In the wake of all the rumours, re-reading Multimodal Llama in the Llama 3 paper today! 🦙
Comprises of three main components
1. Image Encoder
2. Image Adapter
3. Video Adapter
Image Encoder
> A ViT-H/14 variant of the vision transformer with 630M parameters, trained on 2.5B image-text pairs
> Processes images of resolution 224x224, divided into 16x16 patches of 14x14 pixels each
> Leverages multi-layer feature extraction utilizing features from the 4th, 8th, 16th, 24th, and 31st layers, in addition to the final layer features
> 8 gated self-attention layers, resulting in a total of 40 transformer blocks, and with 850M total parameters
> Encoder generates a 7680-dimensional representation for each patch, producing a total of 256 patches
Image Adapter
> Cross-attention layers b/w visual and language model token representations, applied after every fourth self-attention layer in the language model, utilizing GQA
> Initial pre-training: Trained on ≈6B image-text pairs, with images resized to fit within four tiles of 336x336 pixels, arranged to accommodate various aspect ratios.
> Annealing: Training continues on ≈500M images, increasing per-tile resolution to enhance performance on downstream tasks
Video Adapter
> Input: Up to 64 video frames, each encoded
> Temporal modeling: Combines 32 consecutive frames and adds video cross-attention layers
> Aggregator: Implemented as a perceiver resampler
> Parameters: 0.6B and 4.6B for Llama 3 7B and 70B, respectively
If/ when it gets released - it'd be a leap forward for VLMs! 🤗
