fix os.execute, fix hacks for data_parallel, fix COCO image loading, fix yaml missing size

Author: ssnl

extract.py

@@ -26,14 +26,14 @@
                     help='dir dataset to download or/and load images')
 parser.add_argument('--data_split', default='train', type=str,
                     help='Options: (default) train | val | test')
-parser.add_argument('--arch', '-a', default='resnet152',
+parser.add_argument('--arch', '-a', default='fbresnet152',
                     choices=convnets.model_names,
                     help='model architecture: ' +
                         ' | '.join(convnets.model_names) +
                         ' (default: fbresnet152)')
-parser.add_argument('--workers', default=4, type=int, 
+parser.add_argument('--workers', default=4, type=int,
                     help='number of data loading workers (default: 4)')
-parser.add_argument('--batch_size', '-b', default=80, type=int, 
+parser.add_argument('--batch_size', '-b', default=80, type=int,
                     help='mini-batch size (default: 80)')
 parser.add_argument('--mode', default='both', type=str,
                     help='Options: att | noatt | (default) both')
@@ -56,7 +56,7 @@ def main():
     if args.dataset == 'coco':
         if 'coco' not in args.dir_data:
             raise ValueError('"coco" string not in dir_data')
-        dataset = datasets.COCOImages(args.data_split, dict(dir=args.dir_data), 
+        dataset = datasets.COCOImages(args.data_split, dict(dir=args.dir_data),
             transform=transforms.Compose([
                 transforms.Scale(args.size),
                 transforms.CenterCrop(args.size),
@@ -68,7 +68,7 @@ def main():
             raise ValueError('train split is required for vgenome')
         if 'vgenome' not in args.dir_data:
             raise ValueError('"vgenome" string not in dir_data')
-        dataset = datasets.VisualGenomeImages(args.data_split, dict(dir=args.dir_data), 
+        dataset = datasets.VisualGenomeImages(args.data_split, dict(dir=args.dir_data),
             transform=transforms.Compose([
                 transforms.Scale(args.size),
                 transforms.CenterCrop(args.size),
@@ -122,7 +122,7 @@ def extract(data_loader, model, path_file, mode):
 
         nb_regions = output_att.size(2) * output_att.size(3)
         output_noatt = output_att.sum(3).sum(2).div(nb_regions).view(-1, 2048)
-        
+
         batch_size = output_att.size(0)
         if mode == 'both' or mode == 'att':
             hdf5_att[idx:idx+batch_size]   = output_att.data.cpu().numpy()
@@ -141,7 +141,7 @@ def extract(data_loader, model, path_file, mode):
                    i, len(data_loader),
                    batch_time=batch_time,
                    data_time=data_time,))
-            
+
     hdf5_file.close()
 
     # Saving image names in the same order than extraction
@@ -154,4 +154,4 @@ def extract(data_loader, model, path_file, mode):
 
 
 if __name__ == '__main__':
-    main()
+    main()

options/vqa/mutan_att_trainval.yaml

@@ -12,8 +12,9 @@ vqa:
     samplingans: True
 coco:
     dir: data/coco
-    arch: fbresnet152torch
+    arch: fbresnet152
     mode: att
+    size: 448
 model:
     arch: MutanAtt
     dim_v: 2048

vqa/datasets/coco.py

@@ -23,18 +23,21 @@ class COCOImages(AbstractImagesDataset):
     def __init__(self, data_split, opt, transform=None, loader=default_loader):
         self.split_name = split_name(data_split)
         super(COCOImages, self).__init__(data_split, opt, transform, loader)
-        self.dir_split = os.path.join(self.dir_raw, self.split_name)
+        self.dir_split = self.get_dir_data()
         self.dataset = ImagesFolder(self.dir_split, transform=self.transform, loader=self.loader)
         self.name_to_index = self._load_name_to_index()
 
+    def get_dir_data(self):
+        return os.path.join(self.dir_raw, self.split_name)
+
     def _raw(self):
         if self.data_split in ['train', 'val']:
             os.system('wget http://msvocds.blob.core.windows.net/coco2014/{}.zip -P {}'.format(self.split_name, self.dir_raw))
         elif self.data_split == 'test':
-            os.execute('wget http://msvocds.blob.core.windows.net/coco2015/test2015.zip -P '+self.dir_raw)
+            os.system('wget http://msvocds.blob.core.windows.net/coco2015/test2015.zip -P '+self.dir_raw)
         else:
             assert False, 'data_split {} not exists'.format(self.data_split)
-        os.execute('unzip '+os.path.join(self.dir_raw, self.split_name+'.zip')+' -d '+self.dir_raw)
+        os.system('unzip '+os.path.join(self.dir_raw, self.split_name+'.zip')+' -d '+self.dir_raw)
 
     def _load_name_to_index(self):
         self.name_to_index = {name:index for index, name in enumerate(self.dataset.imgs)}

vqa/datasets/images.py

@@ -38,7 +38,7 @@ def __init__(self, root, transform=None, loader=default_loader):
 
     def __getitem__(self, index):
         item = {}
-        item['name'] = self.imgs[index] 
+        item['name'] = self.imgs[index]
         item['path'] = os.path.join(self.root, item['name'])
         if self.loader is not None:
             item['visual']  = self.loader(item['path'])
@@ -57,11 +57,14 @@ def __init__(self, data_split, opt, transform=None, loader=default_loader):
         self.opt = opt
         self.transform = transform
         self.loader = loader
-
         self.dir_raw = os.path.join(self.opt['dir'], 'raw')
-        if not os.path.exists(self.dir_raw):
+
+        if not os.path.exists(self.get_dir_data()):
             self._raw()
 
+    def get_dir_data(self):
+        return self.dir_raw
+
     def get_by_name(self, image_name):
         index = self.name_to_index[image_name]
         return self[index]
@@ -73,4 +76,4 @@ def __getitem__(self, index):
         raise NotImplementedError
 
     def __len__(self):
-        raise NotImplementedError
+        raise NotImplementedError

vqa/models/convnets.py

@@ -2,7 +2,6 @@
 import torch
 import torch.nn as nn
 import torchvision.models as pytorch_models
-
 import sys
 sys.path.append('vqa/external/pretrained-models.pytorch')
 import pretrainedmodels as torch7_models
@@ -21,7 +20,21 @@
 def factory(opt, cuda=True, data_parallel=True):
     opt = copy.copy(opt)
 
-    # forward_* will be better handle in futur release
+    class WrapperModule(nn.Module):
+        def __init__(self, net, forward_fn):
+            super(WrapperModule, self).__init__()
+            self.net = net
+            self.forward_fn = forward_fn
+
+        def forward(self, x):
+            return self.forward_fn(self.net, x)
+
+        def __getattr__(self, attr):
+            try:
+                return super(WrapperModule, self).__getattr__(attr)
+            except AttributeError:
+                return getattr(self.net, attr)
+
     def forward_resnet(self, x):
         x = self.conv1(x)
         x = self.bn1(x)
@@ -58,26 +71,23 @@ def forward_resnext(self, x):
     if opt['arch'] in pytorch_resnet_names:
         model = pytorch_models.__dict__[opt['arch']](pretrained=True)
 
-        convnet = model # ugly hack in case of DataParallel wrapping
-        model.forward = lambda x: forward_resnet(convnet, x)
+        model = WrapperModule(model, forward_resnet) # ugly hack in case of DataParallel wrapping
 
     elif opt['arch'] == 'fbresnet152':
         model = torch7_models.__dict__[opt['arch']](num_classes=1000,
                                                     pretrained='imagenet')
 
-        convnet = model # ugly hack in case of DataParallel wrapping
-        model.forward = lambda x: forward_resnet(convnet, x)
+        model = WrapperModule(model, forward_resnet) # ugly hack in case of DataParallel wrapping
 
     elif opt['arch'] in torch7_resnet_names:
         model = torch7_models.__dict__[opt['arch']](num_classes=1000,
                                                     pretrained='imagenet')
-        
-        convnet = model # ugly hack in case of DataParallel wrapping
-        model.forward = lambda x: forward_resnext(convnet, x)
+
+        model = WrapperModule(model, forward_resnext) # ugly hack in case of DataParallel wrapping
 
     else:
         raise ValueError
-    
+
     if data_parallel:
         model = nn.DataParallel(model).cuda()
         if not cuda:
@@ -86,4 +96,4 @@ def forward_resnext(self, x):
     if cuda:
         model.cuda()
 
-    return model
+    return model