training on gqa

src/base_config.py

@@ -144,7 +144,7 @@ def parse_args_and_set_config(__c=__C, blacklist=None):
             if allowed_type is bool:
 
                 def allowed_type(x):
-                    return bool(parsing.strtobool(x))
+                    return bool(parsing.str_to_bool(x))
 
             parser.add_argument(
                 f"--{arg}",

src/config.py

@@ -3,7 +3,7 @@
 from base_config import __C, parse_args_and_set_config, edict, _to_values_only
 
 
-parse_bool = lambda x: bool(parsing.strtobool(x))
+parse_bool = lambda x: bool(parsing.str_to_bool(x))
 
 if torch.cuda.is_available():
     __C.orig_dir = (
@@ -26,7 +26,7 @@
 
 __C.train.num_plot_samples = (32, edict(type=int))
 __C.train.augment = (False, edict(type=parse_bool))
-__C.train.dataset = ("orig", edict(choices=["orig", "uni"]))
+__C.train.dataset = ("orig", edict(choices=["orig", "uni", "gqa"]))
 __C.train.gradient_clip_val = (8.0, edict(type=float))
 
 __C.train.optimizers = edict()
@@ -37,7 +37,7 @@
 __C.model.encoder = edict()
 __C.model.encoder.type = (
     "resnet50",
-    edict(choices=["scratch", "pretrained", "resnet50"]),
+    edict(choices=["none", "scratch", "pretrained", "resnet50", "resnet101"]),
 )
 __C.model.encoder.ckpt_fp = ("", edict(type=str))
 __C.model.encoder.out_nc = 512
@@ -48,16 +48,16 @@
     init_mem="random",
     max_step=(12, edict(type=int)),
     separate_syntax_semantics=False,
-    use_feats="spatial",
+    use_feats=("spatial", edict(type=str, choices=["spatial", "objects", "pixels"])),
     num_gt_lobs=0,
     common=edict(
-        module_dim=512,
+        module_dim=(512, edict(type=int)),
         # use_feats='spatial',
     ),
     input_unit=edict(
-        in_channels=512,
+        in_channels=(512, edict(type=int)),
         wordvec_dim=300,
-        rnn_dim=512,
+        rnn_dim=(512, edict(type=int)),
         bidirectional=True,
         separate_syntax_semantics_embeddings=False,
         stem_act="ELU",