0，有時間看看源碼還是看看源碼吧，雖然看了也還是菜雞。。。
https://github.com/PaddlePaddle
1，常用方法總結
'''========================================1，資源配置========================================'''
paddle.fluid.is_compiled_with_cuda()
paddle.set_device('gpu:0'/'cpu')
paddle.set_device('gpu:0') if use_gpu else paddle.set_device('cpu')
'''========================================2，tensor========================================'''
#************************************轉為tensor************************************
paddle.to_tensor(**, stop_gradient=False)
#************************************轉變tensor的dtype************************************
paddle.cast(tensor, 'float32')
tensor.astype('float32')
#************************************維度擴展************************************
tensor.unsqueeze([0, 1])
#************************************取隨機************************************
paddle.randn([2,10])
'''========================================3，數據加載========================================'''
paddle.io.Dataset/DataLoader
class myDataset(Dataset):
def __init__(self, files):
...
def __len__(self):
return len(...)/...shape[0]
def __getitem__(self,index):
...
return data, label
traingData = myDataset(files)
trainingDataloader = Dataloader(traingData, batch_size=, shuffle=True)
#************************************獲取數據和標簽************************************
data, label = next(iter(trainingDataloader))
for (data, label) in trainingDataloader:
for batch, (data, label) in enumerate(trainingDataloader):
'''========================================4，基本模型庫========================================'''
#************************************模型構造方法1 nn基礎模型************************************
myNet = paddle.nn.Linear(10, 10)
myNet.weight
myNet.weight.grad
myNet.bias
myNet.bias.grad
myNet.parameters()
#************************************模型構造方法2.1 nn.Sequential************************************
myNet = paddle.nn.Sequential(
nn.Linear(10, 10),
nn.Tanh(),
nn.Linear(10, 10)
)
[param.shape for param in myNet.parameters()]
[(name, param.shape) for (name, param) in myNet.named_parameters()]
#************************************模型構造方法2.2 collections.OrderedDict 與 nn.Sequential 結合，為子模塊命名************************************
myNet = nn.Sequential(('hidden_linear',nn.Linear(10,10)),
('hidden_activation', nn.Tanh()),
('output_linear', nn.Linear(10,10)))
[param.shape for param in myNet.parameters()]
[(name, param.shape) for (name, param) in myNet.named_parameters()]
myNet.hidden_linear.weight
myNet.hidden_linear.weight.grad
myNet.hidden_linear.bias
myNet.hidden_linear.bias.grad
#************************************動態添加子模塊************************************
nn.Sequential().add_sublayer()
#************************************模型構造方法3 nn.Module************************************
paddle.nn.Layer
class myModule(nn.Layer):
def __init__(self):
super().__init__()
def forward(self, inputs):
...
myNet = myModule()
[(name, param.shape) for (name, param) in myNet.named_parameters()]
'''========================================5，優化器========================================'''
#************************************四大類優化器************************************
opt = paddle.optimizer.SGD(learning_rate=lr, parameters=myNet.parameters())
paddle.optimizer.Momentum(learning_rate=0.01, momentum=0.9, parameters=model.parameters())
paddle.optimizer.Adagrad(learning_rate=0.01, parameters=model.parameters())
paddle.optimizer.Adam(learning_rate=0.01, parameters=model.parameters())
#************************************5.1：針對模型不同層設置不同的學習率************************************
#************************************5.2：自定義根據 epoch 改變學習率************************************
#************************************5.3：手動設置學習率衰減區間************************************
#************************************5.4：變學習率API，與pytorch相比不用step************************************
lr = paddle.optimizer.lr.PolynomialDecay(learning_rate=0.01, decay_steps=total_steps, end_lr=0.001)
opt = paddle.optimizer.Momentum(learning_rate=lr, parameters=model.parameters())
'''========================================6，損失函數========================================'''
paddle.nn.CrossEntropyLoss()
paddle.nn.functional.cross_entropy()
'''========================================7，訓練========================================'''
myNet.train()
loss.backward()
opt.step()
opt.clear_grad()
'''========================================8，保存========================================'''
#************************************方法一************************************
paddle.save(opt.state_dict(), '')
paddle.save(model.state_dict(), '')
#************************************方法二************************************
paddle.save(optimizer, '') #報錯，不支持
paddle.save(myNet, '') #報錯，不支持
'''========================================9.加載========================================'''
#保存方法一的對應加載
myNet = myModule()
model_dict = paddle.load()
myNet.set_state_dict(model_dict)
optimizer同理
#保存方法二的對應加載，不支持
myNet = paddle.load()
optimizer = paddle.load()
'''========================================10，測試========================================'''
myNet.eval()
'''========================================11，計算准確率========================================'''
paddle.metric.Accuracy()
'''========================================12，視覺庫========================================'''
paddle.vision
paddle.vision.models
paddle.vision.transforms
'''========================================其他========================================'''
model.parameters()
model.named_parameters()
model.state_dict()
optimizer.state_dict()
2，paddle API文檔
3，paddle 組件
