2d:null 2e:["$","title",null,{"children":"Documentation of New CharActers | KarlieDoesNotExist | Observable"}] 33:T5ca,md`
### 15/5 As shown below, when the structure of the character becomes complex, the model trained for 12000 steps will have problems in recovering from the damages.

CA behaviour before/after recovery at 12000 steps:

CA behaviour before/after recovery at 16000 steps:

In some cases, the generation failed even when the structure is simple. This is a visualisation of the loss history when the model was trained to rebuild the character

This is because each generation starts from a seed (i.e. one black pixel in the centre of the canvas) and gradually grows to the target character/pattern, meaning that the centre pixel of our input image cannot be transparent, and other non-transparent pixels cannot be too far away from it. In the cases like the character

, the generation failed even when we manually set the centre pixel to white, as it is isolated from other coloured pixels. `34:[{"id":"0c7e29b724b0879aa8d4c240404bcdcec87608a915a583a0bf1586d5a9ec87e4fc5bb8046239b1d34bc941670e6add8b67fc42fd7b435ba0352a93e9ad6e371a","href":"https://static.observableusercontent.com/files/0c7e29b724b0879aa8d4c240404bcdcec87608a915a583a0bf1586d5a9ec87e4fc5bb8046239b1d34bc941670e6add8b67fc42fd7b435ba0352a93e9ad6e371a","name":"xi_12000.png","type":"image/png","lastModified":1621313176109.732,"size":146742},{"id":"74d187fc3be49a3412f144e94bca274c3e9ce248331088c4f81f981f710e2f554b5f07650ca84a31e3b2a7d1b1996830e8777be72e52ef28352bfc6219ba082d","href":"https://static.observableusercontent.com/files/74d187fc3be49a3412f144e94bca274c3e9ce248331088c4f81f981f710e2f554b5f07650ca84a31e3b2a7d1b1996830e8777be72e52ef28352bfc6219ba082d","name":"wang_failed.png","type":"image/png","lastModified":1621313175461.943,"size":25966},{"id":"aad6df63e79b5bc1d7f0b5a890a3f6e74cf0ff8383096d0953a327b78a507ffc24eb636e9b461f1f613443b99c7cab7871f1e3ddc95535192f92d603602a03d7","href":"https://static.observableusercontent.com/files/aad6df63e79b5bc1d7f0b5a890a3f6e74cf0ff8383096d0953a327b78a507ffc24eb636e9b461f1f613443b99c7cab7871f1e3ddc95535192f92d603602a03d7","name":"loss_history_fei.png","type":"image/png","lastModified":1621315900704.409,"size":603667},{"id":"629a7c3d45f6c3ffa7b11037d6e02cf41bf2fb7f2b04d5f8e068014f4e347dc3381b79cef8d64fd4a7db3abe64bc7c984872a669beae1fd943d9ad1a9b4341bc","href":"https://static.observableusercontent.com/files/629a7c3d45f6c3ffa7b11037d6e02cf41bf2fb7f2b04d5f8e068014f4e347dc3381b79cef8d64fd4a7db3abe64bc7c984872a669beae1fd943d9ad1a9b4341bc","name":"title.png","type":"image/png","lastModified":1621358407537.017,"size":78853},{"id":"ecf35ffb18a6a3c19035ab12c12286c3262709acf7f5011bddaf5af22d2844aa887a9a4fcb86a637c38d2a2d5ba2cad1d82117cdd7bcbce4233dc88e077c4caa","href":"https://static.observableusercontent.com/files/ecf35ffb18a6a3c19035ab12c12286c3262709acf7f5011bddaf5af22d2844aa887a9a4fcb86a637c38d2a2d5ba2cad1d82117cdd7bcbce4233dc88e077c4caa","name":"transformation_3.gif","type":"image/gif","lastModified":1622207865966.037,"size":12094196},{"id":"9a7fc32180f0e38d499475bda278e04b87e25d4cf849b6306b05efeb3cb157e4e5c418f772c4dc90ba0e85704e68f44f149ac641701fcbaaba4d7f0261c2387b","href":"https://static.observableusercontent.com/files/9a7fc32180f0e38d499475bda278e04b87e25d4cf849b6306b05efeb3cb157e4e5c418f772c4dc90ba0e85704e68f44f149ac641701fcbaaba4d7f0261c2387b","name":"infect_sim_to_trad.gif","type":"image/gif","lastModified":1622207364122.699,"size":8013866},{"id":"2207fc03582d48b23952ac5a1fb2ec9400db3eca37d6a133022fe65d09ff02ce4f8950bf58ff236bb22831a39a6e8ce60fd48c9525270aaea562c19767e23791","href":"https://static.observableusercontent.com/files/2207fc03582d48b23952ac5a1fb2ec9400db3eca37d6a133022fe65d09ff02ce4f8950bf58ff236bb22831a39a6e8ce60fd48c9525270aaea562c19767e23791","name":"word_concact.png","type":"image/png","lastModified":1622301820023.569,"size":1037079},{"id":"ee32c12109f15fd4a2a556a58bde5ffafff11a59eeac88169fdb7bc7ae9f6819e536a82bcec5486fe01d88a8843007a8da7c977fb7c5b0efdf0b30cb07edc345","href":"https://static.observableusercontent.com/files/ee32c12109f15fd4a2a556a58bde5ffafff11a59eeac88169fdb7bc7ae9f6819e536a82bcec5486fe01d88a8843007a8da7c977fb7c5b0efdf0b30cb07edc345","name":"ji_traininglog_15000@1.jfif","type":"image/jpeg","lastModified":1621313130469.458,"size":32238},{"id":"336aebcc611acd3b9f3738ac3baf05378ed5045f6f7bfe28fd014ffc6ea292c4036c9b9d58e5159b74532170d48db94690c69cc1a82bd7ccbae4518b3074eef4","href":"https://static.observableusercontent.com/files/336aebcc611acd3b9f3738ac3baf05378ed5045f6f7bfe28fd014ffc6ea292c4036c9b9d58e5159b74532170d48db94690c69cc1a82bd7ccbae4518b3074eef4","name":"fei4.mp4","type":"video/mp4","lastModified":1621316163879.024,"size":180848},{"id":"e258fe9d1a797642241aa3fd91515286f3aee552fb82cabe21b6b992c74e5e23973bc85aee7787d34ed4c77d3eb0a831d76821bf377112b98b1ca603460d0ac2","href":"https://static.observableusercontent.com/files/e258fe9d1a797642241aa3fd91515286f3aee552fb82cabe21b6b992c74e5e23973bc85aee7787d34ed4c77d3eb0a831d76821bf377112b98b1ca603460d0ac2","name":"fei2.mp4","type":"video/mp4","lastModified":1621316152624.872,"size":393876},{"id":"612b07b06a04ad0195e152c766b5052c515aa81c2f040ade77545e42de72911d4ef833d77aa180751d8afb402125c3a7d6549d656c5b2532079e16c33f17bd2c","href":"https://static.observableusercontent.com/files/612b07b06a04ad0195e152c766b5052c515aa81c2f040ade77545e42de72911d4ef833d77aa180751d8afb402125c3a7d6549d656c5b2532079e16c33f17bd2c","name":"fei3.mp4","type":"video/mp4","lastModified":1621316143460.093,"size":346694},{"id":"636d55a884aee7f5f8807c98fee2159a1e3a443fe942c83d90d3ff008342b58ba8b4495d999feaba3343f2ba97c56ce74eabe4f3c5d3a4d75c5ba6920809757d","href":"https://static.observableusercontent.com/files/636d55a884aee7f5f8807c98fee2159a1e3a443fe942c83d90d3ff008342b58ba8b4495d999feaba3343f2ba97c56ce74eabe4f3c5d3a4d75c5ba6920809757d","name":"fei.jpeg","type":"image/jpeg","lastModified":1621315991579.653,"size":77447},{"id":"9fef0dc04d44576e548e0ed18fc7dab683995c6579e444c3b4a1cff02be240ec8845765d9b8118356cb01ce8e353b77bc228628905f0f7aedc2ede512289f66c","href":"https://static.observableusercontent.com/files/9fef0dc04d44576e548e0ed18fc7dab683995c6579e444c3b4a1cff02be240ec8845765d9b8118356cb01ce8e353b77bc228628905f0f7aedc2ede512289f66c","name":"2420_pool.jpeg","type":"image/jpeg","lastModified":1621315971546.948,"size":83858},{"id":"97e4ebc6efce881c3bf84ee92f366fc4add31cd883c735a2c4f7e2167a567becf46d8057e05683815b0dc95917d9cbfe72ed536905ced3c5c3a57727855028a0","href":"https://static.observableusercontent.com/files/97e4ebc6efce881c3bf84ee92f366fc4add31cd883c735a2c4f7e2167a567becf46d8057e05683815b0dc95917d9cbfe72ed536905ced3c5c3a57727855028a0","name":"qiong.jpeg","type":"image/jpeg","lastModified":1621315955763.148,"size":10092},{"id":"cf16f4a2e7761e50dd2fc4268fe0c6e8e838d84f903c0c1ec5b55a9e28d5a75de98e6c3da7b02ec86b8862565131deaed546c764f1dc8bbc78754f3ad3b2b27d","href":"https://static.observableusercontent.com/files/cf16f4a2e7761e50dd2fc4268fe0c6e8e838d84f903c0c1ec5b55a9e28d5a75de98e6c3da7b02ec86b8862565131deaed546c764f1dc8bbc78754f3ad3b2b27d","name":"qiong_failed.mp4","type":"video/mp4","lastModified":1621315954976.449,"size":246698},{"id":"49a66e59e78b4e0b245b59298882eb94612076368cd7c3c94d042e139cfd88b5fc20e5d740e1064699b155ae6ee847b22c7e9980c7859810615b4736a57a20a2","href":"https://static.observableusercontent.com/files/49a66e59e78b4e0b245b59298882eb94612076368cd7c3c94d042e139cfd88b5fc20e5d740e1064699b155ae6ee847b22c7e9980c7859810615b4736a57a20a2","name":"loss_history_qiong.png","type":"image/png","lastModified":1621315902038.16,"size":705018},{"id":"82fc47332751a12e30b34ddbb1d9a416aa901626a7036ef92a876c34cf917c84262c8c684810e8f911584d24bf0574d7b4ab1d8c20be277a6f013775544addc9","href":"https://static.observableusercontent.com/files/82fc47332751a12e30b34ddbb1d9a416aa901626a7036ef92a876c34cf917c84262c8c684810e8f911584d24bf0574d7b4ab1d8c20be277a6f013775544addc9","name":"fei_trainingLoop.png","type":"image/png","lastModified":1621315900635.192,"size":517963},{"id":"ad3f0d968948f9ad600cb216e7e090ee794e0960d3e083a023001c7d3ea2341ac4cdc5ca87628abe015ffdf8908bd40626a194d4987b9bd5de43d70764b8ede0","href":"https://static.observableusercontent.com/files/ad3f0d968948f9ad600cb216e7e090ee794e0960d3e083a023001c7d3ea2341ac4cdc5ca87628abe015ffdf8908bd40626a194d4987b9bd5de43d70764b8ede0","name":"han_CA.png","type":"image/png","lastModified":1621314477807.693,"size":713180},{"id":"81f587142bfe9f81c7f892c189db982250bcb3842f2ab7a4ebcddd6eb190de34f78ab87f96b532099dff40f38796a5a134317071e5fff569d27bc1a83b93eaee","href":"https://static.observableusercontent.com/files/81f587142bfe9f81c7f892c189db982250bcb3842f2ab7a4ebcddd6eb190de34f78ab87f96b532099dff40f38796a5a134317071e5fff569d27bc1a83b93eaee","name":"yue_CA.png","type":"image/png","lastModified":1621314416708.308,"size":667196},{"id":"520af01c842e3aff2ffaf348902c8a191c663aa26f1465da96e9fc29810a2627d6a0935cad4fe1d4aed9c385bedf0f8720a98c46e132f99f9ad531d4febc9ee9","href":"https://static.observableusercontent.com/files/520af01c842e3aff2ffaf348902c8a191c663aa26f1465da96e9fc29810a2627d6a0935cad4fe1d4aed9c385bedf0f8720a98c46e132f99f9ad531d4febc9ee9","name":"screenshot-2.png","type":"image/png","lastModified":1621314237440.805,"size":12539},{"id":"32171e99fe79eb1f7eaf7649564d7cab01826b251bd2e1591e2bb7d6137b6c037db06f2ddaf59397dd4889d675b08c2154d7333bc9abe13f5bf8fdc7ef966b4d","href":"https://static.observableusercontent.com/files/32171e99fe79eb1f7eaf7649564d7cab01826b251bd2e1591e2bb7d6137b6c037db06f2ddaf59397dd4889d675b08c2154d7333bc9abe13f5bf8fdc7ef966b4d","name":"hua_CA_behaviour.png","type":"image/png","lastModified":1621314214876.4,"size":14814},{"id":"694cdfd112ee63ac9c7c0354121a25a38718c7711a995427a3c72f95a5d0ead7a3ec5ba16da1f9f96d8b9014a828d6d507e9995e42b132f4a4945a482d226228","href":"https://static.observableusercontent.com/files/694cdfd112ee63ac9c7c0354121a25a38718c7711a995427a3c72f95a5d0ead7a3ec5ba16da1f9f96d8b9014a828d6d507e9995e42b132f4a4945a482d226228","name":"ji_traininglog_12000.jfif","type":"image/jpeg","lastModified":1621313034542.209,"size":31845},{"id":"6e30e3f009e26c94ff73c6233364a40d72dc755a25ddd9d7e4be370f18e5ece1ac95e0d59db083fb7d4025405da8265789a3b826ff31a4888c4501ce6543d190","href":"https://static.observableusercontent.com/files/6e30e3f009e26c94ff73c6233364a40d72dc755a25ddd9d7e4be370f18e5ece1ac95e0d59db083fb7d4025405da8265789a3b826ff31a4888c4501ce6543d190","name":"transfer_compressed@1.gif","type":"image/gif","lastModified":1621311498993.932,"size":3963632}] 31:{"id":"01ff246d06ed3d4f","slug":"final-project-documentation","title":"Documentation of New CharActers","thumbnail":"cfafdb6c150893cae753a503f92449aa3eb8ce7edf579a46d5a57967f2be82ed","access_level":"public","create_time":"$D2021-05-18T03:59:51.222Z","update_time":"$D2021-07-15T13:35:22.219Z","view_time":null,"first_public_time":"$D2021-05-18T05:40:06.648Z","has_unpublished_edits":null,"fork":false,"stars":2,"starred":false,"owner":{"type":"team","id":"b34c988451399e98","login":"wenrazhao2","name":"KarlieDoesNotExist","avatar_url":"https://avatars-next.observableusercontent.com/64edd542e55cff350a7503e5050267fb1768e8cc5052184b8a8cb49162d172f3","tier":"starter_2024"},"version":252,"publish_version":null,"role":null,"authored":false,"authors":[{"type":"individual","id":"302ba7635bb9b25d","login":"wenrazhao2","name":"KarlieDoesNotExist","avatar_url":"https://avatars-next.observableusercontent.com/64edd542e55cff350a7503e5050267fb1768e8cc5052184b8a8cb49162d172f3","tier":null}],"members":[{"type":"individual","id":"302ba7635bb9b25d","login":"wenrazhao2","name":"KarlieDoesNotExist","avatar_url":"https://avatars-next.observableusercontent.com/64edd542e55cff350a7503e5050267fb1768e8cc5052184b8a8cb49162d172f3","tier":null,"role":"editor"}],"comments":[],"custom_thumbnail":true,"trashed":false,"lapsed":null,"subscribed":null,"forks":1,"collection":null,"type":"notebook","body":{"title":"Documentation of New CharActers","cells":[{"id":0,"value":"# Documentation of New CharActers","pinned":false,"hidden":false,"mode":"md","author":"$undefined","summary":"$undefined","output":"$undefined","database":"$undefined"},{"id":4,"value":"md`\nThis is a page documenting some of my neural cellular automata network training process, outcomes, experiments, and problems I met.\n
\n## [Take me to the work (in progress)](https://karliezhao.github.io/NewCharActers/)\n
\n## References:\n### [Growing Cellular Automata original paper](https://distill.pub/2020/growing-ca/)\n### [Self-classifying MNIST Digits](https://distill.pub/2020/selforg/mnist/)
\n### [Differentiable-morphogenesis](https://github.com/Mayukhdeb/differentiable-morphogenesis) (infections and morphing from image to image)\n`","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","output":"$undefined","database":"$undefined"},{"id":11,"value":"md`\n
\n## Settings
\n- ### Nvdia P100 on google cloud
\n👉🏻 It takes around 35-40 mins for each target to go through 12000 steps. I increased the step number to 16000 for some explicitly complex-structure characters.\n\n- ### Dataset\n👉🏻 For the most basic generation and recovery models, the target/input data is an image with a transparent background, ranging in size from 40 by 40 pixels to 100 by 100.
\n👉🏻 For the \"multi-target\" model that I linked below, the target is a sequence of images that have the same resolution. However, pytorch will throw a 'GPU ran out of memory' error when you feed too many images or the resolution is too high.\n

\n\n`","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","output":"$undefined","database":"$undefined"},{"id":167,"value":"md`## Training log & other notes\n\n`","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","output":"$undefined","database":"$undefined"},{"id":210,"value":"md`\n### 28/5\nSome experiments on word N-grams...\n
I think the content of the smaller characters that form the outline of the \"bigger character\" is important and should be meaningful to Chinese readers. However, it's quite hard to meet this requirement when the bigger character's silhouette needs to be clear and sharp.\n\n- What difference can CA bring to this 'ASCII art'?\n- What benefits can machine learning bring to this?\n- In terms of the content and meaning, what's the relationship between the larger character and the smaller ones that form the large one?\n\n

\n\n`","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","output":"$undefined","database":"$undefined"},{"id":192,"value":"md`\n### 27/5\nSuccessfully trained the \"infective model\" that can transform the simplified 风 to traditional 風.\nThe starting point of the infection seems to be any pixel that is active (i.e. not transparent) in the initial state. This might create some sorts of uncertainty that we are looking forward to.\n\n

\n\nWhen I tried to recover the original 风 in simplified from the fully \"infected\" state, I ended up with a very vague image showing something between the two stages.\n\n\n

\n`","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","output":"$undefined","database":"$undefined"},{"id":170,"value":"md`\n### 18/5\nI think I found a way to run multiple rules on one canvas, but sadly that the cells in different rules can't, and probably also won't be able to, communicate with the cells on the other side...it's just very hard to get everything work as expected when the program runs on tensors...\n\nHere's a line of title where the \"细\" gets an extra arm 🥰\n\n\n

\n\n`","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","output":"$undefined","database":"$undefined"},{"id":39,"value":"md`\n
\n### 17/5\nIn this experiment I trained two models that are able to \"infect\" a given CA state. Given a random point, the infection model's effect spreads through the image and turns it into another character. It will be cool if we can make one script \"infect and kill\" the other script. However the outcome doesn't seem stable, it can be due to the complexity of the character's structure, but maybe there are some problems with my training procedure.\n\n

\n\n### [Experiments on Multiple Targets](https://github.com/PWhiddy/Growing-Neural-Cellular-Automata-Pytorch) \nThis is an interesting experiment where multiple target images are trained in one model. \n\n

\n`","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","output":"$undefined","database":"$undefined"},{"id":61,"value":"$33","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","output":"$undefined","database":"$undefined"},{"id":95,"value":"md`\n
\n### 14/5\nSuccesfull recreated the CA model with tensorflow and tensorflow.js.
\nSome interesting patterns:
\n

\n\n

\n\n\n`","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","output":"$undefined","database":"$undefined"},{"id":129,"value":"md`
\n### On/before 13/5\n\n### Examples of unsuccessful & successful cases\n

\nTarget:

\n\n

\n\nTarget:

(\"窮\" in traditional Chinese)
\n\n

\n

\n

The patterns in the pool during training in the 2420th loop.

\n\n

CA behaviour at training steps 100, 500, 1000, 4000:\n\n\n\n`","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","output":"$undefined","database":"$undefined"},{"id":143,"value":"md`\n
\n

Example of the successful cases:

\n\n

Training process:

\n\n

CA behaviour at training steps 100, 500, 1000, 4000:

\n

\n

A random sample of the patterns in the pool during training, sampled every 20 training steps👇🏻

\n\n`","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","output":"$undefined","database":"$undefined"}],"files":"$W34","stdlib":"1","titling":"auto"}} 32:"$31:body"

The patterns in the pool during training in the 2420th loop.

CA behaviour at training steps 100, 500, 1000, 4000:\n\n \n\n\n\n`","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","output":"$undefined","database":"$undefined"},{"id":143,"value":"md`\n\n

Example of the successful cases:

Training process:

CA behaviour at training steps 100, 500, 1000, 4000:

A random sample of the patterns in the pool during training, sampled every 20 training steps👇🏻

CA behaviour at training steps 100, 500, 1000, 4000:\n\n\n\n`","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","output":"$undefined","database":"$undefined"},{"id":143,"value":"md`\n
\n