API Reference¶

Complete reference for the public API of the dg SDK. Grouped by namespace.

Models¶

Class	Description
`DLGModel`	Base class for DeepGate models. Subclass and override `build_model_graph()`.

Class	Description
`Linear`	Affine linear transformation with optional ReLU.
`Norm`	Element-wise input normalization. Casts integer inputs to `float32`.
`UpperTriangularLinear`	Linear with upper-triangular weights, in-place-friendly.
`LowerTriangularLinear`	Linear with lower-triangular weights, in-place-friendly.

Class	Description
`QuantLinear`	Quantized affine linear transformation.
`QuantUpperTriangularLinear`	Quantized `UpperTriangularLinear`.
`QuantLowerTriangularLinear`	Quantized `LowerTriangularLinear`.
`QuantConv2d`	Quantized 2D convolution with optional BN folding and ReLU fusion.
`QuantDepthwiseConv2d`	Quantized depthwise 2D convolution.
`QuantAvgPool2d`	Quantized 2D average pooling; shares quantization params with its input.
`QuantResidualAdd`	Quantized residual block with two parallel streams.

Class	Description
`Flatten`	`NCHW → NHWC` permute then flatten.
`Pad`	Right-pad the last dimension with zeros.
`Slice`	Keep the first `keep_bits` entries of the last dimension.
`Expand`	Repeat the last dimension `repeat` times.
`BlockShuffle`	Fixed-permutation block shuffle of the last dimension.
`MaxPool2d`	Non-quantized 2D max pooling.
`BitShift`	Bit-shift with integer truncation and straight-through gradients.
`Reshape`	`NCHW → NHWC` permute then reshape to a target shape.
`Concat`	Concatenate a list of tensors along a non-batch dimension.

Symbol	Description
`Train`	Orchestrates training over epochs. Alias for `dg.train.Trainer`.
`freeze_backbone`	Freeze every layer except the head, for fine-tuning.

Function	Description
`get_acc`	Top-1 classification accuracy over a dataset.
`compute_metric`	Generic per-batch metric loop under `no_grad`.

Function	Description
`from_pretrained`	Load a model from a local directory or from the DeepGate hub.