661 Matching Annotations
  1. Oct 2020
    1. left

      .

    2. causes the analog sticks on the controllers to randomly move around and input commands to the console, even when they’re not being physically moved.

      .

  2. Sep 2020
    1. no donations received for 5 months

      only accept bitcoin ?

  3. Aug 2020
    1. test the uniqueness, completeness, and distinctness of features (Deequ);

      we need this stuff

    2. AWS Deequ

      work with Apache Spark

    3. a low latency online feature store (typically a key-value store or real-time database) and a scale-out SQL database to store large volumes of feature data for training and batch applications.

      -key-value -sql

    4. provide low latency access to features for online applications

      2

    5. store both large volumes of feature data

      1

    6. transactional data-lake that provides versioned, structured datasets.

      .

    7. Without time-travel, they cannot support incremental feature engineering

      .

    8. ML pipeline

      .

    9. we need both versioned code and versioned data

      CODE, DATA

    10. Git is not suitable as a platform for versioning data, as it does not scale to store large volumes of data.

      Still work with small CSV, but what is other option ?

    11. need to version data

      like put CSV on git

    1. Metaflow follows the dataflow paradigm which models a program as a directed graph of operations.

      seem like luigi

    1. pixel-wise tasks can poten-tially overfit when the network is continually exposedto a given set of pixels during training

      why we should not overlap in pixel-wise task

    1. We believe that whilethemaxandaveragefunctions are rather similar, the useof average pooling encourages the network to identify thecomplete extent of the object. The basic intuition behindthis is that the loss for average pooling benefits when thenetwork identifiesalldiscriminative regions of an object ascompared to max pooling.

      average pooling vs max pooling

    1. We introduce the use of rectified linear units (ReLU)

      is it first paper talk about ReLU ?

  4. arxiv.org arxiv.org
    1. We add the BN transform immediately before thenonlinearity

      before relu

    1. Americans with low income have been more mobile during the pandemic than the wealthy

      .

    2. Without much savings in the bank, you have to make frequent trips to the grocery story to get supplies little by little.

      poor people can't stock stufff

  5. warhammer40k.fandom.com warhammer40k.fandom.com
    1. homes of the wealthiest and most powerful people in the city, including the planetary Imperial nobility, the wealthiest merchants, industrial barons and any visitors of importance from off-world.

      VIP

    2. middle-class or even upper middle class

      white collar !! But why call middle-class while majority of hive city in Lower hive ?

    3. power is free

      Lower Hive pay, Upper Hive free :D

    4. the bureaucrats

      white collar worker

    5. the people of the Lower Hive must pay a premium for every watt of power they consume

      electricity bill is high :D

    1. Adam generally requires more regularization than SGD

      apart from weight_decay, what else regularization do we have ?

    2. amsgrad are a poor “fix

      so default amsgrad = False ?!

    3. as long as it’s properly tuned

      how the f

    1. the best value ofλ′is tightly coupled withthe learning rateα

      if you are going to schedule learning rate, then schedule weight decay also. Correct ? please comment

    2. However, as we will demonstrate later in this section, this equiva-lence doesnothold for adaptive gradient methods.

      So, whether deep learning library implement weight decay or L2 ? Not sure until we read and understand source code. Documentation named parameters as weight_decay but annotated as L2 penalty

    3. L2regularization is very frequently referred to as weight decay

      In pytorch documentation, weight_decay is annotated as L2, but weight decay is not L2.

    4. practitioners do not need to switch between Adam and SGD anymore

      yeah, no need test SGD, just all-in AdamW

    5. larger the runtime/number of batchpasses to be performed, the smaller the optimal weight decay.

      Run long -> small weight decay

    6. adaptive gradient methods do not generalize as well as SGD with momentum when tested on adiverse set of deep learning tasks, such as image classification, character-level language modelingand constituency parsing.

      SGD for hard-core state-of-the-art people

    1. θθ\theta is a parameter (theta), e.g. your weights, biases and activations.

      trainablle parameter

  6. Jul 2020
    1. cross-correlation

      It is convolution without rotating stuff to save computation resource. Same effect if real convolution was use.

    1. not caused by overfitting

      then Regularization not help, isn't it ?

    2. depth increasing, accuracy gets saturated

      what the fuck ><

    3. intermediate normalization

      batch norm

    1. Sons of Dorn (Novel) by Chris Roberson

      Maybe it have something to do with Feudal World.

    1. cardiovascular diseases

      .

    2. natto

      "Nattō (納豆) is a traditional Japanese food made from soybeans that have been fermented with Bacillus subtilis var. natto" (Wiki)

    3. prescription drug to help improve memory, thinking, and brain function in people who are healing from a stroke.

      stroke recover

    4. no good scientific evidence

      .

    5. use vinpocetine for improving memory and thinking skills, boosting energy, for weight loss,

      No scientific evident

    6. not meet the criteria

      .

    7. prescription drug

      Found in Cerevit, over-the-counter medicine. Smart drug

    1. exceptionally tricky

      more bug ?

    2. rewrite the computational graph

      potential make calculation longer

    3. very few occasions when in-place operations actually lower memory usage by any significant amount

      in-place does not help for most of the case

    4. breaking the training procedure of the model

      difficult for state of the art

    1. —had removed the human discourse data and added voice mimicry software

      praise the omnissiah

    1. hungry, tired, and fully aware of their dependence

      .

    2. withholds food, medicine, and necessary supplies

      .

    3. manufacturing a crisis of scarcity

      make thing expensive so poor people have nothing

    1. UsernamePasswordToken

      what if we don't use Username and password ? What is other choice ?

    1. in a sin-gle training run using stochastic gradient descent

      1 epoch with sgd ? ONE EPOCH ?

    1. The youth were sent out to clear land for farming.

      đi kinh tế mới

    2. “One Child Policy.”

      after no comdom, then boy only, (discard if baby is girl)

    3. Contraceptives

      no comdom, huh ?

    4. environmentally unsustainable

      Tư bản giãy chết

    5. tư bản giãy chết

    1. A Servo-skull is a drone-like robotic device

      Yeah, an UAV

    2. To have one's skull chosen to serve as a Servo-skull is a great honour in the Imperium

      I want my skull to be chosen to make UAV too.

    1. Customized collation

      maybe sample generated from sample (data augmentation) need to be flatten,

    1. extend the life cycle of Nintendo Switch

      what about hardware upgrade ? If not, how long Nintendo Switch can survive before become obsolete ?

  7. www.pythonmorsels.com www.pythonmorsels.com
    1. you'll want to comment out the noted lines of code in the tests file to test them properly

      Comment out @unittest.expectedFailure in unittest file

    1. exposed immediately

      no HTTPS = fuck yourself

    2. the username and password are sent with every request, potentially exposing them

      man in middle

    1. you can consume vegetables by simply taking it in supplement form

      vegetable in pill, what the fuck ?>

    1. “M-SegNet” (Modified-SegNet).

      Segmentation

    2. “M-SFAgNet” (Modified-SFANet)

      multi-scale

  8. Jun 2020
    1. atrous rate

      dilated rate

    2. the modelâ ̆A ́Zs field-of-view

      what the fuck ?

    3. he importance

      weight

    4. the receptive field size should be changed across the imagedue to perspective distortion.

      Context-aware crowd counting

    5. loss of spatial information

      dilated convolutional layers prevent loss of information

    6. atrous convolutional layers

      dilated convolutional layer

    7. proposed in

      CSRNET

    8. he study in [22]

      Csrnet: Dilated convolutional neuralnetworks for understanding the highly congested scenes.

    9. [27]

      Bayesian loss for crowdcount estimation with point supervision.

    10. SegNet
    11. SFANet
    1. 68(+189%)

      why + 189% to what ? SHB MAE is around ~ 25

    2. estimated for each scale and the meanis taken as the overall estimate

      image1 = full_size image2 = full_size*0.8

      result1 = model(image1) result2 = model(image2) result = (result1 + result2) / 2

    3. 23.7633.12

      FCN only work best on spared crowd scene, so, only part B

    4. 2D integrator

      Is 2D integrator have to work on training state ? or only when evaluation ?

    5. D integration (

      is it tensor.sum()

    6. 1 overlapping centrecrop

      The authors said overlap does not help model, so here is figure said that overlap centre make it worse

    7. pixel-wise tasks can poten-tially overfit when the network is continually exposedto a given set of pixels during training

      This is pixel-wise tasks, so crop does not help much.

    8. 9:1 ratio

      Training set is split to => 9 training : 1 validate

    9. Zhang et al., 2016)

      What the hell is this ?

    10. (Zhang et al., 2016).

      MCNN

    11. Single-image crowd counting via multi-column convo-lutional neural network

      MCNN

    1. “Do you know what my kind’s greatest advantage is?” the vampire said, “you people tell stories about us from the time you’re little.

      Female vampire character

    1. j

      observation

    2. i

      level of treatment (your method apply on the sample)

    3. alternative hypothesis is what we think is going on

      we want Ha (alternative hypothesis) is true.

    1. Tiêu điểm: Vay nợ để tiêu xài - Quả bom nổ chậm | VTV24

      Cá nhân tuyên bố phá sản là cc gì ?

    1. treatment

      factor

    2. alternative hypothesis

      the hypothesis we want to prove.

    3. state a null hypothesis

      instead say "our method better", say "our method does not make any different".

    4. Plants with different fertilizers will grow to different heights

      hypothesis

    1. Pi={P1,...,PC(i)}

      Annotation is set of point (2d coordination) of object. Point, not box.

    2. kernel density estimate

      KDE

    1. On Undernet, you can either go with "Bookz" or "ebooks", while on IRChighway, you can join "ebooks". Type in your channel and hit join. Example

      piracy ebook IRC

    1. expensive and slow [23]

      paid wifi hotspot

    2. the government has prioritized education and health initiatives across the countryand it is estimated that 99.8% of the 11 million inhabitants in Cuba are literate

      really good thing for communism

    1. a third of the average Cuban’s monthly salary — about $5 — to use a computer for an hour.

      15$ salary per month

    1. dùi mài kinh sử

      dư tiền vãi lồn

    2. Phạm hai chữ húy, mất cả tiến sĩ lẫn cử nhân

      lỡ phạm húy mất hết :v

    1. Đã vào trường thi là bình đẳng

      Nhưng mà khi chưa vào trường thi thì vẫn đéo bình đẳng, xét thân nhân, xxx

    2. Thân thuộc với những người phạm tội đã bị chém, giảo (thắt cổ), đi đày, sung quân (dù những người này đã được tha về)...

      Xét thân nhân nha, thân nhân đéo tốt thì đéo đi thi

    3. thí sinh muốn dự thi phải ghi danh tại lý trưởng của làng để xem xét tư cách.

      xét lý lịch :v

    4. hoàn toàn bằng chữ Hán

      vậy chữ Nôm nằm ở đâu ?

  9. May 2020
    1. a scale label is obtained to represent the average size of human headsin one patch

      annotate scale

    2. Specifically, SS-Net and LS-Net is trained on TSand TL

      How do you classify the training data to Ts and Tl ?

    1. fraction of the top k ranked candidates that received and accepted an InMail (viewed as precision@k)

      evaluation metric

    2. two-way interest

      InMail interactive is a metrics for two way interest

    3. InMail Accept

      objective to optimize (i.e: increase InMail accept rate)

    4. not just that a candidate shown must be relevant to the recruiter’s query, but also that the candidate contacted by the recruiter must show interest in the job opportunity.

      candidate -> recruiter AND recruiter -> candidate

      Two way preference.

    1. trừng phạt bằng roi vọt và học thuộc lòng

      giống giờ

    2. không cấm con em nhà thường dân

      vậy là đời trước cấm con nhà thường dân đi thi à :v

    1. isometric strength
    2. It does not appear that CrS increases maximal isometric strength, the rate of maximal force production, nor aerobic exercise performance.

      it does not make you stronnger.

    3. 20-30 g·day

      human body create 1-2g per day

    1. inflatable ball, and after you are positioned inside of the big transparent ball, you can literally walk on water

      how it work ?

    1. for every kilowatt hour of electricity we consumed, we purchased a kilowatt hour of renewable energy.

      So total Google buy 2 kilowatt hour for each kilowatt hour of electricity used ?

    1. khẩu pháo Oto Melara 76 mm phía mũi tàu

      còn lại ổ ship-to-ship cùi bắp

    2. tháo dỡ dàn pháo đánh chặn tầm gần Phalanx (điều khiển bằng radar) ở đuôi tàu và dàn radar tìm kiếm mục tiêu trên không SPS-40 (tầm trinh sát tối đa 450 km, ở trên đỉnh tháp tín hiệu phía sau tàu).

      cho tàu éo cho súng

    1. number of shards are immutable

      IMMUTABLE ?! we cannot change number of shard ?

    2. how to split an index was to just add another index and grow your shard count that way.

      yeah, create new index

    3. shrink an index into a new index with fewer shards than the original index.

      How about the opposite ?

    1. Meanwhile, a worm makes its way to the young king's brain, while in the north, Crown Prince Chang and Seo-bi track down a mysterious merchant woman who was selling the resurrection plant to find out the culprit behind the zombie invasions.

      Oh fuck, why Crown Prince Chang is not become a king ? but a brat ?

    1. Ashley Too is activated by a news report. She learns of her real self's coma,

      the doll is real AI that aware itself is a doll of another "real" self.

    2. Catherine discovers that Ashley has not been taking medication intended to counter her darker moods.

      tự kỉ ?

    3. Jack hides Ashley Too and tells Rachel that she threw it away

      why ?