Mistral3VisionEncoder classkeras_hub.models.Mistral3VisionEncoder(
image_size=1540,
patch_size=14,
num_channels=3,
hidden_dim=1024,
num_layers=24,
num_heads=16,
head_dim=64,
intermediate_dim=4096,
rope_theta=10000.0,
layer_norm_epsilon=1e-05,
activation="silu",
attention_dropout=0.0,
**kwargs
)
Vision encoder used by Mistral3.
This is not exposed as a standalone model. It is the vision tower consumed by the Mistral3 multimodal architecture.
pixel_values follows the Hugging Face Pixtral layout:
(num_images, num_channels, height, width).
When image_sizes is provided, each image is cropped to its effective
image size after patchification, the resulting patch sequences are
concatenated into a single sequence, and attention is restricted to
patches belonging to the same image.