Visualization

[dongheechoi]

bss_prediction_1019_16.pdf

Embedding plot

• http://163.152.20.6:8888/lab/tree/dh.ipynb
• https://github.com/DaehanKim/bss_demand_prediction/blob/master/dh.ipynb

Demand plot

• http://gstar1:8888/lab/tree/demand.ipynb
• https://github.com/DaehanKim/bss_demand_prediction/blob/master/demand.ipynb

암호 : 1234

시각화

Station Embedding 시각화

• Station Embedding vector를 t-SNE로 시각화
• 특정 statition을 다르게 ( 색깔이라던가 ) 표시할 수 있는 기능
• 지역별 위치를 나타낸 fig랑 비교할 예정

time embedding 시각화

• hour( 24dim ) * weeday or end ( 2dim ) t-SNE시각화

Demand 시각화

• x축 시간 y축 Demand - Plot 예시 )

Todo-list

• [x] git에서 받아서 서버에서 실행
• 모델 저장하는 부분 구현
  • [x] 그냥 Jupyter로 돌려서 시작
• [x] Station 시각화
• [x] time-embedding 시각화
• [ ] Demand 시각화

[dongheechoi]

http://163.152.20.6:8888/lab/tree/dh.ipynb 암호 : 1234

• tSNE plot했음.
• Embedding에서 값 가져 오는 것 필요.
  • 대충 코드 읽고 m.hour_embedding(torch.FloatTensor([0.,1.]).cuda()) 이런식으로 해봤는데 안됨;;

[dongheechoi]

해결 -> LongTensor였어야 함. m.hour_embedding(torch.LongTensor([0.,1.]).cuda())

[dongheechoi]

• Hour embedding

  Code

  from sklearn.manifold import TSNE
  %matplotlib inline
  import matplotlib.pyplot as plt
  from mpl_toolkits.mplot3d import Axes3D
  arr = m.hour_embedding(torch.LongTensor([float(i) for i in range(24)]).cuda()).cpu().detach().numpy()
  tsne = TSNE(n_components=2, verbose=1, perplexity=40, n_iter=300)
  tsne_results = tsne.fit_transform(arr)

res array([[ 6.52357407e+01, 2.97109871e+01], [ 7.22931480e+00, -1.17531822e+02], [ 9.38050938e+00, 1.25519302e+02], [-5.00310135e+01, -2.23163605e+01], [-1.00831932e+02, -7.36254959e+01], [ 8.21195374e+01, -1.00142166e+02], [ 1.58067169e+02, -8.27860413e+01], [ 4.22641830e+01, -5.91689301e+01], [-4.04240417e+01, 8.13821793e+01], [-8.41493912e+01, 1.18714699e+02], [-5.41209459e+00, -4.35924416e+01], [-4.62892342e+01, 1.74711166e+02], [-1.54264862e+02, -1.80793495e+01], [ 2.49544926e+01, 1.54471397e-02], [ 7.10910492e+01, 9.26024246e+01], [ 1.06514442e+02, -2.95543060e+02], [-2.05398598e+01, 2.42003193e+01], [-7.22683792e+01, 3.53312378e+01], [-1.01585526e+02, -1.00038815e+01], [ 1.22630280e+02, 3.59977837e+01], [-1.23820366e+02, 6.02001953e+01], [ 9.57163010e+01, -2.74828587e+01], [ 1.43090353e+01, 6.53545151e+01], [-4.44189644e+01, -8.75185699e+01]], dtype=float32)

[dongheechoi]

조금씩 다르게 나옴

[dongheechoi]

숫자 붙여봄

[dongheechoi]

• Day / Night concat day embedding에서 0 index를 d로 나타내고, 1을 n으로 나타냄

def plot2(arr):
    tsne = TSNE(n_components=2, verbose=1, perplexity=40, n_iter=300)
    tsne_results = tsne.fit_transform(arr)
    tr = tsne_results.transpose()
    x1 = tr[0][:24]
    y1 = tr[1][:24]
    x2 = tr[0][24:]
    y2 = tr[1][24:]

    plt.figure(figsize=(16,10))
    fig, ax = plt.subplots()
    ax.scatter(x=x1, y=y1, marker='^')
    ax.scatter(x=x2, y=y2, marker='o')
    for i in range(len(x1)):
        ax.annotate('d'+str(i), (x1[i], y1[i]))
        ax.annotate('n'+str(i), (x2[i], y2[i]))
    return tsne_results
hd_arr = np.vstack(
    [[np.array([np.append(he, d_arr[0]) for he in h_arr])],
    [np.array([np.append(he, d_arr[1]) for he in h_arr])]]
)

# 0~23 : 
arr = hd_arr.reshape(-1, 64)
tnse_res = plot2(arr)

[dongheechoi]

• locatioin embedding

  arr = m.location_embedding(torch.LongTensor([float(i) for i in range(171)]).cuda())
  arr = arr.cpu().detach().numpy()
  tnse_res = plot(arr)

  

[dongheechoi]

• demand는 어떤걸 봐야 하는지 잘 모르겠음.

[dongheechoi]

조정하니 좀 나뉘긴함.

[dongheechoi]

이거도 좀 신기하게 나뉘네

[DaehanKim]

시각화 고생하셨습니다!