文章目錄

• GAN的原理圖：
• GAN的原版算法描述：
• pytorch實現(xiàn)
• • 構建generator和discriminator:
  • 生成fake data:
  • 生成real data:
  • 定義訓練D的loss，定義訓練G的loss, 實際就是forward pass：
  • 優(yōu)化過程, 實際就是backward pass：
  • • 為了實現(xiàn)fixedGtrainD，fixedDtrainG，我們設計優(yōu)化器更新指定區(qū)域的參數(shù):
    • fixedG trainD：
    • fixedDtrainG:
  • 完整版訓練過程：

GAN的原理圖：

GAN的原版算法描述：

pytorch實現(xiàn)

構建generator和discriminator:

G = nn.Sequential( # Generatornn.Linear(N_IDEAS, 128), # random ideas (could from normal distribution)nn.ReLU(),nn.Linear(128, ART_COMPONENTS), # making a painting from these random ideas )D = nn.Sequential( # Discriminatornn.Linear(ART_COMPONENTS, 128), # receive art work either from the famous artist or a newbie like Gnn.ReLU(),nn.Linear(128, 1),nn.Sigmoid(), # tell the probability that the art work is made by artist )

生成fake data:

G_ideas = torch.randn(BATCH_SIZE, N_IDEAS) # random ideasG_paintings = G(G_ideas) # fake painting from G (random ideas)

生成real data:

def artist_works(): # painting from the famous artist (real target)a = np.random.uniform(1, 2, size=BATCH_SIZE)[:, np.newaxis]paintings = a * np.power(PAINT_POINTS, 2) + (a-1)paintings = torch.from_numpy(paintings).float()return paintingsartist_paintings = artist_works() # real painting from artist

定義訓練D的loss，定義訓練G的loss, 實際就是forward pass：

這個loss就相當于把G和D連接起來了，形成通路了，這里實際上體現(xiàn)了pytorch動態(tài)圖的思想。

prob_artist0 = D(artist_paintings) # D try to increase this probprob_artist1 = D(G_paintings) # D try to reduce this probD_loss = - torch.mean(torch.log(prob_artist0) + torch.log(1. - prob_artist1))G_loss = torch.mean(torch.log(1. - prob_artist1))

優(yōu)化過程, 實際就是backward pass：

為了實現(xiàn)fixedGtrainD，fixedDtrainG，我們設計優(yōu)化器更新指定區(qū)域的參數(shù):

opt_D = torch.optim.Adam(D.parameters(), lr=LR_D) opt_G = torch.optim.Adam(G.parameters(), lr=LR_G)

fixedG trainD：

opt_D.zero_grad()：需要先初始化opt_D，避免前面的數(shù)據(jù)影響當前更新。
D_loss.backward(retain_graph=True) ：計算整個graph梯度，retain_graph=True，需要保持計算圖，啥意思？pytorch默認計算一次backward就釋放當前graph，釋放了就是你必須從頭開始走forward pass ，而這里我們需要重新走一遍原圖的D部分。
opt_D.step()：根據(jù)梯度更新指定區(qū)域的參數(shù)

opt_D.zero_grad()D_loss.backward(retain_graph=True) # reusing computational graphopt_D.step()

fixedDtrainG:

G_loss.backward()：G_loss = torch.mean(torch.log(1. - prob_artist1))，prob_artist1 = D(G_paintings) 可以看出我們需要重新走一遍D（G不需要走），這個是在原來graph上操作的，這就是為什么需要retain graph的原因

opt_G.zero_grad()G_loss.backward()opt_G.step()

完整版訓練過程：

def artist_works(): # painting from the famous artist (real target)a = np.random.uniform(1, 2, size=BATCH_SIZE)[:, np.newaxis]paintings = a * np.power(PAINT_POINTS, 2) + (a-1)paintings = torch.from_numpy(paintings).float()return paintingsG = nn.Sequential( # Generatornn.Linear(N_IDEAS, 128), # random ideas (could from normal distribution)nn.ReLU(),nn.Linear(128, ART_COMPONENTS), # making a painting from these random ideas )D = nn.Sequential( # Discriminatornn.Linear(ART_COMPONENTS, 128), # receive art work either from the famous artist or a newbie like Gnn.ReLU(),nn.Linear(128, 1),nn.Sigmoid(), # tell the probability that the art work is made by artist )opt_D = torch.optim.Adam(D.parameters(), lr=LR_D) opt_G = torch.optim.Adam(G.parameters(), lr=LR_G)plt.ion() # something about continuous plottingfor step in range(10000):artist_paintings = artist_works() # real painting from artistG_ideas = torch.randn(BATCH_SIZE, N_IDEAS) # random ideasG_paintings = G(G_ideas) # fake painting from G (random ideas)prob_artist0 = D(artist_paintings) # D try to increase this probprob_artist1 = D(G_paintings) # D try to reduce this probD_loss = - torch.mean(torch.log(prob_artist0) + torch.log(1. - prob_artist1))G_loss = torch.mean(torch.log(1. - prob_artist1))opt_D.zero_grad()D_loss.backward(retain_graph=True) # reusing computational graphopt_D.step()opt_G.zero_grad()G_loss.backward()opt_G.step()

總結

以上是生活随笔為你收集整理的深度学习总结：GAN，原理，算法描述，pytoch实现的全部內(nèi)容，希望文章能夠幫你解決所遇到的問題。

如果覺得生活随笔網(wǎng)站內(nèi)容還不錯，歡迎將生活随笔推薦給好友。