rllab-snu / TopologicalSemanticGraphMemory

Topological Semantic Graph Memory for Image Goal Navigation (CoRL 2022 oral)
MIT License
109 stars 8 forks source link

Config file might be missing? #9

Closed Yonggie closed 1 year ago

Yonggie commented 1 year ago

Hi, exellenct work and thanks for the opensourcing.

We see two config files in ./configs/: TSGM.yaml and imggoalnav_gibson.yaml, but it seems some configs are missing.

in TSGM runner initialization:

TSGMPolicy(
            observation_space=observation_space,
            action_space=action_space,
            hidden_size=config.features.hidden_size,
            rnn_type=config.features.rnn_type,
            num_recurrent_layers=config.features.num_recurrent_layers,
            backbone=config.features.backbone,
            goal_sensor_uuid=config.TASK_CONFIG.TASK.GOAL_SENSOR_UUID,
            normalize_visual_inputs=True,
            cfg=config
        )

In the two existing yaml files, I cannot find config.features.* .

In perception.py, there are even more config that cannot find in the 2 existing yamls:

self.goal_Decoder = Attblock(cfg.transformer.hidden_dim,
                                     cfg.transformer.nheads,
                                     cfg.transformer.dim_feedforward,
                                     cfg.transformer.dropout)
        self.cGCN = CrossGCN(cfg)
        self.vis_Decoder = Attblock(cfg.transformer.hidden_dim,
                                     cfg.transformer.nheads,
                                     cfg.transformer.dim_feedforward,
                                     cfg.transformer.dropout)
        self.obj_Decoder = Attblock(cfg.features.object_feature_dim,
                                    cfg.transformer.nheads,
                                    cfg.transformer.dim_feedforward,
                                    cfg.transformer.dropout)
        self.goal_obj_Decoder = Attblock(cfg.features.object_feature_dim,
                                         cfg.transformer.nheads,
                                         cfg.transformer.dim_feedforward,
                                         cfg.transformer.dropout)

I'm just wondering where can I find a valid config, for example config.features.*, so that I can change several parameters?

Yonggie commented 1 year ago

For me, I just need a runnable ok performance config, is it available? It would very grateful if you kindly offer: yinzecheng.cuhk@gmail.com.

Yonggie commented 1 year ago

I just found more parameters in ./config/default.py! It almost have every param I need~