使用 VisualDL 可视化模型，数据和训练¶

# 数据载入
class MNISTDataset():
  def __init__(self, mode='train'):
    self.mnist_data = paddle.vision.datasets.MNIST(mode=mode)

  def __getitem__(self, idx):
    data, label = self.mnist_data[idx]
    data = np.reshape(data, [1, 28, 28]).astype('float32') / 255
    label = np.reshape(label, [1]).astype('int64')
    return (data, label)

  def __len__(self):
    return len(self.mnist_data)

# 查看 9 张输入的训练图像的样例
dataset = MNISTDataset(mode='train')
image_matrix = []
for i in range(9):
  image, label = dataset[i]
  # 将 dataset 中的 CHW 排列的图像转换成 HWC 排列再写入 VisualDL
  image_matrix.append(image.transpose([1,2,0]))
# 将九张输入图像合成长宽相同的图像网格，即 3X3 的图像网格
logwriter.add_image_matrix(tag='input_images', step=1, imgs=image_matrix, rows=-1)

# 将九张输入图像以向量的形式写入 embeddings，查看数据降维后的关系
tags = ['image_{}'.format(i) for i in range(9)]
logwriter.add_embeddings('input_image_embeddings', mat=[img.reshape(-1) for img in image_matrix], metadata=tags)

#创建模型
model = MNIST()
#保存模型，获取模型结构文件
paddle.jit.save(model, './runs/mnist_experiment/model', [paddle.static.InputSpec([-1,1,28,28])])

logwriter.add_hparams(hparams_dict={'lr': 0.001, 'batch_size': 16, 'opt': 'sgd'},
                           metrics_list=['train_avg_loss', 'test_avg_loss', 'test_avg_acc'])

logwriter.add_hparams(hparams_dict={'lr': 0.0001, 'batch_size': 16, 'opt': 'adam'},
                           metrics_list=['train_avg_loss', 'test_avg_loss', 'test_avg_acc'])

  for epoch_id in range(EPOCH_NUM):
      model.train()
      train_batchs_per_epoch = len(train_loader)
      for batch_id, data in enumerate(train_loader):
          #准备数据
          images, labels = data

          #前向计算的过程
          predicts = model(images)

          #计算损失，取一个批次样本损失的平均值
          loss = F.cross_entropy(predicts, labels)
          avg_loss = paddle.mean(loss)

          #记录当前训练 Loss 到 VisualDL
          logwriter.add_scalar("train_avg_loss", value=avg_loss.numpy(), step=batch_id+epoch_id*(train_batchs_per_epoch))

          #记录网络中最后一个 fc 层的参数到 VisualDL
          logwriter.add_histogram("fc_weight", values=model.fc.weight.numpy(), step=batch_id+epoch_id*(train_batchs_per_epoch))

          #每训练了 100 批次的数据，打印下当前 Loss 的情况
          if batch_id % 200 == 0:
              print("epoch: {}, batch: {}, loss is: {}".format(epoch_id, batch_id, avg_loss.numpy()))

          #后向传播，更新参数的过程
          avg_loss.backward()
          # 最小化 loss,更新参数
          opt.step()
          # 清除梯度
          opt.clear_grad()

for epoch_id in range(EPOCH_NUM):
      model.train()
      train_batchs_per_epoch = len(train_loader)
      for batch_id, data in enumerate(train_loader):
          #准备数据
          images, labels = data

          #前向计算的过程
          predicts = model(images)

          #计算损失，取一个批次样本损失的平均值
          loss = F.cross_entropy(predicts, labels)
          avg_loss = paddle.mean(loss)

          #记录当前训练 Loss 到 VisualDL
          logwriter.add_scalar("train_avg_loss", value=avg_loss.numpy(), step=batch_id+epoch_id*(train_batchs_per_epoch))

          #记录网络中最后一个 fc 层的参数到 VisualDL
          logwriter.add_histogram("fc_weight", values=model.fc.weight.numpy(), step=batch_id+epoch_id*(train_batchs_per_epoch))

          #每训练了 100 批次的数据，打印下当前 Loss 的情况
          if batch_id % 200 == 0:
              print("epoch: {}, batch: {}, loss is: {}".format(epoch_id, batch_id, avg_loss.numpy()))

          #后向传播，更新参数的过程
          avg_loss.backward()
          # 最小化 loss,更新参数
          opt.step()
          # 清除梯度
          opt.clear_grad()

for batch_id, data in enumerate(test_loader):
      #准备数据
      images, labels = data
      #前向计算的过程
      predicts = model(images)
      #计算损失
      loss = F.cross_entropy(predicts, labels)
      #计算准确率
      acc = paddle.metric.accuracy(predicts, labels)
      accuracies.append(acc.numpy())
      losses.append(loss.numpy())

avg_acc, avg_loss = np.mean(accuracies), np.mean(losses)
print("[validation]After epoch {}: accuracy/loss: {}/{}".format(epoch_id, avg_acc, avg_loss))
#记录当前测试集平均 Loss 和准确率到 VisualDL
logwriter.add_scalar("test_avg_loss", value=avg_acc, step=epoch_id)
logwriter.add_scalar("test_avg_acc", value=avg_loss, step=epoch_id)

# evaluate model after one epoch
    model.eval()
    accuracies = []
    losses = []
    class_probs = []
    class_preds = []
    for batch_id, data in enumerate(test_loader):
        #准备数据
        images, labels = data
        #前向计算的过程
        predicts = model(images)
        #计算损失
        loss = F.cross_entropy(predicts, labels)
        #计算准确率
        acc = paddle.metric.accuracy(predicts, labels)
        accuracies.append(acc.numpy())
        losses.append(loss.numpy())
        #记录用于画 pr 曲线需要的预测概率和类别
        class_probs_batch = [F.softmax(predict, axis=0) for predict in predicts]
        class_preds_batch = paddle.argmax(predicts, 1)

        class_probs.append(class_probs_batch)
        class_preds.append(class_preds_batch)

    test_probs = paddle.concat([paddle.stack(batch) for batch in class_probs]).numpy()
    test_preds = paddle.concat(class_preds).numpy()

    for i in range(10):
      logwriter.add_pr_curve('class_{}'.format(i), labels=(test_preds == i),predictions=test_probs[:,i], step=epoch_id)