Fix optimizer parms in fit.py + Don't repeatedly call slow prepare_mk…

…l.sh script (#7545) (#7547)

* Only call MKL script once

* Fix 'momentum' and 'multi_precision' optimizer args

* fix working

Makefile

@@ -36,8 +36,8 @@ include $(config)
 
 ifeq ($(USE_MKL2017), 1)
 # must run ./prepare_mkl before including mshadow.mk
-	RETURN_STRING = $(shell ./prepare_mkl.sh $(MKLML_ROOT))
-	MKLROOT = $(firstword $(RETURN_STRING))
+	RETURN_STRING := $(shell ./prepare_mkl.sh $(MKLML_ROOT))
+	MKLROOT := $(firstword $(RETURN_STRING))
 	export USE_MKLML = $(lastword $(RETURN_STRING))
 endif
 

example/image-classification/common/fit.py

@@ -163,10 +163,17 @@ def fit(args, network, data_loader, **kwargs):
     lr_scheduler  = lr_scheduler
     optimizer_params = {
             'learning_rate': lr,
-            'momentum' : args.mom,
             'wd' : args.wd,
-            'lr_scheduler': lr_scheduler,
-            'multi_precision': True}
+            'lr_scheduler': lr_scheduler}
+
+    # Add 'multi_precision' parameter only for SGD optimizer
+    if args.optimizer == 'sgd':
+        optimizer_params['multi_precision'] = True
+
+    # Only a limited number of optimizers have 'momentum' property
+    has_momentum = {'sgd', 'dcasgd'}
+    if args.optimizer in has_momentum:
+        optimizer_params['momentum'] = args.mom
 
     monitor = mx.mon.Monitor(args.monitor, pattern=".*") if args.monitor > 0 else None