feat: validate the baseline model's detection ability

map_nav_src/r2r/data_utils.py

@@ -37,7 +37,7 @@ def construct_instrs(anno_dir, dataset, splits, tokenizer, max_instr_len=512, is
         # Split multiple instructions into separate entries
         for j, instr in enumerate(item['instructions']):
             new_item = dict(item)
-            new_item['instr_id'] = '%s_%d' % (item['path_id'], j)
+            new_item['instr_id'] = '%s_%d' % (item['id'], j)
             new_item['instruction'] = instr
             new_item['instr_encoding'] = item['instr_encodings'][j][:max_instr_len]
             del new_item['instructions']

map_nav_src/r2r/main_nav.py

@@ -58,12 +58,12 @@ def build_dataset(args, rank=0, is_test=False):
     )
 
     # val_env_names = ['val_train_seen']
-    val_env_names = ['val_train_seen', 'val_seen', 'val_unseen']
+    val_env_names = ['val_seen', 'val_unseen']
     if args.dataset == 'r4r' and (not args.test):
         val_env_names[-1] == 'val_unseen_sampled'
     
-    if args.submit and args.dataset != 'r4r':
-        val_env_names.append('test')
+    # if args.submit and args.dataset != 'r4r':
+    #     val_env_names.append('test')
         
     val_envs = {}
     for split in val_env_names:

map_nav_src/reverie/agent_obj.py

@@ -324,7 +324,8 @@ class GMapObjectNavAgent(Seq2SeqAgent):
         ml_loss = 0.     
         og_loss = 0.   
 
-        for t in range(self.args.max_action_len):
+        # for t in range(self.args.max_action_len):
+        for t in range(1):
             for i, gmap in enumerate(gmaps):
                 if not ended[i]:
                     gmap.node_step_ids[obs[i]['viewpoint']] = t + 1

map_nav_src/reverie/env.py

@@ -339,7 +339,7 @@ class ReverieObjectNavBatch(object):
         self._next_minibatch(**kwargs)
         
         scanIds = [item['scan'] for item in self.batch]
-        viewpointIds = [item['path'][0] for item in self.batch]
+        viewpointIds = [item['path'][-1] for item in self.batch]
         headings = [item['heading'] for item in self.batch]
         self.env.newEpisodes(scanIds, viewpointIds, headings)
         return self._get_obs()
@@ -357,7 +357,7 @@ class ReverieObjectNavBatch(object):
         shortest_distances = self.shortest_distances[scan]
 
         path = sum(pred_path, [])
-        assert gt_path[0] == path[0], 'Result trajectories should include the start position'
+        # assert gt_path[0] == path[0], 'Result trajectories should include the start position'
 
         scores['action_steps'] = len(pred_path) - 1
         scores['trajectory_steps'] = len(path) - 1

map_nav_src/reverie/main_nav_obj.py

@@ -66,10 +66,10 @@ def build_dataset(args, rank=0):
         )
 
     # val_env_names = ['val_train_seen']
-    val_env_names = ['val_train_seen', 'val_seen', 'val_unseen']
+    val_env_names = ['val_seen', 'val_unseen']
 
-    if args.submit:
-        val_env_names.append('test')
+    # if args.submit:
+    #     val_env_names.append('test')
         
     val_envs = {}
     for split in val_env_names:

map_nav_src/scripts/run_r2r.sh

@@ -10,7 +10,7 @@ obj_ft_dim=768
 ngpus=1
 seed=0
 
-name=${train_alg}-${features}
+name=${train_alg}-${features}-reverie-glip-adversarial
 name=${name}-seed.${seed}
 name=${name}-init.aug.45k
 
@@ -57,11 +57,11 @@ flag="--root_dir ${DATA_ROOT}
 # train
 CUDA_VISIBLE_DEVICES='0' python r2r/main_nav.py $flag  \
       --tokenizer bert \
-      --bert_ckpt_file 'put the pretrained model (see pretrain_src) here' \
+      --bert_ckpt_file '../datasets/REVERIE/exprs_map/pretrain/cmt-vitbase-mlm.mrc.sap.og-init.lxmert-aug.speaker/ckpts/model_step_100000.pt' \
       --eval_first
 
 # test
-CUDA_VISIBLE_DEVICES='0' python r2r/main_nav.py $flag  \
-      --tokenizer bert \
-      --resume_file ../datasets/R2R/trained_models/best_val_unseen \
-      --test --submit
+# CUDA_VISIBLE_DEVICES='0' python r2r/main_nav.py $flag  \
+#       --tokenizer bert \
+#       --resume_file ../datasets/R2R/trained_models/best_val_unseen \
+#      --test --submit

map_nav_src/scripts/run_reverie.sh

@@ -10,7 +10,7 @@ obj_ft_dim=768
 ngpus=1
 seed=0
 
-name=${train_alg}-${features}
+name=${train_alg}-${features}-adversarial-but-original-model-with-glip-filter
 name=${name}-seed.${seed} #-${ngpus}gpus
 
 outdir=${DATA_ROOT}/REVERIE/exprs_map/finetune/${name}
@@ -59,13 +59,14 @@ flag="--root_dir ${DATA_ROOT}
       --gamma 0."
 
 # train
-CUDA_VISIBLE_DEVICES='0' python reverie/main_nav_obj.py $flag  \
-      --tokenizer bert \
-      --bert_ckpt_file 'put the pretrained model (see pretrain_src) here' \
-      --eval_first
+# CUDA_VISIBLE_DEVICES='0' python reverie/main_nav_obj.py $flag  \
+#       --tokenizer bert \
+#       --resume_file ../datasets/REVERIE/exprs_map/finetune/dagger-vitbase-adversarial-but-original-model-with-glip-filter-seed.0/ckpts/best_val_unseen \
+#       --bert_ckpt_file '../datasets/REVERIE/exprs_map/pretrain/cmt-vitbase-mlm.mrc.sap.og-init.lxmert-aug.speaker/ckpts/model_step_100000.pt' \
+#       --eval_first
 
 # test
 CUDA_VISIBLE_DEVICES='0' python reverie/main_nav_obj.py $flag  \
       --tokenizer bert \
-      --resume_file ../datasets/REVERIE/trained_models/best_val_unseen \
+      --resume_file ../datasets/REVERIE/exprs_map/finetune/dagger-vitbase-adversarial-but-original-model-with-glip-filter-seed.0/ckpts/best_val_unseen \
       --test --submit