• CS224W Colab_2 笔记

    目录

    1、PyG中数据集的属性查询（Question 1，类别数目和特征数目 Question 2，3）

    2、OGB包的使用（Question 4：How many features are in the ogbn-arxiv graph?） 

    3、节点分类任务GNN框架的实现与初步理解

    ① torch.nn.ModuleList()类、torch.nn.LogSoftmax()类的使用

    ② PyG中GCNConv()类、BatchNorm()类的使用

    ③ 此处GNN类的作用/构成理解：

    ④ torch.flatten(input, start_dim=0, end_dim=-1)：去掉start~end的维度，把结果包装成一个“一维Tensor”并返回。

    ​编辑

    ⑤ train函数的理解，def train(model, data, train_idx, optimizer, loss_fn):

    ⑥ loss_fn函数：

    ⑦ test(...)方法中的out.argmax(......)函数使用：如代码注释

    4、图级别预测

    ① 框架构建思路：

    ② DataLoader与batch：

    ③ train函数中的tips：

    1、PyG中数据集的属性查询（Question 1，类别数目和特征数目 Question 2，3）

    1. # Question 1：What is the number of classes and number of features in the ENZYMES dataset?
    2. def get_num_classes(pyg_dataset):
    3.   # TODO: Implement a function that takes a PyG dataset object
    4.   # and returns the number of classes for that dataset.
    5.  
    6.   num_classes = 0
    7.  
    8.   ############# Your code here ############
    9.   ## (~1 line of code)
    10.   ## Note
    11.   ## 1. Colab autocomplete functionality might be useful.
    12.  
    13.   #########################################
    14.   num_classes = pyg_dataset.num_classes
    15.   return num_classes
    16.  
    17. def get_num_features(pyg_dataset):
    18.   # TODO: Implement a function that takes a PyG dataset object
    19.   # and returns the number of features for that dataset.
    20.  
    21.   num_features = 0
    22.  
    23.   ############# Your code here ############
    24.   ## (~1 line of code)
    25.   ## Note
    26.   ## 1. Colab autocomplete functionality might be useful.
    27.  
    28.   #########################################
    29.   num_features = pyg_dataset.num_features
    30.   return num_features
    31.  
    32. if 'IS_GRADESCOPE_ENV' not in os.environ:
    33.   num_classes = get_num_classes(pyg_dataset)
    34.   num_features = get_num_features(pyg_dataset)
    35.   print("{} dataset has {} classes".format(name, num_classes))
    36.   print("{} dataset has {} features".format(name, num_features))
    1. # Question 2：What is the label of the graph with index 100 in the ENZYMES dataset? 
    2. def get_graph_class(pyg_dataset, idx):
    3.   # TODO: Implement a function that takes a PyG dataset object,
    4.   # an index of a graph within the dataset, and returns the class/label
    5.   # of the graph (as an integer).
    6.  
    7.   label = -1
    8.  
    9.   ############# Your code here ############
    10.   ## (~1 line of code)
    11.  
    12.   #########################################
    13.   label = pyg_dataset[idx].y.item()
    14.   return label
    15.  
    16. # Here pyg_dataset is a dataset for graph classification
    17. if 'IS_GRADESCOPE_ENV' not in os.environ:
    18.   graph_0 = pyg_dataset[0]
    19.   print(graph_0)
    20.   idx = 100
    21.   label = get_graph_class(pyg_dataset, idx)
    22.   print('Graph with index {} has label {}'.format(idx, label))
    1. # Question 3：How many edges does the graph with index 200 have?
    2.  
    3. print(pyg_dataset[200].num_edges)
    4. def get_graph_num_edges(pyg_dataset, idx):
    5.   # TODO: Implement a function that takes a PyG dataset object,
    6.   # the index of a graph in the dataset, and returns the number of
    7.   # edges in the graph (as an integer). You should not count an edge
    8.   # twice if the graph is undirected. For example, in an undirected
    9.   # graph G, if two nodes v and u are connected by an edge, this edge
    10.   # should only be counted once.
    11.  
    12.   num_edges = 0
    13.  
    14.   ############# Your code here ############
    15.   ## Note:
    16.   ## 1. You can't return the data.num_edges directly
    17.   ## 2. We assume the graph is undirected
    18.   ## 3. Look at the PyG dataset built in functions
    19.   ## (~4 lines of code)
    20.  
    21.   #########################################
    22.   num_edges = pyg_dataset[idx].num_edges / 2
    23.   return num_edges
    25. if 'IS_GRADESCOPE_ENV' not in os.environ:
    26.   idx = 200
    27.   num_edges = get_graph_num_edges(pyg_dataset, idx)
    28.   print('Graph with index {} has {} edges'.format(idx, num_edges))

    2、OGB包的使用（Question 4：How many features are in the ogbn-arxiv graph?） 

    1. #Question 4
    2. import torch
    3. import pandas as pd
    4. import os
    5. print("aa1")
    6. import torch_geometric.transforms as T
    7. print("aa2")
    8. from ogb.nodeproppred import PygNodePropPredDataset
    9. print("aa3")
    10.  
    11. if 'IS_GRADESCOPE_ENV' not in os.environ:
    12.   dataset_name = 'ogbn-arxiv'
    13.   print("bb")
    14.   # Load the dataset and transform it to sparse tensor
    15.   dataset = PygNodePropPredDataset(name=dataset_name,
    16.                                   transform=T.ToSparseTensor(), root="./Arxiv")
    17.   print('The {} dataset has {} graph'.format(dataset_name, len(dataset)))
    18.  
    19.   # Extract the graph
    20.   data = dataset[0]
    21.   print(data)
    22.  
    23. #Question 4
    24. def graph_num_features(data):
    25.   # TODO: Implement a function that takes a PyG data object,
    26.   # and returns the number of features in the graph (as an integer).
    27.  
    28.   num_features = 0
    29.  
    30.   ############# Your code here ############
    31.   ## (~1 line of code)
    32.   num_features = data.num_features
    33.   #########################################
    34.   
    35.   return num_features
    36.  
    37. if 'IS_GRADESCOPE_ENV' not in os.environ:
    38.   num_features = graph_num_features(data)
    39.   print('The graph has {} features'.format(num_features))

    3、节点分类任务GNN框架的实现与初步理解

    ① torch.nn.ModuleList()类、torch.nn.LogSoftmax()类的使用

    ② PyG中GCNConv()类、BatchNorm()类的使用

    ③ 此处GNN类的作用/构成理解：

            __init__(...)方法中定义上图中的操作，foward(...)方法负责调用这些操作，将图中的数据X、adj等最终映射为一个N维向量（N：节点数）。

    ④ torch.flatten(input, start_dim=0, end_dim=-1)：去掉start~end的维度，把结果包装成一个“一维Tensor”并返回。

    举例：一个3维矩阵，压缩0~1维，最终返回一个“1维向量”。

    1. x = np.arange(27)
    2. x = np.reshape(x, (3, 3, 3))
    3. x = torch.from_numpy(x)
    4. print('before flatten', x)
    5. x = torch.flatten(x, start_dim=0, end_dim=1)
    6. print('after flatten', x)
    7.  
    8. 输出如下图：

    结果分析：
            首先，去掉一个维度后，变量数目是变多的，例如变量为一个3*3的二维矩阵，去掉第0维后，变成3个有三个元素的向量，1到3，增多。
            回到本例，压缩/去除第0维，x[0]、x[1]和x[2]之间的联系取消，即最外边的中括号去掉，压缩第1维同理。得到的结果为[1,2,3], [4,5,6]...一个个“变量”，最终把这一个个变量封装成一个“一维向量”便是下图结果所示，实际上是2维的。

    ⑤ train函数的理解，def train(model, data, train_idx, optimizer, loss_fn):

    作用：完成一个epoch的训练（即所有数据走一遍）。

    过程：model.train()函数（修改self.training变量的值，使得drop等操作有选择的执行，drop函数定义过程如下所示，需要用到self.training变量）——“优化器”梯度清零——data喂到model里，得到输出结果out——利用loss函数和out以及data标签计算损失——反向传播调整模型参数。

    ⑥ loss_fn函数：

            第二个参数需要为一维向量，不能为二维，内部函数处理时会自动将其上升一个维度，因此要使用torch.flatten(...)函数来降维。

    ⑦ test(...)方法中的out.argmax(......)函数使用：如代码注释

    1. class GCN(torch.nn.Module):
    2.     def __init__(self, input_dim, hidden_dim, output_dim, num_layers,
    3.                  dropout, return_embeds=False):
    4.         # TODO: Implement a function that initializes self.convs,
    5.         # self.bns, and self.softmax.
    6.  
    7.         super(GCN, self).__init__()
    8.  
    9.         # A list of GCNConv layers
    10.         self.convs = None
    11.  
    12.         # A list of 1D batch normalization layers
    13.         self.bns = None
    14.  
    15.         # The log softmax layer
    16.         self.softmax = None
    17.  
    18.         ############# Your code here ############
    19.         ## Note:
    20.         ## 1. You should use torch.nn.ModuleList for self.convs and self.bns
    21.         ## 2. self.convs has num_layers GCNConv layers
    22.         ## 3. self.bns has num_layers - 1 BatchNorm1d layers
    23.         ## 4. You should use torch.nn.LogSoftmax for self.softmax
    24.         ## 5. The parameters you can set for GCNConv include 'in_channels' and
    25.         ## 'out_channels'. For more information please refer to the documentation:
    26.         ## https://pytorch-geometric.readthedocs.io/en/latest/modules/nn.html#torch_geometric.nn.conv.GCNConv
    27.         ## 6. The only parameter you need to set for BatchNorm1d is 'num_features'
    28.         ## For more information please refer to the documentation:
    29.         ## https://pytorch.org/docs/stable/generated/torch.nn.BatchNorm1d.html
    30.         ## (~10 lines of code)
    31.  
    32.         #########################################
    33.         self.convs = torch.nn.ModuleList()
    34.         self.bns = torch.nn.ModuleList()
    35.         self.softmax = torch.nn.LogSoftmax()
    36.         tmp1 = input_dim
    37.         tmp2 = hidden_dim
    38.         for i in range(num_layers - 1):
    39.             self.convs.append(GCNConv(tmp1, tmp2))
    40.             self.bns.append(BatchNorm(tmp2))
    41.             tmp1 = tmp2
    42.         self.convs.append(GCNConv(hidden_dim, output_dim))
    43.  
    44.         # Probability of an element getting zeroed
    45.         self.dropout = dropout
    46.  
    47.         # Skip classification layer and return node embeddings
    48.         self.return_embeds = return_embeds
    49.  
    50.     def reset_parameters(self):
    51.         for conv in self.convs:
    52.             conv.reset_parameters()
    53.         for bn in self.bns:
    54.             bn.reset_parameters()
    55.  
    56.     def forward(self, x, adj_t):
    57.         # TODO: Implement a function that takes the feature tensor x and
    58.         # edge_index tensor adj_t and returns the output tensor as
    59.         # shown in the figure.
    60.  
    61.         out = None
    62.  
    63.         ############# Your code here ############
    64.         ## Note:
    65.         ## 1. Construct the network as shown in the figure
    66.         ## 2. torch.nn.functional.relu and torch.nn.functional.dropout are useful
    67.         ## For more information please refer to the documentation:
    68.         ## https://pytorch.org/docs/stable/nn.functional.html
    69.         ## 3. Don't forget to set F.dropout training to self.training
    70.         ## 4. If return_embeds is True, then skip the last softmax layer
    71.         ## (~7 lines of code)
    72.  
    73.         #########################################
    74.         out = x
    75.         for i in range(len(self.bns)):
    76.             out = self.convs[i](out, adj_t)
    77.             out = self.bns[i](out)
    78.             out = torch.nn.functional.relu(out)
    79.             out = torch.nn.functional.dropout(out, p=self.dropout, training=self.training)
    80.  
    81.         out = self.convs[-1](out, adj_t)
    82.         out = self.softmax(out) if (not self.return_embeds) else out
    83.         return out
    84.  
    85.  
    86. def train(model, data, train_idx, optimizer, loss_fn):
    87.     # TODO: Implement a function that trains the model by
    88.     # using the given optimizer and loss_fn.
    89.     model.train()
    90.     loss = 0
    91.  
    92.     ############# Your code here ############
    93.     ## Note:
    94.     ## 1. Zero grad the optimizer
    95.     ## 2. Feed the data into the model
    96.     ## 3. Slice the model output and label by train_idx
    97.     ## 4. Feed the sliced output and label to loss_fn
    98.     ## (~4 lines of code)
    99.  
    100.     optimizer.zero_grad()
    101.     out = model(data.x, data.adj_t)
    102.  
    103.     loss = loss_fn(out[train_idx], torch.flatten(data.y[train_idx]))
    104.     #########################################
    105.  
    106.     loss.backward()
    107.     optimizer.step()
    108.  
    109.     return loss.item()
    110.  
    111.  
    112. # Test function here
    113. @torch.no_grad()
    114. def test(model, data, split_idx, evaluator, save_model_results=False):
    115.     # TODO: Implement a function that tests the model by
    116.     # using the given split_idx and evaluator.
    117.     model.eval()
    118.  
    119.     # The output of model on all data
    120.     out = None
    121.  
    122.     ############# Your code here ############
    123.     ## (~1 line of code)
    124.     ## Note:
    125.     ## 1. No index slicing here
    126.  
    127.     #########################################
    128.     out = model(data.x, data.adj_t)
    129.  
    130.     # wjunjie
    131.     # argmax在指定维度dim上操作，选择最大值的dim维度上的索引，keepdim=True相当于保留原始维度
    132.     # out是一个n*m的矩阵，m等于类别数目.
    133.     # 这里最后要得到一个n*1个pred矩阵，所以在dim=1上操作argmax(a[0][0],...,a[0][m-1])
    134.     y_pred = out.argmax(dim=-1, keepdim=True)
    135.  
    136.     train_acc = evaluator.eval({
    137.         'y_true': data.y[split_idx['train']],
    138.         'y_pred': y_pred[split_idx['train']],
    139.     })['acc']
    140.     valid_acc = evaluator.eval({
    141.         'y_true': data.y[split_idx['valid']],
    142.         'y_pred': y_pred[split_idx['valid']],
    143.     })['acc']
    144.     test_acc = evaluator.eval({
    145.         'y_true': data.y[split_idx['test']],
    146.         'y_pred': y_pred[split_idx['test']],
    147.     })['acc']
    148.  
    149.     if save_model_results:
    150.         print("Saving Model Predictions")
    151.  
    152.         data = {}
    153.         # 二维y_pred转为一维、cuda数据转为cpu、去掉tensor的梯度、转为numpy
    154.         data['y_pred'] = y_pred.view(-1).cpu().detach().numpy()
    155.  
    156.         df = pd.DataFrame(data=data)
    157.         # Save locally as csv
    158.         df.to_csv('ogbn-arxiv_node.csv', sep=',', index=False)
    159.  
    160.     return train_acc, valid_acc, test_acc
    161.  
    162.  
    163. # Please do not change the args
    164. if 'IS_GRADESCOPE_ENV' not in os.environ:
    165.     args = {
    166.         'device': device,
    167.         'num_layers': 3,
    168.         'hidden_dim': 256,
    169.         'dropout': 0.5,
    170.         'lr': 0.01,
    171.         'epochs': 100,
    172.     }
    173.  
    174. if 'IS_GRADESCOPE_ENV' not in os.environ:
    175.     model = GCN(data.num_features, args['hidden_dim'],
    176.                 dataset.num_classes, args['num_layers'],
    177.                 args['dropout']).to(device)
    178.     evaluator = Evaluator(name='ogbn-arxiv')
    179.  
    180.  
    181. import copy
    182. if 'IS_GRADESCOPE_ENV' not in os.environ:
    183.   # reset the parameters to initial random value
    184.   model.reset_parameters()
    185.  
    186.   optimizer = torch.optim.Adam(model.parameters(), lr=args['lr'])
    187.   loss_fn = F.nll_loss
    188.  
    189.   best_model = None
    190.   best_valid_acc = 0
    191.  
    192.   for epoch in range(1, 1 + args["epochs"]):
    193.     loss = train(model, data, train_idx, optimizer, loss_fn)
    194.     result = test(model, data, split_idx, evaluator)
    195.     train_acc, valid_acc, test_acc = result
    196.     if valid_acc > best_valid_acc:
    197.         best_valid_acc = valid_acc
    198.         best_model = copy.deepcopy(model)
    199.     print(f'Epoch: {epoch:02d}, '
    200.           f'Loss: {loss:.4f}, '
    201.           f'Train: {100 * train_acc:.2f}%, '
    202.           f'Valid: {100 * valid_acc:.2f}% '
    203.           f'Test: {100 * test_acc:.2f}%')
    204.  
    205. if 'IS_GRADESCOPE_ENV' not in os.environ:
    206.   best_result = test(best_model, data, split_idx, evaluator, save_model_results=True)
    207.   train_acc, valid_acc, test_acc = best_result
    208.   print(f'Best model: '
    209.         f'Train: {100 * train_acc:.2f}%, '
    210.         f'Valid: {100 * valid_acc:.2f}% '
    211.         f'Test: {100 * test_acc:.2f}%')

    4、图级别预测

    ① 框架构建思路：

    一个图级别卷积过程包括：先利用节点GNN得到节点的表征——然后池化——最后进行一些适当的线性or非线性变化。

    因此init函数和foward函数的coding过程就按照上述思路来。

    ② DataLoader与batch：

    教程将obg的datset先转换为了DataLoader类型，并设定了batch。难道DataLoader是使用batch的一种方式？

    ③ train函数中的tips：

    A、Batch：一个batch相当于一个小的dataset，里面包含了“32”（batch size）个图的信息。

    B、python list解析式：其他语法糖还有dict解析式等...

     C、tensor的索引可以是list和LongTensor，还可以bool数组（如train函数中注释所示）。

    1. ### GCN to predict graph property
    2. class GCN_Graph(torch.nn.Module):
    3.     def __init__(self, hidden_dim, output_dim, num_layers, dropout):
    4.         super(GCN_Graph, self).__init__()
    5.  
    6.         # Load encoders for Atoms in molecule graphs
    7.         self.node_encoder = AtomEncoder(hidden_dim)
    8.  
    9.         # Node embedding model
    10.         # Note that the input_dim and output_dim are set to hidden_dim
    11.         self.gnn_node = GCN(hidden_dim, hidden_dim,
    12.                             hidden_dim, num_layers, dropout, return_embeds=True)
    13.  
    14.         self.pool = None
    15.  
    16.         ############# Your code here ############
    17.  
    18.         self.pool = global_mean_pool
    19.         ## Note:
    20.         ## 1. Initialize self.pool as a global mean pooling layer
    21.         ## For more information please refer to the documentation:
    22.         ## https://pytorch-geometric.readthedocs.io/en/latest/modules/nn.html#global-pooling-layers
    23.  
    24.         #########################################
    25.  
    26.         # Output layer
    27.         self.linear = torch.nn.Linear(hidden_dim, output_dim)
    28.  
    29.     def reset_parameters(self):
    30.         self.gnn_node.reset_parameters()
    31.         self.linear.reset_parameters()
    32.  
    33.     def forward(self, batched_data):
    34.         # TODO: Implement a function that takes as input a
    35.         # mini-batch of graphs (torch_geometric.data.Batch) and
    36.         # returns the predicted graph property for each graph.
    37.         #
    38.         # NOTE: Since we are predicting graph level properties,
    39.         # your output will be a tensor with dimension equaling
    40.         # the number of graphs in the mini-batch
    41.  
    42.         # Extract important attributes of our mini-batch
    43.         x, edge_index, batch = batched_data.x, batched_data.edge_index, batched_data.batch
    44.         embed = self.node_encoder(x)
    45.  
    46.         out = None
    47.  
    48.         ############# Your code here ############
    49.         ## Note:
    50.         ## 1. Construct node embeddings using existing GCN model
    51.         ## 2. Use the global pooling layer to aggregate features for each individual graph
    52.         ## For more information please refer to the documentation:
    53.         ## https://pytorch-geometric.readthedocs.io/en/latest/modules/nn.html#global-pooling-layers
    54.         ## 3. Use a linear layer to predict each graph's property
    55.         ## (~3 lines of code)
    56.         out = self.gnn_node(embed, edge_index)
    57.         out = self.pool(out, batch)
    58.         out = self.linear(out)
    59.         #########################################
    60.  
    61.         return out
    62.  
    63.  
    64. def train(model, device, data_loader, optimizer, loss_fn):
    65.  
    66.     # TODO: Implement a function that trains your model by
    67.     # using the given optimizer and loss_fn.
    68.     model.train()
    69.     loss = 0
    70.  
    71.     for step, batch in enumerate(tqdm(data_loader, desc="Iteration")):
    72.         batch = batch.to(device)
    73.         # 一个batch相当于一个小的dataset，里面包含了“32”（batch size）个图的信息
    74.         if batch.x.shape[0] == 1 or batch.batch[-1] == 0: # 这个batch是孤立节点或者该batch里只有一个图
    75.             pass
    76.         else:
    77.             ## ignore nan targets (unlabeled) when computing training loss.
    78.             is_labeled = batch.y == batch.y # 不出意外全是true 【32，1】
    79.  
    80.             ############# Your code here ############
    81.             ## Note:
    82.             ## 1. Zero grad the optimizer
    83.             ## 2. Feed the data into the model
    84.             ## 3. Use `is_labeled` mask to filter output and labels
    85.             ## 4. You may need to change the type of label to torch.float32
    86.             ## 5. Feed the output and label to the loss_fn
    87.             ## (~3 lines of code)
    88.             optimizer.zero_grad()
    89.             out = model(batch)
    90.  
    91.             # python list解析式，其他语法糖还有dict解析式等...
    92.             tmp_index = [index for index in range(is_labeled.shape[0]) if is_labeled[index]]
    93.             loss = loss_fn(out[tmp_index], batch.y[tmp_index].type(torch.float32))
    94.  
    95.             # 看了其他人的，这里index直接用is_labeled代替就可以，所以tensor的索引可以是list和LongTensor，还可以bool数组。
    96.  
    97.             #########################################
    98.  
    99.             loss.backward()
    100.             optimizer.step()
    101.  
    102.     return loss.item()
    103.  
    104.  
    105. # The evaluation function
    106. def eval(model, device, loader, evaluator, save_model_results=False, save_file=None):
    107.     model.eval()
    108.     y_true = []
    109.     y_pred = []
    110.  
    111.     for step, batch in enumerate(tqdm(loader, desc="Iteration")):
    112.         batch = batch.to(device)
    113.  
    114.         if batch.x.shape[0] == 1:
    115.             pass
    116.         else:
    117.             with torch.no_grad():
    118.                 pred = model(batch)
    119.  
    120.             y_true.append(batch.y.view(pred.shape).detach().cpu())
    121.             y_pred.append(pred.detach().cpu())
    122.  
    123.     y_true = torch.cat(y_true, dim=0).numpy()
    124.     y_pred = torch.cat(y_pred, dim=0).numpy()
    125.  
    126.     input_dict = {"y_true": y_true, "y_pred": y_pred}
    127.  
    128.     if save_model_results:
    129.         print("Saving Model Predictions")
    130.  
    131.         # Create a pandas dataframe with a two columns
    132.         # y_pred | y_true
    133.         data = {}
    134.         data['y_pred'] = y_pred.reshape(-1)
    135.         data['y_true'] = y_true.reshape(-1)
    136.  
    137.         df = pd.DataFrame(data=data)
    138.         # Save to csv
    139.         df.to_csv('ogbg-molhiv_graph_' + save_file + '.csv', sep=',', index=False)
    140.  
    141.     return evaluator.eval(input_dict)
    142.  
    143. if 'IS_GRADESCOPE_ENV' not in os.environ:
    144.   model = GCN_Graph(args['hidden_dim'],
    145.               dataset.num_tasks, args['num_layers'],
    146.               args['dropout']).to(device)
    147.   evaluator = Evaluator(name='ogbg-molhiv')
    148.  
    149. import copy
    150.  
    151. if 'IS_GRADESCOPE_ENV' not in os.environ:
    152.   model.reset_parameters()
    153.  
    154.   optimizer = torch.optim.Adam(model.parameters(), lr=args['lr'])
    155.   loss_fn = torch.nn.BCEWithLogitsLoss()
    156.  
    157.   best_model = None
    158.   best_valid_acc = 0
    159.  
    160.   for epoch in range(1, 1 + args["epochs"]):
    161.     print('Training...')
    162.     loss = train(model, device, train_loader, optimizer, loss_fn)
    163.  
    164.     print('Evaluating...')
    165.     train_result = eval(model, device, train_loader, evaluator)
    166.     val_result = eval(model, device, valid_loader, evaluator)
    167.     test_result = eval(model, device, test_loader, evaluator)
    168.  
    169.     train_acc, valid_acc, test_acc = train_result[dataset.eval_metric], val_result[dataset.eval_metric], test_result[dataset.eval_metric]
    170.     if valid_acc > best_valid_acc:
    171.         best_valid_acc = valid_acc
    172.         best_model = copy.deepcopy(model)
    173.     print(f'Epoch: {epoch:02d}, '
    174.           f'Loss: {loss:.4f}, '
    175.           f'Train: {100 * train_acc:.2f}%, '
    176.           f'Valid: {100 * valid_acc:.2f}% '
    177.           f'Test: {100 * test_acc:.2f}%')

    End...

  • 相关阅读:
    EN 13707防水用柔性薄板-屋顶防水用增强沥青薄板—CE认证
    第五篇 python 基本语法(一)
    SE(Squeeze and Excitation)模块的理解以及代码实现
    [apue] 标准 I/O 库那些事儿
    Mybatis-Plus开发提速器mybatis-plus-generator-ui
    Python第一阶段-第十章-Python基础综合案例-数据可视化
    四、ref与DOM-findDomNode-unmountComponentAtNode
    2022前端面试题整理
    Android Studio导入aosp源码
    springboot 四大组件
  • 原文地址：https://blog.csdn.net/qq_41661919/article/details/126426970