Kaynağa Gözat

data(default): 调整数据配置并添加新数据集

- 将 batch_size 从 256 调整为 32
- 将 shuffle_buffer 从 10000 减少到 1250
- 在训练数据中添加 cuhkpedes 数据集
Yijun Fu 1 ay önce
ebeveyn
işleme
1dc1d11531
1 değiştirilmiş dosya ile 8 ekleme ve 2 silme
  1. 8 2
      configs/default.yml

+ 8 - 2
configs/default.yml

@@ -1,10 +1,10 @@
 data:
-  batch_size: 256
+  batch_size: 32
   pin_memory: true
   num_workers: 6
   # Thomas said it should be at least about 5-10x your batch size; beyond that,
   # the differences become academic.
-  shuffle_buffer: 10000
+  shuffle_buffer: 1250
   seed: ${train.seed}
   dataset:
     meta:
@@ -33,10 +33,16 @@ data:
         path: local_data/imagenet_shards
         prefix: imagenet-val-{000000..000049}.tar
         length: 50000
+      cuhkpedes:
+        type: img_txt_pair
+        path: local_data/cuhkpedes_shards
+        prefix: cuhkpedes-train-{000000..000255}.tar
+        length: 34054
     train:
       - gcc3m
       - gcc12m
       - yfcc14m
+      - cuhkpedes
     val:
       - imagenet