古狗7月中MLE新轩面筋

Round 1,  General Coding:

Q1:  input Array: [ 0.1, 0.2, 18.9, 17.5, 14.3, 12.1, 0.9,10.9] Output: 0.1

find X in Array, in the range [X, X+1) contains maximum numbers of the array.  

contains 3 numbers: 0.1, 0.2, 0.9…….[10.9, 11.9) contains 10.9...... return 0.1

Q2.

Round 2, General Coding:

Q1,

Q2. genrate a binary string which have  k 1, and the length is n. For example, if n = 4 and k = 2, the answer is:
'0011', '1100', '1010'.............

def generateStrings(prefix, n, k) -> String:
    if (k == 0):
        print prefix + (n) * '0''
        return
    if (n == k):
        print prefix + (n * "1")
        return
    generateStrings(prefix + "0", n-1, k)
    generateStrings(prefix + "1", n-1, k-1)

generateStrings('', n, k)
Follow ups: How to accelerate it if you have thousands of macines?   

Round 3, Machine Learning:
Give a lot of car images, classify  whether the image includes wheel.
It's a binary classification problem. I proposal to use CNN, the interviewer asked me why not use SVM, I elaborated SVM needs manually design features, and raw image would be too large to input it to the SVM.. I feel the answer is not the best, but interviewer not continue asks.  The interviewer also asked how to locate the the place of the wheel, I propose faster-cnn, then the interviewer asked me how to classify it's a front left  wheel  or right bottom wheel, I answered faster-rcnn with 4 classes.

After that,the interviewer mainly asked me to  implement details for how to train a CNN model, after I finsihed the code, the interviewer asked me what if your model have a good accuracy during training but performance poorly in the customer side? I propose to investigate the reason, not sure if it is the answer he is looking for.

   class LinearClassifier(object):

    def __init__(self):
        self.W = None

    def train(self, X, y, learning_rate=1e-3, reg=1e-5, num_iters=100,
              batch_size=200, verbose=False):
        """
        Train this linear classifier using stochastic gradient descent.

        Inputs:
        - X: A numpy array of shape (N, D) containing training data; there are N
          training samples each of dimension D.
        - y: A numpy array of shape (N,) containing training labels; y[i] = c
          means that X[i] has label 0 <= c < C for C classes.
        - learning_rate: (float) learning rate for optimizat

您好！
本帖隐藏的内容需要积分高于 188 才可浏览
您当前积分为 0。
使用VIP即刻解锁阅读权限或查看其他获取积分的方式

游客，您好！
本帖隐藏的内容需要积分高于 188 才可浏览
您当前积分为 0。
VIP即刻解锁阅读权限 或 查看其他获取积分的方式
Unlock interview details and practice with AI
Curated Interview Questions from Top Companies

+1) * (x2 - x1 + 1)
           var = 0

           /n repeat (y2-y1+1) * (x2-x1+1)
           for i in range(y1, y2+1):
                 for j in range(x1, x2+1):
                      var += (image[i][j] - mean)**2
var /= n
            return var

(image[i][j] - mean)**2 = image[i][j] ** 2 -   mean ** 2

region = [x1, y1, x2, y2]

mean = (dp[x2, y2] - dp[x1, y2] - dp[x2, y1] + dp[x1, x2]) / (（x2 - x1 + 1) * (y2 - y1 +1))
squaredsum = (dp2[x2, y2] - dp2[x1, y2] - dp2[x2, y1] + dp2[x1, x2]) / (（x2 - x1 + 1) * (y2 - y1 +1))


follow ups, How to scale it up if we need to compute it 1 billion times?

I propose use to dp table store the accumulate sum of the left top elements,
such as dp, dp2 with same size of the image[hxw], dp[i][j] is the accumulate sum of image[:i+1][:j+1]
dp2[i][j] is the accumulate squared sum of image[:i+1][:j+1]**2….

Round 5, Behavior:

A difficult situation in one year, how you combat it?
How you make a different in you expertise area?
What you will do now, regarding your past project?
A time you have very negative feedback to your colleague?
A time you received a negative feedback?
How you handle tight deadline?
How you implement a general requirement from your manager? I answered firstly discuss the details with manager, figure out what my manager really want, how I can do in practicaly.
What if you want high quality code but your manager wants the project deliver fast?

看着这题目要窒息了。。。https://leetcode.com/problems/parse-lisp-expression/

dsx2018 发表于 2020-07-27 13:24:48
谢谢分享，Round 3是只写了个Fully connected Layer?

嗯是的，还包括每个epoch里面怎么迭代每个iteration，binary cross entropy loss的计算，sgd更新权重的公式。

谢谢分享。

Svm这种传统的模型也能用来做图像分类。不需要做feature engineering。只不过效果不会特别好。

谢谢分享，Round 3是只写了个Fully connected Layer?

一轮做一道hard题还不够，还要做两题才行

感觉g家面的好难啊233

楼主面的职位 job title是什么? 是Software Engineer, Machine Learning吗？

我也想当程序员 发表于 2020-08-13 12:31:11
楼主面的职位 job title是什么? 是Software Engineer, Machine Learning吗？

嗯是Machine Learning Engineer.

tensorboy 发表于 2020-8-14 03:50
嗯是Machine Learning Engineer.

第三轮这种implement model的形式是最近才有的吗？我印象中ML轮 多数是ml design的case study，少数是各种知识点考查。当然我的印象可能是错的。。

看着问题瑟瑟发抖。。