baselines/rae_depth.py

import os
import sys
from typing import *
import math

import click
import torch
import torch.nn.functional as F

from moge.test.baseline import MGEBaselineInterface


class Baseline(MGEBaselineInterface):

    def __init__(self, repo_path: str, config_path: str, checkpoint_path: str,
                 image_size: int, num_steps: int, use_fp16: bool, device: str = 'cuda:0'):
        super().__init__()
        repo_path = os.path.abspath(repo_path)
        src_path = os.path.join(repo_path, 'src')
        if src_path not in sys.path:
            sys.path.insert(0, src_path)

        from omegaconf import OmegaConf
        from stage2.transport import create_transport, Sampler
        from utils.model_utils import instantiate_from_config

        # Load config
        full_cfg = OmegaConf.load(config_path)
        rae_config = full_cfg.get("stage_1", None)
        model_config = full_cfg.get("stage_2", None)
        transport_config = full_cfg.get("transport", {})
        sampler_config = full_cfg.get("sampler", {})
        misc_config = full_cfg.get("misc", {})

        transport_cfg = OmegaConf.to_container(transport_config, resolve=True) if transport_config else {}
        sampler_cfg = OmegaConf.to_container(sampler_config, resolve=True) if sampler_config else {}
        misc = OmegaConf.to_container(misc_config, resolve=True) if misc_config else {}

        latent_size = tuple(int(dim) for dim in misc.get("latent_size", (768, 32, 32)))
        shift_dim = misc.get("time_dist_shift_dim", math.prod(latent_size))
        shift_base = misc.get("time_dist_shift_base", 4096)
        time_dist_shift = math.sqrt(shift_dim / shift_base)

        # Load RAE (DepthRAE)
        rae = instantiate_from_config(rae_config).to(device)
        rae.eval()

        # Load Stage-2 model
        model = instantiate_from_config(model_config).to(device)

        # Load checkpoint
        checkpoint = torch.load(checkpoint_path, map_location='cpu', weights_only=False)
        if 'ema' in checkpoint:
            state_dict = checkpoint['ema']
        elif 'model' in checkpoint:
            state_dict = checkpoint['model']
        else:
            state_dict = checkpoint
        model.load_state_dict(state_dict)
        model.eval()

        # Create transport sampler
        transport_params = dict(transport_cfg.get("params", {}))
        transport_params.pop("time_dist_shift", None)
        transport = create_transport(**transport_params, time_dist_shift=time_dist_shift)
        transport_sampler = Sampler(transport)

        sampler_mode = sampler_cfg.get("mode", "ODE").upper()
        sampler_params = dict(sampler_cfg.get("params", {}))
        sampler_params['num_steps'] = num_steps

        if sampler_mode == "ODE":
            eval_sampler = transport_sampler.sample_ode(**sampler_params)
        else:
            eval_sampler = transport_sampler.sample_sde(**sampler_params)

        self.rae = rae
        self.model = model
        self.eval_sampler = eval_sampler
        self.latent_size = latent_size
        self.image_size = image_size
        self.device = torch.device(device)
        self.use_fp16 = use_fp16

    @click.command()
    @click.option('--repo', 'repo_path', type=str, default='/home/ywan0794/RAE')
    @click.option('--rae_config', 'config_path', type=str, required=True)
    @click.option('--checkpoint', 'checkpoint_path', type=str, required=True)
    @click.option('--image_size', type=int, default=512)
    @click.option('--num_steps', type=int, default=2)
    @click.option('--fp16', 'use_fp16', is_flag=True)
    @click.option('--device', type=str, default='cuda:0')
    @staticmethod
    def load(repo_path, config_path, checkpoint_path, image_size, num_steps, use_fp16, device):
        return Baseline(repo_path, config_path, checkpoint_path, image_size, num_steps, use_fp16, device)

    def _predict_depth(self, image: torch.FloatTensor):
        original_height, original_width = image.shape[-2:]

        if image.ndim == 3:
            image = image.unsqueeze(0)
            omit_batch_dim = True
        else:
            omit_batch_dim = False

        b = image.shape[0]

        # Resize to model input size
        image_resized = F.interpolate(
            image, size=(self.image_size, self.image_size),
            mode='bilinear', align_corners=False, antialias=True,
        )

        # Encode RGB
        z_rgb = self.rae.encode(image_resized)

        # Sample depth from noise
        z_noise = torch.randn(b, *self.latent_size, device=self.device)
        y = torch.zeros(b, dtype=torch.long, device=self.device)

        # Marigold-style: concat z_rgb with xt before passing to model
        def model_fn(xt, t, y):
            x_input = torch.cat([xt, z_rgb], dim=1)
            return self.model(x_input, t, y)

        # Run diffusion sampling
        z_pred = self.eval_sampler(z_noise, model_fn, y=y)[-1]

        # Decode to depth (pass z_rgb for conditioning)
        depth_pred = self.rae.decode(z_pred.float(), z_rgb)
        depth_pred = depth_pred.mean(dim=1)  # (B, H, W)

        # Resize back to original size
        depth_pred = F.interpolate(
            depth_pred.unsqueeze(1), size=(original_height, original_width),
            mode='bilinear', align_corners=False,
        )[:, 0]

        if omit_batch_dim:
            depth_pred = depth_pred.squeeze(0)

        return depth_pred

    @torch.inference_mode()
    def infer(self, image: torch.FloatTensor, intrinsics: Optional[torch.FloatTensor] = None):
        depth_pred = self._predict_depth(image)
        return {
            'depth_affine_invariant': depth_pred,
        }

    @torch.inference_mode()
    def infer_for_evaluation(self, image: torch.FloatTensor, intrinsics: torch.FloatTensor = None):
        with torch.cuda.amp.autocast(enabled=self.use_fp16, dtype=torch.float16):
            depth_pred = self._predict_depth(image)
        return {
            'depth_affine_invariant': depth_pred,
        }