rm job_conf.num_of_batches_in_snapshot

oneflow/core/actor/actor.cpp

@@ -30,11 +30,6 @@ bool IsLastRegstInPieceWithOrder(const Regst* regst, ColIdOrder order) {
          || (order == ColIdOrder::kDescending && regst->col_id() == 0);
 }
 
-bool NeedModelSave(const JobDesc& job_desc, int64_t model_version_id) {
-  return model_version_id + 1 == job_desc.TotalBatchNum()
-         || (model_version_id + 1) % job_desc.NumOfBatchesInSnapshot() == 0;
-}
-
 void Actor::Init(const JobDesc* job_desc, const TaskProto& task_proto,
                  const ThreadCtx& thread_ctx) {
   job_desc_ = job_desc;

oneflow/core/actor/actor.h

@@ -19,7 +19,7 @@ enum class ColIdOrder { kUnCertain = 0, kAscending, kDescending };
 
 bool IsFirstRegstInPieceWithOrder(const Regst*, ColIdOrder);
 bool IsLastRegstInPieceWithOrder(const Regst*, ColIdOrder);
-bool NeedModelSave(const JobDesc& job_desc, int64_t model_version_id);
+inline bool NeedModelSave(const JobDesc& job_desc, int64_t model_version_id) { return false; }
 
 class Actor {
  public:

oneflow/core/job/improver.cpp

@@ -277,10 +277,6 @@ std::function<const HashMap<int64_t, double>&(int64_t)> MakeGetterPathDurations4
   };
 }
 
-uint64_t NumOfPiecesInSnapshot() {
-  return GlobalJobDesc().NumOfBatchesInSnapshot() * GlobalJobDesc().NumOfPiecesInBatch();
-}
-
 std::function<const HashMap<int64_t, double>&(int64_t)> MakeGetterPathIIScales4RegstDescId(
     const ChainActGraph& graph) {
   auto regst_desc_id2consumer_id2ii_scale =

oneflow/core/job/job.proto

@@ -14,7 +14,6 @@ import "oneflow/core/job/sbp_parallel.proto";
 message TrainConf {
   required int64 batch_size = 1; // batch_size % piece_size = 0
   required NormalModelUpdateOpUserConf model_update_conf = 3;
-  required int32 num_of_batches_in_snapshot = 5;
   repeated string loss_lbn = 6;
   optional int32 loss_scale_factor = 7 [default = 1];
   optional string train_step_lbn = 8;

oneflow/core/job/job_desc.cpp

@@ -46,9 +46,6 @@ bool JobDesc::enable_experiment_run() const {
   return job_conf_.exp_run_conf().enable_experiment_run();
 }
 
-int32_t JobDesc::NumOfBatchesInSnapshot() const {
-  return job_conf_.train_conf().num_of_batches_in_snapshot();
-}
 int64_t JobDesc::TotalBatchNum() const { return job_conf_.total_batch_num(); }
 int64_t JobDesc::BatchSize() const { return job_conf_.train_conf().batch_size(); }
 int64_t JobDesc::NumOfPiecesInBatch() const {

oneflow/core/job/job_desc.h

@@ -52,7 +52,6 @@ class JobDesc final {
   int64_t cudnn_buf_limit_mbyte() const { return job_conf_.cudnn_buf_limit_mbyte(); }
 
   // Train conf
-  int32_t NumOfBatchesInSnapshot() const;
   int64_t TotalBatchNum() const;
   int64_t BatchSize() const;
   int64_t NumOfPiecesInBatch() const;

oneflow/python/model/alexnet.py

@@ -210,7 +210,6 @@ def alexnet_train_job():
     job_conf.train_conf()
     job_conf.train_conf().batch_size = 12
     job_conf.train_conf().primary_lr = 0.00001
-    job_conf.train_conf().num_of_batches_in_snapshot = 100
     job_conf.train_conf().model_update_conf.naive_conf.SetInParent()
     job_conf.train_conf().loss_lbn.extend(["softmax_loss/out"])
 

oneflow/python/test/deprecated/bert/run_pretraining.py

@@ -131,7 +131,6 @@ def PretrainJob():
     job_conf.train_conf()
     job_conf.train_conf().primary_lr = 1e-4
     job_conf.train_conf().weight_l2 = 0.01
-    job_conf.train_conf().num_of_batches_in_snapshot = 1000
     job_conf.model_update_conf(_BERT_MODEL_UPDATE_CONF)
     job_conf.train_conf().loss_lbn.extend(["identity_loss/loss"])
     job_conf.enable_inplace(False)

oneflow/python/test/deprecated/enable_all_reduce_group.py

@@ -15,7 +15,6 @@ def UpdateVariable(x, scope_name, enable_all_reduce_group = True):
         job_conf.batch_size(1).data_part_num(1).default_data_type(flow.float)
         job_conf.train_conf()
         job_conf.train_conf().primary_lr = 0.01
-        job_conf.train_conf().num_of_batches_in_snapshot = 100
         job_conf.train_conf().model_update_conf.naive_conf.SetInParent()
         job_conf.train_conf().loss_lbn.extend([scope_name + "-loss_op/out"])
         job_conf.enable_all_reduce_group(enable_all_reduce_group)

oneflow/python/test/deprecated/multi_machine_bert/run_pretraining.py

@@ -131,7 +131,6 @@ def PretrainJob():
     job_conf.train_conf()
     job_conf.train_conf().primary_lr = 1e-4
     job_conf.train_conf().weight_l2 = 0.01
-    job_conf.train_conf().num_of_batches_in_snapshot = 1000
     job_conf.model_update_conf(_BERT_MODEL_UPDATE_CONF)
     job_conf.train_conf().loss_lbn.extend(["identity_loss/loss"])
     job_conf.enable_inplace(False)