fixed problems with variational dropout

Author: Zach Teed

RAFT.png

@@ -4,6 +4,8 @@ This repository contains the source code for our paper:
 [RAFT: Recurrent All Pairs Field Transforms for Optical Flow](https://arxiv.org/pdf/2003.12039.pdf)<br/>
 Zachary Teed and Jia Deng<br/>
 
+<img src="RAFT.png">
+
 ## Requirements
 Our code was tested using PyTorch 1.3.1 and Python 3. The following additional packages need to be installed
 
@@ -84,11 +86,11 @@ python train.py --name=kitti_ft --image_size 288 896 --dataset=kitti --num_steps
 You can evaluate a model on Sintel and KITTI by running
 
 ```Shell
-python evaluate.py --model=checkpoints/chairs+things.pth
+python evaluate.py --model=models/chairs+things.pth
 ```
 
 or the small model by including the `small` flag
 
 ```Shell
-python evaluate.py --model=checkpoints/small.pth --small
+python evaluate.py --model=models/small.pth --small
 ```

README.md

@@ -133,8 +133,20 @@ def __init__(self, args, hidden_dim=96):
         self.gru = ConvGRU(hidden_dim=hidden_dim, input_dim=82+64)
         self.flow_head = FlowHead(hidden_dim, hidden_dim=128)
 
+        self.drop_inp = VariationalHidDropout(dropout=args.dropout)
+        self.drop_net = VariationalHidDropout(dropout=args.dropout)
+
+    def reset_mask(self, net, inp):
+        self.drop_inp.reset_mask(inp)
+        self.drop_net.reset_mask(net)
+
     def forward(self, net, inp, corr, flow):
         motion_features = self.encoder(flow, corr)
+
+        if self.training:
+            net = self.drop_net(net)
+            inp = self.drop_inp(inp)
+
         inp = torch.cat([inp, motion_features], dim=1)
         net = self.gru(net, inp)
         delta_flow = self.flow_head(net)
@@ -157,12 +169,12 @@ def reset_mask(self, net, inp):
 
     def forward(self, net, inp, corr, flow):
         motion_features = self.encoder(flow, corr)
-        inp = torch.cat([inp, motion_features], dim=1)
 
         if self.training:
             net = self.drop_net(net)
             inp = self.drop_inp(inp)
-
+        
+        inp = torch.cat([inp, motion_features], dim=1)
         net = self.gru(net, inp)
         delta_flow = self.flow_head(net)
 

core/modules/update.py

@@ -26,7 +26,7 @@ def __init__(self, args):
             args.corr_levels = 4
             args.corr_radius = 4
 
-        if 'dropout' not in args._get_kwargs():
+        if not hasattr(args, 'dropout'):
             args.dropout = 0
 
         # feature network, context network, and update block

core/raft.py

@@ -21,7 +21,7 @@
 
 # exclude extremly large displacements
 MAX_FLOW = 1000
-SUM_FREQ = 100
+SUM_FREQ = 200
 VAL_FREQ = 5000
 
 
@@ -56,7 +56,7 @@ def sequence_loss(flow_preds, flow_gt, valid):
 
 
 def fetch_dataloader(args):
-    """ Create the data loader for the corresponding trainign set """
+    """ Create the data loader for the corresponding training set """
 
     if args.dataset == 'chairs':
         train_dataset = datasets.FlyingChairs(args, image_size=args.image_size)
@@ -86,7 +86,7 @@ def fetch_optimizer(args, model):
     optimizer = optim.AdamW(model.parameters(), lr=args.lr, weight_decay=args.wdecay, eps=args.epsilon)
 
     scheduler = optim.lr_scheduler.OneCycleLR(optimizer, args.lr, args.num_steps,
-        pct_start=0.2, cycle_momentum=False, anneal_strategy='linear', final_div_factor=1.0)
+        pct_start=0.2, cycle_momentum=False, anneal_strategy='linear')
 
     return optimizer, scheduler