Add checkpoint

Author: Randl

README.md

@@ -22,4 +22,14 @@ python imagenet.py --dataroot "/path/to/imagenet/" --resume "/path/to/checkpoint
 ```
 ## Results
 
-TODO
+
+For x0.5 model I achieved 0.4% lower top-1 accuracy than claimed.
+
+|Classification Checkpoint| MACs (M)   | Parameters (M)| Top-1 Accuracy| Top-5 Accuracy|  Claimed top-1|  Claimed top-5|
+|-------------------------|------------|---------------|---------------|---------------|---------------|---------------|
+|      [shufflenet_v2_0.5]|41          |1.37           |          59.86|          81.63|           60.3|              -|
+
+You can test it with
+```bash
+python imagenet.py --dataroot "/path/to/imagenet/" --resume "results/shufflenet_v2_0.5/model_best.pth.tar" -e --scaling 0.5
+```

imagenet.py

@@ -20,7 +20,7 @@
 from model import ShuffleNetV2
 from run import train, test, save_checkpoint, find_bounds_clr
 
-parser = argparse.ArgumentParser(description='MobileNetv2 training with PyTorch')
+parser = argparse.ArgumentParser(description='ShuffleNetv2 training with PyTorch')
 parser.add_argument('--dataroot', required=True, metavar='PATH',
                     help='Path to ImageNet train and val folders, preprocessed as described in '
                          'https://github.com/facebook/fb.resnet.torch/blob/master/INSTALL.md#download-the-imagenet-dataset')

logger.py

@@ -51,7 +51,7 @@ def write_text(self, text, print_t=True):
         if print_t:
             print(text)
 
-    def plot_progress_errk(self, claimed_acc=None, title='MobileNetv2', k=1):
+    def plot_progress_errk(self, claimed_acc=None, title='ShuffleNetv2', k=1):
         tr_str = 'train_error{}'.format(k)
         val_str = 'val_error{}'.format(k)
         plt.figure(figsize=(9, 8), dpi=300)
@@ -70,7 +70,7 @@ def plot_progress_errk(self, claimed_acc=None, title='MobileNetv2', k=1):
         plt.xlim(0, len(self.data[tr_str]) + 1)
         plt.savefig(os.path.join(self.log_path, 'top{}.png'.format(k)))
 
-    def plot_progress_loss(self, title='MobileNetv2'):
+    def plot_progress_loss(self, title='ShuffleNetv2'):
         plt.figure(figsize=(9, 8), dpi=300)
         plt.plot(self.data['train_loss'], label='Training')
         plt.plot(self.data['val_loss'], label='Validation')
@@ -81,7 +81,7 @@ def plot_progress_loss(self, title='MobileNetv2'):
         plt.xlim(0, len(self.data['train_loss']) + 1)
         plt.savefig(os.path.join(self.log_path, 'loss.png'))
 
-    def plot_progress(self, claimed_acc1=None, claimed_acc5=None, title='MobileNetv2'):
+    def plot_progress(self, claimed_acc1=None, claimed_acc5=None, title='ShuffleNetv2'):
         self.plot_progress_errk(claimed_acc1, title, 1)
         self.plot_progress_errk(claimed_acc5, title, 5)
         self.plot_progress_loss(title)

results/shufflenet_v2_0.5/loss.png

@@ -0,0 +1,4 @@
+/home/chaimb/research/ShuffleNetV2-pytorch/imagenet.py --dataroot /home/chaimb/ILSVRC/Data/CLS-LOC --seed 42 --gpus 3,4,5,6 -j 20 --scaling 0.5 -b 1024 --clr --max-lr 1
+Namespace(SE=False, batch_size=1024, c_tag=0.5, clr=True, dataroot='/home/chaimb/ILSVRC/Data/CLS-LOC', decay=4e-05, epochs=400, epochs_per_step=20, evaluate=False, find_clr=False, gamma=0.1, gpus=[3, 4, 5, 6], groups=2, input_size=224, learning_rate=0.01, log_interval=100, max_lr=1.0, min_lr=1e-05, mode='triangular2', momentum=0.9, residual=False, results_dir='./results', resume='', save='2018-08-07_17-05-01', scaling=0.5, schedule=[200, 300], seed=42, start_epoch=0, type='float32', workers=20)
+Claimed accuracy is 60.30% top-1
+Best accuracy is 59.86% top-1