0

我在正确加载 .csv 文件以用作非常简单的密集 NN 模型的输入时遇到问题。csv 文件包含所有输入特征和一个“目标”列，用作回归的输出。

到目前为止，这就是我正在做的事情：

def main():

    batch_size = 500

    ## load input file
    df_data = pd.read_csv('some_file.csv', index_col=0)
    ## random train/test split
    df_train = df_data.sample(frac=0.8,random_state=200)
    df_test = df_data.drop(df_train.index)

    ## data pre-processing
    df_train.reset_index(drop=True, inplace=True)
    df_test.reset_index(drop=True, inplace=True)    
    y_train = df_train['target'].to_numpy(dtype=np.float64)
    y_test = df_test['target'].to_numpy(dtype=np.float64)
    X_train = df_train.drop(['target'], axis=1).to_numpy(dtype=np.float64)
    X_test = df_test.drop(['target'], axis=1).to_numpy(dtype=np.float64)


    dataset = mx.gluon.data.dataset.ArrayDataset(X_train, y_train)
    data_loader = mx.gluon.data.DataLoader(dataset, batch_size=batch_size, shuffle=True)

    ##   building model 
    model = nn.Sequential()
    model.add(nn.Dense(150))
    model.add(nn.Dense(1))
    model.initialize(init.Normal(sigma=0.01))

    ## loss function (squared loss)
    loss = gloss.L2Loss()

    ## optimization algorithm, specify:
    trainer = gluon.Trainer(model.collect_params(), 'sgd', {'learning_rate': 0.03})

    ##   training   #
    num_epochs = 10
    for epoch in range(1, num_epochs + 1):
        for X_batch, Y_batch in data_loader:
            with autograd.record():
                l = loss(model(X_batch), Y_batch)
            l.backward()
            trainer.step(batch_size)
        # overall (entire dataset) loss after epoch
        l = loss(model(X_train), y_train)
        print(f'\nEpoch {epoch}, loss: {l.mean().asnumpy()}')

我收到了错误：

mxnet.base.MXNetError: [16:09:03] src/operator/numpy/linalg/./../../tensor/../elemwise_op_common.h:135: Check failed: assign(&dattr, vec.at(i)): Incompatible attr in node  at 1-th input: expected float64, got float32

因此，我尝试通过将 np.float64 切换为 np.float32 来转换数据，但我得到：

File "/home/lews/anaconda3/envs/gluon/lib/python3.7/site-packages/mxnet/gluon/block.py", line 1136, in forward
raise ValueError('In HybridBlock, there must be one NDArray or one Symbol in the input.'
ValueError: In HybridBlock, there must be one NDArray or one Symbol in the input. Please check the type of the args.

加载此数据的正确方法是什么？

4

1 回答 1

0

我通过使用修复它

 ## data pre-processing
y_train = np.array(df_train['target'].to_numpy().reshape(-1,1), dtype=np.float32)
y_test = np.array(df_test['target'].to_numpy().reshape(-1,1), dtype=np.float32)
X_train = np.array(df_train.drop(['target'], axis=1).to_numpy(), dtype=np.float32)
X_test = np.array(df_test.drop(['target'], axis=1).to_numpy(), dtype=np.float32)
于 2020-12-29T13:57:15.467 回答