resnet readme

examples/ResNet/README.md

 
-ImageNet training code coming soon.
+## imagenet-resnet.py
+
+ImageNet training code of pre-activation ResNet. It follows the setup in
+[fb.resnet.torch](https://github.com/facebook/fb.resnet.torch) and get similar performance (with much fewer lines of code),
+
+| Model (WIP)        | Top 5 Error | Top 1 Error |
+|:-------------------|-------------|------------:|
+| ResNet 18          |      10.67% |      29.50% |
+| ResNet 50          |      7.13%  |      24.12% |
 
 ## load-resnet.py
 
@@ -24,8 +32,7 @@ The per-pixel mean used here is slightly different from the original.
 
 ## cifar10-resnet.py
 
-Reproduce the results in paper "Deep Residual Learning for Image Recognition", [http://arxiv.org/abs/1512.03385](http://arxiv.org/abs/1512.03385)
-with the variants proposed in "Identity Mappings in Deep Residual Networks", [https://arxiv.org/abs/1603.05027](https://arxiv.org/abs/1603.05027) on CIFAR10.
+Reproduce pre-activation ResNet on CIFAR10.
 
 The train error shown here is a moving average of the error rate of each batch in training.
 The validation error here is computed on test set.

examples/ResNet/imagenet-resnet.py

@@ -22,11 +22,9 @@ Training code of Pre-Activation version of ResNet on ImageNet.
 Mainly follow the setup in fb.resnet.torch
 """
 
-
-NR_GPU = 4
 TOTAL_BATCH_SIZE = 256
-BATCH_SIZE = TOTAL_BATCH_SIZE / NR_GPU
 INPUT_SHAPE = 224
+DEPTH = None
 
 class Model(ModelDesc):
     def _get_input_vars(self):
@@ -93,7 +91,7 @@ class Model(ModelDesc):
             50: ([3,4,6,3], bottleneck),
             101: ([3,4,23,3], bottleneck)
         }
-        defs, block_func = cfg[34]
+        defs, block_func = cfg[DEPTH]
 
         with argscope(Conv2D, nl=tf.identity, use_bias=False,
                 W_init=variance_scaling_initializer(mode='FAN_OUT')):
@@ -231,22 +229,29 @@ def eval_on_ILSVRC12(model_file, data_dir):
 
 if __name__ == '__main__':
     parser = argparse.ArgumentParser()
-    parser.add_argument('--gpu', help='comma separated list of GPU(s) to use.') # nargs='*' in multi mode
+    parser.add_argument('--gpu', help='comma separated list of GPU(s) to use.')
     parser.add_argument('--data', help='ILSVRC dataset dir')
     parser.add_argument('--load', help='load model')
+    parser.add_argument('-d', '--depth', help='resnet depth',
+            type=int, default=18, choices=[18, 34, 50, 101])
     parser.add_argument('--eval', action='store_true')
     args = parser.parse_args()
+
+    DEPTH = args.depth
     if args.gpu:
         os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
     if args.eval:
+        BATCH_SIZE = 128    # something that can run on one gpu
         eval_on_ILSVRC12(args.load, args.data)
         sys.exit()
 
-    logger.auto_set_dir()
+    assert args.gpu is not None, "Need to specify a list of gpu for training!"
+    NR_GPU = len(args.gpu.split(','))
+    BATCH_SIZE = TOTAL_BATCH_SIZE / NR_GPU
 
+    logger.auto_set_dir()
     config = get_config()
     if args.load:
         config.session_init = SaverRestore(args.load)
-    if args.gpu:
-        config.nr_tower = len(args.gpu.split(','))
+    config.nr_tower = NR_GPU
     SyncMultiGPUTrainer(config).train()

examples/ResNet/load-resnet.py

@@ -86,8 +86,8 @@ class Model(ModelDesc):
                 .GlobalAvgPooling('gap')
                 .FullyConnected('fc1000', 1000, nl=tf.identity)())
         prob = tf.nn.softmax(fc1000, name='prob')
-        nr_wrong = tf.reduce_sum(prediction_incorrect(fc1000, label), name='wrong-top1')
-        nr_wrong = tf.reduce_sum(prediction_incorrect(fc1000, label, 5), name='wrong-top5')
+        nr_wrong = prediction_incorrect(fc1000, label, name='wrong-top1')
+        nr_wrong = prediction_incorrect(fc1000, label, 5, name='wrong-top5')
 
 def get_inference_augmentor():
     # load ResNet mean from Kaiming:
@@ -136,18 +136,20 @@ def eval_on_ILSVRC12(params, data_dir):
         model=Model(),
         input_var_names=['input', 'label'],
         session_init=ParamRestore(params),
-        output_var_names=['prob', 'wrong-top1', 'wrong-top5']
+        output_var_names=['wrong-top1', 'wrong-top5']
     )
     pred = SimpleDatasetPredictor(pred_config, ds)
     acc1, acc5 = RatioCounter(), RatioCounter()
     for o in pred.get_result():
         batch_size = o[0].shape[0]
-        acc1.feed(o[1], batch_size)
-        acc5.feed(o[2], batch_size)
+        acc1.feed(o[0].sum(), batch_size)
+        acc5.feed(o[1].sum(), batch_size)
     print("Top1 Error: {}".format(acc1.ratio))
     print("Top5 Error: {}".format(acc5.ratio))
 
 def name_conversion(caffe_layer_name):
+    """ Convert a caffe parameter name to a tensorflow parameter name as
+        defined in the above model """
     # beginning & end mapping
     NAME_MAP = {'bn_conv1/beta': 'conv0/bn/beta',
             'bn_conv1/gamma': 'conv0/bn/gamma',
@@ -190,7 +192,7 @@ if __name__ == '__main__':
     parser.add_argument('--gpu', help='comma separated list of GPU(s) to use.') # nargs='*' in multi mode
     parser.add_argument('--load', required=True,
                         help='.npy model file generated by tensorpack.utils.loadcaffe')
-    parser.add_argument('--depth', help='resnet depth', required=True, type=int, choices=[50, 101, 152])
+    parser.add_argument('-d', '--depth', help='resnet depth', required=True, type=int, choices=[50, 101, 152])
     parser.add_argument('--input', help='an input image')
     parser.add_argument('--eval', help='ILSVRC dir to run validation on')
 

tensorpack/RL/gymenv.py

@@ -8,7 +8,8 @@ import time
 from ..utils import logger
 try:
     import gym
-    gym.undo_logger_setup()
+    # TODO
+    #gym.undo_logger_setup()
     # https://github.com/openai/gym/pull/199
     # not sure does it cause other problems
     __all__ = ['GymEnv']