Training error

First of all, thank you for publishing this work.
I would like to only retrain the Q table again by manipulating the reward function. So, I download the 100GB dataset and ran the data_phase2 module. However, because the dataset that you have provided doesn't have any *.mdb files, lmdb produces an error here because the read-only flag is true and says the no such file or directory. The path that I set is the root of this 1M record dataset.

        # Load dataset
        for full_path in glob.glob(f'{data_dir}/**'):
            txn = lmdb.open(
                full_path,
                max_readers=1, readonly=True,
                lock=False, readahead=False, meminit=False).begin(write=False)

Is there anything that I am doing wrong? Thanks.

[dotchen]

Hi,

Thank you for your interest in our project. Please refer to the FAQ on this issue.

Thanks for your answer. I'd like to actually use the generated trajectory dataset [100GB] and implement my own data loader but in a similar way to the existing data loader main_dataset.py. However, I couldn't understand a few details in this file.

1. What is the number of plans as self.T?
2. Why spds and rots and locs are plural or array in the code below? Aren't they pointing to a single recorded sample in the dataset? Based on what I see in the data.json file and RGB folder, each sample has 12 label images, 6 semantic images for 3 cameras, 6 RGB images for 3 cameras, speed, location, rotation, cmd, action0, action. Is this like we are looking for the next self.T speeds at idx? If that's the case, why self.T+1 for locs?

Thank you in advance.

def __getitem__(self, idx):
        
        if not self.multi_cam:
            idx *= len(self.camera_yaws)

        lmdb_txn = self.txn_map[idx]
        index = self.idx_map[idx]
        cam_index = self.yaw_map[idx]        

#####################This part######################
        locs = self.__class__.access('loc', lmdb_txn, index, self.T+1, dtype=np.float32)
        rots = self.__class__.access('rot', lmdb_txn, index, self.T, dtype=np.float32)
        spds = self.__class__.access('spd', lmdb_txn, index, self.T, dtype=np.float32).flatten()
        lbls = self.__class__.access('lbl', lmdb_txn, index+1, self.T, dtype=np.uint8).reshape(-1,96,96,12)
###################################################

        wide_rgb = self.__class__.access('wide_rgb_{}'.format(cam_index),  lmdb_txn, index, 1, dtype=np.uint8).reshape(240,480,3)
        wide_sem = self.__class__.access('wide_sem_{}'.format(cam_index),  lmdb_txn, index, 1, dtype=np.uint8).reshape(240,480)
        narr_rgb = self.__class__.access('narr_rgb_{}'.format(cam_index),  lmdb_txn, index, 1, dtype=np.uint8).reshape(240,384,3)
        cmd = self.__class__.access('cmd', lmdb_txn, index, 1, dtype=np.float32).flatten()

        wide_sem = filter_sem(wide_sem, self.seg_channels)

        # Crop cameras
        wide_rgb = wide_rgb[self.wide_crop_top:,:,::-1]
        wide_sem = wide_sem[self.wide_crop_top:]
        narr_rgb = narr_rgb[:-self.narr_crop_bottom,:,::-1]

        return wide_rgb, wide_sem, narr_rgb, lbls, locs, rots, spds, int(cmd)

[dotchen]

What is the number of plans as self.T?

This is the number of Bellman updates one does to get the Q/V values.

Why spds and rots and locs are plural or array in the code below?

They are array because we asked for array in the code (asked for self.T). You can get more details by looking at what the access function does.