拾遗

一、钉钉文件小助手，2020 年 6 月 3 日 16:23

build_main_part() 用于构建出静态图 (Variable及Operation), 并在训练过程中不变；
session.run() 一次每个batch跑一个step，run 里面的参数才会参与计算与更新，一般只需要填 train_op 与 loss；
需要在 tf.summary.scalar 中画图的一定要作为 run 的参数；
optimizer.apply_gradients() 与 minimize() 内部 global step都会自增1；

二、钉钉文件小助手，2020 年 5 月 19 日 10:50

定积分 → 黎曼可积 → 任意分割区间任意取值 → 上下限互换 →（or Newton-Leibniz formula）→ 定积分值相反。

三、钉钉文件小助手，2020 年 3 月 5 日 16:47

foobar2k for mac：无暗黑模式，功能过于单一甚至残废，竟然将我辛苦下的Jay的歌识别成日文版的，且无歌词。弃。

网易云音乐 for mac：出乎意料的精炼小巧，皮肤倒是很容易换，但竟然“高尚”地无法导入本地音乐文件夹，单独打开一首歌可爱女人也识别成了日文版。虽有歌词，且能匹配iTunes。弃。留着听许嵩吧。

iTunes：导入整张专辑挺方便，但不喜欢官方，不能播放flac文件。弃。

几乎折腾了一天从百度云下载的Jay的歌利用U盘转运至MBA，最后还是删了。用另听吧。
另：bilibili或者YouTube上Jay的专辑MV或歌曲合集。

四、钉钉文件小助手，2019年12月20日 11:23

sublime text → pycharm → vscode
sublime text 3弃坑原因：Jedi插件跳转定义失效（同时代码提示太卡），全局搜索鸡肋，git push需要命令行
pycharm弃坑原因：太重
vscode 优点：能同时写LaTeX，装了Mariana Pro主题，file specific 2/4 spaces，obvious code changes，but no multi-row tabs，全局搜索稍微不太习惯

后续 2019年12月23日 14:24 补充：

sublime text3：安装插件需要翻墙；vscode：补全和跳转竟然也是依赖Jedi。

五、KDD paper auxiliary loss：

def loss_op(self):
        super(Base_Two_Tower_Hete_Graph_Text, self).loss_op()
        with tf.name_scope("{}_Loss_Op".format(self.name)):
            if self.training_config.use_neigh_hinge_loss:
                margin = 0.2
                alpha = 0.1
                logging.info('calculate neighbor proximity hinge loss, margin = {}, alpha = {}'.format(margin, alpha))

                for i in [self.query_self_and_neigh_emb, self.shop_self_and_neigh_emb]:
                    self.dis = tf.matmul(i[0], tf.transpose(i[1]))
                    self.positive_distance = tf.reshape(tf.diag_part(self.dis), [self.config.batch_size, 1])
                    self.term1 = term1 = tf.reduce_mean(
                        tf.maximum(0., - self.positive_distance + (self.dis + margin * tf.eye(self.config.batch_size)) + margin))
                    self.loss = self.loss + alpha * term1

            elif self.training_config.use_neigh_nce_loss:
                alpha = 1
                neighbor_cnt = self.training_config.igraph_neighbor_cnt
                logging.info('calculate neighbor proximity NCE loss, alpha = {}'.format(alpha))

                for self_and_neigh_emb in [self.query_self_and_neigh_emb, self.shop_self_and_neigh_emb]:
                    temp0 = tf.split(self_and_neigh_emb, neighbor_cnt * 2 + 1, axis=1)
                    self.term = 0.0
                    for i in range(neighbor_cnt):
                        self.dis = tf.matmul(temp0[0], tf.transpose(temp0[i+1]))
                        self.positive_distance = tf.reshape(tf.diag_part(self.dis), [self.config.batch_size, 1])
                        self.negtive_distance = self.dis - (self.positive_distance * tf.eye(self.config.batch_size))
                        term = tf.reduce_mean(-tf.log(tf.sigmoid(self.positive_distance)+1e-24) -
                                                    tf.reduce_mean(tf.log(1 - tf.sigmoid(self.negtive_distance)+1e-24), axis=1, keep_dims=True))
                        self.term += term
                    self.loss = self.loss + alpha * self.term / neighbor_cnt
            else:
                self.loss = self.loss + 0.0

六、meta learning papers：

七：some public materials：

www2020-DTL-Tutorial-LinkedIn.pptx
2019-Intelligent Feed Recommendation-TangJie.pptx
Deep Semantic Matching in Amazon Product Search-YiweiSong.pdf
DeepGCNs.pptx
Future of Personalized Recommendation Systems_xiexing.pdf
KDD17-metapath2vec.pdf
Neural Architecture Search and Beyond_Barret Zoph_Google.pdf
tianshou_n+e_thu.pdf

八、currently reading papers：

九、currently reading repositories：

十、VScode settings json

{
    "files.autoSave": "off",
    "latex-workshop.message.update.show": false,
    "latex-workshop.view.pdf.viewer": "tab",
    "latex-workshop.message.badbox.show": false,
    "latex-workshop.latex.tools": [
        {
            // 编译工具和命令
            "name": "xelatex",
            "command": "xelatex",
            "args": [
                "-synctex=1",
                "-interaction=nonstopmode",
                "-file-line-error",
                "-pdf",
                "%DOCFILE%"
            ]
        },
        {
            "name": "pdflatex",
            "command": "pdflatex",
            "args": [
                "-synctex=1",
                "-interaction=nonstopmode",
                "-file-line-error",
                "%DOCFILE%"
            ]
        },
        {
            "name": "bibtex",
            "command": "bibtex",
            "args": [
                "%DOCFILE%"
            ]
        }
    ],
    "latex-workshop.latex.recipes": [
        {
            "name": "xelatex",
            "tools": [
                "xelatex"
            ]
        },
        {
            "name": "pdflatex",
            "tools": [
                "pdflatex"
            ]
        },
        {
            "name": "xe->bib->xe->xe",
            "tools": [
                "xelatex",
                "bibtex",
                "xelatex",
                "xelatex"
            ]
        },
        {
            "name": "pdf->bib->pdf->pdf",
            "tools": [
                "pdflatex",
                "bibtex",
                "pdflatex",
                "pdflatex"
            ]
        }
    ],
    "editor.fontSize": 14,
    "editor.codeLens": true,
    "terminal.integrated.inheritEnv": false,
    "[python]": {
        
    },
    "python.linting.enabled": false,
    "workbench.colorTheme": "Mariana Pro (Warm)",
    "editor.wordWrap": "on",
    "workbench.editor.enablePreview": false,
    "breadcrumbs.enabled": true,
    "workbench.activityBar.visible": true,
    "workbench.list.horizontalScrolling": true,
    "editor.minimap.enabled": false,
    "git.enableSmartCommit": true,
    "git.confirmSync": false,
    "update.mode": "manual",
    "explorer.confirmDelete": false,
    "extensions.autoUpdate": false,
    "extensions.ignoreRecommendations": false,
    "window.zoomLevel": 0,
    "workbench.statusBar.visible": true
}