Pretrained CLIP-Vision-Marian pre-trained on a subset of Spanish-translated Conceptual-12M image-text pairs using a seq2seq model training objective. 2.5M cleaned English image-text pairs are translated using Spanish Marian Model. We trained CLIP-Vision-Marian model during community week hosted by Huggingface 🤗 using JAX/Flax.

CLIP-Vision-Marian is a modified transformers model which takes in visual embeddings from CLIP-Vision transformer and feeds into the encoder_hidden_states of a Marian decoder. This is done for deep cross-modal interaction via cross-attention between the two modes. The decoder then predicts logits for the input_ids provided and can be used for generation.

You can use the raw model for encoder-decoder network where you want the encoder to encode images and the decoder to decode text.

Note that this model is primarily aimed at being fine-tuned on tasks like Spanish image captioning.

You will need to clone the model from here. An example of usage is shown below:

The Spanish image captioning model was trained on a subset of Conceptual 12M dataset by Google:

Conceptual 12M, Introduced by Changpinyo et al. in Conceptual 12M: Pushing Web-Scale Image-Text Pre-Training To Recognize Long-Tail Visual Concepts.

The translated dataset can be downloaded from conceptual-12m-multilingual-marian-es. We do not provide images as we do not own any of them. One can download images from the image_url section of the original Conceptual 12M dataset.

Though the original dataset contains 12M image-text pairs, a lot of the URLs are invalid now, and in some cases, images are corrupt or broken. We remove such examples from our data, which leaves us with approximately 10M image-text pairs, out of which we took only 2.5M image, caption pairs.

Total data:
2475000 captions
2475000 images

Total data:
25000 captions
25000 images

The model was trained on Google Cloud Engine TPUv3-8 machine (with 335 GB of RAM, 1000 GB of hard drive, 96 CPU cores) 8 v3 TPU cores for 42K steps with a batch size of 128 and a sequence length of 128. The optimizer used is Adam with a learning rate of 3e-4, β1 = 0.9, β2 = 0.98 and ε = 1e-8, a weight decay of 0.01, learning rate warmup for 1,000 steps and linear decay of the learning rate after.

We tracked experiments using Tensorboard which can be found in Training Metrics tab.

Our model reached eval loss of ~3.1 around ~20K steps. Here are the BLEU^ scores for different languages:

^BLEU scores are out of 1

You can try out our model on 🤗 Huggingface's spaces 🪐 : Streamlit app of Spanish Image Captioning model on Huggingface Spaces

Thanks to Huggingface 🤗 & Google JAX/Flax team for such a wonderful community week. Big thanks to @patrickvonplaten and @patil-suraj for helping us with our solution during the community week.

<img src=https://pbs.twimg.com/media/E443fPjX0AY1BsR.jpg:large>