Conv1d with asymmetric or causal padding and normalization.

Conv1d with asymmetric or causal padding and normalization.

Creates a ResNet model for CIFAR-10, as specified in the original paper.

An attention layer that allows for downscaling the size of the embedding after projection to queries, keys, and values.

Implements the Keyword Transformer (KWT) [1] model. KWT is essentially a vision transformer [2] with minor modifications: - Instead of s

A structure for storing masks and their related data in batched format. Implements basic filtering and concatenation.

A convolutional layer that upsamples the input by a factor of 2. MLX does not yet support transposed convolutions, so we approximate them wit

A single head unmasked attention for use with the VAE.

Residual block from SEANet model as used by EnCodec.

Residual block from SEANet model as used by EnCodec.

A transformer model for inputs with 2 spatial dimensions.

Multi-head Attention block with relative position embeddings.

The autoencoder that allows us to perform diffusion in the latent space.

Transformer blocks with support of window attention and residual propagation blocks

Pre-norm transformer block with self-attention and MLP.

The transformer encoder layer from CLIP.

The transformer encoder layer from CLIP.

A simple port of https://github.com/huggingface/transformers/blob/main/src/transformers/models/clip/image_processing_clip.py.