84
84
5 . <img src =" img/bilibili32.png " width =" 18 " /> [ Transformer] ( https://www.bilibili.com/video/BV1pu411o7BE )
85
85
6 . <img src =" img/bilibili32.png " width =" 18 " /> [ 对比学习] ( https://www.bilibili.com/video/BV19S4y1M7hm )
86
86
87
- ### PyTorch
87
+ ### 深度学习框架: PyTorch
88
88
修订这段文字的时候已经是2023年,PyTorch无论是在工业界还是学术界,都已经碾压了其他的框架,例如TensorFlow、Keras。如果是入坑不久的朋友,我建议你直接学PyTorch就好了。其他框架基本上可以仅follow up即可。
89
89
* [ <img src =" img/bilibili32.png " width =" 18 " /> PyTorch视频集合(32集)] ( https://www.bilibili.com/video/BV197411Z7CE/ )
90
90
* [ <img src =" img/zhihu32.png " width =" 18 " /> PyTorch的安装与Tutorial] ( https://zhuanlan.zhihu.com/p/60526007 )
94
94
### 分布式训练
95
95
* [ <img src =" img/zhihu32.png " width =" 18 " />《分布式训练》] ( https://zhuanlan.zhihu.com/p/129912419 )
96
96
97
- ## 强化学习
98
- * Reinforcement learning (RL) is a type of machine learning, in which an agent explores an environment to learn how to perform desired tasks by taking actions with good outcomes and avoiding actions with bad outcomes.
99
- A reinforcement learning model will learn from its experience and over time will be able to identify which actions lead to the best rewards.
100
-
101
- * TBD
102
97
103
- ## Advanced Topics
104
-
105
- ### 大模型
98
+ ## 大模型
106
99
综述:[ <img src =" img/zhihu32.png " width =" 18 " /> 2022 年中回顾 | 大模型技术最新进展] ( https://zhuanlan.zhihu.com/p/545709881?theme=dark )
107
100
108
- #### LLM 语言大模型
101
+ ### LLM 语言大模型
109
102
语言大模型(LLM)可以通过学习大量的语料来模拟人类语言处理的能力,如文本生成、翻译、问答等。相比普通的模型,LLM具有更高的准确性和更强的适用性。在最近几年,LLM取得了长足的发展,并在各种应用中取得了显著成果。LLM的发展有许多关键节点,下面列举几个重要的节点:
110
103
111
104
* 2014年,Google提出了Word2Vec模型,它能够将单词映射到一个低维向量空间中,并且能够在这个空间中表示单词之间的语义关系。这个模型为深度学习语言模型的发展奠定了基础。
@@ -118,21 +111,35 @@ A reinforcement learning model will learn from its experience and over time will
118
111
119
112
* [ 语言大模型介绍 <img src =" img/bilibili32.png " width =" 18 " />] ( https://www.bilibili.com/video/BV1Hj41177fb ) , by Andrej Karpathy
120
113
121
- #### LVM 视觉大模型
114
+ ### LVM 视觉大模型
122
115
* DINO-v2
123
116
* SAM (Segment-Anything)
124
117
125
- ### 多模态
126
- * BLIP-2, InstructBLIP
127
- * CogVLM
128
- * GPT-4V
118
+ ### 多模态 (MLLM/VLM)
119
+ 随着语言大模型的发展,图文多模态领域从23年开始也获得了巨大的突破。其中BLIP-2架构 巧妙的利用冻结的LLM和ViT,仅训练Adapter层的架构,成为后来图文多模态领域的主流。
120
+ * [ <img src =" img/github32.png " width =" 18 " /> BLIP-2] ( https://github.com/salesforce/LAVIS/tree/main/projects/blip2 )   ;  ;  ;  ;  ;  ;  ; (2023年1月发布)
121
+ * [ <img src =" img/github32.png " width =" 18 " /> InstructBLIP] ( https://github.com/salesforce/LAVIS/tree/main/projects/instructblip ) (2023年5月发布)
122
+ * [ <img src =" img/github32.png " width =" 18 " /> Next-GPT] ( https://github.com/NExT-GPT/NExT-GPT )   ;  ; (2023年9月发布)
123
+ * [ <img src =" img/github32.png " width =" 18 " /> CogVLM] ( https://github.com/THUDM/CogVLM )   ;  ;  ;  ;  ; (2023年10月发布)
124
+ * [ <img src =" img/github32.png " width =" 18 " /> LLaVA-1.5] ( https://github.com/haotian-liu/LLaVA/ )   ;  ; (2023年10月发布)
125
+ * [ <img src =" img/github32.png " width =" 18 " /> Unified-IO 2] ( https://github.com/allenai/unified-io-2/ ) (2023年12月发布)
126
+
127
+ ### 大模型微调
128
+ - [ <img src =" img/colab32.png " width =" 18 " /> 微调 Mixtral-8x7b, QLoRA, 数据集ultrachat] ( https://colab.research.google.com/drive/1VDa0lIfqiwm16hBlIlEaabGVTNB3dN1A )
129
+ - [ <img src =" img/colab32.png " width =" 18 " /> 微调LLaMA-2, LoRA, 数据集guanaco-llama2-1k] ( https://colab.research.google.com/drive/1PEQyJO1-f6j0S_XJ8DV50NkpzasXkrzd )
130
+
131
+ ## 其他主题
129
132
130
133
### 视频理解
131
134
* <img src =" img/bilibili32.png " width =" 18 " /> [ 视频理解论文串讲(上)【论文精读】] ( https://www.bilibili.com/video/BV1fL4y157yA )
132
135
* <img src =" img/bilibili32.png " width =" 18 " /> [ 视频理解论文串讲(下)【论文精读】] ( https://www.bilibili.com/video/BV11Y411P7ep )
133
136
* <img src =" img/bilibili32.png " width =" 18 " /> [ 双流网络:视频理解开山之作【论文精读】] ( https://www.bilibili.com/video/BV1mq4y1x7RU )
134
137
* <img src =" img/bilibili32.png " width =" 18 " /> [ I3D:3D卷积网络【论文精读】] ( https://www.bilibili.com/video/BV1tY4y1p7hq )
135
138
139
+ ### 强化学习
140
+ * Reinforcement learning (RL) is a type of machine learning, in which an agent explores an environment to learn how to perform desired tasks by taking actions with good outcomes and avoiding actions with bad outcomes.
141
+ A reinforcement learning model will learn from its experience and over time will be able to identify which actions lead to the best rewards.
142
+
136
143
## 工欲善其事,必先利其器
137
144
### 推荐的书
138
145
@@ -178,7 +185,7 @@ A reinforcement learning model will learn from its experience and over time will
178
185
### 欢迎反馈
179
186
* 如果发现内容的错误,欢迎在GitHub提交Issue或者Pull Request
180
187
* 个人精力有限,欢迎感兴趣的朋友一起来完善和补充内容
181
- * 欢迎Star 和Share 此Repository
188
+ * 欢迎 : star : Star : star : 和Share 此Repository
182
189
183
190
## Backup
184
191
0 commit comments