"""
This module contains functions dedicated to process the json dictionaries
"""
import itertools
import logging
import pandas as pd
log = logging.getLogger(__name__)
# Arbitrary parameters that we need to think about
fail_threshold = 75
fail_reward = -100
loss_threshold = 2.5 # Anything below this is as good as a failure
KEY_GOOD = "good"
KEY_LOSS = "loss"
[docs]def compute_reward(mdict, biggest_ntotal):
"""
Given a combination dictionary computes the reward function:
If the fail rate for this combination is above the fail threshold, rewards is -100
The formula below for the reward takes into account:
- The rate of ok fits that have a loss below the loss_threshold
- The rate of fits that failed
- The std deviation
- How far away is the median from the best loss
- How far away are median and average
"""
# Check the fail rate to see whether this combination is useles
fail_rate = mdict["fail_rate"]
if fail_rate > fail_threshold:
return fail_reward
# Compute from the points that are not explicitly failures
# the ones that are truly good
true_good = mdict["true_good"]
n_good = mdict["n_good"]
n_total = mdict["n_total"]
rate_good = true_good / n_good * 100.0
# Punish the combination which the dispersion
median = mdict["median"]
avg = mdict["avg"]
std = mdict["std"]
dispersion = abs(avg - median) + std / 2.0 + (median - mdict["best_loss"])
# Compute and weight the outliers
combination_weight = 1.0 + n_total / biggest_ntotal
# The most important thing is the true_good_rate
reward = rate_good / 100.0
# Punish a bit using the failure rate, but not that much
reward -= fail_rate / 400.0
# and punish further using the dispersion
reward -= dispersion
# Finally scale it with the total number of points
return reward * combination_weight
[docs]def bin_generator(df_values, max_n=10):
"""
Receives a dataframe with a list of unique values
. If there are more than `max_n` of them
and they are numeric, create `max_n` bins.
If they are already discrete values or there are less than `max_n` options,
output the same input
# Arguments:
- `df_values`: dataframe with unique values
- `maximum`: maximum number of allowed different values
# Returns:
- `new_vals`: list of tuples with (initial, end) value of the bin
"""
values = df_values.values
lval = len(values)
if lval <= max_n:
return values
if not all(isinstance(i, (int, float)) for i in values):
return values
bins = pd.cut(values, max_n, include_lowest=True)
return bins.categories
[docs]def parse_keys(dataframe, keys):
"""
Receives a dataframe and a set of keys
Looks into the dataframe to read the possible values of the keys
Returns a dictionary { 'key' : [possible values] },
If the values are not discrete then we need to bin it
let's do this for anything with two many numerical values
# Arguments:
- `dataframe`: a pandas dataframe
- `keys`: keys to combine
# Returns:
- `key_info`: a dictionary with the possible values for each key
"""
key_info = {}
for key_name in keys:
# Remove duplicates and nans
all_possible_values = dataframe[key_name].dropna().drop_duplicates()
# If there's anything left, add it to the dictionary
if not all_possible_values.empty:
# But bin it first in case we find a continous variable
key_info[key_name] = bin_generator(all_possible_values)
return key_info
[docs]def get_combinations(key_info, ncomb):
"""
Given a dictionary mapping keys to iterables of possible values (`key_info`),
return a list of the product of all possible mappings of a subset of `ncomb`
keys to single values out of the corresponding possible values, for all such subsets.
For instance,
key_info = {
'key1' : [val1-1, val1-2, ...],
'key2' : [val2-1, val2-2, ...],
}
ncomb = 2
will return a list of dictionaries:
[
{'key1' : val1-1, 'key2', val2-1 ... },
{'key1' : val1-1, 'key2', val2-2 ... },
{'key1' : val1-2, 'key2', val2-1 ... },
{'key1' : val1-2, 'key2', val2-2 ... },
]
Get all combinations of ncomb elements for the keys and values given in the dictionary key_info:
# Arguments:
- `key_info`: dictionary with the possible values for each key
- `ncomb`: elements to combine
# Returns:
- `all_combinations`: A list of dictionaries of parameters
"""
# If we don't have enough keys to produce n combinations, return empty
if len(key_info) < ncomb:
return []
# First generate the combinations of keys
key_combinations = itertools.combinations(key_info, ncomb)
all_combinations = []
# Now, for each combination of keys we have to give values
for keys in key_combinations:
# Generate a list of tuples with the values of the keys
# i.e., something like [ (val1-1, val1-2, val1-3...), (val2-1, val2-2...) ... ]
list_of_items = [key_info[key] for key in keys]
# Now combine all these items, which is what we actually want
items_combinations = itertools.product(*list_of_items)
# Now we want to put things back in the form of a dictionary of parameters
for values in items_combinations:
# Each values comes in the same order as `keys`
new_dictionary = dict(zip(keys, values))
all_combinations.append(new_dictionary)
return all_combinations
[docs]def get_slice(dataframe, query_dict):
"""
Returns a slice of the dataframe where some keys match some values
keys_info must be a dictionary {key1 : value1, key2, value2 ...}
# Arguments:
- `dataframe`: a pandas dataframe
- `query_dict`: a dictionary of combination as given by `get_combinations`
"""
df_slice = dataframe
for key, value in query_dict.items():
key_column = df_slice[key]
# Check whether all values of this slice are NaN
if not key_column.empty and key_column.dropna().empty:
return None
# We need to act differently in the case of continous values we discretized before
# The way we have to check whether something was continous is to check whether the value
# is now a pandas interval
if isinstance(value, pd.Interval):
mask = [i in value for i in key_column]
df_slice = df_slice[mask]
else:
df_slice = df_slice[key_column == value]
return df_slice
[docs]def process_slice(df_slice):
"""
Function to process a slice into a dictionary with useful stats
If the slice is None it means the combination does not apply
# Arguments:
- `df_slice`: a slice of a pandas dataframe
# Returns:
- `proc_dict`: a dictionary of stats
"""
# First check whether there's anything inside the slice
n_total = len(df_slice)
if df_slice is None or n_total == 0:
proc_dict = {"skip": True}
return proc_dict
else:
proc_dict = {"skip": False}
# Get the good values
good = df_slice[df_slice[KEY_GOOD]]
# Get raw stats
n_good = len(good)
n_failed = n_total - n_good
fail_rate = n_failed / n_total * 100.0
# Now get the distribution of the (good) losses
good_losses = good[KEY_LOSS]
best_loss = good_losses.min()
std_dev = good_losses.std()
median = good_losses.median()
avg = good_losses.mean()
# Check how many points are under the loss_threshold
true_good = 0
for i in good_losses:
true_good += int(i < loss_threshold)
# Fill the dictionary
proc_dict["n_failed"] = n_failed
proc_dict["n_good"] = n_good
proc_dict["n_total"] = n_total
proc_dict["fail_rate"] = fail_rate
proc_dict["best_loss"] = best_loss
proc_dict["true_good"] = true_good
proc_dict["std"] = std_dev
proc_dict["avg"] = avg
proc_dict["median"] = median
return proc_dict
[docs]def study_combination(dataframe, query_dict):
"""
Given a dataframe and a dictionary of {key1 : value1, key2: value2}
returns a dictionary with a number of stats for that combination
# Arguments:
- `dataframe`: a pandas dataframe
- `query_dict`: a dictionary for a combination as given by `get_combinations`
# Returns:
- `proc_dict`: a dictionary of the "statistics" for this combination
"""
# Get the slice corresponding to this combination
df_slice = get_slice(dataframe, query_dict)
proc_dict = process_slice(df_slice)
proc_dict["slice"] = df_slice
proc_dict["combination"] = query_dict
return proc_dict
[docs]def dataframe_removal(dataframe, hit_list):
"""
Removes all combinations defined in hit_list from the dataframe.
The hit list is list of dictionaries containing the 'slice' key
where 'slice' must be a slice of 'dataframe'
# Arguments:
- `dataframe`: a pandas dataframe
- `hit_list`: the list of element to remove
# Returns:
- `new_dataframe`: the same dataframe with all elements from hit_list removed
"""
if not hit_list:
return dataframe
# I think I am failing to understand how the index object works
# this looks unnecesaryly verbose
# I am just getting all the indices from all the different combinations
# making sure there are no duplicates and then dropping them from the dataframe
indices_to_remove = hit_list[0]["slice"].index
for victim in hit_list[1:]:
indices_to_remove = indices_to_remove.append(victim["slice"].index)
indices_to_drop = indices_to_remove.drop_duplicates()
new_dataframe = dataframe.drop(indices_to_drop)
return new_dataframe
[docs]def autofilter_dataframe(dataframe, keys, n_to_combine=1, n_to_kill=1, threshold=-1):
"""
Receives a dataframe and a list of keys.
Creates combinations of `n_to_combine` keys and computes the reward
Finally removes from the dataframe the `n_to_kill` worse combinations
Anything under `threshold` will be removed and will not count towards the `n_to_kill`
(by default `threshold` = -50 so only things which are really bad will be removed)
# Arguments:
- `dataframe`: a pandas dataframe
- `keys`: keys to combine
- `n_to_combine`: how many keys do we want to combine
- `n_to_kill`: how many combinations to kill
- `threshold`: anything under this reward will be removed
# Returns:
- `dataframe_sliced`: a slice of the dataframe with the weakest combinations
removed
"""
# Step 0: read from the dataframe the content of the keys
key_info = parse_keys(dataframe, keys)
# Step 1: get the combinations
combinations = get_combinations(key_info, n_to_combine)
# Step 2: run through all possible combinations and compute stats
result_list = []
biggest_ntotal = 1
for combination in combinations:
processed_dict = study_combination(dataframe, combination)
if processed_dict["skip"]:
continue
if processed_dict["n_total"] > biggest_ntotal:
biggest_ntotal = processed_dict["n_total"]
result_list.append(processed_dict)
# Step 3: compute reward
n_to_remove = n_to_kill
for processed_dict in result_list:
reward = compute_reward(processed_dict, biggest_ntotal)
log.debug("Combination %s, reward %f", processed_dict["combination"], reward)
if reward <= threshold:
n_to_remove += 1
processed_dict["reward"] = reward
# Step 4: Order the results by reward
result_list.sort(key=lambda i: i["reward"])
# Step 5: Add the n-last to the list of combinations to remove
hit_list = result_list[:n_to_remove]
for i in hit_list:
log.info("Removing %s with reward %f", i["combination"], i["reward"])
# Step 6: remove the bad guys from the dataframe
new_dataframe = dataframe_removal(dataframe, hit_list)
return new_dataframe