More fix

eric-haibin-lin · Aug 11, 2017 · 81d867e · 81d867e
1 parent 5279aa1
commit 81d867e
Show file tree

Hide file tree

Showing 2 changed files with 11 additions and 5 deletions.
diff --git a/benchmark/python/sparse_end2end.py b/benchmark/python/sparse_end2end.py
@@ -79,7 +79,6 @@ def next(self):
 
 
 def get_sym(feature_dim):
-     initializer = mx.initializer.Normal()
      x = mx.symbol.Variable("data", stype='csr')
      norm_init = mx.initializer.Normal(sigma=0.01)
      w = mx.symbol.Variable("w", shape=(feature_dim, args.output_dim), init=norm_init, stype='row_sparse')
@@ -97,7 +96,7 @@ def get_sym(feature_dim):
     num_batch = args.num_batch
     kvstore = args.kvstore
     profiler = args.profiler > 0
-    batch_size = args.batch_size
+    batch_size = args.batch_size if args.num_gpu == 0 else args.num_gpu * args.batch_size
     dummy_iter = args.dummy_iter
     dataset = args.dataset
     log_level = args.log_level
@@ -160,8 +159,11 @@ def get_sym(feature_dim):
 
     # start profiler
     if profiler:
-        import random
-        name = 'profile_output_' + str(num_worker) + '.json'
+        device = 'cpu'
+        if args.num_gpu > 0:
+            device = 'gpu' + str(args.num_gpu)
+        name = 'profile_' + args.dataset + '_' + device + '_nworker' + str(num_worker)\
+               + '_batchsize' + str(args.batch_size) + '_outdim' + str(args.output_dim) + '.json'
         mx.profiler.profiler_set_config(mode='all', filename=name)
         mx.profiler.profiler_set_state('run')
 
@@ -207,6 +209,8 @@ def get_sym(feature_dim):
             # accumulate prediction accuracy
             mod.update_metric(metric, batch.label)
         logging.info('epoch %d, %s' % (epoch, metric.get()))
+        if epoch == 0:
+            print "num_batches = ", nbatch
     if profiler:
         mx.profiler.profiler_set_state('stop')
     end = time.time()

diff --git a/src/executor/graph_executor.cc b/src/executor/graph_executor.cc
@@ -1151,7 +1151,9 @@ void GraphExecutor::InitDataEntryMemory(std::vector<NDArray>* shared_pool) {
       CHECK_LE(nword, std::numeric_limits<nnvm::dim_t>::max());
       // allocate float arrays
       TShape shape{static_cast<nnvm::dim_t>(nword)};
-      NDArray nd(shape, ctx);
+      // TODO(junwu): adding delay_alloc=true to create nd
+      // is a temporary solution.
+      NDArray nd(shape, ctx, true);
       data_pool_[i] = nd;
       // put the new allocated arrays to shared pool
       if (shared_pool != nullptr)  {