Update config

configs/recognition/tsm/README.md

@@ -7,7 +7,7 @@
 |config | pretrain | top1 acc| top5 acc | gpu_mem(M) | iter time(s) | ckpt | log|
 |-|-|-|-|-|-|-|-|
 |[tsm_r50_1x1x8_100e_kinetics400_rgb](/configs/recognition/tsm/tsm_r50_1x1x8_50e_kinetics400_rgb.py) | ResNet50 |70.24|89.56| x | x | [ckpt]() | [log]()|
-|[tsm_r50_dense_1x1x8_100e_kinetics400_rgb](/configs/recognition/tsm/tsm_r50_dense_1x1x8_50e_kinetics400_rgb.py) | ResNet50 |71.84|90.18| x | x | [ckpt]() | [log]()|
+|[tsm_r50_dense_1x1x8_100e_kinetics400_rgb](/configs/recognition/tsm/tsm_r50_dense_1x1x8_100e_kinetics400_rgb.py) | ResNet50 |71.84|90.18| x | x | [ckpt]() | [log]()|
 |[tsm_r50_1x1x16_50e_kinetics400_rgb](/configs/recognition/tsm/tsm_r50_1x1x16_50e_kinetics400_rgb.py) | ResNet50 |71.69|90.4| x | x | [ckpt]() | [log]()|
 |[tsm_r50_video_1x1x8_100e_kinetics400_rgb](/configs/recognition/tsm/tsm_r50_video_1x1x8_100e_kinetics400_rgb.py) | ResNet50 | x | x | x | x | [ckpt]() | [log]()|
 

configs/recognition/tsm/tsm_r101_1x1x8_50e_sthv2_rgb.py

@@ -105,7 +105,7 @@ optimizer = dict(
     type='SGD',
     constructor='TSMOptimizerConstructor',
     paramwise_cfg=dict(fc_lr5=True),
-    lr=0.01,
+    lr=0.02,
     momentum=0.9,
     weight_decay=0.0005)
 optimizer_config = dict(grad_clip=dict(max_norm=20, norm_type=2))

configs/recognition/tsm/tsm_r50_1x1x8_50e_sthv1_rgb.py

@@ -105,7 +105,7 @@ optimizer = dict(
     type='SGD',
     constructor='TSMOptimizerConstructor',
     paramwise_cfg=dict(fc_lr5=True),
-    lr=0.005,
+    lr=0.01,
     momentum=0.9,
     weight_decay=0.0005)
 optimizer_config = dict(grad_clip=dict(max_norm=20, norm_type=2))

configs/recognition/tsm/tsm_r50_dense_1x1x8_50e_kinetics400_rgb.py → configs/recognition/tsm/tsm_r50_dense_1x1x8_100e_kinetics400_rgb.py

@@ -108,11 +108,11 @@ optimizer = dict(
     paramwise_cfg=dict(fc_lr5=True),
     lr=0.02,
     momentum=0.9,
-    weight_decay=0.0005)
+    weight_decay=0.0001)
 optimizer_config = dict(grad_clip=dict(max_norm=20, norm_type=2))
 # learning policy
-lr_config = dict(policy='step', step=[20, 40])
-total_epochs = 50
+lr_config = dict(policy='step', step=[40, 80])
+total_epochs = 100
 checkpoint_config = dict(interval=1)
 evaluation = dict(
     interval=2, metrics=['top_k_accuracy', 'mean_class_accuracy'], topk=(1, 5))
@@ -125,7 +125,7 @@ log_config = dict(
 # runtime settings
 dist_params = dict(backend='nccl')
 log_level = 'INFO'
-work_dir = './work_dirs/tsm_r50_dense_1x1x8_50e_kinetics400_rgb/'
+work_dir = './work_dirs/tsm_r50_dense_1x1x8_100e_kinetics400_rgb/'
 load_from = None
 resume_from = None
 workflow = [('train', 1)]

configs/recognition/tsn/README.md

@@ -6,7 +6,7 @@
 
 |config | pretrain | top1 acc| top5 acc | gpu_mem(M) | iter time(s) | ckpt | log|
 |-|-|-|-|-|-|-|-|
-|[tsn_r50_1x1x3_100e_ucf101_rgb](/configs/recognition/tsn/tsn_r50_1x1x3_100e_ucf101_rgb.py) | ResNet50 |80.12|96.09| x | x | [ckpt]() | [log]()|
+|[tsn_r50_1x1x3_100e_ucf101_rgb](/configs/recognition/tsn/tsn_r50_1x1x3_80e_ucf101_rgb.py) | ResNet50 |80.12|96.09| x | x | [ckpt]() | [log]()|
 
 ### Kinetics-400
 
@@ -14,7 +14,7 @@
 |-|-|-|-|-|-|-|-|
 |[tsn_r50_1x1x3_100e_kinetics400_rgb](/configs/recognition/tsn/tsn_r50_1x1x3_100e_kinetics400_rgb.py) | ResNet50 |70.60|89.26| x | x | [ckpt]() | [log]()|
 |[tsn_r50_1x1x5_50e_kinetics400_rgb](/configs/recognition/tsn/tsn_r50_1x1x5_50e_kinetics400_rgb.py) | ResNet50 |68.64|88.19| x | x | [ckpt]() | [log]()|
-|[tsn_r50_dense_1x1x5_50e_kinetics400_rgb](/configs/recognition/tsn/tsn_r50_dense_1x1x5_50e_kinetics400_rgb.py) | ResNet50 |68.59|88.31| x | x | [ckpt]() | [log]()|
+|[tsn_r50_dense_1x1x5_50e_kinetics400_rgb](/configs/recognition/tsn/tsn_r50_dense_1x1x5_100e_kinetics400_rgb.py) | ResNet50 |68.59|88.31| x | x | [ckpt]() | [log]()|
 |[tsn_r50_1x1x8_100e_kinetics400_rgb](/configs/recognition/tsn/tsn_r50_1x1x8_100e_kinetics400_rgb.py) | ResNet50 |69.41|88.37| x | x | [ckpt]() | [log]()|
 |[tsn_r50_320p_1x1x3_100e_kinetics400_rgb](/configs/recognition/tsn/tsn_r50_320p_1x1x3_100e_kinetics400_rgb.py) | ResNet50 |70.91|89.51| x | x | [ckpt]() | [log]() |
 |[tsn_r50_320p_1x1x3_110e_kinetics400_flow](/configs/recognition/tsn/tsn_r50_320p_1x1x3_110e_kinetics400_flow.py) | ResNet50 |55.70|79.85| x | x | [ckpt]() | [log]() |

configs/recognition/tsn/tsn_r50_1x1x16_50e_sthv1_rgb.py

@@ -43,6 +43,7 @@ train_pipeline = [
         max_wh_scale_gap=1,
         num_fixed_crops=13),
     dict(type='Resize', scale=(224, 224), keep_ratio=False),
+    dict(type='Flip', flip_ratio=0.5),
     dict(type='Normalize', **img_norm_cfg),
     dict(type='FormatShape', input_format='NCHW'),
     dict(type='Collect', keys=['imgs', 'label'], meta_keys=[]),

configs/recognition/tsn/tsn_r50_1x1x16_50e_sthv2_rgb.py

@@ -41,6 +41,7 @@ train_pipeline = [
         random_crop=False,
         max_wh_scale_gap=1),
     dict(type='Resize', scale=(224, 224), keep_ratio=False),
+    dict(type='Flip', flip_ratio=0.5),
     dict(type='Normalize', **img_norm_cfg),
     dict(type='FormatShape', input_format='NCHW'),
     dict(type='Collect', keys=['imgs', 'label'], meta_keys=[]),

configs/recognition/tsn/tsn_r50_1x1x3_100e_ucf101_rgb.py → configs/recognition/tsn/tsn_r50_1x1x3_80e_ucf101_rgb.py

@@ -114,7 +114,7 @@ log_config = dict(
 # runtime settings
 dist_params = dict(backend='nccl')
 log_level = 'INFO'
-work_dir = './work_dirs/tsn_r50_1x1x3_100e_ucf101_rgb/'
+work_dir = './work_dirs/tsn_r50_1x1x3_80e_ucf101_rgb/'
 load_from = None
 resume_from = None
 workflow = [('train', 1)]

configs/recognition/tsn/tsn_r50_1x1x8_50e_sthv1_rgb.py

@@ -43,7 +43,6 @@ train_pipeline = [
         max_wh_scale_gap=1,
         num_fixed_crops=13),
     dict(type='Resize', scale=(224, 224), keep_ratio=False),
-    dict(type='Flip', flip_ratio=0.5),
     dict(type='Normalize', **img_norm_cfg),
     dict(type='FormatShape', input_format='NCHW'),
     dict(type='Collect', keys=['imgs', 'label'], meta_keys=[]),
@@ -82,7 +81,7 @@ test_pipeline = [
     dict(type='ToTensor', keys=['imgs'])
 ]
 data = dict(
-    videos_per_gpu=16,
+    videos_per_gpu=8,
     workers_per_gpu=4,
     train=dict(
         type=dataset_type,
@@ -100,7 +99,7 @@ data = dict(
         data_prefix=data_root_val,
         pipeline=test_pipeline))
 # optimizer
-optimizer = dict(type='SGD', lr=0.02, momentum=0.9, weight_decay=0.0001)
+optimizer = dict(type='SGD', lr=0.01, momentum=0.9, weight_decay=0.0005)
 optimizer_config = dict(grad_clip=dict(max_norm=20, norm_type=2))
 # learning policy
 lr_config = dict(policy='step', step=[20, 40])

configs/recognition/tsn/tsn_r50_dense_1x1x5_50e_kinetics400_rgb.py → configs/recognition/tsn/tsn_r50_dense_1x1x5_100e_kinetics400_rgb.py

@@ -83,7 +83,7 @@ test_pipeline = [
 data = dict(
     videos_per_gpu=16,
     workers_per_gpu=4,
-    val_dataloader=dict(videos_per_gpu=8),
+    val_dataloader=dict(videos_per_gpu=4),
     train=dict(
         type=dataset_type,
         ann_file=ann_file_train,
@@ -100,11 +100,11 @@ data = dict(
         data_prefix=data_root_val,
         pipeline=test_pipeline))
 # optimizer
-optimizer = dict(type='SGD', lr=0.02, momentum=0.9, weight_decay=0.0005)
+optimizer = dict(type='SGD', lr=0.01, momentum=0.9, weight_decay=0.0001)
 optimizer_config = dict(grad_clip=dict(max_norm=20, norm_type=2))
 # learning policy
-lr_config = dict(policy='step', step=[20, 40])
-total_epochs = 50
+lr_config = dict(policy='step', step=[40, 80])
+total_epochs = 100
 checkpoint_config = dict(interval=1)
 evaluation = dict(
     interval=2, metrics=['top_k_accuracy', 'mean_class_accuracy'], topk=(1, 5))
@@ -117,7 +117,7 @@ log_config = dict(
 # runtime settings
 dist_params = dict(backend='nccl')
 log_level = 'INFO'
-work_dir = './work_dirs/tsn_r50_dense_1x1x5_50e_kinetics400_rgb/'
+work_dir = './work_dirs/tsn_r50_dense_1x1x5_100e_kinetics400_rgb/'
 load_from = None
 resume_from = None
 workflow = [('train', 1)]

configs/recognition/tsn/tsn_r50_dense_1x1x8_100e_kinetics400_rgb.py

@@ -99,7 +99,7 @@ data = dict(
         data_prefix=data_root_val,
         pipeline=test_pipeline))
 # optimizer
-optimizer = dict(type='SGD', lr=0.015, momentum=0.9, weight_decay=0.0001)
+optimizer = dict(type='SGD', lr=0.005, momentum=0.9, weight_decay=0.0001)
 optimizer_config = dict(grad_clip=dict(max_norm=40, norm_type=2))
 # learning policy
 lr_config = dict(policy='step', step=[40, 80])