# note for davis dataloader later: temporally consistent depth estimator: https://github.com/yu-li/TCMonoDepth
# note for cool idea of not even downloading data and just streaming from youtube:https://gist.github.com/Mxhmovd/41e7690114e7ddad8bcd761a76272cc3
import matplotlib.pyplot as plt; 
import cv2
import os
import multiprocessing as mp
import torch.nn.functional as F
import torch
import random
import imageio
import numpy as np
from glob import glob
from collections import defaultdict
from pdb import set_trace as pdb
from itertools import combinations
from random import choice
import matplotlib.pyplot as plt
import imageio.v3 as iio

from torchvision import transforms

import sys

from glob import glob
import os
import gzip
import json
import numpy as np

from PIL import Image


from einops import rearrange, repeat
ch_sec = lambda x: rearrange(x,"... c x y -> ... (x y) c")
hom = lambda x, i=-1: torch.cat((x, torch.ones_like(x.unbind(i)[0].unsqueeze(i))), i)

class SUNRGBD(torch.utils.data.Dataset):
    """Dataset for a class of objects, where each datapoint is a SceneInstanceDataset."""

    def __init__(
        self,
        overfit_size=11000,
    ):
        self.overfit_size=overfit_size
        self.depth_paths=sorted(glob("/data/MYSUN/depth/*.png"))
        self.img_paths=[x.replace("depth","image").replace(".png",".jpg") for x in self.depth_paths]

        train_idx=int(len(self.depth_paths)*.9)
        self.depth_paths_train,self.img_paths_train=self.depth_paths[:train_idx],self.img_paths[:train_idx]
        self.depth_paths_test,self.img_paths_test=self.depth_paths[train_idx:],self.img_paths[train_idx:]

    def train(self):
        self.depth_paths,self.img_paths=self.depth_paths_train,self.img_paths_train
    def val(self):
        self.depth_paths,self.img_paths=self.depth_paths_test,self.img_paths_test

    def __len__(self):
        return min(4000,self.overfit_size)
        return len(self.depth_paths)

    def collate_fn(batch):
      return {
          'pixel_values': torch.stack([x['pixel_values'] for x in batch]),
          'labels': torch.tensor([x['labels'] for x in batch])
    }

    def __getitem__(self, idx):
        depth=torch.from_numpy(plt.imread(self.depth_paths[idx])).float()
        depthmask=(depth!=0).float()
        rgb=torch.from_numpy(plt.imread(self.img_paths[idx])/255).permute(2,0,1).float()

        res=(256,256)
        depth=F.interpolate(depth[None,None],res,antialias=True,mode="bilinear")[0,0]
        depthmask=F.interpolate(depthmask[None,None],res,mode="bilinear")[0,0]
        depthmask=(depthmask==1).float()
        rgb=F.interpolate(rgb[None],res,antialias=True,mode="bilinear")[0]

        # if self.val: idx=idx+len(self.train_depths) # so don't have to use separate embedding tables
    
        return {"rgb":rgb,"depth":depth,"depth_mask":depthmask,"idx":torch.tensor([idx])}