调用resnet50模型

import paddle.vision as vision import paddle.nn as nn paddle.vision.set_image_backend('cv2') model = vision.models.resnet50(pretrained=True, num_classes=12)

定义数据迭代器

train_dataloader = DataLoader(train_data, batch_size=256, shuffle=True, drop_last=False) valid_dataloader = DataLoader(valid_data, batch_size=256, shuffle=False, drop_last=False)

定义优化器

optimizer = paddle.optimizer.Adam(learning_rate=1e-4, parameters=model.parameters(), weight_decay=paddle.regularizer.L2Decay(1e-4))

定义损失函数

criterion = nn.CrossEntropyLoss()

设置gpu环境

paddle.set_device('gpu')

整体训练流程

for epoch_id in range(15): model.train() for batch_id, data in enumerate(train_dataloader()): # 读取数据 features, labels = data features = paddle.to_tensor(features) labels = paddle.to_tensor(labels)

    # 前向传播
    predicts = model(features)

    # 损失计算
    loss = criterion(predicts, labels)

    # 反向传播
    loss.backward()

    # 更新
    optimizer.step()

    # 清零梯度
    optimizer.clear_grad()

    # 打印训练损失
    if batch_id % 10 == 0:
        print('epoch_id:{}, batch_id:{}, loss:{}'.format(epoch_id, batch_id, loss.numpy()[0]))

# 每个epoch评估一次模型性能
model.eval()
acc = 0
total = 0
with paddle.no_grad():
    for data in valid_dataloader():
        images, labels = data
        logits = model(images)
        pred = paddle.argmax(logits, axis=1)
        acc += paddle.sum(pred == labels)
        total += len(labels)
accuracy = acc / total
print('epoch_id:{}, accuracy:{}'.format(epoch_id, accuracy))

# 保存最佳模型
if accuracy > best_acc:
    best_acc = accuracy
    paddle.save(model.state_dict(), 'best_model.pdparams'