Updates and fixes for newer PyTorch versions. Using PyTorch's tensorboard instead of TensorboardX.

Author: martin-danelljan

INSTALL.md

@@ -23,10 +23,10 @@ conda install pytorch torchvision cudatoolkit=10.0 -c pytorch
 - It is possible to use any PyTorch supported version of CUDA (not necessarily v10).   
 - For more details about PyTorch installation, see https://pytorch.org/get-started/previous-versions/.  
 
-#### Install matplotlib, pandas, opencv, visdom and tensorboadX  
+#### Install matplotlib, pandas, opencv, visdom and tensorboad  
 ```bash
 conda install matplotlib pandas
-pip install opencv-python tensorboardX visdom
+pip install opencv-python visdom tb-nightly
 ```
 
 

INSTALL_win.md

@@ -25,10 +25,10 @@ conda install pytorch torchvision cudatoolkit=10.0 -c pytorch
 - It is possible to use any PyTorch supported version of CUDA (not necessarily v10), but better be the same version with your preinstalled CUDA (if you have one)    
 - For more details about PyTorch installation, see https://pytorch.org/get-started/previous-versions/.  
 
-#### Install matplotlib, pandas, opencv, visdom and tensorboadX  
+#### Install matplotlib, pandas, opencv, visdom and tensorboad  
 ```bash
 conda install matplotlib pandas
-pip install opencv-python tensorboardX visdom
+pip install opencv-python visdom tb-nightly
 ```
 
 

README.md

@@ -1,15 +1,5 @@
 # PyTracking
 A general python framework for training and running visual object trackers, based on **PyTorch**.
-
-### **News:** Code released for **DiMP**!!!
-Code now released for our new tracker **DiMP**, accepted as an Oral at ICCV 2019.  
-This release also includes many **new features**, including:  
-* Visualization with Visdom  
-* VOT integration  
-* Many new network modules  
-* Multi GPU training  
-* PyTorch v1.2 support  
-
 
 ## Highlights
 

install.sh

@@ -39,8 +39,8 @@ pip install opencv-python
 
 echo ""
 echo ""
-echo "****************** Installing tensorboardX ******************"
-pip install tensorboardX
+echo "****************** Installing tensorboard ******************"
+pip install tb-nightly
 
 echo ""
 echo ""

ltr/admin/loading.py

@@ -3,8 +3,16 @@
 import sys
 from pathlib import Path
 import importlib
+import inspect
 
 
+def load_trained_network(workspace_dir, network_path, checkpoint=None):
+    checkpoint_dir = os.path.join(workspace_dir, 'checkpoints')
+    directory = '{}/{}'.format(checkpoint_dir, network_path)
+
+    net, _ = load_network(directory, checkpoint)
+    return net
+
 def load_network(network_dir=None, checkpoint=None, constructor_fun_name=None, constructor_module=None, **kwargs):
         """Loads a network checkpoint file.
 
@@ -18,6 +26,7 @@ def load_network(network_dir=None, checkpoint=None, constructor_fun_name=None, c
         The extra keyword arguments are supplied to the network constructor to replace saved ones.
         """
 
+
         if network_dir is not None:
             net_path = Path(network_dir)
         else:
@@ -58,14 +67,16 @@ def load_network(network_dir=None, checkpoint=None, constructor_fun_name=None, c
                 net_constr.fun_name = constructor_fun_name
             if constructor_module is not None:
                 net_constr.fun_module = constructor_module
-            for arg, val in kwargs.items():
-                if arg in net_constr.kwds.keys():
-                    net_constr.kwds[arg] = val
-                else:
-                    print('WARNING: Keyword argument "{}" not found when loading network.'.format(arg))
             # Legacy networks before refactoring
             if net_constr.fun_module.startswith('dlframework.'):
                 net_constr.fun_module = net_constr.fun_module[len('dlframework.'):]
+            net_fun = getattr(importlib.import_module(net_constr.fun_module), net_constr.fun_name)
+            net_fun_args = list(inspect.signature(net_fun).parameters.keys())
+            for arg, val in kwargs.items():
+                if arg in net_fun_args:
+                    net_constr.kwds[arg] = val
+                else:
+                    print('WARNING: Keyword argument "{}" not found when loading network. It was ignored.'.format(arg))
             net = net_constr.get()
         else:
             raise RuntimeError('No constructor for the given network.')
@@ -118,4 +129,4 @@ def _cleanup_legacy_env():
         if m.startswith('dlframework'):
             del_modules.append(m)
     for m in del_modules:
-        del sys.modules[m]
+        del sys.modules[m]

ltr/admin/tensorboard.py

@@ -1,6 +1,6 @@
 import os
 from collections import OrderedDict
-from tensorboardX import SummaryWriter
+from torch.utils.tensorboard import SummaryWriter
 
 
 class TensorboardWriter:

ltr/models/target_classifier/optimizer.py

@@ -148,7 +148,7 @@ def forward(self, weights, feat, bb, sample_weight=None, num_iter=None, compute_
             scores_grad = sample_weight * (score_mask * scores_grad)
 
             # Compute optimal step length
-            alpha_num = (weights_grad * weights_grad).view(num_sequences, -1).sum(dim=1)
+            alpha_num = (weights_grad * weights_grad).sum(dim=(1,2,3))
             alpha_den = ((scores_grad * scores_grad).view(num_images, num_sequences, -1).sum(dim=(0,2)) + reg_weight * alpha_num).clamp(1e-8)
             alpha = alpha_num / alpha_den
 

pytracking/libs/operation.py

@@ -39,5 +39,4 @@ def conv1x1(input: torch.Tensor, weight: torch.Tensor):
     if weight is None:
         return input
 
-    return torch.matmul(weight.view(weight.shape[0], weight.shape[1]),
-                        input.view(input.shape[0], input.shape[1], -1)).view(input.shape[0], weight.shape[0], input.shape[2], input.shape[3])
+    return torch.conv2d(input, weight)