(require '[clojupyter.misc.helper :as helper])
(helper/add-dependencies '[panthera "0.1-alpha.13"])
:ok
:ok
(require '[panthera.panthera :as pt])
(require '[libpython-clj.python :as py])
nil
The show
function is a helper to render data-frames
(require '[clojupyter.display :as display])
(defn show
[obj]
(display/html
(py/call-attr obj "to_html")))
#'user/show
Serieses are like vectors that act also as columns for data-frames (see Data-frame section). One series must have all the contained data with the same data type and if there is more than one type when you create a series than this one takes the most relaxed one.
(pt/series [1 2 3])
0 1 1 2 2 3 dtype: int64
If we print the series we see on the left its index and on the right its values. As you can see below the series itself we get the underlying data type (dtype) as well. Let's swap 3 with "a" and see what happens.
(pt/series [1 2 "a"])
0 1 1 2 2 a dtype: object
Now the dtype it's become object
, which in panthera means either string
or something that can be represented with a string
and is not a primitive.
If we get this data back to Clojure we'll see that we get the underlying original representation with mixed data types.
(vec (pt/series [1 2 "a"]))
[1 2 "a"]
This means that we can always move from a representation to another without many problems. A series can be treated as a Clojure vector if we want to:
(map inc (pt/series (range 3)))
(1 2 3)
But when we do this we lose metadata tied to it. The difference with regular vectors is mostly this metadata, a series specifically:
Let's see a few examples:
(pt/series {:name "my-series"})
name my-series dtype: object
We just created an empty series with the name "my-series" to show that it can exist even with just metadata. The map passed as an argument lets you add other options to the function call without bothering about their position (in Python there is a clear distinction between arguments and keyword arguments, more info).
We can combine arguments together to get the wanted outcome
(pt/series 1 {:name "my-series" :index ["idx"]})
idx 1 Name: my-series, dtype: int64
Now "my-series" has a name, a value and a named index. This distinction is very important in panthera: indexing can be done by name and by position.
(-> (pt/series (range 5) {:name "my-series" :index ["a" "b" "c" "d" "e"]})
(pt/select-rows [0 3]))
a 0 d 3 Name: my-series, dtype: int64
(-> (pt/series (range 5) {:name "my-series" :index ["a" "b" "c" "d" "e"]})
(pt/select-rows ["a" "d"] :loc))
a 0 d 3 Name: my-series, dtype: int64
As you can see above we were able to get the same values from the series, but the first time we used pure positional indexing, while the second one we used named indexing.
This isn't something logical, it just works like this in pandas. So you'll have to memorize:
:iloc
: positional indexing:loc
: named indexing or booleansBe aware that the result of this can be this behaviour:
(-> (pt/series (range 5) {:name "my-series" :index (map #(+ 100 %) (range 5))})
(pt/select-rows [0 3] :iloc))
100 0 103 3 Name: my-series, dtype: int64
(-> (pt/series (range 5) {:name "my-series" :index (map #(+ 100 %) (range 5))})
(pt/select-rows [100 103] :loc))
100 0 103 3 Name: my-series, dtype: int64
(-> (pt/series (range 5) {:name "my-series"})
(pt/select-rows [0 3] :loc))
0 0 3 3 Name: my-series, dtype: int64
What happens above is that somewhat unexpectedly we get always the same values. Let's review every cell by itself:
{100 0 101 1 102 2 ...}
, but this in panthera doesn't change the fact that the first value is 0, the second is 1 and so on. So by getting [0 3]
the result is a series with the first and fourth values(select-keys {100 0 101 1 102 2 ...} [100 103])
would give the same result:loc
), but they are integers and they are positional. This happens because when we don't have named indices both serieses and data-frames assign a monotonically increasing index that has the value of the index itself as a label. If we had to represent a panthera index in pure Clojure it would be something like {0 "0" 1 "1" 2 "2" ...}
There's another way to subset by index: slicing
(-> (pt/series (range 10))
(pt/select-rows (pt/slice 3 6)))
3 3 4 4 5 5 dtype: int64
(-> (pt/series (range 5) {:name "my-series" :index ["a" "b" "c" "d" "e"]})
(pt/select-rows (pt/slice "a" "d") :loc))
a 0 b 1 c 2 d 3 Name: my-series, dtype: int64
Math is easy with panthera! The only thing to keep in mind is that operations are vectorized, so something like (+ [1 2 3] 1)
would result in [2 3 4]
.
To avoid confusion the panthera operations are named differently than the core functions (+
, -
, *
, etc).
(pt/add (pt/series [1 2 3]) 1)
0 2 1 3 2 4 dtype: int64
(pt/pow (pt/series (range 5)) 3)
0 0 1 1 2 8 3 27 4 64 dtype: int64
(pt/add (pt/series [1 2 3]) 1 (pt/series [-1 -2 -3]))
0 1 1 1 2 1 dtype: int64
The only note about these operations is that in order to work the first argument has to be a panthera data structure.
There are more advanced stats functions besides the more regular ones:
(pt/mean (pt/series (range 10)))
4.5
(pt/kurtosis (pt/series (concat (range 10) [100])))
10.712688874485469
(pt/skew (pt/series (concat (range 10) [100])))
3.2568924988901746
(pt/var (pt/series (concat (range 10) [100])))
837.3636363636364
(pt/corr (pt/series (range 10)) (pt/series (range 9 0 -1)))
-1.0
It might happen that you'd like to work with different data types than the ones inferred by panthera. The advice here is to do this only on the Python side of things.
(pt/->numeric (pt/series ["1" "2"]))
0 1 1 2 dtype: int64
(pt/->datetime "2019-01-01")
2019-01-01 00:00:00
(pt/->datetime (pt/series ["2019-01-01" "2019-02-01"]))
0 2019-01-01 1 2019-02-01 dtype: datetime64[ns]
Below an example of why you should be careful to deal with different data types in panthera
(-> (pt/series ["2019-01-01" "2019-02-01"])
pt/->datetime
pt/->clj)
[{:unnamed 2019-01-01 00:00:00} {:unnamed 2019-02-01 00:00:00}]
(-> (pt/series ["2019-01-01" "2019-02-01"])
pt/->datetime
pt/->clj
first
:unnamed
type)
:pyobject
The safest way to deal with dates on the Clojure side of things is to convert them to strings
(-> (pt/series ["2019-01-01" "2019-02-01"])
pt/->datetime
pt/->clj
first
:unnamed
str)
"2019-01-01 00:00:00"
You can have fun with regular numeric types as well
(pt/astype (pt/series [1 2 3]) :float32)
0 1.0 1 2.0 2 3.0 dtype: float32
There are many facilities to let you hack 'n' slash data almost however you want
(pt/cut (pt/series (range 10)) 3)
0 (-0.009, 3.0] 1 (-0.009, 3.0] 2 (-0.009, 3.0] 3 (-0.009, 3.0] 4 (3.0, 6.0] 5 (3.0, 6.0] 6 (3.0, 6.0] 7 (6.0, 9.0] 8 (6.0, 9.0] 9 (6.0, 9.0] dtype: category Categories (3, interval[float64]): [(-0.009, 3.0] < (3.0, 6.0] < (6.0, 9.0]]
Intervals aren't handled (yet) on the Clojure side, so keep 'em strictly in Python if you want to deal with them.
With factorize
you can convert values to ints, so basically you get categories.
(pt/factorize (pt/series [:a :b :c]))
(array([0, 1, 2]), Index(['a', 'b', 'c'], dtype='object'))
With remap
yu can, well, remap your values however you like. Just be aware that you have to pass remap
every value present in the series in the new encoding, otherwise those not specified will be interpreted as NaNs.
(pt/remap (pt/series [:a :b :c]) {:a "this" :b "that"})
0 this 1 that 2 NaN dtype: object
An example on one way to deal with remap
when you want to remap only some values
(def remapper
(-> (pt/series [:a :b :c :d :e :f :g :h :i :j])
pt/unique
(#(zipmap % %))
(assoc "e" "only-this-one")))
(pt/remap
(pt/series [:a :b :c :d :e :f :g :h :i :j])
remapper)
0 a 1 b 2 c 3 d 4 only-this-one 5 f 6 g 7 h 8 i 9 j dtype: object
rolling
lets you calculate statistics on a rolling window basis
(pt/rolling (pt/series (range 10)) 2)
Rolling [window=2,center=False,axis=0]
(-> (pt/series (range 10))
(pt/rolling 2)
pt/mean)
0 NaN 1 0.5 2 1.5 3 2.5 4 3.5 5 4.5 6 5.5 7 6.5 8 7.5 9 8.5 dtype: float64
Dealing with missing values is what really makes the difference between a full-fledged data analysis framework and much more limited solutions.
panthera gives you many options to try easing the pain a bit
(pt/dropna (pt/series [1 2 nil 3]))
0 1.0 1 2.0 3 3.0 dtype: float64
Note that though the name might let you think that we're mutating the original series, this is similar to Clojure's drop
(def my-srs (pt/series [1 2 nil 3]))
(pt/dropna my-srs)
my-srs
0 1.0 1 2.0 2 NaN 3 3.0 dtype: float64
There are various ways to check if your data contains some missing observation. The easiest and fastest one is hasnans?
.
(pt/hasnans? (pt/series (concat (range 1000) [nil])))
true
hasnans?
is a cached value, but this shouldn't be an issue considering that everything is as immutable as possible.
This is another potentially slower way to do the same thing
(pt/all? (pt/not-na? (pt/series (concat (range 1000) [nil]))))
false
Of course not-na?
and all?
have their uses (for instance if you pass the result of not-na?
to select-rows
you'll filter NaNs out of the series).
panthera's workhorse to deal with missing observations is fill-na
which lets you assign a value to NaNs
(pt/fill-na (pt/series [1 2 nil 4]) 3)
0 1.0 1 2.0 2 3.0 3 4.0 dtype: float64
(pt/data-frame [{:a 1 :b 2} {:a 3 :b 4}])
a b 0 1 2 1 3 4
The easiest way to create a data-frame is with a vector of maps where every map is a row, keys are columns names and values, well, are corresponding values.
As we saw earlier data-frames are a collection of serieses, so we can create one starting from a bunch of them.
(pt/data-frame [(pt/series [1 2 3]) (pt/series [4 5 6])])
0 1 2 0 1 2 3 1 4 5 6
(pt/data-frame {:a (pt/series [1 2 3]) :b (pt/series [:x :y :z])})
a b 0 1 x 1 2 y 2 3 z
(pt/dtype (pt/data-frame {:a (pt/series [1 2 3]) :b (pt/series [:x :y :z])}))
a int64 b object dtype: object
Above we see that panthera doesn't complain about the column a
having type int64
and b
having type object
and we can keep working on them as much as we want.
Now we have two dimensions to work with! No worries, it is always possible to operate on both of them. But first let's check all the metadata available to us
(def df (pt/data-frame [{:a 1 :b 2} {:a 3 :b 4}]))
(pt/index df)
RangeIndex(start=0, stop=2, step=1)
(pt/names df)
Index(['a', 'b'], dtype='object')
So, what we saw for serieses works for data-frames as well
(def df (pt/data-frame (map #(zipmap [:a :b :c] %) (partition 3 (range 30)))))
(pt/select-rows df [0 5])
a b c 0 0 1 2 5 15 16 17
(pt/select-rows df (pt/slice 2 5))
a b c 2 6 7 8 3 9 10 11 4 12 13 14
The new thing is subsetting columns, we can do this by name with subset-cols
. You can select any number of columns in this way, as long as they are in the given data-frame
(pt/subset-cols df :a :c)
a c 0 0 2 1 3 5 2 6 8 3 9 11 4 12 14 5 15 17 6 18 20 7 21 23 8 24 26 9 27 29
(pt/mean df)
a 13.5 b 14.5 c 15.5 dtype: float64