Blocks
The eight prebuilt blocks on the Blocks tab, and how to create, expand, and reuse your own.
Blocks A block groups several nodes into one, so a common pattern an attention sub-layer, a U-Net stage drops onto the canvas as a single tile. Studio ships eight prebuilt blocks, and you can turn any selection of nodes into a block of your own. Prebuilt blocks The Blocks tab in the left sidebar lists these eight under Pre-built Blocks : Label What it does ----------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------- SE Block Recalibrates channel-wise feature responses using global average pooling and a gating mechanism. FFN Block A feed-forward network with a SwiGLU activation and dropout, the standard sub-layer used in Transformer architectures. Transformer Encoder Block Multi-head self-attention, a feed-forward sub-layer, layer normalization, and residual connections. Transformer Decoder Block Causal self-attention, cross-attention over an encoder's output, a feed-forward sub-layer, and three residual connections. U-Net Down Block Two convolution, batch normalization, and activation stages followed by max pooling, for a U-Net's encoder path. U-Net Mid Block Two convolution, batch normalization, and activation stages with optional cross-attention, between a U-Net's encoder and decoder paths. U-Net Up Block A transposed convolution, a skip-connection concatenation, and two convolution, batch normalization, and activation stages, for a U-Net's decoder path. Audio Encoder Block Three 1D convolutional layers with activation and batch normalization, followed by global average pooling, for audio feature extraction. Create a block Choose Create Block . , < Type a Block name and click Create . , / The selected nodes must be connected to each other; otherwise Studio shows "Selected nodes must be connected to form a block". The new block replaces the selected nodes on the canvas and appears under Custom Blocks on the Blocks tab. Expand, compact, and ungroup Right click a block for Expand Block to see its internal nodes, or Compact Block to collapse it back once expanded. Unblock Nodes ungroups a block back into its individual nodes on the canvas. Save and reuse your blocks Drag a block from Custom Blocks onto the canvas to reuse it in this project, the same way as a prebuilt one. Custom blocks are saved with the project when you save a version, and load with it when you open it again. Right click a block on the canvas and choose Save as new block⦠to copy its current graph into a new, independent block under a name you give it; the original block is left untouched. To rename or delete a custom block, right click it under Custom Blocks and choose Rename or Delete . Delete asks you to confirm in a Delete Block dialog; deleting a block does not remove the block nodes already on a canvas, but leaves them unresolved, so remove those first if you no longer want them.
Open in Dagnam.AI docs