文档

• 安装
  • 从Conda安装
  • 从源码安装
    • 由conda创建SINGA
    • 使用本地工具在Ubuntu上创建SINGA
    • 在Windows上编译SINGA
    • 更多编译选择
      • USE_MODULES
      • USE_PYTHON
      • USE_CUDA
      • USE_OPENCL
      • PACKAGE
  • FAQ
• 软件架构
  • 核心
  • 模型
  • 输入输出
• 设备(Device)
  • 特定设备
  • Python API
    • singa.device.create_cuda_gpus(num)
    • singa.device.create_cuda_gpus_on(device_ids)
    • singa.device.get_default_device()
• 张量(Tensor)
  • PYTHON API
    • class singa.tensor.Tensor(shape=None, device=None, dtype=0)
      • T()
      • add_column(v)
      • add_row(v)
      • bernoulli(p)
      • clone()
      • copy()
      • copy_data(t)
      • copy_from_numpy(np_array, offset=0)
      • deepcopy()
      • div_column(v)
      • div_row(v)
      • gaussian(mean, std)
      • is_empty()
      • is_transpose()
      • l1()
      • l2()
      • memsize()
      • mult_column(v)
      • mult_row(v)
      • ndim()
      • reset_like(t)
      • set_value(x)
      • size()
      • to_device(device)
      • to_host()
      • uniform(low, high)
      • singa.tensor.abs(t)
      • singa.tensor.add(lhs, rhs, ret=None)
      • singa.tensor.add_column(alpha, v, beta, M)
      • singa.tensor.add_row(alpha, v, beta, M)
      • singa.tensor.average(t, axis=None)
      • singa.tensor.axpy(alpha, x, y)
      • singa.tensor.bernoulli(p, t)
      • singa.tensor.copy_data_to_from(dst, src, size, dst_offset=0, src_offset=0)
      • singa.tensor.div(lhs, rhs, ret=None)
      • singa.tensor.eltwise_mult(lhs, rhs, ret=None)
      • singa.tensor.exp(t)
      • singa.tensor.from_numpy(np_array)
      • singa.tensor.gaussian(mean, std, t)
      • singa.tensor.ge(t, x)
      • singa.tensor.gt(t, x)
      • singa.tensor.le(t, x)
      • singa.tensor.lt(t, x)
      • singa.tensor.log(t)
      • singa.tensor.mult(A, B, C=None, alpha=1.0, beta=0.0)
      • singa.tensor.pow(t, x, out=None)
      • singa.tensor.relu(t)
      • singa.tensor.reshape(t, s)
      • singa.tensor.sigmoid(t)
      • singa.tensor.sign(t)
      • singa.tensor.sizeof(dtype)
      • singa.tensor.softmax(t, out=None)
      • singa.tensor.sqrt(t)
      • singa.tensor.square(t)
      • singa.tensor.sub(lhs, rhs, ret=None)
      • singa.tensor.sum(t, axis=None)
      • singa.tensor.sum_columns(M)
      • singa.tensor.sum_rows(M)
      • singa.tensor.tanh(t)
      • singa.tensor.to_host(t)
      • singa.tensor.to_numpy(t)
      • singa.tensor.uniform(low, high, t)
• 层(Layer)
  • Python API
    • singa.layer.engine = ‘cudnn’
    • class singa.layer.Layer(name, conf=None, **kwargs)
      • setup(in_shapes)
      • caffe_layer()
      • get_output_sample_shape()
      • param_names()
      • param_values()
      • forward(flag, x)
      • backward(flag, dy)
      • to_device(device)
    • class singa.layer.Dummy(name, input_sample_shape=None)
      • forward(flag, x)
      • backward(falg, dy)
    • class singa.layer.Conv2D(name, nb_kernels, kernel=3, stride=1, border_mode=’same’, cudnn_prefer=’fatest’, data_format=’NCHW’, use_bias=True, W_specs=None, b_specs=None, pad=None, input_sample_shape=None)
    • class singa.layer.Conv1D(name, nb_kernels, kernel=3, stride=1, border_mode=’same’, cudnn_prefer=’fatest’, use_bias=True, W_specs={‘init’: ‘Xavier’}, b_specs={‘init’: ‘Constant’, ‘value’: 0}, pad=None, input_sample_shape=None)
      • get_output_sample_shape()
    • class singa.layer.Pooling2D(name, mode, kernel=3, stride=2, border_mode=’same’, pad=None, data_format=’NCHW’, input_sample_shape=None)
    • class singa.layer.MaxPooling2D(name, kernel=3, stride=2, border_mode=’same’, pad=None, data_format=’NCHW’, input_sample_shape=None)
    • class singa.layer.AvgPooling2D(name, kernel=3, stride=2, border_mode=’same’, pad=None, data_format=’NCHW’, input_sample_shape=None)
    • class singa.layer.MaxPooling1D(name, kernel=3, stride=2, border_mode=’same’, pad=None, data_format=’NCHW’, input_sample_shape=None)
    • class singa.layer.AvgPooling1D(name, kernel=3, stride=2, border_mode=’same’, pad=None, data_format=’NCHW’, input_sample_shape=None)
    • class singa.layer.BatchNormalization(name, momentum=0.9, beta_specs=None, gamma_specs=None, input_sample_shape=None)
    • class singa.layer.LRN(name, size=5, alpha=1, beta=0.75, mode=’cross_channel’, k=1, input_sample_shape=None)
    • class singa.layer.Dense(name, num_output, use_bias=True, W_specs=None, b_specs=None, W_transpose=False, input_sample_shape=None)
    • class singa.layer.Dropout(name, p=0.5, input_sample_shape=None)
    • class singa.layer.Activation(name, mode=’relu’, input_sample_shape=None)
    • class singa.layer.Softmax(name, axis=1, input_sample_shape=None)
    • class singa.layer.Flatten(name, axis=1, input_sample_shape=None)
    • class singa.layer.Merge(name, input_sample_shape=None)
      • setup(in_shape)
      • get_output_sample_shape()
      • forward(flag, inputs)
      • backward(flag, grad)
    • class singa.layer.Split(name, num_output, input_sample_shape=None)
      • setup(in_shape)
      • get_output_sample_shape()
      • forward()
      • backward()
    • class singa.layer.Concat(name, axis, input_sample_shapes=None)
      • forward(flag, inputs)
      • backward(flag, dy)
    • class singa.layer.Slice(name, axis, slice_point, input_sample_shape=None)
      • get_output_sample_shape()
      • forward(flag, x)
      • backward(flag, grads)
    • class singa.layer.RNN(name, hidden_size, rnn_mode=’lstm’, dropout=0.0, num_stacks=1, input_mode=’linear’, bidirectional=False, param_specs=None, input_sample_shape=None)
      • forward(flag, inputs)
      • backward(flag, grad)
    • class singa.layer.LSTM(name, hidden_size, dropout=0.0, num_stacks=1, input_mode=’linear’, bidirectional=False, param_specs=None, input_sample_shape=None)
    • class singa.layer.GRU(name, hidden_size, dropout=0.0, num_stacks=1, input_mode=’linear’, bidirectional=False, param_specs=None, input_sample_shape=None)
      • singa.layer.get_layer_list()
• 前馈网络
  • class singa.net.FeedForwardNet(loss=None, metric=None)
    • to_device(dev)
    • add(lyr, src=None)
    • param_values()
    • param_specs()
    • param_names()
    • train(x, y)
    • evaluate(x, y)
    • predict(x)
    • topo_sort(layers, src_of_layer)
    • forward(flag, x, output=[])
    • backward()
    • save(f, buffer_size=10, use_pickle=False)
    • load(f, buffer_size=10, use_pickle=False)
• 初始化器(Initializer)
  • Python API
    • singa.initializer.uniform(t, fan_in=0, fan_out=0)
    • singa.initializer.gaussian(t, fan_in=0, fan_out=0)
• 损失(Loss)
  • class singa.loss.Loss
    • backward()
    • evaluate(flag, x, y)
    • forward(flag, x, y)
  • class singa.loss.SoftmaxCrossEntropy
  • class singa.loss.SquaredError
    • backward()
    • evaluate(flag, x, y)
    • forward(flag, x, y)
• 度量(Metric)
  • class singa.metric.Metric
    • forward(x, y)
    • evaluate(x, y)
  • class singa.metric.Accuracy
• 优化器(Optimizer)
  • class singa.optimizer.Optimizer(lr=None, momentum=None, weight_decay=None, regularizer=None, constraint=None)
    • register(name, specs)
    • apply_regularizer_constraint(epoch, value, grad, name=None, step=-1)
    • apply_with_lr(epoch, lr, grad, value, name=None, step=-1)
    • apply(epoch, grad, value, name=None, step=-1)
  • class singa.optimizer.SGD(lr=None, momentum=None, weight_decay=None, regularizer=None, constraint=None)
    • apply_with_lr(epoch, lr, grad, value, name, step=-1)
  • class singa.optimizer.Nesterov(lr=None, momentum=0.9, weight_decay=None, regularizer=None, constraint=None)
    • apply_with_lr(epoch, lr, grad, value, name, step=-1)
  • class singa.optimizer.RMSProp(rho=0.9, epsilon=1e-08, lr=None, weight_decay=None, regularizer=None, constraint=None)
    • apply_with_lr(epoch, lr, grad, value, name, step=-1)
  • class singa.optimizer.AdaGrad(epsilon=1e-08, lr=None, weight_decay=None, lr_gen=None, regularizer=None, constraint=None)
    • apply_with_lr(epoch, lr, grad, value, name, step=-1)
  • class singa.optimizer.Adam(beta_1=0.9, beta_2=0.999, epsilon=1e-08, lr=None, weight_decay=None, regularizer=None, constraint=None)
    • apply_with_lr(epoch, lr, grad, value, name, step=-1)
  • class singa.optimizer.Regularizer
    • apply(epoch, value, grad, step=-1)
  • class singa.optimizer.CppRegularizer(conf)
    • apply(epoch, value, grad, step=-1)
  • class singa.optimizer.L2Regularizer(coefficient)
    • apply(epoch, value, grad, step=-1)
  • class singa.optimizer.Constraint
    • apply(epoch, value, grad, step=-1)
  • class singa.optimizer.CppConstraint(conf)
    • apply(epoch, value, grad, step=-1)
  • class singa.optimizer.L2Constraint(threshold=None)
    • apply(epoch, value, grad, step=-1)
• 数据(Data)
  • class singa.data.ImageBatchIter(img_list_file, batch_size, image_transform, shuffle=True, delimiter=’ ‘, image_folder=None, capacity=10)
• 图像工具
  • class singa.image_tool.ImageTool
    • color_cast(offset=20, inplace=True)
    • crop3(patch, num_case=1, inplace=True)
    • crop5(patch, num_case=1, inplace=True)
    • crop8(patch, num_case=1, inplace=True)
    • enhance(scale=0.2, inplace=True)
    • flip(num_case=1, inplace=True)
    • num_augmentation()
    • random_crop(patch, inplace=True)
    • resize_by_list(size_list, num_case=1, inplace=True)
    • resize_by_range(rng, inplace=True)
    • rotate_by_list(angle_list, num_case=1, inplace=True)
    • rotate_by_range(rng, inplace=True)
    • singa.image_tool.color_cast(img, offset)
    • singa.image_tool.crop(img, patch, position)
    • singa.image_tool.crop_and_resize(img, patch, position)
    • singa.image_tool.enhance(img, scale)
    • singa.image_tool.load_img(path, grayscale=False)
    • singa.image_tool.resize(img, small_size)
• Snapshot
  • class singa.snapshot.Snapshot(f, mode, buffer_size=10)
    • read()
    • write(param_name, param_val)
• Utils
  • singa.utils.update_progress(progress, info)
• 模型库
  • 在Cifar-10上训练CNN
    • 操作说明
      • SINGA安装
      • 数据准备
      • 训练
      • 预测
  • 在文本上训练Char-RNN
    • 操作说明
  • 在MNIST数据集上训练RBM模型
    • 操作说明
  • 在ImageNet上训练AlexNet
    • 操作说明
      • 编译SINGA
      • 数据下载
      • 数据预处理
      • 训练
  • 用DenseNet做图像分类
    • 操作说明
    • 详细信息
  • 用GoogleNet做图像分类
    • 操作说明
    • 详细信息
  • 用Inception V4做图像分类
    • 操作说明
    • 详细信息
  • 用ResNet做图像分类
    • 操作说明
    • 详细信息
  • 用VGG做图像分类
    • 操作说明
    • 详细信息