add Transformer

Tianrui-Li · Aug 22, 2023 · 1c76cef · 1c76cef
1 parent ca308e7
commit 1c76cef
Show file tree

Hide file tree

Showing 3 changed files with 13 additions and 12 deletions.
diff --git a/configs/transformer/j2sample.py b/configs/transformer/j2sample.py
@@ -7,40 +7,43 @@
         type='ViViT2',
         graph_cfg=dict(layout='nturgb+d', mode='spatial'),
         max_position_embeddings_1=26,  # 25*40+1=1001
-        max_position_embeddings_2=121,
-        # dropout=0.1,
+        max_position_embeddings_2=101,
+        # dropout=0.1
         dim=256,
     ),
     cls_head=dict(type='vit2Head', num_classes=60, in_channels=256))
 
 dataset_type = 'PoseDataset'
 ann_file = 'data/nturgbd/ntu60_3danno.pkl'
-clip_len = 120
+clip_len = 100
 train_pipeline = [
     dict(type='PreNormalize3D'),
-    dict(type='STTSample', clip_len=clip_len, p_interval=(0.5, 1)),
+    dict(type='STTSample', clip_len=120, p_interval=(0.5, 1)),
     dict(type='RandomScale', scale=0.1),
     dict(type='RandomRot'),
     # dict(type='RandomRot', theta=0.2),
     dict(type='GenSkeFeat', dataset='nturgb+d', feats=['j']),
+    dict(type='UniformSample', clip_len=clip_len),
     dict(type='PoseDecode'),
     dict(type='FormatGCNInput', num_person=2),
     dict(type='Collect', keys=['keypoint', 'label'], meta_keys=[]),
     dict(type='ToTensor', keys=['keypoint'])
 ]
 val_pipeline = [
     dict(type='PreNormalize3D'),
+    dict(type='STTSample', clip_len=120, p_interval=(0.5, 1)),
     dict(type='GenSkeFeat', dataset='nturgb+d', feats=['j']),
-    dict(type='STTSample', clip_len=clip_len, p_interval=(0.5, 1)),
+    dict(type='UniformSample', clip_len=clip_len, num_clips=1),
     dict(type='PoseDecode'),
     dict(type='FormatGCNInput', num_person=2),
     dict(type='Collect', keys=['keypoint', 'label'], meta_keys=[]),
     dict(type='ToTensor', keys=['keypoint'])
 ]
 test_pipeline = [
     dict(type='PreNormalize3D'),
+    dict(type='STTSample', clip_len=120, p_interval=(0.5, 1)),
     dict(type='GenSkeFeat', dataset='nturgb+d', feats=['j']),
-    dict(type='STTSample', clip_len=clip_len, p_interval=(0.5, 1)),
+    dict(type='UniformSample', clip_len=clip_len, num_clips=10),
     dict(type='PoseDecode'),
     dict(type='FormatGCNInput', num_person=2),
     dict(type='Collect', keys=['keypoint', 'label'], meta_keys=[]),
@@ -71,7 +74,7 @@
 
 # runtime settings
 log_level = 'INFO'
-work_dir = './work_dirs/transformer/j2/8.21-tm2-3'
+work_dir = './work_dirs/transformer/j2/8.21-tm2-4'
 
 auto_resume = False
 seed = 88
diff --git a/pyskl/datasets/pipelines/sampling.py b/pyskl/datasets/pipelines/sampling.py
@@ -64,9 +64,7 @@ def __call__(self, results):
         data_numpy = data_numpy.transpose(3, 1, 2, 0)
         results['keypoint'] = data_numpy
         results['frame_inds'] = self.inds.astype(int)
-        results['clip_len'] = self.clip_len
-        results['frame_interval'] = None
-        results['num_clips'] = self.num_clips
+        results['total_frames'] = self.clip_len
         return results
 
     def __repr__(self):

diff --git a/tools/dist_train.sh b/tools/dist_train.sh
@@ -2,8 +2,8 @@
 
 export MASTER_PORT=$((12000 + $RANDOM % 20000))
 # for server run
-# export CUDA_VISIBLE_DEVICES=1,4,6
- export CUDA_VISIBLE_DEVICES=4,5,6
+ export CUDA_VISIBLE_DEVICES=0,2,3
+# export CUDA_VISIBLE_DEVICES=4,5,6
 # export CUDA_VISIBLE_DEVICES=5,6
 
 set -x