[add]上传训练benchmark by z00560161

2020-10-19 20:22:23 +08:00
parent 22b83024f5
commit 82522e2f61
1225 changed files with 345421 additions and 0 deletions
@@ -0,0 +1,47 @@
+tensorflow_config:
+    #中文数据用 bert_config_large_cn.json 英文用bert_config_large_en.json
+    bert_config_file: bert_config_large_cn.json
+    #数据集句子长度是256时 设置为 256,40，句子长度是128时设置为128,20 
+    max_seq_length: 128
+    max_predictions_per_seq: 20
+    
+    # 最佳性能train_batch_size为96，如果超显存，可调小至32 
+    train_batch_size: 96
+    learning_rate: 3.125e-5
+    num_warmup_steps: 100
+    num_train_steps: 1000
+    optimizer_type: adam
+    manual_fp16: True
+    use_fp16_cls: True
+    input_files_dir: /home/data/bert_nv/dataset/cn-wiki-128/
+    eval_files_dir: /home/data/bert_nv/dataset/cn-wiki-128/ 
+    do_train: True
+    do_eval: False
+    num_accumulation_steps: 1
+    iterations_per_loop: 100
+    npu_bert_loss_scale: 0
+    distributed: True
+    graph_memory_max_size: 27917287424 
+    variable_memory_max_size: 5368709120 
+    npu_bert_tail_optimize: True
+    save_checkpoints_steps: 1000
+    npu_bert_clip_by_global_norm: False
+
+    # docker 镜像名称:版本号
+    docker_image: c75:b031
+
+    # 仅多机执行需要配置: ip1:卡数量1,ip2:卡数量2
+    mpirun_ip: 90.90.140.199:8,90.90.140.229:8
+
+    # 指定 device id, 多个 id 使用空格分隔, 数量需与 rank_size 相同
+    device_group_1p: 0
+    device_group_2p: 1 4
+    device_group_4p: 0 1 2 3
+    
+    #profiling 配置
+    PROFILING_MODE: false
+    AICPU_PROFILING_MODE: false
+    PROFILING_OPTIONS: training_trace
+    FP_POINT: bert/embeddings/GatherV2
+    BP_POINT: gradients/bert/embeddings/IdentityN_1_grad/UnsortedSegmentSum
+