Upload splitters.py with huggingface_hub
Browse files- splitters.py +72 -87
splitters.py
CHANGED
|
@@ -1,19 +1,10 @@
|
|
| 1 |
import itertools
|
| 2 |
from abc import abstractmethod
|
| 3 |
-
from
|
| 4 |
-
from typing import Dict, List, Optional
|
| 5 |
|
| 6 |
from .artifact import Artifact
|
| 7 |
-
from .
|
| 8 |
-
from .
|
| 9 |
-
from .stream import MultiStream
|
| 10 |
-
|
| 11 |
-
|
| 12 |
-
class Splitter(MultiStreamOperator):
|
| 13 |
-
pass
|
| 14 |
-
|
| 15 |
-
|
| 16 |
-
from .random_utils import random
|
| 17 |
from .split_utils import (
|
| 18 |
parse_random_mix_string,
|
| 19 |
parse_slices_string,
|
|
@@ -21,6 +12,11 @@ from .split_utils import (
|
|
| 21 |
rename_split,
|
| 22 |
slice_streams,
|
| 23 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 24 |
|
| 25 |
|
| 26 |
class RenameSplits(Splitter):
|
|
@@ -41,8 +37,8 @@ class SplitRandomMix(Splitter):
|
|
| 41 |
|
| 42 |
|
| 43 |
class SeparateSplit(Splitter):
|
| 44 |
-
"""
|
| 45 |
-
|
| 46 |
sizes must indicate the size of every split except the last. If no size is give for the last split,
|
| 47 |
it includes all the examples not allocated to any split.
|
| 48 |
"""
|
|
@@ -59,9 +55,15 @@ class SeparateSplit(Splitter):
|
|
| 59 |
return super().verify()
|
| 60 |
|
| 61 |
def process(self, multi_stream: MultiStream) -> MultiStream:
|
| 62 |
-
mapping = {
|
|
|
|
|
|
|
|
|
|
|
|
|
| 63 |
so_far = 0
|
| 64 |
-
for name, size in itertools.zip_longest(
|
|
|
|
|
|
|
| 65 |
mapping[name] = {self.from_split: [(so_far, size)]}
|
| 66 |
if size:
|
| 67 |
so_far += size
|
|
@@ -87,19 +89,25 @@ class Sampler(Artifact):
|
|
| 87 |
|
| 88 |
def set_size(self, size):
|
| 89 |
if isinstance(size, str):
|
| 90 |
-
assert
|
|
|
|
|
|
|
| 91 |
size = int(size)
|
| 92 |
self.sample_size = size
|
| 93 |
|
| 94 |
@abstractmethod
|
| 95 |
-
def sample(
|
|
|
|
|
|
|
| 96 |
pass
|
| 97 |
|
| 98 |
|
| 99 |
class RandomSampler(Sampler):
|
| 100 |
-
def sample(
|
|
|
|
|
|
|
| 101 |
instances_pool = list(instances_pool)
|
| 102 |
-
return
|
| 103 |
|
| 104 |
|
| 105 |
class DiverseLabelsSampler(Sampler):
|
|
@@ -110,14 +118,29 @@ class DiverseLabelsSampler(Sampler):
|
|
| 110 |
self.labels = None
|
| 111 |
|
| 112 |
def examplar_repr(self, examplar):
|
| 113 |
-
|
| 114 |
-
"inputs
|
| 115 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 116 |
examplar_outputs = next(iter(examplar["outputs"].values()))
|
| 117 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 118 |
|
| 119 |
def divide_by_repr(self, examplars_pool):
|
| 120 |
-
labels =
|
| 121 |
for examplar in examplars_pool:
|
| 122 |
label_repr = self.examplar_repr(examplar)
|
| 123 |
if label_repr not in labels:
|
|
@@ -125,11 +148,13 @@ class DiverseLabelsSampler(Sampler):
|
|
| 125 |
labels[label_repr].append(examplar)
|
| 126 |
return labels
|
| 127 |
|
| 128 |
-
def sample(
|
|
|
|
|
|
|
| 129 |
if self.labels is None:
|
| 130 |
self.labels = self.divide_by_repr(instances_pool)
|
| 131 |
all_labels = list(self.labels.keys())
|
| 132 |
-
|
| 133 |
from collections import Counter
|
| 134 |
|
| 135 |
total_allocated = 0
|
|
@@ -146,22 +171,21 @@ class DiverseLabelsSampler(Sampler):
|
|
| 146 |
|
| 147 |
result = []
|
| 148 |
for label, allocation in allocations.items():
|
| 149 |
-
sample =
|
| 150 |
result.extend(sample)
|
| 151 |
|
| 152 |
-
|
| 153 |
return result
|
| 154 |
|
| 155 |
|
| 156 |
-
class SpreadSplit(
|
| 157 |
source_stream: str = None
|
| 158 |
target_field: str = None
|
| 159 |
sampler: Sampler = None
|
| 160 |
|
| 161 |
def prepare(self):
|
| 162 |
-
self.accessible_streams = [self.source_stream]
|
| 163 |
-
self.cache_accessible_streams = True
|
| 164 |
self.local_cache = None
|
|
|
|
| 165 |
|
| 166 |
def verify(self):
|
| 167 |
assert self.source_stream is not None, "Source stream must be specified"
|
|
@@ -169,58 +193,19 @@ class SpreadSplit(InstanceOperatorWithGlobalAccess):
|
|
| 169 |
assert self.sampler is not None, "Sampler must be specified"
|
| 170 |
return super().verify()
|
| 171 |
|
| 172 |
-
def process(
|
| 173 |
-
|
| 174 |
-
|
| 175 |
-
|
| 176 |
-
|
| 177 |
-
|
| 178 |
-
|
| 179 |
-
|
| 180 |
-
|
| 181 |
-
|
| 182 |
-
|
| 183 |
-
|
| 184 |
-
|
| 185 |
-
|
| 186 |
-
|
| 187 |
-
|
| 188 |
-
splitter = SplitRandomMix(
|
| 189 |
-
mix={
|
| 190 |
-
"train": "train[90%]+validation[50%]",
|
| 191 |
-
"validation": "train[10%]+validation[50%]",
|
| 192 |
-
"test": "test",
|
| 193 |
-
}
|
| 194 |
-
)
|
| 195 |
-
|
| 196 |
-
def generator(name, size):
|
| 197 |
-
for i in range(size):
|
| 198 |
-
yield {"text": f"{name}_{i}"}
|
| 199 |
-
|
| 200 |
-
stream = MultiStream.from_generators(
|
| 201 |
-
{
|
| 202 |
-
"train": ReusableGenerator(generator, gen_kwargs={"name": "train", "size": 10}),
|
| 203 |
-
"validation": ReusableGenerator(generator, gen_kwargs={"name": "validation", "size": 10}),
|
| 204 |
-
"test": ReusableGenerator(generator, gen_kwargs={"name": "test", "size": 10}),
|
| 205 |
-
}
|
| 206 |
-
)
|
| 207 |
-
|
| 208 |
-
ds = splitter(stream)
|
| 209 |
-
for key, value in ds.items():
|
| 210 |
-
print(key)
|
| 211 |
-
for item in value:
|
| 212 |
-
print(item)
|
| 213 |
-
|
| 214 |
-
splitter = SliceSplit(
|
| 215 |
-
slices={
|
| 216 |
-
"train": "train[:2]+train[2:4]",
|
| 217 |
-
"validation": "train[4:6]",
|
| 218 |
-
"test": "train[6:]+test",
|
| 219 |
-
}
|
| 220 |
-
)
|
| 221 |
-
|
| 222 |
-
ds = splitter(stream)
|
| 223 |
-
for key, value in ds.items():
|
| 224 |
-
print(key)
|
| 225 |
-
for item in value:
|
| 226 |
-
print(item)
|
|
|
|
| 1 |
import itertools
|
| 2 |
from abc import abstractmethod
|
| 3 |
+
from typing import Dict, List
|
|
|
|
| 4 |
|
| 5 |
from .artifact import Artifact
|
| 6 |
+
from .operator import InstanceOperatorWithMultiStreamAccess, MultiStreamOperator
|
| 7 |
+
from .random_utils import get_random
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 8 |
from .split_utils import (
|
| 9 |
parse_random_mix_string,
|
| 10 |
parse_slices_string,
|
|
|
|
| 12 |
rename_split,
|
| 13 |
slice_streams,
|
| 14 |
)
|
| 15 |
+
from .stream import MultiStream
|
| 16 |
+
|
| 17 |
+
|
| 18 |
+
class Splitter(MultiStreamOperator):
|
| 19 |
+
pass
|
| 20 |
|
| 21 |
|
| 22 |
class RenameSplits(Splitter):
|
|
|
|
| 37 |
|
| 38 |
|
| 39 |
class SeparateSplit(Splitter):
|
| 40 |
+
"""Separates a split (e.g. train) into several splits (e.g. train1, train2).
|
| 41 |
+
|
| 42 |
sizes must indicate the size of every split except the last. If no size is give for the last split,
|
| 43 |
it includes all the examples not allocated to any split.
|
| 44 |
"""
|
|
|
|
| 55 |
return super().verify()
|
| 56 |
|
| 57 |
def process(self, multi_stream: MultiStream) -> MultiStream:
|
| 58 |
+
mapping = {
|
| 59 |
+
key: {key: [(None, None)]}
|
| 60 |
+
for key in multi_stream.keys()
|
| 61 |
+
if key != self.from_split
|
| 62 |
+
}
|
| 63 |
so_far = 0
|
| 64 |
+
for name, size in itertools.zip_longest(
|
| 65 |
+
self.to_split_names, self.to_split_sizes
|
| 66 |
+
):
|
| 67 |
mapping[name] = {self.from_split: [(so_far, size)]}
|
| 68 |
if size:
|
| 69 |
so_far += size
|
|
|
|
| 89 |
|
| 90 |
def set_size(self, size):
|
| 91 |
if isinstance(size, str):
|
| 92 |
+
assert (
|
| 93 |
+
size.isdigit()
|
| 94 |
+
), f"sample_size must be a natural number, got {self.sample_size}"
|
| 95 |
size = int(size)
|
| 96 |
self.sample_size = size
|
| 97 |
|
| 98 |
@abstractmethod
|
| 99 |
+
def sample(
|
| 100 |
+
self, instances_pool: List[Dict[str, object]]
|
| 101 |
+
) -> List[Dict[str, object]]:
|
| 102 |
pass
|
| 103 |
|
| 104 |
|
| 105 |
class RandomSampler(Sampler):
|
| 106 |
+
def sample(
|
| 107 |
+
self, instances_pool: List[Dict[str, object]]
|
| 108 |
+
) -> List[Dict[str, object]]:
|
| 109 |
instances_pool = list(instances_pool)
|
| 110 |
+
return get_random().sample(instances_pool, self.sample_size)
|
| 111 |
|
| 112 |
|
| 113 |
class DiverseLabelsSampler(Sampler):
|
|
|
|
| 118 |
self.labels = None
|
| 119 |
|
| 120 |
def examplar_repr(self, examplar):
|
| 121 |
+
if "inputs" not in examplar:
|
| 122 |
+
raise ValueError(f"'inputs' field is missing from '{examplar}'.")
|
| 123 |
+
inputs = examplar["inputs"]
|
| 124 |
+
if self.choices not in inputs:
|
| 125 |
+
raise ValueError(f"{self.choices} field is missing from '{inputs}'.")
|
| 126 |
+
choices = inputs[self.choices]
|
| 127 |
+
if not isinstance(choices, list):
|
| 128 |
+
raise ValueError(
|
| 129 |
+
f"Unexpected input choices value '{choices}'. Expected a list."
|
| 130 |
+
)
|
| 131 |
+
|
| 132 |
+
if "outputs" not in examplar:
|
| 133 |
+
raise ValueError(f"'outputs' field is missing from '{examplar}'.")
|
| 134 |
examplar_outputs = next(iter(examplar["outputs"].values()))
|
| 135 |
+
if not isinstance(examplar_outputs, list):
|
| 136 |
+
raise ValueError(
|
| 137 |
+
f"Unexpected examplar_outputs value '{examplar_outputs}'. Expected a list."
|
| 138 |
+
)
|
| 139 |
+
|
| 140 |
+
return str([choice for choice in choices if choice in examplar_outputs])
|
| 141 |
|
| 142 |
def divide_by_repr(self, examplars_pool):
|
| 143 |
+
labels = {}
|
| 144 |
for examplar in examplars_pool:
|
| 145 |
label_repr = self.examplar_repr(examplar)
|
| 146 |
if label_repr not in labels:
|
|
|
|
| 148 |
labels[label_repr].append(examplar)
|
| 149 |
return labels
|
| 150 |
|
| 151 |
+
def sample(
|
| 152 |
+
self, instances_pool: List[Dict[str, object]]
|
| 153 |
+
) -> List[Dict[str, object]]:
|
| 154 |
if self.labels is None:
|
| 155 |
self.labels = self.divide_by_repr(instances_pool)
|
| 156 |
all_labels = list(self.labels.keys())
|
| 157 |
+
get_random().shuffle(all_labels)
|
| 158 |
from collections import Counter
|
| 159 |
|
| 160 |
total_allocated = 0
|
|
|
|
| 171 |
|
| 172 |
result = []
|
| 173 |
for label, allocation in allocations.items():
|
| 174 |
+
sample = get_random().sample(self.labels[label], allocation)
|
| 175 |
result.extend(sample)
|
| 176 |
|
| 177 |
+
get_random().shuffle(result)
|
| 178 |
return result
|
| 179 |
|
| 180 |
|
| 181 |
+
class SpreadSplit(InstanceOperatorWithMultiStreamAccess):
|
| 182 |
source_stream: str = None
|
| 183 |
target_field: str = None
|
| 184 |
sampler: Sampler = None
|
| 185 |
|
| 186 |
def prepare(self):
|
|
|
|
|
|
|
| 187 |
self.local_cache = None
|
| 188 |
+
self.sampler.prepare()
|
| 189 |
|
| 190 |
def verify(self):
|
| 191 |
assert self.source_stream is not None, "Source stream must be specified"
|
|
|
|
| 193 |
assert self.sampler is not None, "Sampler must be specified"
|
| 194 |
return super().verify()
|
| 195 |
|
| 196 |
+
def process(
|
| 197 |
+
self, instance: Dict[str, object], multi_stream: MultiStream
|
| 198 |
+
) -> Dict[str, object]:
|
| 199 |
+
try:
|
| 200 |
+
if self.local_cache is None:
|
| 201 |
+
self.local_cache = list(multi_stream[self.source_stream])
|
| 202 |
+
|
| 203 |
+
source_stream = self.local_cache
|
| 204 |
+
|
| 205 |
+
sampled_instances = self.sampler.sample(source_stream)
|
| 206 |
+
instance[self.target_field] = sampled_instances
|
| 207 |
+
return instance
|
| 208 |
+
except Exception as e:
|
| 209 |
+
raise Exception(
|
| 210 |
+
f"Unable to fetch instances from '{self.source_stream}' to '{self.target_field}'"
|
| 211 |
+
) from e
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|