Skip to content

C
chatCSDN

汀、人工智能 / chatCSDN
与 Fork 源项目一致

Fork自 CSDN 技术社区 / ai / chatCSDN

通知 1
Star 0
Fork 0
C
chatCSDN
提交 1cdafdef 编写于 作者: CSDN-Ada助手's avatar CSDN-Ada助手
浏览文件
操作

add blog url

上级 a0f6ec61
显示空白变更内容
内联 并排
Showing with 2 addition and 165 deletion
README.md
浏览文件 @ 1cdafdef
......@@ -4,6 +4,8 @@ ChatCSDN基于RWKV1.5B基模型
源码来源于：https://github.com/BlinkDL/RWKV-LM/tree/main/RWKV-v4neo
主要是在RWKV提供的1.5B参数的基础之上，使用CSDN的问答数据和博客数据进行增量预训练，经过指令微调，得到拥有IT行业知识体系的大语言模型。
原始模型参数地址：https://huggingface.co/BlinkDL/rwkv-4-pile-1b5
微调后的模型参数地址：https://huggingface.co/zxm2023/ChatCSDN
介绍博客地址：https://blog.csdn.net/zxm2015/article/details/130227450
## 预处理数据
使用项目 https://github.com/EleutherAI/gpt-neox 提供的数据转换工具将.jsonl文件转换为.bin和.idx文件，目前已经将代码全部移植过来，放在tools文件夹中。词典使用的是20B_tokenizer.json。
......
img_demoAE.py 已删除 100644 → 0
浏览文件 @ a0f6ec61
########################################################################################################
# The RWKV Language Model - https://github.com/BlinkDL/RWKV-LM
########################################################################################################
import torch, types, os
import numpy as np
from PIL import Image
import torch.nn as nn
from torch.nn import functional as F
import torchvision as vision
import torchvision.transforms as transforms
np.set_printoptions(precision=4, suppress=True, linewidth=200)
print(f'loading...')
########################################################################################################
model_prefix = 'test/image_trained/out-v7c_d8_256-224-13bit-OB32x0.5-201'
input_img = 'test/img_ae_test/test0.png'
########################################################################################################
class ToBinary(torch.autograd.Function):
@staticmethod
def forward(ctx, x):
return torch.floor(x + 0.5) # no need for noise when we have plenty of data
@staticmethod
def backward(ctx, grad_output):
return grad_output.clone() # pass-through
class R_ENCODER(nn.Module):
def __init__(self, args):
super().__init__()
self.args = args
dd = 8
self.Bxx = nn.BatchNorm2d(dd*64)
self.CIN = nn.Conv2d(3, dd, kernel_size=3, padding=1)
self.Cx0 = nn.Conv2d(dd, 32, kernel_size=3, padding=1)
self.Cx1 = nn.Conv2d(32, dd, kernel_size=3, padding=1)
self.B00 = nn.BatchNorm2d(dd*4)
self.C00 = nn.Conv2d(dd*4, 256, kernel_size=3, padding=1)
self.C01 = nn.Conv2d(256, dd*4, kernel_size=3, padding=1)
self.C02 = nn.Conv2d(dd*4, 256, kernel_size=3, padding=1)
self.C03 = nn.Conv2d(256, dd*4, kernel_size=3, padding=1)
self.B10 = nn.BatchNorm2d(dd*16)
self.C10 = nn.Conv2d(dd*16, 256, kernel_size=3, padding=1)
self.C11 = nn.Conv2d(256, dd*16, kernel_size=3, padding=1)
self.C12 = nn.Conv2d(dd*16, 256, kernel_size=3, padding=1)
self.C13 = nn.Conv2d(256, dd*16, kernel_size=3, padding=1)
self.B20 = nn.BatchNorm2d(dd*64)
self.C20 = nn.Conv2d(dd*64, 256, kernel_size=3, padding=1)
self.C21 = nn.Conv2d(256, dd*64, kernel_size=3, padding=1)
self.C22 = nn.Conv2d(dd*64, 256, kernel_size=3, padding=1)
self.C23 = nn.Conv2d(256, dd*64, kernel_size=3, padding=1)
self.COUT = nn.Conv2d(dd*64, args.my_img_bit, kernel_size=3, padding=1)
def forward(self, img):
ACT = F.mish
x = self.CIN(img)
xx = self.Bxx(F.pixel_unshuffle(x, 8))
x = x + self.Cx1(ACT(self.Cx0(x)))
x = F.pixel_unshuffle(x, 2)
x = x + self.C01(ACT(self.C00(ACT(self.B00(x)))))
x = x + self.C03(ACT(self.C02(x)))
x = F.pixel_unshuffle(x, 2)
x = x + self.C11(ACT(self.C10(ACT(self.B10(x)))))
x = x + self.C13(ACT(self.C12(x)))
x = F.pixel_unshuffle(x, 2)
x = x + self.C21(ACT(self.C20(ACT(self.B20(x)))))
x = x + self.C23(ACT(self.C22(x)))
x = self.COUT(x + xx)
return torch.sigmoid(x)
class R_DECODER(nn.Module):
def __init__(self, args):
super().__init__()
self.args = args
dd = 8
self.CIN = nn.Conv2d(args.my_img_bit, dd*64, kernel_size=3, padding=1)
self.B00 = nn.BatchNorm2d(dd*64)
self.C00 = nn.Conv2d(dd*64, 256, kernel_size=3, padding=1)
self.C01 = nn.Conv2d(256, dd*64, kernel_size=3, padding=1)
self.C02 = nn.Conv2d(dd*64, 256, kernel_size=3, padding=1)
self.C03 = nn.Conv2d(256, dd*64, kernel_size=3, padding=1)
self.B10 = nn.BatchNorm2d(dd*16)
self.C10 = nn.Conv2d(dd*16, 256, kernel_size=3, padding=1)
self.C11 = nn.Conv2d(256, dd*16, kernel_size=3, padding=1)
self.C12 = nn.Conv2d(dd*16, 256, kernel_size=3, padding=1)
self.C13 = nn.Conv2d(256, dd*16, kernel_size=3, padding=1)
self.B20 = nn.BatchNorm2d(dd*4)
self.C20 = nn.Conv2d(dd*4, 256, kernel_size=3, padding=1)
self.C21 = nn.Conv2d(256, dd*4, kernel_size=3, padding=1)
self.C22 = nn.Conv2d(dd*4, 256, kernel_size=3, padding=1)
self.C23 = nn.Conv2d(256, dd*4, kernel_size=3, padding=1)
self.Cx0 = nn.Conv2d(dd, 32, kernel_size=3, padding=1)
self.Cx1 = nn.Conv2d(32, dd, kernel_size=3, padding=1)
self.COUT = nn.Conv2d(dd, 3, kernel_size=3, padding=1)
def forward(self, code):
ACT = F.mish
x = self.CIN(code)
x = x + self.C01(ACT(self.C00(ACT(self.B00(x)))))
x = x + self.C03(ACT(self.C02(x)))
x = F.pixel_shuffle(x, 2)
x = x + self.C11(ACT(self.C10(ACT(self.B10(x)))))
x = x + self.C13(ACT(self.C12(x)))
x = F.pixel_shuffle(x, 2)
x = x + self.C21(ACT(self.C20(ACT(self.B20(x)))))
x = x + self.C23(ACT(self.C22(x)))
x = F.pixel_shuffle(x, 2)
x = x + self.Cx1(ACT(self.Cx0(x)))
x = self.COUT(x)
return torch.sigmoid(x)
########################################################################################################
print(f'building model...')
args = types.SimpleNamespace()
args.my_img_bit = 13
encoder = R_ENCODER(args).eval().cuda()
decoder = R_DECODER(args).eval().cuda()
zpow = torch.tensor([2**i for i in range(0,13)]).reshape(13,1,1).cuda().long()
encoder.load_state_dict(torch.load(f'{model_prefix}-E.pth'))
decoder.load_state_dict(torch.load(f'{model_prefix}-D.pth'))
########################################################################################################
print(f'test image...')
img_transform = transforms.Compose([
transforms.PILToTensor(),
transforms.ConvertImageDtype(torch.float),
transforms.Resize((224, 224))
])
with torch.no_grad():
img = img_transform(Image.open(input_img)).unsqueeze(0).cuda()
z = encoder(img)
z = ToBinary.apply(z)
zz = torch.sum(z.squeeze().long() * zpow, dim=0)
print(f'Code shape = {zz.shape}\n{zz.cpu().numpy()}\n')
out = decoder(z)
vision.utils.save_image(out, f"{input_img.split('.')[0]}-out-13bit.jpg")
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或