仓库 Issues 博客
语言: 全部
排序: 最佳匹配
找到相关 Issues 约为157
搜索工具
排序方式: 最佳匹配
sum(cumsum_logits > top_p) but it should be top_p_num = sum(cumsum_logits < top_p) + 1
accuracy is {'top_1_accuracy': 0.24097062643620742, 'top_5_accuracy': 0.4638348923468878} ``` 文档给定精度标准如下: 八卡:top1:78.33%, top5:93.96%
mobilenetV3_small_x1_0执行8p训练结果评估,相同的ckpt文件精度不达标: 精度结果: result: {'Loss': 8.284313969734388, 'Top_1_Acc': 0.0014022435897435897, 'Top_5_Acc': 0
--help 2.执行不支持的top 命令, toybox top -H; toybox top -k; toybox top -o; toybox top -O; toybox top -s; toybox top -b; toybox top -d; toybox
: result on upright images: {'top_1_accuracy': 0.0} result on 180 degrees rotated images: {'top_1_accuracy': 1.0} 二、软件版本: --CANN
: result on upright images: {'top_1_accuracy': 0.0} result on 180 degrees rotated images: {'top_1_accuracy': 1.0} 二、软件版本: --CANN
mobilenetV3_small_x1_0执行8p训练结果评估,相同的ckpt文件精度不达标: 精度结果: result: {'Loss': 8.284313969734388, 'Top_1_Acc': 0.0014022435897435897, 'Top_5_Acc': 0
the current behavior metric: {'Loss': nan, 'Top1-Acc': 0.001001602564102564, 'Top5-Acc': 0.005008012820512821} ## Describe the
filtered_logits = top_k_top_p_filtering(next_token_logits, top_k=top_k, top_p=top_p) File "generate.py", line 74, in top_k_top_p_filtering
/img/016_00.png) top center no-repeat, url(./img/016_01.png) top center no-repeat, url(./img/016_02.png) top center no-repeat,
GPU比仍不达标,此处针对TOP2算子分析 二、软件版本: 20210825_001144024_newest 10.246.246.59 arm_eulor2.9 三、测试步骤: SG2IM_for_PyTorch 网络单P训练,根据profiling定位性能问题

推荐博客

...