hotfix: fix bugs (#20)

* hotfix: update script * hotfix: comment out....
bytedance · Jun 23, 2020 · 6c45c79 · 6c45c79
1 parent 7f12e15
commit 6c45c79
Show file tree

Hide file tree

Showing 2 changed files with 17 additions and 17 deletions.
diff --git a/byteps/mxnet/__init__.py b/byteps/mxnet/__init__.py
@@ -277,10 +277,10 @@ def _register_compressor(self, params, optimizer_params, compression_params):
         # change
         if compression_params.get("momentum"):
             # 1bit compressor use an additional momentum for weight decay
-            # if compressor == "onebit" and "wd" in optimizer_params:
-            #     intra_compressor = Compression.wdmom(
-            #         intra_compressor, optimizer_params["momentum"], optimizer_params["wd"])
-            #     del optimizer_params["wd"]
+            if compressor == "onebit" and "wd" in optimizer_params:
+                intra_compressor = Compression.wdmom(
+                    intra_compressor, optimizer_params["momentum"], optimizer_params["wd"])
+                del optimizer_params["wd"]
 
             del optimizer_params['momentum']
 

diff --git a/example/mxnet/train_gluon_imagenet_byteps_gc.py b/example/mxnet/train_gluon_imagenet_byteps_gc.py
@@ -434,7 +434,7 @@ def train(ctx):
 
         best_val_score = 1
 
-        bps.byteps_declare_tensor("acc")
+        # bps.byteps_declare_tensor("acc")
         for epoch in range(opt.resume_epoch, opt.num_epochs):
             tic = time.time()
             if opt.use_rec:
@@ -505,18 +505,18 @@ def train(ctx):
 
             err_top1_val, err_top5_val = test(ctx, val_data)
 
-            acc = mx.nd.array([train_metric_score, err_top1_val, err_top5_val],
-                              ctx=ctx[0])
-            bps.byteps_push_pull(acc, name="acc", is_average=False)
-            acc /= bps.size()
-            train_metric_score, err_top1_val, err_top5_val = acc[0].asscalar(
-            ), acc[1].asscalar(), acc[2].asscalar()
-
-            if bps.rank() == 0:
-                logger.info('[Epoch %d] training: %s=%f' %
-                            (epoch, train_metric_name, train_metric_score))
-                logger.info('[Epoch %d] validation: err-top1=%f err-top5=%f' %
-                            (epoch, err_top1_val, err_top5_val))
+            # acc = mx.nd.array([train_metric_score, err_top1_val, err_top5_val],
+            #                   ctx=ctx[0])
+            # bps.byteps_push_pull(acc, name="acc", is_average=False)
+            # acc /= bps.size()
+            # train_metric_score, err_top1_val, err_top5_val = acc[0].asscalar(
+            # ), acc[1].asscalar(), acc[2].asscalar()
+
+            # if bps.rank() == 0:
+            logger.info('[Epoch %d] training: %s=%f' %
+                        (epoch, train_metric_name, train_metric_score))
+            logger.info('[Epoch %d] validation: err-top1=%f err-top5=%f' %
+                        (epoch, err_top1_val, err_top5_val))
 
             if err_top1_val < best_val_score:
                 best_val_score = err_top1_val