[Fix] delete the useless load_from_pretrained flag.

internnav/configs/trainer/il.py

@@ -26,7 +26,6 @@ class IlCfg(BaseModel, extra='allow'):
     save_filter_frozen_weights: Optional[bool] = None
     load_from_ckpt: Optional[bool]
     ckpt_to_load: Optional[str]
-    load_from_pretrain: Optional[bool]
     dataset_r2r_root_dir: Optional[str] = None
     dataset_3dgs_root_dir: Optional[str] = None
     dataset_grutopia10_root_dir: Optional[str] = None

internnav/model/__init__.py

@@ -8,18 +8,17 @@
 
 from .basemodel.cma.cma_clip_policy import CMA_CLIP_Net, CMACLIPModelConfig
 from .basemodel.cma.cma_policy import CMAModelConfig, CMANet
-from .basemodel.rdp.rdp_policy import RDPNet, RDPModelConfig
+from .basemodel.internvla_n1.internvla_n1_policy import (
+    InternVLAN1ModelConfig,
+    InternVLAN1Net,
+)
 from .basemodel.navdp.navdp_policy import NavDPModelConfig, NavDPNet
-
-from .basemodel.seq2seq.seq2seq_policy import Seq2SeqNet, Seq2SeqModelConfig
-
-from .basemodel.internvla_n1.internvla_n1_policy import InternVLAN1Net, InternVLAN1ModelConfig
-
+from .basemodel.rdp.rdp_policy import RDPModelConfig, RDPNet
+from .basemodel.seq2seq.seq2seq_policy import Seq2SeqModelConfig, Seq2SeqNet
 from .utils.misc import set_cuda, set_random_seed, wrap_model
 from .utils.save import load_checkpoint
 
 
-
 def get_policy(policy_name):
     if policy_name == 'CMA_CLIP_Policy':
         return CMA_CLIP_Net
@@ -52,128 +51,3 @@ def get_config(policy_name):
         return NavDPModelConfig
     else:
         raise ValueError(f'Policy {policy_name} not found')
-
-
-def initialize_policy(
-    config,
-    device=None,
-    action_stats=None,
-) -> None:
-    from internnav.utils.common_log_util import common_logger as logger  # 延迟导入
-
-    load_from_ckpt = config.il.load_from_ckpt
-    load_from_pretrain = config.il.load_from_pretrain
-
-    default_gpu, n_gpu, device = set_cuda(config, device)
-    if default_gpu:
-        logger.info(
-            'device: {} n_gpu: {}, distributed training: {}'.format(device, n_gpu, bool(config.local_rank != -1))
-        )
-
-    seed = config.seed
-    if config.ddp.use:
-        seed += config.local_rank
-    set_random_seed(seed)
-
-    # if default_gpu:
-    #     save_training_meta(config)
-
-    observation_space = spaces.Box(
-        low=0.0,
-        high=1.0,
-        shape=(256, 256, 1),
-        dtype=np.float32,
-    )
-
-    policy = get_policy(config.model.policy_name)
-
-    self_policy = policy(
-        config=config.model,
-        observation_space=observation_space,
-    )
-
-    if load_from_pretrain:
-        new_ckpt_weights = {}
-        model_config = config.model
-        self_policy.load_state_dict(new_ckpt_weights, strict=False)
-
-    start_epoch = 0
-    if load_from_ckpt:
-        ckpt_path = config.il.ckpt_to_load
-        ckpt_dict = load_checkpoint(ckpt_path, map_location='cpu')
-        if 'state_dict' in ckpt_dict:
-            state_dict = ckpt_dict['state_dict']
-        else:
-            state_dict = ckpt_dict
-        if 'epoch' in ckpt_dict:
-            start_epoch = ckpt_dict['epoch']
-        new_state_dict = {}
-        # Iterate through the state dictionary items
-        for k, v in state_dict.items():
-            new_key = k.replace('module.', '')
-            if config.model.policy_name != 'RDP_Policy':
-                new_key = new_key.replace('net.', '')  # this is for cma policy
-            new_state_dict[new_key] = v
-        del state_dict[k]  # Remove the old key with 'module.'
-
-        incompatible_keys, _ = self_policy.load_state_dict(new_state_dict, strict=False)
-        if len(incompatible_keys) > 0:
-            logger.warning(f'Incompatible keys: {incompatible_keys}')
-        logger.info(f'Loaded weights from checkpoint: {ckpt_path}')
-
-    params = sum(param.numel() for param in self_policy.parameters())
-    params_t = sum(p.numel() for p in self_policy.parameters() if p.requires_grad)
-    logger.info(f'Agent parameters: {params / 1e6:.2f}M. Trainable: {params_t / 1e6:.2f}M')
-    logger.info('Finished setting up policy.')
-
-    if len(config.torch_gpu_ids) == 1:
-        config.ddp.use = False
-    if config.ddp.use:
-        if config.ddp.use_dp:
-            # Data parallel
-            self_policy = wrap_model(
-                self_policy,
-                config.torch_gpu_ids,
-                config.local_rank,
-                logger,
-                config.world_size,
-                use_dp=config.ddp.use_dp,
-            )
-        else:
-            # Distributed data parallel
-            self_policy = wrap_model(
-                self_policy,
-                torch.device(f'cuda:{config.local_rank}'),
-                config.local_rank,
-                logger,
-                config.world_size,
-                use_dp=config.ddp.use_dp,
-            )
-    else:
-        self_policy.to(device)
-
-    optimizer = torch.optim.Adam(self_policy.parameters(), lr=float(config.il.lr))
-
-    lr_scheduler = None
-    if config.il.lr_schedule.use:
-        if config.il.lr_schedule.type == 'cosine':
-            lr_scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
-                optimizer,
-                T_max=config.il.epochs,
-                eta_min=float(config.il.lr_schedule.min_lr),
-            )
-        elif config.il.lr_schedule.type == 'step':
-            lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(
-                optimizer,
-                milestones=config.il.lr_schedule.decay_epochs,
-                gamma=float(config.il.lr_schedule.decay_factor),
-            )
-        elif config.il.lr_schedule.type == 'linear':
-            lr_scheduler = torch.optim.lr_scheduler.LinearLR(
-                optimizer,
-                start_factor=config.il.lr_schedule.warmup_factor,
-                end_factor=1.0,
-                total_iters=config.il.lr_schedule.warmup_epochs,
-            )
-
-    return self_policy, optimizer, lr_scheduler, start_epoch

scripts/train/configs/challenge_train_kujiale_cfg.py

@@ -39,16 +39,15 @@
         warmup_ratio=0.1,
         save_filter_frozen_weights=True,
         load_from_ckpt=False,
-        ckpt_to_load='checkpoints/r2r/fine_tuned/rdp/checkpoint-104150',
-        load_from_pretrain=True,
+        ckpt_to_load='',
         dataset_r2r_root_dir='interiornav_data/raw_data/',
         dataset_3dgs_root_dir='',
         dataset_grutopia10_root_dir='',
         lmdb_features_dir='r2r',
         lerobot_features_dir='data/vln_pe/traj_data/interior',
         camera_name='pano_camera_0',
         report_to='wandb',  # wandb, tensorboard, none
-        ddp_find_unused_parameters = True,
+        ddp_find_unused_parameters=True,
         filter_failure=FilterFailure(
             use=True,
             min_rgb_nums=15,

scripts/train/configs/challenge_train_mp3d_cfg.py

@@ -40,15 +40,14 @@
         save_filter_frozen_weights=True,
         load_from_ckpt=False,
         ckpt_to_load='',
-        load_from_pretrain=True,
         dataset_r2r_root_dir='data/vln_pe/raw_data/r2r',
         dataset_3dgs_root_dir='',
         dataset_grutopia10_root_dir='',
         lmdb_features_dir='r2r',
         lerobot_features_dir='data/vln_pe/traj_data/r2r',
         camera_name='pano_camera_0',
         report_to='wandb',  # wandb, tensorboard, none
-        ddp_find_unused_parameters = True,
+        ddp_find_unused_parameters=True,
         filter_failure=FilterFailure(
             use=True,
             min_rgb_nums=15,

scripts/train/configs/cma.py

@@ -30,14 +30,13 @@
         batch_size=2,
         lr=1e-4,
         num_workers=8,
-        weight_decay=1e-5, 
-        warmup_ratio=0.05,  
+        weight_decay=1e-5,
+        warmup_ratio=0.05,
         use_iw=True,
         inflection_weight_coef=3.2,
         save_filter_frozen_weights=False,
         load_from_ckpt=False,
         ckpt_to_load='',
-        load_from_pretrain=False,
         lmdb_map_size=1e12,
         dataset_r2r_root_dir='data/vln_pe/raw_data/r2r',
         dataset_3dgs_root_dir='',
@@ -46,7 +45,7 @@
         lerobot_features_dir='data/vln_pe/traj_data/r2r',
         camera_name='pano_camera_0',
         report_to='wandb',  # wandb, tensorboard, none
-        ddp_find_unused_parameters = True,
+        ddp_find_unused_parameters=True,
         filter_failure=FilterFailure(
             use=True,
             min_rgb_nums=15,

scripts/train/configs/cma_plus.py

@@ -30,14 +30,13 @@
         batch_size=2,
         lr=1e-4,
         num_workers=8,
-        weight_decay=1e-5, 
-        warmup_ratio=0.05,  
+        weight_decay=1e-5,
+        warmup_ratio=0.05,
         use_iw=True,
         inflection_weight_coef=3.2,
         save_filter_frozen_weights=False,
         load_from_ckpt=False,
         ckpt_to_load='checkpoints/r2r/zero_shot/cma',
-        load_from_pretrain=False,
         lmdb_map_size=1e12,
         dataset_r2r_root_dir='data/vln_pe/raw_data/r2r',
         dataset_3dgs_root_dir='',
@@ -46,7 +45,7 @@
         lerobot_features_dir='data/vln_pe/traj_data/r2r',
         camera_name='pano_camera_0',
         report_to='wandb',  # wandb, tensorboard, none
-        ddp_find_unused_parameters = True,
+        ddp_find_unused_parameters=True,
         filter_failure=FilterFailure(
             use=True,
             min_rgb_nums=15,

scripts/train/configs/navdp.py

@@ -2,7 +2,6 @@
 from internnav.configs.trainer.eval import EvalCfg
 from internnav.configs.trainer.exp import ExpCfg
 from internnav.configs.trainer.il import FilterFailure, IlCfg, Loss
-import os
 
 navdp_exp_cfg = ExpCfg(
     name='navdp_train',
@@ -14,7 +13,7 @@
     tensorboard_dir='checkpoints/%s/tensorboard',
     checkpoint_folder='checkpoints/%s/ckpts',
     log_dir='checkpoints/%s/logs',
-    local_rank= 0,
+    local_rank=0,
     # device = None,
     seed=0,
     eval=EvalCfg(
@@ -41,7 +40,6 @@
         save_filter_frozen_weights=False,
         load_from_ckpt=False,
         ckpt_to_load='',
-        load_from_pretrain=False,
         lmdb_map_size=1e12,
         dataset_r2r_root_dir='data/vln_pe/raw_data/r2r',
         dataset_3dgs_root_dir='',
@@ -50,8 +48,8 @@
         lerobot_features_dir='data/vln_pe/traj_data/r2r',
         camera_name='pano_camera_0',
         report_to='tensorboard',  # wandb, tensorboard, none
-        dataset_navdp = 'data/datasets/navdp_dataset_lerobot.json',
-        root_dir = 'data/datasets/InternData-N1/vln_n1/traj_data',
+        dataset_navdp='data/datasets/navdp_dataset_lerobot.json',
+        root_dir='data/datasets/InternData-N1/vln_n1/traj_data',
         image_size=224,
         scene_scale=1.0,
         preload=False,
@@ -66,7 +64,7 @@
         dropout=0.1,
         scratch=False,
         finetune=False,
-        ddp_find_unused_parameters = True,
+        ddp_find_unused_parameters=True,
         filter_failure=FilterFailure(
             use=True,
             min_rgb_nums=15,

scripts/train/configs/rdp.py

@@ -40,15 +40,14 @@
         save_filter_frozen_weights=True,
         load_from_ckpt=False,
         ckpt_to_load='',
-        load_from_pretrain=True,
         dataset_r2r_root_dir='data/vln_pe/raw_data/r2r',
         dataset_3dgs_root_dir='',
         dataset_grutopia10_root_dir='',
         lmdb_features_dir='r2r',
         lerobot_features_dir='data/vln_pe/traj_data/r2r',
         camera_name='pano_camera_0',
         report_to='wandb',  # wandb, tensorboard, none
-        ddp_find_unused_parameters = True,
+        ddp_find_unused_parameters=True,
         filter_failure=FilterFailure(
             use=True,
             min_rgb_nums=15,

scripts/train/configs/seq2seq.py

@@ -37,7 +37,6 @@
         save_filter_frozen_weights=False,
         load_from_ckpt=False,
         ckpt_to_load='',
-        load_from_pretrain=True,
         lmdb_map_size=1e12,
         dataset_r2r_root_dir='data/vln_pe/raw_data/r2r',
         dataset_3dgs_root_dir='',
@@ -46,7 +45,7 @@
         lerobot_features_dir='data/vln_pe/traj_data/r2r',
         camera_name='pano_camera_0',
         report_to='wandb',  # wandb, tensorboard, none
-        ddp_find_unused_parameters = True,
+        ddp_find_unused_parameters=True,
         filter_failure=FilterFailure(
             use=True,
             min_rgb_nums=15,

scripts/train/configs/seq2seq_plus.py

@@ -37,7 +37,6 @@
         save_filter_frozen_weights=False,
         load_from_ckpt=False,
         ckpt_to_load='checkpoints/r2r/zero_shot/seq2seq',
-        load_from_pretrain=True,
         lmdb_map_size=1e12,
         dataset_r2r_root_dir='data/vln_pe/raw_data/r2r',
         dataset_3dgs_root_dir='',
@@ -46,7 +45,7 @@
         lerobot_features_dir='data/vln_pe/traj_data/r2r',
         camera_name='pano_camera_0',
         report_to='wandb',  # wandb, tensorboard, none
-        ddp_find_unused_parameters = True,
+        ddp_find_unused_parameters=True,
         filter_failure=FilterFailure(
             use=True,
             min_rgb_nums=15,