基本配置

print(torch.version)

torch.set_num_threads(1) print(torch.config.parallel_info())

模型操作

模型文件
- 有多种不同的叫法，bin/pt/pth，其本质都是zip压缩形式的pickle数据。
模型导出
模型加载

Tensor数据类型

https://pytorch.org/docs/stable/tensors.html torch.Tensor 根据dtype和device参数的不同，可以表达几十种数据类型

nn.Linear

nn.Embedding

Tensor 操作

生成：torch.empty, torch.tensor, torch.rand, torch.zeros

Resize: torch.view

reshape

CUDA Tensor: tensor.to可以将一个tensor转移到指定的device上

register_buffer

unsqueeze 在指定的位置增维度

einsum 爱因斯坦求和

bmm 三维tensor的矩阵乘法运算，矩阵必须是3维数据

masked_fill(mask, value) 将mask中为True的元素所对应的基础tensor元素设置为值value

tril 返回一个矩阵主对角线以下的下三角矩阵，其它元素全部为0

expand

view

permute

contiguous

custom PyTorch OP implementation

import torch
from torch.autograd import Function
from torch.utils.cpp_extension import load

# 加载自定义操作的C++扩展模块
custom_op = load(name='custom_op', sources=['custom_op.cpp'])

# 自定义操作的前向传播和反向传播函数
class CustomOpFunction(Function):
    @staticmethod
    def forward(ctx, input):
        output = custom_op.custom_forward(input)  # 调用C++扩展模块的前向传播函数
        ctx.save_for_backward(input)  # 保存输入，以备反向传播使用
        return output

    @staticmethod
    def backward(ctx, grad_output):
        input, = ctx.saved_tensors  # 获取保存的输入
        grad_input = custom_op.custom_backward(grad_output, input)  # 调用C++扩展模块的反向传播函数
        return grad_input

# 使用自定义操作
input = torch.randn(3, 3, requires_grad=True)
custom_op_function = CustomOpFunction.apply
output = custom_op_function(input)
loss = output.sum()
loss.backward()

print("Input:", input)
print("Output:", output)
print("Gradient:", input.grad)

自动求导

TORCH.FX

模型量化

分法1
- Dynamic Quantization
- Post-Training Static Quantization(PTQ)
- Quantization Aware Training(QAT)
分法2
- Eager Mode Quantization
- FX Graph Mode Quantization

Dynamic Quantization

在推理过程中跟踪输入数据的分布来动态地量化权重. 方法：weights quantized with activations read/stored in floating point and quantized for compute.

quantized_model = torch.quantization.quantize_dynamic(
    model, {torch.nn.Linear}, dtype=torch.qint8
)
print(quantized_model)

Post-Training Static Quantization(PTQ)

# set quantization config for server (x86)
deploymentmyModel.qconfig = torch.quantization.get_default_config('fbgemm')

# insert observers
torch.quantization.prepare(myModel, inplace=True)
# Calibrate the model and collect statistics

# convert to quantized version
torch.quantization.convert(myModel, inplace=True)

Quantization Aware Training(QAT)

# specify quantization config for QAT
qat_model.qconfig=torch.quantization.get_default_qat_qconfig('fbgemm')

# prepare QAT
torch.quantization.prepare_qat(qat_model, inplace=True)

# convert to quantized version, removing dropout, to check for accuracy on each epoch
quantized_model=torch.quantization.convert(qat_model.eval(), inplace=False)