The VisionEncoderDecoderModel can be used to initialize an image-to-text model with any pretrained Transformer-based vision model as the encoder (e.g. ViT, BEiT, DeiT, Swin) and any pretrained language model as the decoder (e.g. RoBERTa, GPT2, BERT, DistilBERT).
The effectiveness of initializing image-to-text-sequence models with pretrained checkpoints has been shown in (for example) TrOCR: Transformer-based Optical Character Recognition with Pre-trained Models by Minghao Li, Tengchao Lv, Lei Cui, Yijuan Lu, Dinei Florencio, Cha Zhang, Zhoujun Li, Furu Wei.
After such a VisionEncoderDecoderModel has been trained/fine-tuned, it can be saved/loaded just like any other models (see the examples below for more information).
An example application is image captioning, in which the encoder is used to encode the image, after which an autoregressive language model generates the caption. Another example is optical character recognition. Refer to TrOCR, which is an instance of VisionEncoderDecoderModel.
VisionEncoderDecoderModel can be randomly initialized from an encoder and a decoder config. In the following example, we show how to do this using the default ViTModel configuration for the encoder
and the default BertForCausalLM configuration for the decoder.
>>> from transformers import BertConfig, ViTConfig, VisionEncoderDecoderConfig, VisionEncoderDecoderModel
>>> config_encoder = ViTConfig()
>>> config_decoder = BertConfig()
>>> config = VisionEncoderDecoderConfig.from_encoder_decoder_configs(config_encoder, config_decoder)
>>> model = VisionEncoderDecoderModel(config=config)VisionEncoderDecoderModel can be initialized from a pretrained encoder checkpoint and a pretrained decoder checkpoint. Note that any pretrained Transformer-based vision model, e.g. Swin, can serve as the encoder and both pretrained auto-encoding models, e.g. BERT, pretrained causal language models, e.g. GPT2, as well as the pretrained decoder part of sequence-to-sequence models, e.g. decoder of BART, can be used as the decoder.
Depending on which architecture you choose as the decoder, the cross-attention layers might be randomly initialized.
Initializing VisionEncoderDecoderModel from a pretrained encoder and decoder checkpoint requires the model to be fine-tuned on a downstream task, as has been shown in the Warm-starting-encoder-decoder blog post.
To do so, the VisionEncoderDecoderModel class provides a VisionEncoderDecoderModel.from_encoder_decoder_pretrained() method.
>>> from transformers import VisionEncoderDecoderModel
>>> model = VisionEncoderDecoderModel.from_encoder_decoder_pretrained(
... "microsoft/swin-base-patch4-window7-224-in22k", "google-bert/bert-base-uncased"
... )To load fine-tuned checkpoints of the VisionEncoderDecoderModel class, VisionEncoderDecoderModel provides the from_pretrained(...) method just like any other model architecture in Transformers.
To perform inference, one uses the generate method, which allows to autoregressively generate text. This method supports various forms of decoding, such as greedy, beam search and multinomial sampling.
>>> import requests
>>> from PIL import Image
>>> from transformers import GPT2TokenizerFast, ViTImageProcessor, VisionEncoderDecoderModel
>>> # load a fine-tuned image captioning model and corresponding tokenizer and image processor
>>> model = VisionEncoderDecoderModel.from_pretrained("nlpconnect/vit-gpt2-image-captioning")
>>> tokenizer = GPT2TokenizerFast.from_pretrained("nlpconnect/vit-gpt2-image-captioning")
>>> image_processor = ViTImageProcessor.from_pretrained("nlpconnect/vit-gpt2-image-captioning")
>>> # let's perform inference on an image
>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
>>> image = Image.open(requests.get(url, stream=True).raw)
>>> pixel_values = image_processor(image, return_tensors="pt").pixel_values
>>> # autoregressively generate caption (uses greedy decoding by default)
>>> generated_ids = model.generate(pixel_values)
>>> generated_text = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
>>> print(generated_text)
a cat laying on a blanket next to a cat laying on a bedTFVisionEncoderDecoderModel.from_pretrained() currently doesn’t support initializing the model from a
PyTorch checkpoint. Passing from_pt=True to this method will throw an exception. If there are only PyTorch
checkpoints for a particular vision encoder-decoder model, a workaround is:
>>> from transformers import VisionEncoderDecoderModel, TFVisionEncoderDecoderModel
>>> _model = VisionEncoderDecoderModel.from_pretrained("nlpconnect/vit-gpt2-image-captioning")
>>> _model.encoder.save_pretrained("./encoder")
>>> _model.decoder.save_pretrained("./decoder")
>>> model = TFVisionEncoderDecoderModel.from_encoder_decoder_pretrained(
... "./encoder", "./decoder", encoder_from_pt=True, decoder_from_pt=True
... )
>>> # This is only for copying some specific attributes of this particular model.
>>> model.config = _model.configOnce the model is created, it can be fine-tuned similar to BART, T5 or any other encoder-decoder model on a dataset of (image, text) pairs.
As you can see, only 2 inputs are required for the model in order to compute a loss: pixel_values (which are the
images) and labels (which are the input_ids of the encoded target sequence).
>>> from transformers import ViTImageProcessor, BertTokenizer, VisionEncoderDecoderModel
>>> from datasets import load_dataset
>>> image_processor = ViTImageProcessor.from_pretrained("google/vit-base-patch16-224-in21k")
>>> tokenizer = BertTokenizer.from_pretrained("google-bert/bert-base-uncased")
>>> model = VisionEncoderDecoderModel.from_encoder_decoder_pretrained(
... "google/vit-base-patch16-224-in21k", "google-bert/bert-base-uncased"
... )
>>> model.config.decoder_start_token_id = tokenizer.cls_token_id
>>> model.config.pad_token_id = tokenizer.pad_token_id
>>> dataset = load_dataset("huggingface/cats-image")
>>> image = dataset["test"]["image"][0]
>>> pixel_values = image_processor(image, return_tensors="pt").pixel_values
>>> labels = tokenizer(
... "an image of two cats chilling on a couch",
... return_tensors="pt",
... ).input_ids
>>> # the forward function automatically creates the correct decoder_input_ids
>>> loss = model(pixel_values=pixel_values, labels=labels).lossThis model was contributed by nielsr. This model’s TensorFlow and Flax versions were contributed by ydshieh.
( **kwargs )
Parameters
VisionEncoderDecoderConfig is the configuration class to store the configuration of a VisionEncoderDecoderModel. It is used to instantiate a Vision-Encoder-Text-Decoder model according to the specified arguments, defining the encoder and decoder configs.
Configuration objects inherit from PretrainedConfig and can be used to control the model outputs. Read the documentation from PretrainedConfig for more information.
Examples:
>>> from transformers import BertConfig, ViTConfig, VisionEncoderDecoderConfig, VisionEncoderDecoderModel
>>> # Initializing a ViT & BERT style configuration
>>> config_encoder = ViTConfig()
>>> config_decoder = BertConfig()
>>> config = VisionEncoderDecoderConfig.from_encoder_decoder_configs(config_encoder, config_decoder)
>>> # Initializing a ViTBert model (with random weights) from a ViT & google-bert/bert-base-uncased style configurations
>>> model = VisionEncoderDecoderModel(config=config)
>>> # Accessing the model configuration
>>> config_encoder = model.config.encoder
>>> config_decoder = model.config.decoder
>>> # set decoder config to causal lm
>>> config_decoder.is_decoder = True
>>> config_decoder.add_cross_attention = True
>>> # Saving the model, including its configuration
>>> model.save_pretrained("my-model")
>>> # loading model and config from pretrained folder
>>> encoder_decoder_config = VisionEncoderDecoderConfig.from_pretrained("my-model")
>>> model = VisionEncoderDecoderModel.from_pretrained("my-model", config=encoder_decoder_config)( encoder_config: PretrainedConfig decoder_config: PretrainedConfig **kwargs ) → VisionEncoderDecoderConfig
Instantiate a VisionEncoderDecoderConfig (or a derived class) from a pre-trained encoder model configuration and decoder model configuration.
( config: Optional = None encoder: Optional = None decoder: Optional = None )
Parameters
This class can be used to initialize an image-to-text-sequence model with any pretrained vision autoencoding model as the encoder and any pretrained text autoregressive model as the decoder. The encoder is loaded via from_pretrained() function and the decoder is loaded via from_pretrained() function. Cross-attention layers are automatically added to the decoder and should be fine-tuned on a downstream generative task, like image captioning.
The effectiveness of initializing sequence-to-sequence models with pretrained checkpoints for sequence generation tasks was shown in Leveraging Pre-trained Checkpoints for Sequence Generation Tasks by Sascha Rothe, Shashi Narayan, Aliaksei Severyn. Michael Matena, Yanqi Zhou, Wei Li, Peter J. Liu.
Additionally, in TrOCR: Transformer-based Optical Character Recognition with Pre-trained Models it is shown how leveraging large pretrained vision models for optical character recognition (OCR) yields a significant performance improvement.
After such a Vision-Encoder-Text-Decoder 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.
VisionEncoderDecoderModel is a generic model class that will be instantiated as a transformer architecture with one of the base vision model classes of the library as encoder and another one as decoder when created with the :meth~transformers.AutoModel.from_pretrained class method for the encoder and :meth~transformers.AutoModelForCausalLM.from_pretrained class method for the decoder.
( pixel_values: Optional = None decoder_input_ids: Optional = None decoder_attention_mask: Optional = None encoder_outputs: Optional = None past_key_values: Optional = None decoder_inputs_embeds: Optional = None labels: Optional = None use_cache: Optional = None output_attentions: Optional = None output_hidden_states: Optional = None return_dict: Optional = None **kwargs ) → transformers.modeling_outputs.Seq2SeqLMOutput or tuple(torch.FloatTensor)
Parameters
torch.FloatTensor of shape (batch_size, num_channels, height, width)) —
Pixel values. 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. torch.LongTensor of shape (batch_size, target_sequence_length), optional) —
Indices of decoder input sequence tokens in the vocabulary.
Indices can be obtained using PreTrainedTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.
If past_key_values is used, optionally only the last decoder_input_ids have to be input (see
past_key_values).
For training, decoder_input_ids are automatically created by the model by shifting the labels to the
right, replacing -100 by the pad_token_id and prepending them with the decoder_start_token_id.
torch.BoolTensor of shape (batch_size, target_sequence_length), optional) —
Default behavior: generate a tensor that ignores pad tokens in decoder_input_ids. Causal mask will also
be used by default. tuple(torch.FloatTensor), optional) —
This tuple must consist of (last_hidden_state, optional: hidden_states, optional: attentions)
last_hidden_state (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size)) is a tensor
of hidden-states at the output of the last layer of the encoder. Used in the cross-attention of the
decoder. tuple(tuple(torch.FloatTensor)) of length config.n_layers with each tuple having 4 tensors of shape (batch_size, num_heads, sequence_length - 1, embed_size_per_head)) —
Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
If past_key_values are used, the user can optionally input only the last decoder_input_ids (those that
don’t have their past key value states given to this model) of shape (batch_size, 1) instead of all
decoder_input_ids of shape (batch_size, sequence_length).
torch.FloatTensor of shape (batch_size, target_sequence_length, hidden_size), optional) —
Optionally, instead of passing decoder_input_ids you can choose to directly pass an embedded
representation. This is useful if you want more control over how to convert decoder_input_ids indices
into associated vectors than the model’s internal embedding lookup matrix. torch.LongTensor of shape (batch_size, sequence_length), optional) —
Labels for computing the masked language modeling loss for the decoder. Indices should be in [-100, 0, ..., config.vocab_size] (see input_ids docstring) Tokens with indices set to -100 are ignored
(masked), the loss is only computed for the tokens with labels in [0, ..., config.vocab_size] bool, optional) —
If set to True, past_key_values key value states are returned and can be used to speed up decoding (see
past_key_values). bool, optional) —
Whether or not to return the attentions tensors of all attention layers. See attentions under returned
tensors for more detail. bool, optional) —
Whether or not to return the hidden states of all layers. See hidden_states under returned tensors for
more detail. bool, optional) —
If set to True, the model will return a ~utils.Seq2SeqLMOutput instead of a plain tuple. **encoder_kwargs for the encoder forward function.**decoder_kwargs for the decoder forward function.Returns
transformers.modeling_outputs.Seq2SeqLMOutput or tuple(torch.FloatTensor)
A transformers.modeling_outputs.Seq2SeqLMOutput 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 (VisionEncoderDecoderConfig) and inputs.
loss (torch.FloatTensor of shape (1,), optional, returned when labels is provided) — Language modeling loss.
logits (torch.FloatTensor of shape (batch_size, sequence_length, config.vocab_size)) — Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
past_key_values (tuple(tuple(torch.FloatTensor)), optional, returned when use_cache=True is passed or when config.use_cache=True) — Tuple of tuple(torch.FloatTensor) of length config.n_layers, with each tuple having 2 tensors of shape
(batch_size, num_heads, sequence_length, embed_size_per_head)) and 2 additional tensors of shape
(batch_size, num_heads, encoder_sequence_length, embed_size_per_head).
Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
blocks) that can be used (see past_key_values input) to speed up sequential decoding.
decoder_hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of torch.FloatTensor (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape (batch_size, sequence_length, hidden_size).
Hidden-states of the decoder at the output of each layer plus the initial embedding outputs.
decoder_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length).
Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the self-attention heads.
cross_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length).
Attentions weights of the decoder’s cross-attention layer, after the attention softmax, used to compute the weighted average in the cross-attention heads.
encoder_last_hidden_state (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size), optional) — Sequence of hidden-states at the output of the last layer of the encoder of the model.
encoder_hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of torch.FloatTensor (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape (batch_size, sequence_length, hidden_size).
Hidden-states of the encoder at the output of each layer plus the initial embedding outputs.
encoder_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length).
Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the self-attention heads.
The VisionEncoderDecoderModel 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:
>>> from transformers import AutoProcessor, VisionEncoderDecoderModel
>>> import requests
>>> from PIL import Image
>>> import torch
>>> processor = AutoProcessor.from_pretrained("microsoft/trocr-base-handwritten")
>>> model = VisionEncoderDecoderModel.from_pretrained("microsoft/trocr-base-handwritten")
>>> # load image from the IAM dataset
>>> url = "https://fki.tic.heia-fr.ch/static/img/a01-122-02.jpg"
>>> image = Image.open(requests.get(url, stream=True).raw).convert("RGB")
>>> # training
>>> model.config.decoder_start_token_id = processor.tokenizer.cls_token_id
>>> model.config.pad_token_id = processor.tokenizer.pad_token_id
>>> model.config.vocab_size = model.config.decoder.vocab_size
>>> pixel_values = processor(image, return_tensors="pt").pixel_values
>>> text = "hello world"
>>> labels = processor.tokenizer(text, return_tensors="pt").input_ids
>>> outputs = model(pixel_values=pixel_values, labels=labels)
>>> loss = outputs.loss
>>> # inference (generation)
>>> generated_ids = model.generate(pixel_values)
>>> generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]( encoder_pretrained_model_name_or_path: str = None decoder_pretrained_model_name_or_path: str = None *model_args **kwargs )
Parameters
str, optional) —
Information necessary to initiate the image encoder. Can be either:
google/vit-base-patch16-224-in21k../my_model_directory/../tf_model/model.ckpt.index). In
this case, from_tf should be set to True and a configuration object should be provided as
config argument. This loading path is slower than converting the TensorFlow checkpoint in a
PyTorch model using the provided conversion scripts and loading the PyTorch model afterwards.str, optional, defaults to None) —
Information necessary to initiate the text decoder. Can be either:
./my_model_directory/../tf_model/model.ckpt.index). In
this case, from_tf should be set to True and a configuration object should be provided as
config argument. This loading path is slower than converting the TensorFlow checkpoint in a
PyTorch model using the provided conversion scripts and loading the PyTorch model afterwards.__init__ method. output_attentions=True).
Behaves differently depending on whether a config is provided or automatically loaded.
Instantiate an encoder and a decoder from one or two base classes of the library from pretrained model checkpoints.
The model is set in evaluation mode by default using model.eval() (Dropout modules are deactivated). To train
the model, you need to first set it back in training mode with model.train().
Example:
>>> from transformers import VisionEncoderDecoderModel
>>> # initialize a vit-bert from a pretrained ViT and a pretrained BERT model. Note that the cross-attention layers will be randomly initialized
>>> model = VisionEncoderDecoderModel.from_encoder_decoder_pretrained(
... "google/vit-base-patch16-224-in21k", "google-bert/bert-base-uncased"
... )
>>> # saving model after fine-tuning
>>> model.save_pretrained("./vit-bert")
>>> # load fine-tuned model
>>> model = VisionEncoderDecoderModel.from_pretrained("./vit-bert")( config: Optional[PretrainedConfig] = None encoder: Optional[TFPreTrainedModel] = None decoder: Optional[TFPreTrainedModel] = None )
Parameters
This class can be used to initialize an image-to-text-sequence model with any pretrained vision autoencoding model as the encoder and any pretrained text autoregressive model as the decoder. The encoder is loaded via from_pretrained() function and the decoder is loaded via from_pretrained() function. Cross-attention layers are automatically added to the decoder and should be fine-tuned on a downstream generative task, like image captioning.
The effectiveness of initializing sequence-to-sequence models with pretrained checkpoints for sequence generation tasks was shown in Leveraging Pre-trained Checkpoints for Sequence Generation Tasks by Sascha Rothe, Shashi Narayan, Aliaksei Severyn. Michael Matena, Yanqi Zhou, Wei Li, Peter J. Liu.
Additionally, in TrOCR: Transformer-based Optical Character Recognition with Pre-trained Models it is shown how leveraging large pretrained vision models for optical character recognition (OCR) yields a significant performance improvement.
After such a Vision-Encoder-Text-Decoder 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 TF 2.0 Keras Model and refer to the TF 2.0 documentation for all matter related to general usage and behavior.
TFVisionEncoderDecoderModel is a generic model class that will be instantiated as a transformer architecture with one of the base vision model classes of the library as encoder and another one of the base model classes as decoder when created with the from_pretrained() class method for the encoder and from_pretrained() class method for the decoder.
( pixel_values: np.ndarray | tf.Tensor | None = None decoder_input_ids: np.ndarray | tf.Tensor | None = None decoder_attention_mask: np.ndarray | tf.Tensor | None = None encoder_outputs: Optional[Union[Tuple, TFBaseModelOutput]] = None past_key_values: Optional[Tuple[Tuple[Union[np.ndarray, tf.Tensor]]]] = None decoder_inputs_embeds: np.ndarray | tf.Tensor | None = None labels: np.ndarray | tf.Tensor | None = None use_cache: Optional[bool] = None output_attentions: Optional[bool] = None output_hidden_states: Optional[bool] = None return_dict: Optional[bool] = None training: bool = False **kwargs ) → transformers.modeling_tf_outputs.TFSeq2SeqLMOutput or tuple(tf.Tensor)
Parameters
np.ndarray, tf.Tensor, List[tf.Tensor] `Dict[str, tf.Tensor] or Dict[str, np.ndarray] and each example must have the shape (batch_size, num_channels, height, width)) —
Pixel values. Pixel values can be obtained using the vision’s model’s image processor. For example, using
AutoImageProcessor. See ViTImageProcessor.call() for details. np.ndarray or tf.Tensor of shape (batch_size, target_sequence_length), optional) —
Indices of decoder input sequence tokens in the vocabulary.
Indices can be obtained using PreTrainedTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.
If past_key_values is used, optionally only the last decoder_input_ids have to be input (see
past_key_values).
Provide for sequence to sequence training to the decoder. Indices can be obtained using PreTrainedTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.
np.ndarray or tf.Tensor of shape (batch_size, target_sequence_length), optional) —
Default behavior: generate a tensor that ignores pad tokens in decoder_input_ids. Causal mask will also
be used by default. tuple(tuple(tf.Tensor), optional) —
This tuple must consist of (last_hidden_state, optional: hidden_states, optional: attentions)
last_hidden_state (tf.Tensor of shape (batch_size, sequence_length, hidden_size)) is a tensor of hidden-states at the output
of the last layer of the encoder. Used in the cross-attention of the decoder. tuple(tuple(tf.Tensor)) of length config.n_layers with each tuple having 4 tensors of shape (batch_size, num_heads, sequence_length - 1, embed_size_per_head)) —
Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
If past_key_values are used, the user can optionally input only the last decoder_input_ids (those that
don’t have their past key value states given to this model) of shape (batch_size, 1) instead of all
decoder_input_ids of shape (batch_size, sequence_length).
np.ndarray or tf.Tensor of shape (batch_size, target_sequence_length, hidden_size), optional) —
Optionally, instead of passing decoder_input_ids you can choose to directly pass an embedded
representation. This is useful if you want more control over how to convert decoder_input_ids indices
into associated vectors than the model’s internal embedding lookup matrix. np.ndarray or tf.Tensor of shape (batch_size, sequence_length), optional) —
Labels for computing the masked language modeling loss for the decoder. Indices should be in [-100, 0, ..., config.vocab_size] (see input_ids docstring) Tokens with indices set to -100 are ignored
(masked), the loss is only computed for the tokens with labels in [0, ..., config.vocab_size] bool, optional) —
If set to True, past_key_values key value states are returned and can be used to speed up decoding (see
past_key_values). bool, optional) —
Whether or not to return the attentions tensors of all attention layers. See attentions under returned
tensors for more detail. bool, optional) —
Whether or not to return the hidden states of all layers. See hidden_states under returned tensors for
more detail. bool, optional) —
If set to True, the model will return a ~utils.Seq2SeqLMOutput instead of a plain tuple. bool, optional, defaults to False) —
Whether or not to use the model in training mode (some modules like dropout modules have different
behaviors between training and evaluation). **encoder_kwargs for the encoder forward function.**decoder_kwargs for the decoder forward function.Returns
transformers.modeling_tf_outputs.TFSeq2SeqLMOutput or tuple(tf.Tensor)
A transformers.modeling_tf_outputs.TFSeq2SeqLMOutput 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 (VisionEncoderDecoderConfig) and inputs.
loss (tf.Tensor of shape (n,), optional, where n is the number of non-masked labels, returned when labels is provided) — Language modeling loss.
logits (tf.Tensor of shape (batch_size, sequence_length, config.vocab_size)) — Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
past_key_values (List[tf.Tensor], optional, returned when use_cache=True is passed or when config.use_cache=True) — List of tf.Tensor of length config.n_layers, with each tensor of shape (2, batch_size, num_heads, sequence_length, embed_size_per_head)).
Contains pre-computed hidden-states (key and values in the attention blocks) of the decoder that can be
used (see past_key_values input) to speed up sequential decoding.
decoder_hidden_states (tuple(tf.Tensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of tf.Tensor (one for the output of the embeddings + one for the output of each layer) of shape
(batch_size, sequence_length, hidden_size).
Hidden-states of the decoder at the output of each layer plus the initial embedding outputs.
decoder_attentions (tuple(tf.Tensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of tf.Tensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length).
Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the self-attention heads.
cross_attentions (tuple(tf.Tensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of tf.Tensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length).
Attentions weights of the decoder’s cross-attention layer, after the attention softmax, used to compute the weighted average in the cross-attention heads.
encoder_last_hidden_state (tf.Tensor of shape (batch_size, sequence_length, hidden_size), optional) — Sequence of hidden-states at the output of the last layer of the encoder of the model.
encoder_hidden_states (tuple(tf.Tensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of tf.Tensor (one for the output of the embeddings + one for the output of each layer) of shape
(batch_size, sequence_length, hidden_size).
Hidden-states of the encoder at the output of each layer plus the initial embedding outputs.
encoder_attentions (tuple(tf.Tensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of tf.Tensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length).
Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the self-attention heads.
The TFVisionEncoderDecoderModel 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:
>>> from transformers import AutoImageProcessor, AutoTokenizer, TFVisionEncoderDecoderModel
>>> from PIL import Image
>>> import requests
>>> image_processor = AutoImageProcessor.from_pretrained("google/vit-base-patch16-224-in21k")
>>> decoder_tokenizer = AutoTokenizer.from_pretrained("openai-community/gpt2")
>>> # initialize a bert2gpt2 from a pretrained BERT and GPT2 models. Note that the cross-attention layers will be randomly initialized
>>> model = TFVisionEncoderDecoderModel.from_encoder_decoder_pretrained(
... "google/vit-base-patch16-224-in21k", "openai-community/gpt2"
... )
>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
>>> img = Image.open(requests.get(url, stream=True).raw)
>>> # forward
>>> pixel_values = image_processor(images=img, return_tensors="tf").pixel_values # Batch size 1
>>> decoder_input_ids = decoder_tokenizer("Linda Davis", return_tensors="tf").input_ids # Batch size 1
>>> outputs = model(pixel_values=pixel_values, decoder_input_ids=decoder_input_ids)
>>> # training
>>> outputs = model(pixel_values=pixel_values, decoder_input_ids=decoder_input_ids, labels=decoder_input_ids)
>>> loss, logits = outputs.loss, outputs.logits
>>> # save and load from pretrained
>>> model.save_pretrained("vit-gpt2")
>>> model = TFVisionEncoderDecoderModel.from_pretrained("vit-gpt2")
>>> # generation
>>> generated = model.generate(pixel_values, decoder_start_token_id=model.config.decoder.bos_token_id)( encoder_pretrained_model_name_or_path: str = None decoder_pretrained_model_name_or_path: str = None *model_args **kwargs )
Parameters
str, optional) —
Information necessary to initiate the encoder. Can be either:
google/vit-base-patch16-224-in21k../my_model_directory/../pt_model/). In this case,
encoder_from_pt should be set to True.str, optional, defaults to None) —
Information necessary to initiate the decoder. Can be either:
./my_model_directory/../pt_model/). In this case,
decoder_from_pt should be set to True.__init__ method. output_attentions=True).
Behaves differently depending on whether a config is provided or automatically loaded.
Instantiate an encoder and a decoder from one or two base classes of the library from pretrained model checkpoints.
Example:
>>> from transformers import TFVisionEncoderDecoderModel
>>> # initialize a vit-bert from a pretrained ViT and a pretrained BERT model. Note that the cross-attention layers will be randomly initialized
>>> model = TFVisionEncoderDecoderModel.from_encoder_decoder_pretrained(
... "google/vit-base-patch16-224-in21k", "google-bert/bert-base-uncased"
... )
>>> # saving model after fine-tuning
>>> model.save_pretrained("./vit-bert")
>>> # load fine-tuned model
>>> model = TFVisionEncoderDecoderModel.from_pretrained("./vit-bert")( config: VisionEncoderDecoderConfig input_shape: Optional = None seed: int = 0 dtype: dtype = <class 'jax.numpy.float32'> _do_init: bool = True **kwargs )
Parameters
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 an image-to-text-sequence model with any pretrained vision autoencoding model as the encoder and any pretrained text autoregressive model as the decoder. The encoder is loaded via from_pretrained() function and the decoder is loaded via from_pretrained() function. Cross-attention layers are automatically added to the decoder and should be fine-tuned on a downstream generative task, like image captioning.
The effectiveness of initializing sequence-to-sequence models with pretrained checkpoints for sequence generation tasks was shown in Leveraging Pre-trained Checkpoints for Sequence Generation Tasks by Sascha Rothe, Shashi Narayan, Aliaksei Severyn. Michael Matena, Yanqi Zhou, Wei Li, Peter J. Liu.
Additionally, in TrOCR: Transformer-based Optical Character Recognition with Pre-trained Models it is shown how leveraging large pretrained vision models for optical character recognition (OCR) yields a significant performance improvement.
After such a Vision-Encoder-Text-Decoder 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 FlaxPreTrainedModel. 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.nn.Module subclass. Use it as a regular Flax Module and refer to the Flax documentation for all matter related to general usage and behavior.
FlaxVisionEncoderDecoderModel is a generic model class that will be instantiated as a transformer architecture with the module (flax.nn.Module) of one of the base vision model classes of the library as encoder module and another one as decoder module when created with the :meth~transformers.FlaxAutoModel.from_pretrained class method for the encoder and :meth~transformers.FlaxAutoModelForCausalLM.from_pretrained class method for the decoder.
( pixel_values: Array decoder_input_ids: Optional = None decoder_attention_mask: Optional = None decoder_position_ids: Optional = None output_attentions: Optional = None output_hidden_states: Optional = None return_dict: Optional = None train: bool = False params: dict = None dropout_rng: PRNGKey = None ) → transformers.modeling_flax_outputs.FlaxSeq2SeqLMOutput or tuple(torch.FloatTensor)
Parameters
jnp.ndarray of shape (batch_size, num_channels, height, width)) —
Pixel values. Pixel values can be obtained using the vision model’s image processor. For example, using
AutoImageProcessor. See ViTImageProcessor.call() for details. jnp.ndarray of shape (batch_size, target_sequence_length), optional) —
Indices of decoder input sequence tokens in the vocabulary.
Indices can be obtained using PreTrainedTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.
jnp.ndarray of shape (batch_size, target_sequence_length), optional) —
Default behavior: generate a tensor that ignores pad tokens in decoder_input_ids. Causal mask will also
be used by default. jnp.ndarray of shape (batch_size, sequence_length), optional) —
Indices of positions of each decoder input sequence tokens in the position embeddings. Selected in the
range [0, config.decoder.max_position_embeddings - 1]. bool, optional) —
Whether or not to return the attentions tensors of all attention layers. See attentions under returned
tensors for more detail. bool, optional) —
Whether or not to return the hidden states of all layers. See hidden_states under returned tensors for
more detail. bool, optional) —
If set to True, the model will return a ~utils.FlaxSeq2SeqLMOutput instead of a plain tuple. Returns
transformers.modeling_flax_outputs.FlaxSeq2SeqLMOutput or tuple(torch.FloatTensor)
A transformers.modeling_flax_outputs.FlaxSeq2SeqLMOutput 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 (VisionEncoderDecoderConfig) and inputs.
logits (jnp.ndarray of shape (batch_size, sequence_length, config.vocab_size)) — Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
past_key_values (tuple(tuple(jnp.ndarray)), optional, returned when use_cache=True is passed or when config.use_cache=True) — Tuple of tuple(jnp.ndarray) of length config.n_layers, with each tuple having 2 tensors of shape
(batch_size, num_heads, sequence_length, embed_size_per_head)) and 2 additional tensors of shape
(batch_size, num_heads, encoder_sequence_length, embed_size_per_head).
Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
blocks) that can be used (see past_key_values input) to speed up sequential decoding.
decoder_hidden_states (tuple(jnp.ndarray), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of jnp.ndarray (one for the output of the embeddings + one for the output of each layer) of shape
(batch_size, sequence_length, hidden_size).
Hidden-states of the decoder at the output of each layer plus the initial embedding outputs.
decoder_attentions (tuple(jnp.ndarray), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of jnp.ndarray (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length).
Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the self-attention heads.
cross_attentions (tuple(jnp.ndarray), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of jnp.ndarray (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length).
Attentions weights of the decoder’s cross-attention layer, after the attention softmax, used to compute the weighted average in the cross-attention heads.
encoder_last_hidden_state (jnp.ndarray of shape (batch_size, sequence_length, hidden_size), optional) — Sequence of hidden-states at the output of the last layer of the encoder of the model.
encoder_hidden_states (tuple(jnp.ndarray), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of jnp.ndarray (one for the output of the embeddings + one for the output of each layer) of shape
(batch_size, sequence_length, hidden_size).
Hidden-states of the encoder at the output of each layer plus the initial embedding outputs.
encoder_attentions (tuple(jnp.ndarray), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of jnp.ndarray (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length).
Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the self-attention heads.
The FlaxVisionEncoderDecoderModel 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:
>>> from transformers import FlaxVisionEncoderDecoderModel, AutoImageProcessor, AutoTokenizer
>>> from PIL import Image
>>> import requests
>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
>>> image = Image.open(requests.get(url, stream=True).raw)
>>> image_processor = AutoImageProcessor.from_pretrained("google/vit-base-patch16-224-in21k")
>>> # load output tokenizer
>>> tokenizer_output = AutoTokenizer.from_pretrained("openai-community/gpt2")
>>> # initialize a vit-gpt2 from pretrained ViT and GPT2 models. Note that the cross-attention layers will be randomly initialized
>>> model = FlaxVisionEncoderDecoderModel.from_encoder_decoder_pretrained(
... "google/vit-base-patch16-224-in21k", "openai-community/gpt2"
... )
>>> pixel_values = image_processor(images=image, return_tensors="np").pixel_values
>>> # use GPT2's eos_token as the pad as well as eos token
>>> model.config.eos_token_id = model.config.decoder.eos_token_id
>>> model.config.pad_token_id = model.config.eos_token_id
>>> # generation
>>> sequences = model.generate(pixel_values, num_beams=4, max_length=12).sequences
>>> captions = tokenizer_output.batch_decode(sequences, skip_special_tokens=True)( encoder_pretrained_model_name_or_path: Union = None decoder_pretrained_model_name_or_path: Union = None *model_args **kwargs )
Parameters
Union[str, os.PathLike], optional) —
Information necessary to initiate the encoder. Can be either:
google/vit-base-patch16-224-in21k../my_model_directory/.Union[str, os.PathLike], optional, defaults to None) —
Information necessary to initiate the decoder. Can be either:
./my_model_directory/.__init__ method. output_attentions=True).
Behaves differently depending on whether a config is provided or automatically loaded.
Instantiate an encoder and a decoder from one or two base classes of the library from pretrained model checkpoints.
Example:
>>> from transformers import FlaxVisionEncoderDecoderModel
>>> # initialize a vit-gpt2 from a pretrained ViT and a pretrained GPT2 model. Note that the cross-attention layers will be randomly initialized
>>> model = FlaxVisionEncoderDecoderModel.from_encoder_decoder_pretrained(
... "google/vit-base-patch16-224-in21k", "openai-community/gpt2"
... )
>>> # saving model after fine-tuning
>>> model.save_pretrained("./vit-gpt2")
>>> # load fine-tuned model
>>> model = FlaxVisionEncoderDecoderModel.from_pretrained("./vit-gpt2")