Vision Text Dual Encoder

VisionTextDualEncoder

Overview

The VisionTextDualEncoderModel can be used to initialize a vision-text dual encoder model with any pretrained vision autoencoding model as the vision encoder (e.g. ViT, BEiT, DeiT) and any pretrained text autoencoding model as the text encoder (e.g. RoBERTa, BERT). Two projection layers are added on top of both the vision and text encoder to project the output embeddings to a shared latent space. The projection layers are randomly initialized so the model should be fine-tuned on a downstream task. This model can be used to align the vision-text embeddings using CLIP like contrastive image-text training and then can be used for zero-shot vision tasks such image-classification or retrieval.

In LiT: Zero-Shot Transfer with Locked-image Text Tuning it is shown how leveraging pre-trained (locked/frozen) image and text model for contrastive learning yields significant improvement on new zero-shot vision tasks such as image classification or retrieval.

VisionTextDualEncoderConfig

class transformers.VisionTextDualEncoderConfig

<source>

( projection_dim = 512logit_scale_init_value = 2.6592**kwargs )

Parameters

• text_config (dict) — Dictionary of configuration options that defines text model config.

• vision_config (dict) — Dictionary of configuration options that defines vison model config.

• projection_dim (int, optional, defaults to 512) — Dimentionality of text and vision projection layers.

• logit_scale_init_value (float, optional, defaults to 2.6592) — The inital value of the logit_scale paramter. Default is used as per the original CLIP implementation.

• kwargs (optional) — Dictionary of keyword arguments.

VisionTextDualEncoderConfig is the configuration class to store the configuration of a VisionTextDualEncoderModel. It is used to instantiate VisionTextDualEncoderModel model according to the specified arguments, defining the text model and vision model configs.

Configuration objects inherit from PretrainedConfig and can be used to control the model outputs. Read the documentation from PretrainedConfig for more information.

Examples:

Copied

>>> from transformers import ViTConfig, BertConfig, VisionTextDualEncoderConfig, VisionTextDualEncoderModel

>>> # Initializing a BERT and ViT configuration
>>> config_vision = ViTConfig()
>>> config_text = BertConfig()

>>> config = VisionTextDualEncoderConfig.from_vision_text_configs(config_vision, config_text, projection_dim=512)

>>> # Initializing a BERT and ViT model (with random weights)
>>> model = VisionTextDualEncoderModel(config=config)

>>> # Accessing the model configuration
>>> config_vision = model.config.vision_config
>>> config_text = model.config.text_config

>>> # Saving the model, including its configuration
>>> model.save_pretrained("vit-bert")

>>> # loading model and config from pretrained folder
>>> vision_text_config = VisionTextDualEncoderConfig.from_pretrained("vit-bert")
>>> model = VisionTextDualEncoderModel.from_pretrained("vit-bert", config=vision_text_config)

from_vision_text_configs

<source>

( vision_config: PretrainedConfigtext_config: PretrainedConfig**kwargs ) → VisionTextDualEncoderConfig

Returns

VisionTextDualEncoderConfig

An instance of a configuration object

Instantiate a VisionTextDualEncoderConfig (or a derived class) from text model configuration and vision model configuration.

VisionTextDualEncoderProcessor

class transformers.VisionTextDualEncoderProcessor

<source>

( image_processor = Nonetokenizer = None**kwargs )

Parameters

• image_processor (AutoImageProcessor) — The image processor is a required input.

• tokenizer (PreTrainedTokenizer) — The tokenizer is a required input.

Constructs a VisionTextDualEncoder processor which wraps an image processor and a tokenizer into a single processor.

VisionTextDualEncoderProcessor offers all the functionalities of AutoImageProcessor and AutoTokenizer. See the __call__() and decode() for more information.

batch_decode

<source>

( *args**kwargs )

This method forwards all its arguments to VisionTextDualEncoderTokenizer’s batch_decode(). Please refer to the docstring of this method for more information.

decode

<source>

( *args**kwargs )

This method forwards all its arguments to VisionTextDualEncoderTokenizer’s decode(). Please refer to the docstring of this method for more information.

VisionTextDualEncoderModel

class transformers.VisionTextDualEncoderModel

<source>

( config: typing.Optional[transformers.models.vision_text_dual_encoder.configuration_vision_text_dual_encoder.VisionTextDualEncoderConfig] = Nonevision_model: typing.Optional[transformers.modeling_utils.PreTrainedModel] = Nonetext_model: typing.Optional[transformers.modeling_utils.PreTrainedModel] = None )

Parameters

• config (VisionEncoderDecoderConfig) — Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the from_pretrained() method to load the model weights.

This class can be used to initialize a vision-text dual encoder model with any pretrained vision autoencoding model as the vision encoder and any pretrained text model as the text encoder. The vision and text encoders are loaded via the from_pretrained() method. The projection layers are automatically added to the model and should be fine-tuned on a downstream task, like contrastive image-text modeling.

In LiT: Zero-Shot Transfer with Locked-image Text Tuning it is shown how leveraging pre-trained (locked/frozen) image and text model for contrastive learning yields significant improvment on new zero-shot vision tasks such as image classification or retrieval.

After such a Vision-Text-Dual-Encoder model has been trained/fine-tuned, it can be saved/loaded just like any other models (see the examples for more information).

This model inherits from PreTrainedModel. Check the superclass documentation for the generic methods the library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads etc.)

This model is also a PyTorch torch.nn.Module subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior.

forward

<source>

( input_ids: typing.Optional[torch.LongTensor] = Nonepixel_values: typing.Optional[torch.FloatTensor] = Noneattention_mask: typing.Optional[torch.Tensor] = Noneposition_ids: typing.Optional[torch.LongTensor] = Nonereturn_loss: typing.Optional[bool] = Nonetoken_type_ids: typing.Optional[torch.LongTensor] = Noneoutput_attentions: typing.Optional[bool] = Noneoutput_hidden_states: typing.Optional[bool] = Nonereturn_dict: typing.Optional[bool] = None ) → transformers.models.clip.modeling_clip.CLIPOutput or tuple(torch.FloatTensor)

Parameters

• input_ids (torch.LongTensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide it.

  Indices can be obtained using AutoTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.

  What are input IDs?

• attention_mask (torch.Tensor of shape (batch_size, sequence_length), optional) — Mask to avoid performing attention on padding token indices. Mask values selected in [0, 1]:

  • 1 for tokens that are not masked,

  • 0 for tokens that are masked.

  What are attention masks?

• position_ids (torch.LongTensor of shape (batch_size, sequence_length), optional) — Indices of positions of each input sequence tokens in the position embeddings. Selected in the range [0, config.max_position_embeddings - 1].

  What are position IDs?

• pixel_values (torch.FloatTensor of shape (batch_size, num_channels, height, width)) — Pixel values. Padding will be ignored by default should you provide it. Pixel values can be obtained using an image processor (e.g. if you use ViT as the encoder, you should use AutoImageProcessor). See ViTImageProcessor.call() for details.

• return_loss (bool, optional) — Whether or not to return the contrastive loss.

• output_attentions (bool, optional) — Whether or not to return the attentions tensors of all attention layers. See attentions under returned tensors for more detail.

• output_hidden_states (bool, optional) — Whether or not to return the hidden states of all layers. See hidden_states under returned tensors for more detail.

• return_dict (bool, optional) — Whether or not to return a ModelOutput instead of a plain tuple.

Returns

transformers.models.clip.modeling_clip.CLIPOutput or tuple(torch.FloatTensor)

A transformers.models.clip.modeling_clip.CLIPOutput or a tuple of torch.FloatTensor (if return_dict=False is passed or when config.return_dict=False) comprising various elements depending on the configuration (VisionTextDualEncoderConfig) and inputs.

• loss (torch.FloatTensor of shape (1,), optional, returned when return_loss is True) — Contrastive loss for image-text similarity.

• logits_per_image:(torch.FloatTensor of shape (image_batch_size, text_batch_size)) — The scaled dot product scores between image_embeds and text_embeds. This represents the image-text similarity scores.

• logits_per_text:(torch.FloatTensor of shape (text_batch_size, image_batch_size)) — The scaled dot product scores between text_embeds and image_embeds. This represents the text-image similarity scores.

• text_embeds(torch.FloatTensor of shape (batch_size, output_dim) — The text embeddings obtained by applying the projection layer to the pooled output of CLIPTextModel.

• image_embeds(torch.FloatTensor of shape (batch_size, output_dim) — The image embeddings obtained by applying the projection layer to the pooled output of CLIPVisionModel.

• text_model_output(BaseModelOutputWithPooling): The output of the CLIPTextModel.

• vision_model_output(BaseModelOutputWithPooling): The output of the CLIPVisionModel.

The VisionTextDualEncoderModel forward method, overrides the __call__ special method.

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the pre and post processing steps while the latter silently ignores them.

Examples:

Copied

>>> from PIL import Image
>>> import requests
>>> from transformers import (
...     VisionTextDualEncoderModel,
...     VisionTextDualEncoderProcessor,
...     AutoImageProcessor,
...     AutoTokenizer,
... )

>>> tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
>>> image_processor = AutoImageProcessor.from_pretrained("google/vit-base-patch16-224")
>>> processor = VisionTextDualEncoderProcessor(image_processor, tokenizer)
>>> model = VisionTextDualEncoderModel.from_vision_text_pretrained(
...     "google/vit-base-patch16-224", "bert-base-uncased"
... )

>>> # contrastive training
>>> urls = [
...     "http://images.cocodataset.org/val2017/000000039769.jpg",
...     "https://farm3.staticflickr.com/2674/5850229113_4fe05d5265_z.jpg",
... ]
>>> images = [Image.open(requests.get(url, stream=True).raw) for url in urls]
>>> inputs = processor(
...     text=["a photo of a cat", "a photo of a dog"], images=images, return_tensors="pt", padding=True
... )
>>> outputs = model(
...     input_ids=inputs.input_ids,
...     attention_mask=inputs.attention_mask,
...     pixel_values=inputs.pixel_values,
...     return_loss=True,
... )
>>> loss, logits_per_image = outputs.loss, outputs.logits_per_image  # this is the image-text similarity score

>>> # save and load from pretrained
>>> model.save_pretrained("vit-bert")
>>> model = VisionTextDualEncoderModel.from_pretrained("vit-bert")

>>> # inference
>>> outputs = model(**inputs)
>>> logits_per_image = outputs.logits_per_image  # this is the image-text similarity score
>>> probs = logits_per_image.softmax(dim=1)  # we can take the softmax to get the label probabilities

FlaxVisionTextDualEncoderModel

class transformers.FlaxVisionTextDualEncoderModel

<source>

( config: VisionTextDualEncoderConfiginput_shape: typing.Optional[typing.Tuple] = Noneseed: int = 0dtype: dtype = <class 'jax.numpy.float32'>_do_init: bool = True**kwargs )

Parameters

• config (VisionTextDualEncoderConfig) — Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the from_pretrained() method to load the model weights.

• dtype (jax.numpy.dtype, optional, defaults to jax.numpy.float32) — The data type of the computation. Can be one of jax.numpy.float32, jax.numpy.float16 (on GPUs) and jax.numpy.bfloat16 (on TPUs).

  This can be used to enable mixed-precision training or half-precision inference on GPUs or TPUs. If specified all the computation will be performed with the given dtype.

  Note that this only specifies the dtype of the computation and does not influence the dtype of model parameters.

  If you wish to change the dtype of the model parameters, see to_fp16() and to_bf16().

This class can be used to initialize a vision-text dual encoder model with any pretrained vision autoencoding model as the vision encoder and any pretrained text model as the text encoder. The vision and text encoders are loaded via the from_pretrained() method. The projection layers are automatically added to the model and should be fine-tuned on a downstream task, like contrastive image-text modeling.

In LiT: Zero-Shot Transfer with Locked-image Text Tuning it is shown how leveraging pre-trained (locked/frozen) image and text model for contrastive learning yields significant improvment on new zero-shot vision tasks such as image classification or retrieval.

After such a Vision-Text-Dual-Encoder model has been trained/fine-tuned, it can be saved/loaded just like any other models (see the examples for more information).

This model inherits from PreTrainedModel. Check the superclass documentation for the generic methods the library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads etc.)

This model is also a Flax Linen flax.linen.Module subclass. Use it as a regular Flax linen Module and refer to the Flax documentation for all matter related to general usage and behavior.

Finally, this model supports inherent JAX features such as:

• Just-In-Time (JIT) compilation

• Automatic Differentiation

• Vectorization

• Parallelization

__call__

<source>

( input_idspixel_valuesattention_mask = Noneposition_ids = Nonetoken_type_ids = Noneparams: dict = Nonedropout_rng: PRNGKey = Nonetrain: bool = Falseoutput_attentions: typing.Optional[bool] = Noneoutput_hidden_states: typing.Optional[bool] = Nonereturn_dict: typing.Optional[bool] = None ) → transformers.models.clip.modeling_flax_clip.FlaxCLIPOutput or tuple(torch.FloatTensor)

Parameters

• input_ids (numpy.ndarray of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide it.

  Indices can be obtained using AutoTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.

  What are input IDs?

• attention_mask (torch.Tensor of shape (batch_size, sequence_length), optional) — Mask to avoid performing attention on padding token indices. Mask values selected in [0, 1]:

  • 1 for tokens that are not masked,

  • 0 for tokens that are masked.

  What are attention masks?

• position_ids (numpy.ndarray of shape (batch_size, sequence_length), optional) — Indices of positions of each input sequence tokens in the position embeddings. Selected in the range [0, config.max_position_embeddings - 1].

  What are position IDs?

• pixel_values (torch.FloatTensor of shape (batch_size, num_channels, height, width)) — Pixel values. Padding will be ignored by default should you provide it. Pixel values can be obtained using an image processor (e.g. if you use ViT as the encoder, you should use AutoImageProcessor). See ViTImageProcessor.call() for details.

• output_attentions (bool, optional) — Whether or not to return the attentions tensors of all attention layers. See attentions under returned tensors for more detail.

• output_hidden_states (bool, optional) — Whether or not to return the hidden states of all layers. See hidden_states under returned tensors for more detail.

• return_dict (bool, optional) — Whether or not to return a ModelOutput instead of a plain tuple.

Returns

transformers.models.clip.modeling_flax_clip.FlaxCLIPOutput or tuple(torch.FloatTensor)

A transformers.models.clip.modeling_flax_clip.FlaxCLIPOutput or a tuple of torch.FloatTensor (if return_dict=False is passed or when config.return_dict=False) comprising various elements depending on the configuration (VisionTextDualEncoderConfig) and inputs.

• logits_per_image:(jnp.ndarray of shape (image_batch_size, text_batch_size)) — The scaled dot product scores between image_embeds and text_embeds. This represents the image-text similarity scores.

• logits_per_text:(jnp.ndarray of shape (text_batch_size, image_batch_size)) — The scaled dot product scores between text_embeds and image_embeds. This represents the text-image similarity scores.

• text_embeds(jnp.ndarray of shape (batch_size, output_dim) — The text embeddings obtained by applying the projection layer to the pooled output of FlaxCLIPTextModel.

• image_embeds(jnp.ndarray of shape (batch_size, output_dim) — The image embeddings obtained by applying the projection layer to the pooled output of FlaxCLIPVisionModel.

• text_model_output(FlaxBaseModelOutputWithPooling): The output of the FlaxCLIPTextModel.

• vision_model_output(FlaxBaseModelOutputWithPooling): The output of the FlaxCLIPVisionModel.

The FlaxVisionTextDualEncoderModel forward method, overrides the __call__ special method.

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the pre and post processing steps while the latter silently ignores them.

Examples:

Copied

>>> from PIL import Image
>>> import requests
>>> import jax
>>> from transformers import (
...     FlaxVisionTextDualEncoderModel,
...     VisionTextDualEncoderProcessor,
...     AutoImageProcessor,
...     AutoTokenizer,
... )

>>> tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
>>> image_processor = AutoImageProcesor.from_pretrained("google/vit-base-patch16-224")
>>> processor = VisionTextDualEncoderProcessor(image_processor, tokenizer)
>>> model = FlaxVisionTextDualEncoderModel.from_vision_text_pretrained(
...     "google/vit-base-patch16-224", "bert-base-uncased"
... )

>>> # contrastive training
>>> urls = [
...     "http://images.cocodataset.org/val2017/000000039769.jpg",
...     "https://farm3.staticflickr.com/2674/5850229113_4fe05d5265_z.jpg",
... ]
>>> images = [Image.open(requests.get(url, stream=True).raw) for url in urls]
>>> inputs = processor(
...     text=["a photo of a cat", "a photo of a dog"], images=images, return_tensors="np", padding=True
... )
>>> outputs = model(
...     input_ids=inputs.input_ids,
...     attention_mask=inputs.attention_mask,
...     pixel_values=inputs.pixel_values,
... )
>>> logits_per_image = outputs.logits_per_image  # this is the image-text similarity score

>>> # save and load from pretrained
>>> model.save_pretrained("vit-bert")
>>> model = FlaxVisionTextDualEncoderModel.from_pretrained("vit-bert")

>>> # inference
>>> outputs = model(**inputs)
>>> logits_per_image = outputs.logits_per_image  # this is the image-text similarity score
>>> probs = jax.nn.softmax(logits_per_image, axis=1)  # we can take the softmax to get the label probabilities

TFVisionTextDualEncoderModel

class transformers.TFVisionTextDualEncoderModel

<source>

( *args**kwargs )

Parameters

• config (VisionEncoderDecoderConfig) — Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the from_pretrained() method to load the model weights.

This class can be used to initialize a vision-text dual encoder model with any pretrained vision autoencoding model as the vision encoder and any pretrained text model as the text encoder. The vision and text encoders are loaded via the from_pretrained() method. The projection layers are automatically added to the model and should be fine-tuned on a downstream task, like contrastive image-text modeling.

In LiT: Zero-Shot Transfer with Locked-image Text Tuning it is shown how leveraging pre-trained (locked/frozen) image and text model for contrastive learning yields significant improvment on new zero-shot vision tasks such as image classification or retrieval.

After such a Vision-Text-Dual-Encoder model has been trained/fine-tuned, it can be saved/loaded just like any other models (see the examples for more information).

This model inherits from TFPreTrainedModel. Check the superclass documentation for the generic methods the library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads etc.)

This model is also a Keras Model subclass. Use it as a regular Keras Model and refer to the TF documentation for all matter related to general usage and behavior.

call

<source>

( input_ids: tf.Tensor | None = Nonepixel_values: tf.Tensor | None = Noneattention_mask: tf.Tensor | None = Noneposition_ids: tf.Tensor | None = Nonereturn_loss: Optional[bool] = Nonetoken_type_ids: tf.Tensor | None = Noneoutput_attentions: Optional[bool] = Noneoutput_hidden_states: Optional[bool] = Nonereturn_dict: Optional[bool] = Nonetraining: bool = False ) → transformers.models.clip.modeling_tf_clip.TFCLIPOutput or tuple(tf.Tensor)

Parameters

• input_ids (tf.Tensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide it.

  Indices can be obtained using AutoTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.

  What are input IDs?

• attention_mask (tf.Tensor of shape (batch_size, sequence_length), optional) — Mask to avoid performing attention on padding token indices. Mask values selected in [0, 1]:

  • 1 for tokens that are not masked,

  • 0 for tokens that are masked.

  What are attention masks?

• position_ids (tf.Tensor of shape (batch_size, sequence_length), optional) — Indices of positions of each input sequence tokens in the position embeddings. Selected in the range [0, config.max_position_embeddings - 1].

  What are position IDs?

• pixel_values (tf.Tensor of shape (batch_size, num_channels, height, width)) — Pixel values. Padding will be ignored by default should you provide it. Pixel values can be obtained using an image processor (e.g. if you use ViT as the encoder, you should use AutoImageProcessor). See ViTImageProcessor.call() for details.

• return_loss (bool, optional) — Whether or not to return the contrastive loss.

• output_attentions (bool, optional) — Whether or not to return the attentions tensors of all attention layers. See attentions under returned tensors for more detail.

• output_hidden_states (bool, optional) — Whether or not to return the hidden states of all layers. See hidden_states under returned tensors for more detail.

• return_dict (bool, optional) — Whether or not to return a ModelOutput instead of a plain tuple.

Returns

transformers.models.clip.modeling_tf_clip.TFCLIPOutput or tuple(tf.Tensor)

A transformers.models.clip.modeling_tf_clip.TFCLIPOutput or a tuple of tf.Tensor (if return_dict=False is passed or when config.return_dict=False) comprising various elements depending on the configuration (VisionTextDualEncoderConfig) and inputs.

• loss (tf.Tensor of shape (1,), optional, returned when return_loss is True) — Contrastive loss for image-text similarity.

• logits_per_image:(tf.Tensor of shape (image_batch_size, text_batch_size)) — The scaled dot product scores between image_embeds and text_embeds. This represents the image-text similarity scores.

• logits_per_text:(tf.Tensor of shape (text_batch_size, image_batch_size)) — The scaled dot product scores between text_embeds and image_embeds. This represents the text-image similarity scores.

• text_embeds(tf.Tensor of shape (batch_size, output_dim) — The text embeddings obtained by applying the projection layer to the pooled output of TFCLIPTextModel.

• image_embeds(tf.Tensor of shape (batch_size, output_dim) — The image embeddings obtained by applying the projection layer to the pooled output of TFCLIPVisionModel.

• text_model_output(~modeling_tf_utils.TFBaseModelOutputWithPooling): The output of the TFCLIPTextModel.

• vision_model_output(~modeling_tf_utils.TFBaseModelOutputWithPooling): The output of the TFCLIPVisionModel.

The TFVisionTextDualEncoderModel forward method, overrides the __call__ special method.

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the pre and post processing steps while the latter silently ignores them.

Examples:

Copied

>>> from PIL import Image
>>> import requests
>>> from transformers import (
...     TFVisionTextDualEncoderModel,
...     VisionTextDualEncoderProcessor,
...     AutoImageProcessor,
...     AutoTokenizer,
... )

>>> tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
>>> image_processor = AutoImageProcessor.from_pretrained("google/vit-base-patch16-224")
>>> processor = VisionTextDualEncoderProcessor(image_processor, tokenizer)
>>> model = TFVisionTextDualEncoderModel.from_vision_text_pretrained(
...     "google/vit-base-patch16-224", "bert-base-uncased"
... )

>>> # contrastive training
>>> urls = [
...     "http://images.cocodataset.org/val2017/000000039769.jpg",
...     "https://farm3.staticflickr.com/2674/5850229113_4fe05d5265_z.jpg",
... ]
>>> images = [Image.open(requests.get(url, stream=True).raw) for url in urls]
>>> inputs = processor(
...     text=["a photo of a cat", "a photo of a dog"], images=images, return_tensors="np", padding=True
... )
>>> outputs = model(
...     input_ids=inputs.input_ids,
...     attention_mask=inputs.attention_mask,
...     pixel_values=inputs.pixel_values,
...     return_loss=True,
... )
>>> loss, logits_per_image = outputs.loss, outputs.logits_per_image  # this is the image-text similarity score

>>> # save and load from pretrained
>>> model.save_pretrained("vit-bert")
>>> model = TFVisionTextDualEncoderModel.from_pretrained("vit-bert")

>>> # inference
>>> outputs = model(**inputs)
>>> logits_per_image = outputs.logits_per_image  # this is the image-text similarity score
>>> probs = tf.nn.softmax(logits_per_image, axis=1)  # we can take the softmax to get the label probabilities