Upload standard.py with huggingface_hub
Browse files- standard.py +81 -17
standard.py
CHANGED
@@ -1,11 +1,12 @@
|
|
|
|
1 |
from typing import List
|
2 |
|
3 |
from .card import TaskCard
|
4 |
from .dataclass import InternalField, OptionalField
|
5 |
from .formats import ICLFormat
|
6 |
from .instructions import Instruction
|
7 |
-
from .operator import
|
8 |
-
from .operators import StreamRefiner
|
9 |
from .recipe import Recipe
|
10 |
from .renderers import StandardRenderer
|
11 |
from .schema import ToUnitxtGroup
|
@@ -13,21 +14,30 @@ from .splitters import Sampler, SeparateSplit, SpreadSplit
|
|
13 |
from .templates import Template
|
14 |
|
15 |
|
16 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
17 |
card: TaskCard
|
18 |
template: Template = None
|
19 |
instruction: Instruction = None
|
20 |
format: ICLFormat = ICLFormat()
|
21 |
|
|
|
|
|
22 |
max_train_instances: int = None
|
23 |
max_validation_instances: int = None
|
24 |
max_test_instances: int = None
|
25 |
|
26 |
-
train_refiner: StreamRefiner = OptionalField(default_factory=
|
27 |
-
validation_refiner: StreamRefiner = OptionalField(
|
28 |
-
|
29 |
-
)
|
30 |
-
test_refiner: StreamRefiner = OptionalField(default_factory=lambda: StreamRefiner(apply_to_streams=["test"]))
|
31 |
|
32 |
demos_pool_size: int = None
|
33 |
num_demos: int = 0
|
@@ -37,6 +47,8 @@ class BaseRecipe(Recipe, SourceSequntialOperator):
|
|
37 |
demos_field: str = "demos"
|
38 |
sampler: Sampler = None
|
39 |
|
|
|
|
|
40 |
steps: List[StreamingOperator] = InternalField(default_factory=list)
|
41 |
|
42 |
def verify(self):
|
@@ -48,7 +60,31 @@ class BaseRecipe(Recipe, SourceSequntialOperator):
|
|
48 |
)
|
49 |
if self.demos_pool_size < self.num_demos:
|
50 |
raise ValueError(
|
51 |
-
f"demos_pool_size must be bigger than num_demos
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
52 |
)
|
53 |
|
54 |
def prepare(self):
|
@@ -56,14 +92,23 @@ class BaseRecipe(Recipe, SourceSequntialOperator):
|
|
56 |
self.card.loader,
|
57 |
]
|
58 |
|
|
|
|
|
|
|
|
|
|
|
59 |
if self.card.preprocess_steps is not None:
|
60 |
self.steps.extend(self.card.preprocess_steps)
|
61 |
|
62 |
self.steps.append(self.card.task)
|
63 |
|
|
|
|
|
|
|
|
|
64 |
if self.demos_pool_size is not None:
|
65 |
self.steps.append(
|
66 |
-
|
67 |
from_split=self.demos_taken_from,
|
68 |
to_split_names=[self.demos_pool_name, self.demos_taken_from],
|
69 |
to_split_sizes=[int(self.demos_pool_size)],
|
@@ -79,7 +124,7 @@ class BaseRecipe(Recipe, SourceSequntialOperator):
|
|
79 |
sampler.set_size(self.num_demos)
|
80 |
|
81 |
self.steps.append(
|
82 |
-
|
83 |
source_stream=self.demos_pool_name,
|
84 |
target_field=self.demos_field,
|
85 |
sampler=sampler,
|
@@ -87,12 +132,15 @@ class BaseRecipe(Recipe, SourceSequntialOperator):
|
|
87 |
)
|
88 |
|
89 |
self.train_refiner.max_instances = self.max_train_instances
|
|
|
90 |
self.steps.append(self.train_refiner)
|
91 |
|
92 |
self.validation_refiner.max_instances = self.max_validation_instances
|
|
|
93 |
self.steps.append(self.validation_refiner)
|
94 |
|
95 |
self.test_refiner.max_instances = self.max_test_instances
|
|
|
96 |
self.steps.append(self.test_refiner)
|
97 |
|
98 |
render = StandardRenderer(
|
@@ -104,6 +152,9 @@ class BaseRecipe(Recipe, SourceSequntialOperator):
|
|
104 |
|
105 |
self.steps.append(render)
|
106 |
|
|
|
|
|
|
|
107 |
postprocessors = render.get_postprocessors()
|
108 |
|
109 |
self.steps.append(
|
@@ -122,10 +173,21 @@ class StandardRecipeWithIndexes(BaseRecipe):
|
|
122 |
def prepare(self):
|
123 |
assert (
|
124 |
self.template_card_index is None or self.template is None
|
125 |
-
), "Specify either template or template_card_index"
|
|
|
|
|
|
|
126 |
if self.template_card_index is not None:
|
127 |
-
|
128 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
129 |
assert (
|
130 |
self.instruction_card_index is None or self.instruction is None
|
131 |
), "Specify either instruction or instruction_card_index"
|
@@ -136,9 +198,9 @@ class StandardRecipeWithIndexes(BaseRecipe):
|
|
136 |
|
137 |
|
138 |
class StandardRecipe(StandardRecipeWithIndexes):
|
139 |
-
"""
|
140 |
-
|
141 |
-
This class can be used to prepare a recipe
|
142 |
with all necessary steps, refiners and renderers included. It allows to set various
|
143 |
parameters and steps in a sequential manner for preparing the recipe.
|
144 |
|
@@ -146,6 +208,7 @@ class StandardRecipe(StandardRecipeWithIndexes):
|
|
146 |
card (TaskCard): TaskCard object associated with the recipe.
|
147 |
template (Template, optional): Template object to be used for the recipe.
|
148 |
instruction (Instruction, optional): Instruction object to be used for the recipe.
|
|
|
149 |
format (ICLFormat, optional): ICLFormat object to be used for the recipe.
|
150 |
train_refiner (StreamRefiner, optional): Train refiner to be used in the recipe.
|
151 |
max_train_instances (int, optional): Maximum training instances for the refiner.
|
@@ -160,6 +223,7 @@ class StandardRecipe(StandardRecipeWithIndexes):
|
|
160 |
demos_field (str, optional): Field name for demos. Default is "demos".
|
161 |
sampler (Sampler, optional): Sampler object to be used in the recipe.
|
162 |
steps (List[StreamingOperator], optional): List of StreamingOperator objects to be used in the recipe.
|
|
|
163 |
instruction_card_index (int, optional): Index of instruction card to be used
|
164 |
for preparing the recipe.
|
165 |
template_card_index (int, optional): Index of template card to be used for
|
|
|
1 |
+
import logging
|
2 |
from typing import List
|
3 |
|
4 |
from .card import TaskCard
|
5 |
from .dataclass import InternalField, OptionalField
|
6 |
from .formats import ICLFormat
|
7 |
from .instructions import Instruction
|
8 |
+
from .operator import SourceSequentialOperator, StreamingOperator
|
9 |
+
from .operators import Augmentor, NullAugmentor, StreamRefiner
|
10 |
from .recipe import Recipe
|
11 |
from .renderers import StandardRenderer
|
12 |
from .schema import ToUnitxtGroup
|
|
|
14 |
from .templates import Template
|
15 |
|
16 |
|
17 |
+
# Used to give meaningful name to recipe steps
|
18 |
+
class CreateDemosPool(SeparateSplit):
|
19 |
+
pass
|
20 |
+
|
21 |
+
|
22 |
+
class AddDemosField(SpreadSplit):
|
23 |
+
pass
|
24 |
+
|
25 |
+
|
26 |
+
class BaseRecipe(Recipe, SourceSequentialOperator):
|
27 |
card: TaskCard
|
28 |
template: Template = None
|
29 |
instruction: Instruction = None
|
30 |
format: ICLFormat = ICLFormat()
|
31 |
|
32 |
+
loader_limit: int = None
|
33 |
+
|
34 |
max_train_instances: int = None
|
35 |
max_validation_instances: int = None
|
36 |
max_test_instances: int = None
|
37 |
|
38 |
+
train_refiner: StreamRefiner = OptionalField(default_factory=StreamRefiner)
|
39 |
+
validation_refiner: StreamRefiner = OptionalField(default_factory=StreamRefiner)
|
40 |
+
test_refiner: StreamRefiner = OptionalField(default_factory=StreamRefiner)
|
|
|
|
|
41 |
|
42 |
demos_pool_size: int = None
|
43 |
num_demos: int = 0
|
|
|
47 |
demos_field: str = "demos"
|
48 |
sampler: Sampler = None
|
49 |
|
50 |
+
augmentor: Augmentor = OptionalField(default_factory=NullAugmentor)
|
51 |
+
|
52 |
steps: List[StreamingOperator] = InternalField(default_factory=list)
|
53 |
|
54 |
def verify(self):
|
|
|
60 |
)
|
61 |
if self.demos_pool_size < self.num_demos:
|
62 |
raise ValueError(
|
63 |
+
f"demos_pool_size must be bigger than num_demos ({self.num_demos}), Got demos_pool_size={self.demos_pool_size}"
|
64 |
+
)
|
65 |
+
if self.loader_limit and self.demos_pool_size > self.loader_limit:
|
66 |
+
raise ValueError(
|
67 |
+
f"demos_pool_size must be bigger than loader_limit ({self.loader_limit}), Got demos_pool_size={self.demos_pool_size}"
|
68 |
+
)
|
69 |
+
|
70 |
+
if self.loader_limit:
|
71 |
+
if self.max_test_instances and self.max_test_instances > self.loader_limit:
|
72 |
+
raise ValueError(
|
73 |
+
f"max_test_instances must be bigger than loader_limit ({self.loader_limit}), Got max_test_instances={self.max_test_instances}"
|
74 |
+
)
|
75 |
+
if (
|
76 |
+
self.max_validation_instances
|
77 |
+
and self.max_validation_instances > self.loader_limit
|
78 |
+
):
|
79 |
+
raise ValueError(
|
80 |
+
f"max_validation_instances must be bigger than loader_limit ({self.loader_limit}), Got max_validation_instances={self.max_validation_instances}"
|
81 |
+
)
|
82 |
+
if (
|
83 |
+
self.max_train_instances
|
84 |
+
and self.max_train_instances > self.loader_limit
|
85 |
+
):
|
86 |
+
raise ValueError(
|
87 |
+
f"max_train_instances must be bigger than loader_limit ({self.loader_limit}), Got max_train_instances={self.max_train_instances}"
|
88 |
)
|
89 |
|
90 |
def prepare(self):
|
|
|
92 |
self.card.loader,
|
93 |
]
|
94 |
|
95 |
+
if self.loader_limit:
|
96 |
+
self.card.loader.loader_limit = self.loader_limit
|
97 |
+
logging.info(f"Loader line limit was set to {self.loader_limit}")
|
98 |
+
self.steps.append(StreamRefiner(max_instances=self.loader_limit))
|
99 |
+
|
100 |
if self.card.preprocess_steps is not None:
|
101 |
self.steps.extend(self.card.preprocess_steps)
|
102 |
|
103 |
self.steps.append(self.card.task)
|
104 |
|
105 |
+
if self.augmentor.augment_task_input:
|
106 |
+
self.augmentor.set_task_input_fields(self.card.task.augmentable_inputs)
|
107 |
+
self.steps.append(self.augmentor)
|
108 |
+
|
109 |
if self.demos_pool_size is not None:
|
110 |
self.steps.append(
|
111 |
+
CreateDemosPool(
|
112 |
from_split=self.demos_taken_from,
|
113 |
to_split_names=[self.demos_pool_name, self.demos_taken_from],
|
114 |
to_split_sizes=[int(self.demos_pool_size)],
|
|
|
124 |
sampler.set_size(self.num_demos)
|
125 |
|
126 |
self.steps.append(
|
127 |
+
AddDemosField(
|
128 |
source_stream=self.demos_pool_name,
|
129 |
target_field=self.demos_field,
|
130 |
sampler=sampler,
|
|
|
132 |
)
|
133 |
|
134 |
self.train_refiner.max_instances = self.max_train_instances
|
135 |
+
self.train_refiner.apply_to_streams = ["train"]
|
136 |
self.steps.append(self.train_refiner)
|
137 |
|
138 |
self.validation_refiner.max_instances = self.max_validation_instances
|
139 |
+
self.validation_refiner.apply_to_streams = ["validation"]
|
140 |
self.steps.append(self.validation_refiner)
|
141 |
|
142 |
self.test_refiner.max_instances = self.max_test_instances
|
143 |
+
self.test_refiner.apply_to_streams = ["test"]
|
144 |
self.steps.append(self.test_refiner)
|
145 |
|
146 |
render = StandardRenderer(
|
|
|
152 |
|
153 |
self.steps.append(render)
|
154 |
|
155 |
+
if self.augmentor.augment_model_input:
|
156 |
+
self.steps.append(self.augmentor)
|
157 |
+
|
158 |
postprocessors = render.get_postprocessors()
|
159 |
|
160 |
self.steps.append(
|
|
|
173 |
def prepare(self):
|
174 |
assert (
|
175 |
self.template_card_index is None or self.template is None
|
176 |
+
), f"Specify either template ({self.template}) or template_card_index ({self.template_card_index}) but not both"
|
177 |
+
assert not (
|
178 |
+
self.template_card_index is None and self.template is None
|
179 |
+
), "Specify either template or template_card_index in card"
|
180 |
if self.template_card_index is not None:
|
181 |
+
try:
|
182 |
+
self.template = self.card.templates[self.template_card_index]
|
183 |
+
except Exception as e:
|
184 |
+
if isinstance(self.card.templates, dict):
|
185 |
+
options = self.card.templates.keys()
|
186 |
+
else:
|
187 |
+
options = list(range(0, len(self.card.templates)))
|
188 |
+
raise ValueError(
|
189 |
+
f"card_template_index '{self.template_card_index}' is not in card. Available options: {options}"
|
190 |
+
) from e
|
191 |
assert (
|
192 |
self.instruction_card_index is None or self.instruction is None
|
193 |
), "Specify either instruction or instruction_card_index"
|
|
|
198 |
|
199 |
|
200 |
class StandardRecipe(StandardRecipeWithIndexes):
|
201 |
+
"""This class represents a standard recipe for data processing and preperation.
|
202 |
+
|
203 |
+
This class can be used to prepare a recipe.
|
204 |
with all necessary steps, refiners and renderers included. It allows to set various
|
205 |
parameters and steps in a sequential manner for preparing the recipe.
|
206 |
|
|
|
208 |
card (TaskCard): TaskCard object associated with the recipe.
|
209 |
template (Template, optional): Template object to be used for the recipe.
|
210 |
instruction (Instruction, optional): Instruction object to be used for the recipe.
|
211 |
+
loader_limit (int, optional): Specifies the maximum number of instances per stream to be returned from the loader (used to reduce loading time in large datasets)
|
212 |
format (ICLFormat, optional): ICLFormat object to be used for the recipe.
|
213 |
train_refiner (StreamRefiner, optional): Train refiner to be used in the recipe.
|
214 |
max_train_instances (int, optional): Maximum training instances for the refiner.
|
|
|
223 |
demos_field (str, optional): Field name for demos. Default is "demos".
|
224 |
sampler (Sampler, optional): Sampler object to be used in the recipe.
|
225 |
steps (List[StreamingOperator], optional): List of StreamingOperator objects to be used in the recipe.
|
226 |
+
augmentor (Augmentor) : Augmentor to be used to pseudo randomly augment the source text
|
227 |
instruction_card_index (int, optional): Index of instruction card to be used
|
228 |
for preparing the recipe.
|
229 |
template_card_index (int, optional): Index of template card to be used for
|