Conv2dQAT

Applicability

Product	Supported
Atlas A3 training series products/Atlas A3 inference series products	√
Atlas A2 training products/Atlas A2 inference products	√
Atlas 200I/500 A2 inference product	√
Atlas inference series products	√
Atlas training products	√

Description

Constructs the QAT operator of Conv2d.

Prototype

API for construction from scratch:

qat = amct_pytorch.nn.module.quantization.conv2d.Conv2dQAT(in_channels, out_channels, kernel_size, stride, padding, dilation, groups, bias, padding_mode, device, dtype, config)

API for construction based on the native operator:

qat = amct_pytorch.nn.module.quantization.conv2d.Conv2dQAT.from_float(mod, config)

Parameters

**Table 1** Parameters in the API for operator construction from scratch
Parameter	Input/Output	Description
in_channels	Input	Number of input channels. An int.
out_channels	Input	Number of output channels. An int.
kernel_size	Input	Size of the convolution kernel. An int or a tuple.
stride	Input	Convolution stride. An int or a tuple. Default: 1
padding	Input	Padding size. An int or a tuple. Default: 0
dilation	Input	Spacing between kernel elements. An int or a tuple. Default: 1
groups	Input	Connections between the inputs and outputs. An int. Default: 1
bias	Input	Whether to enable bias items to participate in learning. A bool. Other data types (such as integers, strings, and lists) are converted based on the Python truth value judgment rules. Default: True
padding_mode	Input	Padding mode. Restrictions: Must be zeros.
device	Input	Running device. Default: None
dtype	Input	Torch data type. Torch data type. Only torch.float32 is supported.
config	Input	Quantization configuration. The following is a configuration example. For details about quantization configuration parameters, see Quantization Configuration Parameters. config = { "retrain_enable":true, "retrain_data_config": { "dst_type": "INT8", "batch_num": 10, "fixed_min": False, "clip_min": -1.0, "clip_max": 1.0 }, "retrain_weight_config": { "dst_type": "INT8", "weights_retrain_algo": "arq_retrain", "channel_wise": False } } A dict. Default: None

**Table 2** Parameters in the API for construction based on the native operator
Parameter	Input/Output	Description
mod	Input	Native Conv2d operator to be quantized. A torch.nn.Module.
config	Input	Quantization configuration. The following is a configuration example. For details about quantization configuration parameters, see Quantization Configuration Parameters. config = { "retrain_enable":true, "retrain_data_config": { "dst_type": "INT8", "batch_num": 10, "fixed_min": False, "clip_min": -1.0, "clip_max": 1.0 }, "retrain_weight_config": { "dst_type": "INT8", "weights_retrain_algo": "arq_retrain", "channel_wise": False } } A dict. Default: None

Returns

Construction from scratch: returns the constructed QAT single-operator instance.
Construction based on native operators: returns the QAT single-operator converted from torch.nn.Module.

Example

Construction from scratch:

from amct_pytorch.nn.module.quantization.conv2d import Conv2dQAT

Conv2dQAT(in_channels=1, out_channels=1, kernel_size=1, stride=1,
          padding=0, dilation=1, groups=1, bias=True,
          padding_mode='zeros', device=None, dtype=None, config=None)

Construction based on the native operator:

import torch

from amct_pytorch.nn.module.quantization.conv2d import Conv2dQAT

conv2d_op = torch.nn.Conv2d(in_channels=1, out_channels=1, kernel_size=1, stride=1,
                            padding=0, dilation=1, groups=1, bias=True,
                            padding_mode='zeros', device=None, dtype=None)
Conv2dQAT.from_float(mod=conv2d_op, config=None)

Parent topic: Single-Operator Mode