我可以使用 DataFrame.to_csv 和 pandas.read_csv 来一致地读写浮点类型吗？

Question

I'd like to write float values to a CSV file using DataFrame.to_csv and ensure that upon reading it back with pandas.read_csv, I get the exact same in-memory value. The text representation doesn't have to make sense to a person reading.

Are there common textual representations of Python float values? Or a reliable way to deserialize and serialize float to text?

float_format doesn't guarantee read-write reliability

Answer 1

Yes and no. If your floats are in float64 format, then it doesn't make a difference; this is the default float type for pandas. If you're saving any other float type (such as float32 or float16), you risk losing it unless you know the type in advance and can pass that to read_csv .

df = pd.DataFrame(np.random.randn(5, 2), dtype=np.float16)
df.to_csv('data.csv', index=False)

pd.read_csv('data.csv').dtypes
0    float64  # this should be float16, right?
1    float64
dtype: object

pd.read_csv('data.csv', dtype=pd.np.float16).dtypes # need dtype=... here
0    float16
1    float16
dtype: object

OTOH, pickling your data is a much better option if you intend to preserve the data, it is also more compact and should be a bit faster (not timed).

df.to_pickle('data.pkl')

pd.read_pickle('data.pkl').dtypes
0    float16
1    float16
dtype: object