Merge pull request #45 from Oneflow-Inc/dev-hswang

dev-hswang

Classification/cnns/README.md

@@ -634,5 +634,5 @@ python3 cnn_benchmark/of_cnn_train_val.py \
     --model="vgg" \
 ```
 
-The top1 accuracy and the top5 acuuracy are 69.3359% and 89.1370%, respectively for our oneflow model after 90 epochs of training.
+The top1 accuracy and the top5 acuuracy are 69.7326% and 89.3806%, respectively for our oneflow model after 90 epochs of training.
 For reference, the top1 accuracy and the top5 accuracy are 71.5% and 89.9%, respectively for the model from the tensorflow benchmarks after 90 epochs of training.

Classification/cnns/of_cnn_train_val.py

 from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
-
 import os
 import math
-
 import oneflow as flow
-
 import ofrecord_util
 import config as configs
 from util import Snapshot, Summary, InitNodes, Metric
@@ -15,14 +12,10 @@ import resnet_model
 import vgg_model
 import alexnet_model
 
-os.environ['CUDA_VISIBLE_DEVICES'] = '0,1,2,3'
-
-
 parser = configs.get_parser()
 args = parser.parse_args()
 configs.print_args(args)
 
-
 total_device_num = args.num_nodes * args.gpu_num_per_node
 train_batch_size = total_device_num * args.batch_size_per_device
 val_batch_size = total_device_num * args.val_batch_size_per_device

Classification/cnns/optimizer_util.py

@@ -60,6 +60,14 @@ def gen_model_update_conf(args):
             "decay_batches": decay_batches, 
             "end_learning_rate": 0.00001,
         }}
+    
+    # weight decay
+    # if args.wd > 0:
+    #     assert args.wd < 1.0
+    #     model_update_conf['weight_decay_conf'] = {
+    #         "weight_decay_rate": args.wd, 
+    #         "excludes": {"pattern": ['_bn-']}
+    #     }
 
     pprint.pprint(model_update_conf)
     return model_update_conf

Classification/cnns/vgg_model.py

@@ -40,7 +40,6 @@ def conv2d_layer(
     bn=True,
 ):   
     weight_shape = (filters, input.shape[1], kernel_size, kernel_size)
-    print("weight_shape:{}".format(weight_shape))
     weight = flow.get_variable(
         name + "_weight",
         shape=weight_shape,
@@ -65,7 +64,7 @@ def conv2d_layer(
                 output = _batch_norm(output, name + "_bn")
                 output = flow.nn.relu(output)
             else:
-                output = flow.nn.relu(output)     
+                output = flow.nn.relu(output)
         else:
             raise NotImplementedError