reference: Official Documentation & Cheatsheet
acrynyms:
df: DataFramedfg: DataFrame.GroupBy objectss: Series- UDF: user-defined functions
s.value_counts() # table(R)
s.isna() # NA?
s.isin() # withinselect rows/columns/both: df.loc[] or df.iloc[]
select elements: df.at or df.iat
df[['a', 'b', 'c']] # DataFrame with columns a,b,c
df['a'] # Series
df.a # same as above, do not work with some column names
df[['a']] # DataFrame with a single column
df[df['a'] > 10] # rows where values in column a > 10
df.iloc[10:20] # rows 11-20
df.take(range(10, 20)) # same as above
df.iloc[:, [1, 3]] # columns 1,3
df.take([1, 3], axis='columns') # save as above
df.loc[:, 'a':'c'] # all columns between a to c (both a & c inclusive)
df.loc[df['a'] > 10, ['a', 'c']]
df.iat[1, 2]
df.at[1, 'c']axis: {0 or ‘index’, 1 or ‘columns’}
inplace: bool. Whether to modify the DataFrame rather than creating a new one
# rename
mapper_dict = {'old_name1': 'new_name1', 'old_name2': 'new_name2'}
df.rename(columns=mapper_dict)
df.rename(mapper_dict, axis=1)
# remove
df.drop(columns=['a', 'b', 'c']) # drop selected columns
# create
df.assign(c=lambda df: df.a * df.b)
df['c'] = dff['a'] * df['b'] # same as above
df['c'] = idmap_fca['a'].str.cat(idmap_fca['b'], sep='_') # concatenate two string columns row-wise
df['c'] = df['a'].clip(lower=-3, upper=3) # trim values into range
df['c'] = pd.qcut(df['a'], n, labels=False) # bin columns into n buckets# subset /filter
df.sample(frac=0.5)
df.sample(n=10)
df.head(n)
df.tail(n)
df.nlargest(n, 'value')
df.nsmallest(n, 'value')
df[df.['a'].str.contains('xxx')] # select rows matching a pattern in a column
df.drop_duplicates() # Return DataFrame with duplicate rows removed.
# query
df.query('(a == "xxx") && (b >= 5)') # dplyr::filter alike
df.query('(a == "xxx") && (b + @const >= 5)') # `const` is in the environment
df.query('(`cell type` == "xxx") && (b + @const >= 5)') # quote colnames that are not valid object name with backstick
df.query('a.str.startswith("abc")', engine='python')
# note: column names that are python keywords (like list, import, for, etc) cannot be used.
# sort
df.sort_values('a', assending=False)
df.sort_index()
# other
len(df) # number of rows
df.shape # Tuple (rows, columns)
df.describe() # basic descriptive statistics
df.dropna() # drow rows with NA in any columns
df.fillna() # repalce NA/NULL data with valuedf.reset_index() # convert index(s) to column(s)
df.reset_index(drop=True, inplace=True) # remove index (i.e., set to default integer range)
df.rename(index=mapper_dict, inplace=True) # update index values by dict {old_value: new_value}https://pandas.pydata.org/docs/user_guide/groupby.html
# groupby columns
df.groupby('a') # = df.groupby(['a'])
df.groupby(['a', 'b'])
df.groupby('a', observed=True) # drop unobserved levels if a is `Categorical`
# groupby indexes (int, name or sequence of such)
df.groupby(level=0)
df.groupby(level=['a', 'b']) # multiIndex
# mixed groupby (index + columns)
df.groupby([pd.Grouper(level=0), "a"])- By default the group keys are sorted during the groupby operation. You may however pass
sort=Falsefor potential speedups. - In the agg/transform/filter/apply result, the keys of the groups appear in the index by default. They can be instead included in the columns by passing
as_index=False.
dfg = df.groupby('a')
dfg.groups() # dict {groupname: corresponding_index_values}
len(dfg.groups) # number of groups
dfg['b'] # grouped series
dfg.get_group('group_1') # select a single group
dfg.get_group(('key_1', 'key_2')) # select a single group
# iterate groups
for name, group in dfg:
print(name) # group key value or tuple of group key values if grouped with multiple columns
print(group) # all columns of each group (including grouping columns)Group-wise operations:
| method | input | output |
|---|---|---|
.agg |
pd.Series |
scalar |
.transform |
pd.Series |
pd.Series or scalar |
.filter |
pd.Series or pd.DataFrame |
scalar (boolean) |
.apply |
pd.DataFrame |
scalar or pd.Series or pd.DataFrame |
compute a summary statistic (or statistics) for each group.
# compute sum of each non-grouping columns
df.groupby('a').sum()
df.groupby('a').agg('sum') # .agg is preferred over aggregate, both are same
df.groupby('a').std(numeric_only=True) # excludign non-numeric columns
# builtin agg funs: any, all, size, count (non-NA), nunique, first, last, sum, max, min, mean, median, std, var, ...
# UDF: takes a series, returns a series or scalar
df.group('a').agg(lambda x: x.astype(int).sum()) # supply a function that will applied to each column (x: series)
# multiple functions
df.groupby('a').agg(['sum', 'mean'])
df.groupby('a')['b'].agg(['sum', 'mean'])
df.groupby('a')['b'].agg([lambda x: x.max() - x.min(), lambda x: median])
df.groupby('a').agg({'b': 'sum', 'c': 'median'})
# named agg
df.groupby('a')['b'].agg(min_b='min', max_b='max')
df.groupby('a').agg(
min_b = pd.NamedAgg(column='b', aggfun='min'),
max_b = pd.NamedAgg('b', 'max'),
median_c = pd.NamedAgg('c', 'median')
)perform some group-specific computations and return a like-indexed object. A common use of a transformation is to add the result back into the original DataFrame.
df.groupby('a').cumsum() # cumsum within each group
df.groupby('a').transform('cumsum') # same as above
df['cumsum'] = df.groupby('a').cumsum() # add group-wise cumsum to the original df
# builtin trans: bfill, ffill, shift, diff, rank, cumcount, cumsum, cummax, cummin, cumprod, pct_changeUDFs for .transform should expect a series as input and must:
- Return a result that is either the same size as the group chunk or broadcastable to the size of the group chunk;
- Operate column-by-column on the group chunk. The transform is applied to the first group chunk using chunk.apply.
- Not perform in-place operations on the group chunk.
df.groupby('a').transform(lambda x: (x - x.mean()) / x.std()) # UDF returns a result of the same size as chunk
df.groupby('a').transform(lambda x: x.max() - x.min()) # UDF returns a result broadcastable to the chunk
df.groupby('a').transform(lambda x: x.fillna(x.mean())) # fill NA with group meanexapnding, resample, rolling can be applied to grouped df, too.
discard some groups, according to a group-wise computation that evaluates to True or False. The result of the filter method is then the subset of groups for which the UDF (expects series or dataframe as input) returned True.
# series
s.groupby(s).filter(lambda x: x.sum() > 2)
# For DataFrames with multiple columns, filters should explicitly specify a column as the filter criterion.
df.groupby('a').filter(lambda x: x['b'].sum() > 2)
# works when a single column remains after excluding the group columns
df.groupby('a').filter(lambda x: x.sum() > 2) The function passed to
applymust take a series or dataframe as its first argument and return aDataFrame,Seriesor scalar.applywill then take care of combining the results back together into a single dataframe or series.
df.groupby('a')["b"].apply(lambda x: x.describe())
df.groupby('a')["b"].apply(lambda x: pd.DataFrame({'ori': x, 'new': x - x.mean()})) # returns DataFramedfg.head(1)
dfg.tail(1)
df.pivot_table(index=['col1', 'col2'], # columns to used to grouping and returned as index
columns=['col3', 'col4'], # columns to used to grouping and returned as index
values = ['col5', 'col6'], # columns used to aggregation
aggfunc=len, # aggregation function, default to mean
margins=True # compute row or column margins
fill_value=0 # fill value for empty combinaitons
)
pd.corsstab(df['col1'], df['col2'], margins=True) # similar to table(df$col1, df$col2) in R# merge using columns as keys
# how: inner, left, right, outer
# suffixes: list-like values added to ovelapping column names from two df
pd.merge(df1, df2, how='inner', on='key1')
pd.merge(df1, df2, how='inner', on='key1', suffixes=('_left', '_right'))
pd.merge(df1, df2, how='inner', on=['key1', 'key2'])
pd.merge(df1, df2, how='left', left_on='key1_df1', right_on='key1_df2')
# merge using index as keys
pd.merge(df1, df2, left_index=True, right_index=True)
df1.join(df2)
# keep rows only appear in one df
pd.merge(ydf, zdf, how='outer', indicator=True).
query('_merge == "left_only"').drop(columns=['_merge'])
# concatenate
pd.concat([df1, df2]) # vertical stack
pd.concat([df1, df2], axis=1) # horizontal stack (side by side, but still align rows by index)
pd.concat([df1, df2], axis=1, ignore_index=True) # horizontal stack, ignore row index
pd.concat([df1, df2], axis=1, keys=['df1', 'df2']) # convert column names to multiIndex
pd.concat([s1, s2, s3])
pd.concat([s1, s2, s3], axis=1, join='inner') # outer by default
# concatenate with partial overlap in index or column
s1.combine_first(s2) # keep the first for overlapping index values
df1.combine_first(df2) # keep columns in df1 if overlapping# stack and unstack
df.stack() # convert columns to rows. column names becomes the last level of index
# stack ignore missing vlaues by default, to avoid: `dropna=False`
df.unstack() # convert rows to columns, the last level of index becomes the last level of columns
# `level=` can be specified for both stack and unstack if columns or index is multiIndex.
# wide to long (column names -> column `variable`, column values -> column `value`)
pd.melt(df, id_vars='key')
pd.melt(df, id_vars='key', value_vars=['A', 'B']) # select a subset of volumns to melt
pd.melt(df, value_vars=['A', 'B']) # id_vars not necessary
pd.melt(df) # melt all columns
# long to wide
pd.pivot(df, index=['col_a', 'col_b'], columns='group', values=['col_val1', 'col_val2'])
# equal to : df.set_index(['col_a', 'col_b', 'group']).unstack(level='group')# read
pd.read_table('df.tsv')
pd.read_csv('df.tsv', sep='\t')
pd.read_excel('df.xlsx')df.plot.hist() # hist for each column
df.plot.scatter(x='a', y='b')s = pd.Series(
np.random.uniform(size=9),
index=[["a", "a", "a", "b", "b", "c", "c", "d", "d"], [1, 2, 3, 1, 3, 1, 2, 2, 3]])
s.loc['a']
s['a'] # level 0 == a
s.loc[:, 2]
s[:, 2] # level 1 == 2
s.loc['a':'c']
s['a':'c'] # level 0 == a or b or c
s.loc[['a', 'c']]
s[['a', 'c']] # level 0 == a or c
s.loc[('a', 2)]
s[('a', 2)] # level 0 == a, level 1 == 2
# below we only show simplier version. `s.loc[]` version will be skipped
s[[('a', 2), ('b', 3)]] # select two elements
s[(['a', 'c'], 2)] # select level 0 == a or c, level == 2
s[(slice('a', 'c'), 2)] # select level 0 == a to c, level == 2
s[(slice(None), [1, 3])] # select level 1 == 1 or 3
# for df
df['a'] # column level 0 == a
df[('a', 'b')] # column level 0 == a, level 1 == b
df.loc['a'] # index level 0 == a
df.loc[('a', 'b')] # index level 0 == a, level 1 == b
df.loc[(slice(None), 'b'), :] # index level 1 == b
df.loc[(slice('a', 'c'), 'b'), :] # index level 0 == a to c, level 1 == b
df.loc[(['a', 'c'], 'b'), :] # index level 0 == a or c, level 1 == b
# convert multiIndex to flat index
df.columns = ["_".join(idx) for idx in df.columns.to_flat_index()]