Untitled
最近在看深度学习的东西，一开始看的吴恩达的UFLDL教程，有中文版就直接看了，后来发现有些地方总是不是很明确，又去看英文版，然后又找了些资料看，才发现，中文版的译者在翻译的时候会对省略的公式推导过程进行补充，但是补充的又是错的，难怪觉得有问题。反向传播法其实是神经网络的基础了，但是很多人在学的时候总是会遇到一些问题，或者看到大篇的公式觉得好像很难就退缩了，其实不难，就是一个链式求导法则反复用。如果不想看公式，可以直接把数值带进去，实际的计算一下，体会一下这个过程之后再来推导公式，这样就会觉得很容易了。
说到神经网络，大家看到这个图应该不陌生：
这是典型的三层神经网络的基本构成，Layer L1是输入层，Layer
L2是隐含层，Layer
L3是隐含层，我们现在手里有一堆数据{x1,x2,x3,...,xn},输出也是一堆数据{y1,y2,y3,...,yn},现在要他们在隐含层做某种变换，让你把数据灌进去后得到你期望的输出。如果你希望你的输出和原始输入一样，那么就是最常见的自编码模型（Auto-Encoder）。可能有人会问，为什么要输入输出都一样呢？有什么用啊？其实应 ...
Untitled
原来收敛不了估计是之前的loss函数设置不对,导致学习率相对来说太大了
训练100个epoch的loss
1733140892597
Contrastive Multiview Coding
854efd16088ec0a2949395f17c9b60f9386aedb0ff5a819cd0390912c2b2702d066471cafe8cbf03265d19bb8e0cba2b0eafa60e172aa96b2ccb75f5a1c2be2d234b83df7280d3550c56a7e445c16ec60092887d819fb7984883842fec7169ebf87a2732418183a2a52bc02851ec99ef4c301706b1ac14e4e5d6c2a2b4f40dcf92c2ee61bcaae0ff14eeabc904fcbb7cc03cd5414a7e568af842fb17e72c3a340dedd4d90c39de435d6da16c64efc089a28c94349bcec271daebfed378554f24281d93e15de32f6b162f70b411900c1f017cf998aa5e625c1aee8afc0dd55a94e2728a9e33950085e7d7f944db23fa9ff074e5e6b09e71d50 ...
周报11月10日2
854efd16088ec0a2949395f17c9b60f9618ed3796518aea5e690afa16559fc1f7b0cd4c708f2cc063e655635dd263ab76e4ff4bc1da84ac7ca46b77f350f7658ee204247e1c5a23721d9cc0f6c79974a1cc4d6ff5c7dc3049ef247b0e536e01f3392158aed08516076472078143a571b135f4de26b94ff07a4d5b96c412679bdd70c83a509391d1167ce1cfac36611a678c4beb1c6a605e1b1862789b454955a3b1fb1fe431e4b80987bd001abf12b6f8267bdb24e376bf9c6248ea048e3c53bc0cccbd9bc1cade245e956865b05af46b858b2a32cd864e680eac23f98fcd31bc23056f911535cf4aa86264385fd22487b3fbd5523c32eea8 ...
CSDI:Conditional Score-based Diffusion Models for Probabilistic Time Series Imputation
854efd16088ec0a2949395f17c9b60f9386aedb0ff5a819cd0390912c2b2702d5f321fa588f275c76f882b23be7542bc91fc22134b48b18b8f5cb560fa056139012167e77f0224ee1bc73ef037c53f30248a984d6a7b238e1807e1c5a25868dedd57fb9b57c00f75d4d7c0d746d6fa8859f0f0fc6d632ea6b6494b0ecd199b2918310f16d9b7c84a111f9f907a1d9ce30a66d2c4704bc2c544768e88e72e11e60726d0c6657830fe064cf5568f5b2dd712af6b1b289e00483ca9ea565912124f481c9184208638f7c17b86400d21be2d9274f39a8e2a9a96990af564134014631edce85438eb56df943adc51b1f49bad5c79462b553d1bf93 ...
CSDI代码解释
854efd16088ec0a2949395f17c9b60f9386aedb0ff5a819cd0390912c2b2702d238ea47d1573205f1b957d8433d6f0de9a1bd95ba6f70c2db3cbbefaea755cb8859969bddf71c304d7561a027329a74858baf35ac5184b80e19a311b6796d724a57e1ecb9be93dafcee3ec9f0b6ee176b822ff6af0ff850ea2fee97f418275acfeef96d8188787a6f860eac808800c768b1c10a7b40f2bf0d90691ecc69282e14f272bd9e8154d566c08bc4cca73b91b00934d7d3b4da2474612b476a3fd7123c58435968d6941e46a48ed3b3b18a897eb38e4d2ff066991f727d4077fd96c9676e9d34d9eea86dcaf886125380385796038bb5701b73d1b5 ...
TimesURL:Self-supervised Contrastive Learning for Universal Time Series Representation Learning
854efd16088ec0a2949395f17c9b60f969e90a67b1b28d50d97fce0c9150dfbf175df193eda1c184cc14b0b6c9e96b93cfb4303ec8071863c69bdb5785453c331d1e75436ec3633f2d4d4e912ab014b8ae9812be3cf8710663c172e68ebb751811b92fc9e67d3151f5842ee357d7fbc75ac0d9dc7b749777a24b00ae654c323ecb1fb338efe1b1413cb20eea57dc900133c094dedbb7a2dc6eb0eed373c6658cd7d565a8d6882ae570fd89030ab246d1ac87bb671b98f7f1a5587fb9a8d8da716f71a771781b895b12e3e50efd61198efc31e94c4aa43c7a1707a08bfa0529e7df34c279b109736ffaf84de9491341a7b58cd414ab9187c95 ...
周报11月18日
854efd16088ec0a2949395f17c9b60f9618ed3796518aea5e690afa16559fc1f54427900688f3cbd631196238cd80952bd025411d6c77470e7bfe6b1a81070110487423f046b687f32f03695903a00c213b574020b6405d7f96ff0c7fa3942234266f2928c2c131312ee2effc1d484e9e6ea56d6151378e19fc2d61cad2e8072461ff244f8f2d501fe9f08ae165aa93832e746a1aeb3043c2c285471ba1328dd28c69ccb7b2795bce1c6a0302db4dc2ccb04695ac6e03f2070f5281be921940bc0fa883b61e12c8176526276852cbb84776c66288820af4c5d39fc64d99e0bbb30663cc2aa54b1edaa79490cf2332c369f85e3f58ce17d61a ...
STraT代码阅读
854efd16088ec0a2949395f17c9b60f9386aedb0ff5a819cd0390912c2b2702d2bbf6e15377c49cfb8d488a63fc1174b879e1db4d9837cb43bcaced2c614cf7940ead00bae524cf55319aab6966ecfeb5fc2f2afc44a56d276de2d1dea7d265a4e3c29a1b9a75e6cf242d41dbbb9055feb3bc555fa366708ff1da65afde048efc37ff248b344c494d33853f0995082e8b9f463acd83c0456048fefc042364d6f506500634585264ae39eafa4e075584f247c12b25807a712c7c65efc5e3292a52a1fc5a54a28d50ebfe8dbdac2503d0ca96cdb87c5402423585991bb3b2f5b1d803460be341f9ef142cb3825d784ab1b7068ba04c1b342a62 ...
巴基斯坦
854efd16088ec0a2949395f17c9b60f9ccd15ce29826ada004fc71dd724e293172f6453e1fde581d65fe6651bcd681e5f0f82916defa9f8fb4ae0e2c5be9deb4620af640cdd2ac05f5baa06dc2f29699256fbfe755bfdf651a916a8bd4153d5704c12bc9f0e2a3625a08e16b0275c29f85b6e3d01c6b2aaf6470d35792393c6255978f927f1f18624081b3b8dc14d471701ef8b8cbb5d360bb819e772e1564b0bd7eb0d40bce15663887b01c68293d71d4f8c19f3a1433ad58714de4e2e303535f6818a367fe670979e8263420622c92517d382d7d85eecc59971e602b45ab02a940ca77faaed5baf933d3261a338de8eacdb87b6abca29b1 ...