Add files using upload-large-folder tool

outdoor_v48_4gpu_v2/code/05_02-14:21:58/mytrain.py

@@ -0,0 +1,601 @@
+# --------------------------------------------------------
+# training code for CUT3R
+# --------------------------------------------------------
+# References:
+# DUSt3R: https://github.com/naver/dust3r
+# --------------------------------------------------------
+import argparse
+import datetime
+import json
+import numpy as np
+import os
+import sys
+import time
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Sized
+from itertools import islice
+
+import torch
+import torch.backends.cudnn as cudnn
+import torch.nn.functional as F
+from torch.utils.tensorboard import SummaryWriter
+
+torch.backends.cuda.matmul.allow_tf32 = True  # for gpu >= Ampere and pytorch >= 1.12
+
+from dust3r.model import (
+    PreTrainedModel,
+    ARCroco3DStereo,
+    ARCroco3DStereoConfig,
+    inf,
+    strip_module,
+)  # noqa: F401, needed when loading the model
+from dust3r.datasets import get_data_loader
+from dust3r.losses_noteacher  import *  # noqa: F401, needed when loading the model
+from dust3r.inference import loss_of_one_batch  # noqa
+from dust3r.viz import colorize
+from dust3r.utils.render import get_render_results
+import dust3r.utils.path_to_croco  # noqa: F401
+import croco.utils.misc as misc  # noqa
+from croco.utils.misc import NativeScalerWithGradNormCount as NativeScaler  # noqa
+
+import hydra
+from omegaconf import OmegaConf
+import logging
+import pathlib
+from tqdm import tqdm
+import random
+import builtins
+import shutil
+
+from accelerate import Accelerator
+from accelerate import DistributedDataParallelKwargs, InitProcessGroupKwargs
+from accelerate.logging import get_logger
+from datetime import timedelta
+import torch.multiprocessing
+
+from slamformer.models.slamformer import SLAMFormer  # upstream typo: pi3 → slamformer
+
+
+torch.multiprocessing.set_sharing_strategy("file_system")
+
+printer = get_logger(__name__, log_level="DEBUG")
+
+
+def setup_for_distributed(accelerator: Accelerator):
+    """
+    This function disables printing when not in master process
+    """
+    builtin_print = builtins.print
+
+    def print(*args, **kwargs):
+        force = kwargs.pop("force", False)
+        force = force or (accelerator.num_processes > 8)
+        if accelerator.is_main_process or force:
+            now = datetime.datetime.now().time()
+            builtin_print("[{}] ".format(now), end="")  # print with time stamp
+            builtin_print(*args, **kwargs)
+
+    builtins.print = print
+
+
+def save_current_code(outdir):
+    now = datetime.datetime.now()  # current date and time
+    date_time = now.strftime("%m_%d-%H:%M:%S")
+    src_dir = "."
+    dst_dir = os.path.join(outdir, "code", "{}".format(date_time))
+    shutil.copytree(
+        src_dir,
+        dst_dir,
+        ignore=shutil.ignore_patterns(
+            ".vscode*",
+            "assets*",
+            "example*",
+            "checkpoints*",
+            "OLD*",
+            "logs*",
+            "out*",
+            "runs*",
+            "*.png",
+            "*.mp4",
+            "*__pycache__*",
+            "*.git*",
+            "*.idea*",
+            "*.zip",
+            "*.jpg",
+        ),
+        dirs_exist_ok=True,
+    )
+    return dst_dir
+
+
+def train(args):
+
+    accelerator = Accelerator(
+        gradient_accumulation_steps=args.accum_iter,
+        mixed_precision="bf16",
+        kwargs_handlers=[
+            DistributedDataParallelKwargs(find_unused_parameters=True),
+            InitProcessGroupKwargs(timeout=timedelta(seconds=6000)),
+        ],
+    )
+    device = accelerator.device
+
+    setup_for_distributed(accelerator)
+
+    printer.info("output_dir: " + args.output_dir)
+    if args.output_dir:
+        Path(args.output_dir).mkdir(parents=True, exist_ok=True)
+
+    if accelerator.is_main_process:
+        dst_dir = save_current_code(outdir=args.output_dir)
+        printer.info(f"Saving current code to {dst_dir}")
+
+    # auto resume
+    if not args.resume:
+        last_ckpt_fname = os.path.join(args.output_dir, f"checkpoint-last.pth")
+        #last_ckpt_fname = os.path.join(args.output_dir, f"checkpoint-7.pth")
+
+        args.resume = last_ckpt_fname if os.path.isfile(last_ckpt_fname) else None
+
+    printer.info("job dir: {}".format(os.path.dirname(os.path.realpath(__file__))))
+
+    # fix the seed
+    seed = args.seed + accelerator.state.process_index
+    printer.info(
+        f"Setting seed to {seed} for process {accelerator.state.process_index}"
+    )
+    torch.manual_seed(seed)
+    np.random.seed(seed)
+    random.seed(seed)
+    cudnn.benchmark = args.benchmark
+
+    # training dataset and loader
+    printer.info("Building train dataset %s", args.train_dataset)
+    #  dataset and loader
+    data_loader_train = build_dataset(
+        args.train_dataset,
+        args.batch_size,
+        args.num_workers,
+        accelerator=accelerator,
+        test=False,
+        fixed_length=args.fixed_length
+    )
+    printer.info("Building test dataset %s", args.test_dataset)
+    data_loader_test = {
+        dataset.split("(")[0]: build_dataset(
+            dataset,
+            args.batch_size,
+            args.num_workers,
+            accelerator=accelerator,
+            test=True,
+            fixed_length=True
+        )
+        for dataset in args.test_dataset.split("+")
+    }
+
+    # model
+    printer.info("Loading model")
+    model = SLAMFormer()
+    teacher = None
+
+    # model: PreTrainedModel = eval(args.model)
+    printer.info(f"All model parameters: {sum(p.numel() for p in model.parameters())}")
+
+
+    printer.info(f">> Creating train criterion = {args.train_criterion}")
+    train_criterion = eval(args.train_criterion).to(device)
+    printer.info(
+        f">> Creating test criterion = {args.test_criterion or args.train_criterion}"
+    )
+    test_criterion = eval(args.test_criterion or args.criterion).to(device)
+
+    model.to(device)
+
+    if args.gradient_checkpointing:
+        model.gradient_checkpointing_enable()
+    if args.long_context:
+        model.fixed_input_length = False
+
+    freeze_keys = None
+    print('NOTE:', args.pretrained, args.resume)
+    if args.pretrained and not args.resume:
+        printer.info(f"Loading pretrained: {args.pretrained}")
+        ckpt = torch.load(args.pretrained, map_location=device)
+        '''
+        ckpt_ = dict()
+        for key, v in ckpt.items():
+            ckpt_[key[7:]] = v
+        '''
+        '''
+        freeze_keys = list(ckpt.keys())
+
+        ls = dict()
+        for key, v in ckpt.items():
+            if 'aggregator' in key:
+                key_ = key.replace('aggregator', 'backend_transformer')
+                ls[key_] = key
+        for key_ in ls.keys():
+            key = ls[key_]
+            ckpt[key_] = ckpt[key]
+        '''
+        printer.info(
+            model.load_state_dict(ckpt, strict=False)
+        )
+        del ckpt# in case it occupies memory
+    '''
+    if freeze_keys is None:
+        freeze_keys = []
+
+        for name, param in model.named_parameters(): 
+            if 'backend_transformer' not in name:
+                freeze_keys.append(name)
+    '''
+    '''
+    printer.info("Loading teacher model")
+    ckpt_teacher = torch.load(args.teacher, map_location=device)
+    teacher.load_state_dict(ckpt_teacher, strict=True)
+    teacher = teacher.to("cuda")
+    for p in teacher.parameters():
+        p.requires_grad = False  
+    teacher.eval()
+    del ckpt_teacher
+
+    '''
+    # freeze
+    printer.info("Freezing patch embedding and positional encoding parameters...")
+    frozen_params = 0
+    total_params = 0
+
+    frozen_param_names = []
+
+    for name, param in model.named_parameters():
+        total_params += param.numel()
+        param.requires_grad = True
+
+    if hasattr(model, 'encoder'):# and hasattr(model.aggregator, 'patch_embed'):
+        for param in model.encoder.parameters():#aggregator.patch_embed.parameters():
+            if param.requires_grad:
+                param.requires_grad = False
+
+    if hasattr(model, 'register_token'):
+        model.register_token.requires_grad = False
+
+    # YIJUN: Skip the freezekeys
+    '''
+    for name, param in model.named_parameters():
+        if 'camera_decoder' in name or 'camera_head' in name:
+            print(name)
+            param.requires_grad = False
+    '''
+
+    for name, p in model.named_parameters():
+        if not p.requires_grad:
+            frozen_params += p.numel()
+            frozen_param_names.append(name)
+
+    printer.info(
+        f"Frozen {frozen_params:,} parameters out of {total_params:,} total parameters. ({frozen_params / total_params:.2%})")
+    printer.info(
+        f"Trainable parameters: {total_params - frozen_params:,} ({(total_params - frozen_params) / total_params:.2%})")
+    if frozen_param_names:
+        printer.info(
+            f"Example frozen parameters: {', '.join(frozen_param_names[:5])}{'...' if len(frozen_param_names) > 5 else ''}")
+
+
+
+    # following timm: set wd as 0 for bias and norm layers
+    param_groups = misc.get_parameter_groups(model, args.weight_decay)
+    optimizer = torch.optim.AdamW(param_groups, lr=args.lr, betas=(0.9, 0.95))
+    # print(optimizer)
+    loss_scaler = NativeScaler(accelerator=accelerator)
+
+    best_so_far = misc.load_model(
+        args=args, model_without_ddp=model, optimizer=optimizer, loss_scaler=loss_scaler
+    )
+    if best_so_far is None:
+        best_so_far = float("inf")
+
+    accelerator.even_batches = False
+    optimizer, model, data_loader_train = accelerator.prepare(
+        optimizer, model, data_loader_train
+    )
+
+    def write_log_stats(epoch, train_stats, test_stats):
+        if accelerator.is_main_process:
+            if log_writer is not None:
+                log_writer.flush()
+
+            log_stats = dict(
+                epoch=epoch, **{f"train_{k}": v for k, v in train_stats.items()}
+            )
+            for test_name in data_loader_test:
+                if test_name not in test_stats:
+                    continue
+                log_stats.update(
+                    {test_name + "_" + k: v for k, v in test_stats[test_name].items()}
+                )
+
+            with open(
+                os.path.join(args.output_dir, "log.txt"), mode="a", encoding="utf-8"
+            ) as f:
+                f.write(json.dumps(log_stats) + "\n")
+
+    def save_model(epoch, fname, best_so_far, data_iter_step):
+        misc.save_model(
+            accelerator=accelerator,
+            args=args,
+            model_without_ddp=model,
+            optimizer=optimizer,
+            loss_scaler=loss_scaler,
+            epoch=epoch,
+            step=data_iter_step,
+            fname=fname,
+            best_so_far=best_so_far,
+        )
+
+    log_writer = (
+        SummaryWriter(log_dir=args.output_dir) if accelerator.is_main_process else None
+    )
+
+    printer.info(f"Start training for {args.epochs} epochs")
+    start_time = time.time()
+    train_stats = test_stats = {}
+
+    for epoch in range(args.start_epoch, args.epochs + 1):
+
+        # Save immediately the last checkpoint
+        if epoch > args.start_epoch:
+            if (
+                args.save_freq
+                and np.allclose(epoch / args.save_freq, int(epoch / args.save_freq))
+                or epoch == args.epochs
+            ):
+                save_model(epoch - 1, "last", best_so_far, args.start_step)
+
+        new_best = False
+
+        if epoch > args.start_epoch:
+            if args.keep_freq and epoch % args.keep_freq == 0:
+                save_model(epoch - 1, str(epoch), best_so_far, args.start_step)
+            if new_best:
+                save_model(epoch - 1, "best", best_so_far, args.start_step)
+        if epoch >= args.epochs:
+            break  # exit after writing last test to disk
+
+
+        # Train
+        train_stats = train_one_epoch(
+            model,
+            teacher,
+            train_criterion,
+            data_loader_train,
+            optimizer,
+            accelerator,
+            epoch,
+            loss_scaler,
+            log_writer=log_writer,
+            args=args
+        )
+
+
+    total_time = time.time() - start_time
+    total_time_str = str(datetime.timedelta(seconds=int(total_time)))
+    printer.info("Training time {}".format(total_time_str))
+
+    save_final_model(accelerator, args, args.epochs, model, best_so_far=best_so_far)
+
+
+def save_final_model(accelerator, args, epoch, model_without_ddp, best_so_far=None):
+    output_dir = Path(args.output_dir)
+    checkpoint_path = output_dir / "checkpoint-final.pth"
+    to_save = {
+        "args": args,
+        "model": (
+            model_without_ddp
+            if isinstance(model_without_ddp, dict)
+            else model_without_ddp.cpu().state_dict()
+        ),
+        "epoch": epoch,
+    }
+    if best_so_far is not None:
+        to_save["best_so_far"] = best_so_far
+    printer.info(f">> Saving model to {checkpoint_path} ...")
+    misc.save_on_master(accelerator, to_save, checkpoint_path)
+
+
+def build_dataset(dataset, batch_size, num_workers, accelerator, test=False, fixed_length=False):
+    split = ["Train", "Test"][test]
+    printer.info(f"Building {split} Data loader for dataset: {dataset}")
+    loader = get_data_loader(
+        dataset,
+        batch_size=batch_size,
+        num_workers=num_workers,
+        pin_mem=True,
+        shuffle=not (test),
+        drop_last=not (test),
+        accelerator=accelerator,
+        fixed_length=fixed_length
+    )
+    return loader
+
+
+def train_one_epoch(
+    model: torch.nn.Module,
+    teacher: torch.nn.Module,
+    criterion: torch.nn.Module,
+    data_loader: Sized,
+    optimizer: torch.optim.Optimizer,
+    accelerator: Accelerator,
+    epoch: int,
+    loss_scaler,
+    args,
+    log_writer=None,
+):
+    assert torch.backends.cuda.matmul.allow_tf32 == True
+
+    model.train(True)
+    metric_logger = misc.MetricLogger(delimiter="  ")
+    metric_logger.add_meter("lr", misc.SmoothedValue(window_size=1, fmt="{value:.6f}"))
+    header = "Epoch: [{}]".format(epoch)
+    accum_iter = args.accum_iter
+
+    def save_model(epoch, fname, best_so_far, data_iter_step):
+        unwrapped_model = accelerator.unwrap_model(model)
+        misc.save_model(
+            accelerator=accelerator,
+            args=args,
+            model_without_ddp=unwrapped_model,
+            optimizer=optimizer,
+            loss_scaler=loss_scaler,
+            epoch=epoch,
+            step=data_iter_step,
+            fname=fname,
+            best_so_far=best_so_far,
+        )
+
+    if log_writer is not None:
+        printer.info("log_dir: {}".format(log_writer.log_dir))
+
+    if hasattr(data_loader, "dataset") and hasattr(data_loader.dataset, "set_epoch"):
+        data_loader.dataset.set_epoch(epoch)
+    if (
+        hasattr(data_loader, "batch_sampler")
+        and hasattr(data_loader.batch_sampler, "batch_sampler")
+        and hasattr(data_loader.batch_sampler.batch_sampler, "set_epoch")
+    ):
+        data_loader.batch_sampler.batch_sampler.set_epoch(epoch)
+
+
+    optimizer.zero_grad()
+
+    start_step = args.start_step
+
+    data_iter = metric_logger.log_every(data_loader, args.print_freq, accelerator, header)
+
+    for data_iter_step, batch in enumerate(data_iter):
+            
+        with accelerator.accumulate(model):
+            # change the range of the image to [0, 1]
+            if isinstance(batch, dict) and "img" in batch:
+                batch["img"] = (batch["img"] + 1.0) / 2.0
+            elif isinstance(batch, list) and all(isinstance(v, dict) and "img" in v for v in batch):
+                for view in batch:
+                    view["img"] = (view["img"] + 1.0) / 2.0
+
+            epoch_f = epoch + data_iter_step / len(data_loader)
+            # we use a per iteration (instead of per epoch) lr scheduler
+            if data_iter_step % accum_iter == 0:
+                misc.adjust_learning_rate(optimizer, epoch_f, args)
+
+            epoch_f = epoch + data_iter_step / len(data_loader)
+            step = int(epoch_f * len(data_loader))
+
+            result = loss_of_one_batch(
+                batch,
+                model,
+                criterion,
+                accelerator,
+                teacher=teacher,
+                inference=False,
+                symmetrize_batch=False,
+                use_amp=bool(args.amp),
+            )
+      
+            loss, loss_details = result["loss"]  # criterion returns two values
+
+            loss_value = float(loss)
+
+            if not math.isfinite(loss_value):
+                print(
+                    f"Loss is {loss_value}, stopping training, loss details: {loss_details}"
+                )
+                sys.exit(1)
+            if not result.get("already_backprop", False):
+                loss_scaler(
+                    loss,
+                    optimizer,
+                    parameters=model.parameters(),
+                    update_grad=True,
+                    clip_grad=1.0,
+                )
+                optimizer.zero_grad()
+
+            is_metric = batch[0]["is_metric"]
+            curr_num_view = len(batch)
+
+            del loss
+            
+            tb_vis_img = (data_iter_step + 1) % accum_iter == 0 and (
+                (step + 1) % (args.print_img_freq)
+            ) == 0
+            if not tb_vis_img:
+                del batch
+            else:
+                torch.cuda.empty_cache()
+
+            lr = optimizer.param_groups[0]["lr"]
+            metric_logger.update(epoch=epoch_f)
+            metric_logger.update(lr=lr)
+            metric_logger.update(step=step)
+            #
+            metric_logger.update(loss=loss_value, **loss_details)
+            #
+            if (data_iter_step + 1) % accum_iter == 0 and (
+                (data_iter_step + 1) % (accum_iter * args.print_freq)
+            ) == 0:
+                loss_value_reduce = accelerator.gather(
+                    torch.tensor(loss_value).to(accelerator.device)
+                ).mean()  # MUST BE EXECUTED BY ALL NODES
+
+                if log_writer is None:
+                    continue
+                """ We use epoch_1000x as the x-axis in tensorboard.
+                This calibrates different curves when batch size changes.
+                """
+                epoch_1000x = int(epoch_f * 1000)
+                log_writer.add_scalar("train_loss", loss_value_reduce, step)
+                log_writer.add_scalar("train_lr", lr, step)
+                log_writer.add_scalar("train_iter", epoch_1000x, step)
+                for name, val in loss_details.items():
+                    if isinstance(val, torch.Tensor):
+                        if val.ndim > 0:
+                            continue
+                    if isinstance(val, dict):
+                        continue
+                    log_writer.add_scalar("train_" + name, val, step)
+
+        if (
+            data_iter_step % int(args.save_freq * len(data_loader)) == 0
+            and data_iter_step != 0
+            and data_iter_step != len(data_loader) - 1
+        ):
+            print("saving at step", data_iter_step)
+            save_model(epoch - 1, "last", float("inf"), data_iter_step)
+
+    # gather the stats from all processes
+    metric_logger.synchronize_between_processes(accelerator)
+    printer.info("Averaged stats: %s", metric_logger)
+    return {k: meter.global_avg for k, meter in metric_logger.meters.items()}
+
+def batch_append(original_list, new_list):
+    for sublist, new_item in zip(original_list, new_list):
+        sublist.append(new_item)
+    return original_list
+
+
+@hydra.main(
+    version_base=None,
+    config_path=str(os.path.dirname(os.path.abspath(__file__))) + "/../config",
+    config_name="mytrain.yaml",
+)
+def run(cfg: OmegaConf):
+    OmegaConf.resolve(cfg)
+    logdir = pathlib.Path(cfg.logdir)
+    logdir.mkdir(parents=True, exist_ok=True)
+    train(cfg)
+
+
+if __name__ == "__main__":
+    run()

outdoor_v48_4gpu_v2/code/05_02-14:21:58/streamvggt/heads/camera_head.py

@@ -0,0 +1,175 @@
+import math
+import numpy as np
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from streamvggt.layers import Mlp
+from streamvggt.layers.block import Block
+from streamvggt.heads.head_act import activate_pose
+
+
+class CameraHead(nn.Module):
+    def __init__(
+        self,
+        dim_in: int = 2048,
+        trunk_depth: int = 4,
+        pose_encoding_type: str = "absT_quaR_FoV",
+        num_heads: int = 16,
+        mlp_ratio: int = 4,
+        init_values: float = 0.01,
+        trans_act: str = "linear",
+        quat_act: str = "linear",
+        fl_act: str = "relu",  # Field of view activations: ensures FOV values are positive.
+    ):
+        super().__init__()
+
+        if pose_encoding_type == "absT_quaR_FoV":
+            self.target_dim = 9
+        else:
+            raise ValueError(f"Unsupported camera encoding type: {pose_encoding_type}")
+
+        self.trans_act = trans_act
+        self.quat_act = quat_act
+        self.fl_act = fl_act
+        self.trunk_depth = trunk_depth
+
+        # Build the trunk using a sequence of transformer blocks.
+        self.trunk = nn.Sequential(
+            *[
+                Block(
+                    dim=dim_in,
+                    num_heads=num_heads,
+                    mlp_ratio=mlp_ratio,
+                    init_values=init_values,
+                )
+                for _ in range(trunk_depth)
+            ]
+        )
+
+        # Normalizations for camera token and trunk output.
+        self.token_norm = nn.LayerNorm(dim_in)
+        self.trunk_norm = nn.LayerNorm(dim_in)
+
+        # Learnable empty camera pose token.
+        self.empty_pose_tokens = nn.Parameter(torch.zeros(1, 1, self.target_dim))
+        self.embed_pose = nn.Linear(self.target_dim, dim_in)
+
+        # Module for producing modulation parameters: shift, scale, and a gate.
+        self.poseLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(dim_in, 3 * dim_in, bias=True))
+
+        # Adaptive layer normalization without affine parameters.
+        self.adaln_norm = nn.LayerNorm(dim_in, elementwise_affine=False, eps=1e-6)
+        self.pose_branch = Mlp(
+            in_features=dim_in,
+            hidden_features=dim_in // 2,
+            out_features=self.target_dim,
+            drop=0,
+        )
+
+    def forward(self, aggregated_tokens_list: list, num_iterations: int = 4, past_key_values_camera = None, use_cache: bool = False) -> list:
+        """
+        Forward pass to predict camera parameters.
+
+        Args:
+            aggregated_tokens_list (list): List of token tensors from the network;
+                the last tensor is used for prediction.
+            num_iterations (int, optional): Number of iterative refinement steps. Defaults to 4.
+
+        Returns:
+            list: A list of predicted camera encodings (post-activation) from each iteration.
+        """
+        # Use tokens from the last block for camera prediction.
+        tokens = aggregated_tokens_list[-1]
+
+        # Extract the camera tokens
+        pose_tokens = tokens[:, :, 0]
+        pose_tokens = self.token_norm(pose_tokens)
+
+        if use_cache:
+            pred_pose_enc_list, past_key_values_camera = self.trunk_fn(pose_tokens, num_iterations, past_key_values_camera, use_cache)
+            return pred_pose_enc_list, past_key_values_camera
+        else:
+            pred_pose_enc_list = self.trunk_fn(pose_tokens, num_iterations, past_key_values_camera=None, use_cache=use_cache)
+            return pred_pose_enc_list
+
+    def trunk_fn(self, pose_tokens: torch.Tensor, num_iterations: int, past_key_values_camera, use_cache: bool) -> list:
+        """
+        Iteratively refine camera pose predictions.
+
+        Args:
+            pose_tokens (torch.Tensor): Normalized camera tokens with shape [B, 1, C].
+            num_iterations (int): Number of refinement iterations.
+
+        Returns:
+            list: List of activated camera encodings from each iteration.
+        """
+        B, S, C = pose_tokens.shape  # S is expected to be 1.
+        pred_pose_enc = None
+        pred_pose_enc_list = []
+
+        for _ in range(num_iterations):
+            # Use a learned empty pose for the first iteration.
+            if pred_pose_enc is None:
+                module_input = self.embed_pose(self.empty_pose_tokens.expand(B, S, -1))
+            else:
+                # Detach the previous prediction to avoid backprop through time.
+                pred_pose_enc = pred_pose_enc.detach()
+                module_input = self.embed_pose(pred_pose_enc)
+
+            # Generate modulation parameters and split them into shift, scale, and gate components.
+            shift_msa, scale_msa, gate_msa = self.poseLN_modulation(module_input).chunk(3, dim=-1)
+
+            # Adaptive layer normalization and modulation.
+            pose_tokens_modulated = gate_msa * modulate(self.adaln_norm(pose_tokens), shift_msa, scale_msa)
+            pose_tokens_modulated = pose_tokens_modulated + pose_tokens
+
+            if not use_cache:
+                L = S * 1
+                frame_ids = torch.arange(L, device=pose_tokens_modulated.device) // 1  # [0,0,...,1,1,...,S-1]
+                future_frame = frame_ids.unsqueeze(1) < frame_ids.unsqueeze(0)
+                attn_mask = future_frame.to(pose_tokens_modulated.dtype) * torch.finfo(pose_tokens_modulated.dtype).min
+            else:
+                attn_mask = None
+
+            if use_cache:
+                for idx in range(self.trunk_depth):
+                    pose_tokens_modulated, block_kv = self.trunk[idx](
+                        pose_tokens_modulated,
+                        attn_mask=attn_mask,
+                        past_key_values=past_key_values_camera[idx] if past_key_values_camera[idx] is not None else None,
+                        use_cache=True
+                    )
+                    past_key_values_camera[idx] = block_kv
+            else:
+                for idx in range(self.trunk_depth):
+                    pose_tokens_modulated = self.trunk[idx](pose_tokens_modulated, attn_mask=attn_mask)
+                    
+            # Compute the delta update for the pose encoding.
+            pred_pose_enc_delta = self.pose_branch(self.trunk_norm(pose_tokens_modulated))
+
+            if pred_pose_enc is None:
+                pred_pose_enc = pred_pose_enc_delta
+            else:
+                pred_pose_enc = pred_pose_enc + pred_pose_enc_delta
+
+            # Apply final activation functions for translation, quaternion, and field-of-view.
+            activated_pose = activate_pose(
+                pred_pose_enc,
+                trans_act=self.trans_act,
+                quat_act=self.quat_act,
+                fl_act=self.fl_act,
+            )
+            pred_pose_enc_list.append(activated_pose)
+            
+        if use_cache:
+            return pred_pose_enc_list, past_key_values_camera
+        return pred_pose_enc_list
+
+
+def modulate(x: torch.Tensor, shift: torch.Tensor, scale: torch.Tensor) -> torch.Tensor:
+    """
+    Modulate the input tensor using scaling and shifting parameters.
+    """
+    return x * (1 + scale) + shift

outdoor_v48_4gpu_v2/code/05_02-14:21:58/streamvggt/heads/dpt_head.py

@@ -0,0 +1,471 @@
+import os
+from typing import List, Dict, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from .head_act import activate_head
+from .utils import create_uv_grid, position_grid_to_embed
+
+
+class DPTHead(nn.Module):
+    """
+    Args:
+        dim_in (int): Input dimension (channels).
+        patch_size (int, optional): Patch size. Default is 14.
+        output_dim (int, optional): Number of output channels. Default is 4.
+        activation (str, optional): Activation type. Default is "inv_log".
+        conf_activation (str, optional): Confidence activation type. Default is "expp1".
+        features (int, optional): Feature channels for intermediate representations. Default is 256.
+        out_channels (List[int], optional): Output channels for each intermediate layer.
+        intermediate_layer_idx (List[int], optional): Indices of layers from aggregated tokens used for DPT.
+        pos_embed (bool, optional): Whether to use positional embedding. Default is True.
+        feature_only (bool, optional): If True, return features only without the last several layers and activation head. Default is False.
+        down_ratio (int, optional): Downscaling factor for the output resolution. Default is 1.
+    """
+
+    def __init__(
+        self,
+        dim_in: int,
+        patch_size: int = 14,
+        output_dim: int = 4,
+        activation: str = "inv_log",
+        conf_activation: str = "expp1",
+        features: int = 256,
+        out_channels: List[int] = [256, 512, 1024, 1024],
+        intermediate_layer_idx: List[int] = [4, 11, 17, 23],
+        pos_embed: bool = True,
+        feature_only: bool = False,
+        down_ratio: int = 1,
+    ) -> None:
+        super(DPTHead, self).__init__()
+        self.patch_size = patch_size
+        self.activation = activation
+        self.conf_activation = conf_activation
+        self.pos_embed = pos_embed
+        self.feature_only = feature_only
+        self.down_ratio = down_ratio
+        self.intermediate_layer_idx = intermediate_layer_idx
+
+        self.norm = nn.LayerNorm(dim_in)
+
+        # Projection layers for each output channel from tokens.
+        self.projects = nn.ModuleList(
+            [
+                nn.Conv2d(
+                    in_channels=dim_in,
+                    out_channels=oc,
+                    kernel_size=1,
+                    stride=1,
+                    padding=0,
+                )
+                for oc in out_channels
+            ]
+        )
+
+        # Resize layers for upsampling feature maps.
+        self.resize_layers = nn.ModuleList(
+            [
+                nn.ConvTranspose2d(
+                    in_channels=out_channels[0], out_channels=out_channels[0], kernel_size=4, stride=4, padding=0
+                ),
+                nn.ConvTranspose2d(
+                    in_channels=out_channels[1], out_channels=out_channels[1], kernel_size=2, stride=2, padding=0
+                ),
+                nn.Identity(),
+                nn.Conv2d(
+                    in_channels=out_channels[3], out_channels=out_channels[3], kernel_size=3, stride=2, padding=1
+                ),
+            ]
+        )
+
+        self.scratch = _make_scratch(
+            out_channels,
+            features,
+            expand=False,
+        )
+
+        # Attach additional modules to scratch.
+        self.scratch.stem_transpose = None
+        self.scratch.refinenet1 = _make_fusion_block(features)
+        self.scratch.refinenet2 = _make_fusion_block(features)
+        self.scratch.refinenet3 = _make_fusion_block(features)
+        self.scratch.refinenet4 = _make_fusion_block(features, has_residual=False)
+
+        head_features_1 = features
+        head_features_2 = 32
+
+        if feature_only:
+            self.scratch.output_conv1 = nn.Conv2d(head_features_1, head_features_1, kernel_size=3, stride=1, padding=1)
+        else:
+            self.scratch.output_conv1 = nn.Conv2d(
+                head_features_1, head_features_1 // 2, kernel_size=3, stride=1, padding=1
+            )
+            conv2_in_channels = head_features_1 // 2
+
+            self.scratch.output_conv2 = nn.Sequential(
+                nn.Conv2d(conv2_in_channels, head_features_2, kernel_size=3, stride=1, padding=1),
+                nn.ReLU(inplace=True),
+                nn.Conv2d(head_features_2, output_dim, kernel_size=1, stride=1, padding=0),
+            )
+
+    def forward(
+        self,
+        aggregated_tokens_list: List[torch.Tensor],
+        images: torch.Tensor,
+        patch_start_idx: int,
+        frames_chunk_size: int = 8,
+    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        """
+        Forward pass through the DPT head, supports processing by chunking frames.
+        Args:
+            aggregated_tokens_list (List[Tensor]): List of token tensors from different transformer layers.
+            images (Tensor): Input images with shape [B, S, 3, H, W], in range [0, 1].
+            patch_start_idx (int): Starting index for patch tokens in the token sequence.
+                Used to separate patch tokens from other tokens (e.g., camera or register tokens).
+            frames_chunk_size (int, optional): Number of frames to process in each chunk.
+                If None or larger than S, all frames are processed at once. Default: 8.
+
+        Returns:
+            Tensor or Tuple[Tensor, Tensor]:
+                - If feature_only=True: Feature maps with shape [B, S, C, H, W]
+                - Otherwise: Tuple of (predictions, confidence) both with shape [B, S, 1, H, W]
+        """
+        B, S, _, H, W = images.shape
+
+        # If frames_chunk_size is not specified or greater than S, process all frames at once
+        if frames_chunk_size is None or frames_chunk_size >= S:
+            return self._forward_impl(aggregated_tokens_list, images, patch_start_idx)
+
+        # Otherwise, process frames in chunks to manage memory usage
+        assert frames_chunk_size > 0
+
+        # Process frames in batches
+        all_preds = []
+        all_conf = []
+
+        for frames_start_idx in range(0, S, frames_chunk_size):
+            frames_end_idx = min(frames_start_idx + frames_chunk_size, S)
+
+            # Process batch of frames
+            if self.feature_only:
+                chunk_output = self._forward_impl(
+                    aggregated_tokens_list, images, patch_start_idx, frames_start_idx, frames_end_idx
+                )
+                all_preds.append(chunk_output)
+            else:
+                chunk_preds, chunk_conf = self._forward_impl(
+                    aggregated_tokens_list, images, patch_start_idx, frames_start_idx, frames_end_idx
+                )
+                all_preds.append(chunk_preds)
+                all_conf.append(chunk_conf)
+
+        # Concatenate results along the sequence dimension
+        if self.feature_only:
+            return torch.cat(all_preds, dim=1)
+        else:
+            return torch.cat(all_preds, dim=1), torch.cat(all_conf, dim=1)
+
+    def _forward_impl(
+        self,
+        aggregated_tokens_list: List[torch.Tensor],
+        images: torch.Tensor,
+        patch_start_idx: int,
+        frames_start_idx: int = None,
+        frames_end_idx: int = None,
+    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        """
+        Args:
+            aggregated_tokens_list (List[Tensor]): List of token tensors from different transformer layers.
+            images (Tensor): Input images with shape [B, S, 3, H, W].
+            patch_start_idx (int): Starting index for patch tokens.
+            frames_start_idx (int, optional): Starting index for frames to process.
+            frames_end_idx (int, optional): Ending index for frames to process.
+
+        Returns:
+            Tensor or Tuple[Tensor, Tensor]: Feature maps or (predictions, confidence).
+        """
+        if frames_start_idx is not None and frames_end_idx is not None:
+            images = images[:, frames_start_idx:frames_end_idx].contiguous()
+
+        B, S, _, H, W = images.shape
+
+        patch_h, patch_w = H // self.patch_size, W // self.patch_size
+
+        out = []
+        dpt_idx = 0
+
+        for layer_idx in self.intermediate_layer_idx:
+            x = aggregated_tokens_list[layer_idx][:, :, patch_start_idx:]
+
+            # Select frames if processing a chunk
+            if frames_start_idx is not None and frames_end_idx is not None:
+                x = x[:, frames_start_idx:frames_end_idx]
+
+            x = x.reshape(B * S, -1, x.shape[-1])
+
+            x = self.norm(x)
+            x = x.permute(0, 2, 1).reshape((x.shape[0], x.shape[-1], patch_h, patch_w))
+
+            x = self.projects[dpt_idx](x)
+            if self.pos_embed:
+                x = self._apply_pos_embed(x, W, H)
+            x = self.resize_layers[dpt_idx](x)
+
+            out.append(x)
+            dpt_idx += 1
+
+        # Fuse features from multiple layers.
+        out = self.scratch_forward(out)
+        # Interpolate fused output to match target image resolution.
+        out = custom_interpolate(
+            out,
+            (int(patch_h * self.patch_size / self.down_ratio), int(patch_w * self.patch_size / self.down_ratio)),
+            mode="bilinear",
+            align_corners=True,
+        )
+
+        if self.pos_embed:
+            out = self._apply_pos_embed(out, W, H)
+
+        if self.feature_only:
+            return out.reshape(B, S, *out.shape[1:])
+
+        out = self.scratch.output_conv2(out)
+        preds, conf = activate_head(out, activation=self.activation, conf_activation=self.conf_activation)
+
+        preds = preds.reshape(B, S, *preds.shape[1:])
+        conf = conf.reshape(B, S, *conf.shape[1:])
+        return preds, conf
+
+    def _apply_pos_embed(self, x: torch.Tensor, W: int, H: int, ratio: float = 0.1) -> torch.Tensor:
+        """
+        Apply positional embedding to tensor x.
+        """
+        patch_w = x.shape[-1]
+        patch_h = x.shape[-2]
+        pos_embed = create_uv_grid(patch_w, patch_h, aspect_ratio=W / H, dtype=x.dtype, device=x.device)
+        pos_embed = position_grid_to_embed(pos_embed, x.shape[1])
+        pos_embed = pos_embed * ratio
+        pos_embed = pos_embed.permute(2, 0, 1)[None].expand(x.shape[0], -1, -1, -1)
+        return x + pos_embed
+
+    def scratch_forward(self, features: List[torch.Tensor]) -> torch.Tensor:
+        """
+        Forward pass through the fusion blocks.
+
+        Args:
+            features (List[Tensor]): List of feature maps from different layers.
+
+        Returns:
+            Tensor: Fused feature map.
+        """
+        layer_1, layer_2, layer_3, layer_4 = features
+
+        layer_1_rn = self.scratch.layer1_rn(layer_1)
+        layer_2_rn = self.scratch.layer2_rn(layer_2)
+        layer_3_rn = self.scratch.layer3_rn(layer_3)
+        layer_4_rn = self.scratch.layer4_rn(layer_4)
+
+        out = self.scratch.refinenet4(layer_4_rn, size=layer_3_rn.shape[2:])
+        del layer_4_rn, layer_4
+
+        out = self.scratch.refinenet3(out, layer_3_rn, size=layer_2_rn.shape[2:])
+        del layer_3_rn, layer_3
+
+        out = self.scratch.refinenet2(out, layer_2_rn, size=layer_1_rn.shape[2:])
+        del layer_2_rn, layer_2
+
+        out = self.scratch.refinenet1(out, layer_1_rn)
+        del layer_1_rn, layer_1
+
+        out = self.scratch.output_conv1(out)
+        return out
+
+
+def _make_fusion_block(features: int, size: int = None, has_residual: bool = True, groups: int = 1) -> nn.Module:
+    return FeatureFusionBlock(
+        features,
+        nn.ReLU(inplace=True),
+        deconv=False,
+        bn=False,
+        expand=False,
+        align_corners=True,
+        size=size,
+        has_residual=has_residual,
+        groups=groups,
+    )
+
+
+def _make_scratch(in_shape: List[int], out_shape: int, groups: int = 1, expand: bool = False) -> nn.Module:
+    scratch = nn.Module()
+    out_shape1 = out_shape
+    out_shape2 = out_shape
+    out_shape3 = out_shape
+    if len(in_shape) >= 4:
+        out_shape4 = out_shape
+
+    if expand:
+        out_shape1 = out_shape
+        out_shape2 = out_shape * 2
+        out_shape3 = out_shape * 4
+        if len(in_shape) >= 4:
+            out_shape4 = out_shape * 8
+
+    scratch.layer1_rn = nn.Conv2d(
+        in_shape[0], out_shape1, kernel_size=3, stride=1, padding=1, bias=False, groups=groups
+    )
+    scratch.layer2_rn = nn.Conv2d(
+        in_shape[1], out_shape2, kernel_size=3, stride=1, padding=1, bias=False, groups=groups
+    )
+    scratch.layer3_rn = nn.Conv2d(
+        in_shape[2], out_shape3, kernel_size=3, stride=1, padding=1, bias=False, groups=groups
+    )
+    if len(in_shape) >= 4:
+        scratch.layer4_rn = nn.Conv2d(
+            in_shape[3], out_shape4, kernel_size=3, stride=1, padding=1, bias=False, groups=groups
+        )
+    return scratch
+
+
+class ResidualConvUnit(nn.Module):
+    """Residual convolution module."""
+
+    def __init__(self, features, activation, bn, groups=1):
+        """Init.
+
+        Args:
+            features (int): number of features
+        """
+        super().__init__()
+
+        self.bn = bn
+        self.groups = groups
+        self.conv1 = nn.Conv2d(features, features, kernel_size=3, stride=1, padding=1, bias=True, groups=self.groups)
+        self.conv2 = nn.Conv2d(features, features, kernel_size=3, stride=1, padding=1, bias=True, groups=self.groups)
+
+        self.norm1 = None
+        self.norm2 = None
+
+        self.activation = activation
+        self.skip_add = nn.quantized.FloatFunctional()
+
+    def forward(self, x):
+        """Forward pass.
+
+        Args:
+            x (tensor): input
+
+        Returns:
+            tensor: output
+        """
+
+        out = self.activation(x)
+        out = self.conv1(out)
+        if self.norm1 is not None:
+            out = self.norm1(out)
+
+        out = self.activation(out)
+        out = self.conv2(out)
+        if self.norm2 is not None:
+            out = self.norm2(out)
+
+        return self.skip_add.add(out, x)
+
+
+class FeatureFusionBlock(nn.Module):
+    """Feature fusion block."""
+
+    def __init__(
+        self,
+        features,
+        activation,
+        deconv=False,
+        bn=False,
+        expand=False,
+        align_corners=True,
+        size=None,
+        has_residual=True,
+        groups=1,
+    ):
+        """Init.
+
+        Args:
+            features (int): number of features
+        """
+        super(FeatureFusionBlock, self).__init__()
+
+        self.deconv = deconv
+        self.align_corners = align_corners
+        self.groups = groups
+        self.expand = expand
+        out_features = features
+        if self.expand == True:
+            out_features = features // 2
+
+        self.out_conv = nn.Conv2d(
+            features, out_features, kernel_size=1, stride=1, padding=0, bias=True, groups=self.groups
+        )
+
+        if has_residual:
+            self.resConfUnit1 = ResidualConvUnit(features, activation, bn, groups=self.groups)
+
+        self.has_residual = has_residual
+        self.resConfUnit2 = ResidualConvUnit(features, activation, bn, groups=self.groups)
+
+        self.skip_add = nn.quantized.FloatFunctional()
+        self.size = size
+
+    def forward(self, *xs, size=None):
+        """Forward pass.
+
+        Returns:
+            tensor: output
+        """
+        output = xs[0]
+
+        if self.has_residual:
+            res = self.resConfUnit1(xs[1])
+            output = self.skip_add.add(output, res)
+
+        output = self.resConfUnit2(output)
+
+        if (size is None) and (self.size is None):
+            modifier = {"scale_factor": 2}
+        elif size is None:
+            modifier = {"size": self.size}
+        else:
+            modifier = {"size": size}
+
+        output = custom_interpolate(output, **modifier, mode="bilinear", align_corners=self.align_corners)
+        output = self.out_conv(output)
+
+        return output
+
+
+def custom_interpolate(
+    x: torch.Tensor,
+    size: Tuple[int, int] = None,
+    scale_factor: float = None,
+    mode: str = "bilinear",
+    align_corners: bool = True,
+) -> torch.Tensor:
+    """
+    Custom interpolate to avoid INT_MAX issues in nn.functional.interpolate.
+    """
+    if size is None:
+        size = (int(x.shape[-2] * scale_factor), int(x.shape[-1] * scale_factor))
+
+    INT_MAX = 1610612736
+
+    input_elements = size[0] * size[1] * x.shape[0] * x.shape[1]
+
+    if input_elements > INT_MAX:
+        chunks = torch.chunk(x, chunks=(input_elements // INT_MAX) + 1, dim=0)
+        interpolated_chunks = [
+            nn.functional.interpolate(chunk, size=size, mode=mode, align_corners=align_corners) for chunk in chunks
+        ]
+        x = torch.cat(interpolated_chunks, dim=0)
+        return x.contiguous()
+    else:
+        return nn.functional.interpolate(x, size=size, mode=mode, align_corners=align_corners)

outdoor_v48_4gpu_v2/code/05_02-14:21:58/streamvggt/heads/head_act.py

@@ -0,0 +1,116 @@
+import torch
+import torch.nn.functional as F
+
+
+def activate_pose(pred_pose_enc, trans_act="linear", quat_act="linear", fl_act="linear"):
+    """
+    Args:
+        pred_pose_enc: Tensor containing encoded pose parameters [translation, quaternion, focal length]
+        trans_act: Activation type for translation component
+        quat_act: Activation type for quaternion component
+        fl_act: Activation type for focal length component
+
+    Returns:
+        Activated pose parameters tensor
+    """
+    T = pred_pose_enc[..., :3]
+    quat = pred_pose_enc[..., 3:7]
+    fl = pred_pose_enc[..., 7:]  # or fov
+
+    T = base_pose_act(T, trans_act)
+    quat = base_pose_act(quat, quat_act)
+    fl = base_pose_act(fl, fl_act)  # or fov
+
+    pred_pose_enc = torch.cat([T, quat, fl], dim=-1)
+
+    return pred_pose_enc
+
+
+def base_pose_act(pose_enc, act_type="linear"):
+    """
+    Apply basic activation function to pose parameters.
+
+    Args:
+        pose_enc: Tensor containing encoded pose parameters
+        act_type: Activation type ("linear", "inv_log", "exp", "relu")
+
+    Returns:
+        Activated pose parameters
+    """
+    if act_type == "linear":
+        return pose_enc
+    elif act_type == "inv_log":
+        return inverse_log_transform(pose_enc)
+    elif act_type == "exp":
+        return torch.exp(pose_enc)
+    elif act_type == "relu":
+        return F.relu(pose_enc)
+    else:
+        raise ValueError(f"Unknown act_type: {act_type}")
+
+
+def activate_head(out, activation="norm_exp", conf_activation="expp1"):
+    """
+    Process network output to extract 3D points and confidence values.
+
+    Args:
+        out: Network output tensor (B, C, H, W)
+        activation: Activation type for 3D points
+        conf_activation: Activation type for confidence values
+
+    Returns:
+        Tuple of (3D points tensor, confidence tensor)
+    """
+    # Move channels from last dim to the 4th dimension => (B, H, W, C)
+    fmap = out.permute(0, 2, 3, 1)  # B,H,W,C expected
+
+    # Split into xyz (first C-1 channels) and confidence (last channel)
+    xyz = fmap[:, :, :, :-1]
+    conf = fmap[:, :, :, -1]
+
+    if activation == "norm_exp":
+        d = xyz.norm(dim=-1, keepdim=True).clamp(min=1e-8)
+        xyz_normed = xyz / d
+        pts3d = xyz_normed * torch.expm1(d)
+    elif activation == "norm":
+        pts3d = xyz / xyz.norm(dim=-1, keepdim=True)
+    elif activation == "exp":
+        pts3d = torch.exp(xyz)
+    elif activation == "relu":
+        pts3d = F.relu(xyz)
+    elif activation == "inv_log":
+        pts3d = inverse_log_transform(xyz)
+    elif activation == "xy_inv_log":
+        xy, z = xyz.split([2, 1], dim=-1)
+        z = inverse_log_transform(z)
+        pts3d = torch.cat([xy * z, z], dim=-1)
+    elif activation == "sigmoid":
+        pts3d = torch.sigmoid(xyz)
+    elif activation == "linear":
+        pts3d = xyz
+    else:
+        raise ValueError(f"Unknown activation: {activation}")
+
+    if conf_activation == "expp1":
+        conf_out = 1 + conf.exp()
+    elif conf_activation == "expp0":
+        conf_out = conf.exp()
+    elif conf_activation == "sigmoid":
+        conf_out = torch.sigmoid(conf)
+    else:
+        raise ValueError(f"Unknown conf_activation: {conf_activation}")
+
+    return pts3d, conf_out
+
+
+def inverse_log_transform(y):
+    """
+    Apply inverse log transform: sign(y) * (exp(|y|) - 1)
+
+    Args:
+        y: Input tensor
+
+    Returns:
+        Transformed tensor
+    """
+    return torch.sign(y) * (torch.expm1(torch.abs(y)))

outdoor_v48_4gpu_v2/code/05_02-14:21:58/streamvggt/heads/track_head.py

@@ -0,0 +1,102 @@
+import torch.nn as nn
+from .dpt_head import DPTHead
+from .track_modules.base_track_predictor import BaseTrackerPredictor
+
+
+class TrackHead(nn.Module):
+    """
+    Track head that uses DPT head to process tokens and BaseTrackerPredictor for tracking.
+    The tracking is performed iteratively, refining predictions over multiple iterations.
+    """
+
+    def __init__(
+        self,
+        dim_in,
+        patch_size=14,
+        features=128,
+        iters=4,
+        predict_conf=True,
+        stride=2,
+        corr_levels=7,
+        corr_radius=4,
+        hidden_size=384,
+    ):
+        """
+        Initialize the TrackHead module.
+
+        Args:
+            dim_in (int): Input dimension of tokens from the backbone.
+            patch_size (int): Size of image patches used in the vision transformer.
+            features (int): Number of feature channels in the feature extractor output.
+            iters (int): Number of refinement iterations for tracking predictions.
+            predict_conf (bool): Whether to predict confidence scores for tracked points.
+            stride (int): Stride value for the tracker predictor.
+            corr_levels (int): Number of correlation pyramid levels
+            corr_radius (int): Radius for correlation computation, controlling the search area.
+            hidden_size (int): Size of hidden layers in the tracker network.
+        """
+        super().__init__()
+
+        self.patch_size = patch_size
+
+        # Feature extractor based on DPT architecture
+        # Processes tokens into feature maps for tracking
+        self.feature_extractor = DPTHead(
+            dim_in=dim_in,
+            patch_size=patch_size,
+            features=features,
+            feature_only=True,  # Only output features, no activation
+            down_ratio=2,  # Reduces spatial dimensions by factor of 2
+            pos_embed=False,
+        )
+
+        # Tracker module that predicts point trajectories
+        # Takes feature maps and predicts coordinates and visibility
+        self.tracker = BaseTrackerPredictor(
+            latent_dim=features,  # Match the output_dim of feature extractor
+            predict_conf=predict_conf,
+            stride=stride,
+            corr_levels=corr_levels,
+            corr_radius=corr_radius,
+            hidden_size=hidden_size,
+        )
+
+        self.iters = iters
+
+    def forward(self, aggregated_tokens_list, images, patch_start_idx, query_points=None, iters=None):
+        """
+        Forward pass of the TrackHead.
+
+        Args:
+            aggregated_tokens_list (list): List of aggregated tokens from the backbone.
+            images (torch.Tensor): Input images of shape (B, S, C, H, W) where:
+                                   B = batch size, S = sequence length.
+            patch_start_idx (int): Starting index for patch tokens.
+            query_points (torch.Tensor, optional): Initial query points to track.
+                                                  If None, points are initialized by the tracker.
+            iters (int, optional): Number of refinement iterations. If None, uses self.iters.
+
+        Returns:
+            tuple:
+                - coord_preds (torch.Tensor): Predicted coordinates for tracked points.
+                - vis_scores (torch.Tensor): Visibility scores for tracked points.
+                - conf_scores (torch.Tensor): Confidence scores for tracked points (if predict_conf=True).
+        """
+        B, S, _, H, W = images.shape
+
+        # Extract features from tokens
+        # feature_maps has shape (B, S, C, H//2, W//2) due to down_ratio=2
+        feature_maps = self.feature_extractor(aggregated_tokens_list, images, patch_start_idx)
+
+        # Use default iterations if not specified
+        if iters is None:
+            iters = self.iters
+
+        # Perform tracking using the extracted features
+        coord_preds, vis_scores, conf_scores = self.tracker(
+            query_points=query_points,
+            fmaps=feature_maps,
+            iters=iters,
+        )
+
+        return coord_preds, vis_scores, conf_scores

outdoor_v48_4gpu_v2/code/05_02-14:21:58/streamvggt/heads/track_modules/base_track_predictor.py

@@ -0,0 +1,195 @@
+import torch
+import torch.nn as nn
+from einops import rearrange, repeat
+
+
+from .blocks import EfficientUpdateFormer, CorrBlock
+from .utils import sample_features4d, get_2d_embedding, get_2d_sincos_pos_embed
+from .modules import Mlp
+
+
+class BaseTrackerPredictor(nn.Module):
+    def __init__(
+        self,
+        stride=1,
+        corr_levels=5,
+        corr_radius=4,
+        latent_dim=128,
+        hidden_size=384,
+        use_spaceatt=True,
+        depth=6,
+        max_scale=518,
+        predict_conf=True,
+    ):
+        super(BaseTrackerPredictor, self).__init__()
+        self.stride = stride
+        self.latent_dim = latent_dim
+        self.corr_levels = corr_levels
+        self.corr_radius = corr_radius
+        self.hidden_size = hidden_size
+        self.max_scale = max_scale
+        self.predict_conf = predict_conf
+
+        self.flows_emb_dim = latent_dim // 2
+
+        self.corr_mlp = Mlp(
+            in_features=self.corr_levels * (self.corr_radius * 2 + 1) ** 2,
+            hidden_features=self.hidden_size,
+            out_features=self.latent_dim,
+        )
+
+        self.transformer_dim = self.latent_dim + self.latent_dim + self.latent_dim + 4
+
+        self.query_ref_token = nn.Parameter(torch.randn(1, 2, self.transformer_dim))
+
+        space_depth = depth if use_spaceatt else 0
+        time_depth = depth
+
+        self.updateformer = EfficientUpdateFormer(
+            space_depth=space_depth,
+            time_depth=time_depth,
+            input_dim=self.transformer_dim,
+            hidden_size=self.hidden_size,
+            output_dim=self.latent_dim + 2,
+            mlp_ratio=4.0,
+            add_space_attn=use_spaceatt,
+        )
+
+        self.fmap_norm = nn.LayerNorm(self.latent_dim)
+        self.ffeat_norm = nn.GroupNorm(1, self.latent_dim)
+
+        # A linear layer to update track feats at each iteration
+        self.ffeat_updater = nn.Sequential(nn.Linear(self.latent_dim, self.latent_dim), nn.GELU())
+
+        self.vis_predictor = nn.Sequential(nn.Linear(self.latent_dim, 1))
+
+        if predict_conf:
+            self.conf_predictor = nn.Sequential(nn.Linear(self.latent_dim, 1))
+
+    def forward(self, query_points, fmaps=None, iters=6, return_feat=False, down_ratio=1, apply_sigmoid=True):
+        """
+        query_points: B x N x 2, the number of batches, tracks, and xy
+        fmaps: B x S x C x HH x WW, the number of batches, frames, and feature dimension.
+                note HH and WW is the size of feature maps instead of original images
+        """
+        B, N, D = query_points.shape
+        B, S, C, HH, WW = fmaps.shape
+
+        assert D == 2, "Input points must be 2D coordinates"
+
+        # apply a layernorm to fmaps here
+        fmaps = self.fmap_norm(fmaps.permute(0, 1, 3, 4, 2))
+        fmaps = fmaps.permute(0, 1, 4, 2, 3)
+
+        # Scale the input query_points because we may downsample the images
+        # by down_ratio or self.stride
+        # e.g., if a 3x1024x1024 image is processed to a 128x256x256 feature map
+        # its query_points should be query_points/4
+        if down_ratio > 1:
+            query_points = query_points / float(down_ratio)
+
+        query_points = query_points / float(self.stride)
+
+        # Init with coords as the query points
+        # It means the search will start from the position of query points at the reference frames
+        coords = query_points.clone().reshape(B, 1, N, 2).repeat(1, S, 1, 1)
+
+        # Sample/extract the features of the query points in the query frame
+        query_track_feat = sample_features4d(fmaps[:, 0], coords[:, 0])
+
+        # init track feats by query feats
+        track_feats = query_track_feat.unsqueeze(1).repeat(1, S, 1, 1)  # B, S, N, C
+        # back up the init coords
+        coords_backup = coords.clone()
+
+        fcorr_fn = CorrBlock(fmaps, num_levels=self.corr_levels, radius=self.corr_radius)
+
+        coord_preds = []
+
+        # Iterative Refinement
+        for _ in range(iters):
+            # Detach the gradients from the last iteration
+            # (in my experience, not very important for performance)
+            coords = coords.detach()
+
+            fcorrs = fcorr_fn.corr_sample(track_feats, coords)
+
+            corr_dim = fcorrs.shape[3]
+            fcorrs_ = fcorrs.permute(0, 2, 1, 3).reshape(B * N, S, corr_dim)
+            fcorrs_ = self.corr_mlp(fcorrs_)
+
+            # Movement of current coords relative to query points
+            flows = (coords - coords[:, 0:1]).permute(0, 2, 1, 3).reshape(B * N, S, 2)
+
+            flows_emb = get_2d_embedding(flows, self.flows_emb_dim, cat_coords=False)
+
+            # (In my trials, it is also okay to just add the flows_emb instead of concat)
+            flows_emb = torch.cat([flows_emb, flows / self.max_scale, flows / self.max_scale], dim=-1)
+
+            track_feats_ = track_feats.permute(0, 2, 1, 3).reshape(B * N, S, self.latent_dim)
+
+            # Concatenate them as the input for the transformers
+            transformer_input = torch.cat([flows_emb, fcorrs_, track_feats_], dim=2)
+
+            # 2D positional embed
+            pos_embed = get_2d_sincos_pos_embed(self.transformer_dim, grid_size=(HH, WW)).to(query_points.device)
+            sampled_pos_emb = sample_features4d(pos_embed.expand(B, -1, -1, -1), coords[:, 0])
+
+            sampled_pos_emb = rearrange(sampled_pos_emb, "b n c -> (b n) c").unsqueeze(1)
+
+            x = transformer_input + sampled_pos_emb
+
+            # Add the query ref token to the track feats
+            query_ref_token = torch.cat(
+                [self.query_ref_token[:, 0:1], self.query_ref_token[:, 1:2].expand(-1, S - 1, -1)], dim=1
+            )
+            x = x + query_ref_token.to(x.device).to(x.dtype)
+
+            # B, N, S, C
+            x = rearrange(x, "(b n) s d -> b n s d", b=B)
+
+            # Compute the delta coordinates and delta track features
+            delta, _ = self.updateformer(x)
+
+            # BN, S, C
+            delta = rearrange(delta, " b n s d -> (b n) s d", b=B)
+            delta_coords_ = delta[:, :, :2]
+            delta_feats_ = delta[:, :, 2:]
+
+            track_feats_ = track_feats_.reshape(B * N * S, self.latent_dim)
+            delta_feats_ = delta_feats_.reshape(B * N * S, self.latent_dim)
+
+            # Update the track features
+            track_feats_ = self.ffeat_updater(self.ffeat_norm(delta_feats_)) + track_feats_
+
+            track_feats = track_feats_.reshape(B, N, S, self.latent_dim).permute(0, 2, 1, 3)  # BxSxNxC
+
+            # B x S x N x 2
+            coords = coords + delta_coords_.reshape(B, N, S, 2).permute(0, 2, 1, 3)
+
+            # Force coord0 as query
+            # because we assume the query points should not be changed
+            coords[:, 0] = coords_backup[:, 0]
+
+            # The predicted tracks are in the original image scale
+            if down_ratio > 1:
+                coord_preds.append(coords * self.stride * down_ratio)
+            else:
+                coord_preds.append(coords * self.stride)
+
+        # B, S, N
+        vis_e = self.vis_predictor(track_feats.reshape(B * S * N, self.latent_dim)).reshape(B, S, N)
+        if apply_sigmoid:
+            vis_e = torch.sigmoid(vis_e)
+
+        if self.predict_conf:
+            conf_e = self.conf_predictor(track_feats.reshape(B * S * N, self.latent_dim)).reshape(B, S, N)
+            if apply_sigmoid:
+                conf_e = torch.sigmoid(conf_e)
+        else:
+            conf_e = None
+
+        if return_feat:
+            return coord_preds, vis_e, track_feats, query_track_feat, conf_e
+        else:
+            return coord_preds, vis_e, conf_e

outdoor_v48_4gpu_v2/code/05_02-14:21:58/streamvggt/heads/track_modules/blocks.py

@@ -0,0 +1,237 @@
+import math
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from .utils import bilinear_sampler
+from .modules import Mlp, AttnBlock, CrossAttnBlock, ResidualBlock
+
+
+class EfficientUpdateFormer(nn.Module):
+    """
+    Transformer model that updates track estimates.
+    """
+
+    def __init__(
+        self,
+        space_depth=6,
+        time_depth=6,
+        input_dim=320,
+        hidden_size=384,
+        num_heads=8,
+        output_dim=130,
+        mlp_ratio=4.0,
+        add_space_attn=True,
+        num_virtual_tracks=64,
+    ):
+        super().__init__()
+
+        self.out_channels = 2
+        self.num_heads = num_heads
+        self.hidden_size = hidden_size
+        self.add_space_attn = add_space_attn
+
+        # Add input LayerNorm before linear projection
+        self.input_norm = nn.LayerNorm(input_dim)
+        self.input_transform = torch.nn.Linear(input_dim, hidden_size, bias=True)
+
+        # Add output LayerNorm before final projection
+        self.output_norm = nn.LayerNorm(hidden_size)
+        self.flow_head = torch.nn.Linear(hidden_size, output_dim, bias=True)
+        self.num_virtual_tracks = num_virtual_tracks
+
+        if self.add_space_attn:
+            self.virual_tracks = nn.Parameter(torch.randn(1, num_virtual_tracks, 1, hidden_size))
+        else:
+            self.virual_tracks = None
+
+        self.time_blocks = nn.ModuleList(
+            [
+                AttnBlock(
+                    hidden_size,
+                    num_heads,
+                    mlp_ratio=mlp_ratio,
+                    attn_class=nn.MultiheadAttention,
+                )
+                for _ in range(time_depth)
+            ]
+        )
+
+        if add_space_attn:
+            self.space_virtual_blocks = nn.ModuleList(
+                [
+                    AttnBlock(
+                        hidden_size,
+                        num_heads,
+                        mlp_ratio=mlp_ratio,
+                        attn_class=nn.MultiheadAttention,
+                    )
+                    for _ in range(space_depth)
+                ]
+            )
+            self.space_point2virtual_blocks = nn.ModuleList(
+                [CrossAttnBlock(hidden_size, hidden_size, num_heads, mlp_ratio=mlp_ratio) for _ in range(space_depth)]
+            )
+            self.space_virtual2point_blocks = nn.ModuleList(
+                [CrossAttnBlock(hidden_size, hidden_size, num_heads, mlp_ratio=mlp_ratio) for _ in range(space_depth)]
+            )
+            assert len(self.time_blocks) >= len(self.space_virtual2point_blocks)
+        self.initialize_weights()
+
+    def initialize_weights(self):
+        def _basic_init(module):
+            if isinstance(module, nn.Linear):
+                torch.nn.init.xavier_uniform_(module.weight)
+                if module.bias is not None:
+                    nn.init.constant_(module.bias, 0)
+            torch.nn.init.trunc_normal_(self.flow_head.weight, std=0.001)
+
+        self.apply(_basic_init)
+
+    def forward(self, input_tensor, mask=None):
+        # Apply input LayerNorm
+        input_tensor = self.input_norm(input_tensor)
+        tokens = self.input_transform(input_tensor)
+
+        init_tokens = tokens
+
+        B, _, T, _ = tokens.shape
+
+        if self.add_space_attn:
+            virtual_tokens = self.virual_tracks.repeat(B, 1, T, 1)
+            tokens = torch.cat([tokens, virtual_tokens], dim=1)
+
+        _, N, _, _ = tokens.shape
+
+        j = 0
+        for i in range(len(self.time_blocks)):
+            time_tokens = tokens.contiguous().view(B * N, T, -1)  # B N T C -> (B N) T C
+
+            time_tokens = self.time_blocks[i](time_tokens)
+
+            tokens = time_tokens.view(B, N, T, -1)  # (B N) T C -> B N T C
+            if self.add_space_attn and (i % (len(self.time_blocks) // len(self.space_virtual_blocks)) == 0):
+                space_tokens = tokens.permute(0, 2, 1, 3).contiguous().view(B * T, N, -1)  # B N T C -> (B T) N C
+                point_tokens = space_tokens[:, : N - self.num_virtual_tracks]
+                virtual_tokens = space_tokens[:, N - self.num_virtual_tracks :]
+
+                virtual_tokens = self.space_virtual2point_blocks[j](virtual_tokens, point_tokens, mask=mask)
+                virtual_tokens = self.space_virtual_blocks[j](virtual_tokens)
+                point_tokens = self.space_point2virtual_blocks[j](point_tokens, virtual_tokens, mask=mask)
+
+                space_tokens = torch.cat([point_tokens, virtual_tokens], dim=1)
+                tokens = space_tokens.view(B, T, N, -1).permute(0, 2, 1, 3)  # (B T) N C -> B N T C
+                j += 1
+
+        if self.add_space_attn:
+            tokens = tokens[:, : N - self.num_virtual_tracks]
+
+        tokens = tokens + init_tokens
+
+        # Apply output LayerNorm before final projection
+        tokens = self.output_norm(tokens)
+        flow = self.flow_head(tokens)
+
+        return flow, None
+
+
+class CorrBlock:
+    def __init__(self, fmaps, num_levels=4, radius=4, multiple_track_feats=False, padding_mode="zeros"):
+        """
+        Build a pyramid of feature maps from the input.
+
+        fmaps: Tensor (B, S, C, H, W)
+        num_levels: number of pyramid levels (each downsampled by factor 2)
+        radius: search radius for sampling correlation
+        multiple_track_feats: if True, split the target features per pyramid level
+        padding_mode: passed to grid_sample / bilinear_sampler
+        """
+        B, S, C, H, W = fmaps.shape
+        self.S, self.C, self.H, self.W = S, C, H, W
+        self.num_levels = num_levels
+        self.radius = radius
+        self.padding_mode = padding_mode
+        self.multiple_track_feats = multiple_track_feats
+
+        # Build pyramid: each level is half the spatial resolution of the previous
+        self.fmaps_pyramid = [fmaps]  # level 0 is full resolution
+        current_fmaps = fmaps
+        for i in range(num_levels - 1):
+            B, S, C, H, W = current_fmaps.shape
+            # Merge batch & sequence dimensions
+            current_fmaps = current_fmaps.reshape(B * S, C, H, W)
+            # Avg pool down by factor 2
+            current_fmaps = F.avg_pool2d(current_fmaps, kernel_size=2, stride=2)
+            _, _, H_new, W_new = current_fmaps.shape
+            current_fmaps = current_fmaps.reshape(B, S, C, H_new, W_new)
+            self.fmaps_pyramid.append(current_fmaps)
+
+        # Precompute a delta grid (of shape (2r+1, 2r+1, 2)) for sampling.
+        # This grid is added to the (scaled) coordinate centroids.
+        r = self.radius
+        dx = torch.linspace(-r, r, 2 * r + 1, device=fmaps.device, dtype=fmaps.dtype)
+        dy = torch.linspace(-r, r, 2 * r + 1, device=fmaps.device, dtype=fmaps.dtype)
+        # delta: for every (dy,dx) displacement (i.e. Δx, Δy)
+        self.delta = torch.stack(torch.meshgrid(dy, dx, indexing="ij"), dim=-1)  # shape: (2r+1, 2r+1, 2)
+
+    def corr_sample(self, targets, coords):
+        """
+        Instead of storing the entire correlation pyramid, we compute each level's correlation
+        volume, sample it immediately, then discard it. This saves GPU memory.
+
+        Args:
+          targets: Tensor (B, S, N, C) — features for the current targets.
+          coords: Tensor (B, S, N, 2) — coordinates at full resolution.
+
+        Returns:
+          Tensor (B, S, N, L) where L = num_levels * (2*radius+1)**2 (concatenated sampled correlations)
+        """
+        B, S, N, C = targets.shape
+
+        # If you have multiple track features, split them per level.
+        if self.multiple_track_feats:
+            targets_split = torch.split(targets, C // self.num_levels, dim=-1)
+
+        out_pyramid = []
+        for i, fmaps in enumerate(self.fmaps_pyramid):
+            # Get current spatial resolution H, W for this pyramid level.
+            B, S, C, H, W = fmaps.shape
+            # Reshape feature maps for correlation computation:
+            # fmap2s: (B, S, C, H*W)
+            fmap2s = fmaps.view(B, S, C, H * W)
+            # Choose appropriate target features.
+            fmap1 = targets_split[i] if self.multiple_track_feats else targets  # shape: (B, S, N, C)
+
+            # Compute correlation directly
+            corrs = compute_corr_level(fmap1, fmap2s, C)
+            corrs = corrs.view(B, S, N, H, W)
+
+            # Prepare sampling grid:
+            # Scale down the coordinates for the current level.
+            centroid_lvl = coords.reshape(B * S * N, 1, 1, 2) / (2**i)
+            # Make sure our precomputed delta grid is on the same device/dtype.
+            delta_lvl = self.delta.to(coords.device).to(coords.dtype)
+            # Now the grid for grid_sample is:
+            # coords_lvl = centroid_lvl + delta_lvl   (broadcasted over grid)
+            coords_lvl = centroid_lvl + delta_lvl.view(1, 2 * self.radius + 1, 2 * self.radius + 1, 2)
+
+            # Sample from the correlation volume using bilinear interpolation.
+            # We reshape corrs to (B * S * N, 1, H, W) so grid_sample acts over each target.
+            corrs_sampled = bilinear_sampler(
+                corrs.reshape(B * S * N, 1, H, W), coords_lvl, padding_mode=self.padding_mode
+            )
+            # The sampled output is (B * S * N, 1, 2r+1, 2r+1). Flatten the last two dims.
+            corrs_sampled = corrs_sampled.view(B, S, N, -1)  # Now shape: (B, S, N, (2r+1)^2)
+            out_pyramid.append(corrs_sampled)
+
+        # Concatenate all levels along the last dimension.
+        out = torch.cat(out_pyramid, dim=-1).contiguous()
+        return out
+
+
+def compute_corr_level(fmap1, fmap2s, C):
+    # fmap1: (B, S, N, C)
+    # fmap2s: (B, S, C, H*W)
+    corrs = torch.matmul(fmap1, fmap2s)  # (B, S, N, H*W)
+    corrs = corrs.view(fmap1.shape[0], fmap1.shape[1], fmap1.shape[2], -1)  # (B, S, N, H*W)
+    return corrs / math.sqrt(C)

outdoor_v48_4gpu_v2/code/05_02-14:21:58/streamvggt/heads/track_modules/modules.py

@@ -0,0 +1,211 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from functools import partial
+from typing import Callable
+import collections
+from torch import Tensor
+from itertools import repeat
+
+
+# From PyTorch internals
+def _ntuple(n):
+    def parse(x):
+        if isinstance(x, collections.abc.Iterable) and not isinstance(x, str):
+            return tuple(x)
+        return tuple(repeat(x, n))
+
+    return parse
+
+
+def exists(val):
+    return val is not None
+
+
+def default(val, d):
+    return val if exists(val) else d
+
+
+to_2tuple = _ntuple(2)
+
+
+class ResidualBlock(nn.Module):
+    """
+    ResidualBlock: construct a block of two conv layers with residual connections
+    """
+
+    def __init__(self, in_planes, planes, norm_fn="group", stride=1, kernel_size=3):
+        super(ResidualBlock, self).__init__()
+
+        self.conv1 = nn.Conv2d(
+            in_planes,
+            planes,
+            kernel_size=kernel_size,
+            padding=1,
+            stride=stride,
+            padding_mode="zeros",
+        )
+        self.conv2 = nn.Conv2d(
+            planes,
+            planes,
+            kernel_size=kernel_size,
+            padding=1,
+            padding_mode="zeros",
+        )
+        self.relu = nn.ReLU(inplace=True)
+
+        num_groups = planes // 8
+
+        if norm_fn == "group":
+            self.norm1 = nn.GroupNorm(num_groups=num_groups, num_channels=planes)
+            self.norm2 = nn.GroupNorm(num_groups=num_groups, num_channels=planes)
+            if not stride == 1:
+                self.norm3 = nn.GroupNorm(num_groups=num_groups, num_channels=planes)
+
+        elif norm_fn == "batch":
+            self.norm1 = nn.BatchNorm2d(planes)
+            self.norm2 = nn.BatchNorm2d(planes)
+            if not stride == 1:
+                self.norm3 = nn.BatchNorm2d(planes)
+
+        elif norm_fn == "instance":
+            self.norm1 = nn.InstanceNorm2d(planes)
+            self.norm2 = nn.InstanceNorm2d(planes)
+            if not stride == 1:
+                self.norm3 = nn.InstanceNorm2d(planes)
+
+        elif norm_fn == "none":
+            self.norm1 = nn.Sequential()
+            self.norm2 = nn.Sequential()
+            if not stride == 1:
+                self.norm3 = nn.Sequential()
+        else:
+            raise NotImplementedError
+
+        if stride == 1:
+            self.downsample = None
+        else:
+            self.downsample = nn.Sequential(
+                nn.Conv2d(in_planes, planes, kernel_size=1, stride=stride),
+                self.norm3,
+            )
+
+    def forward(self, x):
+        y = x
+        y = self.relu(self.norm1(self.conv1(y)))
+        y = self.relu(self.norm2(self.conv2(y)))
+
+        if self.downsample is not None:
+            x = self.downsample(x)
+
+        return self.relu(x + y)
+
+
+class Mlp(nn.Module):
+    """MLP as used in Vision Transformer, MLP-Mixer and related networks"""
+
+    def __init__(
+        self,
+        in_features,
+        hidden_features=None,
+        out_features=None,
+        act_layer=nn.GELU,
+        norm_layer=None,
+        bias=True,
+        drop=0.0,
+        use_conv=False,
+    ):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        bias = to_2tuple(bias)
+        drop_probs = to_2tuple(drop)
+        linear_layer = partial(nn.Conv2d, kernel_size=1) if use_conv else nn.Linear
+
+        self.fc1 = linear_layer(in_features, hidden_features, bias=bias[0])
+        self.act = act_layer()
+        self.drop1 = nn.Dropout(drop_probs[0])
+        self.fc2 = linear_layer(hidden_features, out_features, bias=bias[1])
+        self.drop2 = nn.Dropout(drop_probs[1])
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop1(x)
+        x = self.fc2(x)
+        x = self.drop2(x)
+        return x
+
+
+class AttnBlock(nn.Module):
+    def __init__(
+        self,
+        hidden_size,
+        num_heads,
+        attn_class: Callable[..., nn.Module] = nn.MultiheadAttention,
+        mlp_ratio=4.0,
+        **block_kwargs
+    ):
+        """
+        Self attention block
+        """
+        super().__init__()
+
+        self.norm1 = nn.LayerNorm(hidden_size)
+        self.norm2 = nn.LayerNorm(hidden_size)
+
+        self.attn = attn_class(embed_dim=hidden_size, num_heads=num_heads, batch_first=True, **block_kwargs)
+
+        mlp_hidden_dim = int(hidden_size * mlp_ratio)
+
+        self.mlp = Mlp(in_features=hidden_size, hidden_features=mlp_hidden_dim, drop=0)
+
+    def forward(self, x, mask=None):
+        # Prepare the mask for PyTorch's attention (it expects a different format)
+        # attn_mask = mask if mask is not None else None
+        # Normalize before attention
+        x = self.norm1(x)
+
+        # PyTorch's MultiheadAttention returns attn_output, attn_output_weights
+        # attn_output, _ = self.attn(x, x, x, attn_mask=attn_mask)
+
+        attn_output, _ = self.attn(x, x, x)
+
+        # Add & Norm
+        x = x + attn_output
+        x = x + self.mlp(self.norm2(x))
+        return x
+
+
+class CrossAttnBlock(nn.Module):
+    def __init__(self, hidden_size, context_dim, num_heads=1, mlp_ratio=4.0, **block_kwargs):
+        """
+        Cross attention block
+        """
+        super().__init__()
+
+        self.norm1 = nn.LayerNorm(hidden_size)
+        self.norm_context = nn.LayerNorm(hidden_size)
+        self.norm2 = nn.LayerNorm(hidden_size)
+
+        self.cross_attn = nn.MultiheadAttention(
+            embed_dim=hidden_size, num_heads=num_heads, batch_first=True, **block_kwargs
+        )
+
+        mlp_hidden_dim = int(hidden_size * mlp_ratio)
+
+        self.mlp = Mlp(in_features=hidden_size, hidden_features=mlp_hidden_dim, drop=0)
+
+    def forward(self, x, context, mask=None):
+        # Normalize inputs
+        x = self.norm1(x)
+        context = self.norm_context(context)
+
+        # Apply cross attention
+        # Note: nn.MultiheadAttention returns attn_output, attn_output_weights
+        attn_output, _ = self.cross_attn(x, context, context, attn_mask=mask)
+
+        # Add & Norm
+        x = x + attn_output
+        x = x + self.mlp(self.norm2(x))
+        return x

outdoor_v48_4gpu_v2/code/05_02-14:21:58/streamvggt/heads/track_modules/utils.py

@@ -0,0 +1,216 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from typing import Optional, Tuple, Union
+
+
+def get_2d_sincos_pos_embed(embed_dim: int, grid_size: Union[int, Tuple[int, int]], return_grid=False) -> torch.Tensor:
+    """
+    This function initializes a grid and generates a 2D positional embedding using sine and cosine functions.
+    It is a wrapper of get_2d_sincos_pos_embed_from_grid.
+    Args:
+    - embed_dim: The embedding dimension.
+    - grid_size: The grid size.
+    Returns:
+    - pos_embed: The generated 2D positional embedding.
+    """
+    if isinstance(grid_size, tuple):
+        grid_size_h, grid_size_w = grid_size
+    else:
+        grid_size_h = grid_size_w = grid_size
+    grid_h = torch.arange(grid_size_h, dtype=torch.float)
+    grid_w = torch.arange(grid_size_w, dtype=torch.float)
+    grid = torch.meshgrid(grid_w, grid_h, indexing="xy")
+    grid = torch.stack(grid, dim=0)
+    grid = grid.reshape([2, 1, grid_size_h, grid_size_w])
+    pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
+    if return_grid:
+        return (
+            pos_embed.reshape(1, grid_size_h, grid_size_w, -1).permute(0, 3, 1, 2),
+            grid,
+        )
+    return pos_embed.reshape(1, grid_size_h, grid_size_w, -1).permute(0, 3, 1, 2)
+
+
+def get_2d_sincos_pos_embed_from_grid(embed_dim: int, grid: torch.Tensor) -> torch.Tensor:
+    """
+    This function generates a 2D positional embedding from a given grid using sine and cosine functions.
+
+    Args:
+    - embed_dim: The embedding dimension.
+    - grid: The grid to generate the embedding from.
+
+    Returns:
+    - emb: The generated 2D positional embedding.
+    """
+    assert embed_dim % 2 == 0
+
+    # use half of dimensions to encode grid_h
+    emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0])  # (H*W, D/2)
+    emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1])  # (H*W, D/2)
+
+    emb = torch.cat([emb_h, emb_w], dim=2)  # (H*W, D)
+    return emb
+
+
+def get_1d_sincos_pos_embed_from_grid(embed_dim: int, pos: torch.Tensor) -> torch.Tensor:
+    """
+    This function generates a 1D positional embedding from a given grid using sine and cosine functions.
+
+    Args:
+    - embed_dim: The embedding dimension.
+    - pos: The position to generate the embedding from.
+
+    Returns:
+    - emb: The generated 1D positional embedding.
+    """
+    assert embed_dim % 2 == 0
+    omega = torch.arange(embed_dim // 2, dtype=torch.double)
+    omega /= embed_dim / 2.0
+    omega = 1.0 / 10000**omega  # (D/2,)
+
+    pos = pos.reshape(-1)  # (M,)
+    out = torch.einsum("m,d->md", pos, omega)  # (M, D/2), outer product
+
+    emb_sin = torch.sin(out)  # (M, D/2)
+    emb_cos = torch.cos(out)  # (M, D/2)
+
+    emb = torch.cat([emb_sin, emb_cos], dim=1)  # (M, D)
+    return emb[None].float()
+
+
+def get_2d_embedding(xy: torch.Tensor, C: int, cat_coords: bool = True) -> torch.Tensor:
+    """
+    This function generates a 2D positional embedding from given coordinates using sine and cosine functions.
+
+    Args:
+    - xy: The coordinates to generate the embedding from.
+    - C: The size of the embedding.
+    - cat_coords: A flag to indicate whether to concatenate the original coordinates to the embedding.
+
+    Returns:
+    - pe: The generated 2D positional embedding.
+    """
+    B, N, D = xy.shape
+    assert D == 2
+
+    x = xy[:, :, 0:1]
+    y = xy[:, :, 1:2]
+    div_term = (torch.arange(0, C, 2, device=xy.device, dtype=torch.float32) * (1000.0 / C)).reshape(1, 1, int(C / 2))
+
+    pe_x = torch.zeros(B, N, C, device=xy.device, dtype=torch.float32)
+    pe_y = torch.zeros(B, N, C, device=xy.device, dtype=torch.float32)
+
+    pe_x[:, :, 0::2] = torch.sin(x * div_term)
+    pe_x[:, :, 1::2] = torch.cos(x * div_term)
+
+    pe_y[:, :, 0::2] = torch.sin(y * div_term)
+    pe_y[:, :, 1::2] = torch.cos(y * div_term)
+
+    pe = torch.cat([pe_x, pe_y], dim=2)  # (B, N, C*3)
+    if cat_coords:
+        pe = torch.cat([xy, pe], dim=2)  # (B, N, C*3+3)
+    return pe
+
+
+def bilinear_sampler(input, coords, align_corners=True, padding_mode="border"):
+    r"""Sample a tensor using bilinear interpolation
+
+    `bilinear_sampler(input, coords)` samples a tensor :attr:`input` at
+    coordinates :attr:`coords` using bilinear interpolation. It is the same
+    as `torch.nn.functional.grid_sample()` but with a different coordinate
+    convention.
+
+    The input tensor is assumed to be of shape :math:`(B, C, H, W)`, where
+    :math:`B` is the batch size, :math:`C` is the number of channels,
+    :math:`H` is the height of the image, and :math:`W` is the width of the
+    image. The tensor :attr:`coords` of shape :math:`(B, H_o, W_o, 2)` is
+    interpreted as an array of 2D point coordinates :math:`(x_i,y_i)`.
+
+    Alternatively, the input tensor can be of size :math:`(B, C, T, H, W)`,
+    in which case sample points are triplets :math:`(t_i,x_i,y_i)`. Note
+    that in this case the order of the components is slightly different
+    from `grid_sample()`, which would expect :math:`(x_i,y_i,t_i)`.
+
+    If `align_corners` is `True`, the coordinate :math:`x` is assumed to be
+    in the range :math:`[0,W-1]`, with 0 corresponding to the center of the
+    left-most image pixel :math:`W-1` to the center of the right-most
+    pixel.
+
+    If `align_corners` is `False`, the coordinate :math:`x` is assumed to
+    be in the range :math:`[0,W]`, with 0 corresponding to the left edge of
+    the left-most pixel :math:`W` to the right edge of the right-most
+    pixel.
+
+    Similar conventions apply to the :math:`y` for the range
+    :math:`[0,H-1]` and :math:`[0,H]` and to :math:`t` for the range
+    :math:`[0,T-1]` and :math:`[0,T]`.
+
+    Args:
+        input (Tensor): batch of input images.
+        coords (Tensor): batch of coordinates.
+        align_corners (bool, optional): Coordinate convention. Defaults to `True`.
+        padding_mode (str, optional): Padding mode. Defaults to `"border"`.
+
+    Returns:
+        Tensor: sampled points.
+    """
+    coords = coords.detach().clone()
+    ############################################################
+    # IMPORTANT:
+    coords = coords.to(input.device).to(input.dtype)
+    ############################################################
+
+    sizes = input.shape[2:]
+
+    assert len(sizes) in [2, 3]
+
+    if len(sizes) == 3:
+        # t x y -> x y t to match dimensions T H W in grid_sample
+        coords = coords[..., [1, 2, 0]]
+
+    if align_corners:
+        scale = torch.tensor(
+            [2 / max(size - 1, 1) for size in reversed(sizes)], device=coords.device, dtype=coords.dtype
+        )
+    else:
+        scale = torch.tensor([2 / size for size in reversed(sizes)], device=coords.device, dtype=coords.dtype)
+
+    coords.mul_(scale)  # coords = coords * scale
+    coords.sub_(1)  # coords = coords - 1
+
+    return F.grid_sample(input, coords, align_corners=align_corners, padding_mode=padding_mode)
+
+
+def sample_features4d(input, coords):
+    r"""Sample spatial features
+
+    `sample_features4d(input, coords)` samples the spatial features
+    :attr:`input` represented by a 4D tensor :math:`(B, C, H, W)`.
+
+    The field is sampled at coordinates :attr:`coords` using bilinear
+    interpolation. :attr:`coords` is assumed to be of shape :math:`(B, R,
+    2)`, where each sample has the format :math:`(x_i, y_i)`. This uses the
+    same convention as :func:`bilinear_sampler` with `align_corners=True`.
+
+    The output tensor has one feature per point, and has shape :math:`(B,
+    R, C)`.
+
+    Args:
+        input (Tensor): spatial features.
+        coords (Tensor): points.
+
+    Returns:
+        Tensor: sampled features.
+    """
+
+    B, _, _, _ = input.shape
+
+    # B R 2 -> B R 1 2
+    coords = coords.unsqueeze(2)
+
+    # B C R 1
+    feats = bilinear_sampler(input, coords)
+
+    return feats.permute(0, 2, 1, 3).view(B, -1, feats.shape[1] * feats.shape[3])  # B C R 1 -> B R C

outdoor_v48_4gpu_v2/code/05_02-14:21:58/streamvggt/heads/utils.py

@@ -0,0 +1,99 @@
+import torch
+import torch.nn as nn
+
+
+def position_grid_to_embed(pos_grid: torch.Tensor, embed_dim: int, omega_0: float = 100) -> torch.Tensor:
+    """
+    Convert 2D position grid (HxWx2) to sinusoidal embeddings (HxWxC)
+
+    Args:
+        pos_grid: Tensor of shape (H, W, 2) containing 2D coordinates
+        embed_dim: Output channel dimension for embeddings
+
+    Returns:
+        Tensor of shape (H, W, embed_dim) with positional embeddings
+    """
+    H, W, grid_dim = pos_grid.shape
+    assert grid_dim == 2
+    pos_flat = pos_grid.reshape(-1, grid_dim)  # Flatten to (H*W, 2)
+
+    # Process x and y coordinates separately
+    emb_x = make_sincos_pos_embed(embed_dim // 2, pos_flat[:, 0], omega_0=omega_0)  # [1, H*W, D/2]
+    emb_y = make_sincos_pos_embed(embed_dim // 2, pos_flat[:, 1], omega_0=omega_0)  # [1, H*W, D/2]
+
+    # Combine and reshape
+    emb = torch.cat([emb_x, emb_y], dim=-1)  # [1, H*W, D]
+
+    return emb.view(H, W, embed_dim)  # [H, W, D]
+
+
+def make_sincos_pos_embed(embed_dim: int, pos: torch.Tensor, omega_0: float = 100) -> torch.Tensor:
+    """
+    This function generates a 1D positional embedding from a given grid using sine and cosine functions.
+
+    Args:
+    - embed_dim: The embedding dimension.
+    - pos: The position to generate the embedding from.
+
+    Returns:
+    - emb: The generated 1D positional embedding.
+    """
+    assert embed_dim % 2 == 0
+    omega = torch.arange(embed_dim // 2, dtype=torch.double, device=pos.device)
+    omega /= embed_dim / 2.0
+    omega = 1.0 / omega_0**omega  # (D/2,)
+
+    pos = pos.reshape(-1)  # (M,)
+    out = torch.einsum("m,d->md", pos, omega)  # (M, D/2), outer product
+
+    emb_sin = torch.sin(out)  # (M, D/2)
+    emb_cos = torch.cos(out)  # (M, D/2)
+
+    emb = torch.cat([emb_sin, emb_cos], dim=1)  # (M, D)
+    return emb.float()
+
+
+def create_uv_grid(
+    width: int, height: int, aspect_ratio: float = None, dtype: torch.dtype = None, device: torch.device = None
+) -> torch.Tensor:
+    """
+    Create a normalized UV grid of shape (width, height, 2).
+
+    The grid spans horizontally and vertically according to an aspect ratio,
+    ensuring the top-left corner is at (-x_span, -y_span) and the bottom-right
+    corner is at (x_span, y_span), normalized by the diagonal of the plane.
+
+    Args:
+        width (int): Number of points horizontally.
+        height (int): Number of points vertically.
+        aspect_ratio (float, optional): Width-to-height ratio. Defaults to width/height.
+        dtype (torch.dtype, optional): Data type of the resulting tensor.
+        device (torch.device, optional): Device on which the tensor is created.
+
+    Returns:
+        torch.Tensor: A (width, height, 2) tensor of UV coordinates.
+    """
+    # Derive aspect ratio if not explicitly provided
+    if aspect_ratio is None:
+        aspect_ratio = float(width) / float(height)
+
+    # Compute normalized spans for X and Y
+    diag_factor = (aspect_ratio**2 + 1.0) ** 0.5
+    span_x = aspect_ratio / diag_factor
+    span_y = 1.0 / diag_factor
+
+    # Establish the linspace boundaries
+    left_x = -span_x * (width - 1) / width
+    right_x = span_x * (width - 1) / width
+    top_y = -span_y * (height - 1) / height
+    bottom_y = span_y * (height - 1) / height
+
+    # Generate 1D coordinates
+    x_coords = torch.linspace(left_x, right_x, steps=width, dtype=dtype, device=device)
+    y_coords = torch.linspace(top_y, bottom_y, steps=height, dtype=dtype, device=device)
+
+    # Create 2D meshgrid (width x height) and stack into UV
+    uu, vv = torch.meshgrid(x_coords, y_coords, indexing="xy")
+    uv_grid = torch.stack((uu, vv), dim=-1)
+
+    return uv_grid

outdoor_v48_4gpu_v2/code/05_02-14:21:58/streamvggt/layers/__init__.py

@@ -0,0 +1,5 @@
+from .mlp import Mlp
+from .patch_embed import PatchEmbed
+from .swiglu_ffn import SwiGLUFFN, SwiGLUFFNFused
+from .block import NestedTensorBlock
+from .attention import MemEffAttention

outdoor_v48_4gpu_v2/code/05_02-14:21:58/streamvggt/layers/attention.py

@@ -0,0 +1,129 @@
+import logging
+import os
+import warnings
+
+import torch
+from torch import Tensor
+from torch import nn
+import torch.nn.functional as F
+from typing import Union, Tuple, Dict, Optional
+
+from einops import rearrange
+
+XFORMERS_AVAILABLE = False
+
+
+class Attention(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_heads: int = 8,
+        qkv_bias: bool = True,
+        proj_bias: bool = True,
+        attn_drop: float = 0.0,
+        proj_drop: float = 0.0,
+        norm_layer: nn.Module = nn.LayerNorm,
+        qk_norm: bool = False,
+        fused_attn: bool = True,  # use F.scaled_dot_product_attention or not
+        rope=None,
+    ) -> None:
+        super().__init__()
+        assert dim % num_heads == 0, "dim should be divisible by num_heads"
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+        self.scale = self.head_dim**-0.5
+        self.fused_attn = fused_attn
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.q_norm = norm_layer(self.head_dim) if qk_norm else nn.Identity()
+        self.k_norm = norm_layer(self.head_dim) if qk_norm else nn.Identity()
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim, bias=proj_bias)
+        self.proj_drop = nn.Dropout(proj_drop)
+        self.rope = rope
+
+    def forward(self, 
+        x: torch.Tensor, 
+        pos=None, 
+        attn_mask=None, 
+        past_key_values=None, 
+        use_cache=False
+    ) -> Union[torch.Tensor, Tuple[torch.Tensor, Tuple]]:
+        B, N, C = x.shape
+        qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, self.head_dim).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv.unbind(0)
+
+        pos_k = pos
+        if use_cache:
+            k = k.unsqueeze(2)
+            v = v.unsqueeze(2)
+            if past_key_values is not None:
+                past_k, past_v = past_key_values
+                k = torch.cat([past_k, k], dim=2)
+                v = torch.cat([past_v, v], dim=2)
+                
+            new_kv = (k, v)
+            a, b, c, d, e = k.shape
+            k = k.reshape(a, b, c*d, e)
+            v = v.reshape(a, b, c*d, e)
+            if pos_k is not None:
+                #print(pos_k.shape)
+                pos_k = pos_k.repeat(1, c, 1)
+                #print(pos_k.shape)
+
+        q, k = self.q_norm(q), self.k_norm(k)
+
+        if self.rope is not None:
+            q = self.rope(q, pos)
+            k = self.rope(k, pos_k)
+
+        if self.fused_attn:
+            x = F.scaled_dot_product_attention(
+                q,
+                k,
+                v,
+                attn_mask=attn_mask,
+                dropout_p=self.attn_drop.p if self.training else 0.0,
+            )
+
+        else:
+            q = q * self.scale
+            attn = q @ k.transpose(-2, -1)
+
+            # Mask
+            if attn_mask is not None:
+                assert attn_mask.shape[-2:] == (N, N), f"Expected mask shape [..., {N}, {N}], got {attn_mask.shape}"
+                attn = attn + attn_mask
+
+            attn = attn.softmax(dim=-1)
+            attn = self.attn_drop(attn)
+            x = attn @ v
+
+        x = x.transpose(1, 2).reshape(B, N, C)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        if use_cache:
+            return x, new_kv
+        return x
+
+
+class MemEffAttention(Attention):
+    def forward(self, x: Tensor, attn_bias=None, pos=None) -> Tensor:
+        assert pos is None
+        if not XFORMERS_AVAILABLE:
+            if attn_bias is not None:
+                raise AssertionError("xFormers is required for using nested tensors")
+            return super().forward(x)
+
+        B, N, C = x.shape
+        qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads)
+
+        q, k, v = unbind(qkv, 2)
+
+        x = memory_efficient_attention(q, k, v, attn_bias=attn_bias)
+        x = x.reshape([B, N, C])
+
+        x = self.proj(x)
+        x = self.proj_drop(x)
+
+        return x

outdoor_v48_4gpu_v2/code/05_02-14:21:58/streamvggt/layers/block.py

@@ -0,0 +1,263 @@
+import logging
+import os
+from typing import Callable, List, Any, Tuple, Dict, Union
+import warnings
+
+import torch
+from torch import nn, Tensor
+
+from .attention import Attention
+from .drop_path import DropPath
+from .layer_scale import LayerScale
+from .mlp import Mlp
+
+
+XFORMERS_AVAILABLE = False
+
+
+class Block(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_heads: int,
+        mlp_ratio: float = 4.0,
+        qkv_bias: bool = True,
+        proj_bias: bool = True,
+        ffn_bias: bool = True,
+        drop: float = 0.0,
+        attn_drop: float = 0.0,
+        init_values=None,
+        drop_path: float = 0.0,
+        act_layer: Callable[..., nn.Module] = nn.GELU,
+        norm_layer: Callable[..., nn.Module] = nn.LayerNorm,
+        attn_class: Callable[..., nn.Module] = Attention,
+        ffn_layer: Callable[..., nn.Module] = Mlp,
+        qk_norm: bool = False,
+        fused_attn: bool = True,  # use F.scaled_dot_product_attention or not
+        rope=None,
+    ) -> None:
+        super().__init__()
+
+        self.norm1 = norm_layer(dim)
+
+        self.attn = attn_class(
+            dim,
+            num_heads=num_heads,
+            qkv_bias=qkv_bias,
+            proj_bias=proj_bias,
+            attn_drop=attn_drop,
+            proj_drop=drop,
+            qk_norm=qk_norm,
+            fused_attn=fused_attn,
+            rope=rope,
+        )
+
+        self.ls1 = LayerScale(dim, init_values=init_values) if init_values else nn.Identity()
+        self.drop_path1 = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
+
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = ffn_layer(
+            in_features=dim,
+            hidden_features=mlp_hidden_dim,
+            act_layer=act_layer,
+            drop=drop,
+            bias=ffn_bias,
+        )
+        self.ls2 = LayerScale(dim, init_values=init_values) if init_values else nn.Identity()
+        self.drop_path2 = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
+
+        self.sample_drop_ratio = drop_path
+
+    def forward(self, x: Tensor, pos=None, attn_mask=None, past_key_values=None, use_cache=False) -> Union[Tensor, Tuple[Tensor, Dict]]:
+            
+        def attn_residual_func(x: Tensor, pos=None, attn_mask=None, past_key_values=None, use_cache=False) -> Union[Tensor, Tuple[Tensor, Dict]]:
+            if use_cache:
+                output, new_kv = self.attn(self.norm1(x), pos=pos, past_key_values=past_key_values, use_cache=True)
+                return self.ls1(output), new_kv
+            else:
+                if attn_mask is not None:
+                    return self.ls1(self.attn(self.norm1(x), pos=pos, attn_mask=attn_mask))
+                else:
+                    return self.ls1(self.attn(self.norm1(x), pos=pos))
+        def ffn_residual_func(x: Tensor) -> Tensor:
+            return self.ls2(self.mlp(self.norm2(x)))
+        
+        if use_cache:
+            attn_output, new_kv = attn_residual_func(x, pos=pos, past_key_values=past_key_values, use_cache=True)
+            x = x + attn_output
+            x = x + ffn_residual_func(x)
+            return x, new_kv
+
+        if self.training and self.sample_drop_ratio > 0.1:
+            # the overhead is compensated only for a drop path rate larger than 0.1
+            x = drop_add_residual_stochastic_depth(
+                x,
+                pos=pos,
+                residual_func=attn_residual_func,
+                sample_drop_ratio=self.sample_drop_ratio,
+            )
+            x = drop_add_residual_stochastic_depth(
+                x,
+                residual_func=ffn_residual_func,
+                sample_drop_ratio=self.sample_drop_ratio,
+            )
+        elif self.training and self.sample_drop_ratio > 0.0:
+            x = x + self.drop_path1(attn_residual_func(x, pos=pos, attn_mask=attn_mask))
+            x = x + self.drop_path1(ffn_residual_func(x))  # FIXME: drop_path2
+        else:
+            x = x + attn_residual_func(x, pos=pos, attn_mask=attn_mask)
+            x = x + ffn_residual_func(x)
+        return x
+
+
+def drop_add_residual_stochastic_depth(
+    x: Tensor,
+    residual_func: Callable[[Tensor], Tensor],
+    sample_drop_ratio: float = 0.0,
+    pos=None,
+) -> Tensor:
+    # 1) extract subset using permutation
+    b, n, d = x.shape
+    sample_subset_size = max(int(b * (1 - sample_drop_ratio)), 1)
+    brange = (torch.randperm(b, device=x.device))[:sample_subset_size]
+    x_subset = x[brange]
+
+    # 2) apply residual_func to get residual
+    if pos is not None:
+        # if necessary, apply rope to the subset
+        pos = pos[brange]
+        residual = residual_func(x_subset, pos=pos)
+    else:
+        residual = residual_func(x_subset)
+
+    x_flat = x.flatten(1)
+    residual = residual.flatten(1)
+
+    residual_scale_factor = b / sample_subset_size
+
+    # 3) add the residual
+    x_plus_residual = torch.index_add(x_flat, 0, brange, residual.to(dtype=x.dtype), alpha=residual_scale_factor)
+    return x_plus_residual.view_as(x)
+
+
+def get_branges_scales(x, sample_drop_ratio=0.0):
+    b, n, d = x.shape
+    sample_subset_size = max(int(b * (1 - sample_drop_ratio)), 1)
+    brange = (torch.randperm(b, device=x.device))[:sample_subset_size]
+    residual_scale_factor = b / sample_subset_size
+    return brange, residual_scale_factor
+
+
+def add_residual(x, brange, residual, residual_scale_factor, scaling_vector=None):
+    if scaling_vector is None:
+        x_flat = x.flatten(1)
+        residual = residual.flatten(1)
+        x_plus_residual = torch.index_add(x_flat, 0, brange, residual.to(dtype=x.dtype), alpha=residual_scale_factor)
+    else:
+        x_plus_residual = scaled_index_add(
+            x, brange, residual.to(dtype=x.dtype), scaling=scaling_vector, alpha=residual_scale_factor
+        )
+    return x_plus_residual
+
+
+attn_bias_cache: Dict[Tuple, Any] = {}
+
+
+def get_attn_bias_and_cat(x_list, branges=None):
+    """
+    this will perform the index select, cat the tensors, and provide the attn_bias from cache
+    """
+    batch_sizes = [b.shape[0] for b in branges] if branges is not None else [x.shape[0] for x in x_list]
+    all_shapes = tuple((b, x.shape[1]) for b, x in zip(batch_sizes, x_list))
+    if all_shapes not in attn_bias_cache.keys():
+        seqlens = []
+        for b, x in zip(batch_sizes, x_list):
+            for _ in range(b):
+                seqlens.append(x.shape[1])
+        attn_bias = fmha.BlockDiagonalMask.from_seqlens(seqlens)
+        attn_bias._batch_sizes = batch_sizes
+        attn_bias_cache[all_shapes] = attn_bias
+
+    if branges is not None:
+        cat_tensors = index_select_cat([x.flatten(1) for x in x_list], branges).view(1, -1, x_list[0].shape[-1])
+    else:
+        tensors_bs1 = tuple(x.reshape([1, -1, *x.shape[2:]]) for x in x_list)
+        cat_tensors = torch.cat(tensors_bs1, dim=1)
+
+    return attn_bias_cache[all_shapes], cat_tensors
+
+
+def drop_add_residual_stochastic_depth_list(
+    x_list: List[Tensor],
+    residual_func: Callable[[Tensor, Any], Tensor],
+    sample_drop_ratio: float = 0.0,
+    scaling_vector=None,
+) -> Tensor:
+    # 1) generate random set of indices for dropping samples in the batch
+    branges_scales = [get_branges_scales(x, sample_drop_ratio=sample_drop_ratio) for x in x_list]
+    branges = [s[0] for s in branges_scales]
+    residual_scale_factors = [s[1] for s in branges_scales]
+
+    # 2) get attention bias and index+concat the tensors
+    attn_bias, x_cat = get_attn_bias_and_cat(x_list, branges)
+
+    # 3) apply residual_func to get residual, and split the result
+    residual_list = attn_bias.split(residual_func(x_cat, attn_bias=attn_bias))  # type: ignore
+
+    outputs = []
+    for x, brange, residual, residual_scale_factor in zip(x_list, branges, residual_list, residual_scale_factors):
+        outputs.append(add_residual(x, brange, residual, residual_scale_factor, scaling_vector).view_as(x))
+    return outputs
+
+
+class NestedTensorBlock(Block):
+    def forward_nested(self, x_list: List[Tensor]) -> List[Tensor]:
+        """
+        x_list contains a list of tensors to nest together and run
+        """
+        assert isinstance(self.attn, MemEffAttention)
+
+        if self.training and self.sample_drop_ratio > 0.0:
+
+            def attn_residual_func(x: Tensor, attn_bias=None) -> Tensor:
+                return self.attn(self.norm1(x), attn_bias=attn_bias)
+
+            def ffn_residual_func(x: Tensor, attn_bias=None) -> Tensor:
+                return self.mlp(self.norm2(x))
+
+            x_list = drop_add_residual_stochastic_depth_list(
+                x_list,
+                residual_func=attn_residual_func,
+                sample_drop_ratio=self.sample_drop_ratio,
+                scaling_vector=self.ls1.gamma if isinstance(self.ls1, LayerScale) else None,
+            )
+            x_list = drop_add_residual_stochastic_depth_list(
+                x_list,
+                residual_func=ffn_residual_func,
+                sample_drop_ratio=self.sample_drop_ratio,
+                scaling_vector=self.ls2.gamma if isinstance(self.ls1, LayerScale) else None,
+            )
+            return x_list
+        else:
+
+            def attn_residual_func(x: Tensor, attn_bias=None) -> Tensor:
+                return self.ls1(self.attn(self.norm1(x), attn_bias=attn_bias))
+
+            def ffn_residual_func(x: Tensor, attn_bias=None) -> Tensor:
+                return self.ls2(self.mlp(self.norm2(x)))
+
+            attn_bias, x = get_attn_bias_and_cat(x_list)
+            x = x + attn_residual_func(x, attn_bias=attn_bias)
+            x = x + ffn_residual_func(x)
+            return attn_bias.split(x)
+
+    def forward(self, x_or_x_list):
+        if isinstance(x_or_x_list, Tensor):
+            return super().forward(x_or_x_list)
+        elif isinstance(x_or_x_list, list):
+            if not XFORMERS_AVAILABLE:
+                raise AssertionError("xFormers is required for using nested tensors")
+            return self.forward_nested(x_or_x_list)
+        else:
+            raise AssertionError

outdoor_v48_4gpu_v2/code/05_02-14:21:58/streamvggt/layers/drop_path.py

@@ -0,0 +1,24 @@
+from torch import nn
+
+
+def drop_path(x, drop_prob: float = 0.0, training: bool = False):
+    if drop_prob == 0.0 or not training:
+        return x
+    keep_prob = 1 - drop_prob
+    shape = (x.shape[0],) + (1,) * (x.ndim - 1)  # work with diff dim tensors, not just 2D ConvNets
+    random_tensor = x.new_empty(shape).bernoulli_(keep_prob)
+    if keep_prob > 0.0:
+        random_tensor.div_(keep_prob)
+    output = x * random_tensor
+    return output
+
+
+class DropPath(nn.Module):
+    """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks)."""
+
+    def __init__(self, drop_prob=None):
+        super(DropPath, self).__init__()
+        self.drop_prob = drop_prob
+
+    def forward(self, x):
+        return drop_path(x, self.drop_prob, self.training)

outdoor_v48_4gpu_v2/code/05_02-14:21:58/streamvggt/layers/layer_scale.py

@@ -0,0 +1,20 @@
+from typing import Union
+
+import torch
+from torch import Tensor
+from torch import nn
+
+
+class LayerScale(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        init_values: Union[float, Tensor] = 1e-5,
+        inplace: bool = False,
+    ) -> None:
+        super().__init__()
+        self.inplace = inplace
+        self.gamma = nn.Parameter(init_values * torch.ones(dim))
+
+    def forward(self, x: Tensor) -> Tensor:
+        return x.mul_(self.gamma) if self.inplace else x * self.gamma

outdoor_v48_4gpu_v2/code/05_02-14:21:58/streamvggt/layers/mlp.py

@@ -0,0 +1,30 @@
+from typing import Callable, Optional
+
+from torch import Tensor, nn
+
+
+class Mlp(nn.Module):
+    def __init__(
+        self,
+        in_features: int,
+        hidden_features: Optional[int] = None,
+        out_features: Optional[int] = None,
+        act_layer: Callable[..., nn.Module] = nn.GELU,
+        drop: float = 0.0,
+        bias: bool = True,
+    ) -> None:
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features, bias=bias)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features, bias=bias)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x: Tensor) -> Tensor:
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x

outdoor_v48_4gpu_v2/code/05_02-14:21:58/streamvggt/layers/patch_embed.py

@@ -0,0 +1,79 @@
+from typing import Callable, Optional, Tuple, Union
+
+from torch import Tensor
+import torch.nn as nn
+
+
+def make_2tuple(x):
+    if isinstance(x, tuple):
+        assert len(x) == 2
+        return x
+
+    assert isinstance(x, int)
+    return (x, x)
+
+
+class PatchEmbed(nn.Module):
+    """
+    2D image to patch embedding: (B,C,H,W) -> (B,N,D)
+
+    Args:
+        img_size: Image size.
+        patch_size: Patch token size.
+        in_chans: Number of input image channels.
+        embed_dim: Number of linear projection output channels.
+        norm_layer: Normalization layer.
+    """
+
+    def __init__(
+        self,
+        img_size: Union[int, Tuple[int, int]] = 224,
+        patch_size: Union[int, Tuple[int, int]] = 16,
+        in_chans: int = 3,
+        embed_dim: int = 768,
+        norm_layer: Optional[Callable] = None,
+        flatten_embedding: bool = True,
+    ) -> None:
+        super().__init__()
+
+        image_HW = make_2tuple(img_size)
+        patch_HW = make_2tuple(patch_size)
+        patch_grid_size = (
+            image_HW[0] // patch_HW[0],
+            image_HW[1] // patch_HW[1],
+        )
+
+        self.img_size = image_HW
+        self.patch_size = patch_HW
+        self.patches_resolution = patch_grid_size
+        self.num_patches = patch_grid_size[0] * patch_grid_size[1]
+
+        self.in_chans = in_chans
+        self.embed_dim = embed_dim
+
+        self.flatten_embedding = flatten_embedding
+
+        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_HW, stride=patch_HW)
+        self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()
+
+    def forward(self, x: Tensor) -> Tensor:
+        _, _, H, W = x.shape
+        patch_H, patch_W = self.patch_size
+
+        assert H % patch_H == 0, f"Input image height {H} is not a multiple of patch height {patch_H}"
+        assert W % patch_W == 0, f"Input image width {W} is not a multiple of patch width: {patch_W}"
+
+        x = self.proj(x)  # B C H W
+        H, W = x.size(2), x.size(3)
+        x = x.flatten(2).transpose(1, 2)  # B HW C
+        x = self.norm(x)
+        if not self.flatten_embedding:
+            x = x.reshape(-1, H, W, self.embed_dim)  # B H W C
+        return x
+
+    def flops(self) -> float:
+        Ho, Wo = self.patches_resolution
+        flops = Ho * Wo * self.embed_dim * self.in_chans * (self.patch_size[0] * self.patch_size[1])
+        if self.norm is not None:
+            flops += Ho * Wo * self.embed_dim
+        return flops

outdoor_v48_4gpu_v2/code/05_02-14:21:58/streamvggt/layers/rope.py

@@ -0,0 +1,172 @@
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from typing import Dict, Tuple
+
+
+class PositionGetter:
+    """Generates and caches 2D spatial positions for patches in a grid.
+
+    This class efficiently manages the generation of spatial coordinates for patches
+    in a 2D grid, caching results to avoid redundant computations.
+
+    Attributes:
+        position_cache: Dictionary storing precomputed position tensors for different
+            grid dimensions.
+    """
+
+    def __init__(self):
+        """Initializes the position generator with an empty cache."""
+        self.position_cache: Dict[Tuple[int, int], torch.Tensor] = {}
+
+    def __call__(self, batch_size: int, height: int, width: int, device: torch.device) -> torch.Tensor:
+        """Generates spatial positions for a batch of patches.
+
+        Args:
+            batch_size: Number of samples in the batch.
+            height: Height of the grid in patches.
+            width: Width of the grid in patches.
+            device: Target device for the position tensor.
+
+        Returns:
+            Tensor of shape (batch_size, height*width, 2) containing y,x coordinates
+            for each position in the grid, repeated for each batch item.
+        """
+        if (height, width) not in self.position_cache:
+            y_coords = torch.arange(height, device=device)
+            x_coords = torch.arange(width, device=device)
+            positions = torch.cartesian_prod(y_coords, x_coords)
+            self.position_cache[height, width] = positions
+
+        cached_positions = self.position_cache[height, width]
+        return cached_positions.view(1, height * width, 2).expand(batch_size, -1, -1).clone()
+
+
+class RotaryPositionEmbedding2D(nn.Module):
+    """2D Rotary Position Embedding implementation.
+
+    This module applies rotary position embeddings to input tokens based on their
+    2D spatial positions. It handles the position-dependent rotation of features
+    separately for vertical and horizontal dimensions.
+
+    Args:
+        frequency: Base frequency for the position embeddings. Default: 100.0
+        scaling_factor: Scaling factor for frequency computation. Default: 1.0
+
+    Attributes:
+        base_frequency: Base frequency for computing position embeddings.
+        scaling_factor: Factor to scale the computed frequencies.
+        frequency_cache: Cache for storing precomputed frequency components.
+    """
+
+    def __init__(self, frequency: float = 100.0, scaling_factor: float = 1.0):
+        """Initializes the 2D RoPE module."""
+        super().__init__()
+        self.base_frequency = frequency
+        self.scaling_factor = scaling_factor
+        self.frequency_cache: Dict[Tuple, Tuple[torch.Tensor, torch.Tensor]] = {}
+
+    def _compute_frequency_components(
+        self, dim: int, seq_len: int, device: torch.device, dtype: torch.dtype
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Computes frequency components for rotary embeddings.
+
+        Args:
+            dim: Feature dimension (must be even).
+            seq_len: Maximum sequence length.
+            device: Target device for computations.
+            dtype: Data type for the computed tensors.
+
+        Returns:
+            Tuple of (cosine, sine) tensors for frequency components.
+        """
+        cache_key = (dim, seq_len, device, dtype)
+        if cache_key not in self.frequency_cache:
+            # Compute frequency bands
+            exponents = torch.arange(0, dim, 2, device=device).float() / dim
+            inv_freq = 1.0 / (self.base_frequency**exponents)
+
+            # Generate position-dependent frequencies
+            positions = torch.arange(seq_len, device=device, dtype=inv_freq.dtype)
+            angles = torch.einsum("i,j->ij", positions, inv_freq)
+
+            # Compute and cache frequency components
+            angles = angles.to(dtype)
+            angles = torch.cat((angles, angles), dim=-1)
+            cos_components = angles.cos().to(dtype)
+            sin_components = angles.sin().to(dtype)
+            self.frequency_cache[cache_key] = (cos_components, sin_components)
+
+        return self.frequency_cache[cache_key]
+
+    @staticmethod
+    def _rotate_features(x: torch.Tensor) -> torch.Tensor:
+        """Performs feature rotation by splitting and recombining feature dimensions.
+
+        Args:
+            x: Input tensor to rotate.
+
+        Returns:
+            Rotated feature tensor.
+        """
+        feature_dim = x.shape[-1]
+        x1, x2 = x[..., : feature_dim // 2], x[..., feature_dim // 2 :]
+        return torch.cat((-x2, x1), dim=-1)
+
+    def _apply_1d_rope(
+        self, tokens: torch.Tensor, positions: torch.Tensor, cos_comp: torch.Tensor, sin_comp: torch.Tensor
+    ) -> torch.Tensor:
+        """Applies 1D rotary position embeddings along one dimension.
+
+        Args:
+            tokens: Input token features.
+            positions: Position indices.
+            cos_comp: Cosine components for rotation.
+            sin_comp: Sine components for rotation.
+
+        Returns:
+            Tokens with applied rotary position embeddings.
+        """
+        # Embed positions with frequency components
+        cos = F.embedding(positions, cos_comp)[:, None, :, :]
+        sin = F.embedding(positions, sin_comp)[:, None, :, :]
+
+        # Apply rotation
+        return (tokens * cos) + (self._rotate_features(tokens) * sin)
+
+    def forward(self, tokens: torch.Tensor, positions: torch.Tensor) -> torch.Tensor:
+        """Applies 2D rotary position embeddings to input tokens.
+
+        Args:
+            tokens: Input tensor of shape (batch_size, n_heads, n_tokens, dim).
+                   The feature dimension (dim) must be divisible by 4.
+            positions: Position tensor of shape (batch_size, n_tokens, 2) containing
+                      the y and x coordinates for each token.
+
+        Returns:
+            Tensor of same shape as input with applied 2D rotary position embeddings.
+
+        Raises:
+            AssertionError: If input dimensions are invalid or positions are malformed.
+        """
+        # Validate inputs
+        assert tokens.size(-1) % 2 == 0, "Feature dimension must be even"
+        assert positions.ndim == 3 and positions.shape[-1] == 2, "Positions must have shape (batch_size, n_tokens, 2)"
+
+        # Compute feature dimension for each spatial direction
+        feature_dim = tokens.size(-1) // 2
+
+        # Get frequency components
+        max_position = int(positions.max()) + 1
+        cos_comp, sin_comp = self._compute_frequency_components(feature_dim, max_position, tokens.device, tokens.dtype)
+
+        # Split features for vertical and horizontal processing
+        vertical_features, horizontal_features = tokens.chunk(2, dim=-1)
+
+        # Apply RoPE separately for each dimension
+        vertical_features = self._apply_1d_rope(vertical_features, positions[..., 0], cos_comp, sin_comp)
+        horizontal_features = self._apply_1d_rope(horizontal_features, positions[..., 1], cos_comp, sin_comp)
+
+        # Combine processed features
+        return torch.cat((vertical_features, horizontal_features), dim=-1)

outdoor_v48_4gpu_v2/code/05_02-14:21:58/streamvggt/layers/swiglu_ffn.py

@@ -0,0 +1,67 @@
+import os
+from typing import Callable, Optional
+import warnings
+
+from torch import Tensor, nn
+import torch.nn.functional as F
+
+
+class SwiGLUFFN(nn.Module):
+    def __init__(
+        self,
+        in_features: int,
+        hidden_features: Optional[int] = None,
+        out_features: Optional[int] = None,
+        act_layer: Callable[..., nn.Module] = None,
+        drop: float = 0.0,
+        bias: bool = True,
+    ) -> None:
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.w12 = nn.Linear(in_features, 2 * hidden_features, bias=bias)
+        self.w3 = nn.Linear(hidden_features, out_features, bias=bias)
+
+    def forward(self, x: Tensor) -> Tensor:
+        x12 = self.w12(x)
+        x1, x2 = x12.chunk(2, dim=-1)
+        hidden = F.silu(x1) * x2
+        return self.w3(hidden)
+
+
+XFORMERS_ENABLED = os.environ.get("XFORMERS_DISABLED") is None
+# try:
+#     if XFORMERS_ENABLED:
+#         from xformers.ops import SwiGLU
+
+#         XFORMERS_AVAILABLE = True
+#         warnings.warn("xFormers is available (SwiGLU)")
+#     else:
+#         warnings.warn("xFormers is disabled (SwiGLU)")
+#         raise ImportError
+# except ImportError:
+SwiGLU = SwiGLUFFN
+XFORMERS_AVAILABLE = False
+
+# warnings.warn("xFormers is not available (SwiGLU)")
+
+
+class SwiGLUFFNFused(SwiGLU):
+    def __init__(
+        self,
+        in_features: int,
+        hidden_features: Optional[int] = None,
+        out_features: Optional[int] = None,
+        act_layer: Callable[..., nn.Module] = None,
+        drop: float = 0.0,
+        bias: bool = True,
+    ) -> None:
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        hidden_features = (int(hidden_features * 2 / 3) + 7) // 8 * 8
+        super().__init__(
+            in_features=in_features,
+            hidden_features=hidden_features,
+            out_features=out_features,
+            bias=bias,
+        )

outdoor_v48_4gpu_v2/code/05_02-14:21:58/streamvggt/layers/vision_transformer.py

@@ -0,0 +1,398 @@
+from functools import partial
+import math
+import logging
+from typing import Sequence, Tuple, Union, Callable
+
+import torch
+import torch.nn as nn
+from torch.utils.checkpoint import checkpoint
+from torch.nn.init import trunc_normal_
+from . import Mlp, PatchEmbed, SwiGLUFFNFused, MemEffAttention, NestedTensorBlock as Block
+
+logger = logging.getLogger("dinov2")
+
+
+def named_apply(fn: Callable, module: nn.Module, name="", depth_first=True, include_root=False) -> nn.Module:
+    if not depth_first and include_root:
+        fn(module=module, name=name)
+    for child_name, child_module in module.named_children():
+        child_name = ".".join((name, child_name)) if name else child_name
+        named_apply(fn=fn, module=child_module, name=child_name, depth_first=depth_first, include_root=True)
+    if depth_first and include_root:
+        fn(module=module, name=name)
+    return module
+
+
+class BlockChunk(nn.ModuleList):
+    def forward(self, x):
+        for b in self:
+            x = b(x)
+        return x
+
+
+class DinoVisionTransformer(nn.Module):
+    def __init__(
+        self,
+        img_size=224,
+        patch_size=16,
+        in_chans=3,
+        embed_dim=768,
+        depth=12,
+        num_heads=12,
+        mlp_ratio=4.0,
+        qkv_bias=True,
+        ffn_bias=True,
+        proj_bias=True,
+        drop_path_rate=0.0,
+        drop_path_uniform=False,
+        init_values=None,  # for layerscale: None or 0 => no layerscale
+        embed_layer=PatchEmbed,
+        act_layer=nn.GELU,
+        block_fn=Block,
+        ffn_layer="mlp",
+        block_chunks=1,
+        num_register_tokens=0,
+        interpolate_antialias=False,
+        interpolate_offset=0.1,
+        qk_norm=False,
+    ):
+        """
+        Args:
+            img_size (int, tuple): input image size
+            patch_size (int, tuple): patch size
+            in_chans (int): number of input channels
+            embed_dim (int): embedding dimension
+            depth (int): depth of transformer
+            num_heads (int): number of attention heads
+            mlp_ratio (int): ratio of mlp hidden dim to embedding dim
+            qkv_bias (bool): enable bias for qkv if True
+            proj_bias (bool): enable bias for proj in attn if True
+            ffn_bias (bool): enable bias for ffn if True
+            drop_path_rate (float): stochastic depth rate
+            drop_path_uniform (bool): apply uniform drop rate across blocks
+            weight_init (str): weight init scheme
+            init_values (float): layer-scale init values
+            embed_layer (nn.Module): patch embedding layer
+            act_layer (nn.Module): MLP activation layer
+            block_fn (nn.Module): transformer block class
+            ffn_layer (str): "mlp", "swiglu", "swiglufused" or "identity"
+            block_chunks: (int) split block sequence into block_chunks units for FSDP wrap
+            num_register_tokens: (int) number of extra cls tokens (so-called "registers")
+            interpolate_antialias: (str) flag to apply anti-aliasing when interpolating positional embeddings
+            interpolate_offset: (float) work-around offset to apply when interpolating positional embeddings
+        """
+        super().__init__()
+        norm_layer = partial(nn.LayerNorm, eps=1e-6)
+
+        # tricky but makes it work
+        self.use_checkpoint = False
+        #
+
+        self.num_features = self.embed_dim = embed_dim  # num_features for consistency with other models
+        self.num_tokens = 1
+        self.n_blocks = depth
+        self.num_heads = num_heads
+        self.patch_size = patch_size
+        self.num_register_tokens = num_register_tokens
+        self.interpolate_antialias = interpolate_antialias
+        self.interpolate_offset = interpolate_offset
+
+        self.patch_embed = embed_layer(img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim)
+        num_patches = self.patch_embed.num_patches
+
+        self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
+        self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + self.num_tokens, embed_dim))
+        assert num_register_tokens >= 0
+        self.register_tokens = (
+            nn.Parameter(torch.zeros(1, num_register_tokens, embed_dim)) if num_register_tokens else None
+        )
+
+        if drop_path_uniform is True:
+            dpr = [drop_path_rate] * depth
+        else:
+            dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)]  # stochastic depth decay rule
+
+        if ffn_layer == "mlp":
+            logger.info("using MLP layer as FFN")
+            ffn_layer = Mlp
+        elif ffn_layer == "swiglufused" or ffn_layer == "swiglu":
+            logger.info("using SwiGLU layer as FFN")
+            ffn_layer = SwiGLUFFNFused
+        elif ffn_layer == "identity":
+            logger.info("using Identity layer as FFN")
+
+            def f(*args, **kwargs):
+                return nn.Identity()
+
+            ffn_layer = f
+        else:
+            raise NotImplementedError
+
+        blocks_list = [
+            block_fn(
+                dim=embed_dim,
+                num_heads=num_heads,
+                mlp_ratio=mlp_ratio,
+                qkv_bias=qkv_bias,
+                proj_bias=proj_bias,
+                ffn_bias=ffn_bias,
+                drop_path=dpr[i],
+                norm_layer=norm_layer,
+                act_layer=act_layer,
+                ffn_layer=ffn_layer,
+                init_values=init_values,
+                qk_norm=qk_norm,
+            )
+            for i in range(depth)
+        ]
+        if block_chunks > 0:
+            self.chunked_blocks = True
+            chunked_blocks = []
+            chunksize = depth // block_chunks
+            for i in range(0, depth, chunksize):
+                # this is to keep the block index consistent if we chunk the block list
+                chunked_blocks.append([nn.Identity()] * i + blocks_list[i : i + chunksize])
+            self.blocks = nn.ModuleList([BlockChunk(p) for p in chunked_blocks])
+        else:
+            self.chunked_blocks = False
+            self.blocks = nn.ModuleList(blocks_list)
+
+        self.norm = norm_layer(embed_dim)
+        self.head = nn.Identity()
+
+        self.mask_token = nn.Parameter(torch.zeros(1, embed_dim))
+
+        self.init_weights()
+
+    def init_weights(self):
+        trunc_normal_(self.pos_embed, std=0.02)
+        nn.init.normal_(self.cls_token, std=1e-6)
+        if self.register_tokens is not None:
+            nn.init.normal_(self.register_tokens, std=1e-6)
+        named_apply(init_weights_vit_timm, self)
+
+    def interpolate_pos_encoding(self, x, w, h):
+        previous_dtype = x.dtype
+        npatch = x.shape[1] - 1
+        N = self.pos_embed.shape[1] - 1
+        if npatch == N and w == h:
+            return self.pos_embed
+        pos_embed = self.pos_embed.float()
+        class_pos_embed = pos_embed[:, 0]
+        patch_pos_embed = pos_embed[:, 1:]
+        dim = x.shape[-1]
+        w0 = w // self.patch_size
+        h0 = h // self.patch_size
+        M = int(math.sqrt(N))  # Recover the number of patches in each dimension
+        assert N == M * M
+        kwargs = {}
+        if self.interpolate_offset:
+            # Historical kludge: add a small number to avoid floating point error in the interpolation, see https://github.com/facebookresearch/dino/issues/8
+            # Note: still needed for backward-compatibility, the underlying operators are using both output size and scale factors
+            sx = float(w0 + self.interpolate_offset) / M
+            sy = float(h0 + self.interpolate_offset) / M
+            kwargs["scale_factor"] = (sx, sy)
+        else:
+            # Simply specify an output size instead of a scale factor
+            kwargs["size"] = (w0, h0)
+        patch_pos_embed = nn.functional.interpolate(
+            patch_pos_embed.reshape(1, M, M, dim).permute(0, 3, 1, 2),
+            mode="bicubic",
+            antialias=self.interpolate_antialias,
+            **kwargs,
+        )
+        assert (w0, h0) == patch_pos_embed.shape[-2:]
+        patch_pos_embed = patch_pos_embed.permute(0, 2, 3, 1).view(1, -1, dim)
+        return torch.cat((class_pos_embed.unsqueeze(0), patch_pos_embed), dim=1).to(previous_dtype)
+
+    def prepare_tokens_with_masks(self, x, masks=None):
+        B, nc, w, h = x.shape
+        x = self.patch_embed(x)
+        if masks is not None:
+            x = torch.where(masks.unsqueeze(-1), self.mask_token.to(x.dtype).unsqueeze(0), x)
+
+        x = torch.cat((self.cls_token.expand(x.shape[0], -1, -1), x), dim=1)
+        x = x + self.interpolate_pos_encoding(x, w, h)
+
+        if self.register_tokens is not None:
+            x = torch.cat(
+                (
+                    x[:, :1],
+                    self.register_tokens.expand(x.shape[0], -1, -1),
+                    x[:, 1:],
+                ),
+                dim=1,
+            )
+
+        return x
+
+    def forward_features_list(self, x_list, masks_list):
+        x = [self.prepare_tokens_with_masks(x, masks) for x, masks in zip(x_list, masks_list)]
+
+        for blk in self.blocks:
+            if self.use_checkpoint:
+                x = checkpoint(blk, x, use_reentrant=self.use_reentrant)
+            else:
+                x = blk(x)
+
+        all_x = x
+        output = []
+        for x, masks in zip(all_x, masks_list):
+            x_norm = self.norm(x)
+            output.append(
+                {
+                    "x_norm_clstoken": x_norm[:, 0],
+                    "x_norm_regtokens": x_norm[:, 1 : self.num_register_tokens + 1],
+                    "x_norm_patchtokens": x_norm[:, self.num_register_tokens + 1 :],
+                    "x_prenorm": x,
+                    "masks": masks,
+                }
+            )
+        return output
+
+    def forward_features(self, x, masks=None):
+        if isinstance(x, list):
+            return self.forward_features_list(x, masks)
+
+        x = self.prepare_tokens_with_masks(x, masks)
+
+        for blk in self.blocks:
+            if self.use_checkpoint:
+                x = checkpoint(blk, x, use_reentrant=self.use_reentrant)
+            else:
+                x = blk(x)
+
+        x_norm = self.norm(x)
+        return {
+            "x_norm_clstoken": x_norm[:, 0],
+            "x_norm_regtokens": x_norm[:, 1 : self.num_register_tokens + 1],
+            "x_norm_patchtokens": x_norm[:, self.num_register_tokens + 1 :],
+            "x_prenorm": x,
+            "masks": masks,
+        }
+
+    def _get_intermediate_layers_not_chunked(self, x, n=1):
+        x = self.prepare_tokens_with_masks(x)
+        # If n is an int, take the n last blocks. If it's a list, take them
+        output, total_block_len = [], len(self.blocks)
+        blocks_to_take = range(total_block_len - n, total_block_len) if isinstance(n, int) else n
+        for i, blk in enumerate(self.blocks):
+            x = blk(x)
+            if i in blocks_to_take:
+                output.append(x)
+        assert len(output) == len(blocks_to_take), f"only {len(output)} / {len(blocks_to_take)} blocks found"
+        return output
+
+    def _get_intermediate_layers_chunked(self, x, n=1):
+        x = self.prepare_tokens_with_masks(x)
+        output, i, total_block_len = [], 0, len(self.blocks[-1])
+        # If n is an int, take the n last blocks. If it's a list, take them
+        blocks_to_take = range(total_block_len - n, total_block_len) if isinstance(n, int) else n
+        for block_chunk in self.blocks:
+            for blk in block_chunk[i:]:  # Passing the nn.Identity()
+                x = blk(x)
+                if i in blocks_to_take:
+                    output.append(x)
+                i += 1
+        assert len(output) == len(blocks_to_take), f"only {len(output)} / {len(blocks_to_take)} blocks found"
+        return output
+
+    def get_intermediate_layers(
+        self,
+        x: torch.Tensor,
+        n: Union[int, Sequence] = 1,  # Layers or n last layers to take
+        reshape: bool = False,
+        return_class_token: bool = False,
+        norm=True,
+    ) -> Tuple[Union[torch.Tensor, Tuple[torch.Tensor]]]:
+        if self.chunked_blocks:
+            outputs = self._get_intermediate_layers_chunked(x, n)
+        else:
+            outputs = self._get_intermediate_layers_not_chunked(x, n)
+        if norm:
+            outputs = [self.norm(out) for out in outputs]
+        class_tokens = [out[:, 0] for out in outputs]
+        outputs = [out[:, 1 + self.num_register_tokens :] for out in outputs]
+        if reshape:
+            B, _, w, h = x.shape
+            outputs = [
+                out.reshape(B, w // self.patch_size, h // self.patch_size, -1).permute(0, 3, 1, 2).contiguous()
+                for out in outputs
+            ]
+        if return_class_token:
+            return tuple(zip(outputs, class_tokens))
+        return tuple(outputs)
+
+    def forward(self, *args, is_training=True, **kwargs):
+        ret = self.forward_features(*args, **kwargs)
+        if is_training:
+            return ret
+        else:
+            return self.head(ret["x_norm_clstoken"])
+
+
+def init_weights_vit_timm(module: nn.Module, name: str = ""):
+    """ViT weight initialization, original timm impl (for reproducibility)"""
+    if isinstance(module, nn.Linear):
+        trunc_normal_(module.weight, std=0.02)
+        if module.bias is not None:
+            nn.init.zeros_(module.bias)
+
+
+def vit_small(patch_size=16, num_register_tokens=0, **kwargs):
+    model = DinoVisionTransformer(
+        patch_size=patch_size,
+        embed_dim=384,
+        depth=12,
+        num_heads=6,
+        mlp_ratio=4,
+        block_fn=partial(Block, attn_class=MemEffAttention),
+        num_register_tokens=num_register_tokens,
+        **kwargs,
+    )
+    return model
+
+
+def vit_base(patch_size=16, num_register_tokens=0, **kwargs):
+    model = DinoVisionTransformer(
+        patch_size=patch_size,
+        embed_dim=768,
+        depth=12,
+        num_heads=12,
+        mlp_ratio=4,
+        block_fn=partial(Block, attn_class=MemEffAttention),
+        num_register_tokens=num_register_tokens,
+        **kwargs,
+    )
+    return model
+
+
+def vit_large(patch_size=16, num_register_tokens=0, **kwargs):
+    model = DinoVisionTransformer(
+        patch_size=patch_size,
+        embed_dim=1024,
+        depth=24,
+        num_heads=16,
+        mlp_ratio=4,
+        block_fn=partial(Block, attn_class=MemEffAttention),
+        num_register_tokens=num_register_tokens,
+        **kwargs,
+    )
+    return model
+
+
+def vit_giant2(patch_size=16, num_register_tokens=0, **kwargs):
+    """
+    Close to ViT-giant, with embed-dim 1536 and 24 heads => embed-dim per head 64
+    """
+    model = DinoVisionTransformer(
+        patch_size=patch_size,
+        embed_dim=1536,
+        depth=40,
+        num_heads=24,
+        mlp_ratio=4,
+        block_fn=partial(Block, attn_class=MemEffAttention),
+        num_register_tokens=num_register_tokens,
+        **kwargs,
+    )
+    return model

outdoor_v48_4gpu_v2/code/05_02-14:21:58/streamvggt/models/aggregator.py

@@ -0,0 +1,394 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import logging
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from typing import Optional, Tuple, Union, List, Dict, Any
+
+from streamvggt.layers import PatchEmbed
+from streamvggt.layers.block import Block
+from streamvggt.layers.rope import RotaryPositionEmbedding2D, PositionGetter
+from streamvggt.layers.vision_transformer import vit_small, vit_base, vit_large, vit_giant2
+
+logger = logging.getLogger(__name__)
+
+_RESNET_MEAN = [0.485, 0.456, 0.406]
+_RESNET_STD = [0.229, 0.224, 0.225]
+
+
+class Aggregator(nn.Module):
+    """
+    The Aggregator applies alternating-attention over input frames,
+    as described in VGGT: Visual Geometry Grounded Transformer.
+
+
+    Args:
+        img_size (int): Image size in pixels.
+        patch_size (int): Size of each patch for PatchEmbed.
+        embed_dim (int): Dimension of the token embeddings.
+        depth (int): Number of blocks.
+        num_heads (int): Number of attention heads.
+        mlp_ratio (float): Ratio of MLP hidden dim to embedding dim.
+        num_register_tokens (int): Number of register tokens.
+        block_fn (nn.Module): The block type used for attention (Block by default).
+        qkv_bias (bool): Whether to include bias in QKV projections.
+        proj_bias (bool): Whether to include bias in the output projection.
+        ffn_bias (bool): Whether to include bias in MLP layers.
+        patch_embed (str): Type of patch embed. e.g., "conv" or "dinov2_vitl14_reg".
+        aa_order (list[str]): The order of alternating attention, e.g. ["frame", "global"].
+        aa_block_size (int): How many blocks to group under each attention type before switching. If not necessary, set to 1.
+        qk_norm (bool): Whether to apply QK normalization.
+        rope_freq (int): Base frequency for rotary embedding. -1 to disable.
+        init_values (float): Init scale for layer scale.
+    """
+
+    def __init__(
+        self,
+        img_size=518,
+        patch_size=14,
+        embed_dim=1024,
+        depth=24,
+        num_heads=16,
+        mlp_ratio=4.0,
+        num_register_tokens=4,
+        block_fn=Block,
+        qkv_bias=True,
+        proj_bias=True,
+        ffn_bias=True,
+        patch_embed="dinov2_vitl14_reg",
+        aa_order=["frame", "global"],
+        aa_block_size=1,
+        qk_norm=True,
+        rope_freq=100,
+        init_values=0.01,
+    ):
+        super().__init__()
+
+        self.__build_patch_embed__(patch_embed, img_size, patch_size, num_register_tokens, embed_dim=embed_dim)
+
+        # Initialize rotary position embedding if frequency > 0
+        self.rope = RotaryPositionEmbedding2D(frequency=rope_freq) if rope_freq > 0 else None
+        self.position_getter = PositionGetter() if self.rope is not None else None
+
+        self.frame_blocks = nn.ModuleList(
+            [
+                block_fn(
+                    dim=embed_dim,
+                    num_heads=num_heads,
+                    mlp_ratio=mlp_ratio,
+                    qkv_bias=qkv_bias,
+                    proj_bias=proj_bias,
+                    ffn_bias=ffn_bias,
+                    init_values=init_values,
+                    qk_norm=qk_norm,
+                    rope=self.rope,
+                )
+                for _ in range(depth)
+            ]
+        )
+
+        self.global_blocks = nn.ModuleList(
+            [
+                block_fn(
+                    dim=embed_dim,
+                    num_heads=num_heads,
+                    mlp_ratio=mlp_ratio,
+                    qkv_bias=qkv_bias,
+                    proj_bias=proj_bias,
+                    ffn_bias=ffn_bias,
+                    init_values=init_values,
+                    qk_norm=qk_norm,
+                    rope=self.rope,
+                )
+                for _ in range(depth)
+            ]
+        )
+
+        self.depth = depth
+        self.aa_order = aa_order
+        self.patch_size = patch_size
+        self.aa_block_size = aa_block_size
+
+        # Validate that depth is divisible by aa_block_size
+        if self.depth % self.aa_block_size != 0:
+            raise ValueError(f"depth ({depth}) must be divisible by aa_block_size ({aa_block_size})")
+
+        self.aa_block_num = self.depth // self.aa_block_size
+
+        # Note: We have two camera tokens, one for the first frame and one for the rest
+        # The same applies for register tokens
+        self.camera_token = nn.Parameter(torch.randn(1, 2, 1, embed_dim))
+        self.register_token = nn.Parameter(torch.randn(1, 2, num_register_tokens, embed_dim))
+
+        # The patch tokens start after the camera and register tokens
+        self.patch_start_idx = 1 + num_register_tokens
+
+        # Initialize parameters with small values
+        nn.init.normal_(self.camera_token, std=1e-6)
+        nn.init.normal_(self.register_token, std=1e-6)
+
+        # Register normalization constants as buffers
+        for name, value in (
+            ("_resnet_mean", _RESNET_MEAN),
+            ("_resnet_std", _RESNET_STD),
+        ):
+            self.register_buffer(
+                name,
+                torch.FloatTensor(value).reshape(1, 1, 3, 1, 1),
+                persistent=False,
+            )
+
+
+    def __build_patch_embed__(
+        self,
+        patch_embed,
+        img_size,
+        patch_size,
+        num_register_tokens,
+        interpolate_antialias=True,
+        interpolate_offset=0.0,
+        block_chunks=0,
+        init_values=1.0,
+        embed_dim=1024,
+    ):
+        """
+        Build the patch embed layer. If 'conv', we use a
+        simple PatchEmbed conv layer. Otherwise, we use a vision transformer.
+        """
+
+        if "conv" in patch_embed:
+            self.patch_embed = PatchEmbed(img_size=img_size, patch_size=patch_size, in_chans=3, embed_dim=embed_dim)
+        else:
+            vit_models = {
+                "dinov2_vitl14_reg": vit_large,
+                "dinov2_vitb14_reg": vit_base,
+                "dinov2_vits14_reg": vit_small,
+                "dinov2_vitg2_reg": vit_giant2,
+            }
+
+            self.patch_embed = vit_models[patch_embed](
+                img_size=img_size,
+                patch_size=patch_size,
+                num_register_tokens=num_register_tokens,
+                interpolate_antialias=interpolate_antialias,
+                interpolate_offset=interpolate_offset,
+                block_chunks=block_chunks,
+                init_values=init_values,
+            )
+
+            # Disable gradient updates for mask token
+            if hasattr(self.patch_embed, "mask_token"):
+                self.patch_embed.mask_token.requires_grad_(False)
+
+    def forward(
+        self,
+        images: torch.Tensor,
+        past_key_values=None,
+        use_cache=False,
+        past_frame_idx=0
+    ) -> Tuple[List[torch.Tensor], int]:
+        """
+        Args:
+            images (torch.Tensor): Input images with shape [B, S, 3, H, W], in range [0, 1].
+                B: batch size, S: sequence length, 3: RGB channels, H: height, W: width
+
+        Returns:
+            (list[torch.Tensor], int):
+                The list of outputs from the attention blocks,
+                and the patch_start_idx indicating where patch tokens begin.
+        """
+        B, S, C_in, H, W = images.shape
+
+        if use_cache and past_key_values[0] is not None:
+            _, _, S_true, _, _ = past_key_values[0][0].shape
+            S_true += 1
+        else:
+            S_true = S
+        
+        if use_cache and S > 1:
+            print(f"Use KV cache expects S=1, got S={S}")
+
+        if C_in != 3:
+            raise ValueError(f"Expected 3 input channels, got {C_in}")
+
+        # Normalize images and reshape for patch embed
+        images = (images - self._resnet_mean.to(images.device)) / self._resnet_std.to(images.device)
+
+        # Reshape to [B*S, C, H, W] for patch embedding
+        images = images.reshape(B * S, C_in, H, W)
+        patch_tokens = self.patch_embed(images)
+
+        if isinstance(patch_tokens, dict):
+            patch_tokens = patch_tokens["x_norm_patchtokens"]
+
+        _, P, C = patch_tokens.shape
+
+        if use_cache:
+            camera_token_full = slice_expand_and_flatten(self.camera_token, B, S_true)
+            camera_token = camera_token_full[-1:, :, :]
+            
+            register_token_full = slice_expand_and_flatten(self.register_token, B, S_true)
+            register_token = register_token_full[-1:, :, :]
+        else:
+            camera_token = slice_expand_and_flatten(self.camera_token, B, S)
+            register_token = slice_expand_and_flatten(self.register_token, B, S)
+        # Concatenate special tokens with patch tokens
+        tokens = torch.cat([camera_token, register_token, patch_tokens], dim=1)
+
+        pos = None
+        if self.rope is not None:
+            pos = self.position_getter(B * S, H // self.patch_size, W // self.patch_size, device=images.device)
+
+        if self.patch_start_idx > 0:
+            # do not use position embedding for special tokens (camera and register tokens)
+            # so set pos to 0 for the special tokens
+            pos = pos + 1
+            pos_special = torch.zeros(B * S, self.patch_start_idx, 2).to(images.device).to(pos.dtype)
+            pos = torch.cat([pos_special, pos], dim=1)
+
+        # update P because we added special tokens
+        _, P, C = tokens.shape
+
+        frame_idx = 0
+        global_idx = 0
+        output_list = []
+
+        for _ in range(self.aa_block_num):
+            for attn_type in self.aa_order:
+                if attn_type == "frame":
+                    tokens, frame_idx, frame_intermediates = self._process_frame_attention(
+                        tokens, B, S, P, C, frame_idx, pos=pos
+                    )
+                elif attn_type == "global":
+                    if use_cache:
+                        if past_key_values[global_idx] is not None:
+                            k, v = past_key_values[global_idx]
+                        tokens, global_idx, global_intermediates, new_kv = self._process_global_attention(
+                            tokens, B, S, P, C, global_idx, pos=pos,
+                            past_key_values_block=past_key_values[global_idx] if past_key_values[global_idx] is not None else None,
+                            use_cache=True,
+                            past_frame_idx=past_frame_idx
+                        )
+                        past_key_values[global_idx - 1] = new_kv
+                    else: 
+                        tokens, global_idx, global_intermediates = self._process_global_attention(
+                            tokens, B, S, P, C, global_idx, pos=pos
+                        )
+                else:
+                    raise ValueError(f"Unknown attention type: {attn_type}")
+            for i in range(len(frame_intermediates)):
+                # concat frame and global intermediates, [B x S x P x 2C]
+                concat_inter = torch.cat([frame_intermediates[i], global_intermediates[i]], dim=-1)
+                output_list.append(concat_inter)
+
+        del concat_inter
+        del frame_intermediates
+        del global_intermediates
+        if use_cache:      
+            return output_list, self.patch_start_idx, past_key_values
+        return output_list, self.patch_start_idx
+
+    def _process_frame_attention(self, tokens, B, S, P, C, frame_idx, pos=None):
+        """
+        Process frame attention blocks. We keep tokens in shape (B*S, P, C).
+        """
+        # If needed, reshape tokens or positions:
+        if tokens.shape != (B * S, P, C):
+            tokens = tokens.reshape(B, S, P, C).reshape(B * S, P, C)
+
+        if pos is not None and pos.shape != (B * S, P, 2):
+            pos = pos.reshape(B, S, P, 2).reshape(B * S, P, 2)
+
+        intermediates = []
+
+        # by default, self.aa_block_size=1, which processes one block at a time
+        for _ in range(self.aa_block_size):
+            tokens = self.frame_blocks[frame_idx](tokens, pos=pos)
+            frame_idx += 1
+            intermediates.append(tokens.reshape(B, S, P, C))
+
+        return tokens, frame_idx, intermediates
+
+    def _process_global_attention(
+        self,
+        tokens,
+        B,
+        S,
+        P,
+        C,
+        global_idx,
+        pos=None,
+        past_key_values_block=None,
+        use_cache=False,
+        past_frame_idx=0
+    ) -> Union[Tuple[torch.Tensor, int, List[torch.Tensor]], Tuple[torch.Tensor, int, List[torch.Tensor], List]]:
+        """
+        Process global attention blocks. We keep tokens in shape (B, S*P, C).
+                """
+        
+        if tokens.shape != (B, S * P, C):
+            tokens = tokens.reshape(B, S, P, C).reshape(B, S * P, C)
+
+        if pos is not None and pos.shape != (B, S * P, 2):
+            pos = pos.reshape(B, S, P, 2).reshape(B, S * P, 2)
+            
+        intermediates = []
+
+        for _ in range(self.aa_block_size):
+            if not use_cache:
+                L = S * P
+                frame_ids = torch.arange(L, device=tokens.device) // P  # [0,0,...,1,1,...,S-1]
+                future_frame = frame_ids.unsqueeze(1) < frame_ids.unsqueeze(0)
+                attn_mask = future_frame.to(tokens.dtype) * torch.finfo(tokens.dtype).min
+            else:
+                attn_mask = None
+                
+            if use_cache:
+                tokens, block_kv = self.global_blocks[global_idx](
+                    tokens, 
+                    pos=pos, 
+                    attn_mask=attn_mask, 
+                    past_key_values=past_key_values_block,
+                    use_cache=True
+                )
+            else:
+                tokens = self.global_blocks[global_idx](tokens, pos=pos, attn_mask=attn_mask)
+            global_idx += 1
+            intermediates.append(tokens.reshape(B, S, P, C))
+
+            # if self.use_causal_global:
+            #     del attn_mask
+        if use_cache:
+            return tokens, global_idx, intermediates, block_kv
+        return tokens, global_idx, intermediates
+
+
+def slice_expand_and_flatten(token_tensor, B, S):
+    """
+    Processes specialized tokens with shape (1, 2, X, C) for multi-frame processing:
+    1) Uses the first position (index=0) for the first frame only
+    2) Uses the second position (index=1) for all remaining frames (S-1 frames)
+    3) Expands both to match batch size B
+    4) Concatenates to form (B, S, X, C) where each sequence has 1 first-position token
+       followed by (S-1) second-position tokens
+    5) Flattens to (B*S, X, C) for processing
+
+    Returns:
+        torch.Tensor: Processed tokens with shape (B*S, X, C)
+    """
+
+    # Slice out the "query" tokens => shape (1, 1, ...)
+    query = token_tensor[:, 0:1, ...].expand(B, 1, *token_tensor.shape[2:])
+    # Slice out the "other" tokens => shape (1, S-1, ...)
+    others = token_tensor[:, 1:, ...].expand(B, S - 1, *token_tensor.shape[2:])
+    # Concatenate => shape (B, S, ...)
+    combined = torch.cat([query, others], dim=1)
+
+    # Finally flatten => shape (B*S, ...)
+    combined = combined.reshape(B * S, *combined.shape[2:])
+    return combined

outdoor_v48_4gpu_v2/code/05_02-14:21:58/streamvggt/models/streamvggt.py

@@ -0,0 +1,248 @@
+import torch
+import torch.nn as nn
+from huggingface_hub import PyTorchModelHubMixin  # used for model hub
+
+from streamvggt.models.aggregator import Aggregator
+from streamvggt.heads.camera_head import CameraHead
+from streamvggt.heads.dpt_head import DPTHead
+from streamvggt.heads.track_head import TrackHead
+from transformers.file_utils import ModelOutput
+from typing import Optional, Tuple, List, Any
+from dataclasses import dataclass
+import pdb
+
+@dataclass
+class StreamVGGTOutput(ModelOutput):
+    ress: Optional[List[dict]] = None
+    views: Optional[torch.Tensor] = None
+
+class StreamVGGT(nn.Module, PyTorchModelHubMixin):
+    def __init__(self, img_size=518, patch_size=14, embed_dim=1024):
+        super().__init__()
+
+        self.aggregator = Aggregator(img_size=img_size, patch_size=patch_size, embed_dim=embed_dim)
+        self.camera_head = CameraHead(dim_in=2 * embed_dim)
+        self.point_head = DPTHead(dim_in=2 * embed_dim, output_dim=4, activation="inv_log", conf_activation="expp1")
+        self.depth_head = DPTHead(dim_in=2 * embed_dim, output_dim=2, activation="exp", conf_activation="expp1")
+        self.track_head = TrackHead(dim_in=2 * embed_dim, patch_size=patch_size)
+    
+
+
+    def forward(
+        self,
+        views,
+        query_points: torch.Tensor = None,
+        history_info: Optional[dict] = None,
+        past_key_values=None,
+        use_cache=False,
+        past_frame_idx=0
+    ):
+        images = torch.stack(
+            [view["img"] for view in views], dim=0
+        ).permute(1, 0, 2, 3, 4)    # B S C H W
+
+        # If without batch dimension, add it
+        if len(images.shape) == 4:
+            images = images.unsqueeze(0)
+        if query_points is not None and len(query_points.shape) == 2:
+            query_points = query_points.unsqueeze(0)
+
+        if history_info is None:
+            history_info = {"token": None}
+
+        aggregated_tokens_list, patch_start_idx = self.aggregator(images)
+        predictions = {}
+
+        with torch.cuda.amp.autocast(enabled=False):
+            if self.camera_head is not None:
+                pose_enc_list = self.camera_head(aggregated_tokens_list)
+                predictions["pose_enc"] = pose_enc_list[-1]  # pose encoding of the last iteration
+
+            if self.depth_head is not None:
+                depth, depth_conf = self.depth_head(
+                    aggregated_tokens_list, images=images, patch_start_idx=patch_start_idx
+                )
+                predictions["depth"] = depth
+                predictions["depth_conf"] = depth_conf
+
+            if self.point_head is not None:
+                pts3d, pts3d_conf = self.point_head(
+                    aggregated_tokens_list, images=images, patch_start_idx=patch_start_idx
+                )
+                predictions["world_points"] = pts3d
+                predictions["world_points_conf"] = pts3d_conf
+
+            if self.track_head is not None and query_points is not None:
+                track_list, vis, conf = self.track_head(
+                    aggregated_tokens_list, images=images, patch_start_idx=patch_start_idx, query_points=query_points
+                )
+                predictions["track"] = track_list[-1]  # track of the last iteration
+                predictions["vis"] = vis
+                predictions["conf"] = conf
+            predictions["images"] = images
+
+            B, S = images.shape[:2]
+            ress = []
+            for s in range(S):
+                res = {
+                    'pts3d_in_other_view': predictions['world_points'][:, s],  # [B, H, W, 3]
+                    'conf': predictions['world_points_conf'][:, s],  # [B, H, W]
+
+                    'depth': predictions['depth'][:, s],  # [B, H, W, 1]
+                    'depth_conf': predictions['depth_conf'][:, s],  # [B, H, W]
+                    'camera_pose': predictions['pose_enc'][:, s, :],  # [B, 9]
+
+                    **({'valid_mask': views[s]["valid_mask"]}
+                    if 'valid_mask' in views[s] else {}),  # [B, H, W]
+
+                    **({'track': predictions['track'][:, s],  # [B, N, 2]
+                        'vis': predictions['vis'][:, s],  # [B, N]
+                        'track_conf': predictions['conf'][:, s]}
+                    if 'track' in predictions else {})
+                }
+                ress.append(res)
+            return StreamVGGTOutput(ress=ress, views=views)  # [S] [B, C, H, W]
+        
+    def frontendT(self, frame):
+        images = frame[None, None] # 1,1,C,H,W
+        B,S,C,H,W = images.shape
+        #if self.frontend_past_key_values is None:
+        with torch.no_grad():
+            if not hasattr(self,"frontend_past_key_values"):
+                self.frontend_images_size = (B,C,H,W)
+                self.frontend_past_key_values = [None] * self.aggregator.depth
+                self.frontend_kid = 0
+            else:
+                self.frontend_kid += 1
+
+
+            aggregator_output = self.aggregator(
+                images, 
+                past_key_values=self.frontend_past_key_values,
+                use_cache=True, 
+                past_frame_idx=self.frontend_kid
+            )
+            aggregated_tokens, patch_start_idx, self.frontend_past_key_values = aggregator_output
+            aggregated_tokens = [t_.detach() for t_ in aggregated_tokens]
+
+        return aggregated_tokens
+
+
+    def extract(self, map_tokens, query_points=None):
+        B,C,H,W = self.frontend_images_size
+        S = map_tokens[0].shape[1]
+        images = torch.zeros((B,S,C,H,W)).to('cuda')
+
+        aggregated_tokens = map_tokens
+        patch_start_idx = self.aggregator.patch_start_idx 
+        all_ress = []
+
+        with torch.no_grad():
+          with torch.cuda.amp.autocast(enabled=False):
+            if self.camera_head is not None:
+                pose_enc = self.camera_head(aggregated_tokens)
+                pose_enc = pose_enc[-1]
+                camera_pose = pose_enc # 1,S,9
+
+            if self.depth_head is not None:
+                depth, depth_conf = self.depth_head(
+                    aggregated_tokens, images=images, patch_start_idx=patch_start_idx
+                )
+                # 1,S,H,W,1     
+                # 1,S,H,W
+            
+            if self.point_head is not None:
+                pts3d, pts3d_conf = self.point_head(
+                    aggregated_tokens, images=images, patch_start_idx=patch_start_idx
+                )
+                # 1,S,H,W,3
+                # 1.S,H,W
+
+            if self.track_head is not None and query_points is not None:
+                track_list, vis, conf = self.track_head(
+                    aggregated_tokens, images=images, patch_start_idx=patch_start_idx, query_points=query_points
+            )
+                track = track_list[-1][:, 0]  
+                query_points = track
+                vis = vis[:, 0]
+                track_conf = conf[:, 0]
+
+        output = {
+            'pts3d_in_other_view': pts3d,
+            'conf': pts3d_conf,
+            'depth': depth,
+            'depth_conf': depth_conf,
+            'camera_pose': camera_pose,
+            }
+        #output = StreamVGGTOutput(ress=all_ress, views=processed_frames)
+        return output
+        
+
+    def inference(self, frames, query_points: torch.Tensor = None, past_key_values=None):        
+        past_key_values = [None] * self.aggregator.depth
+        past_key_values_camera = [None] * self.camera_head.trunk_depth
+        
+        all_ress = []
+        processed_frames = []
+
+        for i, frame in enumerate(frames):
+            images = frame["img"].unsqueeze(0) 
+            aggregator_output = self.aggregator(
+                images, 
+                past_key_values=past_key_values,
+                use_cache=True, 
+                past_frame_idx=i
+            )
+            
+            if isinstance(aggregator_output, tuple) and len(aggregator_output) == 3:
+                aggregated_tokens, patch_start_idx, past_key_values = aggregator_output
+            else:
+                aggregated_tokens, patch_start_idx = aggregator_output
+            
+            with torch.cuda.amp.autocast(enabled=False):
+                if self.camera_head is not None:
+                    pose_enc, past_key_values_camera = self.camera_head(aggregated_tokens, past_key_values_camera=past_key_values_camera, use_cache=True)
+                    pose_enc = pose_enc[-1]
+                    camera_pose = pose_enc[:, 0, :]
+
+                if self.depth_head is not None:
+                    depth, depth_conf = self.depth_head(
+                        aggregated_tokens, images=images, patch_start_idx=patch_start_idx
+                    )
+                    depth = depth[:, 0] 
+                    depth_conf = depth_conf[:, 0]
+                
+                if self.point_head is not None:
+                    pts3d, pts3d_conf = self.point_head(
+                        aggregated_tokens, images=images, patch_start_idx=patch_start_idx
+                    )
+                    pts3d = pts3d[:, 0] 
+                    pts3d_conf = pts3d_conf[:, 0]
+
+                if self.track_head is not None and query_points is not None:
+                    track_list, vis, conf = self.track_head(
+                        aggregated_tokens, images=images, patch_start_idx=patch_start_idx, query_points=query_points
+                )
+                    track = track_list[-1][:, 0]  
+                    query_points = track
+                    vis = vis[:, 0]
+                    track_conf = conf[:, 0]
+
+            all_ress.append({
+                'pts3d_in_other_view': pts3d,
+                'conf': pts3d_conf,
+                'depth': depth,
+                'depth_conf': depth_conf,
+                'camera_pose': camera_pose,
+                **({'valid_mask': frame["valid_mask"]}
+                    if 'valid_mask' in frame else {}),  
+
+                **({'track': track, 
+                    'vis': vis,  
+                    'track_conf': track_conf}
+                if query_points is not None else {})
+            })
+            processed_frames.append(frame)
+        
+        output = StreamVGGTOutput(ress=all_ress, views=processed_frames)
+        return output

outdoor_v48_4gpu_v2/code/05_02-14:21:58/streamvggt/utils/geometry.py

@@ -0,0 +1,166 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import os
+import torch
+import numpy as np
+
+
+def unproject_depth_map_to_point_map(
+    depth_map: np.ndarray, extrinsics_cam: np.ndarray, intrinsics_cam: np.ndarray
+) -> np.ndarray:
+    """
+    Unproject a batch of depth maps to 3D world coordinates.
+
+    Args:
+        depth_map (np.ndarray): Batch of depth maps of shape (S, H, W, 1) or (S, H, W)
+        extrinsics_cam (np.ndarray): Batch of camera extrinsic matrices of shape (S, 3, 4)
+        intrinsics_cam (np.ndarray): Batch of camera intrinsic matrices of shape (S, 3, 3)
+
+    Returns:
+        np.ndarray: Batch of 3D world coordinates of shape (S, H, W, 3)
+    """
+    if isinstance(depth_map, torch.Tensor):
+        depth_map = depth_map.cpu().numpy()
+    if isinstance(extrinsics_cam, torch.Tensor):
+        extrinsics_cam = extrinsics_cam.cpu().numpy()
+    if isinstance(intrinsics_cam, torch.Tensor):
+        intrinsics_cam = intrinsics_cam.cpu().numpy()
+
+    world_points_list = []
+    for frame_idx in range(depth_map.shape[0]):
+        cur_world_points, _, _ = depth_to_world_coords_points(
+            depth_map[frame_idx].squeeze(-1), extrinsics_cam[frame_idx], intrinsics_cam[frame_idx]
+        )
+        world_points_list.append(cur_world_points)
+    world_points_array = np.stack(world_points_list, axis=0)
+
+    return world_points_array
+
+
+def depth_to_world_coords_points(
+    depth_map: np.ndarray,
+    extrinsic: np.ndarray,
+    intrinsic: np.ndarray,
+    eps=1e-8,
+) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
+    """
+    Convert a depth map to world coordinates.
+
+    Args:
+        depth_map (np.ndarray): Depth map of shape (H, W).
+        intrinsic (np.ndarray): Camera intrinsic matrix of shape (3, 3).
+        extrinsic (np.ndarray): Camera extrinsic matrix of shape (3, 4). OpenCV camera coordinate convention, cam from world.
+
+    Returns:
+        tuple[np.ndarray, np.ndarray]: World coordinates (H, W, 3) and valid depth mask (H, W).
+    """
+    if depth_map is None:
+        return None, None, None
+
+    # Valid depth mask
+    point_mask = depth_map > eps
+
+    # Convert depth map to camera coordinates
+    cam_coords_points = depth_to_cam_coords_points(depth_map, intrinsic)
+
+    # Multiply with the inverse of extrinsic matrix to transform to world coordinates
+    # extrinsic_inv is 4x4 (note closed_form_inverse_OpenCV is batched, the output is (N, 4, 4))
+    cam_to_world_extrinsic = closed_form_inverse_se3(extrinsic[None])[0]
+
+    R_cam_to_world = cam_to_world_extrinsic[:3, :3]
+    t_cam_to_world = cam_to_world_extrinsic[:3, 3]
+
+    # Apply the rotation and translation to the camera coordinates
+    world_coords_points = np.dot(cam_coords_points, R_cam_to_world.T) + t_cam_to_world  # HxWx3, 3x3 -> HxWx3
+    # world_coords_points = np.einsum("ij,hwj->hwi", R_cam_to_world, cam_coords_points) + t_cam_to_world
+
+    return world_coords_points, cam_coords_points, point_mask
+
+
+def depth_to_cam_coords_points(depth_map: np.ndarray, intrinsic: np.ndarray) -> tuple[np.ndarray, np.ndarray]:
+    """
+    Convert a depth map to camera coordinates.
+
+    Args:
+        depth_map (np.ndarray): Depth map of shape (H, W).
+        intrinsic (np.ndarray): Camera intrinsic matrix of shape (3, 3).
+
+    Returns:
+        tuple[np.ndarray, np.ndarray]: Camera coordinates (H, W, 3)
+    """
+    H, W = depth_map.shape
+    assert intrinsic.shape == (3, 3), "Intrinsic matrix must be 3x3"
+    assert intrinsic[0, 1] == 0 and intrinsic[1, 0] == 0, "Intrinsic matrix must have zero skew"
+
+    # Intrinsic parameters
+    fu, fv = intrinsic[0, 0], intrinsic[1, 1]
+    cu, cv = intrinsic[0, 2], intrinsic[1, 2]
+
+    # Generate grid of pixel coordinates
+    u, v = np.meshgrid(np.arange(W), np.arange(H))
+
+    # Unproject to camera coordinates
+    x_cam = (u - cu) * depth_map / fu
+    y_cam = (v - cv) * depth_map / fv
+    z_cam = depth_map
+
+    # Stack to form camera coordinates
+    cam_coords = np.stack((x_cam, y_cam, z_cam), axis=-1).astype(np.float32)
+
+    return cam_coords
+
+
+def closed_form_inverse_se3(se3, R=None, T=None):
+    """
+    Compute the inverse of each 4x4 (or 3x4) SE3 matrix in a batch.
+
+    If `R` and `T` are provided, they must correspond to the rotation and translation
+    components of `se3`. Otherwise, they will be extracted from `se3`.
+
+    Args:
+        se3: Nx4x4 or Nx3x4 array or tensor of SE3 matrices.
+        R (optional): Nx3x3 array or tensor of rotation matrices.
+        T (optional): Nx3x1 array or tensor of translation vectors.
+
+    Returns:
+        Inverted SE3 matrices with the same type and device as `se3`.
+
+    Shapes:
+        se3: (N, 4, 4)
+        R: (N, 3, 3)
+        T: (N, 3, 1)
+    """
+    # Check if se3 is a numpy array or a torch tensor
+    is_numpy = isinstance(se3, np.ndarray)
+
+    # Validate shapes
+    if se3.shape[-2:] != (4, 4) and se3.shape[-2:] != (3, 4):
+        raise ValueError(f"se3 must be of shape (N,4,4), got {se3.shape}.")
+
+    # Extract R and T if not provided
+    if R is None:
+        R = se3[:, :3, :3]  # (N,3,3)
+    if T is None:
+        T = se3[:, :3, 3:]  # (N,3,1)
+
+    # Transpose R
+    if is_numpy:
+        # Compute the transpose of the rotation for NumPy
+        R_transposed = np.transpose(R, (0, 2, 1))
+        # -R^T t for NumPy
+        top_right = -np.matmul(R_transposed, T)
+        inverted_matrix = np.tile(np.eye(4), (len(R), 1, 1))
+    else:
+        R_transposed = R.transpose(1, 2)  # (N,3,3)
+        top_right = -torch.bmm(R_transposed, T)  # (N,3,1)
+        inverted_matrix = torch.eye(4, 4)[None].repeat(len(R), 1, 1)
+        inverted_matrix = inverted_matrix.to(R.dtype).to(R.device)
+
+    inverted_matrix[:, :3, :3] = R_transposed
+    inverted_matrix[:, :3, 3:] = top_right
+
+    return inverted_matrix

outdoor_v48_4gpu_v2/code/05_02-14:21:58/streamvggt/utils/load_fn.py

@@ -0,0 +1,146 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+from PIL import Image
+from torchvision import transforms as TF
+
+
+def load_and_preprocess_images(image_path_list, mode="crop"):
+    """
+    A quick start function to load and preprocess images for model input.
+    This assumes the images should have the same shape for easier batching, but our model can also work well with different shapes.
+
+    Args:
+        image_path_list (list): List of paths to image files
+        mode (str, optional): Preprocessing mode, either "crop" or "pad".
+                             - "crop" (default): Sets width to 518px and center crops height if needed.
+                             - "pad": Preserves all pixels by making the largest dimension 518px
+                               and padding the smaller dimension to reach a square shape.
+
+    Returns:
+        torch.Tensor: Batched tensor of preprocessed images with shape (N, 3, H, W)
+
+    Raises:
+        ValueError: If the input list is empty or if mode is invalid
+
+    Notes:
+        - Images with different dimensions will be padded with white (value=1.0)
+        - A warning is printed when images have different shapes
+        - When mode="crop": The function ensures width=518px while maintaining aspect ratio
+          and height is center-cropped if larger than 518px
+        - When mode="pad": The function ensures the largest dimension is 518px while maintaining aspect ratio
+          and the smaller dimension is padded to reach a square shape (518x518)
+        - Dimensions are adjusted to be divisible by 14 for compatibility with model requirements
+    """
+    # Check for empty list
+    if len(image_path_list) == 0:
+        raise ValueError("At least 1 image is required")
+
+    # Validate mode
+    if mode not in ["crop", "pad"]:
+        raise ValueError("Mode must be either 'crop' or 'pad'")
+
+    images = []
+    shapes = set()
+    to_tensor = TF.ToTensor()
+    target_size = 518
+
+    # First process all images and collect their shapes
+    for image_path in image_path_list:
+
+        # Open image
+        img = Image.open(image_path)
+
+        # If there's an alpha channel, blend onto white background:
+        if img.mode == "RGBA":
+            # Create white background
+            background = Image.new("RGBA", img.size, (255, 255, 255, 255))
+            # Alpha composite onto the white background
+            img = Image.alpha_composite(background, img)
+
+        # Now convert to "RGB" (this step assigns white for transparent areas)
+        img = img.convert("RGB")
+
+        width, height = img.size
+
+        if mode == "pad":
+            # Make the largest dimension 518px while maintaining aspect ratio
+            if width >= height:
+                new_width = target_size
+                new_height = round(height * (new_width / width) / 14) * 14  # Make divisible by 14
+            else:
+                new_height = target_size
+                new_width = round(width * (new_height / height) / 14) * 14  # Make divisible by 14
+        else:  # mode == "crop"
+            # Original behavior: set width to 518px
+            new_width = target_size
+            # Calculate height maintaining aspect ratio, divisible by 14
+            new_height = round(height * (new_width / width) / 14) * 14
+
+        # Resize with new dimensions (width, height)
+        img = img.resize((new_width, new_height), Image.Resampling.BICUBIC)
+        img = to_tensor(img)  # Convert to tensor (0, 1)
+
+        # Center crop height if it's larger than 518 (only in crop mode)
+        if mode == "crop" and new_height > target_size:
+            start_y = (new_height - target_size) // 2
+            img = img[:, start_y: start_y + target_size, :]
+
+        # For pad mode, pad to make a square of target_size x target_size
+        if mode == "pad":
+            h_padding = target_size - img.shape[1]
+            w_padding = target_size - img.shape[2]
+
+            if h_padding > 0 or w_padding > 0:
+                pad_top = h_padding // 2
+                pad_bottom = h_padding - pad_top
+                pad_left = w_padding // 2
+                pad_right = w_padding - pad_left
+
+                # Pad with white (value=1.0)
+                img = torch.nn.functional.pad(
+                    img, (pad_left, pad_right, pad_top, pad_bottom), mode="constant", value=1.0
+                )
+
+        shapes.add((img.shape[1], img.shape[2]))
+        images.append(img)
+
+    # Check if we have different shapes
+    # In theory our model can also work well with different shapes
+    if len(shapes) > 1:
+        print(f"Warning: Found images with different shapes: {shapes}")
+        # Find maximum dimensions
+        max_height = max(shape[0] for shape in shapes)
+        max_width = max(shape[1] for shape in shapes)
+
+        # Pad images if necessary
+        padded_images = []
+        for img in images:
+            h_padding = max_height - img.shape[1]
+            w_padding = max_width - img.shape[2]
+
+            if h_padding > 0 or w_padding > 0:
+                pad_top = h_padding // 2
+                pad_bottom = h_padding - pad_top
+                pad_left = w_padding // 2
+                pad_right = w_padding - pad_left
+
+                img = torch.nn.functional.pad(
+                    img, (pad_left, pad_right, pad_top, pad_bottom), mode="constant", value=1.0
+                )
+            padded_images.append(img)
+        images = padded_images
+
+    images = torch.stack(images)  # concatenate images
+
+    # Ensure correct shape when single image
+    if len(image_path_list) == 1:
+        # Verify shape is (1, C, H, W)
+        if images.dim() == 3:
+            images = images.unsqueeze(0)
+
+    return images

outdoor_v48_4gpu_v2/code/05_02-14:21:58/streamvggt/utils/pose_enc.py

@@ -0,0 +1,130 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+from .rotation import quat_to_mat, mat_to_quat
+
+
+def extri_intri_to_pose_encoding(
+    extrinsics,
+    intrinsics,
+    image_size_hw=None,  # e.g., (256, 512)
+    pose_encoding_type="absT_quaR_FoV",
+):
+    """Convert camera extrinsics and intrinsics to a compact pose encoding.
+
+    This function transforms camera parameters into a unified pose encoding format,
+    which can be used for various downstream tasks like pose prediction or representation.
+
+    Args:
+        extrinsics (torch.Tensor): Camera extrinsic parameters with shape BxSx3x4,
+            where B is batch size and S is sequence length.
+            In OpenCV coordinate system (x-right, y-down, z-forward), representing camera from world transformation.
+            The format is [R|t] where R is a 3x3 rotation matrix and t is a 3x1 translation vector.
+        intrinsics (torch.Tensor): Camera intrinsic parameters with shape BxSx3x3.
+            Defined in pixels, with format:
+            [[fx, 0, cx],
+             [0, fy, cy],
+             [0,  0,  1]]
+            where fx, fy are focal lengths and (cx, cy) is the principal point
+        image_size_hw (tuple): Tuple of (height, width) of the image in pixels.
+            Required for computing field of view values. For example: (256, 512).
+        pose_encoding_type (str): Type of pose encoding to use. Currently only
+            supports "absT_quaR_FoV" (absolute translation, quaternion rotation, field of view).
+
+    Returns:
+        torch.Tensor: Encoded camera pose parameters with shape BxSx9.
+            For "absT_quaR_FoV" type, the 9 dimensions are:
+            - [:3] = absolute translation vector T (3D)
+            - [3:7] = rotation as quaternion quat (4D)
+            - [7:] = field of view (2D)
+    """
+
+    # extrinsics: BxSx3x4
+    # intrinsics: BxSx3x3
+
+    if pose_encoding_type == "absT_quaR_FoV":
+        R = extrinsics[:, :, :3, :3]  # BxSx3x3
+        T = extrinsics[:, :, :3, 3]  # BxSx3
+
+        quat = mat_to_quat(R)
+        # Note the order of h and w here
+        H, W = image_size_hw
+        fov_h = 2 * torch.atan((H / 2) / intrinsics[..., 1, 1])
+        fov_w = 2 * torch.atan((W / 2) / intrinsics[..., 0, 0])
+        pose_encoding = torch.cat([T, quat, fov_h[..., None], fov_w[..., None]], dim=-1).float()
+    else:
+        raise NotImplementedError
+
+    return pose_encoding
+
+
+def pose_encoding_to_extri_intri(
+    pose_encoding,
+    image_size_hw=None,  # e.g., (256, 512)
+    pose_encoding_type="absT_quaR_FoV",
+    build_intrinsics=True,
+):
+    """Convert a pose encoding back to camera extrinsics and intrinsics.
+
+    This function performs the inverse operation of extri_intri_to_pose_encoding,
+    reconstructing the full camera parameters from the compact encoding.
+
+    Args:
+        pose_encoding (torch.Tensor): Encoded camera pose parameters with shape BxSx9,
+            where B is batch size and S is sequence length.
+            For "absT_quaR_FoV" type, the 9 dimensions are:
+            - [:3] = absolute translation vector T (3D)
+            - [3:7] = rotation as quaternion quat (4D)
+            - [7:] = field of view (2D)
+        image_size_hw (tuple): Tuple of (height, width) of the image in pixels.
+            Required for reconstructing intrinsics from field of view values.
+            For example: (256, 512).
+        pose_encoding_type (str): Type of pose encoding used. Currently only
+            supports "absT_quaR_FoV" (absolute translation, quaternion rotation, field of view).
+        build_intrinsics (bool): Whether to reconstruct the intrinsics matrix.
+            If False, only extrinsics are returned and intrinsics will be None.
+
+    Returns:
+        tuple: (extrinsics, intrinsics)
+            - extrinsics (torch.Tensor): Camera extrinsic parameters with shape BxSx3x4.
+              In OpenCV coordinate system (x-right, y-down, z-forward), representing camera from world
+              transformation. The format is [R|t] where R is a 3x3 rotation matrix and t is
+              a 3x1 translation vector.
+            - intrinsics (torch.Tensor or None): Camera intrinsic parameters with shape BxSx3x3,
+              or None if build_intrinsics is False. Defined in pixels, with format:
+              [[fx, 0, cx],
+               [0, fy, cy],
+               [0,  0,  1]]
+              where fx, fy are focal lengths and (cx, cy) is the principal point,
+              assumed to be at the center of the image (W/2, H/2).
+    """
+
+    intrinsics = None
+
+    if pose_encoding_type == "absT_quaR_FoV":
+        T = pose_encoding[..., :3]
+        quat = pose_encoding[..., 3:7]
+        fov_h = pose_encoding[..., 7]
+        fov_w = pose_encoding[..., 8]
+
+        R = quat_to_mat(quat)
+        extrinsics = torch.cat([R, T[..., None]], dim=-1)
+
+        if build_intrinsics:
+            H, W = image_size_hw
+            fy = (H / 2.0) / torch.tan(fov_h / 2.0)
+            fx = (W / 2.0) / torch.tan(fov_w / 2.0)
+            intrinsics = torch.zeros(pose_encoding.shape[:2] + (3, 3), device=pose_encoding.device)
+            intrinsics[..., 0, 0] = fx
+            intrinsics[..., 1, 1] = fy
+            intrinsics[..., 0, 2] = W / 2
+            intrinsics[..., 1, 2] = H / 2
+            intrinsics[..., 2, 2] = 1.0  # Set the homogeneous coordinate to 1
+    else:
+        raise NotImplementedError
+
+    return extrinsics, intrinsics

outdoor_v48_4gpu_v2/code/05_02-14:21:58/streamvggt/utils/rotation.py

@@ -0,0 +1,138 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+# Modified from PyTorch3D, https://github.com/facebookresearch/pytorch3d
+
+import torch
+import numpy as np
+import torch.nn.functional as F
+
+
+def quat_to_mat(quaternions: torch.Tensor) -> torch.Tensor:
+    """
+    Quaternion Order: XYZW or say ijkr, scalar-last
+
+    Convert rotations given as quaternions to rotation matrices.
+    Args:
+        quaternions: quaternions with real part last,
+            as tensor of shape (..., 4).
+
+    Returns:
+        Rotation matrices as tensor of shape (..., 3, 3).
+    """
+    i, j, k, r = torch.unbind(quaternions, -1)
+    # pyre-fixme[58]: `/` is not supported for operand types `float` and `Tensor`.
+    two_s = 2.0 / (quaternions * quaternions).sum(-1)
+
+    o = torch.stack(
+        (
+            1 - two_s * (j * j + k * k),
+            two_s * (i * j - k * r),
+            two_s * (i * k + j * r),
+            two_s * (i * j + k * r),
+            1 - two_s * (i * i + k * k),
+            two_s * (j * k - i * r),
+            two_s * (i * k - j * r),
+            two_s * (j * k + i * r),
+            1 - two_s * (i * i + j * j),
+        ),
+        -1,
+    )
+    return o.reshape(quaternions.shape[:-1] + (3, 3))
+
+
+def mat_to_quat(matrix: torch.Tensor) -> torch.Tensor:
+    """
+    Convert rotations given as rotation matrices to quaternions.
+
+    Args:
+        matrix: Rotation matrices as tensor of shape (..., 3, 3).
+
+    Returns:
+        quaternions with real part last, as tensor of shape (..., 4).
+        Quaternion Order: XYZW or say ijkr, scalar-last
+    """
+    if matrix.size(-1) != 3 or matrix.size(-2) != 3:
+        raise ValueError(f"Invalid rotation matrix shape {matrix.shape}.")
+
+    batch_dim = matrix.shape[:-2]
+    m00, m01, m02, m10, m11, m12, m20, m21, m22 = torch.unbind(matrix.reshape(batch_dim + (9,)), dim=-1)
+
+    q_abs = _sqrt_positive_part(
+        torch.stack(
+            [
+                1.0 + m00 + m11 + m22,
+                1.0 + m00 - m11 - m22,
+                1.0 - m00 + m11 - m22,
+                1.0 - m00 - m11 + m22,
+            ],
+            dim=-1,
+        )
+    )
+
+    # we produce the desired quaternion multiplied by each of r, i, j, k
+    quat_by_rijk = torch.stack(
+        [
+            # pyre-fixme[58]: `**` is not supported for operand types `Tensor` and
+            #  `int`.
+            torch.stack([q_abs[..., 0] ** 2, m21 - m12, m02 - m20, m10 - m01], dim=-1),
+            # pyre-fixme[58]: `**` is not supported for operand types `Tensor` and
+            #  `int`.
+            torch.stack([m21 - m12, q_abs[..., 1] ** 2, m10 + m01, m02 + m20], dim=-1),
+            # pyre-fixme[58]: `**` is not supported for operand types `Tensor` and
+            #  `int`.
+            torch.stack([m02 - m20, m10 + m01, q_abs[..., 2] ** 2, m12 + m21], dim=-1),
+            # pyre-fixme[58]: `**` is not supported for operand types `Tensor` and
+            #  `int`.
+            torch.stack([m10 - m01, m20 + m02, m21 + m12, q_abs[..., 3] ** 2], dim=-1),
+        ],
+        dim=-2,
+    )
+
+    # We floor here at 0.1 but the exact level is not important; if q_abs is small,
+    # the candidate won't be picked.
+    flr = torch.tensor(0.1).to(dtype=q_abs.dtype, device=q_abs.device)
+    quat_candidates = quat_by_rijk / (2.0 * q_abs[..., None].max(flr))
+
+    # if not for numerical problems, quat_candidates[i] should be same (up to a sign),
+    # forall i; we pick the best-conditioned one (with the largest denominator)
+    out = quat_candidates[F.one_hot(q_abs.argmax(dim=-1), num_classes=4) > 0.5, :].reshape(batch_dim + (4,))
+
+    # Convert from rijk to ijkr
+    out = out[..., [1, 2, 3, 0]]
+
+    out = standardize_quaternion(out)
+
+    return out
+
+
+def _sqrt_positive_part(x: torch.Tensor) -> torch.Tensor:
+    """
+    Returns torch.sqrt(torch.max(0, x))
+    but with a zero subgradient where x is 0.
+    """
+    ret = torch.zeros_like(x)
+    positive_mask = x > 0
+    if torch.is_grad_enabled():
+        ret[positive_mask] = torch.sqrt(x[positive_mask])
+    else:
+        ret = torch.where(positive_mask, torch.sqrt(x), ret)
+    return ret
+
+
+def standardize_quaternion(quaternions: torch.Tensor) -> torch.Tensor:
+    """
+    Convert a unit quaternion to a standard form: one in which the real
+    part is non negative.
+
+    Args:
+        quaternions: Quaternions with real part last,
+            as tensor of shape (..., 4).
+
+    Returns:
+        Standardized quaternions as tensor of shape (..., 4).
+    """
+    return torch.where(quaternions[..., 3:4] < 0, -quaternions, quaternions)

outdoor_v48_4gpu_v2/code/05_02-14:21:58/streamvggt/utils/visual_track.py

@@ -0,0 +1,239 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import cv2
+import torch
+import numpy as np
+import os
+
+
+def color_from_xy(x, y, W, H, cmap_name="hsv"):
+    """
+    Map (x, y) -> color in (R, G, B).
+    1) Normalize x,y to [0,1].
+    2) Combine them into a single scalar c in [0,1].
+    3) Use matplotlib's colormap to convert c -> (R,G,B).
+
+    You can customize step 2, e.g., c = (x + y)/2, or some function of (x, y).
+    """
+    import matplotlib.cm
+    import matplotlib.colors
+
+    x_norm = x / max(W - 1, 1)
+    y_norm = y / max(H - 1, 1)
+    # Simple combination:
+    c = (x_norm + y_norm) / 2.0
+
+    cmap = matplotlib.cm.get_cmap(cmap_name)
+    # cmap(c) -> (r,g,b,a) in [0,1]
+    rgba = cmap(c)
+    r, g, b = rgba[0], rgba[1], rgba[2]
+    return (r, g, b)  # in [0,1], RGB order
+
+
+def get_track_colors_by_position(tracks_b, vis_mask_b=None, image_width=None, image_height=None, cmap_name="hsv"):
+    """
+    Given all tracks in one sample (b), compute a (N,3) array of RGB color values
+    in [0,255]. The color is determined by the (x,y) position in the first
+    visible frame for each track.
+
+    Args:
+        tracks_b: Tensor of shape (S, N, 2). (x,y) for each track in each frame.
+        vis_mask_b: (S, N) boolean mask; if None, assume all are visible.
+        image_width, image_height: used for normalizing (x, y).
+        cmap_name: for matplotlib (e.g., 'hsv', 'rainbow', 'jet').
+
+    Returns:
+        track_colors: np.ndarray of shape (N, 3), each row is (R,G,B) in [0,255].
+    """
+    S, N, _ = tracks_b.shape
+    track_colors = np.zeros((N, 3), dtype=np.uint8)
+
+    if vis_mask_b is None:
+        # treat all as visible
+        vis_mask_b = torch.ones(S, N, dtype=torch.bool, device=tracks_b.device)
+
+    for i in range(N):
+        # Find first visible frame for track i
+        visible_frames = torch.where(vis_mask_b[:, i])[0]
+        if len(visible_frames) == 0:
+            # track is never visible; just assign black or something
+            track_colors[i] = (0, 0, 0)
+            continue
+
+        first_s = int(visible_frames[0].item())
+        # use that frame's (x,y)
+        x, y = tracks_b[first_s, i].tolist()
+
+        # map (x,y) -> (R,G,B) in [0,1]
+        r, g, b = color_from_xy(x, y, W=image_width, H=image_height, cmap_name=cmap_name)
+        # scale to [0,255]
+        r, g, b = int(r * 255), int(g * 255), int(b * 255)
+        track_colors[i] = (r, g, b)
+
+    return track_colors
+
+
+def visualize_tracks_on_images(
+    images,
+    tracks,
+    track_vis_mask=None,
+    out_dir="track_visuals_concat_by_xy",
+    image_format="CHW",  # "CHW" or "HWC"
+    normalize_mode="[0,1]",
+    cmap_name="hsv",  # e.g. "hsv", "rainbow", "jet"
+    frames_per_row=4,  # New parameter for grid layout
+    save_grid=True,  # Flag to control whether to save the grid image
+):
+    """
+    Visualizes frames in a grid layout with specified frames per row.
+    Each track's color is determined by its (x,y) position
+    in the first visible frame (or frame 0 if always visible).
+    Finally convert the BGR result to RGB before saving.
+    Also saves each individual frame as a separate PNG file.
+
+    Args:
+        images: torch.Tensor (S, 3, H, W) if CHW or (S, H, W, 3) if HWC.
+        tracks: torch.Tensor (S, N, 2), last dim = (x, y).
+        track_vis_mask: torch.Tensor (S, N) or None.
+        out_dir: folder to save visualizations.
+        image_format: "CHW" or "HWC".
+        normalize_mode: "[0,1]", "[-1,1]", or None for direct raw -> 0..255
+        cmap_name: a matplotlib colormap name for color_from_xy.
+        frames_per_row: number of frames to display in each row of the grid.
+        save_grid: whether to save all frames in one grid image.
+
+    Returns:
+        None (saves images in out_dir).
+    """
+
+    if len(tracks.shape) == 4:
+        tracks = tracks.squeeze(0)
+        images = images.squeeze(0)
+        if track_vis_mask is not None:
+            track_vis_mask = track_vis_mask.squeeze(0)
+
+    import matplotlib
+
+    matplotlib.use("Agg")  # for non-interactive (optional)
+
+    os.makedirs(out_dir, exist_ok=True)
+
+    S = images.shape[0]
+    _, N, _ = tracks.shape  # (S, N, 2)
+
+    # Move to CPU
+    images = images.cpu().clone()
+    tracks = tracks.cpu().clone()
+    if track_vis_mask is not None:
+        track_vis_mask = track_vis_mask.cpu().clone()
+
+    # Infer H, W from images shape
+    if image_format == "CHW":
+        # e.g. images[s].shape = (3, H, W)
+        H, W = images.shape[2], images.shape[3]
+    else:
+        # e.g. images[s].shape = (H, W, 3)
+        H, W = images.shape[1], images.shape[2]
+
+    # Pre-compute the color for each track i based on first visible position
+    track_colors_rgb = get_track_colors_by_position(
+        tracks,  # shape (S, N, 2)
+        vis_mask_b=track_vis_mask if track_vis_mask is not None else None,
+        image_width=W,
+        image_height=H,
+        cmap_name=cmap_name,
+    )
+
+    # We'll accumulate each frame's drawn image in a list
+    frame_images = []
+
+    for s in range(S):
+        # shape => either (3, H, W) or (H, W, 3)
+        img = images[s]
+
+        # Convert to (H, W, 3)
+        if image_format == "CHW":
+            img = img.permute(1, 2, 0)  # (H, W, 3)
+        # else "HWC", do nothing
+
+        img = img.numpy().astype(np.float32)
+
+        # Scale to [0,255] if needed
+        if normalize_mode == "[0,1]":
+            img = np.clip(img, 0, 1) * 255.0
+        elif normalize_mode == "[-1,1]":
+            img = (img + 1.0) * 0.5 * 255.0
+            img = np.clip(img, 0, 255.0)
+        # else no normalization
+
+        # Convert to uint8
+        img = img.astype(np.uint8)
+
+        # For drawing in OpenCV, convert to BGR
+        img_bgr = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)
+
+        # Draw each visible track
+        cur_tracks = tracks[s]  # shape (N, 2)
+        if track_vis_mask is not None:
+            valid_indices = torch.where(track_vis_mask[s])[0]
+        else:
+            valid_indices = range(N)
+
+        cur_tracks_np = cur_tracks.numpy()
+        for i in valid_indices:
+            x, y = cur_tracks_np[i]
+            pt = (int(round(x)), int(round(y)))
+
+            # track_colors_rgb[i] is (R,G,B). For OpenCV circle, we need BGR
+            R, G, B = track_colors_rgb[i]
+            color_bgr = (int(B), int(G), int(R))
+            cv2.circle(img_bgr, pt, radius=3, color=color_bgr, thickness=-1)
+
+        # Convert back to RGB for consistent final saving:
+        img_rgb = cv2.cvtColor(img_bgr, cv2.COLOR_BGR2RGB)
+
+        # Save individual frame
+        frame_path = os.path.join(out_dir, f"frame_{s:04d}.png")
+        # Convert to BGR for OpenCV imwrite
+        frame_bgr = cv2.cvtColor(img_rgb, cv2.COLOR_RGB2BGR)
+        cv2.imwrite(frame_path, frame_bgr)
+
+        frame_images.append(img_rgb)
+
+    # Only create and save the grid image if save_grid is True
+    if save_grid:
+        # Calculate grid dimensions
+        num_rows = (S + frames_per_row - 1) // frames_per_row  # Ceiling division
+
+        # Create a grid of images
+        grid_img = None
+        for row in range(num_rows):
+            start_idx = row * frames_per_row
+            end_idx = min(start_idx + frames_per_row, S)
+
+            # Concatenate this row horizontally
+            row_img = np.concatenate(frame_images[start_idx:end_idx], axis=1)
+
+            # If this row has fewer than frames_per_row images, pad with black
+            if end_idx - start_idx < frames_per_row:
+                padding_width = (frames_per_row - (end_idx - start_idx)) * W
+                padding = np.zeros((H, padding_width, 3), dtype=np.uint8)
+                row_img = np.concatenate([row_img, padding], axis=1)
+
+            # Add this row to the grid
+            if grid_img is None:
+                grid_img = row_img
+            else:
+                grid_img = np.concatenate([grid_img, row_img], axis=0)
+
+        out_path = os.path.join(out_dir, "tracks_grid.png")
+        # Convert back to BGR for OpenCV imwrite
+        grid_img_bgr = cv2.cvtColor(grid_img, cv2.COLOR_RGB2BGR)
+        cv2.imwrite(out_path, grid_img_bgr)
+        print(f"[INFO] Saved color-by-XY track visualization grid -> {out_path}")
+
+    print(f"[INFO] Saved {S} individual frames to {out_dir}/frame_*.png")

outdoor_v48_4gpu_v2/code/05_02-14:21:58/vggt/heads/camera_head.py

@@ -0,0 +1,162 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import math
+import numpy as np
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from vggt.layers import Mlp
+from vggt.layers.block import Block
+from vggt.heads.head_act import activate_pose
+
+
+class CameraHead(nn.Module):
+    """
+    CameraHead predicts camera parameters from token representations using iterative refinement.
+
+    It applies a series of transformer blocks (the "trunk") to dedicated camera tokens.
+    """
+
+    def __init__(
+        self,
+        dim_in: int = 2048,
+        trunk_depth: int = 4,
+        pose_encoding_type: str = "absT_quaR_FoV",
+        num_heads: int = 16,
+        mlp_ratio: int = 4,
+        init_values: float = 0.01,
+        trans_act: str = "linear",
+        quat_act: str = "linear",
+        fl_act: str = "relu",  # Field of view activations: ensures FOV values are positive.
+    ):
+        super().__init__()
+
+        if pose_encoding_type == "absT_quaR_FoV":
+            self.target_dim = 9
+        else:
+            raise ValueError(f"Unsupported camera encoding type: {pose_encoding_type}")
+
+        self.trans_act = trans_act
+        self.quat_act = quat_act
+        self.fl_act = fl_act
+        self.trunk_depth = trunk_depth
+
+        # Build the trunk using a sequence of transformer blocks.
+        self.trunk = nn.Sequential(
+            *[
+                Block(
+                    dim=dim_in,
+                    num_heads=num_heads,
+                    mlp_ratio=mlp_ratio,
+                    init_values=init_values,
+                )
+                for _ in range(trunk_depth)
+            ]
+        )
+
+        # Normalizations for camera token and trunk output.
+        self.token_norm = nn.LayerNorm(dim_in)
+        self.trunk_norm = nn.LayerNorm(dim_in)
+
+        # Learnable empty camera pose token.
+        self.empty_pose_tokens = nn.Parameter(torch.zeros(1, 1, self.target_dim))
+        self.embed_pose = nn.Linear(self.target_dim, dim_in)
+
+        # Module for producing modulation parameters: shift, scale, and a gate.
+        self.poseLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(dim_in, 3 * dim_in, bias=True))
+
+        # Adaptive layer normalization without affine parameters.
+        self.adaln_norm = nn.LayerNorm(dim_in, elementwise_affine=False, eps=1e-6)
+        self.pose_branch = Mlp(
+            in_features=dim_in,
+            hidden_features=dim_in // 2,
+            out_features=self.target_dim,
+            drop=0,
+        )
+
+    def forward(self, aggregated_tokens_list: list, num_iterations: int = 4) -> list:
+        """
+        Forward pass to predict camera parameters.
+
+        Args:
+            aggregated_tokens_list (list): List of token tensors from the network;
+                the last tensor is used for prediction.
+            num_iterations (int, optional): Number of iterative refinement steps. Defaults to 4.
+
+        Returns:
+            list: A list of predicted camera encodings (post-activation) from each iteration.
+        """
+        # Use tokens from the last block for camera prediction.
+        tokens = aggregated_tokens_list[-1]
+
+        # Extract the camera tokens
+        pose_tokens = tokens[:, :, 0]
+        pose_tokens = self.token_norm(pose_tokens)
+
+        pred_pose_enc_list = self.trunk_fn(pose_tokens, num_iterations)
+        return pred_pose_enc_list
+
+    def trunk_fn(self, pose_tokens: torch.Tensor, num_iterations: int) -> list:
+        """
+        Iteratively refine camera pose predictions.
+
+        Args:
+            pose_tokens (torch.Tensor): Normalized camera tokens with shape [B, 1, C].
+            num_iterations (int): Number of refinement iterations.
+
+        Returns:
+            list: List of activated camera encodings from each iteration.
+        """
+        B, S, C = pose_tokens.shape  # S is expected to be 1.
+        pred_pose_enc = None
+        pred_pose_enc_list = []
+
+        for _ in range(num_iterations):
+            # Use a learned empty pose for the first iteration.
+            if pred_pose_enc is None:
+                module_input = self.embed_pose(self.empty_pose_tokens.expand(B, S, -1))
+            else:
+                # Detach the previous prediction to avoid backprop through time.
+                pred_pose_enc = pred_pose_enc.detach()
+                module_input = self.embed_pose(pred_pose_enc)
+
+            # Generate modulation parameters and split them into shift, scale, and gate components.
+            shift_msa, scale_msa, gate_msa = self.poseLN_modulation(module_input).chunk(3, dim=-1)
+
+            # Adaptive layer normalization and modulation.
+            pose_tokens_modulated = gate_msa * modulate(self.adaln_norm(pose_tokens), shift_msa, scale_msa)
+            pose_tokens_modulated = pose_tokens_modulated + pose_tokens
+
+            pose_tokens_modulated = self.trunk(pose_tokens_modulated)
+            # Compute the delta update for the pose encoding.
+            pred_pose_enc_delta = self.pose_branch(self.trunk_norm(pose_tokens_modulated))
+
+            if pred_pose_enc is None:
+                pred_pose_enc = pred_pose_enc_delta
+            else:
+                pred_pose_enc = pred_pose_enc + pred_pose_enc_delta
+
+            # Apply final activation functions for translation, quaternion, and field-of-view.
+            activated_pose = activate_pose(
+                pred_pose_enc,
+                trans_act=self.trans_act,
+                quat_act=self.quat_act,
+                fl_act=self.fl_act,
+            )
+            pred_pose_enc_list.append(activated_pose)
+
+        return pred_pose_enc_list
+
+
+def modulate(x: torch.Tensor, shift: torch.Tensor, scale: torch.Tensor) -> torch.Tensor:
+    """
+    Modulate the input tensor using scaling and shifting parameters.
+    """
+    # modified from https://github.com/facebookresearch/DiT/blob/796c29e532f47bba17c5b9c5eb39b9354b8b7c64/models.py#L19
+    return x * (1 + scale) + shift

outdoor_v48_4gpu_v2/code/05_02-14:21:58/vggt/heads/dpt_head.py

@@ -0,0 +1,497 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+
+# Inspired by https://github.com/DepthAnything/Depth-Anything-V2
+
+
+import os
+from typing import List, Dict, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from .head_act import activate_head
+from .utils import create_uv_grid, position_grid_to_embed
+
+
+class DPTHead(nn.Module):
+    """
+    DPT  Head for dense prediction tasks.
+
+    This implementation follows the architecture described in "Vision Transformers for Dense Prediction"
+    (https://arxiv.org/abs/2103.13413). The DPT head processes features from a vision transformer
+    backbone and produces dense predictions by fusing multi-scale features.
+
+    Args:
+        dim_in (int): Input dimension (channels).
+        patch_size (int, optional): Patch size. Default is 14.
+        output_dim (int, optional): Number of output channels. Default is 4.
+        activation (str, optional): Activation type. Default is "inv_log".
+        conf_activation (str, optional): Confidence activation type. Default is "expp1".
+        features (int, optional): Feature channels for intermediate representations. Default is 256.
+        out_channels (List[int], optional): Output channels for each intermediate layer.
+        intermediate_layer_idx (List[int], optional): Indices of layers from aggregated tokens used for DPT.
+        pos_embed (bool, optional): Whether to use positional embedding. Default is True.
+        feature_only (bool, optional): If True, return features only without the last several layers and activation head. Default is False.
+        down_ratio (int, optional): Downscaling factor for the output resolution. Default is 1.
+    """
+
+    def __init__(
+        self,
+        dim_in: int,
+        patch_size: int = 14,
+        output_dim: int = 4,
+        activation: str = "inv_log",
+        conf_activation: str = "expp1",
+        features: int = 256,
+        out_channels: List[int] = [256, 512, 1024, 1024],
+        intermediate_layer_idx: List[int] = [4, 11, 17, 23],
+        pos_embed: bool = True,
+        feature_only: bool = False,
+        down_ratio: int = 1,
+    ) -> None:
+        super(DPTHead, self).__init__()
+        self.patch_size = patch_size
+        self.activation = activation
+        self.conf_activation = conf_activation
+        self.pos_embed = pos_embed
+        self.feature_only = feature_only
+        self.down_ratio = down_ratio
+        self.intermediate_layer_idx = intermediate_layer_idx
+
+        self.norm = nn.LayerNorm(dim_in)
+
+        # Projection layers for each output channel from tokens.
+        self.projects = nn.ModuleList(
+            [
+                nn.Conv2d(
+                    in_channels=dim_in,
+                    out_channels=oc,
+                    kernel_size=1,
+                    stride=1,
+                    padding=0,
+                )
+                for oc in out_channels
+            ]
+        )
+
+        # Resize layers for upsampling feature maps.
+        self.resize_layers = nn.ModuleList(
+            [
+                nn.ConvTranspose2d(
+                    in_channels=out_channels[0], out_channels=out_channels[0], kernel_size=4, stride=4, padding=0
+                ),
+                nn.ConvTranspose2d(
+                    in_channels=out_channels[1], out_channels=out_channels[1], kernel_size=2, stride=2, padding=0
+                ),
+                nn.Identity(),
+                nn.Conv2d(
+                    in_channels=out_channels[3], out_channels=out_channels[3], kernel_size=3, stride=2, padding=1
+                ),
+            ]
+        )
+
+        self.scratch = _make_scratch(
+            out_channels,
+            features,
+            expand=False,
+        )
+
+        # Attach additional modules to scratch.
+        self.scratch.stem_transpose = None
+        self.scratch.refinenet1 = _make_fusion_block(features)
+        self.scratch.refinenet2 = _make_fusion_block(features)
+        self.scratch.refinenet3 = _make_fusion_block(features)
+        self.scratch.refinenet4 = _make_fusion_block(features, has_residual=False)
+
+        head_features_1 = features
+        head_features_2 = 32
+
+        if feature_only:
+            self.scratch.output_conv1 = nn.Conv2d(head_features_1, head_features_1, kernel_size=3, stride=1, padding=1)
+        else:
+            self.scratch.output_conv1 = nn.Conv2d(
+                head_features_1, head_features_1 // 2, kernel_size=3, stride=1, padding=1
+            )
+            conv2_in_channels = head_features_1 // 2
+
+            self.scratch.output_conv2 = nn.Sequential(
+                nn.Conv2d(conv2_in_channels, head_features_2, kernel_size=3, stride=1, padding=1),
+                nn.ReLU(inplace=True),
+                nn.Conv2d(head_features_2, output_dim, kernel_size=1, stride=1, padding=0),
+            )
+
+    def forward(
+        self,
+        aggregated_tokens_list: List[torch.Tensor],
+        images: torch.Tensor,
+        patch_start_idx: int,
+        frames_chunk_size: int = 8,
+    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        """
+        Forward pass through the DPT head, supports processing by chunking frames.
+        Args:
+            aggregated_tokens_list (List[Tensor]): List of token tensors from different transformer layers.
+            images (Tensor): Input images with shape [B, S, 3, H, W], in range [0, 1].
+            patch_start_idx (int): Starting index for patch tokens in the token sequence.
+                Used to separate patch tokens from other tokens (e.g., camera or register tokens).
+            frames_chunk_size (int, optional): Number of frames to process in each chunk.
+                If None or larger than S, all frames are processed at once. Default: 8.
+
+        Returns:
+            Tensor or Tuple[Tensor, Tensor]:
+                - If feature_only=True: Feature maps with shape [B, S, C, H, W]
+                - Otherwise: Tuple of (predictions, confidence) both with shape [B, S, 1, H, W]
+        """
+        B, S, _, H, W = images.shape
+
+        # If frames_chunk_size is not specified or greater than S, process all frames at once
+        if frames_chunk_size is None or frames_chunk_size >= S:
+            return self._forward_impl(aggregated_tokens_list, images, patch_start_idx)
+
+        # Otherwise, process frames in chunks to manage memory usage
+        assert frames_chunk_size > 0
+
+        # Process frames in batches
+        all_preds = []
+        all_conf = []
+
+        for frames_start_idx in range(0, S, frames_chunk_size):
+            frames_end_idx = min(frames_start_idx + frames_chunk_size, S)
+
+            # Process batch of frames
+            if self.feature_only:
+                chunk_output = self._forward_impl(
+                    aggregated_tokens_list, images, patch_start_idx, frames_start_idx, frames_end_idx
+                )
+                all_preds.append(chunk_output)
+            else:
+                chunk_preds, chunk_conf = self._forward_impl(
+                    aggregated_tokens_list, images, patch_start_idx, frames_start_idx, frames_end_idx
+                )
+                all_preds.append(chunk_preds)
+                all_conf.append(chunk_conf)
+
+        # Concatenate results along the sequence dimension
+        if self.feature_only:
+            return torch.cat(all_preds, dim=1)
+        else:
+            return torch.cat(all_preds, dim=1), torch.cat(all_conf, dim=1)
+
+    def _forward_impl(
+        self,
+        aggregated_tokens_list: List[torch.Tensor],
+        images: torch.Tensor,
+        patch_start_idx: int,
+        frames_start_idx: int = None,
+        frames_end_idx: int = None,
+    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        """
+        Implementation of the forward pass through the DPT head.
+
+        This method processes a specific chunk of frames from the sequence.
+
+        Args:
+            aggregated_tokens_list (List[Tensor]): List of token tensors from different transformer layers.
+            images (Tensor): Input images with shape [B, S, 3, H, W].
+            patch_start_idx (int): Starting index for patch tokens.
+            frames_start_idx (int, optional): Starting index for frames to process.
+            frames_end_idx (int, optional): Ending index for frames to process.
+
+        Returns:
+            Tensor or Tuple[Tensor, Tensor]: Feature maps or (predictions, confidence).
+        """
+        if frames_start_idx is not None and frames_end_idx is not None:
+            images = images[:, frames_start_idx:frames_end_idx].contiguous()
+
+        B, S, _, H, W = images.shape
+
+        patch_h, patch_w = H // self.patch_size, W // self.patch_size
+
+        out = []
+        dpt_idx = 0
+
+        for layer_idx in self.intermediate_layer_idx:
+            x = aggregated_tokens_list[layer_idx][:, :, patch_start_idx:]
+
+            # Select frames if processing a chunk
+            if frames_start_idx is not None and frames_end_idx is not None:
+                x = x[:, frames_start_idx:frames_end_idx]
+
+            x = x.reshape(B * S, -1, x.shape[-1])
+
+            x = self.norm(x)
+
+            x = x.permute(0, 2, 1).reshape((x.shape[0], x.shape[-1], patch_h, patch_w))
+
+            x = self.projects[dpt_idx](x)
+            if self.pos_embed:
+                x = self._apply_pos_embed(x, W, H)
+            x = self.resize_layers[dpt_idx](x)
+
+            out.append(x)
+            dpt_idx += 1
+
+        # Fuse features from multiple layers.
+        out = self.scratch_forward(out)
+        # Interpolate fused output to match target image resolution.
+        out = custom_interpolate(
+            out,
+            (int(patch_h * self.patch_size / self.down_ratio), int(patch_w * self.patch_size / self.down_ratio)),
+            mode="bilinear",
+            align_corners=True,
+        )
+
+        if self.pos_embed:
+            out = self._apply_pos_embed(out, W, H)
+
+        if self.feature_only:
+            return out.reshape(B, S, *out.shape[1:])
+
+        out = self.scratch.output_conv2(out)
+        preds, conf = activate_head(out, activation=self.activation, conf_activation=self.conf_activation)
+
+        preds = preds.reshape(B, S, *preds.shape[1:])
+        conf = conf.reshape(B, S, *conf.shape[1:])
+        return preds, conf
+
+    def _apply_pos_embed(self, x: torch.Tensor, W: int, H: int, ratio: float = 0.1) -> torch.Tensor:
+        """
+        Apply positional embedding to tensor x.
+        """
+        patch_w = x.shape[-1]
+        patch_h = x.shape[-2]
+        pos_embed = create_uv_grid(patch_w, patch_h, aspect_ratio=W / H, dtype=x.dtype, device=x.device)
+        pos_embed = position_grid_to_embed(pos_embed, x.shape[1])
+        pos_embed = pos_embed * ratio
+        pos_embed = pos_embed.permute(2, 0, 1)[None].expand(x.shape[0], -1, -1, -1)
+        return x + pos_embed
+
+    def scratch_forward(self, features: List[torch.Tensor]) -> torch.Tensor:
+        """
+        Forward pass through the fusion blocks.
+
+        Args:
+            features (List[Tensor]): List of feature maps from different layers.
+
+        Returns:
+            Tensor: Fused feature map.
+        """
+        layer_1, layer_2, layer_3, layer_4 = features
+
+        layer_1_rn = self.scratch.layer1_rn(layer_1)
+        layer_2_rn = self.scratch.layer2_rn(layer_2)
+        layer_3_rn = self.scratch.layer3_rn(layer_3)
+        layer_4_rn = self.scratch.layer4_rn(layer_4)
+
+        out = self.scratch.refinenet4(layer_4_rn, size=layer_3_rn.shape[2:])
+        del layer_4_rn, layer_4
+
+        out = self.scratch.refinenet3(out, layer_3_rn, size=layer_2_rn.shape[2:])
+        del layer_3_rn, layer_3
+
+        out = self.scratch.refinenet2(out, layer_2_rn, size=layer_1_rn.shape[2:])
+        del layer_2_rn, layer_2
+
+        out = self.scratch.refinenet1(out, layer_1_rn)
+        del layer_1_rn, layer_1
+
+        out = self.scratch.output_conv1(out)
+        return out
+
+
+################################################################################
+# Modules
+################################################################################
+
+
+def _make_fusion_block(features: int, size: int = None, has_residual: bool = True, groups: int = 1) -> nn.Module:
+    return FeatureFusionBlock(
+        features,
+        nn.ReLU(inplace=True),
+        deconv=False,
+        bn=False,
+        expand=False,
+        align_corners=True,
+        size=size,
+        has_residual=has_residual,
+        groups=groups,
+    )
+
+
+def _make_scratch(in_shape: List[int], out_shape: int, groups: int = 1, expand: bool = False) -> nn.Module:
+    scratch = nn.Module()
+    out_shape1 = out_shape
+    out_shape2 = out_shape
+    out_shape3 = out_shape
+    if len(in_shape) >= 4:
+        out_shape4 = out_shape
+
+    if expand:
+        out_shape1 = out_shape
+        out_shape2 = out_shape * 2
+        out_shape3 = out_shape * 4
+        if len(in_shape) >= 4:
+            out_shape4 = out_shape * 8
+
+    scratch.layer1_rn = nn.Conv2d(
+        in_shape[0], out_shape1, kernel_size=3, stride=1, padding=1, bias=False, groups=groups
+    )
+    scratch.layer2_rn = nn.Conv2d(
+        in_shape[1], out_shape2, kernel_size=3, stride=1, padding=1, bias=False, groups=groups
+    )
+    scratch.layer3_rn = nn.Conv2d(
+        in_shape[2], out_shape3, kernel_size=3, stride=1, padding=1, bias=False, groups=groups
+    )
+    if len(in_shape) >= 4:
+        scratch.layer4_rn = nn.Conv2d(
+            in_shape[3], out_shape4, kernel_size=3, stride=1, padding=1, bias=False, groups=groups
+        )
+    return scratch
+
+
+class ResidualConvUnit(nn.Module):
+    """Residual convolution module."""
+
+    def __init__(self, features, activation, bn, groups=1):
+        """Init.
+
+        Args:
+            features (int): number of features
+        """
+        super().__init__()
+
+        self.bn = bn
+        self.groups = groups
+        self.conv1 = nn.Conv2d(features, features, kernel_size=3, stride=1, padding=1, bias=True, groups=self.groups)
+        self.conv2 = nn.Conv2d(features, features, kernel_size=3, stride=1, padding=1, bias=True, groups=self.groups)
+
+        self.norm1 = None
+        self.norm2 = None
+
+        self.activation = activation
+        self.skip_add = nn.quantized.FloatFunctional()
+
+    def forward(self, x):
+        """Forward pass.
+
+        Args:
+            x (tensor): input
+
+        Returns:
+            tensor: output
+        """
+
+        out = self.activation(x)
+        out = self.conv1(out)
+        if self.norm1 is not None:
+            out = self.norm1(out)
+
+        out = self.activation(out)
+        out = self.conv2(out)
+        if self.norm2 is not None:
+            out = self.norm2(out)
+
+        return self.skip_add.add(out, x)
+
+
+class FeatureFusionBlock(nn.Module):
+    """Feature fusion block."""
+
+    def __init__(
+        self,
+        features,
+        activation,
+        deconv=False,
+        bn=False,
+        expand=False,
+        align_corners=True,
+        size=None,
+        has_residual=True,
+        groups=1,
+    ):
+        """Init.
+
+        Args:
+            features (int): number of features
+        """
+        super(FeatureFusionBlock, self).__init__()
+
+        self.deconv = deconv
+        self.align_corners = align_corners
+        self.groups = groups
+        self.expand = expand
+        out_features = features
+        if self.expand == True:
+            out_features = features // 2
+
+        self.out_conv = nn.Conv2d(
+            features, out_features, kernel_size=1, stride=1, padding=0, bias=True, groups=self.groups
+        )
+
+        if has_residual:
+            self.resConfUnit1 = ResidualConvUnit(features, activation, bn, groups=self.groups)
+
+        self.has_residual = has_residual
+        self.resConfUnit2 = ResidualConvUnit(features, activation, bn, groups=self.groups)
+
+        self.skip_add = nn.quantized.FloatFunctional()
+        self.size = size
+
+    def forward(self, *xs, size=None):
+        """Forward pass.
+
+        Returns:
+            tensor: output
+        """
+        output = xs[0]
+
+        if self.has_residual:
+            res = self.resConfUnit1(xs[1])
+            output = self.skip_add.add(output, res)
+
+        output = self.resConfUnit2(output)
+
+        if (size is None) and (self.size is None):
+            modifier = {"scale_factor": 2}
+        elif size is None:
+            modifier = {"size": self.size}
+        else:
+            modifier = {"size": size}
+
+        output = custom_interpolate(output, **modifier, mode="bilinear", align_corners=self.align_corners)
+        output = self.out_conv(output)
+
+        return output
+
+
+def custom_interpolate(
+    x: torch.Tensor,
+    size: Tuple[int, int] = None,
+    scale_factor: float = None,
+    mode: str = "bilinear",
+    align_corners: bool = True,
+) -> torch.Tensor:
+    """
+    Custom interpolate to avoid INT_MAX issues in nn.functional.interpolate.
+    """
+    if size is None:
+        size = (int(x.shape[-2] * scale_factor), int(x.shape[-1] * scale_factor))
+
+    INT_MAX = 1610612736
+
+    input_elements = size[0] * size[1] * x.shape[0] * x.shape[1]
+
+    if input_elements > INT_MAX:
+        chunks = torch.chunk(x, chunks=(input_elements // INT_MAX) + 1, dim=0)
+        interpolated_chunks = [
+            nn.functional.interpolate(chunk, size=size, mode=mode, align_corners=align_corners) for chunk in chunks
+        ]
+        x = torch.cat(interpolated_chunks, dim=0)
+        return x.contiguous()
+    else:
+        return nn.functional.interpolate(x, size=size, mode=mode, align_corners=align_corners)

outdoor_v48_4gpu_v2/code/05_02-14:21:58/vggt/heads/head_act.py

@@ -0,0 +1,125 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+
+import torch
+import torch.nn.functional as F
+
+
+def activate_pose(pred_pose_enc, trans_act="linear", quat_act="linear", fl_act="linear"):
+    """
+    Activate pose parameters with specified activation functions.
+
+    Args:
+        pred_pose_enc: Tensor containing encoded pose parameters [translation, quaternion, focal length]
+        trans_act: Activation type for translation component
+        quat_act: Activation type for quaternion component
+        fl_act: Activation type for focal length component
+
+    Returns:
+        Activated pose parameters tensor
+    """
+    T = pred_pose_enc[..., :3]
+    quat = pred_pose_enc[..., 3:7]
+    fl = pred_pose_enc[..., 7:]  # or fov
+
+    T = base_pose_act(T, trans_act)
+    quat = base_pose_act(quat, quat_act)
+    fl = base_pose_act(fl, fl_act)  # or fov
+
+    pred_pose_enc = torch.cat([T, quat, fl], dim=-1)
+
+    return pred_pose_enc
+
+
+def base_pose_act(pose_enc, act_type="linear"):
+    """
+    Apply basic activation function to pose parameters.
+
+    Args:
+        pose_enc: Tensor containing encoded pose parameters
+        act_type: Activation type ("linear", "inv_log", "exp", "relu")
+
+    Returns:
+        Activated pose parameters
+    """
+    if act_type == "linear":
+        return pose_enc
+    elif act_type == "inv_log":
+        return inverse_log_transform(pose_enc)
+    elif act_type == "exp":
+        return torch.exp(pose_enc)
+    elif act_type == "relu":
+        return F.relu(pose_enc)
+    else:
+        raise ValueError(f"Unknown act_type: {act_type}")
+
+
+def activate_head(out, activation="norm_exp", conf_activation="expp1"):
+    """
+    Process network output to extract 3D points and confidence values.
+
+    Args:
+        out: Network output tensor (B, C, H, W)
+        activation: Activation type for 3D points
+        conf_activation: Activation type for confidence values
+
+    Returns:
+        Tuple of (3D points tensor, confidence tensor)
+    """
+    # Move channels from last dim to the 4th dimension => (B, H, W, C)
+    fmap = out.permute(0, 2, 3, 1)  # B,H,W,C expected
+
+    # Split into xyz (first C-1 channels) and confidence (last channel)
+    xyz = fmap[:, :, :, :-1]
+    conf = fmap[:, :, :, -1]
+
+    if activation == "norm_exp":
+        d = xyz.norm(dim=-1, keepdim=True).clamp(min=1e-8)
+        xyz_normed = xyz / d
+        pts3d = xyz_normed * torch.expm1(d)
+    elif activation == "norm":
+        pts3d = xyz / xyz.norm(dim=-1, keepdim=True)
+    elif activation == "exp":
+        pts3d = torch.exp(xyz)
+    elif activation == "relu":
+        pts3d = F.relu(xyz)
+    elif activation == "inv_log":
+        pts3d = inverse_log_transform(xyz)
+    elif activation == "xy_inv_log":
+        xy, z = xyz.split([2, 1], dim=-1)
+        z = inverse_log_transform(z)
+        pts3d = torch.cat([xy * z, z], dim=-1)
+    elif activation == "sigmoid":
+        pts3d = torch.sigmoid(xyz)
+    elif activation == "linear":
+        pts3d = xyz
+    else:
+        raise ValueError(f"Unknown activation: {activation}")
+
+    if conf_activation == "expp1":
+        conf_out = 1 + conf.exp()
+    elif conf_activation == "expp0":
+        conf_out = conf.exp()
+    elif conf_activation == "sigmoid":
+        conf_out = torch.sigmoid(conf)
+    else:
+        raise ValueError(f"Unknown conf_activation: {conf_activation}")
+
+    return pts3d, conf_out
+
+
+def inverse_log_transform(y):
+    """
+    Apply inverse log transform: sign(y) * (exp(|y|) - 1)
+
+    Args:
+        y: Input tensor
+
+    Returns:
+        Transformed tensor
+    """
+    return torch.sign(y) * (torch.expm1(torch.abs(y)))

outdoor_v48_4gpu_v2/code/05_02-14:21:58/vggt/heads/track_head.py

@@ -0,0 +1,108 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch.nn as nn
+from .dpt_head import DPTHead
+from .track_modules.base_track_predictor import BaseTrackerPredictor
+
+
+class TrackHead(nn.Module):
+    """
+    Track head that uses DPT head to process tokens and BaseTrackerPredictor for tracking.
+    The tracking is performed iteratively, refining predictions over multiple iterations.
+    """
+
+    def __init__(
+        self,
+        dim_in,
+        patch_size=14,
+        features=128,
+        iters=4,
+        predict_conf=True,
+        stride=2,
+        corr_levels=7,
+        corr_radius=4,
+        hidden_size=384,
+    ):
+        """
+        Initialize the TrackHead module.
+
+        Args:
+            dim_in (int): Input dimension of tokens from the backbone.
+            patch_size (int): Size of image patches used in the vision transformer.
+            features (int): Number of feature channels in the feature extractor output.
+            iters (int): Number of refinement iterations for tracking predictions.
+            predict_conf (bool): Whether to predict confidence scores for tracked points.
+            stride (int): Stride value for the tracker predictor.
+            corr_levels (int): Number of correlation pyramid levels
+            corr_radius (int): Radius for correlation computation, controlling the search area.
+            hidden_size (int): Size of hidden layers in the tracker network.
+        """
+        super().__init__()
+
+        self.patch_size = patch_size
+
+        # Feature extractor based on DPT architecture
+        # Processes tokens into feature maps for tracking
+        self.feature_extractor = DPTHead(
+            dim_in=dim_in,
+            patch_size=patch_size,
+            features=features,
+            feature_only=True,  # Only output features, no activation
+            down_ratio=2,  # Reduces spatial dimensions by factor of 2
+            pos_embed=False,
+        )
+
+        # Tracker module that predicts point trajectories
+        # Takes feature maps and predicts coordinates and visibility
+        self.tracker = BaseTrackerPredictor(
+            latent_dim=features,  # Match the output_dim of feature extractor
+            predict_conf=predict_conf,
+            stride=stride,
+            corr_levels=corr_levels,
+            corr_radius=corr_radius,
+            hidden_size=hidden_size,
+        )
+
+        self.iters = iters
+
+    def forward(self, aggregated_tokens_list, images, patch_start_idx, query_points=None, iters=None):
+        """
+        Forward pass of the TrackHead.
+
+        Args:
+            aggregated_tokens_list (list): List of aggregated tokens from the backbone.
+            images (torch.Tensor): Input images of shape (B, S, C, H, W) where:
+                                   B = batch size, S = sequence length.
+            patch_start_idx (int): Starting index for patch tokens.
+            query_points (torch.Tensor, optional): Initial query points to track.
+                                                  If None, points are initialized by the tracker.
+            iters (int, optional): Number of refinement iterations. If None, uses self.iters.
+
+        Returns:
+            tuple:
+                - coord_preds (torch.Tensor): Predicted coordinates for tracked points.
+                - vis_scores (torch.Tensor): Visibility scores for tracked points.
+                - conf_scores (torch.Tensor): Confidence scores for tracked points (if predict_conf=True).
+        """
+        B, S, _, H, W = images.shape
+
+        # Extract features from tokens
+        # feature_maps has shape (B, S, C, H//2, W//2) due to down_ratio=2
+        feature_maps = self.feature_extractor(aggregated_tokens_list, images, patch_start_idx)
+
+        # Use default iterations if not specified
+        if iters is None:
+            iters = self.iters
+
+        # Perform tracking using the extracted features
+        coord_preds, vis_scores, conf_scores = self.tracker(
+            query_points=query_points,
+            fmaps=feature_maps,
+            iters=iters,
+        )
+
+        return coord_preds, vis_scores, conf_scores

outdoor_v48_4gpu_v2/code/05_02-14:21:58/vggt/heads/track_modules/__init__.py

@@ -0,0 +1,5 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.

outdoor_v48_4gpu_v2/code/05_02-14:21:58/vggt/heads/track_modules/base_track_predictor.py

@@ -0,0 +1,209 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+import torch.nn as nn
+from einops import rearrange, repeat
+
+
+from .blocks import EfficientUpdateFormer, CorrBlock
+from .utils import sample_features4d, get_2d_embedding, get_2d_sincos_pos_embed
+from .modules import Mlp
+
+
+class BaseTrackerPredictor(nn.Module):
+    def __init__(
+        self,
+        stride=1,
+        corr_levels=5,
+        corr_radius=4,
+        latent_dim=128,
+        hidden_size=384,
+        use_spaceatt=True,
+        depth=6,
+        max_scale=518,
+        predict_conf=True,
+    ):
+        super(BaseTrackerPredictor, self).__init__()
+        """
+        The base template to create a track predictor
+        
+        Modified from https://github.com/facebookresearch/co-tracker/
+        and https://github.com/facebookresearch/vggsfm
+        """
+
+        self.stride = stride
+        self.latent_dim = latent_dim
+        self.corr_levels = corr_levels
+        self.corr_radius = corr_radius
+        self.hidden_size = hidden_size
+        self.max_scale = max_scale
+        self.predict_conf = predict_conf
+
+        self.flows_emb_dim = latent_dim // 2
+
+        self.corr_mlp = Mlp(
+            in_features=self.corr_levels * (self.corr_radius * 2 + 1) ** 2,
+            hidden_features=self.hidden_size,
+            out_features=self.latent_dim,
+        )
+
+        self.transformer_dim = self.latent_dim + self.latent_dim + self.latent_dim + 4
+
+        self.query_ref_token = nn.Parameter(torch.randn(1, 2, self.transformer_dim))
+
+        space_depth = depth if use_spaceatt else 0
+        time_depth = depth
+
+        self.updateformer = EfficientUpdateFormer(
+            space_depth=space_depth,
+            time_depth=time_depth,
+            input_dim=self.transformer_dim,
+            hidden_size=self.hidden_size,
+            output_dim=self.latent_dim + 2,
+            mlp_ratio=4.0,
+            add_space_attn=use_spaceatt,
+        )
+
+        self.fmap_norm = nn.LayerNorm(self.latent_dim)
+        self.ffeat_norm = nn.GroupNorm(1, self.latent_dim)
+
+        # A linear layer to update track feats at each iteration
+        self.ffeat_updater = nn.Sequential(nn.Linear(self.latent_dim, self.latent_dim), nn.GELU())
+
+        self.vis_predictor = nn.Sequential(nn.Linear(self.latent_dim, 1))
+
+        if predict_conf:
+            self.conf_predictor = nn.Sequential(nn.Linear(self.latent_dim, 1))
+
+    def forward(self, query_points, fmaps=None, iters=6, return_feat=False, down_ratio=1, apply_sigmoid=True):
+        """
+        query_points: B x N x 2, the number of batches, tracks, and xy
+        fmaps: B x S x C x HH x WW, the number of batches, frames, and feature dimension.
+                note HH and WW is the size of feature maps instead of original images
+        """
+        B, N, D = query_points.shape
+        B, S, C, HH, WW = fmaps.shape
+
+        assert D == 2, "Input points must be 2D coordinates"
+
+        # apply a layernorm to fmaps here
+        fmaps = self.fmap_norm(fmaps.permute(0, 1, 3, 4, 2))
+        fmaps = fmaps.permute(0, 1, 4, 2, 3)
+
+        # Scale the input query_points because we may downsample the images
+        # by down_ratio or self.stride
+        # e.g., if a 3x1024x1024 image is processed to a 128x256x256 feature map
+        # its query_points should be query_points/4
+        if down_ratio > 1:
+            query_points = query_points / float(down_ratio)
+
+        query_points = query_points / float(self.stride)
+
+        # Init with coords as the query points
+        # It means the search will start from the position of query points at the reference frames
+        coords = query_points.clone().reshape(B, 1, N, 2).repeat(1, S, 1, 1)
+
+        # Sample/extract the features of the query points in the query frame
+        query_track_feat = sample_features4d(fmaps[:, 0], coords[:, 0])
+
+        # init track feats by query feats
+        track_feats = query_track_feat.unsqueeze(1).repeat(1, S, 1, 1)  # B, S, N, C
+        # back up the init coords
+        coords_backup = coords.clone()
+
+        fcorr_fn = CorrBlock(fmaps, num_levels=self.corr_levels, radius=self.corr_radius)
+
+        coord_preds = []
+
+        # Iterative Refinement
+        for _ in range(iters):
+            # Detach the gradients from the last iteration
+            # (in my experience, not very important for performance)
+            coords = coords.detach()
+
+            fcorrs = fcorr_fn.corr_sample(track_feats, coords)
+
+            corr_dim = fcorrs.shape[3]
+            fcorrs_ = fcorrs.permute(0, 2, 1, 3).reshape(B * N, S, corr_dim)
+            fcorrs_ = self.corr_mlp(fcorrs_)
+
+            # Movement of current coords relative to query points
+            flows = (coords - coords[:, 0:1]).permute(0, 2, 1, 3).reshape(B * N, S, 2)
+
+            flows_emb = get_2d_embedding(flows, self.flows_emb_dim, cat_coords=False)
+
+            # (In my trials, it is also okay to just add the flows_emb instead of concat)
+            flows_emb = torch.cat([flows_emb, flows / self.max_scale, flows / self.max_scale], dim=-1)
+
+            track_feats_ = track_feats.permute(0, 2, 1, 3).reshape(B * N, S, self.latent_dim)
+
+            # Concatenate them as the input for the transformers
+            transformer_input = torch.cat([flows_emb, fcorrs_, track_feats_], dim=2)
+
+            # 2D positional embed
+            # TODO: this can be much simplified
+            pos_embed = get_2d_sincos_pos_embed(self.transformer_dim, grid_size=(HH, WW)).to(query_points.device)
+            sampled_pos_emb = sample_features4d(pos_embed.expand(B, -1, -1, -1), coords[:, 0])
+
+            sampled_pos_emb = rearrange(sampled_pos_emb, "b n c -> (b n) c").unsqueeze(1)
+
+            x = transformer_input + sampled_pos_emb
+
+            # Add the query ref token to the track feats
+            query_ref_token = torch.cat(
+                [self.query_ref_token[:, 0:1], self.query_ref_token[:, 1:2].expand(-1, S - 1, -1)], dim=1
+            )
+            x = x + query_ref_token.to(x.device).to(x.dtype)
+
+            # B, N, S, C
+            x = rearrange(x, "(b n) s d -> b n s d", b=B)
+
+            # Compute the delta coordinates and delta track features
+            delta, _ = self.updateformer(x)
+
+            # BN, S, C
+            delta = rearrange(delta, " b n s d -> (b n) s d", b=B)
+            delta_coords_ = delta[:, :, :2]
+            delta_feats_ = delta[:, :, 2:]
+
+            track_feats_ = track_feats_.reshape(B * N * S, self.latent_dim)
+            delta_feats_ = delta_feats_.reshape(B * N * S, self.latent_dim)
+
+            # Update the track features
+            track_feats_ = self.ffeat_updater(self.ffeat_norm(delta_feats_)) + track_feats_
+
+            track_feats = track_feats_.reshape(B, N, S, self.latent_dim).permute(0, 2, 1, 3)  # BxSxNxC
+
+            # B x S x N x 2
+            coords = coords + delta_coords_.reshape(B, N, S, 2).permute(0, 2, 1, 3)
+
+            # Force coord0 as query
+            # because we assume the query points should not be changed
+            coords[:, 0] = coords_backup[:, 0]
+
+            # The predicted tracks are in the original image scale
+            if down_ratio > 1:
+                coord_preds.append(coords * self.stride * down_ratio)
+            else:
+                coord_preds.append(coords * self.stride)
+
+        # B, S, N
+        vis_e = self.vis_predictor(track_feats.reshape(B * S * N, self.latent_dim)).reshape(B, S, N)
+        if apply_sigmoid:
+            vis_e = torch.sigmoid(vis_e)
+
+        if self.predict_conf:
+            conf_e = self.conf_predictor(track_feats.reshape(B * S * N, self.latent_dim)).reshape(B, S, N)
+            if apply_sigmoid:
+                conf_e = torch.sigmoid(conf_e)
+        else:
+            conf_e = None
+
+        if return_feat:
+            return coord_preds, vis_e, track_feats, query_track_feat, conf_e
+        else:
+            return coord_preds, vis_e, conf_e

outdoor_v48_4gpu_v2/code/05_02-14:21:58/vggt/heads/track_modules/blocks.py

@@ -0,0 +1,246 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+
+# Modified from https://github.com/facebookresearch/co-tracker/
+
+import math
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from .utils import bilinear_sampler
+from .modules import Mlp, AttnBlock, CrossAttnBlock, ResidualBlock
+
+
+class EfficientUpdateFormer(nn.Module):
+    """
+    Transformer model that updates track estimates.
+    """
+
+    def __init__(
+        self,
+        space_depth=6,
+        time_depth=6,
+        input_dim=320,
+        hidden_size=384,
+        num_heads=8,
+        output_dim=130,
+        mlp_ratio=4.0,
+        add_space_attn=True,
+        num_virtual_tracks=64,
+    ):
+        super().__init__()
+
+        self.out_channels = 2
+        self.num_heads = num_heads
+        self.hidden_size = hidden_size
+        self.add_space_attn = add_space_attn
+
+        # Add input LayerNorm before linear projection
+        self.input_norm = nn.LayerNorm(input_dim)
+        self.input_transform = torch.nn.Linear(input_dim, hidden_size, bias=True)
+
+        # Add output LayerNorm before final projection
+        self.output_norm = nn.LayerNorm(hidden_size)
+        self.flow_head = torch.nn.Linear(hidden_size, output_dim, bias=True)
+        self.num_virtual_tracks = num_virtual_tracks
+
+        if self.add_space_attn:
+            self.virual_tracks = nn.Parameter(torch.randn(1, num_virtual_tracks, 1, hidden_size))
+        else:
+            self.virual_tracks = None
+
+        self.time_blocks = nn.ModuleList(
+            [
+                AttnBlock(
+                    hidden_size,
+                    num_heads,
+                    mlp_ratio=mlp_ratio,
+                    attn_class=nn.MultiheadAttention,
+                )
+                for _ in range(time_depth)
+            ]
+        )
+
+        if add_space_attn:
+            self.space_virtual_blocks = nn.ModuleList(
+                [
+                    AttnBlock(
+                        hidden_size,
+                        num_heads,
+                        mlp_ratio=mlp_ratio,
+                        attn_class=nn.MultiheadAttention,
+                    )
+                    for _ in range(space_depth)
+                ]
+            )
+            self.space_point2virtual_blocks = nn.ModuleList(
+                [CrossAttnBlock(hidden_size, hidden_size, num_heads, mlp_ratio=mlp_ratio) for _ in range(space_depth)]
+            )
+            self.space_virtual2point_blocks = nn.ModuleList(
+                [CrossAttnBlock(hidden_size, hidden_size, num_heads, mlp_ratio=mlp_ratio) for _ in range(space_depth)]
+            )
+            assert len(self.time_blocks) >= len(self.space_virtual2point_blocks)
+        self.initialize_weights()
+
+    def initialize_weights(self):
+        def _basic_init(module):
+            if isinstance(module, nn.Linear):
+                torch.nn.init.xavier_uniform_(module.weight)
+                if module.bias is not None:
+                    nn.init.constant_(module.bias, 0)
+            torch.nn.init.trunc_normal_(self.flow_head.weight, std=0.001)
+
+        self.apply(_basic_init)
+
+    def forward(self, input_tensor, mask=None):
+        # Apply input LayerNorm
+        input_tensor = self.input_norm(input_tensor)
+        tokens = self.input_transform(input_tensor)
+
+        init_tokens = tokens
+
+        B, _, T, _ = tokens.shape
+
+        if self.add_space_attn:
+            virtual_tokens = self.virual_tracks.repeat(B, 1, T, 1)
+            tokens = torch.cat([tokens, virtual_tokens], dim=1)
+
+        _, N, _, _ = tokens.shape
+
+        j = 0
+        for i in range(len(self.time_blocks)):
+            time_tokens = tokens.contiguous().view(B * N, T, -1)  # B N T C -> (B N) T C
+
+            time_tokens = self.time_blocks[i](time_tokens)
+
+            tokens = time_tokens.view(B, N, T, -1)  # (B N) T C -> B N T C
+            if self.add_space_attn and (i % (len(self.time_blocks) // len(self.space_virtual_blocks)) == 0):
+                space_tokens = tokens.permute(0, 2, 1, 3).contiguous().view(B * T, N, -1)  # B N T C -> (B T) N C
+                point_tokens = space_tokens[:, : N - self.num_virtual_tracks]
+                virtual_tokens = space_tokens[:, N - self.num_virtual_tracks :]
+
+                virtual_tokens = self.space_virtual2point_blocks[j](virtual_tokens, point_tokens, mask=mask)
+                virtual_tokens = self.space_virtual_blocks[j](virtual_tokens)
+                point_tokens = self.space_point2virtual_blocks[j](point_tokens, virtual_tokens, mask=mask)
+
+                space_tokens = torch.cat([point_tokens, virtual_tokens], dim=1)
+                tokens = space_tokens.view(B, T, N, -1).permute(0, 2, 1, 3)  # (B T) N C -> B N T C
+                j += 1
+
+        if self.add_space_attn:
+            tokens = tokens[:, : N - self.num_virtual_tracks]
+
+        tokens = tokens + init_tokens
+
+        # Apply output LayerNorm before final projection
+        tokens = self.output_norm(tokens)
+        flow = self.flow_head(tokens)
+
+        return flow, None
+
+
+class CorrBlock:
+    def __init__(self, fmaps, num_levels=4, radius=4, multiple_track_feats=False, padding_mode="zeros"):
+        """
+        Build a pyramid of feature maps from the input.
+
+        fmaps: Tensor (B, S, C, H, W)
+        num_levels: number of pyramid levels (each downsampled by factor 2)
+        radius: search radius for sampling correlation
+        multiple_track_feats: if True, split the target features per pyramid level
+        padding_mode: passed to grid_sample / bilinear_sampler
+        """
+        B, S, C, H, W = fmaps.shape
+        self.S, self.C, self.H, self.W = S, C, H, W
+        self.num_levels = num_levels
+        self.radius = radius
+        self.padding_mode = padding_mode
+        self.multiple_track_feats = multiple_track_feats
+
+        # Build pyramid: each level is half the spatial resolution of the previous
+        self.fmaps_pyramid = [fmaps]  # level 0 is full resolution
+        current_fmaps = fmaps
+        for i in range(num_levels - 1):
+            B, S, C, H, W = current_fmaps.shape
+            # Merge batch & sequence dimensions
+            current_fmaps = current_fmaps.reshape(B * S, C, H, W)
+            # Avg pool down by factor 2
+            current_fmaps = F.avg_pool2d(current_fmaps, kernel_size=2, stride=2)
+            _, _, H_new, W_new = current_fmaps.shape
+            current_fmaps = current_fmaps.reshape(B, S, C, H_new, W_new)
+            self.fmaps_pyramid.append(current_fmaps)
+
+        # Precompute a delta grid (of shape (2r+1, 2r+1, 2)) for sampling.
+        # This grid is added to the (scaled) coordinate centroids.
+        r = self.radius
+        dx = torch.linspace(-r, r, 2 * r + 1, device=fmaps.device, dtype=fmaps.dtype)
+        dy = torch.linspace(-r, r, 2 * r + 1, device=fmaps.device, dtype=fmaps.dtype)
+        # delta: for every (dy,dx) displacement (i.e. Δx, Δy)
+        self.delta = torch.stack(torch.meshgrid(dy, dx, indexing="ij"), dim=-1)  # shape: (2r+1, 2r+1, 2)
+
+    def corr_sample(self, targets, coords):
+        """
+        Instead of storing the entire correlation pyramid, we compute each level's correlation
+        volume, sample it immediately, then discard it. This saves GPU memory.
+
+        Args:
+          targets: Tensor (B, S, N, C) — features for the current targets.
+          coords: Tensor (B, S, N, 2) — coordinates at full resolution.
+
+        Returns:
+          Tensor (B, S, N, L) where L = num_levels * (2*radius+1)**2 (concatenated sampled correlations)
+        """
+        B, S, N, C = targets.shape
+
+        # If you have multiple track features, split them per level.
+        if self.multiple_track_feats:
+            targets_split = torch.split(targets, C // self.num_levels, dim=-1)
+
+        out_pyramid = []
+        for i, fmaps in enumerate(self.fmaps_pyramid):
+            # Get current spatial resolution H, W for this pyramid level.
+            B, S, C, H, W = fmaps.shape
+            # Reshape feature maps for correlation computation:
+            # fmap2s: (B, S, C, H*W)
+            fmap2s = fmaps.view(B, S, C, H * W)
+            # Choose appropriate target features.
+            fmap1 = targets_split[i] if self.multiple_track_feats else targets  # shape: (B, S, N, C)
+
+            # Compute correlation directly
+            corrs = compute_corr_level(fmap1, fmap2s, C)
+            corrs = corrs.view(B, S, N, H, W)
+
+            # Prepare sampling grid:
+            # Scale down the coordinates for the current level.
+            centroid_lvl = coords.reshape(B * S * N, 1, 1, 2) / (2**i)
+            # Make sure our precomputed delta grid is on the same device/dtype.
+            delta_lvl = self.delta.to(coords.device).to(coords.dtype)
+            # Now the grid for grid_sample is:
+            # coords_lvl = centroid_lvl + delta_lvl   (broadcasted over grid)
+            coords_lvl = centroid_lvl + delta_lvl.view(1, 2 * self.radius + 1, 2 * self.radius + 1, 2)
+
+            # Sample from the correlation volume using bilinear interpolation.
+            # We reshape corrs to (B * S * N, 1, H, W) so grid_sample acts over each target.
+            corrs_sampled = bilinear_sampler(
+                corrs.reshape(B * S * N, 1, H, W), coords_lvl, padding_mode=self.padding_mode
+            )
+            # The sampled output is (B * S * N, 1, 2r+1, 2r+1). Flatten the last two dims.
+            corrs_sampled = corrs_sampled.view(B, S, N, -1)  # Now shape: (B, S, N, (2r+1)^2)
+            out_pyramid.append(corrs_sampled)
+
+        # Concatenate all levels along the last dimension.
+        out = torch.cat(out_pyramid, dim=-1).contiguous()
+        return out
+
+
+def compute_corr_level(fmap1, fmap2s, C):
+    # fmap1: (B, S, N, C)
+    # fmap2s: (B, S, C, H*W)
+    corrs = torch.matmul(fmap1, fmap2s)  # (B, S, N, H*W)
+    corrs = corrs.view(fmap1.shape[0], fmap1.shape[1], fmap1.shape[2], -1)  # (B, S, N, H*W)
+    return corrs / math.sqrt(C)

outdoor_v48_4gpu_v2/code/05_02-14:21:58/vggt/heads/track_modules/modules.py

@@ -0,0 +1,218 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from functools import partial
+from typing import Callable
+import collections
+from torch import Tensor
+from itertools import repeat
+
+
+# From PyTorch internals
+def _ntuple(n):
+    def parse(x):
+        if isinstance(x, collections.abc.Iterable) and not isinstance(x, str):
+            return tuple(x)
+        return tuple(repeat(x, n))
+
+    return parse
+
+
+def exists(val):
+    return val is not None
+
+
+def default(val, d):
+    return val if exists(val) else d
+
+
+to_2tuple = _ntuple(2)
+
+
+class ResidualBlock(nn.Module):
+    """
+    ResidualBlock: construct a block of two conv layers with residual connections
+    """
+
+    def __init__(self, in_planes, planes, norm_fn="group", stride=1, kernel_size=3):
+        super(ResidualBlock, self).__init__()
+
+        self.conv1 = nn.Conv2d(
+            in_planes,
+            planes,
+            kernel_size=kernel_size,
+            padding=1,
+            stride=stride,
+            padding_mode="zeros",
+        )
+        self.conv2 = nn.Conv2d(
+            planes,
+            planes,
+            kernel_size=kernel_size,
+            padding=1,
+            padding_mode="zeros",
+        )
+        self.relu = nn.ReLU(inplace=True)
+
+        num_groups = planes // 8
+
+        if norm_fn == "group":
+            self.norm1 = nn.GroupNorm(num_groups=num_groups, num_channels=planes)
+            self.norm2 = nn.GroupNorm(num_groups=num_groups, num_channels=planes)
+            if not stride == 1:
+                self.norm3 = nn.GroupNorm(num_groups=num_groups, num_channels=planes)
+
+        elif norm_fn == "batch":
+            self.norm1 = nn.BatchNorm2d(planes)
+            self.norm2 = nn.BatchNorm2d(planes)
+            if not stride == 1:
+                self.norm3 = nn.BatchNorm2d(planes)
+
+        elif norm_fn == "instance":
+            self.norm1 = nn.InstanceNorm2d(planes)
+            self.norm2 = nn.InstanceNorm2d(planes)
+            if not stride == 1:
+                self.norm3 = nn.InstanceNorm2d(planes)
+
+        elif norm_fn == "none":
+            self.norm1 = nn.Sequential()
+            self.norm2 = nn.Sequential()
+            if not stride == 1:
+                self.norm3 = nn.Sequential()
+        else:
+            raise NotImplementedError
+
+        if stride == 1:
+            self.downsample = None
+        else:
+            self.downsample = nn.Sequential(
+                nn.Conv2d(in_planes, planes, kernel_size=1, stride=stride),
+                self.norm3,
+            )
+
+    def forward(self, x):
+        y = x
+        y = self.relu(self.norm1(self.conv1(y)))
+        y = self.relu(self.norm2(self.conv2(y)))
+
+        if self.downsample is not None:
+            x = self.downsample(x)
+
+        return self.relu(x + y)
+
+
+class Mlp(nn.Module):
+    """MLP as used in Vision Transformer, MLP-Mixer and related networks"""
+
+    def __init__(
+        self,
+        in_features,
+        hidden_features=None,
+        out_features=None,
+        act_layer=nn.GELU,
+        norm_layer=None,
+        bias=True,
+        drop=0.0,
+        use_conv=False,
+    ):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        bias = to_2tuple(bias)
+        drop_probs = to_2tuple(drop)
+        linear_layer = partial(nn.Conv2d, kernel_size=1) if use_conv else nn.Linear
+
+        self.fc1 = linear_layer(in_features, hidden_features, bias=bias[0])
+        self.act = act_layer()
+        self.drop1 = nn.Dropout(drop_probs[0])
+        self.fc2 = linear_layer(hidden_features, out_features, bias=bias[1])
+        self.drop2 = nn.Dropout(drop_probs[1])
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop1(x)
+        x = self.fc2(x)
+        x = self.drop2(x)
+        return x
+
+
+class AttnBlock(nn.Module):
+    def __init__(
+        self,
+        hidden_size,
+        num_heads,
+        attn_class: Callable[..., nn.Module] = nn.MultiheadAttention,
+        mlp_ratio=4.0,
+        **block_kwargs
+    ):
+        """
+        Self attention block
+        """
+        super().__init__()
+
+        self.norm1 = nn.LayerNorm(hidden_size)
+        self.norm2 = nn.LayerNorm(hidden_size)
+
+        self.attn = attn_class(embed_dim=hidden_size, num_heads=num_heads, batch_first=True, **block_kwargs)
+
+        mlp_hidden_dim = int(hidden_size * mlp_ratio)
+
+        self.mlp = Mlp(in_features=hidden_size, hidden_features=mlp_hidden_dim, drop=0)
+
+    def forward(self, x, mask=None):
+        # Prepare the mask for PyTorch's attention (it expects a different format)
+        # attn_mask = mask if mask is not None else None
+        # Normalize before attention
+        x = self.norm1(x)
+
+        # PyTorch's MultiheadAttention returns attn_output, attn_output_weights
+        # attn_output, _ = self.attn(x, x, x, attn_mask=attn_mask)
+
+        attn_output, _ = self.attn(x, x, x)
+
+        # Add & Norm
+        x = x + attn_output
+        x = x + self.mlp(self.norm2(x))
+        return x
+
+
+class CrossAttnBlock(nn.Module):
+    def __init__(self, hidden_size, context_dim, num_heads=1, mlp_ratio=4.0, **block_kwargs):
+        """
+        Cross attention block
+        """
+        super().__init__()
+
+        self.norm1 = nn.LayerNorm(hidden_size)
+        self.norm_context = nn.LayerNorm(hidden_size)
+        self.norm2 = nn.LayerNorm(hidden_size)
+
+        self.cross_attn = nn.MultiheadAttention(
+            embed_dim=hidden_size, num_heads=num_heads, batch_first=True, **block_kwargs
+        )
+
+        mlp_hidden_dim = int(hidden_size * mlp_ratio)
+
+        self.mlp = Mlp(in_features=hidden_size, hidden_features=mlp_hidden_dim, drop=0)
+
+    def forward(self, x, context, mask=None):
+        # Normalize inputs
+        x = self.norm1(x)
+        context = self.norm_context(context)
+
+        # Apply cross attention
+        # Note: nn.MultiheadAttention returns attn_output, attn_output_weights
+        attn_output, _ = self.cross_attn(x, context, context, attn_mask=mask)
+
+        # Add & Norm
+        x = x + attn_output
+        x = x + self.mlp(self.norm2(x))
+        return x

outdoor_v48_4gpu_v2/code/05_02-14:21:58/vggt/heads/track_modules/utils.py

@@ -0,0 +1,226 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+# Modified from https://github.com/facebookresearch/vggsfm
+# and https://github.com/facebookresearch/co-tracker/tree/main
+
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from typing import Optional, Tuple, Union
+
+
+def get_2d_sincos_pos_embed(embed_dim: int, grid_size: Union[int, Tuple[int, int]], return_grid=False) -> torch.Tensor:
+    """
+    This function initializes a grid and generates a 2D positional embedding using sine and cosine functions.
+    It is a wrapper of get_2d_sincos_pos_embed_from_grid.
+    Args:
+    - embed_dim: The embedding dimension.
+    - grid_size: The grid size.
+    Returns:
+    - pos_embed: The generated 2D positional embedding.
+    """
+    if isinstance(grid_size, tuple):
+        grid_size_h, grid_size_w = grid_size
+    else:
+        grid_size_h = grid_size_w = grid_size
+    grid_h = torch.arange(grid_size_h, dtype=torch.float)
+    grid_w = torch.arange(grid_size_w, dtype=torch.float)
+    grid = torch.meshgrid(grid_w, grid_h, indexing="xy")
+    grid = torch.stack(grid, dim=0)
+    grid = grid.reshape([2, 1, grid_size_h, grid_size_w])
+    pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
+    if return_grid:
+        return (
+            pos_embed.reshape(1, grid_size_h, grid_size_w, -1).permute(0, 3, 1, 2),
+            grid,
+        )
+    return pos_embed.reshape(1, grid_size_h, grid_size_w, -1).permute(0, 3, 1, 2)
+
+
+def get_2d_sincos_pos_embed_from_grid(embed_dim: int, grid: torch.Tensor) -> torch.Tensor:
+    """
+    This function generates a 2D positional embedding from a given grid using sine and cosine functions.
+
+    Args:
+    - embed_dim: The embedding dimension.
+    - grid: The grid to generate the embedding from.
+
+    Returns:
+    - emb: The generated 2D positional embedding.
+    """
+    assert embed_dim % 2 == 0
+
+    # use half of dimensions to encode grid_h
+    emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0])  # (H*W, D/2)
+    emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1])  # (H*W, D/2)
+
+    emb = torch.cat([emb_h, emb_w], dim=2)  # (H*W, D)
+    return emb
+
+
+def get_1d_sincos_pos_embed_from_grid(embed_dim: int, pos: torch.Tensor) -> torch.Tensor:
+    """
+    This function generates a 1D positional embedding from a given grid using sine and cosine functions.
+
+    Args:
+    - embed_dim: The embedding dimension.
+    - pos: The position to generate the embedding from.
+
+    Returns:
+    - emb: The generated 1D positional embedding.
+    """
+    assert embed_dim % 2 == 0
+    omega = torch.arange(embed_dim // 2, dtype=torch.double)
+    omega /= embed_dim / 2.0
+    omega = 1.0 / 10000**omega  # (D/2,)
+
+    pos = pos.reshape(-1)  # (M,)
+    out = torch.einsum("m,d->md", pos, omega)  # (M, D/2), outer product
+
+    emb_sin = torch.sin(out)  # (M, D/2)
+    emb_cos = torch.cos(out)  # (M, D/2)
+
+    emb = torch.cat([emb_sin, emb_cos], dim=1)  # (M, D)
+    return emb[None].float()
+
+
+def get_2d_embedding(xy: torch.Tensor, C: int, cat_coords: bool = True) -> torch.Tensor:
+    """
+    This function generates a 2D positional embedding from given coordinates using sine and cosine functions.
+
+    Args:
+    - xy: The coordinates to generate the embedding from.
+    - C: The size of the embedding.
+    - cat_coords: A flag to indicate whether to concatenate the original coordinates to the embedding.
+
+    Returns:
+    - pe: The generated 2D positional embedding.
+    """
+    B, N, D = xy.shape
+    assert D == 2
+
+    x = xy[:, :, 0:1]
+    y = xy[:, :, 1:2]
+    div_term = (torch.arange(0, C, 2, device=xy.device, dtype=torch.float32) * (1000.0 / C)).reshape(1, 1, int(C / 2))
+
+    pe_x = torch.zeros(B, N, C, device=xy.device, dtype=torch.float32)
+    pe_y = torch.zeros(B, N, C, device=xy.device, dtype=torch.float32)
+
+    pe_x[:, :, 0::2] = torch.sin(x * div_term)
+    pe_x[:, :, 1::2] = torch.cos(x * div_term)
+
+    pe_y[:, :, 0::2] = torch.sin(y * div_term)
+    pe_y[:, :, 1::2] = torch.cos(y * div_term)
+
+    pe = torch.cat([pe_x, pe_y], dim=2)  # (B, N, C*3)
+    if cat_coords:
+        pe = torch.cat([xy, pe], dim=2)  # (B, N, C*3+3)
+    return pe
+
+
+def bilinear_sampler(input, coords, align_corners=True, padding_mode="border"):
+    r"""Sample a tensor using bilinear interpolation
+
+    `bilinear_sampler(input, coords)` samples a tensor :attr:`input` at
+    coordinates :attr:`coords` using bilinear interpolation. It is the same
+    as `torch.nn.functional.grid_sample()` but with a different coordinate
+    convention.
+
+    The input tensor is assumed to be of shape :math:`(B, C, H, W)`, where
+    :math:`B` is the batch size, :math:`C` is the number of channels,
+    :math:`H` is the height of the image, and :math:`W` is the width of the
+    image. The tensor :attr:`coords` of shape :math:`(B, H_o, W_o, 2)` is
+    interpreted as an array of 2D point coordinates :math:`(x_i,y_i)`.
+
+    Alternatively, the input tensor can be of size :math:`(B, C, T, H, W)`,
+    in which case sample points are triplets :math:`(t_i,x_i,y_i)`. Note
+    that in this case the order of the components is slightly different
+    from `grid_sample()`, which would expect :math:`(x_i,y_i,t_i)`.
+
+    If `align_corners` is `True`, the coordinate :math:`x` is assumed to be
+    in the range :math:`[0,W-1]`, with 0 corresponding to the center of the
+    left-most image pixel :math:`W-1` to the center of the right-most
+    pixel.
+
+    If `align_corners` is `False`, the coordinate :math:`x` is assumed to
+    be in the range :math:`[0,W]`, with 0 corresponding to the left edge of
+    the left-most pixel :math:`W` to the right edge of the right-most
+    pixel.
+
+    Similar conventions apply to the :math:`y` for the range
+    :math:`[0,H-1]` and :math:`[0,H]` and to :math:`t` for the range
+    :math:`[0,T-1]` and :math:`[0,T]`.
+
+    Args:
+        input (Tensor): batch of input images.
+        coords (Tensor): batch of coordinates.
+        align_corners (bool, optional): Coordinate convention. Defaults to `True`.
+        padding_mode (str, optional): Padding mode. Defaults to `"border"`.
+
+    Returns:
+        Tensor: sampled points.
+    """
+    coords = coords.detach().clone()
+    ############################################################
+    # IMPORTANT:
+    coords = coords.to(input.device).to(input.dtype)
+    ############################################################
+
+    sizes = input.shape[2:]
+
+    assert len(sizes) in [2, 3]
+
+    if len(sizes) == 3:
+        # t x y -> x y t to match dimensions T H W in grid_sample
+        coords = coords[..., [1, 2, 0]]
+
+    if align_corners:
+        scale = torch.tensor(
+            [2 / max(size - 1, 1) for size in reversed(sizes)], device=coords.device, dtype=coords.dtype
+        )
+    else:
+        scale = torch.tensor([2 / size for size in reversed(sizes)], device=coords.device, dtype=coords.dtype)
+
+    coords.mul_(scale)  # coords = coords * scale
+    coords.sub_(1)  # coords = coords - 1
+
+    return F.grid_sample(input, coords, align_corners=align_corners, padding_mode=padding_mode)
+
+
+def sample_features4d(input, coords):
+    r"""Sample spatial features
+
+    `sample_features4d(input, coords)` samples the spatial features
+    :attr:`input` represented by a 4D tensor :math:`(B, C, H, W)`.
+
+    The field is sampled at coordinates :attr:`coords` using bilinear
+    interpolation. :attr:`coords` is assumed to be of shape :math:`(B, R,
+    2)`, where each sample has the format :math:`(x_i, y_i)`. This uses the
+    same convention as :func:`bilinear_sampler` with `align_corners=True`.
+
+    The output tensor has one feature per point, and has shape :math:`(B,
+    R, C)`.
+
+    Args:
+        input (Tensor): spatial features.
+        coords (Tensor): points.
+
+    Returns:
+        Tensor: sampled features.
+    """
+
+    B, _, _, _ = input.shape
+
+    # B R 2 -> B R 1 2
+    coords = coords.unsqueeze(2)
+
+    # B C R 1
+    feats = bilinear_sampler(input, coords)
+
+    return feats.permute(0, 2, 1, 3).view(B, -1, feats.shape[1] * feats.shape[3])  # B C R 1 -> B R C

outdoor_v48_4gpu_v2/code/05_02-14:21:58/vggt/heads/utils.py

@@ -0,0 +1,108 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+import torch.nn as nn
+
+
+def position_grid_to_embed(pos_grid: torch.Tensor, embed_dim: int, omega_0: float = 100) -> torch.Tensor:
+    """
+    Convert 2D position grid (HxWx2) to sinusoidal embeddings (HxWxC)
+
+    Args:
+        pos_grid: Tensor of shape (H, W, 2) containing 2D coordinates
+        embed_dim: Output channel dimension for embeddings
+
+    Returns:
+        Tensor of shape (H, W, embed_dim) with positional embeddings
+    """
+    H, W, grid_dim = pos_grid.shape
+    assert grid_dim == 2
+    pos_flat = pos_grid.reshape(-1, grid_dim)  # Flatten to (H*W, 2)
+
+    # Process x and y coordinates separately
+    emb_x = make_sincos_pos_embed(embed_dim // 2, pos_flat[:, 0], omega_0=omega_0)  # [1, H*W, D/2]
+    emb_y = make_sincos_pos_embed(embed_dim // 2, pos_flat[:, 1], omega_0=omega_0)  # [1, H*W, D/2]
+
+    # Combine and reshape
+    emb = torch.cat([emb_x, emb_y], dim=-1)  # [1, H*W, D]
+
+    return emb.view(H, W, embed_dim)  # [H, W, D]
+
+
+def make_sincos_pos_embed(embed_dim: int, pos: torch.Tensor, omega_0: float = 100) -> torch.Tensor:
+    """
+    This function generates a 1D positional embedding from a given grid using sine and cosine functions.
+
+    Args:
+    - embed_dim: The embedding dimension.
+    - pos: The position to generate the embedding from.
+
+    Returns:
+    - emb: The generated 1D positional embedding.
+    """
+    assert embed_dim % 2 == 0
+    omega = torch.arange(embed_dim // 2, dtype=torch.double, device=pos.device)
+    omega /= embed_dim / 2.0
+    omega = 1.0 / omega_0**omega  # (D/2,)
+
+    pos = pos.reshape(-1)  # (M,)
+    out = torch.einsum("m,d->md", pos, omega)  # (M, D/2), outer product
+
+    emb_sin = torch.sin(out)  # (M, D/2)
+    emb_cos = torch.cos(out)  # (M, D/2)
+
+    emb = torch.cat([emb_sin, emb_cos], dim=1)  # (M, D)
+    return emb.float()
+
+
+# Inspired by https://github.com/microsoft/moge
+
+
+def create_uv_grid(
+    width: int, height: int, aspect_ratio: float = None, dtype: torch.dtype = None, device: torch.device = None
+) -> torch.Tensor:
+    """
+    Create a normalized UV grid of shape (width, height, 2).
+
+    The grid spans horizontally and vertically according to an aspect ratio,
+    ensuring the top-left corner is at (-x_span, -y_span) and the bottom-right
+    corner is at (x_span, y_span), normalized by the diagonal of the plane.
+
+    Args:
+        width (int): Number of points horizontally.
+        height (int): Number of points vertically.
+        aspect_ratio (float, optional): Width-to-height ratio. Defaults to width/height.
+        dtype (torch.dtype, optional): Data type of the resulting tensor.
+        device (torch.device, optional): Device on which the tensor is created.
+
+    Returns:
+        torch.Tensor: A (width, height, 2) tensor of UV coordinates.
+    """
+    # Derive aspect ratio if not explicitly provided
+    if aspect_ratio is None:
+        aspect_ratio = float(width) / float(height)
+
+    # Compute normalized spans for X and Y
+    diag_factor = (aspect_ratio**2 + 1.0) ** 0.5
+    span_x = aspect_ratio / diag_factor
+    span_y = 1.0 / diag_factor
+
+    # Establish the linspace boundaries
+    left_x = -span_x * (width - 1) / width
+    right_x = span_x * (width - 1) / width
+    top_y = -span_y * (height - 1) / height
+    bottom_y = span_y * (height - 1) / height
+
+    # Generate 1D coordinates
+    x_coords = torch.linspace(left_x, right_x, steps=width, dtype=dtype, device=device)
+    y_coords = torch.linspace(top_y, bottom_y, steps=height, dtype=dtype, device=device)
+
+    # Create 2D meshgrid (width x height) and stack into UV
+    uu, vv = torch.meshgrid(x_coords, y_coords, indexing="xy")
+    uv_grid = torch.stack((uu, vv), dim=-1)
+
+    return uv_grid

outdoor_v48_4gpu_v2/code/05_02-14:21:58/vggt/layers/patch_embed.py

@@ -0,0 +1,88 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# found in the LICENSE file in the root directory of this source tree.
+
+# References:
+#   https://github.com/facebookresearch/dino/blob/master/vision_transformer.py
+#   https://github.com/rwightman/pytorch-image-models/tree/master/timm/layers/patch_embed.py
+
+from typing import Callable, Optional, Tuple, Union
+
+from torch import Tensor
+import torch.nn as nn
+
+
+def make_2tuple(x):
+    if isinstance(x, tuple):
+        assert len(x) == 2
+        return x
+
+    assert isinstance(x, int)
+    return (x, x)
+
+
+class PatchEmbed(nn.Module):
+    """
+    2D image to patch embedding: (B,C,H,W) -> (B,N,D)
+
+    Args:
+        img_size: Image size.
+        patch_size: Patch token size.
+        in_chans: Number of input image channels.
+        embed_dim: Number of linear projection output channels.
+        norm_layer: Normalization layer.
+    """
+
+    def __init__(
+        self,
+        img_size: Union[int, Tuple[int, int]] = 224,
+        patch_size: Union[int, Tuple[int, int]] = 16,
+        in_chans: int = 3,
+        embed_dim: int = 768,
+        norm_layer: Optional[Callable] = None,
+        flatten_embedding: bool = True,
+    ) -> None:
+        super().__init__()
+
+        image_HW = make_2tuple(img_size)
+        patch_HW = make_2tuple(patch_size)
+        patch_grid_size = (
+            image_HW[0] // patch_HW[0],
+            image_HW[1] // patch_HW[1],
+        )
+
+        self.img_size = image_HW
+        self.patch_size = patch_HW
+        self.patches_resolution = patch_grid_size
+        self.num_patches = patch_grid_size[0] * patch_grid_size[1]
+
+        self.in_chans = in_chans
+        self.embed_dim = embed_dim
+
+        self.flatten_embedding = flatten_embedding
+
+        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_HW, stride=patch_HW)
+        self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()
+
+    def forward(self, x: Tensor) -> Tensor:
+        _, _, H, W = x.shape
+        patch_H, patch_W = self.patch_size
+
+        assert H % patch_H == 0, f"Input image height {H} is not a multiple of patch height {patch_H}"
+        assert W % patch_W == 0, f"Input image width {W} is not a multiple of patch width: {patch_W}"
+
+        x = self.proj(x)  # B C H W
+        H, W = x.size(2), x.size(3)
+        x = x.flatten(2).transpose(1, 2)  # B HW C
+        x = self.norm(x)
+        if not self.flatten_embedding:
+            x = x.reshape(-1, H, W, self.embed_dim)  # B H W C
+        return x
+
+    def flops(self) -> float:
+        Ho, Wo = self.patches_resolution
+        flops = Ho * Wo * self.embed_dim * self.in_chans * (self.patch_size[0] * self.patch_size[1])
+        if self.norm is not None:
+            flops += Ho * Wo * self.embed_dim
+        return flops

outdoor_v48_4gpu_v2/code/05_02-14:21:58/vggt/models/aggregator.py

@@ -0,0 +1,332 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import logging
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from typing import Optional, Tuple, Union, List, Dict, Any
+
+from vggt.layers import PatchEmbed
+from vggt.layers.block import Block
+from vggt.layers.rope import RotaryPositionEmbedding2D, PositionGetter
+from vggt.layers.vision_transformer import vit_small, vit_base, vit_large, vit_giant2
+
+logger = logging.getLogger(__name__)
+
+_RESNET_MEAN = [0.485, 0.456, 0.406]
+_RESNET_STD = [0.229, 0.224, 0.225]
+
+
+class Aggregator(nn.Module):
+    """
+    The Aggregator applies alternating-attention over input frames,
+    as described in VGGT: Visual Geometry Grounded Transformer.
+
+
+    Args:
+        img_size (int): Image size in pixels.
+        patch_size (int): Size of each patch for PatchEmbed.
+        embed_dim (int): Dimension of the token embeddings.
+        depth (int): Number of blocks.
+        num_heads (int): Number of attention heads.
+        mlp_ratio (float): Ratio of MLP hidden dim to embedding dim.
+        num_register_tokens (int): Number of register tokens.
+        block_fn (nn.Module): The block type used for attention (Block by default).
+        qkv_bias (bool): Whether to include bias in QKV projections.
+        proj_bias (bool): Whether to include bias in the output projection.
+        ffn_bias (bool): Whether to include bias in MLP layers.
+        patch_embed (str): Type of patch embed. e.g., "conv" or "dinov2_vitl14_reg".
+        aa_order (list[str]): The order of alternating attention, e.g. ["frame", "global"].
+        aa_block_size (int): How many blocks to group under each attention type before switching. If not necessary, set to 1.
+        qk_norm (bool): Whether to apply QK normalization.
+        rope_freq (int): Base frequency for rotary embedding. -1 to disable.
+        init_values (float): Init scale for layer scale.
+    """
+
+    def __init__(
+        self,
+        img_size=518,
+        patch_size=14,
+        embed_dim=1024,
+        depth=24,
+        num_heads=16,
+        mlp_ratio=4.0,
+        num_register_tokens=4,
+        block_fn=Block,
+        qkv_bias=True,
+        proj_bias=True,
+        ffn_bias=True,
+        patch_embed="dinov2_vitl14_reg",
+        aa_order=["frame", "global"],
+        aa_block_size=1,
+        qk_norm=True,
+        rope_freq=100,
+        init_values=0.01,
+    ):
+        super().__init__()
+
+        self.__build_patch_embed__(patch_embed, img_size, patch_size, num_register_tokens, embed_dim=embed_dim)
+
+        # Initialize rotary position embedding if frequency > 0
+        self.rope = RotaryPositionEmbedding2D(frequency=rope_freq) if rope_freq > 0 else None
+        self.position_getter = PositionGetter() if self.rope is not None else None
+
+        self.frame_blocks = nn.ModuleList(
+            [
+                block_fn(
+                    dim=embed_dim,
+                    num_heads=num_heads,
+                    mlp_ratio=mlp_ratio,
+                    qkv_bias=qkv_bias,
+                    proj_bias=proj_bias,
+                    ffn_bias=ffn_bias,
+                    init_values=init_values,
+                    qk_norm=qk_norm,
+                    rope=self.rope,
+                )
+                for _ in range(depth)
+            ]
+        )
+
+        self.global_blocks = nn.ModuleList(
+            [
+                block_fn(
+                    dim=embed_dim,
+                    num_heads=num_heads,
+                    mlp_ratio=mlp_ratio,
+                    qkv_bias=qkv_bias,
+                    proj_bias=proj_bias,
+                    ffn_bias=ffn_bias,
+                    init_values=init_values,
+                    qk_norm=qk_norm,
+                    rope=self.rope,
+                )
+                for _ in range(depth)
+            ]
+        )
+
+        self.depth = depth
+        self.aa_order = aa_order
+        self.patch_size = patch_size
+        self.aa_block_size = aa_block_size
+
+        # Validate that depth is divisible by aa_block_size
+        if self.depth % self.aa_block_size != 0:
+            raise ValueError(f"depth ({depth}) must be divisible by aa_block_size ({aa_block_size})")
+
+        self.aa_block_num = self.depth // self.aa_block_size
+
+        # Note: We have two camera tokens, one for the first frame and one for the rest
+        # The same applies for register tokens
+        self.camera_token = nn.Parameter(torch.randn(1, 2, 1, embed_dim))
+        self.register_token = nn.Parameter(torch.randn(1, 2, num_register_tokens, embed_dim))
+
+        # The patch tokens start after the camera and register tokens
+        self.patch_start_idx = 1 + num_register_tokens
+
+        # Initialize parameters with small values
+        nn.init.normal_(self.camera_token, std=1e-6)
+        nn.init.normal_(self.register_token, std=1e-6)
+
+        # Register normalization constants as buffers
+        for name, value in (
+            ("_resnet_mean", _RESNET_MEAN),
+            ("_resnet_std", _RESNET_STD),
+        ):
+            self.register_buffer(
+                name,
+                torch.FloatTensor(value).reshape(1, 1, 3, 1, 1),
+                persistent=False,
+            )
+
+    def __build_patch_embed__(
+        self,
+        patch_embed,
+        img_size,
+        patch_size,
+        num_register_tokens,
+        interpolate_antialias=True,
+        interpolate_offset=0.0,
+        block_chunks=0,
+        init_values=1.0,
+        embed_dim=1024,
+    ):
+        """
+        Build the patch embed layer. If 'conv', we use a
+        simple PatchEmbed conv layer. Otherwise, we use a vision transformer.
+        """
+
+        if "conv" in patch_embed:
+            self.patch_embed = PatchEmbed(img_size=img_size, patch_size=patch_size, in_chans=3, embed_dim=embed_dim)
+        else:
+            vit_models = {
+                "dinov2_vitl14_reg": vit_large,
+                "dinov2_vitb14_reg": vit_base,
+                "dinov2_vits14_reg": vit_small,
+                "dinov2_vitg2_reg": vit_giant2,
+            }
+
+            self.patch_embed = vit_models[patch_embed](
+                img_size=img_size,
+                patch_size=patch_size,
+                num_register_tokens=num_register_tokens,
+                interpolate_antialias=interpolate_antialias,
+                interpolate_offset=interpolate_offset,
+                block_chunks=block_chunks,
+                init_values=init_values,
+            )
+
+            # Disable gradient updates for mask token
+            if hasattr(self.patch_embed, "mask_token"):
+                self.patch_embed.mask_token.requires_grad_(False)
+
+    def forward(
+        self,
+        images: torch.Tensor,
+    ) -> Union[Tuple[List[torch.Tensor], int], Tuple[List[torch.Tensor], int, Dict]]:
+        """
+        Args:
+            images (torch.Tensor): Input images with shape [B, S, 3, H, W], in range [0, 1].
+                B: batch size, S: sequence length, 3: RGB channels, H: height, W: width
+
+        Returns:
+            (list[torch.Tensor], int):
+                The list of outputs from the attention blocks,
+                and the patch_start_idx indicating where patch tokens begin.
+        """
+        B, S, C_in, H, W = images.shape
+        # Normalize images and reshape for patch embed
+        images = (images - self._resnet_mean.to(images.device)) / self._resnet_std.to(images.device)
+
+        # Reshape to [B*S, C, H, W] for patch embedding
+        images = images.reshape(B * S, C_in, H, W)
+        patch_tokens = self.patch_embed(images)
+
+        if isinstance(patch_tokens, dict):
+            patch_tokens = patch_tokens["x_norm_patchtokens"]
+
+        _, P, C = patch_tokens.shape
+
+        camera_token = slice_expand_and_flatten(self.camera_token, B, S)
+        register_token = slice_expand_and_flatten(self.register_token, B, S)
+
+        # Concatenate special tokens with patch tokens
+        tokens = torch.cat([camera_token, register_token, patch_tokens], dim=1)
+
+        pos = None
+        if self.rope is not None:
+            pos = self.position_getter(B * S, H // self.patch_size, W // self.patch_size, device=images.device)
+
+        if self.patch_start_idx > 0:
+            # do not use position embedding for special tokens (camera and register tokens)
+            # so set pos to 0 for the special tokens
+            pos = pos + 1
+            pos_special = torch.zeros(B * S, self.patch_start_idx, 2).to(images.device).to(pos.dtype)
+            pos = torch.cat([pos_special, pos], dim=1)
+
+        # update P because we added special tokens
+        _, P, C = tokens.shape
+
+        frame_idx = 0
+        global_idx = 0
+        output_list = []
+
+        for block_num in range(self.aa_block_num):
+            for attn_type in self.aa_order:
+                if attn_type == "frame":
+                    tokens, frame_idx, frame_intermediates = self._process_frame_attention(
+                        tokens, B, S, P, C, frame_idx, pos=pos
+                    )
+                elif attn_type == "global":
+                    tokens, global_idx, global_intermediates = self._process_global_attention(
+                        tokens, B, S, P, C, global_idx, pos=pos
+                    )
+                else:
+                    raise ValueError(f"Unknown attention type: {attn_type}")
+
+            for i in range(len(frame_intermediates)):
+                # concat frame and global intermediates, [B x S x P x 2C]
+                concat_inter = torch.cat([frame_intermediates[i], global_intermediates[i]], dim=-1)
+                output_list.append(concat_inter)
+
+        del concat_inter
+        del frame_intermediates
+        del global_intermediates
+
+        return output_list, self.patch_start_idx
+
+
+    def _process_frame_attention(self, tokens, B, S, P, C, frame_idx, pos=None):
+        """
+        Process frame attention blocks. We keep tokens in shape (B*S, P, C).
+        """
+        # If needed, reshape tokens or positions:
+        if tokens.shape != (B * S, P, C):
+            tokens = tokens.reshape(B, S, P, C).reshape(B * S, P, C)
+
+        if pos is not None and pos.shape != (B * S, P, 2):
+            pos = pos.reshape(B, S, P, 2).reshape(B * S, P, 2)
+
+        intermediates = []
+
+        # by default, self.aa_block_size=1, which processes one block at a time
+        for _ in range(self.aa_block_size):
+            tokens = self.frame_blocks[frame_idx](tokens, pos=pos)
+            frame_idx += 1
+            intermediates.append(tokens.reshape(B, S, P, C))
+
+        return tokens, frame_idx, intermediates
+
+
+    def _process_global_attention(self, tokens, B, S, P, C, global_idx, pos=None) -> Union[Tuple[torch.Tensor, int, List[torch.Tensor]], Tuple[torch.Tensor, int, List[torch.Tensor], List]]:
+        """
+        Process global attention blocks. We keep tokens in shape (B, S*P, C).
+        """
+        
+        if tokens.shape != (B, S * P, C):
+            tokens = tokens.reshape(B, S, P, C).reshape(B, S * P, C)
+
+        if pos is not None and pos.shape != (B, S * P, 2):
+            pos = pos.reshape(B, S, P, 2).reshape(B, S * P, 2)
+            
+        intermediates = []
+        
+        for _ in range(self.aa_block_size):
+            tokens = self.global_blocks[global_idx](tokens, pos=pos)
+            
+            global_idx += 1
+            intermediates.append(tokens.reshape(B, S, P, C))
+
+        return tokens, global_idx, intermediates
+
+
+
+
+def slice_expand_and_flatten(token_tensor, B, S):
+    """
+    Processes specialized tokens with shape (1, 2, X, C) for multi-frame processing:
+    1) Uses the first position (index=0) for the first frame only
+    2) Uses the second position (index=1) for all remaining frames (S-1 frames)
+    3) Expands both to match batch size B   
+    4) Concatenates to form (B, S, X, C) where each sequence has 1 first-position token
+       followed by (S-1) second-position tokens
+    5) Flattens to (B*S, X, C) for processing
+
+    Returns:
+        torch.Tensor: Processed tokens with shape (B*S, X, C)
+    """
+
+    # Slice out the "query" tokens => shape (1, 1, ...)
+    query = token_tensor[:, 0:1, ...].expand(B, 1, *token_tensor.shape[2:])
+    # Slice out the "other" tokens => shape (1, S-1, ...)
+    others = token_tensor[:, 1:, ...].expand(B, S - 1, *token_tensor.shape[2:])
+    # Concatenate => shape (B, S, ...)
+    combined = torch.cat([query, others], dim=1)
+
+    # Finally flatten => shape (B*S, ...)
+    combined = combined.reshape(B * S, *combined.shape[2:])
+    return combined

outdoor_v48_4gpu_v2/code/05_02-14:21:58/vggt/models/vggt.py

@@ -0,0 +1,228 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+import torch.nn as nn
+from huggingface_hub import PyTorchModelHubMixin  # used for model hub
+
+from vggt.models.aggregator import Aggregator
+from vggt.heads.camera_head import CameraHead
+from vggt.heads.dpt_head import DPTHead
+from vggt.heads.track_head import TrackHead
+from transformers.file_utils import ModelOutput
+from typing import Optional, Tuple, List, Any
+from dataclasses import dataclass
+
+@dataclass
+class VGGTOutput(ModelOutput):
+    ress: Optional[List[dict]] = None
+    views: Optional[torch.Tensor] = None
+
+class VGGT(nn.Module, PyTorchModelHubMixin):
+    def __init__(self, img_size=518, patch_size=14, embed_dim=1024):
+        super().__init__()
+
+        self.aggregator = Aggregator(img_size=img_size, patch_size=patch_size, embed_dim=embed_dim)
+        self.camera_head = CameraHead(dim_in=2 * embed_dim)
+        self.point_head = DPTHead(dim_in=2 * embed_dim, output_dim=4, activation="inv_log", conf_activation="expp1")
+        self.depth_head = DPTHead(dim_in=2 * embed_dim, output_dim=2, activation="exp", conf_activation="expp1")
+        self.track_head = TrackHead(dim_in=2 * embed_dim, patch_size=patch_size)
+
+    def forward(
+        self,
+        views,
+        query_points: torch.Tensor = None,
+    ):
+        """
+        Forward pass of the VGGT model.
+
+        Args:
+            images (torch.Tensor): Input images with shape [S, 3, H, W] or [B, S, 3, H, W], in range [0, 1].
+                B: batch size, S: sequence length, 3: RGB channels, H: height, W: width
+            query_points (torch.Tensor, optional): Query points for tracking, in pixel coordinates.
+                Shape: [N, 2] or [B, N, 2], where N is the number of query points.
+                Default: None
+            history_info (dict, optional): Token history for streaming inference
+            past_key_values (dict, optional): KV cache from previous steps
+            use_cache (bool): Whether to use and return KV cache
+            past_frame_idx (int): Frame index for position encoding in sequence
+
+        Returns:
+            dict: A dictionary containing the following predictions:
+                - pose_enc (torch.Tensor): Camera pose encoding with shape [B, S, 9] (from the last iteration)
+                - depth (torch.Tensor): Predicted depth maps with shape [B, S, H, W, 1]
+                - depth_conf (torch.Tensor): Confidence scores for depth predictions with shape [B, S, H, W]
+                - world_points (torch.Tensor): 3D world coordinates for each pixel with shape [B, S, H, W, 3]
+                - world_points_conf (torch.Tensor): Confidence scores for world points with shape [B, S, H, W]
+                - images (torch.Tensor): Original input images, preserved for visualization
+
+                If query_points is provided, also includes:
+                - track (torch.Tensor): Point tracks with shape [B, S, N, 2] (from the last iteration), in pixel coordinates
+                - vis (torch.Tensor): Visibility scores for tracked points with shape [B, S, N]
+                - conf (torch.Tensor): Confidence scores for tracked points with shape [B, S, N]
+        """
+        images = torch.stack(
+            [view["img"] for view in views], dim=0
+        ).permute(1, 0, 2, 3, 4)    # B S C H 
+
+        # If without batch dimension, add it
+        if len(images.shape) == 4:
+            images = images.unsqueeze(0)
+
+        if query_points is not None and len(query_points.shape) == 2:
+            query_points = query_points.unsqueeze(0)
+        
+        aggregated_tokens_list, patch_start_idx = self.aggregator(images)
+        predictions = {}
+
+        with torch.cuda.amp.autocast(enabled=False):
+            if self.camera_head is not None:
+                pose_enc_list = self.camera_head(aggregated_tokens_list)
+                predictions["pose_enc"] = pose_enc_list[-1]  # pose encoding of the last iteration
+
+            if self.depth_head is not None:
+                depth, depth_conf = self.depth_head(
+                    aggregated_tokens_list, images=images, patch_start_idx=patch_start_idx
+                )
+                predictions["depth"] = depth
+                predictions["depth_conf"] = depth_conf
+
+            if self.point_head is not None:
+                pts3d, pts3d_conf = self.point_head(
+                    aggregated_tokens_list, images=images, patch_start_idx=patch_start_idx
+                )
+                predictions["world_points"] = pts3d
+                predictions["world_points_conf"] = pts3d_conf
+            if self.track_head is not None and query_points is not None:
+                track_list, vis, conf = self.track_head(
+                    aggregated_tokens_list, images=images, patch_start_idx=patch_start_idx,
+                    query_points=query_points
+                )
+                predictions["track"] = track_list[-1]  # track of the last iteration
+                predictions["vis"] = vis
+                predictions["conf"] = conf
+            predictions["images"] = images
+
+            B, S = images.shape[:2]
+            ress = []
+            for s in range(S):
+                res = {
+                    'pts3d_in_other_view': predictions['world_points'][:, s],  # [B, H, W, 3]
+                    'conf': predictions['world_points_conf'][:, s],  # [B, H, W]
+
+                    'depth': predictions['depth'][:, s],  # [B, H, W, 1]
+                    'depth_conf': predictions['depth_conf'][:, s],  # [B, H, W]
+
+                    'camera_pose': predictions['pose_enc'][:, s, :7],  # [B, 7]
+
+                    **({'valid_mask': views[s]["valid_mask"]}
+                        if 'valid_mask' in views[s] else {}),  # [B, H, W]
+
+                    **({'track': predictions['track'][:, s],  # [B, N, 2]
+                        'vis': predictions['vis'][:, s],  # [B, N]
+                        'track_conf': predictions['conf'][:, s]}
+                        if 'track' in predictions else {})
+                }
+                ress.append(res)
+            return VGGTOutput(ress=ress, views=views)  # [S] [B, C, H, W]
+    
+    def inference(
+        self,
+        views,
+        query_points: torch.Tensor = None,
+    ):
+        """
+        Forward pass of the VGGT model.
+
+        Args:
+            images (torch.Tensor): Input images with shape [S, 3, H, W] or [B, S, 3, H, W], in range [0, 1].
+                B: batch size, S: sequence length, 3: RGB channels, H: height, W: width
+            query_points (torch.Tensor, optional): Query points for tracking, in pixel coordinates.
+                Shape: [N, 2] or [B, N, 2], where N is the number of query points.
+                Default: None
+            history_info (dict, optional): Token history for streaming inference
+            past_key_values (dict, optional): KV cache from previous steps
+            use_cache (bool): Whether to use and return KV cache
+            past_frame_idx (int): Frame index for position encoding in sequence
+
+        Returns:
+            dict: A dictionary containing the following predictions:
+                - pose_enc (torch.Tensor): Camera pose encoding with shape [B, S, 9] (from the last iteration)
+                - depth (torch.Tensor): Predicted depth maps with shape [B, S, H, W, 1]
+                - depth_conf (torch.Tensor): Confidence scores for depth predictions with shape [B, S, H, W]
+                - world_points (torch.Tensor): 3D world coordinates for each pixel with shape [B, S, H, W, 3]
+                - world_points_conf (torch.Tensor): Confidence scores for world points with shape [B, S, H, W]
+                - images (torch.Tensor): Original input images, preserved for visualization
+
+                If query_points is provided, also includes:
+                - track (torch.Tensor): Point tracks with shape [B, S, N, 2] (from the last iteration), in pixel coordinates
+                - vis (torch.Tensor): Visibility scores for tracked points with shape [B, S, N]
+                - conf (torch.Tensor): Confidence scores for tracked points with shape [B, S, N]
+        """
+        images = torch.stack(
+            [view["img"] for view in views], dim=0
+        ).permute(1, 0, 2, 3, 4)    # B S C H W
+
+        # If without batch dimension, add it
+        if len(images.shape) == 4:
+            images = images.unsqueeze(0)
+
+        if query_points is not None and len(query_points.shape) == 2:
+            query_points = query_points.unsqueeze(0)
+        
+        aggregated_tokens_list, patch_start_idx = self.aggregator(images)
+        predictions = {}
+
+        with torch.cuda.amp.autocast(enabled=False):
+            if self.camera_head is not None:
+                pose_enc_list = self.camera_head(aggregated_tokens_list)
+                predictions["pose_enc"] = pose_enc_list[-1]  # pose encoding of the last iteration
+
+            if self.depth_head is not None:
+                depth, depth_conf = self.depth_head(
+                    aggregated_tokens_list, images=images, patch_start_idx=patch_start_idx
+                )
+                predictions["depth"] = depth
+                predictions["depth_conf"] = depth_conf
+
+            if self.point_head is not None:
+                pts3d, pts3d_conf = self.point_head(
+                    aggregated_tokens_list, images=images, patch_start_idx=patch_start_idx
+                )
+                predictions["world_points"] = pts3d
+                predictions["world_points_conf"] = pts3d_conf
+            if self.track_head is not None and query_points is not None:
+                track_list, vis, conf = self.track_head(
+                    aggregated_tokens_list, images=images, patch_start_idx=patch_start_idx,
+                    query_points=query_points
+                )
+                predictions["track"] = track_list[-1]  # track of the last iteration
+                predictions["vis"] = vis
+                predictions["conf"] = conf
+            predictions["images"] = images
+            
+            B, S = images.shape[:2]
+            ress = []
+            for s in range(S):
+                res = {
+                    'pts3d_in_other_view': predictions['world_points'][:, s],  # [B, H, W, 3]
+                    'conf': predictions['world_points_conf'][:, s],  # [B, H, W]
+
+                    'depth': predictions['depth'][:, s],  # [B, H, W, 1]
+                    'depth_conf': predictions['depth_conf'][:, s],  # [B, H, W]
+
+                    'camera_pose': predictions['pose_enc'][:, s, :],  # [B, 9]
+
+                    **({'valid_mask': views[s]["valid_mask"]}
+                        if 'valid_mask' in views[s] else {}),  # [B, H, W]
+
+                    **({'track': predictions['track'][:, s],  # [B, N, 2]
+                        'vis': predictions['vis'][:, s],  # [B, N]
+                        'track_conf': predictions['conf'][:, s]}
+                        if 'track' in predictions else {})
+                }
+                ress.append(res)
+            return VGGTOutput(ress=ress, views=views)  # [S] [B, C, H, W]

outdoor_v48_4gpu_v2/code/05_02-14:21:58/vggt/train_utils/augmentation.py

@@ -0,0 +1,72 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from typing import Optional, Dict
+from torchvision import transforms
+
+
+def get_image_augmentation(
+    color_jitter: Optional[Dict[str, float]] = None,
+    gray_scale: bool = True,
+    gau_blur: bool = False
+) -> Optional[transforms.Compose]:
+    """Create a composition of image augmentations.
+
+    Args:
+        color_jitter: Dictionary containing color jitter parameters:
+            - brightness: float (default: 0.5)
+            - contrast: float (default: 0.5)
+            - saturation: float (default: 0.5)
+            - hue: float (default: 0.1)
+            - p: probability of applying (default: 0.9)
+            If None, uses default values
+        gray_scale: Whether to apply random grayscale (default: True)
+        gau_blur: Whether to apply gaussian blur (default: False)
+
+    Returns:
+        A Compose object of transforms or None if no transforms are added
+    """
+    transform_list = []
+    default_jitter = {
+        "brightness": 0.5,
+        "contrast": 0.5,
+        "saturation": 0.5,
+        "hue": 0.1,
+        "p": 0.9
+    }
+
+    # Handle color jitter
+    if color_jitter is not None:
+        # Merge with defaults for missing keys
+        effective_jitter = {**default_jitter, **color_jitter}
+    else:
+        effective_jitter = default_jitter
+
+    transform_list.append(
+        transforms.RandomApply(
+            [
+                transforms.ColorJitter(
+                    brightness=effective_jitter["brightness"],
+                    contrast=effective_jitter["contrast"],
+                    saturation=effective_jitter["saturation"],
+                    hue=effective_jitter["hue"],
+                )
+            ],
+            p=effective_jitter["p"],
+        )
+    )
+
+    if gray_scale:
+        transform_list.append(transforms.RandomGrayscale(p=0.05))
+
+    if gau_blur:
+        transform_list.append(
+            transforms.RandomApply(
+                [transforms.GaussianBlur(5, sigma=(0.1, 1.0))], p=0.05
+            )
+        )
+
+    return transforms.Compose(transform_list) if transform_list else None

outdoor_v48_4gpu_v2/code/05_02-14:21:58/vggt/train_utils/general.py

@@ -0,0 +1,369 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+import torch.nn as nn
+import os
+import math
+import random
+import numpy as np
+from typing import Union, Optional
+import logging
+from iopath.common.file_io import g_pathmgr
+import torch.distributed as dist
+from pathlib import Path
+from typing import Dict, Iterable, List
+
+
+
+from collections import defaultdict
+from dataclasses import fields, is_dataclass
+from typing import Any, Mapping, Protocol, runtime_checkable
+
+
+
+
+def check_and_fix_inf_nan(input_tensor, loss_name="default", hard_max=100):
+    """
+    Checks if 'input_tensor' contains inf or nan values and clamps extreme values.
+    
+    Args:
+        input_tensor (torch.Tensor): The loss tensor to check and fix.
+        loss_name (str): Name of the loss (for diagnostic prints).
+        hard_max (float, optional): Maximum absolute value allowed. Values outside 
+                                  [-hard_max, hard_max] will be clamped. If None, 
+                                  no clamping is performed. Defaults to 100.
+    """
+    if input_tensor is None:
+        return input_tensor
+    
+    # Check for inf/nan values
+    has_inf_nan = torch.isnan(input_tensor).any() or torch.isinf(input_tensor).any()
+    if has_inf_nan:
+        logging.warning(f"Tensor {loss_name} contains inf or nan values. Replacing with zeros.")
+        input_tensor = torch.where(
+            torch.isnan(input_tensor) | torch.isinf(input_tensor),
+            torch.zeros_like(input_tensor),
+            input_tensor
+        )
+
+    # Apply hard clamping if specified
+    if hard_max is not None:
+        input_tensor = torch.clamp(input_tensor, min=-hard_max, max=hard_max)
+
+    return input_tensor
+
+
+def get_resume_checkpoint(checkpoint_save_dir):
+    if not g_pathmgr.isdir(checkpoint_save_dir):
+        return None
+    ckpt_file = os.path.join(checkpoint_save_dir, "checkpoint.pt")
+    if not g_pathmgr.isfile(ckpt_file):
+        return None
+
+    return ckpt_file
+
+class DurationMeter:
+    def __init__(self, name, device, fmt=":f"):
+        self.name = name
+        self.device = device
+        self.fmt = fmt
+        self.val = 0
+
+    def reset(self):
+        self.val = 0
+
+    def update(self, val):
+        self.val = val
+
+    def add(self, val):
+        self.val += val
+
+    def __str__(self):
+        return f"{self.name}: {human_readable_time(self.val)}"
+
+
+def human_readable_time(time_seconds):
+    time = int(time_seconds)
+    minutes, seconds = divmod(time, 60)
+    hours, minutes = divmod(minutes, 60)
+    days, hours = divmod(hours, 24)
+    return f"{days:02}d {hours:02}h {minutes:02}m"
+
+
+
+class ProgressMeter:
+    def __init__(self, num_batches, meters, real_meters, prefix=""):
+        self.batch_fmtstr = self._get_batch_fmtstr(num_batches)
+        self.meters = meters
+        self.real_meters = real_meters
+        self.prefix = prefix
+
+    def display(self, batch):
+        entries = [self.prefix + self.batch_fmtstr.format(batch)]
+        entries += [str(meter) for meter in self.meters]
+        entries += [
+            " | ".join(
+                [
+                    f"{os.path.join(name, subname)}: {val:.4f}"
+                    for subname, val in meter.compute().items()
+                ]
+            )
+            for name, meter in self.real_meters.items()
+        ]
+        logging.info(" | ".join(entries))
+
+    def _get_batch_fmtstr(self, num_batches):
+        num_digits = len(str(num_batches // 1))
+        fmt = "{:" + str(num_digits) + "d}"
+        return "[" + fmt + "/" + fmt.format(num_batches) + "]"
+
+
+
+@runtime_checkable
+class _CopyableData(Protocol):
+    def to(self, device: torch.device, *args: Any, **kwargs: Any):
+        """Copy data to the specified device"""
+        ...
+
+
+def _is_named_tuple(x) -> bool:
+    return isinstance(x, tuple) and hasattr(x, "_asdict") and hasattr(x, "_fields")
+
+
+def copy_data_to_device(data, device: torch.device, *args: Any, **kwargs: Any):
+    """Function that recursively copies data to a torch.device.
+
+    Args:
+        data: The data to copy to device
+        device: The device to which the data should be copied
+        args: positional arguments that will be passed to the `to` call
+        kwargs: keyword arguments that will be passed to the `to` call
+
+    Returns:
+        The data on the correct device
+    """
+
+    if _is_named_tuple(data):
+        return type(data)(
+            **copy_data_to_device(data._asdict(), device, *args, **kwargs)
+        )
+    elif isinstance(data, (list, tuple)):
+        return type(data)(copy_data_to_device(e, device, *args, **kwargs) for e in data)
+    elif isinstance(data, defaultdict):
+        return type(data)(
+            data.default_factory,
+            {
+                k: copy_data_to_device(v, device, *args, **kwargs)
+                for k, v in data.items()
+            },
+        )
+    elif isinstance(data, Mapping) and not is_dataclass(data):  # handing FrameData-like things
+        return type(data)(
+            {
+                k: copy_data_to_device(v, device, *args, **kwargs)
+                for k, v in data.items()
+            }
+        )
+    elif is_dataclass(data) and not isinstance(data, type):
+        new_data_class = type(data)(
+            **{
+                field.name: copy_data_to_device(
+                    getattr(data, field.name), device, *args, **kwargs
+                )
+                for field in fields(data)
+                if field.init
+            }
+        )
+        for field in fields(data):
+            if not field.init:
+                setattr(
+                    new_data_class,
+                    field.name,
+                    copy_data_to_device(
+                        getattr(data, field.name), device, *args, **kwargs
+                    ),
+                )
+        return new_data_class
+    elif isinstance(data, _CopyableData):
+        return data.to(device, *args, **kwargs)
+    return data
+
+
+
+def safe_makedirs(path: str):
+    if not path:
+        logging.warning("safe_makedirs called with an empty path. No operation performed.")
+        return False
+
+    try:
+        os.makedirs(path, exist_ok=True)
+        return True
+    except OSError as e:
+        logging.error(f"Failed to create directory '{path}'. Reason: {e}")
+        raise
+    except Exception as e:
+        # Catch any other unexpected errors.
+        logging.error(f"An unexpected error occurred while creating directory '{path}'. Reason: {e}")
+        raise
+
+
+
+def set_seeds(seed_value, max_epochs, dist_rank):
+    """
+    Set the python random, numpy and torch seed for each gpu. Also set the CUDA
+    seeds if the CUDA is available. This ensures deterministic nature of the training.
+    """
+    seed_value = (seed_value + dist_rank) * max_epochs
+    logging.info(f"GPU SEED: {seed_value}")
+    random.seed(seed_value)
+    np.random.seed(seed_value)
+    torch.manual_seed(seed_value)
+
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed(seed_value)
+        torch.cuda.manual_seed_all(seed_value)  # for multi-GPU
+
+
+
+
+def log_env_variables():
+    env_keys = sorted(list(os.environ.keys()))
+    st = ""
+    for k in env_keys:
+        v = os.environ[k]
+        st += f"{k}={v}\n"
+    logging.info("Logging ENV_VARIABLES")
+    logging.info(st)
+
+
+def is_dist_avail_and_initialized():
+    if not dist.is_available():
+        return False
+    if not dist.is_initialized():
+        return False
+    return True
+
+
+
+class AverageMeter:
+    """Computes and stores the average and current value.
+    Args:
+        name (str): Name of the metric being tracked
+        device (torch.device, optional): Device for tensor operations. Defaults to None.
+        fmt (str): Format string for displaying values. Defaults to ":f"
+    """
+
+    def __init__(self, name: str, device: Optional[torch.device] = None, fmt: str = ":f"):
+        self.name = name
+        self.fmt = fmt
+        self.device = device
+        self.reset()
+
+    def reset(self):
+        self.val = 0
+        self.avg = 0
+        self.sum = 0
+        self.count = 0
+        self._allow_updates = True
+
+    def update(self, val, n=1):
+        if n <= 0:
+            raise ValueError(f"n must be positive, got {n}")
+
+        self.val = val
+        self.sum += val * n
+        self.count += n
+        self.avg = self.sum / self.count if self.count > 0 else 0.0
+
+    def __str__(self) -> str:
+        """String representation showing current and average values."""
+        fmtstr = "{name}: {val" + self.fmt + "} ({avg" + self.fmt + "})"
+        return fmtstr.format(**self.__dict__)
+
+    @property
+    def value(self) -> float:
+        """Get the current value."""
+        return self.val
+
+    @property
+    def average(self) -> float:
+        """Get the running average."""
+        return self.avg
+
+#################
+
+
+_UNITS = ('', ' K', ' M', ' B', ' T')          # U+202F = thin-space for nicer look
+
+def pretty_int(n: int) -> str:
+    """Abbreviate a non-negative integer (0 → 0, 12_345 → '12.3 K')."""
+    assert n >= 0, 'pretty_int() expects a non-negative int'
+    if n < 1_000:
+        return f'{n:,}'
+    exp = int(math.log10(n) // 3)        # group of 3 digits
+    exp = min(exp, len(_UNITS) - 1)      # cap at trillions
+    value = n / 10 ** (3 * exp)
+    return f'{value:.1f}'.rstrip('0').rstrip('.') + _UNITS[exp]
+
+
+def model_summary(model: torch.nn.Module,
+                  *,
+                  log_file = None,
+                  prefix: str = '') -> None:
+    """
+    Print / save a compact parameter summary.
+
+    Args
+    ----
+    model      : The PyTorch nn.Module to inspect.
+    log_file   : Optional path – if given, the full `str(model)` and per-parameter
+                 lists are written there (three separate *.txt files).
+    prefix     : Optional string printed at the beginning of every log line
+                 (handy when several models share the same stdout).
+    """
+    if get_rank():          # only rank-0 prints
+        return
+
+    # --- counts -------------------------------------------------------------
+    trainable = sum(p.numel() for p in model.parameters() if p.requires_grad)
+    total     = sum(p.numel() for p in model.parameters())
+    frozen    = total - trainable
+
+    print(prefix + '='*60)
+    print(prefix + f'Model type : {model.__class__.__name__}')
+    print(prefix + f'Total      : {pretty_int(total)} parameters')
+    print(prefix + f'  trainable: {pretty_int(trainable)}')
+    print(prefix + f'  frozen   : {pretty_int(frozen)}')
+    print(prefix + '='*60)
+
+    # --- optional file dump -------------------------------------------------
+    if log_file is None:
+        return
+
+    log_file = Path(log_file)
+    log_file.write_text(str(model))                      # full architecture
+
+    # two extra detailed lists
+    def _dump(names: Iterable[str], fname: str):
+        """Write a formatted per-parameter list to *log_file.with_name(fname)*."""
+        with open(log_file.with_name(fname), 'w') as f:
+            for n in names:
+                p = dict(model.named_parameters())[n]
+                shape = str(tuple(p.shape))
+                f.write(f'{n:<60s} {shape:<20} {p.numel()}\n')
+
+    named = dict(model.named_parameters())
+    _dump([n for n,p in named.items() if p.requires_grad],  'trainable.txt')
+    _dump([n for n,p in named.items() if not p.requires_grad], 'frozen.txt')
+
+
+def get_rank():
+    if not is_dist_avail_and_initialized():
+        return 0
+    return dist.get_rank()
+
+

outdoor_v48_4gpu_v2/code/05_02-14:21:58/vggt/train_utils/normalization.py

@@ -0,0 +1,130 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+import logging
+from typing import Optional, Tuple
+from vggt.utils.geometry import closed_form_inverse_se3
+from vggt.train_utils.general import check_and_fix_inf_nan
+
+
+def check_valid_tensor(input_tensor: Optional[torch.Tensor], name: str = "tensor") -> None:
+    """
+    Check if a tensor contains NaN or Inf values and log a warning if found.
+    
+    Args:
+        input_tensor: The tensor to check
+        name: Name of the tensor for logging purposes
+    """
+    if input_tensor is not None:
+        if torch.isnan(input_tensor).any() or torch.isinf(input_tensor).any():
+            logging.warning(f"NaN or Inf found in tensor: {name}")
+
+
+def normalize_camera_extrinsics_and_points_batch(
+    extrinsics: torch.Tensor,
+    cam_points: Optional[torch.Tensor] = None,
+    world_points: Optional[torch.Tensor] = None,
+    depths: Optional[torch.Tensor] = None,
+    scale_by_points: bool = True,
+    point_masks: Optional[torch.Tensor] = None,
+) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[torch.Tensor], Optional[torch.Tensor]]:
+    """
+    Normalize camera extrinsics and corresponding 3D points.
+    
+    This function transforms the coordinate system to be centered at the first camera
+    and optionally scales the scene to have unit average distance.
+    
+    Args:
+        extrinsics: Camera extrinsic matrices of shape (B, S, 3, 4)
+        cam_points: 3D points in camera coordinates of shape (B, S, H, W, 3) or (*,3)
+        world_points: 3D points in world coordinates of shape (B, S, H, W, 3) or (*,3)
+        depths: Depth maps of shape (B, S, H, W)
+        scale_by_points: Whether to normalize the scale based on point distances
+        point_masks: Boolean masks for valid points of shape (B, S, H, W)
+    
+    Returns:
+        Tuple containing:
+        - Normalized camera extrinsics of shape (B, S, 3, 4)
+        - Normalized camera points (same shape as input cam_points)
+        - Normalized world points (same shape as input world_points)
+        - Normalized depths (same shape as input depths)
+    """
+    # Validate inputs
+    check_valid_tensor(extrinsics, "extrinsics")
+    check_valid_tensor(cam_points, "cam_points")
+    check_valid_tensor(world_points, "world_points")
+    check_valid_tensor(depths, "depths")
+
+
+    B, S, _, _ = extrinsics.shape
+    device = extrinsics.device
+    #assert device == torch.device("cpu")
+
+
+    # Convert extrinsics to homogeneous form: (B, N,4,4)
+    extrinsics_homog = torch.cat(
+        [
+            extrinsics,
+            torch.zeros((B, S, 1, 4), device=device),
+        ],
+        dim=-2,
+    )
+    extrinsics_homog[:, :, -1, -1] = 1.0
+
+    # first_cam_extrinsic_inv, the inverse of the first camera's extrinsic matrix
+    # which can be also viewed as the cam_to_world extrinsic matrix
+    first_cam_extrinsic_inv = closed_form_inverse_se3(extrinsics_homog[:, 0])
+    # new_extrinsics = torch.matmul(extrinsics_homog, first_cam_extrinsic_inv)
+    new_extrinsics = torch.matmul(extrinsics_homog, first_cam_extrinsic_inv.unsqueeze(1))  # (B,N,4,4)
+
+
+    if world_points is not None:
+        # since we are transforming the world points to the first camera's coordinate system
+        # we directly use the cam_from_world extrinsic matrix of the first camera
+        # instead of using the inverse of the first camera's extrinsic matrix
+        R = extrinsics[:, 0, :3, :3]
+        t = extrinsics[:, 0, :3, 3]
+        new_world_points = (world_points @ R.transpose(-1, -2).unsqueeze(1).unsqueeze(2)) + t.unsqueeze(1).unsqueeze(2).unsqueeze(3)
+    else:
+        new_world_points = None
+
+
+    if scale_by_points:
+        new_depths = depths.clone()
+
+        dist = new_world_points.norm(dim=-1)
+        dist_sum = (dist * point_masks).sum(dim=[1,2,3])
+        valid_count = point_masks.sum(dim=[1,2,3])
+        avg_scale = (dist_sum / (valid_count + 1e-3)).clamp(min=1e-6, max=1e6)
+
+
+        new_world_points = new_world_points / avg_scale.view(-1, 1, 1, 1, 1)
+        new_extrinsics[:, :, :3, 3] = new_extrinsics[:, :, :3, 3] / avg_scale.view(-1, 1, 1)
+        if depths is not None:
+            new_depths = new_depths / avg_scale.view(-1, 1, 1, 1)
+        if cam_points is not None:
+            new_cam_points = cam_points.clone()
+            new_cam_points = new_cam_points / avg_scale.view(-1, 1, 1, 1, 1)
+    else:
+        return new_extrinsics[:, :, :3], cam_points, new_world_points, depths
+
+    new_extrinsics = new_extrinsics[:, :, :3] # 4x4 -> 3x4
+    new_extrinsics = check_and_fix_inf_nan(new_extrinsics, "new_extrinsics", hard_max=None)
+    if cam_points is not None:
+        new_cam_points = check_and_fix_inf_nan(new_cam_points, "new_cam_points", hard_max=None)
+    else:
+        new_cam_points = None
+    new_world_points = check_and_fix_inf_nan(new_world_points, "new_world_points", hard_max=None)
+    new_depths = check_and_fix_inf_nan(new_depths, "new_depths", hard_max=None)
+
+
+    return new_extrinsics, new_cam_points, new_world_points, new_depths
+
+
+
+
+

outdoor_v48_4gpu_v2/code/05_02-14:21:58/vggt/train_utils/normalization_v37.py

@@ -0,0 +1,130 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+import logging
+from typing import Optional, Tuple
+from vggt.utils.geometry import closed_form_inverse_se3
+from vggt.train_utils.general import check_and_fix_inf_nan
+
+
+def check_valid_tensor(input_tensor: Optional[torch.Tensor], name: str = "tensor") -> None:
+    """
+    Check if a tensor contains NaN or Inf values and log a warning if found.
+    
+    Args:
+        input_tensor: The tensor to check
+        name: Name of the tensor for logging purposes
+    """
+    if input_tensor is not None:
+        if torch.isnan(input_tensor).any() or torch.isinf(input_tensor).any():
+            logging.warning(f"NaN or Inf found in tensor: {name}")
+
+
+def normalize_camera_extrinsics_and_points_batch(
+    extrinsics: torch.Tensor,
+    cam_points: Optional[torch.Tensor] = None,
+    world_points: Optional[torch.Tensor] = None,
+    depths: Optional[torch.Tensor] = None,
+    scale_by_points: bool = True,
+    point_masks: Optional[torch.Tensor] = None,
+) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[torch.Tensor], Optional[torch.Tensor]]:
+    """
+    Normalize camera extrinsics and corresponding 3D points.
+    
+    This function transforms the coordinate system to be centered at the first camera
+    and optionally scales the scene to have unit average distance.
+    
+    Args:
+        extrinsics: Camera extrinsic matrices of shape (B, S, 3, 4)
+        cam_points: 3D points in camera coordinates of shape (B, S, H, W, 3) or (*,3)
+        world_points: 3D points in world coordinates of shape (B, S, H, W, 3) or (*,3)
+        depths: Depth maps of shape (B, S, H, W)
+        scale_by_points: Whether to normalize the scale based on point distances
+        point_masks: Boolean masks for valid points of shape (B, S, H, W)
+    
+    Returns:
+        Tuple containing:
+        - Normalized camera extrinsics of shape (B, S, 3, 4)
+        - Normalized camera points (same shape as input cam_points)
+        - Normalized world points (same shape as input world_points)
+        - Normalized depths (same shape as input depths)
+    """
+    # Validate inputs
+    check_valid_tensor(extrinsics, "extrinsics")
+    check_valid_tensor(cam_points, "cam_points")
+    check_valid_tensor(world_points, "world_points")
+    check_valid_tensor(depths, "depths")
+
+
+    B, S, _, _ = extrinsics.shape
+    device = extrinsics.device
+    #assert device == torch.device("cpu")
+
+
+    # Convert extrinsics to homogeneous form: (B, N,4,4)
+    extrinsics_homog = torch.cat(
+        [
+            extrinsics,
+            torch.zeros((B, S, 1, 4), device=device),
+        ],
+        dim=-2,
+    )
+    extrinsics_homog[:, :, -1, -1] = 1.0
+
+    # first_cam_extrinsic_inv, the inverse of the first camera's extrinsic matrix
+    # which can be also viewed as the cam_to_world extrinsic matrix
+    first_cam_extrinsic_inv = closed_form_inverse_se3(extrinsics_homog[:, 0])
+    # new_extrinsics = torch.matmul(extrinsics_homog, first_cam_extrinsic_inv)
+    new_extrinsics = torch.matmul(extrinsics_homog, first_cam_extrinsic_inv.unsqueeze(1))  # (B,N,4,4)
+
+
+    if world_points is not None:
+        # since we are transforming the world points to the first camera's coordinate system
+        # we directly use the cam_from_world extrinsic matrix of the first camera
+        # instead of using the inverse of the first camera's extrinsic matrix
+        R = extrinsics[:, 0, :3, :3]
+        t = extrinsics[:, 0, :3, 3]
+        new_world_points = (world_points @ R.transpose(-1, -2).unsqueeze(1).unsqueeze(2)) + t.unsqueeze(1).unsqueeze(2).unsqueeze(3)
+    else:
+        new_world_points = None
+
+
+    if scale_by_points:
+        new_depths = depths.clone()
+
+        dist = new_world_points.norm(dim=-1)
+        dist_sum = (dist * point_masks).sum(dim=[1,2,3])
+        valid_count = point_masks.sum(dim=[1,2,3])
+        avg_scale = (dist_sum / (valid_count + 1e-3)).clamp(min=1e-6, max=1e6)
+
+
+        new_world_points = new_world_points / avg_scale.view(-1, 1, 1, 1, 1)
+        new_extrinsics[:, :, :3, 3] = new_extrinsics[:, :, :3, 3] / avg_scale.view(-1, 1, 1)
+        if depths is not None:
+            new_depths = new_depths / avg_scale.view(-1, 1, 1, 1)
+        if cam_points is not None:
+            new_cam_points = cam_points.clone()
+            new_cam_points = new_cam_points / avg_scale.view(-1, 1, 1, 1, 1)
+    else:
+        return new_extrinsics[:, :, :3], cam_points, new_world_points, depths
+
+    new_extrinsics = new_extrinsics[:, :, :3] # 4x4 -> 3x4
+    new_extrinsics = check_and_fix_inf_nan(new_extrinsics, "new_extrinsics", hard_max=None)
+    if cam_points is not None:
+        new_cam_points = check_and_fix_inf_nan(new_cam_points, "new_cam_points", hard_max=None)
+    else:
+        new_cam_points = None
+    new_world_points = check_and_fix_inf_nan(new_world_points, "new_world_points", hard_max=None)
+    new_depths = check_and_fix_inf_nan(new_depths, "new_depths", hard_max=None)
+
+
+    return new_extrinsics, new_cam_points, new_world_points, new_depths
+
+
+
+
+

outdoor_v48_4gpu_v2/code/05_02-14:21:58/vggt/utils/geometry.py

@@ -0,0 +1,166 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import os
+import torch
+import numpy as np
+
+
+def unproject_depth_map_to_point_map(
+    depth_map: np.ndarray, extrinsics_cam: np.ndarray, intrinsics_cam: np.ndarray
+) -> np.ndarray:
+    """
+    Unproject a batch of depth maps to 3D world coordinates.
+
+    Args:
+        depth_map (np.ndarray): Batch of depth maps of shape (S, H, W, 1) or (S, H, W)
+        extrinsics_cam (np.ndarray): Batch of camera extrinsic matrices of shape (S, 3, 4)
+        intrinsics_cam (np.ndarray): Batch of camera intrinsic matrices of shape (S, 3, 3)
+
+    Returns:
+        np.ndarray: Batch of 3D world coordinates of shape (S, H, W, 3)
+    """
+    if isinstance(depth_map, torch.Tensor):
+        depth_map = depth_map.cpu().numpy()
+    if isinstance(extrinsics_cam, torch.Tensor):
+        extrinsics_cam = extrinsics_cam.cpu().numpy()
+    if isinstance(intrinsics_cam, torch.Tensor):
+        intrinsics_cam = intrinsics_cam.cpu().numpy()
+
+    world_points_list = []
+    for frame_idx in range(depth_map.shape[0]):
+        cur_world_points, _, _ = depth_to_world_coords_points(
+            depth_map[frame_idx].squeeze(-1), extrinsics_cam[frame_idx], intrinsics_cam[frame_idx]
+        )
+        world_points_list.append(cur_world_points)
+    world_points_array = np.stack(world_points_list, axis=0)
+
+    return world_points_array
+
+
+def depth_to_world_coords_points(
+    depth_map: np.ndarray,
+    extrinsic: np.ndarray,
+    intrinsic: np.ndarray,
+    eps=1e-8,
+) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
+    """
+    Convert a depth map to world coordinates.
+
+    Args:
+        depth_map (np.ndarray): Depth map of shape (H, W).
+        intrinsic (np.ndarray): Camera intrinsic matrix of shape (3, 3).
+        extrinsic (np.ndarray): Camera extrinsic matrix of shape (3, 4). OpenCV camera coordinate convention, cam from world.
+
+    Returns:
+        tuple[np.ndarray, np.ndarray]: World coordinates (H, W, 3) and valid depth mask (H, W).
+    """
+    if depth_map is None:
+        return None, None, None
+
+    # Valid depth mask
+    point_mask = depth_map > eps
+
+    # Convert depth map to camera coordinates
+    cam_coords_points = depth_to_cam_coords_points(depth_map, intrinsic)
+
+    # Multiply with the inverse of extrinsic matrix to transform to world coordinates
+    # extrinsic_inv is 4x4 (note closed_form_inverse_OpenCV is batched, the output is (N, 4, 4))
+    cam_to_world_extrinsic = closed_form_inverse_se3(extrinsic[None])[0]
+
+    R_cam_to_world = cam_to_world_extrinsic[:3, :3]
+    t_cam_to_world = cam_to_world_extrinsic[:3, 3]
+
+    # Apply the rotation and translation to the camera coordinates
+    world_coords_points = np.dot(cam_coords_points, R_cam_to_world.T) + t_cam_to_world  # HxWx3, 3x3 -> HxWx3
+    # world_coords_points = np.einsum("ij,hwj->hwi", R_cam_to_world, cam_coords_points) + t_cam_to_world
+
+    return world_coords_points, cam_coords_points, point_mask
+
+
+def depth_to_cam_coords_points(depth_map: np.ndarray, intrinsic: np.ndarray) -> tuple[np.ndarray, np.ndarray]:
+    """
+    Convert a depth map to camera coordinates.
+
+    Args:
+        depth_map (np.ndarray): Depth map of shape (H, W).
+        intrinsic (np.ndarray): Camera intrinsic matrix of shape (3, 3).
+
+    Returns:
+        tuple[np.ndarray, np.ndarray]: Camera coordinates (H, W, 3)
+    """
+    H, W = depth_map.shape
+    assert intrinsic.shape == (3, 3), "Intrinsic matrix must be 3x3"
+    assert intrinsic[0, 1] == 0 and intrinsic[1, 0] == 0, "Intrinsic matrix must have zero skew"
+
+    # Intrinsic parameters
+    fu, fv = intrinsic[0, 0], intrinsic[1, 1]
+    cu, cv = intrinsic[0, 2], intrinsic[1, 2]
+
+    # Generate grid of pixel coordinates
+    u, v = np.meshgrid(np.arange(W), np.arange(H))
+
+    # Unproject to camera coordinates
+    x_cam = (u - cu) * depth_map / fu
+    y_cam = (v - cv) * depth_map / fv
+    z_cam = depth_map
+
+    # Stack to form camera coordinates
+    cam_coords = np.stack((x_cam, y_cam, z_cam), axis=-1).astype(np.float32)
+
+    return cam_coords
+
+
+def closed_form_inverse_se3(se3, R=None, T=None):
+    """
+    Compute the inverse of each 4x4 (or 3x4) SE3 matrix in a batch.
+
+    If `R` and `T` are provided, they must correspond to the rotation and translation
+    components of `se3`. Otherwise, they will be extracted from `se3`.
+
+    Args:
+        se3: Nx4x4 or Nx3x4 array or tensor of SE3 matrices.
+        R (optional): Nx3x3 array or tensor of rotation matrices.
+        T (optional): Nx3x1 array or tensor of translation vectors.
+
+    Returns:
+        Inverted SE3 matrices with the same type and device as `se3`.
+
+    Shapes:
+        se3: (N, 4, 4)
+        R: (N, 3, 3)
+        T: (N, 3, 1)
+    """
+    # Check if se3 is a numpy array or a torch tensor
+    is_numpy = isinstance(se3, np.ndarray)
+
+    # Validate shapes
+    if se3.shape[-2:] != (4, 4) and se3.shape[-2:] != (3, 4):
+        raise ValueError(f"se3 must be of shape (N,4,4), got {se3.shape}.")
+
+    # Extract R and T if not provided
+    if R is None:
+        R = se3[:, :3, :3]  # (N,3,3)
+    if T is None:
+        T = se3[:, :3, 3:]  # (N,3,1)
+
+    # Transpose R
+    if is_numpy:
+        # Compute the transpose of the rotation for NumPy
+        R_transposed = np.transpose(R, (0, 2, 1))
+        # -R^T t for NumPy
+        top_right = -np.matmul(R_transposed, T)
+        inverted_matrix = np.tile(np.eye(4), (len(R), 1, 1))
+    else:
+        R_transposed = R.transpose(1, 2)  # (N,3,3)
+        top_right = -torch.bmm(R_transposed, T)  # (N,3,1)
+        inverted_matrix = torch.eye(4, 4)[None].repeat(len(R), 1, 1)
+        inverted_matrix = inverted_matrix.to(R.dtype).to(R.device)
+
+    inverted_matrix[:, :3, :3] = R_transposed
+    inverted_matrix[:, :3, 3:] = top_right
+
+    return inverted_matrix

outdoor_v48_4gpu_v2/code/05_02-14:21:58/vggt/utils/load_fn.py

@@ -0,0 +1,147 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+from PIL import Image
+from torchvision import transforms as TF
+
+
+def load_and_preprocess_images(image_path_list, mode="crop"):
+    """
+    A quick start function to load and preprocess images for model input.
+    This assumes the images should have the same shape for easier batching, but our model can also work well with different shapes.
+
+    Args:
+        image_path_list (list): List of paths to image files
+        mode (str, optional): Preprocessing mode, either "crop" or "pad".
+                             - "crop" (default): Sets width to 518px and center crops height if needed.
+                             - "pad": Preserves all pixels by making the largest dimension 518px
+                               and padding the smaller dimension to reach a square shape.
+
+    Returns:
+        torch.Tensor: Batched tensor of preprocessed images with shape (N, 3, H, W)
+
+    Raises:
+        ValueError: If the input list is empty or if mode is invalid
+
+    Notes:
+        - Images with different dimensions will be padded with white (value=1.0)
+        - A warning is printed when images have different shapes
+        - When mode="crop": The function ensures width=518px while maintaining aspect ratio
+          and height is center-cropped if larger than 518px
+        - When mode="pad": The function ensures the largest dimension is 518px while maintaining aspect ratio
+          and the smaller dimension is padded to reach a square shape (518x518)
+        - Dimensions are adjusted to be divisible by 14 for compatibility with model requirements
+    """
+    # Check for empty list
+    if len(image_path_list) == 0:
+        raise ValueError("At least 1 image is required")
+
+    # Validate mode
+    if mode not in ["crop", "pad"]:
+        raise ValueError("Mode must be either 'crop' or 'pad'")
+
+    images = []
+    shapes = set()
+    to_tensor = TF.ToTensor()
+    target_size = 224
+
+    # First process all images and collect their shapes
+    for image_path in image_path_list:
+
+        # Open image
+        img = Image.open(image_path)
+
+        # If there's an alpha channel, blend onto white background:
+        if img.mode == "RGBA":
+            # Create white background
+            background = Image.new("RGBA", img.size, (255, 255, 255, 255))
+            # Alpha composite onto the white background
+            img = Image.alpha_composite(background, img)
+
+        # Now convert to "RGB" (this step assigns white for transparent areas)
+        img = img.convert("RGB")
+
+        width, height = img.size
+
+        if mode == "pad":
+            # Make the largest dimension 518px while maintaining aspect ratio
+            if width >= height:
+                new_width = target_size
+                new_height = round(height * (new_width / width) / 14) * 14  # Make divisible by 14
+            else:
+                new_height = target_size
+                new_width = round(width * (new_height / height) / 14) * 14  # Make divisible by 14
+        else:  # mode == "crop"
+            # Original behavior: set width to 518px
+            new_width = target_size
+            # Calcu late height maintaining aspect ratio, divisible by 14
+            # new_height = round(height * (new_width / width) / 14) * 14
+            new_height = target_size
+
+        # Resize with new dimensions (width, height)
+        img = img.resize((new_width, new_height), Image.Resampling.BICUBIC)
+        img = to_tensor(img)  # Convert to tensor (0, 1)
+
+        # Center crop height if it's larger than 518 (only in crop mode)
+        if mode == "crop" and new_height > target_size:
+            start_y = (new_height - target_size) // 2
+            img = img[:, start_y: start_y + target_size, :]
+
+        # For pad mode, pad to make a square of target_size x target_size
+        if mode == "pad":
+            h_padding = target_size - img.shape[1]
+            w_padding = target_size - img.shape[2]
+
+            if h_padding > 0 or w_padding > 0:
+                pad_top = h_padding // 2
+                pad_bottom = h_padding - pad_top
+                pad_left = w_padding // 2
+                pad_right = w_padding - pad_left
+
+                # Pad with white (value=1.0)
+                img = torch.nn.functional.pad(
+                    img, (pad_left, pad_right, pad_top, pad_bottom), mode="constant", value=1.0
+                )
+
+        shapes.add((img.shape[1], img.shape[2]))
+        images.append(img)
+
+    # Check if we have different shapes
+    # In theory our model can also work well with different shapes
+    if len(shapes) > 1:
+        print(f"Warning: Found images with different shapes: {shapes}")
+        # Find maximum dimensions
+        max_height = max(shape[0] for shape in shapes)
+        max_width = max(shape[1] for shape in shapes)
+
+        # Pad images if necessary
+        padded_images = []
+        for img in images:
+            h_padding = max_height - img.shape[1]
+            w_padding = max_width - img.shape[2]
+
+            if h_padding > 0 or w_padding > 0:
+                pad_top = h_padding // 2
+                pad_bottom = h_padding - pad_top
+                pad_left = w_padding // 2
+                pad_right = w_padding - pad_left
+
+                img = torch.nn.functional.pad(
+                    img, (pad_left, pad_right, pad_top, pad_bottom), mode="constant", value=1.0
+                )
+            padded_images.append(img)
+        images = padded_images
+
+    images = torch.stack(images)  # concatenate images
+
+    # Ensure correct shape when single image
+    if len(image_path_list) == 1:
+        # Verify shape is (1, C, H, W)
+        if images.dim() == 3:
+            images = images.unsqueeze(0)
+
+    return images

outdoor_v48_4gpu_v2/code/05_02-14:21:58/vggt/utils/pose_enc.py

@@ -0,0 +1,130 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+from .rotation import quat_to_mat, mat_to_quat
+
+
+def extri_intri_to_pose_encoding(
+    extrinsics,
+    intrinsics,
+    image_size_hw=None,  # e.g., (256, 512)
+    pose_encoding_type="absT_quaR_FoV",
+):
+    """Convert camera extrinsics and intrinsics to a compact pose encoding.
+
+    This function transforms camera parameters into a unified pose encoding format,
+    which can be used for various downstream tasks like pose prediction or representation.
+
+    Args:
+        extrinsics (torch.Tensor): Camera extrinsic parameters with shape BxSx3x4,
+            where B is batch size and S is sequence length.
+            In OpenCV coordinate system (x-right, y-down, z-forward), representing camera from world transformation.
+            The format is [R|t] where R is a 3x3 rotation matrix and t is a 3x1 translation vector.
+        intrinsics (torch.Tensor): Camera intrinsic parameters with shape BxSx3x3.
+            Defined in pixels, with format:
+            [[fx, 0, cx],
+             [0, fy, cy],
+             [0,  0,  1]]
+            where fx, fy are focal lengths and (cx, cy) is the principal point
+        image_size_hw (tuple): Tuple of (height, width) of the image in pixels.
+            Required for computing field of view values. For example: (256, 512).
+        pose_encoding_type (str): Type of pose encoding to use. Currently only
+            supports "absT_quaR_FoV" (absolute translation, quaternion rotation, field of view).
+
+    Returns:
+        torch.Tensor: Encoded camera pose parameters with shape BxSx9.
+            For "absT_quaR_FoV" type, the 9 dimensions are:
+            - [:3] = absolute translation vector T (3D)
+            - [3:7] = rotation as quaternion quat (4D)
+            - [7:] = field of view (2D)
+    """
+
+    # extrinsics: BxSx3x4
+    # intrinsics: BxSx3x3
+
+    if pose_encoding_type == "absT_quaR_FoV":
+        R = extrinsics[:, :, :3, :3]  # BxSx3x3
+        T = extrinsics[:, :, :3, 3]  # BxSx3
+
+        quat = mat_to_quat(R)
+        # Note the order of h and w here
+        H, W = image_size_hw
+        fov_h = 2 * torch.atan((H / 2) / intrinsics[..., 1, 1])
+        fov_w = 2 * torch.atan((W / 2) / intrinsics[..., 0, 0])
+        pose_encoding = torch.cat([T, quat, fov_h[..., None], fov_w[..., None]], dim=-1).float()
+    else:
+        raise NotImplementedError
+
+    return pose_encoding
+
+
+def pose_encoding_to_extri_intri(
+    pose_encoding,
+    image_size_hw=None,  # e.g., (256, 512)
+    pose_encoding_type="absT_quaR_FoV",
+    build_intrinsics=True,
+):
+    """Convert a pose encoding back to camera extrinsics and intrinsics.
+
+    This function performs the inverse operation of extri_intri_to_pose_encoding,
+    reconstructing the full camera parameters from the compact encoding.
+
+    Args:
+        pose_encoding (torch.Tensor): Encoded camera pose parameters with shape BxSx9,
+            where B is batch size and S is sequence length.
+            For "absT_quaR_FoV" type, the 9 dimensions are:
+            - [:3] = absolute translation vector T (3D)
+            - [3:7] = rotation as quaternion quat (4D)
+            - [7:] = field of view (2D)
+        image_size_hw (tuple): Tuple of (height, width) of the image in pixels.
+            Required for reconstructing intrinsics from field of view values.
+            For example: (256, 512).
+        pose_encoding_type (str): Type of pose encoding used. Currently only
+            supports "absT_quaR_FoV" (absolute translation, quaternion rotation, field of view).
+        build_intrinsics (bool): Whether to reconstruct the intrinsics matrix.
+            If False, only extrinsics are returned and intrinsics will be None.
+
+    Returns:
+        tuple: (extrinsics, intrinsics)
+            - extrinsics (torch.Tensor): Camera extrinsic parameters with shape BxSx3x4.
+              In OpenCV coordinate system (x-right, y-down, z-forward), representing camera from world
+              transformation. The format is [R|t] where R is a 3x3 rotation matrix and t is
+              a 3x1 translation vector.
+            - intrinsics (torch.Tensor or None): Camera intrinsic parameters with shape BxSx3x3,
+              or None if build_intrinsics is False. Defined in pixels, with format:
+              [[fx, 0, cx],
+               [0, fy, cy],
+               [0,  0,  1]]
+              where fx, fy are focal lengths and (cx, cy) is the principal point,
+              assumed to be at the center of the image (W/2, H/2).
+    """
+
+    intrinsics = None
+
+    if pose_encoding_type == "absT_quaR_FoV":
+        T = pose_encoding[..., :3]
+        quat = pose_encoding[..., 3:7]
+        fov_h = pose_encoding[..., 7]
+        fov_w = pose_encoding[..., 8]
+
+        R = quat_to_mat(quat)
+        extrinsics = torch.cat([R, T[..., None]], dim=-1)
+
+        if build_intrinsics:
+            H, W = image_size_hw
+            fy = (H / 2.0) / torch.tan(fov_h / 2.0)
+            fx = (W / 2.0) / torch.tan(fov_w / 2.0)
+            intrinsics = torch.zeros(pose_encoding.shape[:2] + (3, 3), device=pose_encoding.device)
+            intrinsics[..., 0, 0] = fx
+            intrinsics[..., 1, 1] = fy
+            intrinsics[..., 0, 2] = W / 2
+            intrinsics[..., 1, 2] = H / 2
+            intrinsics[..., 2, 2] = 1.0  # Set the homogeneous coordinate to 1
+    else:
+        raise NotImplementedError
+
+    return extrinsics, intrinsics

outdoor_v48_4gpu_v2/code/05_02-14:21:58/vggt/utils/visual_track.py

@@ -0,0 +1,239 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import cv2
+import torch
+import numpy as np
+import os
+
+
+def color_from_xy(x, y, W, H, cmap_name="hsv"):
+    """
+    Map (x, y) -> color in (R, G, B).
+    1) Normalize x,y to [0,1].
+    2) Combine them into a single scalar c in [0,1].
+    3) Use matplotlib's colormap to convert c -> (R,G,B).
+
+    You can customize step 2, e.g., c = (x + y)/2, or some function of (x, y).
+    """
+    import matplotlib.cm
+    import matplotlib.colors
+
+    x_norm = x / max(W - 1, 1)
+    y_norm = y / max(H - 1, 1)
+    # Simple combination:
+    c = (x_norm + y_norm) / 2.0
+
+    cmap = matplotlib.cm.get_cmap(cmap_name)
+    # cmap(c) -> (r,g,b,a) in [0,1]
+    rgba = cmap(c)
+    r, g, b = rgba[0], rgba[1], rgba[2]
+    return (r, g, b)  # in [0,1], RGB order
+
+
+def get_track_colors_by_position(tracks_b, vis_mask_b=None, image_width=None, image_height=None, cmap_name="hsv"):
+    """
+    Given all tracks in one sample (b), compute a (N,3) array of RGB color values
+    in [0,255]. The color is determined by the (x,y) position in the first
+    visible frame for each track.
+
+    Args:
+        tracks_b: Tensor of shape (S, N, 2). (x,y) for each track in each frame.
+        vis_mask_b: (S, N) boolean mask; if None, assume all are visible.
+        image_width, image_height: used for normalizing (x, y).
+        cmap_name: for matplotlib (e.g., 'hsv', 'rainbow', 'jet').
+
+    Returns:
+        track_colors: np.ndarray of shape (N, 3), each row is (R,G,B) in [0,255].
+    """
+    S, N, _ = tracks_b.shape
+    track_colors = np.zeros((N, 3), dtype=np.uint8)
+
+    if vis_mask_b is None:
+        # treat all as visible
+        vis_mask_b = torch.ones(S, N, dtype=torch.bool, device=tracks_b.device)
+
+    for i in range(N):
+        # Find first visible frame for track i
+        visible_frames = torch.where(vis_mask_b[:, i])[0]
+        if len(visible_frames) == 0:
+            # track is never visible; just assign black or something
+            track_colors[i] = (0, 0, 0)
+            continue
+
+        first_s = int(visible_frames[0].item())
+        # use that frame's (x,y)
+        x, y = tracks_b[first_s, i].tolist()
+
+        # map (x,y) -> (R,G,B) in [0,1]
+        r, g, b = color_from_xy(x, y, W=image_width, H=image_height, cmap_name=cmap_name)
+        # scale to [0,255]
+        r, g, b = int(r * 255), int(g * 255), int(b * 255)
+        track_colors[i] = (r, g, b)
+
+    return track_colors
+
+
+def visualize_tracks_on_images(
+    images,
+    tracks,
+    track_vis_mask=None,
+    out_dir="track_visuals_concat_by_xy",
+    image_format="CHW",  # "CHW" or "HWC"
+    normalize_mode="[0,1]",
+    cmap_name="hsv",  # e.g. "hsv", "rainbow", "jet"
+    frames_per_row=4,  # New parameter for grid layout
+    save_grid=True,  # Flag to control whether to save the grid image
+):
+    """
+    Visualizes frames in a grid layout with specified frames per row.
+    Each track's color is determined by its (x,y) position
+    in the first visible frame (or frame 0 if always visible).
+    Finally convert the BGR result to RGB before saving.
+    Also saves each individual frame as a separate PNG file.
+
+    Args:
+        images: torch.Tensor (S, 3, H, W) if CHW or (S, H, W, 3) if HWC.
+        tracks: torch.Tensor (S, N, 2), last dim = (x, y).
+        track_vis_mask: torch.Tensor (S, N) or None.
+        out_dir: folder to save visualizations.
+        image_format: "CHW" or "HWC".
+        normalize_mode: "[0,1]", "[-1,1]", or None for direct raw -> 0..255
+        cmap_name: a matplotlib colormap name for color_from_xy.
+        frames_per_row: number of frames to display in each row of the grid.
+        save_grid: whether to save all frames in one grid image.
+
+    Returns:
+        None (saves images in out_dir).
+    """
+
+    if len(tracks.shape) == 4:
+        tracks = tracks.squeeze(0)
+        images = images.squeeze(0)
+        if track_vis_mask is not None:
+            track_vis_mask = track_vis_mask.squeeze(0)
+
+    import matplotlib
+
+    matplotlib.use("Agg")  # for non-interactive (optional)
+
+    os.makedirs(out_dir, exist_ok=True)
+
+    S = images.shape[0]
+    _, N, _ = tracks.shape  # (S, N, 2)
+
+    # Move to CPU
+    images = images.cpu().clone()
+    tracks = tracks.cpu().clone()
+    if track_vis_mask is not None:
+        track_vis_mask = track_vis_mask.cpu().clone()
+
+    # Infer H, W from images shape
+    if image_format == "CHW":
+        # e.g. images[s].shape = (3, H, W)
+        H, W = images.shape[2], images.shape[3]
+    else:
+        # e.g. images[s].shape = (H, W, 3)
+        H, W = images.shape[1], images.shape[2]
+
+    # Pre-compute the color for each track i based on first visible position
+    track_colors_rgb = get_track_colors_by_position(
+        tracks,  # shape (S, N, 2)
+        vis_mask_b=track_vis_mask if track_vis_mask is not None else None,
+        image_width=W,
+        image_height=H,
+        cmap_name=cmap_name,
+    )
+
+    # We'll accumulate each frame's drawn image in a list
+    frame_images = []
+
+    for s in range(S):
+        # shape => either (3, H, W) or (H, W, 3)
+        img = images[s]
+
+        # Convert to (H, W, 3)
+        if image_format == "CHW":
+            img = img.permute(1, 2, 0)  # (H, W, 3)
+        # else "HWC", do nothing
+
+        img = img.numpy().astype(np.float32)
+
+        # Scale to [0,255] if needed
+        if normalize_mode == "[0,1]":
+            img = np.clip(img, 0, 1) * 255.0
+        elif normalize_mode == "[-1,1]":
+            img = (img + 1.0) * 0.5 * 255.0
+            img = np.clip(img, 0, 255.0)
+        # else no normalization
+
+        # Convert to uint8
+        img = img.astype(np.uint8)
+
+        # For drawing in OpenCV, convert to BGR
+        img_bgr = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)
+
+        # Draw each visible track
+        cur_tracks = tracks[s]  # shape (N, 2)
+        if track_vis_mask is not None:
+            valid_indices = torch.where(track_vis_mask[s])[0]
+        else:
+            valid_indices = range(N)
+
+        cur_tracks_np = cur_tracks.numpy()
+        for i in valid_indices:
+            x, y = cur_tracks_np[i]
+            pt = (int(round(x)), int(round(y)))
+
+            # track_colors_rgb[i] is (R,G,B). For OpenCV circle, we need BGR
+            R, G, B = track_colors_rgb[i]
+            color_bgr = (int(B), int(G), int(R))
+            cv2.circle(img_bgr, pt, radius=3, color=color_bgr, thickness=-1)
+
+        # Convert back to RGB for consistent final saving:
+        img_rgb = cv2.cvtColor(img_bgr, cv2.COLOR_BGR2RGB)
+
+        # Save individual frame
+        frame_path = os.path.join(out_dir, f"frame_{s:04d}.png")
+        # Convert to BGR for OpenCV imwrite
+        frame_bgr = cv2.cvtColor(img_rgb, cv2.COLOR_RGB2BGR)
+        cv2.imwrite(frame_path, frame_bgr)
+
+        frame_images.append(img_rgb)
+
+    # Only create and save the grid image if save_grid is True
+    if save_grid:
+        # Calculate grid dimensions
+        num_rows = (S + frames_per_row - 1) // frames_per_row  # Ceiling division
+
+        # Create a grid of images
+        grid_img = None
+        for row in range(num_rows):
+            start_idx = row * frames_per_row
+            end_idx = min(start_idx + frames_per_row, S)
+
+            # Concatenate this row horizontally
+            row_img = np.concatenate(frame_images[start_idx:end_idx], axis=1)
+
+            # If this row has fewer than frames_per_row images, pad with black
+            if end_idx - start_idx < frames_per_row:
+                padding_width = (frames_per_row - (end_idx - start_idx)) * W
+                padding = np.zeros((H, padding_width, 3), dtype=np.uint8)
+                row_img = np.concatenate([row_img, padding], axis=1)
+
+            # Add this row to the grid
+            if grid_img is None:
+                grid_img = row_img
+            else:
+                grid_img = np.concatenate([grid_img, row_img], axis=0)
+
+        out_path = os.path.join(out_dir, "tracks_grid.png")
+        # Convert back to BGR for OpenCV imwrite
+        grid_img_bgr = cv2.cvtColor(grid_img, cv2.COLOR_RGB2BGR)
+        cv2.imwrite(out_path, grid_img_bgr)
+        print(f"[INFO] Saved color-by-XY track visualization grid -> {out_path}")
+
+    print(f"[INFO] Saved {S} individual frames to {out_dir}/frame_*.png")