0

我正在升级到 caffe2，我尝试了一个玩具问题：

# Read the data from the database
def read_input(model, batch_size, db, db_type):
    # load the data
    data, binary, dist = model.TensorProtosDBInput(
        [], ['data', 'binary'], batch_size=batch_size,
        db=db, db_type=db_type)
    # Get the absolute distance.
    data = model.StopGradient(data, data)
    binary = model.StopGradient(binary, binary)
    return data, binary


def define_network(model, inp, b_classifier):
    fc1 = brew.fc(model, inp, 'fc1', dim_in=2, dim_out=10)
    fc1 = brew.relu(model, fc1, fc1)
    fc2 = brew.fc(model, fc1, 'fc2', dim_in=10, dim_out=20)
    fc2 = brew.relu(model, fc2, fc2)
    fc3 = brew.fc(model, fc2, 'fc3', dim_in=20, dim_out=10)
    fc3 = brew.relu(model, fc3, fc3)

    nDimOut = 1+int(b_classifier)
    predict = brew.fc(model, fc3, 'predict', dim_in=10, dim_out=nDimOut)
    if b_classifier:
        softmax = brew.softmax(model, predict, 'softmax')
        return softmax
    return predict


def add_classifier_loos_and_acc(model, softmax, gt):
    crossEnt = model.LabelCrossEntropy([softmax, gt], 'crossEnt')
    loss = model.AveragedLoss(crossEnt, "loss")
    accuracy = brew.accuracy(model, [softmax, gt], "accuracy")

    return loss, accuracy


def add_training_ops(model, loss):
    model.AddGradientOperators([loss])
    optimizer.build_rms_prop(model, base_learning_rate=0.001)

batchSize = 128
# Create a training network
train_model = model_helper.ModelHelper(name="binary_train")
data, binary = read_input(train_model, batch_size=batchSize, db='Data/' + folder + '/train.minidb', db_type='minidb')
softmax = define_network(train_model, data, True)
loss, _ = add_classifier_loos_and_acc(train_model, softmax, binary)
add_training_ops(train_model, loss)

# Train and check.
workspace.RunNetOnce(train_model.param_init_net)
# creating the network
workspace.CreateNet(train_model.net, overwrite=True)

# Train once
workspace.RunNet(train_model.net)
testRes = workspace.FetchBlob('softmax')
gt = workspace.FetchBlob('binary')
crossEnt = workspace.FetchBlob('crossEnt')
avgCrossEnt = np.mean(crossEnt)
loss = workspace.FetchBlob('loss')

当我运行代码并到达workspace.RunNet(train_model.net) 行时，我的代码崩溃了：

RuntimeError: [enforce fail at accuracy_op.cc:29] label.ndim() == 1. 2 vs 1 Error from operator: 
input: "softmax" input: "binary" output: "accuracy" name: "" type: "Accuracy"
We've got an error while stopping in post-mortem: <type 'exceptions.KeyboardInterrupt'>

我试图理解它大约 2 天，但我什么也没得到。有谁知道我做错了什么？

谢谢 ！

4

1 回答 1

0

我发现了问题。问题是二进制文件的形状是 (batchSize,1)，它的大小应该是 (batchSize,)。添加 FlattenToVec() 解决了这个问题

于 2018-01-29T12:55:32.763 回答