tao-finetune-clip

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

CLIP

Contrastive Language-Image Pre-training model for zero-shot and fine-tuned image classification, image-text retrieval, and embedding extraction. Fine-tuning adapts CLIP's shared image-text embedding space to domain-specific image-caption data.

No default NGC pretrained checkpoint is required. When

train.pretrained_model_path

evaluate.checkpoint

inference.checkpoint

, or

export.checkpoint

is unset, TAO loads pretrained weights from HuggingFace for SigLIP2/OpenCLIP variants or

torch.hub

for Radio-CLIP, so first use needs network access or a local mirror.

Supported actions:

train

evaluate

inference

export

gen_trt_engine

CLIP是用于零样本与微调图像分类、图文检索及嵌入提取的对比式语言-图像预训练模型。微调可让CLIP共享的图文嵌入空间适配特定领域的图像-标题数据。

无需默认NGC预训练检查点。当

train.pretrained_model_path

、

evaluate.checkpoint

、

inference.checkpoint

或

export.checkpoint

未设置时，TAO会从HuggingFace加载SigLIP2/OpenCLIP变体的预训练权重，或从

torch.hub

加载Radio-CLIP的预训练权重，因此首次使用需要网络访问或本地镜像。

支持的操作：

train

、

evaluate

、

inference

、

export

、

gen_trt_engine

。

Train Action Policy

训练操作策略

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，需读取

references/skill_info.yaml

，并通过显式

automl_policy

值或用户工作流请求解析运行覆盖项。将“关闭AutoML”“禁用AutoML”“无HPO”或“普通训练”这类表述视为本次运行的

automl_policy: off

；否则默认设为

auto

。当

automl_policy: auto

、

automl_enabled: true

，且

schemas/train.schema.json

和

references/spec_template_train.yaml

已打包时，默认将训练操作通过

tao-skill-bank:tao-run-automl

路由，并使用该模型的

skill_dir

。保留数据集、配置、输出目录、GPU/平台设置、父检查点及

automl_policy

的工作流/应用覆盖项。仅当

automl_policy: off

或打包的训练配置文件/模板缺失时，才使用直接模型训练；若配置文件缺失，需报告AutoML已启用但当前模型无法运行，需生成配置文件后方可使用。

非训练操作（如

evaluate

、

inference

、

export

及部署流程）仍保留在该模型技能中。单次运行的

automl_policy

覆盖项不会更改模型元数据。

Instructions

使用说明

Use this skill for NVIDIA TAO CLIP jobs: training, evaluation, embedding inference, ONNX export, and TensorRT engine generation. Start by identifying the requested action, then load only the referenced files needed for that action:

defaults.json

for default parameters,

config.json

for action/data-source wiring,

references/spec_template.yaml

for full spec shape, and

references/model_info.yaml

for SDK metadata.

For dataset-backed actions, collect the required image, caption, list, or prompt files from the user and place the resolved paths in

spec_overrides

. For

export

and

gen_trt_engine

, infer parent artifacts from the upstream job when available; otherwise require explicit checkpoint, ONNX, or engine paths. Run

gen_trt_engine

, TensorRT

evaluate

, and TensorRT

inference

in the TAO Deploy image.

For TAO Deploy TensorRT actions (

gen_trt_engine

, TensorRT

evaluate

, and TensorRT

inference

), read

references/tao-deploy-clip.md

first. Deploy spec templates live in this skill's

references/

folder with the

spec_template_deploy_*.yaml

prefix.

本技能适用于NVIDIA TAO CLIP任务：训练、评估、嵌入推理、ONNX导出及TensorRT引擎生成。首先确定所需操作，然后仅加载该操作所需的参考文件：

defaults.json

（默认参数）、

config.json

（操作/数据源关联）、

references/spec_template.yaml

（完整配置结构）及

references/model_info.yaml

（SDK元数据）。

对于基于数据集的操作，需从用户处收集所需的图像、标题、列表或提示文件，并将解析后的路径放入

spec_overrides

。对于

export

和

gen_trt_engine

操作，若上游任务存在则推断父工件；否则需要显式提供检查点、ONNX或引擎路径。

gen_trt_engine

、TensorRT

evaluate

及TensorRT

inference

需在TAO Deploy镜像中运行。

对于TAO Deploy TensorRT操作（

gen_trt_engine

、TensorRT

evaluate

及TensorRT

inference

），需先阅读

references/tao-deploy-clip.md

。部署配置模板存放在本技能的

references/

文件夹中，前缀为

spec_template_deploy_*.yaml

。

Training Requirements

训练要求

Dataset type: image_text
Formats: custom image/caption folders or WebDataset shards
Monitoring metric: val/t2i_mAP

数据集类型： image_text
格式： 自定义图像/标题文件夹或WebDataset分片
监控指标： val/t2i_mAP

Supported Models

支持的模型

SigLIP2:

siglip2-so400m-patch16-256

(default),

siglip2-so400m-patch14-224

siglip2-so400m-patch14-384

siglip2-so400m-patch16-384

siglip2-so400m-patch16-512

siglip2-so400m-patch16-naflex

Radio-CLIP:

c-radio_v3-b

c-radio_v3-l

c-radio_v3-h

c-radio_v3-g

OpenCLIP / NV-CLIP:

ViT-L-14-SigLIP-CLIPA-224

ViT-L-14-SigLIP-CLIPA-336

ViT-H-14-SigLIP-CLIPA-224

ViT-H-14-SigLIP-CLIPA-336

ViT-H-14-SigLIP-CLIPA-574

Radio-CLIP requires

model.adaptor_name

to be set to

siglip

clip

SigLIP2：

siglip2-so400m-patch16-256

（默认）、

siglip2-so400m-patch14-224

、

siglip2-so400m-patch14-384

、

siglip2-so400m-patch16-384

、

siglip2-so400m-patch16-512

、

siglip2-so400m-patch16-naflex

Radio-CLIP：

c-radio_v3-b

、

c-radio_v3-l

、

c-radio_v3-h

、

c-radio_v3-g

OpenCLIP / NV-CLIP：

ViT-L-14-SigLIP-CLIPA-224

、

ViT-L-14-SigLIP-CLIPA-336

、

ViT-H-14-SigLIP-CLIPA-224

、

ViT-H-14-SigLIP-CLIPA-336

、

ViT-H-14-SigLIP-CLIPA-574

Radio-CLIP需设置

model.adaptor_name

为

siglip

或

clip

。

Per-Action Dataset Requirements

各操作的数据集要求

Action	Spec Key	Source	Files	List?
train	dataset.train.datasets	train_datasets	image_dir: images.tar.gz, image_list_file: image_list.txt, caption_dir: captions.tar.gz	Yes
train	dataset.train.wds.root_dir	train_wds_dataset	root directory containing `.tar` shards	No
train	dataset.train.wds.shard_list_file	train_wds_dataset	shards.txt listing shard paths	No
train	dataset.val.datasets	eval_dataset	image_dir: images.tar.gz, image_list_file: image_list.txt, caption_dir: captions.tar.gz	Yes
evaluate	dataset.val.datasets	eval_dataset	image_dir: images.tar.gz, image_list_file: image_list.txt, caption_dir: captions.tar.gz	Yes
inference	inference.datasets	inference_dataset	image_dir: images.tar.gz	Yes
inference	inference.text_file	inference_dataset	prompts.txt	No
export	export.checkpoint	parent train job or explicit checkpoint	checkpoint .pth, optional for pretrained export	No
gen_trt_engine	gen_trt_engine.onnx_file	parent export job or explicit ONNX	clip_model.onnx	No

For custom training, set

dataset.train.type: custom

and provide

dataset.train.datasets

entries. Image and caption files must share the same base name.

caption_file_suffix

defaults to

.txt

, and

image_list_file

is optional.

For WDS training, set

dataset.train.type: wds

and provide at least one of

dataset.train.wds.root_dir

dataset.train.wds.shard_list_file

root_dir

is scanned recursively for

.tar

shards.

shard_list_file

is a text file with one shard path per line; relative lines resolve under the list-file directory unless

root_dir

is also supplied, in which case they resolve under

root_dir

. Validation/evaluation data remains custom format via

dataset.val.datasets

操作	配置键	来源	文件	是否需要列表？
train	dataset.train.datasets	train_datasets	image_dir: images.tar.gz, image_list_file: image_list.txt, caption_dir: captions.tar.gz	是
train	dataset.train.wds.root_dir	train_wds_dataset	包含 `.tar` 分片的根目录	否
train	dataset.train.wds.shard_list_file	train_wds_dataset	列出分片路径的shards.txt	否
train	dataset.val.datasets	eval_dataset	image_dir: images.tar.gz, image_list_file: image_list.txt, caption_dir: captions.tar.gz	是
evaluate	dataset.val.datasets	eval_dataset	image_dir: images.tar.gz, image_list_file: image_list.txt, caption_dir: captions.tar.gz	是
inference	inference.datasets	inference_dataset	image_dir: images.tar.gz	是
inference	inference.text_file	inference_dataset	prompts.txt	否
export	export.checkpoint	父训练任务或显式检查点	检查点.pth，预训练导出可选	否
gen_trt_engine	gen_trt_engine.onnx_file	父导出任务或显式ONNX文件	clip_model.onnx	否

自定义训练需设置

dataset.train.type: custom

并提供

dataset.train.datasets

条目。图像和标题文件必须共享相同的基础名称。

caption_file_suffix

默认为

.txt

，

image_list_file

为可选。

WDS训练需设置

dataset.train.type: wds

并至少提供

dataset.train.wds.root_dir

或

dataset.train.wds.shard_list_file

中的一个。

root_dir

会递归扫描

.tar

分片。

shard_list_file

是每行一个分片路径的文本文件；相对路径会在列表文件目录下解析，若同时提供

root_dir

则在

root_dir

下解析。验证/评估数据仍通过

dataset.val.datasets

使用自定义格式。

Typical Spec Overrides

典型配置覆盖项

Data source overrides are mandatory for dataset-backed actions. Construct paths from the Per-Action Dataset Requirements table and include them in

spec_overrides

. For inference, provide at least one of

inference.datasets

inference.text_file

python

S3_TRAIN = "s3://bucket/data/train"
S3_WDS = "s3://bucket/data/wds"
S3_EVAL = "s3://bucket/data/eval"
S3_INFER = "s3://bucket/data/infer"

train, custom dataset:

python

{
    "train.num_epochs": 10,
    "dataset.train.type": "custom",
    "dataset.train.datasets": [{"image_dir": f"{S3_TRAIN}/images.tar.gz", "image_list_file": f"{S3_TRAIN}/image_list.txt", "caption_dir": f"{S3_TRAIN}/captions.tar.gz"}],
    "dataset.val.datasets": [{"image_dir": f"{S3_EVAL}/images.tar.gz", "image_list_file": f"{S3_EVAL}/image_list.txt", "caption_dir": f"{S3_EVAL}/captions.tar.gz"}],
}

train, WDS dataset:

python

{
    "train.num_epochs": 10,
    "dataset.train.type": "wds",
    "dataset.train.wds.root_dir": f"{S3_WDS}",
    "dataset.train.wds.shard_list_file": f"{S3_WDS}/shards.txt",
    "dataset.train.wds.samples_per_shard": 10000,
    "dataset.val.datasets": [{"image_dir": f"{S3_EVAL}/images.tar.gz", "image_list_file": f"{S3_EVAL}/image_list.txt", "caption_dir": f"{S3_EVAL}/captions.tar.gz"}],
}

evaluate:

python

{
    "dataset.val.datasets": [{"image_dir": f"{S3_EVAL}/images.tar.gz", "image_list_file": f"{S3_EVAL}/image_list.txt", "caption_dir": f"{S3_EVAL}/captions.tar.gz"}],
}

Leave

evaluate.checkpoint

unset for zero-shot evaluation with pretrained weights. Set

evaluate.trt_engine

instead of

evaluate.checkpoint

for TensorRT evaluation.

inference:

python

{
    "inference.datasets": [{"image_dir": f"{S3_INFER}/images.tar.gz"}],
    "inference.text_file": f"{S3_INFER}/prompts.txt",
}

Inference writes

image_embeddings.h5

and/or

text_embeddings.h5

under

results_dir

. The saved embeddings are L2-normalized.

export:

python

{
    "export.onnx_file": "${results_dir}/export/clip_model.onnx",
    "export.encoder_type": "combined",
    "export.batch_size": -1,
}

Set

export.encoder_type: separate

when deployment should use independent vision and text encoders. Separate export writes

_vision.onnx

and

_text.onnx

variants derived from the base

export.onnx_file

gen_trt_engine:

python

{
    "gen_trt_engine.onnx_file": "${results_dir}/export/clip_model.onnx",
    "gen_trt_engine.trt_engine": "${results_dir}/deploy/clip_model.engine",
    "gen_trt_engine.batch_size": -1,
    "gen_trt_engine.tensorrt.data_type": "fp16",
    "gen_trt_engine.tensorrt.min_batch_size": 1,
    "gen_trt_engine.tensorrt.opt_batch_size": 1,
    "gen_trt_engine.tensorrt.max_batch_size": 16,
}

基于数据集的操作必须覆盖数据源。根据“各操作的数据集要求”表格构造路径并放入

spec_overrides

。推理操作需至少提供

inference.datasets

或

inference.text_file

中的一个。

python

S3_TRAIN = "s3://bucket/data/train"
S3_WDS = "s3://bucket/data/wds"
S3_EVAL = "s3://bucket/data/eval"
S3_INFER = "s3://bucket/data/infer"

训练（自定义数据集）：

python

{
    "train.num_epochs": 10,
    "dataset.train.type": "custom",
    "dataset.train.datasets": [{"image_dir": f"{S3_TRAIN}/images.tar.gz", "image_list_file": f"{S3_TRAIN}/image_list.txt", "caption_dir": f"{S3_TRAIN}/captions.tar.gz"}],
    "dataset.val.datasets": [{"image_dir": f"{S3_EVAL}/images.tar.gz", "image_list_file": f"{S3_EVAL}/image_list.txt", "caption_dir": f"{S3_EVAL}/captions.tar.gz"}],
}

训练（WDS数据集）：

python

{
    "train.num_epochs": 10,
    "dataset.train.type": "wds",
    "dataset.train.wds.root_dir": f"{S3_WDS}",
    "dataset.train.wds.shard_list_file": f"{S3_WDS}/shards.txt",
    "dataset.train.wds.samples_per_shard": 10000,
    "dataset.val.datasets": [{"image_dir": f"{S3_EVAL}/images.tar.gz", "image_list_file": f"{S3_EVAL}/image_list.txt", "caption_dir": f"{S3_EVAL}/captions.tar.gz"}],
}

评估：

python

{
    "dataset.val.datasets": [{"image_dir": f"{S3_EVAL}/images.tar.gz", "image_list_file": f"{S3_EVAL}/image_list.txt", "caption_dir": f"{S3_EVAL}/captions.tar.gz"}],
}

若使用预训练权重进行零样本评估，无需设置

evaluate.checkpoint

。若使用TensorRT评估，需设置

evaluate.trt_engine

而非

evaluate.checkpoint

。

推理：

python

{
    "inference.datasets": [{"image_dir": f"{S3_INFER}/images.tar.gz"}],
    "inference.text_file": f"{S3_INFER}/prompts.txt",
}

推理会在

results_dir

下生成

image_embeddings.h5

和/或

text_embeddings.h5

。保存的嵌入经过L2归一化处理。

导出：

python

{
    "export.onnx_file": "${results_dir}/export/clip_model.onnx",
    "export.encoder_type": "combined",
    "export.batch_size": -1,
}

若部署需使用独立的视觉和文本编码器，设置

export.encoder_type: separate

。分离导出会基于基础

export.onnx_file

生成

_vision.onnx

和

_text.onnx

变体。

生成TensorRT引擎：

python

{
    "gen_trt_engine.onnx_file": "${results_dir}/export/clip_model.onnx",
    "gen_trt_engine.trt_engine": "${results_dir}/deploy/clip_model.engine",
    "gen_trt_engine.batch_size": -1,
    "gen_trt_engine.tensorrt.data_type": "fp16",
    "gen_trt_engine.tensorrt.min_batch_size": 1,
    "gen_trt_engine.tensorrt.opt_batch_size": 1,
    "gen_trt_engine.tensorrt.max_batch_size": 16,
}

Eval Dataset

评估数据集

Optional for training. If provided, validation metrics are computed at validation intervals. Required for

evaluate

训练时可选。若提供，会在验证间隔计算验证指标。

evaluate

操作必须提供。

Deploy Workflow

部署工作流

The skill exposes

gen_trt_engine

as the deploy action. In generated SDK runners, use

model_info["actions"]["gen_trt_engine"]

and run it in the TAO Deploy image, not the PyTorch training image. The in-container command is

clip gen_trt_engine -e {config_path}

; direct TAO Launcher usage spells the same action as

tao deploy clip gen_trt_engine -e /path/to/spec.yaml

TAO Deploy supports both combined and separate encoder formats. For separate encoders, pass the base path without

_vision

_text

gen_trt_engine.onnx_file

and

gen_trt_engine.trt_engine

; TAO detects or writes the suffixed vision/text files.

Use

evaluate.trt_engine

for TensorRT evaluation and

inference.trt_engine

for TensorRT embedding extraction. These TensorRT paths also run in the TAO Deploy image. Direct TAO Launcher usage spells these as

tao deploy clip evaluate

and

tao deploy clip inference

Full TAO Deploy reference: tao-deploy-clip.

本技能将

gen_trt_engine

作为部署操作。在生成的SDK运行器中，使用

model_info["actions"]["gen_trt_engine"]

并在TAO Deploy镜像（而非PyTorch训练镜像）中运行。容器内命令为

clip gen_trt_engine -e {config_path}

；直接使用TAO Launcher的命令为

tao deploy clip gen_trt_engine -e /path/to/spec.yaml

。

TAO Deploy支持组合式和分离式编码器格式。对于分离式编码器，将不带

_vision

或

_text

的基础路径传入

gen_trt_engine.onnx_file

和

gen_trt_engine.trt_engine

；TAO会检测或生成带后缀的视觉/文本文件。

使用

evaluate.trt_engine

进行TensorRT评估，使用

inference.trt_engine

进行TensorRT嵌入提取。这些TensorRT操作同样在TAO Deploy镜像中运行。直接使用TAO Launcher的命令为

tao deploy clip evaluate

和

tao deploy clip inference

。

完整TAO Deploy参考：tao-deploy-clip。

Important Parameters

重要参数

model.type: Backbone family and resolution. Use fixed-resolution SigLIP2/OpenCLIP variants for deployment.
model.adaptor_name: Required for Radio-CLIP. Set to
```
siglip
```
or
```
clip
```
.
model.image_size: Training transform image resolution. Keep it aligned with the selected fixed-resolution backbone.
train.num_epochs: CLIP fine-tuning often converges quickly. Start with 10-20 epochs for domain adaptation, then increase only if validation loss is still improving.
train.optim.vision_lr / train.optim.text_lr: Learning rates for the two encoders. CLIP is sensitive to high learning rates; reduce both if loss is unstable.
model.freeze_vision_encoder / model.freeze_text_encoder: Defaults are false. Freezing one encoder can help when the dataset is small or only one modality needs adaptation.
train.loss_type:
```
siglip
```
is recommended for SigLIP2 and Radio-CLIP. Use
```
clip
```
for CLIP-style softmax loss.
export.encoder_type:
```
combined
```
exports one ONNX graph.
```
separate
```
exports independent vision and text graphs.
gen_trt_engine.tensorrt.data_type: TensorRT deployment supports
```
fp16
```
and
```
fp32
```
.

model.type： 骨干网络系列及分辨率。部署时使用固定分辨率的SigLIP2/OpenCLIP变体。
model.adaptor_name： Radio-CLIP必填。设置为
```
siglip
```
或
```
clip
```
。
model.image_size： 训练变换的图像分辨率。需与所选固定分辨率骨干网络保持一致。
train.num_epochs： CLIP微调通常收敛较快。领域适配先从10-20个epoch开始，仅当验证损失仍在改善时再增加。
train.optim.vision_lr / train.optim.text_lr： 两个编码器的学习率。CLIP对高学习率敏感；若损失不稳定，需降低两者的学习率。
model.freeze_vision_encoder / model.freeze_text_encoder： 默认值为false。当数据集较小时或仅需适配一种模态时，冻结一个编码器会有所帮助。
train.loss_type： SigLIP2和Radio-CLIP推荐使用
```
siglip
```
。CLIP风格的softmax损失使用
```
clip
```
。
export.encoder_type：
```
combined
```
导出单个ONNX图。
```
separate
```
导出独立的视觉和文本图。
gen_trt_engine.tensorrt.data_type： TensorRT部署支持
```
fp16
```
和
```
fp32
```
。

Hardware

硬件要求

Single-GPU training works for small datasets. Use 4+ GPUs for datasets with more than 100k images or large backbones. Use 16GB+ VRAM per GPU for small/fixed-resolution runs and larger GPUs for Radio-CLIP or high-resolution OpenCLIP variants.

单GPU适用于小型数据集训练。数据集超过10万张图像或使用大型骨干网络时，需使用4个及以上GPU。小型/固定分辨率运行需每GPU至少16GB显存，Radio-CLIP或高分辨率OpenCLIP变体需使用更大显存的GPU。

Error Patterns

错误模式

CUDA out of memory: Reduce

dataset.train.batch_size

dataset.val.batch_size

, or the TensorRT opt/max batch sizes. For export/deploy, check

export.input_height

and

export.input_width

against the selected fixed-resolution backbone.

NaN loss: Learning rate is too high for fine-tuning. Reduce

train.optim.vision_lr

and

train.optim.text_lr

, increase

train.optim.warmup_steps

, and verify that captions are valid non-empty text.

Zero retrieval or classification quality: Check that captions and prompts match the target label vocabulary. CLIP compares image and text embeddings, so prompt wording matters.

Dataset size smaller than total batch size: The total batch size is

batch_size * num_gpus

. If the dataset, especially validation, has fewer samples than this, reduce

dataset.val.batch_size

dataset.train.batch_size

Radio-CLIP config validation error: Set

model.adaptor_name

explicitly to

siglip

clip

Naflex export failure:

siglip2-so400m-patch16-naflex

is training-only in the current TAO docs and cannot be exported to ONNX or TensorRT. Use a fixed-resolution variant such as

siglip2-so400m-patch16-384

ONNX external data missing: Models larger than 2 GB export an ONNX file plus an external data file. Keep both files in the same directory and do not rename the external data file before

gen_trt_engine

TensorRT shape mismatch: When using dynamic batch export, provide min/opt/max shape profiles for every input. Text sequence length must match the tokenizer length, commonly 77 for CLIP tokenizers and 64 for SigLIP2 tokenizers.

attention_mask warning:

attention_mask

is currently accepted by exported graphs for compatibility, but TAO ignores its values and may remove it in a future release. Do not build new direct-ONNX inference code that depends on mask values.

Error merging spec.yaml with schema: A Hydra/OmegaConf config validation error. Common causes are putting

num_epochs

num_gpus

at the spec root instead of under

train.*

, or mixing up training image size (

model.image_size

) with export dimensions (

export.input_height

and

export.input_width

CUDA内存不足： 减小

dataset.train.batch_size

、

dataset.val.batch_size

或TensorRT的最优/最大批量大小。导出/部署时，检查

export.input_height

和

export.input_width

是否与所选固定分辨率骨干网络匹配。

损失为NaN： 微调学习率过高。降低

train.optim.vision_lr

和

train.optim.text_lr

，增加

train.optim.warmup_steps

，并验证标题为有效的非空文本。

检索或分类质量为零： 检查标题和提示是否匹配目标标签词汇表。CLIP会比较图像和文本嵌入，因此提示措辞至关重要。

数据集大小小于总批量大小： 总批量大小为

batch_size * num_gpus

。若数据集（尤其是验证集）样本数小于该值，需减小

dataset.val.batch_size

或

dataset.train.batch_size

。

Radio-CLIP配置验证错误： 需显式设置

model.adaptor_name

为

siglip

或

clip

。

Naflex导出失败： 当前TAO文档中

siglip2-so400m-patch16-naflex

仅支持训练，无法导出到ONNX或TensorRT。请使用固定分辨率变体，如

siglip2-so400m-patch16-384

。

ONNX外部数据缺失： 大于2GB的模型会导出一个ONNX文件加一个外部数据文件。需将两个文件放在同一目录下，且在

gen_trt_engine

前不要重命名外部数据文件。

TensorRT形状不匹配： 使用动态批量导出时，需为每个输入提供最小/最优/最大形状配置文件。文本序列长度必须与分词器长度匹配，CLIP分词器通常为77，SigLIP2分词器通常为64。

attention_mask警告： 导出的图目前为兼容性接受

attention_mask

，但TAO会忽略其值，未来版本可能移除该参数。请勿构建依赖掩码值的新直接ONNX推理代码。

spec.yaml与配置结构合并错误： Hydra/OmegaConf配置验证错误。常见原因是将

num_epochs

或

num_gpus

放在配置根目录而非

train.*

下，或混淆训练图像大小（

model.image_size

）与导出维度（

export.input_height

和

export.input_width

）。

Spec Param / Parent Model Inference

配置参数/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

clip.config.json

Action	Spec Field	Inference Function	Meaning
evaluate	`encryption_key`	`key`	encryption key
evaluate	`evaluate.checkpoint`	`parent_model`	model file inferred from the parent job results folder
evaluate	`evaluate.trt_engine`	`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
gen_trt_engine	`encryption_key`	`key`	encryption key
gen_trt_engine	`gen_trt_engine.onnx_file`	`parent_model`	model file inferred from the parent job results folder
gen_trt_engine	`gen_trt_engine.trt_engine`	`create_engine_file`	output TensorRT engine path
gen_trt_engine	`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	`inference.trt_engine`	`parent_model`	model file inferred from the parent job results folder
inference	`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

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本MD文件中，而非

config.json

。生成的运行器应读取本节内容，并在

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。

TAO Core

clip.config.json

中的推理映射：

操作	配置字段	推理函数	含义
evaluate	`encryption_key`	`key`	加密密钥
evaluate	`evaluate.checkpoint`	`parent_model`	从父任务结果文件夹推断的模型文件
evaluate	`evaluate.trt_engine`	`parent_model`	从父任务结果文件夹推断的模型文件
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
gen_trt_engine	`encryption_key`	`key`	加密密钥
gen_trt_engine	`gen_trt_engine.onnx_file`	`parent_model`	从父任务结果文件夹推断的模型文件
gen_trt_engine	`gen_trt_engine.trt_engine`	`create_engine_file`	输出TensorRT引擎路径
gen_trt_engine	`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的模型文件
inference	`inference.trt_engine`	`parent_model`	从父任务结果文件夹推断的模型文件
inference	`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`	无恢复检查点时的预训练模型
train	`train.resume_training_checkpoint_path`	`resume_model`	从当前任务结果文件夹推断的模型文件

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，并返回所选模型文件或文件夹。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以猜测检查点路径。