update readme

FlagOpen · Nov 10, 2023 · 57d9f16 · 57d9f16
1 parent dbf90a4
commit 57d9f16
Show file tree

Hide file tree

Showing 2 changed files with 4 additions and 5 deletions.
diff --git a/training/benchmarks/bigtransfer/pytorch/config/_base.py b/training/benchmarks/bigtransfer/pytorch/config/_base.py
@@ -41,8 +41,7 @@
 eval_batch_size: int = batch_size
 
 target_mAP: float = 0.83
-# max_steps: int = 40000
-max_steps: int = 50
+max_steps: int = 40000
 # step per epoch=1281167/global batchsize
 # 40000 steps with bs==24 1*8 are 4 epoch
 # indeed, transfer model should only fine-tune 1 epoch

diff --git a/training/nvidia/bigtransfer-pytorch/README.md b/training/nvidia/bigtransfer-pytorch/README.md
@@ -42,7 +42,7 @@
 | 总吞吐量       | p_whole,见“性能指标”                            | 实际训练样本数除以总时间(performance_whole) |
 | 训练吞吐量     | p_train,见“性能指标”                            | 不包含每个epoch末尾的评估部分耗时           |
 | **计算吞吐量** | **p_core,见“性能指标”**                         | 不包含数据IO部分的耗时(p3>p2>p1)            |
-| 训练结果       | val_loss,见“性能指标”                           | 验证loss                                    |
+| 训练结果       | acc,见“性能指标”                                | 单位为top1分类准确率(acc1)                  |
 | 额外修改项     | 无                                              |                                             |
 
 * 性能指标
@@ -51,8 +51,8 @@
 | ----------------- | --------- | ------ | -------- | ------- | ------- | ------ | ---------- | --------- |
 | A100单机8卡(1x8)  | fp32      | /      | 5869     | 222     | 225     | 228    | 0.84192    | 31.4/40.0 |
 | A100单机8卡(1x8)  | fp32      | bs=20  | 5505     | 236     | 240     | 243    | 0.84016    | 37.4/40.0 |
-| A100单机单卡(1x1) | fp32      | bs=16  |          | 23.1    | 29.5    | 29.8   |            | 38.1/40.0 |
-| A100两机8卡(2x8)  | fp32      | bs=20  |          | 459     | 465     | 470    |            | 36.6/40.0 |
+| A100单机单卡(1x1) | fp32      | bs=16  | /     | 23.1    | 29.5    | 29.8   | /          | 38.1/40.0 |
+| A100两机8卡(2x8)  | fp32      | bs=20  | /     | 459     | 465     | 470    | /          | 36.6/40.0 |
 
 
 训练精度来源：https://paperswithcode.com/paper/large-scale-learning-of-general-visual