機器學習中訓練和驗證指標曲線圖能告訴我們什麼？

在本文中將對訓練和驗證可能產生的情況進行總結並介紹這些圖表到底能為我們提供什麼樣的資訊。

讓我們從一些簡單的程式碼開始以下程式碼建立了一個基本的訓練流程框架。

from sklearn.model_selection import train_test_split<br>from sklearn.datasets import make_classification<br>import torch<br>from torch.utils.data import Dataset, DataLoader<br>import torch.optim as torch_optim<br>import torch.nn as nn<br>import torch.nn.functional as F<br>import numpy as np<br>import matplotlib.pyplot as pltclass MyCustomDataset(Dataset):<br>def __init__(self, X, Y, scale=False):<br>self.X = torch.from_numpy(X.astype(np.float32))<br>self.y = torch.from_numpy(Y.astype(np.int64))<br><br>def __len__(self):<br>return len(self.y)<br><br>def __getitem__(self, idx):<br>return self.X[idx], self.y[idx]def get_optimizer(model, lr=0.001, wd=0.0):<br>parameters = filter(lambda p: p.requires_grad, model.parameters())<br>optim = torch_optim.Adam(parameters, lr=lr, weight_decay=wd)<br>return optimdef train_model(model, optim, train_dl, loss_func):<br># Ensure the model is in Training mode<br>model.train()<br>total = 0<br>sum_loss = 0<br>for x, y in train_dl:<br>batch = y.shape[0]<br># Train the model for this batch worth of data<br>logits = model(x)<br># Run the loss function. We will decide what this will be when we call our Training Loop<br>loss = loss_func(logits, y)<br># The next 3 lines do all the PyTorch back propagation goodness<br>optim.zero_grad()<br>loss.backward()<br>optim.step()<br># Keep a running check of our total number of samples in this epoch<br>total += batch<br># And keep a running total of our loss<br>sum_loss += batch*(loss.item())<br>return sum_loss/total<br>def train_loop(model, train_dl, valid_dl, epochs, loss_func, lr=0.1, wd=0):<br>optim = get_optimizer(model, lr=lr, wd=wd)<br>train_loss_list = []<br>val_loss_list = []<br>acc_list = []<br>for i in range(epochs): <br>loss = train_model(model, optim, train_dl, loss_func)<br># After training this epoch, keep a list of progress of <br># the loss of each epoch <br>train_loss_list.append(loss)<br>val, acc = val_loss(model, valid_dl, loss_func)<br># Likewise for the validation loss and accuracy<br>val_loss_list.append(val)<br>acc_list.append(acc)<br>print("training loss: %.5f valid loss: %.5f accuracy: %.5f" % (loss, val, acc))<br><br>return train_loss_list, val_loss_list, acc_list<br>def val_loss(model, valid_dl, loss_func):<br># Put the model into evaluation mode, not training mode<br>model.eval()<br>total = 0<br>sum_loss = 0<br>correct = 0<br>batch_count = 0<br>for x, y in valid_dl:<br>batch_count += 1<br>current_batch_size = y.shape[0]<br>logits = model(x)<br>loss = loss_func(logits, y)<br>sum_loss += current_batch_size*(loss.item())<br>total += current_batch_size<br># All of the code above is the same, in essence, to<br># Training, so see the comments there<br># Find out which of the returned predictions is the loudest<br># of them all, and that's our prediction(s)<br>preds = logits.sigmoid().argmax(1)<br># See if our predictions are right<br>correct += (preds == y).float().mean().item()<br>return sum_loss/total, correct/batch_count<br>def view_results(train_loss_list, val_loss_list, acc_list):<br>plt.rcParams["figure.figsize"] = (15, 5)<br>plt.figure()<br>epochs = np.arange(0, len(train_loss_list)) plt.subplot(1, 2, 1)<br>plt.plot(epochs-0.5, train_loss_list)<br>plt.plot(epochs, val_loss_list)<br>plt.title('model loss')<br>plt.ylabel('loss')<br>plt.xlabel('epoch')<br>plt.legend(['train', 'val', 'acc'], loc = 'upper left')<br><br>plt.subplot(1, 2, 2)<br>plt.plot(acc_list)<br>plt.title('accuracy')<br>plt.ylabel('accuracy')<br>plt.xlabel('epoch')<br>plt.legend(['train', 'val', 'acc'], loc = 'upper left')<br>plt.show()<br><br>def get_data_train_and_show(model, batch_size=128, n_samples=10000, n_classes=2, n_features=30, val_size=0.2, epochs=20, lr=0.1, wd=0, break_it=False):<br># We'll make a fictitious dataset, assuming all relevant<br># EDA / Feature Engineering has been done and this is our <br># resultant data<br>X, y = make_classification(n_samples=n_samples, n_classes=n_classes, n_features=n_features, n_informative=n_features, n_redundant=0, random_state=1972)<br><br>if break_it: # Specifically mess up the data<br>X = np.random.rand(n_samples,n_features)<br>X_train, X_val, y_train, y_val = train_test_split(X, y, test_size=val_size, random_state=1972) train_ds = MyCustomDataset(X_train, y_train)<br>valid_ds = MyCustomDataset(X_val, y_val)<br>train_dl = DataLoader(train_ds, batch_size=batch_size, shuffle=True)<br>valid_dl = DataLoader(valid_ds, batch_size=batch_size, shuffle=True) train_loss_list, val_loss_list, acc_list = train_loop(model, train_dl, valid_dl, epochs=epochs, loss_func=F.cross_entropy, lr=lr, wd=wd)<br>view_results(train_loss_list, val_loss_list, acc_list)

以上的程式碼很簡單，就是取得數據，訓練，驗證這樣一個基本的流程，下面我們開始進入正題。

場景1 - 模型似乎可以學習，但在驗證或準確性方面表現不佳

無論超參數如何，模型Train loss 都會緩慢下降，但Val loss 不會下降，並且其Accuracy 並沒有表明它正在學習任何東西。

例如在這種情況下，二進位分類的準確率徘徊在 50% 左右。

class Scenario_1_Model_1(nn.Module):<br>def __init__(self, in_features=30, out_features=2):<br>super().__init__()<br>self.lin1 = nn.Linear(in_features, out_features)<br>def forward(self, x):<br>x = self.lin1(x)<br>return x<br>get_data_train_and_show(Scenario_1_Model_1(), lr=0.001, break_it=True)

資料中沒有足夠的資訊來允許‘學習’，訓練資料可能沒有包含足夠的資訊來讓模型「學習」。

在這種情況下（程式碼中訓練資料時隨機資料），這意味著它無法學習任何實質內容。

數據必須有足夠的資訊可以從中學習。 EDA 和特徵工程是關鍵！模型學習可以學到的東西，而不是不是編造不存在的東西。

場景2 — 訓練、驗證和準確度曲線都非常不穩定

例如下面程式碼： lr=0.1，bs=128

class Scenario_2_Model_1(nn.Module):<br>def __init__(self, in_features=30, out_features=2):<br>super().__init__()<br>self.lin1 = nn.Linear(in_features, out_features)<br>def forward(self, x):<br>x = self.lin1(x)<br>return x<br>get_data_train_and_show(Scenario_2_Model_1(), lr=0.1)

#“學習率太高”或“批量太小”可以嘗試將學習率從0.1 降低到0.001，這意味著它不會“反彈”，而是會平穩地降低。

get_data_train_and_show(Scenario_1_Model_1(), lr=0.001)

除了降低學習率外，增加批次大小也會使其更平滑。

get_data_train_and_show(Scenario_1_Model_1(), lr=0.001, batch_size=256)

場景3－訓練損失接近零，準確率看起來還不錯，但驗證並沒有下降，而且還上升了

class Scenario_3_Model_1(nn.Module):<br>def __init__(self, in_features=30, out_features=2):<br>super().__init__()<br>self.lin1 = nn.Linear(in_features, 50)<br>self.lin2 = nn.Linear(50, 150)<br>self.lin3 = nn.Linear(150, 50)<br>self.lin4 = nn.Linear(50, out_features)<br>def forward(self, x):<br>x = F.relu(self.lin1(x))<br>x = F.relu(self.lin2(x))<br>x = F.relu(self.lin3(x))<br>x = self.lin4(x)<br>return x<br>get_data_train_and_show(Scenario_3_Model_1(), lr=0.001)

這肯定是過度擬合了：訓練損失低且準確率高，而驗證損失和訓練損失越來越大，都是經典的過度擬合指標。

從根本上來說，你的模型學習能力太強了。它對訓練資料的記憶太好，這意味著它也不能泛化到新資料。

我們可以嘗試的第一件事是降低模型的複雜性。

class Scenario_3_Model_2(nn.Module):<br>def __init__(self, in_features=30, out_features=2):<br>super().__init__()<br>self.lin1 = nn.Linear(in_features, 50)<br>self.lin2 = nn.Linear(50, out_features)<br>def forward(self, x):<br>x = F.relu(self.lin1(x))<br>x = self.lin2(x)<br>return x<br>get_data_train_and_show(Scenario_3_Model_2(), lr=0.001)

這讓它變得更好了，還可以引入 L2 權重衰減正則化，讓它再次變得更好（適用於較淺的模型）。

get_data_train_and_show(Scenario_3_Model_2(), lr=0.001, wd=0.02)

如果我們想保持模型的深度和大小，可以嘗試使用 dropout（適用於更深的模型）。

class Scenario_3_Model_3(nn.Module):<br>def __init__(self, in_features=30, out_features=2):<br>super().__init__()<br>self.lin1 = nn.Linear(in_features, 50)<br>self.lin2 = nn.Linear(50, 150)<br>self.lin3 = nn.Linear(150, 50)<br>self.lin4 = nn.Linear(50, out_features)<br>self.drops = nn.Dropout(0.4)<br>def forward(self, x):<br>x = F.relu(self.lin1(x))<br>x = self.drops(x)<br>x = F.relu(self.lin2(x))<br>x = self.drops(x)<br>x = F.relu(self.lin3(x))<br>x = self.drops(x)<br>x = self.lin4(x)<br>return x<br>get_data_train_and_show(Scenario_3_Model_3(), lr=0.001)

场景 4 - 训练和验证表现良好，但准确度没有提高

lr = 0.001，bs = 128（默认，分类类别= 5

class Scenario_4_Model_1(nn.Module):<br>def __init__(self, in_features=30, out_features=2):<br>super().__init__()<br>self.lin1 = nn.Linear(in_features, 2)<br>self.lin2 = nn.Linear(2, out_features)<br>def forward(self, x):<br>x = F.relu(self.lin1(x))<br>x = self.lin2(x)<br>return x<br>get_data_train_and_show(Scenario_4_Model_1(out_features=5), lr=0.001, n_classes=5)

没有足够的学习能力：模型中的其中一层的参数少于模型可能输出中的类。 在这种情况下，当有 5 个可能的输出类时，中间的参数只有 2 个。

这意味着模型会丢失信息，因为它不得不通过一个较小的层来填充它，因此一旦层的参数再次扩大，就很难恢复这些信息。

所以需要记录层的参数永远不要小于模型的输出大小。

class Scenario_4_Model_2(nn.Module):<br>def __init__(self, in_features=30, out_features=2):<br>super().__init__()<br>self.lin1 = nn.Linear(in_features, 50)<br>self.lin2 = nn.Linear(50, out_features)<br>def forward(self, x):<br>x = F.relu(self.lin1(x))<br>x = self.lin2(x)<br>return x<br>get_data_train_and_show(Scenario_4_Model_2(out_features=5), lr=0.001, n_classes=5)

总结

以上就是一些常见的训练、验证时的曲线的示例，希望你在遇到相同情况时可以快速定位并且改进。

以上是機器學習中訓練和驗證指標曲線圖能告訴我們什麼？的詳細內容。更多資訊請關注PHP中文網其他相關文章！