UPDATE

Author: Xuanyi Dong

README.md

@@ -51,7 +51,7 @@ python ./exps/basic_main.py [<required arguments>]
 The argument list is loaded by `./lib/config_utils/basic_args.py`.
 An examples script can is `./scripts/300W-DET.sh`, and you can simple run to train the base detector on the `300-W` dataset.
 ```
-sh scripts/300W-DET.sh
+bash scripts/300W-DET.sh
 ```
 
 ### Improving the Detector by SBR
@@ -64,7 +64,7 @@ The argument list is loaded by `./lib/config_utils/lk_args.py`.
 #### An example to train SBR on the unlabeled sequences
 The `init_model` parameter is the path to the detector trained in the `Basic Training` section.
 ```
-sh scripts/demo_sbr.sh
+bash scripts/demo_sbr.sh
 ```
 To see visualization results use the commands in `Visualization`.
 
@@ -104,7 +104,7 @@ ffmpeg -start_number 3 -i cache_data/cache/demo-sbr-vis/image%04d.png -b:v 30000
 supervision-by-registration is released under the [CC-BY-NC license](https://github.com/facebookresearch/supervision-by-registration/blob/master/LICENSE).
 
 
-## Useful information
+## Useful Information
 
 ### 1. train on your own video data
 You should look at the `./lib/datasets/VideoDataset.py` and `./lib/datasets/parse_utils.py`, and add how to find the neighbour frames when giving one image path.

exps/basic_main.py

@@ -159,6 +159,7 @@ def main(args):
           'args' : deepcopy(args),
           'arch' : model_config.arch,
           'state_dict': net.state_dict(),
+          'detector'  : net.state_dict(),
           'scheduler' : scheduler.state_dict(),
           'optimizer' : optimizer.state_dict(),
           }, logger.path('model') / '{:}-{:}.pth'.format(model_config.arch, epoch_str), logger)
@@ -169,6 +170,7 @@ def main(args):
           }, logger.last_info(), logger)
 
     eval_results = eval_all(args, eval_loaders, net, criterion, epoch_str, logger, opt_config)
+    logger.log('NME Results : {:}'.format( eval_results ))
 
     # measure elapsed time
     epoch_time.update(time.time() - start_time)

exps/eval.py

@@ -20,6 +20,7 @@
 from models import obtain_model, remove_module_dict
 from config_utils import load_configure
 
+
 def evaluate(args):
   assert torch.cuda.is_available(), 'CUDA is not available.'
   torch.backends.cudnn.enabled   = True
@@ -39,14 +40,15 @@ def evaluate(args):
                                       std=[0.229, 0.224, 0.225])
 
   param = snapshot['args']
-  eval_transform  = transforms.Compose([transforms.PreCrop(param.pre_crop_expand), transforms.TrainScale2WH((param.crop_width, param.crop_height)),  transforms.ToTensor(), normalize])
+  import pdb; pdb.set_trace()
+  eval_transform  = transforms.Compose([transforms.PreCrop(param.pre_crop_expand), transforms.TrainScale2WH((param.crop_width, param.crop_height)), transforms.ToTensor(), normalize])
   model_config = load_configure(param.model_config, None)
   dataset = Dataset(eval_transform, param.sigma, model_config.downsample, param.heatmap_type, param.data_indicator)
   dataset.reset(param.num_pts)
 
   net = obtain_model(model_config, param.num_pts + 1)
   net = net.cuda()
-  weights = remove_module_dict(snapshot['state_dict'])
+  weights = remove_module_dict(snapshot['detector'])
   net.load_state_dict(weights)
   print ('Prepare input data')
   [image, _, _, _, _, _, cropped_size], meta = dataset.prepare_input(args.image, args.face)

exps/lk_main.py

@@ -169,6 +169,7 @@ def main(args):
           'args' : deepcopy(args),
           'arch' : model_config.arch,
           'state_dict': net.state_dict(),
+          'detector'  : detector.state_dict(),
           'scheduler' : scheduler.state_dict(),
           'optimizer' : optimizer.state_dict(),
           }, logger.path('model') / '{:}-{:}.pth'.format(model_config.arch, epoch_str), logger)

lib/procedure/basic_eval.py

@@ -23,8 +23,8 @@ def basic_eval_all(args, loaders, net, criterion, epoch_str, logger, opt_config)
     nme, _, _ = eval_meta.compute_mse(logger)
     meta_path = logger.path('meta') / 'eval-{:}-{:02d}-{:02d}.pth'.format(epoch_str, i, len(loaders))
     eval_meta.save(meta_path)
-    nmes.append(nme)
-  return ', '.join(['{:.1f}'.format(x) for x in nmes])
+    nmes.append(nme*100)
+  return ', '.join(['{:.2f}'.format(x) for x in nmes])
 
 
 def basic_eval(args, loader, net, criterion, epoch_str, logger, opt_config):

lib/xvision/common_eval.py

@@ -62,7 +62,7 @@ def evaluate_normalized_mean_error(predictions, groundtruth, log, extra_faces):
   accuracy_under_007 = np.sum(error_per_image<0.07) * 100. / error_per_image.size
   accuracy_under_008 = np.sum(error_per_image<0.08) * 100. / error_per_image.size
 
-  print_log('Compute NME and AUC for {:} images with {:} points :: [(nms): mean={:.3f}, std={:.3f}], [email protected]={:.3f}, [email protected]{:.3f}, [email protected]={:.3f}, [email protected]={:.3f}'.format(num_images, num_points, normalise_mean_error*100, error_per_image.std()*100, area_under_curve07*100, area_under_curve08*100, accuracy_under_007, accuracy_under_008), log)
+  print_log('Compute NME and AUC for {:} images with {:} points :: [(NME): mean={:.3f}, std={:.3f}], [email protected]={:.3f}, [email protected]{:.3f}, [email protected]={:.3f}, [email protected]={:.3f}'.format(num_images, num_points, normalise_mean_error*100, error_per_image.std()*100, area_under_curve07*100, area_under_curve08*100, accuracy_under_007, accuracy_under_008), log)
 
   for_pck_curve = []
   for x in range(0, 3501, 1):