, max_epochs=310)
+val_cfg = dict(interval=1)  # validate every epoch
+test_cfg = dict()

configs/tnt/tnt-s-p16_16xb64_in1k.py

@@ -30,11 +30,13 @@ data = dict(
 optimizer = dict(type='AdamW', lr=1e-3, weight_decay=0.05)
 optimizer_config = dict(grad_clip=None)
 
-lr_config = dict(
-    policy='CosineAnnealing',
-    min_lr=0,
-    warmup_by_epoch=True,
-    warmup='linear',
-    warmup_iters=5,
-    warmup_ratio=1e-3)
-runner = dict(type='EpochBasedRunner', max_epochs=300)
+# learning policy
+param_scheduler = [
+    dict(type='LinearLR', start_factor=1e-3, by_epoch=True, begin=0, end=5),
+    dict(type='CosineAnnealingLR', T_max=295, by_epoch=True, begin=5, end=300)
+]
+
+# train, val, test setting
+train_cfg = dict(by_epoch=True, max_epochs=300)
+val_cfg = dict(interval=1)  # validate every epoch
+test_cfg = dict()

configs/twins/twins-pcpvt-base_8xb128_in1k.py

@@ -23,13 +23,15 @@ optimizer = dict(
     paramwise_cfg=paramwise_cfg)
 
 # learning policy
-lr_config = dict(
-    policy='CosineAnnealing',
-    by_epoch=True,
-    min_lr_ratio=1e-2,
-    warmup='linear',
-    warmup_ratio=1e-3,
-    warmup_iters=5,
-    warmup_by_epoch=True)
+param_scheduler = [
+    dict(type='LinearLR', start_factor=1e-3, by_epoch=True, begin=0, end=5),
+    dict(
+        type='CosineAnnealingLR',
+        T_max=295,
+        eta_min=1e-2,
+        by_epoch=True,
+        begin=5,
+        end=300)
+]
 
 evaluation = dict(interval=1, metric='accuracy')

configs/twins/twins-svt-base_8xb128_in1k.py

@@ -22,13 +22,16 @@ optimizer = dict(
     betas=(0.9, 0.999),
     paramwise_cfg=paramwise_cfg)
 
-lr_config = dict(
-    policy='CosineAnnealing',
-    by_epoch=True,
-    min_lr_ratio=1e-2,
-    warmup='linear',
-    warmup_ratio=1e-3,
-    warmup_iters=5,
-    warmup_by_epoch=True)
+# learning policy
+param_scheduler = [
+    dict(type='LinearLR', start_factor=1e-3, by_epoch=True, begin=0, end=5),
+    dict(
+        type='CosineAnnealingLR',
+        T_max=295,
+        eta_min=1e-2,
+        by_epoch=True,
+        begin=5,
+        end=300)
+]
 
 evaluation = dict(interval=1, metric='accuracy')

configs/vgg/vgg16_8xb16_voc.py

@@ -21,6 +21,9 @@ optimizer = dict(
     paramwise_cfg=dict(custom_keys={'.backbone.classifier': dict(lr_mult=10)}))
 optimizer_config = dict(grad_clip=None)
 # learning policy
-lr_config = dict(policy='step', step=20, gamma=0.1)
+param_scheduler = dict(type='StepLR', by_epoch=True, step_size=20, gamma=0.1)
 
-runner = dict(type='EpochBasedRunner', max_epochs=40)
+# train, val, test setting
+train_cfg = dict(by_epoch=True, max_epochs=40)
+val_cfg = dict(interval=1)  # validate every epoch
+test_cfg = dict()

configs/vision_transformer/vit-base-p16_ft-4xb544-ipu_in1k.py

@@ -72,13 +72,15 @@ optimizer = dict(
 )
 
 # learning policy
-lr_config = dict(
-    policy='CosineAnnealing',
-    min_lr=0,
-    warmup='linear',
-    warmup_iters=800,
-    warmup_ratio=0.02,
-)
+param_scheduler = [
+    dict(type='LinearLR', start_factor=0.02, by_epoch=False, begin=0, end=800),
+    dict(
+        type='CosineAnnealingLR',
+        T_max=4200,
+        by_epoch=False,
+        begin=800,
+        end=5000)
+]
 
 # ipu cfg
 # model partition config