有效地序列化/反序列化SparseDataFrame

Question

有没有人有效地对熊猫SparseDataFrame进行序列化/反序列化？

import pandas as pd
import scipy
from scipy import sparse
dfs = pd.SparseDataFrame(scipy.sparse.random(1000, 1000).toarray())
# just for testing

挑剔不是答案

这太慢了。

import pickle, time
start = time.time()
# serialization
msg = list(pickle.dumps(dfs, protocol=pickle.HIGHEST_PROTOCOL))
# deserialization
dfs = pickle.loads(bytes(msg))
stop = time.time()
stop - start
# 0.4420337677001953
# This is with Python 3.5 so it's using cPickle

与之相比，msgpack在密集版本上的速度更快

df = dfs.to_dense()
start = time.time()
# serialization
msg = list(df.to_msgpack(compress='zlib'))
# deserialization
df = pd.read_msgpack(bytes(msg))
stop = time.time()
stop - start
# 0.09514737129211426

msgpack

Msgpack就是答案，但我找不到SparseDataFrame（related）的实现

# serialization
dfs.to_msgpack(compress='zlib')
# Returns: NotImplementedError: msgpack sparse frame is not implemented

坐标格式

通过scipy.sparse.coo_matrix采用坐标格式的msgpack似乎值得考虑，但转换为python.sparse.coo_matrix的过程很慢

from scipy.sparse import coo_matrix
start = time.time()

# serialization
columns = dfs.columns
shape = dfs.shape
start_to_coo = time.time()
dfc = dfs.to_coo()
stop_to_coo = time.time()
start_comprehension = time.time()
row = [x.item() for x in df.row]
col = [x.item() for x in df.col]
data = [x.item() for x in df.data]
stop_comprehension = time.time()
start_packing = time.time()
msg = list(msgpack.packb({'columns':list(columns), 'shape':shape, 'row':row, 'col':col, 'data':data}))
stop_packing = time.time()

# deserialization
start_unpacking = time.time()
dict = msgpack.unpackb(bytes(msg))
stop_unpacking = time.time()
columns=dict[b'columns']
index=range(dict[b'shape'][0])
dfc = coo_matrix((dict[b'data'], (dict[b'row'], dict[b'col'])), shape=dict[b'shape'])

stop = time.time()
print('total: ' + str(stop - start))
print('  to_coo: ' + str(stop_to_coo - start_to_coo))
print('  comprehension: ' + str(stop_comprehension - start_comprehension))
print('  packing: ' + str(stop_packing - start_packing))
print('  unpacking: ' + str(stop_unpacking - start_unpacking))

#total: 0.2799222469329834
#  to_coo:               0.22925591468811035
#  comprehension & cast: 0.02356100082397461 (msgpack does not support all numpy formats)
#  packing:              0.004893064498901367
#  unpacking:            0.001984834671020508

从那里开始，似乎需要经历一种密集的格式。

start = time.time()
dfs = pd.SparseDataFrame(dfc.toarray())
stop = time.time()
stop - start
# 2.8947737216949463

Answer 1

时间上的开销来自dumps和loads中的字符串处理。

使用dumps/loads：

def pickle_dumps():
    msg = list(pickle.dumps(dfs, protocol=pickle.HIGHEST_PROTOCOL))
    pickle.loads(bytes(msg))

%timeit pickle_dumps()
# 212 ms ± 2.19 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

使用dump/load：

def pickle_file():
    with open('dump.pickle', 'wb') as f:
        pickle.dump(dfs, f, protocol=pickle.HIGHEST_PROTOCOL)

    with open('dump.pickle', 'rb') as f:
        return pickle.load(f)

%timeit pickle_file()
# 82.7 ms ± 1.25 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

或者使用内置的pandas更短：

def to_pickle():    
    dfs.to_pickle('./dump.pickle')
    pd.read_pickle('./dump.pickle')

%timeit to_pickle()
# 86.8 ms ± 1.54 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

Answer 2

我的考试有问题

dfs = pd.SparseDataFrame(scipy.sparse.random(1000, 1000).toarray())

并没有真正存储稀疏表示。代替

dfs = pd.DataFrame(scipy.sparse.random(1000, 1000).toarray()).to_sparse(fill_value=0)

确实。

此后，稀疏表示中的pickle效果比密集表示中的msgpack更好。

此外，我使用df.row代替了dfc.row。 df指向另一个数据框。 msgpack可能将结果保存在缓存中，并且没有执行任何操作。

更正此错误后，基于coo_matrix的表示形式上的msgpack不会比数据框上的pickle有所改善。