Back to Details

tao-train-sparse4d

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Sparse4D

Sparse4D

Sparse4D for multi-camera temporal 3D object detection and tracking. Uses sparse queries with deformable attention across camera views and time for end-to-end 3D perception. Includes instance bank for temporal tracking.
Requires pretrained ResNet-101 backbone. Set train.pretrained_model_path.
Sparse4D用于多相机时序3D目标检测与跟踪。它采用稀疏查询结合跨相机视角与时序的可变形注意力机制，实现端到端3D感知，并内置实例库以支持时序跟踪。
需要预训练的ResNet-101骨干网络。请设置train.pretrained_model_path参数。

Dataclass Schemas

数据类模式

Generated TAO Core schemas are packaged in 
schemas/<action>.schema.json
, with 
schemas/manifest.json
listing available actions. Each generated schema also emits 
references/spec_template_<action>.yaml
from the schema top-level 
default
field. AutoML enablement is declared at the model layer in 
references/skill_info.yaml
via 
automl_enabled
. Runnable AutoML still requires 
schemas/train.schema.json
and 
references/spec_template_train.yaml
to exist and parse. Use the packaged train schema for 
automl_default_parameters
, 
automl_disabled_parameters
, defaults, min/max bounds, enums, option weights, math conditions, dependencies, and popular parameters. Do not expect 
~/tao-core
at runtime; maintainers regenerate schemas/templates before packaging the skill bank.
生成的TAO Core模式打包在
schemas/<action>.schema.json
中，
schemas/manifest.json
会列出所有可用动作。每个生成的模式还会从模式顶层的
default
字段生成
references/spec_template_<action>.yaml
文件。AutoML支持在
references/skill_info.yaml
的模型层通过
automl_enabled
声明。可运行的AutoML仍要求
schemas/train.schema.json
和
references/spec_template_train.yaml
存在且可解析。使用打包的训练模式来配置
automl_default_parameters
、
automl_disabled_parameters
、默认值、最小/最大边界、枚举值、选项权重、数学条件、依赖关系以及常用参数。运行时不要依赖
~/tao-core
；维护人员会在打包技能库前重新生成模式/模板。

Train Action Policy

训练动作策略

This model is AutoML-enabled at the model layer. Before handling any train-stage request, read 
references/skill_info.yaml
and resolve the run override from either an explicit 
automl_policy
value or the user's workflow request. Treat phrases like "turn off AutoML", "disable AutoML", "no HPO", or "plain training" as 
automl_policy: off
for this run only; otherwise default to 
auto
. When 
automl_policy: auto
, 
automl_enabled: true
, and both 
schemas/train.schema.json
and 
references/spec_template_train.yaml
are packaged, route the train action through 
tao-skill-bank:tao-run-automl
by default with this model's 
skill_dir
. Preserve workflow/application overrides for datasets, specs, output directories, GPU/platform settings, parent checkpoints, and 
automl_policy
. Use direct model training only when 
automl_policy: off
or the packaged train schema/template is missing; in the missing-schema case, report that AutoML is enabled but not runnable for this model until schemas are generated.
Non-train actions such as 
evaluate
, 
inference
, 
export
, and deploy flows stay in this model skill. The per-run 
automl_policy
override does not change model metadata.
该模型在模型层支持AutoML。处理任何训练阶段请求前，请读取
references/skill_info.yaml
，并通过显式的
automl_policy
值或用户的工作流请求确定运行覆盖配置。将“turn off AutoML”、“disable AutoML”、“no HPO”或“plain training”这类短语视为本次运行的
automl_policy: off
；否则默认使用
auto
。当
automl_policy: auto
、
automl_enabled: true
，且
schemas/train.schema.json
和
references/spec_template_train.yaml
已打包时，默认将训练动作通过
tao-skill-bank:tao-run-automl
路由，并传入该模型的
skill_dir
。保留数据集、规格、输出目录、GPU/平台设置、父检查点和
automl_policy
的工作流/应用覆盖配置。仅当
automl_policy: off
或打包的训练模式/模板缺失时，才使用直接模型训练；若模式缺失，需报告该模型已启用AutoML但无法运行，直到生成对应的模式。
非训练动作（如
evaluate
、
inference
、
export
以及部署流程）仍在该模型技能中处理。单次运行的
automl_policy
覆盖配置不会改变模型元数据。

Training Requirements

训练要求

  • Dataset type: sparse4d
  • Formats: ovpkl
  • Monitoring metric: val_mAP
  • 数据集类型： sparse4d
  • 格式： ovpkl
  • 监控指标： val_mAP

Per-Action Dataset Requirements

各动作的数据集要求

ActionSpec KeySourceFilesList?
dataset_convertaicity.rootidNo
evaluatedataset.data_rooteval_dataset(from convert job, spec: aicity.split)No
evaluatemodel.head.instance_bank.anchortrain_datasets/results/{dataset_convert_job_id}/anchor_init.npyNo
evaluatedataset.train_dataset.ann_filetrain_datasets(from convert job, spec: aicity.split)No
evaluatedataset.val_dataset.ann_fileeval_dataset(from convert job, spec: aicity.split)No
evaluatedataset.test_dataset.ann_fileinference_dataset(from convert job, spec: aicity.split)No
exportmodel.head.instance_bank.anchortrain_datasets/results/{dataset_convert_job_id}/anchor_init.npyNo
inferencedataset.data_rootinference_dataset(from convert job, spec: aicity.split)No
inferencemodel.head.instance_bank.anchortrain_datasets/results/{dataset_convert_job_id}/anchor_init.npyNo
inferencedataset.train_dataset.ann_filetrain_datasets(from convert job, spec: aicity.split)No
inferencedataset.val_dataset.ann_fileeval_dataset(from convert job, spec: aicity.split)No
inferencedataset.test_dataset.ann_fileinference_dataset(from convert job, spec: aicity.split)No
quantizedataset.data_roottrain_datasets(from convert job, spec: aicity.split)No
quantizemodel.head.instance_bank.anchortrain_datasets/results/{dataset_convert_job_id}/anchor_init.npyNo
quantizedataset.train_dataset.ann_filetrain_datasets(from convert job, spec: aicity.split)No
quantizedataset.val_dataset.ann_fileeval_dataset(from convert job, spec: aicity.split)No
quantizedataset.test_dataset.ann_fileinference_dataset(from convert job, spec: aicity.split)No
quantizedataset.quant_calibration_dataset.images_dirtrain_datasetsNo
traindataset.data_roottrain_datasets(from convert job, spec: aicity.split)No
trainmodel.head.instance_bank.anchortrain_datasets/results/{dataset_convert_job_id}/anchor_init.npyNo
traindataset.train_dataset.ann_filetrain_datasets(from convert job, spec: aicity.split)No
traindataset.val_dataset.ann_fileeval_dataset(from convert job, spec: aicity.split)No
traindataset.test_dataset.ann_fileinference_dataset(from convert job, spec: aicity.split)No
动作规格键来源文件是否为列表？
dataset_convertaicity.rootid否
evaluatedataset.data_rooteval_dataset（来自转换任务，规格：aicity.split）否
evaluatemodel.head.instance_bank.anchortrain_datasets/results/{dataset_convert_job_id}/anchor_init.npy否
evaluatedataset.train_dataset.ann_filetrain_datasets（来自转换任务，规格：aicity.split）否
evaluatedataset.val_dataset.ann_fileeval_dataset（来自转换任务，规格：aicity.split）否
evaluatedataset.test_dataset.ann_fileinference_dataset（来自转换任务，规格：aicity.split）否
exportmodel.head.instance_bank.anchortrain_datasets/results/{dataset_convert_job_id}/anchor_init.npy否
inferencedataset.data_rootinference_dataset（来自转换任务，规格：aicity.split）否
inferencemodel.head.instance_bank.anchortrain_datasets/results/{dataset_convert_job_id}/anchor_init.npy否
inferencedataset.train_dataset.ann_filetrain_datasets（来自转换任务，规格：aicity.split）否
inferencedataset.val_dataset.ann_fileeval_dataset（来自转换任务，规格：aicity.split）否
inferencedataset.test_dataset.ann_fileinference_dataset（来自转换任务，规格：aicity.split）否
quantizedataset.data_roottrain_datasets（来自转换任务，规格：aicity.split）否
quantizemodel.head.instance_bank.anchortrain_datasets/results/{dataset_convert_job_id}/anchor_init.npy否
quantizedataset.train_dataset.ann_filetrain_datasets（来自转换任务，规格：aicity.split）否
quantizedataset.val_dataset.ann_fileeval_dataset（来自转换任务，规格：aicity.split）否
quantizedataset.test_dataset.ann_fileinference_dataset（来自转换任务，规格：aicity.split）否
quantizedataset.quant_calibration_dataset.images_dirtrain_datasets否
traindataset.data_roottrain_datasets（来自转换任务，规格：aicity.split）否
trainmodel.head.instance_bank.anchortrain_datasets/results/{dataset_convert_job_id}/anchor_init.npy否
traindataset.train_dataset.ann_filetrain_datasets（来自转换任务，规格：aicity.split）否
traindataset.val_dataset.ann_fileeval_dataset（来自转换任务，规格：aicity.split）否
traindataset.test_dataset.ann_fileinference_dataset（来自转换任务，规格：aicity.split）否

Typical Spec Overrides

典型规格覆盖配置

Data source overrides are mandatory for every action — the agent MUST construct data source paths from the Per-Action Dataset Requirements table above and include them in 
spec_overrides
.
python
S3_TRAIN = "s3://bucket/data/train"
S3_EVAL = "s3://bucket/data/eval"
train (mandatory data sources):
python
{
    "train.num_epochs": 30,
    "train.checkpoint_interval": 10,
    "train.validation_interval": 10,
    "train.num_gpus": 1,
    "dataset.sequences.split_num": 90,
    "train_dataset.sequences_split_num": 90,
    "dataset.data_root": {"spec": f"{S3_TRAIN}/aicity.split)"},
    "model.head.instance_bank.anchor": f"{S3_TRAIN}//results/{dataset_convert_job_id}/anchor_init.npy",
    "dataset.train_dataset.ann_file": {"spec": f"{S3_TRAIN}/aicity.split)"},
    "dataset.val_dataset.ann_file": {"spec": f"{S3_EVAL}/aicity.split)"},
    "dataset.test_dataset.ann_file": {"spec": f"{S3_EVAL}/aicity.split)"},
}
evaluate (mandatory data sources):
python
{
    "dataset.data_root": {"spec": f"{S3_EVAL}/aicity.split)"},
    "model.head.instance_bank.anchor": f"{S3_TRAIN}//results/{dataset_convert_job_id}/anchor_init.npy",
    "dataset.train_dataset.ann_file": {"spec": f"{S3_TRAIN}/aicity.split)"},
    "dataset.val_dataset.ann_file": {"spec": f"{S3_EVAL}/aicity.split)"},
    "dataset.test_dataset.ann_file": {"spec": f"{S3_EVAL}/aicity.split)"},
}
export (mandatory data sources):
python
{
    "model.head.instance_bank.anchor": f"{S3_TRAIN}//results/{dataset_convert_job_id}/anchor_init.npy",
}
inference (mandatory data sources):
python
{
    "dataset.data_root": {"spec": f"{S3_EVAL}/aicity.split)"},
    "model.head.instance_bank.anchor": f"{S3_TRAIN}//results/{dataset_convert_job_id}/anchor_init.npy",
    "dataset.train_dataset.ann_file": {"spec": f"{S3_TRAIN}/aicity.split)"},
    "dataset.val_dataset.ann_file": {"spec": f"{S3_EVAL}/aicity.split)"},
    "dataset.test_dataset.ann_file": {"spec": f"{S3_EVAL}/aicity.split)"},
}
quantize (mandatory data sources):
python
{
    "dataset.data_root": {"spec": f"{S3_TRAIN}/aicity.split)"},
    "model.head.instance_bank.anchor": f"{S3_TRAIN}//results/{dataset_convert_job_id}/anchor_init.npy",
    "dataset.train_dataset.ann_file": {"spec": f"{S3_TRAIN}/aicity.split)"},
    "dataset.val_dataset.ann_file": {"spec": f"{S3_EVAL}/aicity.split)"},
    "dataset.test_dataset.ann_file": {"spec": f"{S3_EVAL}/aicity.split)"},
    "dataset.quant_calibration_dataset.images_dir": f"{S3_TRAIN}",
}
数据源覆盖配置对每个动作都是必填项——Agent必须根据上述“各动作的数据集要求”表格构建数据源路径，并将其包含在
spec_overrides
中。
python
S3_TRAIN = "s3://bucket/data/train"
S3_EVAL = "s3://bucket/data/eval"
训练（必填数据源）：
python
{
    "train.num_epochs": 30,
    "train.checkpoint_interval": 10,
    "train.validation_interval": 10,
    "train.num_gpus": 1,
    "dataset.sequences.split_num": 90,
    "train_dataset.sequences_split_num": 90,
    "dataset.data_root": {"spec": f"{S3_TRAIN}/aicity.split)"},
    "model.head.instance_bank.anchor": f"{S3_TRAIN}//results/{dataset_convert_job_id}/anchor_init.npy",
    "dataset.train_dataset.ann_file": {"spec": f"{S3_TRAIN}/aicity.split)"},
    "dataset.val_dataset.ann_file": {"spec": f"{S3_EVAL}/aicity.split)"},
    "dataset.test_dataset.ann_file": {"spec": f"{S3_EVAL}/aicity.split)"},
}
评估（必填数据源）：
python
{
    "dataset.data_root": {"spec": f"{S3_EVAL}/aicity.split)"},
    "model.head.instance_bank.anchor": f"{S3_TRAIN}//results/{dataset_convert_job_id}/anchor_init.npy",
    "dataset.train_dataset.ann_file": {"spec": f"{S3_TRAIN}/aicity.split)"},
    "dataset.val_dataset.ann_file": {"spec": f"{S3_EVAL}/aicity.split)"},
    "dataset.test_dataset.ann_file": {"spec": f"{S3_EVAL}/aicity.split)"},
}
导出（必填数据源）：
python
{
    "model.head.instance_bank.anchor": f"{S3_TRAIN}//results/{dataset_convert_job_id}/anchor_init.npy",
}
推理（必填数据源）：
python
{
    "dataset.data_root": {"spec": f"{S3_EVAL}/aicity.split)"},
    "model.head.instance_bank.anchor": f"{S3_TRAIN}//results/{dataset_convert_job_id}/anchor_init.npy",
    "dataset.train_dataset.ann_file": {"spec": f"{S3_TRAIN}/aicity.split)"},
    "dataset.val_dataset.ann_file": {"spec": f"{S3_EVAL}/aicity.split)"},
    "dataset.test_dataset.ann_file": {"spec": f"{S3_EVAL}/aicity.split)"},
}
量化（必填数据源）：
python
{
    "dataset.data_root": {"spec": f"{S3_TRAIN}/aicity.split)"},
    "model.head.instance_bank.anchor": f"{S3_TRAIN}//results/{dataset_convert_job_id}/anchor_init.npy",
    "dataset.train_dataset.ann_file": {"spec": f"{S3_TRAIN}/aicity.split)"},
    "dataset.val_dataset.ann_file": {"spec": f"{S3_EVAL}/aicity.split)"},
    "dataset.test_dataset.ann_file": {"spec": f"{S3_EVAL}/aicity.split)"},
    "dataset.quant_calibration_dataset.images_dir": f"{S3_TRAIN}",
}

Eval Dataset

评估数据集

Optional. Val/test splits configured via dataset ann_file paths.
可选。验证/测试集通过数据集ann_file路径配置。

Important Parameters

重要参数

  • model.backbone: Backbone. Default resnet_101.
  • model.neck.out_channels: FPN output channels. Default 256. num_outs=4.
  • model.input_shape: Input image shape [W, H]. Default [1408, 512].
  • model.head.num_output: Number of detection output queries. Default 300.
  • model.head.num_decoder: Number of decoder layers. Default 6.
  • model.head.temporal: Enable temporal reasoning. Default True.
  • model.head.instance_bank.num_anchor: Instance bank anchors. Default 900.
  • model.head.instance_bank.num_temp_instances: Temporal instance count. Default 600.
  • model.depth_branch.loss_weight: Depth supervision loss weight. Default 0.2.
  • dataset.batch_size: Per-GPU batch size. Default 2.
  • dataset.num_frames: Sequence length. Default 200.
  • dataset.classes: Detection classes. Default [person, gr1_t2, agility_digit, nova_carter]. num_ids=70 for tracking.
  • train.optim.lr: Learning rate. Default 5e-5. img_backbone lr_mult=0.2.
  • train.lr_scheduler: Cosine scheduler with linear warmup (500 iters, ratio 0.333).
  • train.grad_clip.max_norm: Gradient clipping. Default 25.
  • train.precision: Options: bf16, fp16, fp32. Default bf16.
  • evaluate.metrics: Eval metrics. Default ["detection"]. Optional tracking evaluation.
  • evaluate.tracking.enabled: Enable tracking evaluation. tracking_threshold=0.2.
  • model.backbone: 骨干网络。默认值为resnet_101。
  • model.neck.out_channels: FPN输出通道数。默认值为256。num_outs=4。
  • model.input_shape: 输入图像尺寸 [W, H]。默认值为[1408, 512]。
  • model.head.num_output: 检测输出查询数。默认值为300。
  • model.head.num_decoder: 解码器层数。默认值为6。
  • model.head.temporal: 启用时序推理。默认值为True。
  • model.head.instance_bank.num_anchor: 实例库锚点数。默认值为900。
  • model.head.instance_bank.num_temp_instances: 时序实例数量。默认值为600。
  • model.depth_branch.loss_weight: 深度监督损失权重。默认值为0.2。
  • dataset.batch_size: 单GPU批次大小。默认值为2。
  • dataset.num_frames: 序列长度。默认值为200。
  • dataset.classes: 检测类别。默认值为[person, gr1_t2, agility_digit, nova_carter]。跟踪任务的num_ids=70。
  • train.optim.lr: 学习率。默认值为5e-5。img_backbone lr_mult=0.2。
  • train.lr_scheduler: 带线性预热的余弦调度器（500次迭代，比例0.333）。
  • train.grad_clip.max_norm: 梯度裁剪。默认值为25。
  • train.precision: 选项：bf16, fp16, fp32。默认值为bf16。
  • evaluate.metrics: 评估指标。默认值为["detection"]。可选跟踪评估。
  • evaluate.tracking.enabled: 启用跟踪评估。tracking_threshold=0.2。

Multi-GPU / Multi-Node

多GPU / 多节点

Launch method: Lightning-managed (single 
python
process, Lightning spawns workers).
Spec KeyDescriptionDefault
train.num_gpus
Number of GPUs1
train.gpu_ids
GPU device indices[0]
train.num_nodes
Number of nodes1
  • Multi-GPU strategy: 
    ddp_find_unused_parameters_true
    (no fsdp support)
  • sync_batchnorm
    is always enabled (True)
  • Iterations per epoch computed as: 
    num_frames * num_bev_groups / (num_nodes * num_gpus * batch_size)
  • Scaling: When increasing GPUs, effective batch size grows and iterations-per-epoch shrinks proportionally
Multi-node env vars (set by orchestrator): 
WORLD_SIZE
, 
NODE_RANK
, 
MASTER_ADDR
, 
MASTER_PORT
, 
NUM_GPU_PER_NODE
.
启动方式： Lightning管理（单个
python
进程，Lightning生成工作进程）。
规格键描述默认值
train.num_gpus
GPU数量1
train.gpu_ids
GPU设备索引[0]
train.num_nodes
节点数量1
  • 多GPU策略：
    ddp_find_unused_parameters_true
    （不支持fsdp）
  • sync_batchnorm
    始终启用（True）
  • 每轮迭代数计算公式：
    num_frames * num_bev_groups / (num_nodes * num_gpus * batch_size)
  • 缩放规则： 增加GPU数量时，有效批次大小会成比例增加，每轮迭代数会成比例减少
多节点环境变量（由编排器设置）：
WORLD_SIZE
, 
NODE_RANK
, 
MASTER_ADDR
, 
MASTER_PORT
, 
NUM_GPU_PER_NODE
。

Hardware

硬件要求

Minimum 2 GPU(s), recommended 8 GPU(s). 40GB+ (A100 recommended) VRAM per GPU. Multi-camera temporal model is memory intensive. bf16 required for practical training. Multi-GPU strongly recommended. Instance bank requires substantial memory for temporal reasoning.
最少2块GPU，推荐8块GPU。每块GPU需40GB以上显存（推荐A100）。多相机时序模型对内存要求较高。实际训练需要bf16精度。强烈推荐使用多GPU。实例库进行时序推理需要大量内存。

Error Patterns

错误模式

dataset_convert required: Must run dataset_convert first to produce annotation pickles and anchor_init.npy.
Missing anchor file: Set model.head.instance_bank.anchor to the anchor_init.npy path from dataset_convert results.
Temporal OOM: Reduce dataset.num_frames or dataset.batch_size if running out of memory during temporal training.
需先执行dataset_convert：必须先运行dataset_convert生成标注pickle文件和anchor_init.npy。
缺失锚点文件：将model.head.instance_bank.anchor设置为dataset_convert结果中的anchor_init.npy路径。
时序训练内存不足（OOM）：若时序训练时内存不足，可减少dataset.num_frames或dataset.batch_size。

Spec Param / Parent Model Inference

规格参数 / 父模型推理

Model-specific inference mappings belong in this MD file, not in 
config.json
. Generated runners should read this section and apply the mappings with SDK helpers before 
create_job()
. This mirrors the old microservices 
infer_params.py
flow.
Inference mappings from TAO Core 
sparse4d.config.json
:
ActionSpec FieldInference FunctionMeaning
dataset_convert
results_dir
output_dir
current job results directory
evaluate
encryption_key
key
encryption key
evaluate
evaluate.checkpoint
parent_model
model file inferred from the parent job results folder
evaluate
results_dir
output_dir
current job results directory
export
encryption_key
key
encryption key
export
export.checkpoint
parent_model
model file inferred from the parent job results folder
export
export.onnx_file
create_onnx_file
output ONNX path
export
results_dir
output_dir
current job results directory
inference
encryption_key
key
encryption key
inference
inference.checkpoint
parent_model
model file inferred from the parent job results folder
inference
results_dir
output_dir
current job results directory
quantize
encryption_key
key
encryption key
quantize
quantize.model_path
parent_model
model file inferred from the parent job results folder
quantize
results_dir
output_dir
current job results directory
train
encryption_key
key
encryption key
train
results_dir
output_dir
current job results directory
train
train.pretrained_model_path
ptm_if_no_resume_model
PTM when no resume checkpoint exists
train
train.resume_training_checkpoint_path
resume_model
model file inferred from the current job results folder
For 
parent_model
or 
parent_model_folder
, pass the upstream train/export/AutoML child job id as 
parent_job_id
. The SDK lists the parent result folder, filters checkpoint artifacts, and returns the selected model file or folder. Do not add these mappings back to 
config.json
and do not patch generated runner scripts to guess checkpoint paths.
模型特定的推理映射应放在此MD文件中，而非
config.json
。生成的运行器应读取此部分内容，并在调用
create_job()
前使用SDK助手应用映射。这与旧微服务的
infer_params.py
流程一致。
来自TAO Core 
sparse4d.config.json
的推理映射：
动作规格字段推理函数含义
dataset_convert
results_dir
output_dir
当前任务结果目录
evaluate
encryption_key
key
加密密钥
evaluate
evaluate.checkpoint
parent_model
从父任务结果文件夹推断出的模型文件
evaluate
results_dir
output_dir
当前任务结果目录
export
encryption_key
key
加密密钥
export
export.checkpoint
parent_model
从父任务结果文件夹推断出的模型文件
export
export.onnx_file
create_onnx_file
输出ONNX路径
export
results_dir
output_dir
当前任务结果目录
inference
encryption_key
key
加密密钥
inference
inference.checkpoint
parent_model
从父任务结果文件夹推断出的模型文件
inference
results_dir
output_dir
当前任务结果目录
quantize
encryption_key
key
加密密钥
quantize
quantize.model_path
parent_model
从父任务结果文件夹推断出的模型文件
quantize
results_dir
output_dir
当前任务结果目录
train
encryption_key
key
加密密钥
train
results_dir
output_dir
当前任务结果目录
train
train.pretrained_model_path
ptm_if_no_resume_model
无恢复检查点时使用的预训练模型（PTM）
train
train.resume_training_checkpoint_path
resume_model
从当前任务结果文件夹推断出的模型文件
对于
parent_model
或
parent_model_folder
，传入上游训练/导出/AutoML子任务ID作为
parent_job_id
。SDK会列出父结果文件夹，过滤检查点工件，并返回选中的模型文件或文件夹。请勿将这些映射添加回
config.json
，也不要修改生成的运行器脚本来猜测检查点路径。