pathes updated

README.md

@@ -35,7 +35,7 @@ d.	`pip install opencv-python`
 
 The file `requirements.txt` is found in the root of the repository.
 
-The additional files required for operating the repository are available at  `https://ibm.box.com/s/jbpddqmlmcjq4hnaw6rr10dn60wm4vl9` or at `https://drive.google.com/drive/folders/1MZ6HWQpR_Oseo5_v5gmrlAyubrPL-ciO?usp=sharing`. The folders provided in 
+The additional files required for operating the repository are available at `https://drive.google.com/drive/folders/1MZ6HWQpR_Oseo5_v5gmrlAyubrPL-ciO?usp=sharing`. The folders provided in 
 this link ('data' and 'output') should be placed  under the RepMet root of the git package.
 The 'data' folder contains the pre-trained model and associated files, and the /output/benchmarks contains the the benchmark files 
 
@@ -112,7 +112,7 @@ The code for pretraining the model on a large dataset of auxiliary categories is
 `python ./experiments/fpn_end2end_train_test.py --cfg=./experiments/cfgs/resnet_v1_101_voc0712_trainval_fpn_dcn_oneshot_end2end_ohem_8.yaml`
 
 The datasets for the training is defined in the DATASET section of the configuration .yaml file. In the config file present in the command above, two datasets (PascalVOC;ImageNet) are used, and all related fields ahave two corresponding values, separated by `;`  
-
+The restriction of the model to a subset of all the classes is enabled via the `DATASET.cls_filter_files` argument, where pathes to files with class name-to-id LUT and the list of selected classes are provided.  
 The code can use multiple GPUs. List of their ordinals is given in the 'gpus' argument in the `.yaml` file.
 
 ## Options and features of this code package

experiments/cfgs/resnet_v1_101_voc0712_trainval_fpn_dcn_oneshot_end2end_ohem_8.yaml

@@ -13,8 +13,8 @@ default:
   frequent: 10
   kvstore: device
 network:
-#  pretrained: "/dccstor/jsdata1/dev/Deformable-ConvNets/model/fpn_dcn_coco" #"./model/pretrained_model/resnet_v1_101"
-  pretrained: '/dccstor/jsdata1/dev/RepMet/model/pretrained_model/resnet_v1_101' # "./model/pretrained_model/resnet_v1_101"
+  pretrained: "./data/fpn_dcn_coco" 
+
   pretrained_epoch: 0
   pretrained_weights_are_priority: true
   PIXEL_MEANS:
@@ -73,7 +73,7 @@ dataset:
   max_num_extra_classes: 122
   dataset: PascalVOC;ImageNet
   dataset_path: "/dccstor/leonidka1/data/VOCdevkit;/dccstor/leonidka1/data/imagenet/ILSVRC"
-  cls_filter_files: './data/Imagenet_LOC/inloc_cls2id_map.pkl:./data/Imagenet_LOC/inloc_cls2id_map.pkl/inloc_first101_categories.txt'
+  cls_filter_files: './data/Imagenet_LOC/Pascal_inloc_cls2id_map.pkl:./data/Imagenet_LOC/Pascal_inloc_first101_categories.txt'
   image_set: 2007_trainval+2012_trainval;train_loc
   per_category_epoch_max: 0;10
   root_path: "./data"

fpn/train_end2end.py

@@ -245,6 +245,7 @@ def train_net(args, ctx, pretrained, epoch, prefix, begin_epoch, end_epoch, lr,
     lr_epoch = [float(epoch) for epoch in lr_step.split(',')]
     lr_epoch_diff = [epoch - begin_epoch for epoch in lr_epoch if epoch > begin_epoch]
     lr = base_lr * (lr_factor ** (len(lr_epoch) - len(lr_epoch_diff)))
+#    lr_iters = [int(epoch * len(roidb) / batch_size) for epoch in lr_epoch_diff]
     lr_iters = [int(epoch * len(roidb) / batch_size) for epoch in lr_epoch_diff]
     print('lr', lr, 'lr_epoch_diff', lr_epoch_diff, 'lr_iters', lr_iters)
     lr_scheduler = WarmupMultiFactorScheduler(lr_iters, lr_factor, config.TRAIN.warmup, config.TRAIN.warmup_lr, config.TRAIN.warmup_step)