python - 如何以编程方式在python中为caffe生成deploy.txt

Question

我编写了 python 代码以编程方式生成卷积神经网络 (CNN)，用于在 caffe 中训练和验证 .prototxt 文件。以下是我的功能:

def custom_net(lmdb, batch_size):

    # define your own net!
    n = caffe.NetSpec()

    # keep this data layer for all networks
    n.data, n.label = L.Data(batch_size=batch_size, backend=P.Data.LMDB, source=lmdb,
                             ntop=2, transform_param=dict(scale=1. / 255))

    n.conv1 = L.Convolution(n.data, kernel_size=6,
                            num_output=48, weight_filler=dict(type='xavier'))
    n.pool1 = L.Pooling(n.conv1, kernel_size=2, stride=2, pool=P.Pooling.MAX)

    n.conv2 = L.Convolution(n.pool1, kernel_size=5,
                            num_output=48, weight_filler=dict(type='xavier'))
    n.pool2 = L.Pooling(n.conv2, kernel_size=2, stride=2, pool=P.Pooling.MAX)

    n.conv3 = L.Convolution(n.pool2, kernel_size=4,
                            num_output=48, weight_filler=dict(type='xavier'))
    n.pool3 = L.Pooling(n.conv3, kernel_size=2, stride=2, pool=P.Pooling.MAX)

    n.conv4 = L.Convolution(n.pool3, kernel_size=2,
                            num_output=48, weight_filler=dict(type='xavier'))
    n.pool4 = L.Pooling(n.conv4, kernel_size=2, stride=2, pool=P.Pooling.MAX)

    n.fc1 = L.InnerProduct(n.pool4, num_output=50,
                           weight_filler=dict(type='xavier'))

    n.drop1 = L.Dropout(n.fc1, dropout_param=dict(dropout_ratio=0.5))

    n.score = L.InnerProduct(n.drop1, num_output=2,
                             weight_filler=dict(type='xavier'))

    # keep this loss layer for all networks
    n.loss = L.SoftmaxWithLoss(n.score, n.label)

    return n.to_proto()

with open('net_train.prototxt', 'w') as f:
    f.write(str(custom_net(train_lmdb_path, train_batch_size)))

with open('net_test.prototxt', 'w') as f:
    f.write(str(custom_net(test_lmdb_path, test_batch_size)))

有没有一种方法可以类似地生成 deploy.prototxt 以测试不在 lmdb 文件中的看不见的数据？如果是这样，如果有人能指出我的引用，我将不胜感激。

Answer 1

很简单:

from caffe import layers as L, params as P
def custom_net(lmdb, batch_size):
    # define your own net!
    n = caffe.NetSpec()

    if lmdb is None: # "deploy" flavor
        # assuming your data is of shape 3x224x224
        n.data = L.Input(input_param={'shape':{'dim':[1,3,224,224]}})
    else:
        # keep this data layer for all networks
        n.data, n.label = L.Data(batch_size=batch_size, backend=P.Data.LMDB, source=lmdb,
                         ntop=2, transform_param=dict(scale=1. / 255))
    # the other layers common to all flavors: train/val/deploy...
    n.conv1 = L.Convolution(n.data, kernel_size=6,
                        num_output=48, weight_filler=dict(type='xavier'))
    n.pool1 = L.Pooling(n.conv1, kernel_size=2, stride=2, pool=P.Pooling.MAX)

    n.conv2 = L.Convolution(n.pool1, kernel_size=5,
                        num_output=48, weight_filler=dict(type='xavier'))
    n.pool2 = L.Pooling(n.conv2, kernel_size=2, stride=2, pool=P.Pooling.MAX)

    n.conv3 = L.Convolution(n.pool2, kernel_size=4,
                        num_output=48, weight_filler=dict(type='xavier'))
    n.pool3 = L.Pooling(n.conv3, kernel_size=2, stride=2, pool=P.Pooling.MAX)

    n.conv4 = L.Convolution(n.pool3, kernel_size=2,
                        num_output=48, weight_filler=dict(type='xavier'))
    n.pool4 = L.Pooling(n.conv4, kernel_size=2, stride=2, pool=P.Pooling.MAX)

    n.fc1 = L.InnerProduct(n.pool4, num_output=50,
                       weight_filler=dict(type='xavier'))
    # do you "drop" i deploy as well? up to you to decide...
    n.drop1 = L.Dropout(n.fc1, dropout_param=dict(dropout_ratio=0.5))
    n.score = L.InnerProduct(n.drop1, num_output=2,
                         weight_filler=dict(type='xavier'))

    if lmdb is None:
        n.prob = L.Softmax(n.score)
    else:
        # keep this loss layer for all networks apart from "Deploy"
        n.loss = L.SoftmaxWithLoss(n.score, n.label)

    return n.to_proto()

现在调用函数:

with open('net_deploy.prototxt', 'w') as f:
    f.write(str(custom_net(None, None)))

如您所见，有 two modifications到 prototxt(以 lmdb 为 None 为条件):
第一个不是 "Data" 层，而是声明性的 "Input" layer仅声明 "data" 而没有声明 "label"。
第二个变化是输出层:你有一个预测层而不是损失层(参见，例如，this answer)。