TranslatedLayer

class paddle.jit. TranslatedLayer ( programs, persistable_vars ) [源代码]

TranslatedLayer 是一个命令式编程模式 Layer 的继承类， 通过 load 载入构建。能够像一般 Layer 一样在train或者eval模式下使用。

注解

TranslatedLayer 对象不能够通过构造函数创建，仅能够通过 load 接口载入构建。

示例代码：

  1. import numpy as np
  2. import paddle
  3. import paddle.nn as nn
  4. import paddle.optimizer as opt
  6. BATCH_SIZE = 16
  7. BATCH_NUM = 4
  8. EPOCH_NUM = 4
  10. IMAGE_SIZE = 784
  11. CLASS_NUM = 10
  13. # define a random dataset
  14. class RandomDataset(paddle.io.Dataset):
  15.     def __init__(self, num_samples):
  16.         self.num_samples = num_samples
  18.     def __getitem__(self, idx):
  19.         image = np.random.random([IMAGE_SIZE]).astype('float32')
  20.         label = np.random.randint(0, CLASS_NUM - 1, (1, )).astype('int64')
  21.         return image, label
  23.     def __len__(self):
  24.         return self.num_samples
  26. class LinearNet(nn.Layer):
  27.     def __init__(self):
  28.         super(LinearNet, self).__init__()
  29.         self._linear = nn.Linear(IMAGE_SIZE, CLASS_NUM)
  31.     @paddle.jit.to_static
  32.     def forward(self, x):
  33.         return self._linear(x)
  35. def train(layer, loader, loss_fn, opt):
  36.     for epoch_id in range(EPOCH_NUM):
  37.         for batch_id, (image, label) in enumerate(loader()):
  38.             out = layer(image)
  39.             loss = loss_fn(out, label)
  40.             loss.backward()
  41.             opt.step()
  42.             opt.clear_grad()
  43.             print("Epoch {} batch {}: loss = {}".format(
  44.                 epoch_id, batch_id, np.mean(loss.numpy())))
  46. # 1. train & save model.
  48. # create network
  49. layer = LinearNet()
  50. loss_fn = nn.CrossEntropyLoss()
  51. adam = opt.Adam(learning_rate=0.001, parameters=layer.parameters())
  53. # create data loader
  54. dataset = RandomDataset(BATCH_NUM * BATCH_SIZE)
  55. loader = paddle.io.DataLoader(dataset,
  56.     batch_size=BATCH_SIZE,
  57.     shuffle=True,
  58.     drop_last=True,
  59.     num_workers=2)
  61. # train
  62. train(layer, loader, loss_fn, adam)
  64. # save
  65. model_path = "linear.example.model"
  66. paddle.jit.save(layer, model_path)
  68. # 2. load model as TranslatedLayer
  70. # load
  71. translated_layer = paddle.jit.load(model_path)
  73. # inference
  74. translated_layer.eval()
  75. x = paddle.randn([1, IMAGE_SIZE], 'float32')
  76. pred = translated_layer(x)
  78. # fine-tune
  79. translated_layer.train()
  80. adam = opt.Adam(learning_rate=0.001, parameters=translated_layer.parameters())
  81. train(translated_layer, loader, loss_fn, adam)

program(method_name=’forward’):

获取TranslatedLayer中指定方法对应的Program。

参数：

  • method_name (string) - 要获取的Porgram对应的方法名。默认值为”forward”。

返回：Program

返回类型：Program

示例代码：

  1. import numpy as np
  2. import paddle
  3. import paddle.nn as nn
  4. import paddle.optimizer as opt
  6. BATCH_SIZE = 16
  7. BATCH_NUM = 4
  8. EPOCH_NUM = 4
  10. IMAGE_SIZE = 784
  11. CLASS_NUM = 10
  13. # define a random dataset
  14. class RandomDataset(paddle.io.Dataset):
  15.     def __init__(self, num_samples):
  16.         self.num_samples = num_samples
  18.     def __getitem__(self, idx):
  19.         image = np.random.random([IMAGE_SIZE]).astype('float32')
  20.         label = np.random.randint(0, CLASS_NUM - 1, (1, )).astype('int64')
  21.         return image, label
  23.     def __len__(self):
  24.         return self.num_samples
  26. class LinearNet(nn.Layer):
  27.     def __init__(self):
  28.         super(LinearNet, self).__init__()
  29.         self._linear = nn.Linear(IMAGE_SIZE, CLASS_NUM)
  31.     @paddle.jit.to_static
  32.     def forward(self, x):
  33.         return self._linear(x)
  35. def train(layer, loader, loss_fn, opt):
  36.     for epoch_id in range(EPOCH_NUM):
  37.         for batch_id, (image, label) in enumerate(loader()):
  38.             out = layer(image)
  39.             loss = loss_fn(out, label)
  40.             loss.backward()
  41.             opt.step()
  42.             opt.clear_grad()
  43.             print("Epoch {} batch {}: loss = {}".format(
  44.                 epoch_id, batch_id, np.mean(loss.numpy())))
  46. # create network
  47. layer = LinearNet()
  48. loss_fn = nn.CrossEntropyLoss()
  49. adam = opt.Adam(learning_rate=0.001, parameters=layer.parameters())
  51. # create data loader
  52. dataset = RandomDataset(BATCH_NUM * BATCH_SIZE)
  53. loader = paddle.io.DataLoader(dataset,
  54.     batch_size=BATCH_SIZE,
  55.     shuffle=True,
  56.     drop_last=True,
  57.     num_workers=2)
  59. # train
  60. train(layer, loader, loss_fn, adam)
  62. # save
  63. model_path = "linear.example.model"
  64. paddle.jit.save(layer, model_path)
  66. # load
  67. translated_layer = paddle.jit.load(model_path)
  69. # get program
  70. program = translated_layer.program()
  71. print(program)