03wk-2: (로지스틱) – sig(linr(x)), BCELoss, Adam, 로지스틱의 한계

1. 강의영상

2. Imports

import torch
import matplotlib.pyplot as plt 
import numpy as np
import pandas as pd

plt.rcParams['figure.figsize'] = (4.5, 3.0)

3. 로지스틱 – sig(linr(x))

A. 로지스틱 모형

- \(x\)가 커질수록 (혹은 작아질수록) \(y=1\)이 잘나오는 모형은 아래와 같이 설계할 수 있음 <— 외우세요!!!

• \(y_i \sim {\cal B}(\pi_i),\quad\) where \(\pi_i = \frac{\exp(w_0+w_1x_i)}{1+\exp(w_0+w_1x_i)} = \frac{1}{1+\exp(-w_0-w_1x_i)}\)

• \(\hat{y}_i= \frac{\exp(\hat{w}_0+\hat{w}_1x_i)}{1+\exp(\hat{w}_0+\hat{w}_1x_i)}=\frac{1}{1+\exp(-\hat{w}_0-\hat{w}_1x_i)}\)

- 회귀모형과 로지스틱 모형의 비교

• 회귀모형: \(y_i \sim {\cal N}(w_0+w_1x_i, \sigma^2)\)¹
• 로지스틱: \(y_i \sim {\cal B}\big(\frac{\exp(w_0+w_1x_i)}{1+\exp(w_0+w_1x_i)}\big)\)

¹ 원래는 이렇게 썼었지.. \(y_i = w_0 + w_1x_i + \epsilon_i \quad \epsilon_i \sim {\cal N}(0,\sigma^2)\)

- 우리가 예측하고 싶은것

• 회귀모형: 정규분포의 평균을 예측하고 싶음. 즉 \(w_0+w_1x_i\)를 예측하고 싶음. 예측값으로는 \(\hat{w}_0 + \hat{w}_1x_i\)를 사용!
• 로지스틱: 베르누이의 평균을 예측하고 싶음. 즉 \(\frac{\exp(w_0+w_1x_i)}{1+\exp(w_0+w_1x_i)}\)를 예측하고 싶음. 예측값으로는 \(\frac{\exp(\hat{w}_0+\hat{w}_1x_i)}{1+\exp(\hat{w}_0+\hat{w}_1x_i)}\)를 사용!

B. 데이터 – 스펙과 취업

torch.manual_seed(43052)
x = torch.linspace(-1,1,2000).reshape(2000,1)
w0,w1 = -1, 5
prob = torch.exp(w0+w1*x) / (1+torch.exp(w0+w1*x)) 
y = torch.bernoulli(prob)

plt.plot(x,y,'.',alpha=0.03)
plt.plot(x[0],y[0],'.',label=r"$(x_i,y_i)$",color="C0")
plt.plot(x,prob,'--r',label=r"prob (true, unknown) = $\frac{exp(-1+5x)}{1+exp(-1+5x)}$")
plt.legend()

C. Step1: net 설계 (모델링)

- 최초의 곡선을 그려보자.

• 최초의직선: \(\hat{y}_i= \hat{w}_0+\hat{w}_1x_i\) 에서 아무 \(\hat{w}_0\), \(\hat{w}_1\) 을 설정하면 된다.
• 최초의곡선: \(\hat{y}_i= \frac{\exp(\hat{w}_0+\hat{w}_1x_i)}{1+\exp(\hat{w}_0+\hat{w}_1x_i)}=\frac{1}{1+\exp(-\hat{w}_0-\hat{w}_1x_i)}\) 에서 아무 \(\hat{w}_0\), \(\hat{w}_1\) 을 설정하면 된다.

Note

일단은 초기 설정값을 \(\hat{w}_0 = -0.8\), \(\hat{w}_1 = -0.3\) 으로 하자. (실제값은 \(w_0=-1\), \(w_1=5\) 이다)

# 방법1 – l1, sigmoid

l1 = torch.nn.Linear(1,1)
l1(x) # w0hat + w1hat*x 

tensor([[ 0.6311],
        [ 0.6304],
        [ 0.6297],
        ...,
        [-0.6902],
        [-0.6909],
        [-0.6916]], grad_fn=<AddmmBackward0>)

l1.weight.data = torch.tensor([[-0.3]])
l1.bias.data = torch.tensor([-0.8])

def sigmoid(x):
    return torch.exp(x)/(1+torch.exp(x))

plt.plot(x,y,'.',alpha=0.03)
plt.plot(x[0],y[0],'o',label=r"$(x_i,y_i)$",color="C0")
plt.plot(x,prob,'--r',label=r"prob (true, unknown) = $\frac{exp(-1+5x)}{1+exp(-1+5x)}$")
plt.plot(x,sigmoid(l1(x)).data,'--b', label=r"prob (estimated) = $(x_i,\hat{y}_i)$ -- first curve")
plt.legend()

#

# 방법2 – l1, a1

l1 = torch.nn.Linear(1,1)
l1.weight.data = torch.tensor([[-0.3]])
l1.bias.data = torch.tensor([-0.8])

a1 = torch.nn.Sigmoid()

sigmoid(l1(x)), a1(l1(x)) # 똑같아요

(tensor([[0.3775],
         [0.3775],
         [0.3774],
         ...,
         [0.2499],
         [0.2498],
         [0.2497]], grad_fn=<DivBackward0>),
 tensor([[0.3775],
         [0.3775],
         [0.3774],
         ...,
         [0.2499],
         [0.2498],
         [0.2497]], grad_fn=<SigmoidBackward0>))

- 지금까지의 구현 확인

plt.plot(x,y,'.',alpha=0.03)
plt.plot(x[0],y[0],'o',label=r"$(x_i,y_i)$",color="C0")
plt.plot(x,prob,'--r',label=r"prob (true, unknown) = $\frac{exp(-1+5x)}{1+exp(-1+5x)}$")
plt.plot(x,a1(l1(x)).data,'--b', label=r"prob (estimated) = $(x_i,\hat{y}_i)$ -- first curve with $(a_1 \circ l_1)(x)$")
plt.legend()

#

# 방법3 - l1, a1 만들고 \(\to\) net

- 관찰: 지금 아래의 구조이다.

\[{\bf x} \overset{l_1}{\to} {\bf u} \overset{a_1}{\to} {\bf v} = \hat{\bf y}\]

- 소망: 함수 \(l_1, a_1\) 의 합성을 하나로 묶어서

\[(a_1\circ l_1)({\bf x}) := net({\bf x})\]

이러한 기능을 하는 하나의 함수 \(net\)을 만들 수 없을까?

l1 = torch.nn.Linear(1,1)
l1.weight.data = torch.tensor([[-0.3]])
l1.bias.data = torch.tensor([-0.8])
a1 = torch.nn.Sigmoid()

net = torch.nn.Sequential(l1,a1) #l1을 취하고 그다음에 a1을 취하라는 의미

net(x), a1(l1(x)), sigmoid(l1(x))

(tensor([[0.3775],
         [0.3775],
         [0.3774],
         ...,
         [0.2499],
         [0.2498],
         [0.2497]], grad_fn=<SigmoidBackward0>),
 tensor([[0.3775],
         [0.3775],
         [0.3774],
         ...,
         [0.2499],
         [0.2498],
         [0.2497]], grad_fn=<SigmoidBackward0>),
 tensor([[0.3775],
         [0.3775],
         [0.3774],
         ...,
         [0.2499],
         [0.2498],
         [0.2497]], grad_fn=<DivBackward0>))

* net 구조 잠깐 살펴보기

net[0], net[1]

(Linear(in_features=1, out_features=1, bias=True), Sigmoid())

l1 is net[0]

True

a1 is net[1]

True

#

# 방법4 – net을 바로 만들기

net = torch.nn.Sequential(
    torch.nn.Linear(1,1),
    torch.nn.Sigmoid()
)
net[0].weight.data = torch.tensor([[-0.3]])
net[0].bias.data = torch.tensor([-0.8])
yhat = net(x)

net(x)

tensor([[0.3775],
        [0.3775],
        [0.3774],
        ...,
        [0.2499],
        [0.2498],
        [0.2497]], grad_fn=<SigmoidBackward0>)

#

D. Step1~4

net = torch.nn.Sequential(
    torch.nn.Linear(in_features=1, out_features=1),
    torch.nn.Sigmoid()
)
l1, a1 = net 
l1.weight.data = torch.tensor([[-0.3]])
l1.bias.data = torch.tensor([-0.8])
optimizr = torch.optim.SGD(net.parameters(),lr=0.25)
#---#
for epoc in range(100):
    ## 1
    yhat = net(x) 
    ## 2 
    loss = torch.mean((y-yhat)**2)
    ## 3
    loss.backward()
    ## 4 
    optimizr.step()
    optimizr.zero_grad()

plt.plot(x,y,'.',alpha=0.05)
plt.plot(x,prob,'--r')
plt.plot(x,yhat.data,'--b')
plt.title('after 100 epochs')

Text(0.5, 1.0, 'after 100 epochs')

for epoc in range(4900):
    ## 1
    yhat = net(x) 
    ## 2 
    loss = torch.mean((y-yhat)**2)
    ## 3
    loss.backward()
    ## 4 
    optimizr.step()
    optimizr.zero_grad()

plt.plot(x,y,'.',alpha=0.05)
plt.plot(x,prob,'--r')
plt.plot(x,yhat.data,'--b')
plt.title('after 5000 epochs')

Text(0.5, 1.0, 'after 5000 epochs')

4. 학습과정 시각화 및 문제인식

A. 시각화를 위한 준비

def plot_loss(loss_fn, ax=None, Wstar=[-1,5]):
    w0hat,w1hat =torch.meshgrid(torch.arange(-10,3,0.1),torch.arange(-1,10,0.1),indexing='ij')
    w0hat = w0hat.reshape(-1)
    w1hat = w1hat.reshape(-1)
    def l(w0hat,w1hat):
        yhat = torch.exp(w0hat+w1hat*x)/(1+torch.exp(w0hat+w1hat*x))
        return loss_fn(yhat,y) 
    loss = list(map(l,w0hat,w1hat))
    #---#
    if ax is None: 
        fig = plt.figure()
        ax = fig.add_subplot(1,1,1,projection='3d')
    ax.scatter(w0hat,w1hat,loss,s=0.001) 
    ax.scatter(w0hat[::20],w1hat[::20],loss[::20],s=0.1,color='C0') 
    w0star,w1star = np.array(Wstar).reshape(-1)
    ax.scatter(w0star,w1star,l(w0star,w1star),s=200,marker='*',color='red',label=f"W=[{w0star:.1f},{w1star:.1f}]")
    #---#
    ax.elev = 15
    ax.dist = -20
    ax.azim = 75    
    ax.legend()
    ax.set_xlabel(r'$w_0$')  # x축 레이블 설정
    ax.set_ylabel(r'$w_1$')  # y축 레이블 설정
    ax.set_xticks([-10,-5,0])  # x축 틱 간격 설정
    ax.set_yticks([-10,0,10])  # y축 틱 간격 설정

def _learn_and_record(net, loss_fn, optimizr):
    yhat_history = [] 
    loss_history = []
    What_history = []
    Whatgrad_history = []
    What_history.append([net[0].bias.data.item(), net[0].weight.data.item()])
    for epoc in range(100): 
        ## step1 
        yhat = net(x)
        ## step2 
        loss = loss_fn(yhat,y)
        ## step3
        loss.backward() 
        ## step4 
        optimizr.step()
        ## record 
        if epoc % 5 ==0: 
            yhat_history.append(yhat.reshape(-1).data.tolist())
            loss_history.append(loss.item())
            What_history.append([net[0].bias.data.item(), net[0].weight.data.item()])
            Whatgrad_history.append([net[0].bias.grad.item(), net[0].weight.grad.item()])
        optimizr.zero_grad() 
        
    return yhat_history, loss_history, What_history, Whatgrad_history
    
def show_animation(net, loss_fn, optimizr):
    yhat_history,loss_history,What_history,Whatgrad_history = _learn_and_record(net,loss_fn,optimizr)
    
    fig = plt.figure(figsize=(7.5,3.5))
    ax1 = fig.add_subplot(1, 2, 1)
    ax2 = fig.add_subplot(1, 2, 2, projection='3d')
    ## ax1: 왼쪽그림 
    ax1.scatter(x,y,alpha=0.01)
    ax1.scatter(x[0],y[0],color='C0',label=r"observed data = $(x_i,y_i)$")
    ax1.plot(x,prob,'--',label=r"prob (true) = $(x_i,\frac{exp(-1+5x_i)}{1+exp(-1+5x_i)})$")    
    line, = ax1.plot(x,yhat_history[0],'--',label=r"prob (estimated) = $(x_i,\hat{y}_i)$") 
    ax1.legend()
    ## ax2: 오른쪽그림 
    plot_loss(loss_fn,ax2)
    ax2.scatter(np.array(What_history)[0,0],np.array(What_history)[0,1],loss_history[0],color='blue',s=200,marker='*')    
    def animate(epoc):
        line.set_ydata(yhat_history[epoc])
        w0hat = np.array(What_history)[epoc,0]
        w1hat = np.array(What_history)[epoc,1]
        w0hatgrad = np.array(Whatgrad_history)[epoc,0]
        w1hatgrad = np.array(Whatgrad_history)[epoc,1]
        ax2.scatter(w0hat,w1hat,loss_history[epoc],color='grey')
        ax2.set_title(f"What.grad=[{w0hatgrad:.4f},{w1hatgrad:.4f}]",y=0.8)
        fig.suptitle(f"epoch={epoc*5} // What=[{w0hat:.2f},{w1hat:.2f}] // Loss={loss_fn.__class__.__name__} // Opt={optimizr.__class__.__name__}")
        return line
    ani = animation.FuncAnimation(fig, animate, frames=20)    
    plt.close()
    return ani

from matplotlib import animation
plt.rcParams["animation.html"] = "jshtml"

함수사용법

loss_fn = torch.nn.MSELoss()
plot_loss(loss_fn)

torch.manual_seed(42)
net = torch.nn.Sequential(
    torch.nn.Linear(1,1),
    torch.nn.Sigmoid()
) 
loss_fn = torch.nn.MSELoss()
optimizr = torch.optim.SGD(net.parameters(),lr=0.25) 
show_animation(net,loss_fn,optimizr)

Once Loop Reflect

B. 좋은 초기값

net = torch.nn.Sequential(
    torch.nn.Linear(1,1),
    torch.nn.Sigmoid()
) 
net[0].bias.data = torch.tensor([-0.8])
net[0].weight.data = torch.tensor([[-0.3]])
loss_fn = torch.nn.MSELoss()
optimizr = torch.optim.SGD(net.parameters(),lr=0.25) 
#---#
show_animation(net,loss_fn,optimizr)

Once Loop Reflect

C. 가능성 있는 초기값

net = torch.nn.Sequential(
    torch.nn.Linear(1,1),
    torch.nn.Sigmoid()
) 
net[0].bias.data = torch.tensor([-3.0])
net[0].weight.data = torch.tensor([[-1.0]])
loss_fn = torch.nn.MSELoss()
optimizr = torch.optim.SGD(net.parameters(),lr=0.25) 
#---#
show_animation(net,loss_fn,optimizr)

Once Loop Reflect

D. 최악의 초기값

net = torch.nn.Sequential(
    torch.nn.Linear(1,1),
    torch.nn.Sigmoid()
) 
net[0].bias.data = torch.tensor([-10.0])
net[0].weight.data = torch.tensor([[-1.0]])
loss_fn = torch.nn.MSELoss()
optimizr = torch.optim.SGD(net.parameters(),lr=0.25) 
#---#
show_animation(net,loss_fn,optimizr)

Once Loop Reflect

해결하는 접근법:

• 컴공스타일: 에폭을 늘려볼까?
• 산공스타일: 옵티마이저를 바꿔볼까?
• 통계스타일: Loss를 바꿔볼까?

5. 손실함수의 개선

A. BCE Loss를 사용하여 학습

- BCE loss라는게 있음.

• \(loss= - \sum_{i=1}^{n} \big(y_i\log(\hat{y}_i)+(1-y_i)\log(1-\hat{y}_i)\big)\)
• https://en.wikipedia.org/wiki/Cross-entropy

net = torch.nn.Sequential(
    torch.nn.Linear(in_features=1, out_features=1),
    torch.nn.Sigmoid()
)
l1, a1 = net 
l1.weight.data = torch.tensor([[-0.3]])
l1.bias.data = torch.tensor([-0.8])
optimizr = torch.optim.SGD(net.parameters(),lr=0.25)
#---#
for epoc in range(100):
    ## 1
    yhat = net(x) 
    ## 2 
    #loss = torch.mean((y-yhat)**2) # loss_fn(yhat,y)
    loss = -torch.mean(y*torch.log(yhat) + (1-y)*torch.log(1-yhat))
    ## 3
    loss.backward()
    ## 4 
    optimizr.step()
    optimizr.zero_grad()

plt.plot(x,y,'.',alpha=0.05)
plt.plot(x,prob,'--r')
plt.plot(x,yhat.data,'--b')
plt.title('after 100 epochs')

Text(0.5, 1.0, 'after 100 epochs')

같은 100 에폭인데 훨씬 잘맞춤..

- loss수식을 못외우겠다면?

net = torch.nn.Sequential(
    torch.nn.Linear(in_features=1, out_features=1),
    torch.nn.Sigmoid()
)
l1, a1 = net 
l1.weight.data = torch.tensor([[-0.3]])
l1.bias.data = torch.tensor([-0.8])
loss_fn = torch.nn.BCELoss()
optimizr = torch.optim.SGD(net.parameters(),lr=0.25)
#---#
for epoc in range(100):
    ## 1
    yhat = net(x) 
    ## 2 
    loss = loss_fn(yhat,y) # yhat부터 써야함
    ## 3
    loss.backward()
    ## 4 
    optimizr.step()
    optimizr.zero_grad()

plt.plot(x,y,'.',alpha=0.05)
plt.plot(x,prob,'--r')
plt.plot(x,yhat.data,'--b')
plt.title('after 100 epochs')

Text(0.5, 1.0, 'after 100 epochs')

B. Loss Function 시각화

plot_loss(torch.nn.MSELoss())

plot_loss(torch.nn.BCELoss())

- 비교해보자.

fig = plt.figure()
ax1 = fig.add_subplot(1,2,1,projection='3d')
ax2 = fig.add_subplot(1,2,2,projection='3d')
plot_loss(torch.nn.MSELoss(),ax1)
plot_loss(torch.nn.BCELoss(),ax2)

C. 학습과정 시각화 – 좋은 초기값

- MSELoss

net = torch.nn.Sequential(
    torch.nn.Linear(1,1),
    torch.nn.Sigmoid()
) 
net[0].bias.data = torch.tensor([-0.8])
net[0].weight.data = torch.tensor([[-0.3]])
loss_fn = torch.nn.MSELoss()
optimizr = torch.optim.SGD(net.parameters(),lr=0.25) 
#---#
show_animation(net,loss_fn,optimizr)

Once Loop Reflect

- BCELoss

net = torch.nn.Sequential(
    torch.nn.Linear(1,1),
    torch.nn.Sigmoid()
) 
net[0].bias.data = torch.tensor([-0.8])
net[0].weight.data = torch.tensor([[-0.3]])
loss_fn = torch.nn.BCELoss()
optimizr = torch.optim.SGD(net.parameters(),lr=0.25) 
#---#
show_animation(net,loss_fn,optimizr)

Once Loop Reflect

D. 학습과정 시각화 – 가능성 있는 초기값

- MSELoss

net = torch.nn.Sequential(
    torch.nn.Linear(1,1),
    torch.nn.Sigmoid()
) 
net[0].bias.data = torch.tensor([-3.0])
net[0].weight.data = torch.tensor([[-1.0]])
loss_fn = torch.nn.MSELoss()
optimizr = torch.optim.SGD(net.parameters(),lr=0.25) 
#---#
show_animation(net,loss_fn,optimizr)

Once Loop Reflect

- BCELoss

net = torch.nn.Sequential(
    torch.nn.Linear(1,1),
    torch.nn.Sigmoid()
) 
net[0].bias.data = torch.tensor([-3.0])
net[0].weight.data = torch.tensor([[-1.0]])
loss_fn = torch.nn.BCELoss()
optimizr = torch.optim.SGD(net.parameters(),lr=0.25) 
#---#
show_animation(net,loss_fn,optimizr)

Once Loop Reflect

E. 학습과정 시각화 – 최악의 초기값

- MSELoss

net = torch.nn.Sequential(
    torch.nn.Linear(1,1),
    torch.nn.Sigmoid()
) 
net[0].bias.data = torch.tensor([-10.0])
net[0].weight.data = torch.tensor([[-1.0]])
loss_fn = torch.nn.MSELoss()
optimizr = torch.optim.SGD(net.parameters(),lr=0.25) 
#---#
show_animation(net,loss_fn,optimizr)

Once Loop Reflect

- BCELoss

net = torch.nn.Sequential(
    torch.nn.Linear(1,1),
    torch.nn.Sigmoid()
) 
net[0].bias.data = torch.tensor([-10.0])
net[0].weight.data = torch.tensor([[-1.0]])
loss_fn = torch.nn.BCELoss()
optimizr = torch.optim.SGD(net.parameters(),lr=0.25) 
#---#
show_animation(net,loss_fn,optimizr)

Once Loop Reflect

6. 옵티마이저의 개선

A. 학습과정 시각화 – 좋은 초기값

- MSELoss + SGD

net = torch.nn.Sequential(
    torch.nn.Linear(1,1),
    torch.nn.Sigmoid()
) 
net[0].bias.data = torch.tensor([-0.8470])
net[0].weight.data = torch.tensor([[-0.3467]])
loss_fn = torch.nn.MSELoss()
optimizr = torch.optim.SGD(net.parameters(),lr=0.25) 
#---#
show_animation(net,loss_fn,optimizr)

Once Loop Reflect

- MSELoss + Adam

net = torch.nn.Sequential(
    torch.nn.Linear(1,1),
    torch.nn.Sigmoid()
) 
net[0].bias.data = torch.tensor([-0.8])
net[0].weight.data = torch.tensor([[-0.3]])
loss_fn = torch.nn.MSELoss()
optimizr = torch.optim.Adam(net.parameters(),lr=0.25) 
#---#
show_animation(net,loss_fn,optimizr)

Once Loop Reflect

B. 학습과정 시각화 – 가능성 있는 초기값

- MSELoss + SGD

net = torch.nn.Sequential(
    torch.nn.Linear(1,1),
    torch.nn.Sigmoid()
) 
net[0].bias.data = torch.tensor([-3.0])
net[0].weight.data = torch.tensor([[-1.0]])
loss_fn = torch.nn.MSELoss()
optimizr = torch.optim.SGD(net.parameters(),lr=0.25) 
#---#
show_animation(net,loss_fn,optimizr)

Once Loop Reflect

- MSELoss + Adam

net = torch.nn.Sequential(
    torch.nn.Linear(1,1),
    torch.nn.Sigmoid()
) 
net[0].bias.data = torch.tensor([-3.0])
net[0].weight.data = torch.tensor([[-1.0]])
loss_fn = torch.nn.MSELoss()
optimizr = torch.optim.Adam(net.parameters(),lr=0.25) 
#---#
show_animation(net,loss_fn,optimizr)

Once Loop Reflect

C. 학습과정 시각화 – 최악의 초기값

- MSELoss + SGD

net = torch.nn.Sequential(
    torch.nn.Linear(1,1),
    torch.nn.Sigmoid()
) 
net[0].bias.data = torch.tensor([-10.0])
net[0].weight.data = torch.tensor([[-1.0]])
loss_fn = torch.nn.MSELoss()
optimizr = torch.optim.SGD(net.parameters(),lr=0.05) 
#---#
show_animation(net,loss_fn,optimizr)

Once Loop Reflect

- MSELoss + Adam

net = torch.nn.Sequential(
    torch.nn.Linear(1,1),
    torch.nn.Sigmoid()
) 
net[0].bias.data = torch.tensor([-10.0])
net[0].weight.data = torch.tensor([[-1.0]])
loss_fn = torch.nn.MSELoss()
optimizr = torch.optim.Adam(net.parameters(),lr=0.25) 
#---#
show_animation(net,loss_fn,optimizr)

Once Loop Reflect

D. 참고자료

https://www.youtube.com/watch?v=MD2fYip6QsQ

• 11:50 – Momentum
• 12:30 – RMSprop
• 15:55 – Adam

7. 로지스틱의 한계

A. 신문기사 (데이터의 모티브)

- 스펙이 높아도 취업이 안된다고 합니다..

중소·지방 기업 “뽑아봤자 그만두니까”

중소기업 관계자들은 고스펙 지원자를 꺼리는 이유로 높은 퇴직률을 꼽는다. 여건이 좋은 대기업으로 이직하거나 회사를 관두는 경우가 많다는 하소연이다. 고용정보원이 지난 3일 공개한 자료에 따르면 중소기업 청년취업자 가운데 49.5%가 2년 내에 회사를 그만두는 것으로 나타났다.

중소 IT업체 관계자는 “기업 입장에서 가장 뼈아픈 게 신입사원이 그만둬서 새로 뽑는 일”이라며 “명문대 나온 스펙 좋은 지원자를 뽑아놔도 1년을 채우지 않고 그만두는 사원이 대부분이라 우리도 눈을 낮춰 사람을 뽑는다”고 말했다.

B. 가짜데이터 – 스펙의 역설

df = pd.read_csv("https://raw.githubusercontent.com/guebin/DL2025/main/posts/ironyofspec.csv")
df

	x	prob	y
0	-1.000000	0.000045	0.0
1	-0.998999	0.000046	0.0
2	-0.997999	0.000047	0.0
3	-0.996998	0.000047	0.0
4	-0.995998	0.000048	0.0
...	...	...	...
1995	0.995998	0.505002	0.0
1996	0.996998	0.503752	0.0
1997	0.997999	0.502501	0.0
1998	0.998999	0.501251	1.0
1999	1.000000	0.500000	1.0

2000 rows × 3 columns

x = torch.tensor(df.x).float().reshape(-1,1)
y = torch.tensor(df.y).float().reshape(-1,1)
prob = torch.tensor(df.prob).float().reshape(-1,1)

plt.plot(x,y,'o',alpha=0.02)
plt.plot(x[0],y[0],'o',label= r"observed data = $(x_i,y_i)$",color="C0")
plt.plot(x,prob,'--b',label= r"prob (true, unknown)")
plt.legend()

C. 로지스틱으로 적합

torch.manual_seed(43052)
net = torch.nn.Sequential(
    torch.nn.Linear(1,1),
    torch.nn.Sigmoid()
)
loss_fn = torch.nn.BCELoss()
optimizr = torch.optim.Adam(net.parameters())
#---# 
for epoc in range(5000):
    ## 1 
    yhat = net(x)
    ## 2 
    loss = loss_fn(yhat,y)
    ## 3 
    loss.backward()
    ## 4 
    optimizr.step()
    optimizr.zero_grad()

plt.plot(x,y,'o',alpha=0.02)
plt.plot(x[0],y[0],'o',label= r"observed data = $(x_i,y_i)$",color="C0")
plt.plot(x,prob,'--b',label= r"prob (true, unknown)")
plt.plot(x,net(x).data, '--', label= r"prob (estimated) = $(x_i,\hat{y}_i)$")
plt.legend()

- Epoch을 10억번으로 설정해도 이건 못 맞출것 같음.

D. 로지스틱 한계극복 – 아이디어만

- sigmoid를 넣기 전의 상태가 직선이 아니라 꺽이는 직선이야 한다.

a = torch.nn.Sigmoid()

fig,ax = plt.subplots(4,2,figsize=(8,8))
u1 = torch.tensor([-6,-4,-2,0,2,4,6])
u2 = torch.tensor([6,4,2,0,-2,-4,-6])
u3 = torch.tensor([-6,-2,2,6,2,-2,-6])
u4 = torch.tensor([-6,-2,2,6,4,2,0])
ax[0,0].plot(u1,'--o',color='C0',label = r"$u_1$")
ax[0,0].legend()
ax[0,1].plot(a(u1),'--o',color='C0',label = r"$a(u_1)=\frac{exp(u_1)}{exp(u_1)+1}$")
ax[0,1].legend()
ax[1,0].plot(u2,'--o',color='C1',label = r"$u_2$")
ax[1,0].legend()
ax[1,1].plot(a(u2),'--o',color='C1',label = r"$a(u_2)=\frac{exp(u_2)}{exp(u_2)+1}$")
ax[1,1].legend()
ax[2,0].plot(u3,'--o',color='C2', label = r"$u_3$")
ax[2,0].legend()
ax[2,1].plot(a(u3),'--o',color='C2', label = r"$a(u_3)=\frac{exp(u_3)}{exp(u_3)+1}$")
ax[2,1].legend()
ax[3,0].plot(u4,'--o',color='C3', label = r"$u_4$")
ax[3,0].legend()
ax[3,1].plot(a(u4),'--o',color='C3', label = r"$a(u_4)=\frac{exp(u_4)}{exp(u_4)+1}$")
ax[3,1].legend()