File size: 9,161 Bytes
f5776d3 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 |
import dsp
import tqdm
import random
import threading
import dspy
from dspy.predict.retry import Retry
from dspy.primitives import Example
from .teleprompt import Teleprompter
from .vanilla import LabeledFewShot
from dspy.evaluate.evaluate import Evaluate
# TODO: metrics should return an object with __bool__ basically, but fine if they're more complex.
# They can also be sortable.
# TODO: Switch here from dsp.Example to dspy.Example. Right now, it's okay because it's internal only (predictors).
# NOTE: Notice the places where we don't shuffle examples. I do like that this one doesn't shuffle.
# Other ones that consider options may want to use both unshuffled and then shuffle a few times, when considering candidates.
# TODO: the max_rounds via branch_idx to get past the cache, not just temperature.
# In principle, we can also sample multiple outputs from the final generation step
# (or even each step, in case the validation function just wants *one* thing that works, but nah)
# and try them all. Having a pretty solid guess on the "final step" of each example isn't hard by the second round,
# in the sense that we have the trace from the first round. (Yes it may change but that's an edge case that
# won't hurt our "best effort" guarantees.)
# TODO: When this bootstraps for another teleprompter like finetune, we want all demos we gather.
# But when it's for direct use we may want to sample ONE demo per predictor--example pair. This is important for "multi-use" modules.
# TODO: Add baselines=[...]
class BootstrapFewShot(Teleprompter):
def __init__(self, metric=None, teacher_settings={}, max_bootstrapped_demos=4, max_labeled_demos=16, max_rounds=1, max_errors=5):
self.metric = metric
self.teacher_settings = teacher_settings
self.max_bootstrapped_demos = max_bootstrapped_demos
self.max_labeled_demos = max_labeled_demos
self.max_rounds = max_rounds
self.max_errors= max_errors
self.error_count = 0
self.error_lock = threading.Lock()
def compile(self, student, *, teacher=None, trainset, valset=None):
self.trainset = trainset
self.valset = valset
self._prepare_student_and_teacher(student, teacher)
self._prepare_predictor_mappings()
self._bootstrap()
self.student = self._train()
self.student._compiled = True
# set assert_failures and suggest_failures as attributes of student w/ value 0
setattr(self.student, '_assert_failures', 0)
setattr(self.student, '_suggest_failures', 0)
return self.student
def _prepare_student_and_teacher(self, student, teacher):
self.student = student.reset_copy()
self.teacher = teacher.deepcopy() if teacher is not None else student.reset_copy()
assert getattr(self.student, '_compiled', False) is False, "Student must be uncompiled."
if self.max_labeled_demos and getattr(self.teacher, '_compiled', False) is False:
teleprompter = LabeledFewShot(k=self.max_labeled_demos)
self.teacher = teleprompter.compile(self.teacher.reset_copy(), trainset=self.trainset)
def _prepare_predictor_mappings(self):
name2predictor, predictor2name = {}, {}
student, teacher = self.student, self.teacher
assert len(student.predictors()) == len(teacher.predictors()), "Student and teacher must have the same number of predictors."
for (name1, predictor1), (name2, predictor2) in zip(student.named_predictors(), teacher.named_predictors()):
assert name1 == name2, "Student and teacher must have the same program structure."
assert predictor1.signature == predictor2.signature, f"Student and teacher must have the same signatures. {type(predictor1.signature)} != {type(predictor2.signature)}"
assert id(predictor1) != id(predictor2), "Student and teacher must be different objects."
name2predictor[name1] = None # dict(student=predictor1, teacher=predictor2)
predictor2name[id(predictor1)] = name1
# FIXME(shangyint): This is an ugly hack to bind traces of
# retry.module to retry
# if isinstance(predictor1, Retry):
# predictor2name[id(predictor1.module)] = name1
predictor2name[id(predictor2)] = name2
self.name2predictor = name2predictor
self.predictor2name = predictor2name
def _bootstrap(self, *, max_bootstraps=None):
max_bootstraps = max_bootstraps or self.max_bootstrapped_demos
bootstrapped = {}
self.name2traces = {name: [] for name in self.name2predictor}
for round_idx in range(self.max_rounds):
for example_idx, example in enumerate(tqdm.tqdm(self.trainset)):
if len(bootstrapped) >= max_bootstraps:
break
if example_idx not in bootstrapped:
success = self._bootstrap_one_example(example, round_idx)
if success:
bootstrapped[example_idx] = True
print(f'Bootstrapped {len(bootstrapped)} full traces after {example_idx+1} examples in round {round_idx}.')
# Unbootstrapped training examples
self.validation = [x for idx, x in enumerate(self.trainset) if idx not in bootstrapped]
random.Random(0).shuffle(self.validation)
self.validation = self.valset or self.validation
# NOTE: Can't yet use evaluate because we need to trace *per example*
# evaluate = Evaluate(program=self.teacher, metric=self.metric, num_threads=12)
# score = evaluate(self.metric, display_table=False, display_progress=True)
def _bootstrap_one_example(self, example, round_idx=0):
name2traces = self.name2traces
teacher = self.teacher #.deepcopy()
predictor_cache = {}
try:
with dsp.settings.context(trace=[], **self.teacher_settings):
lm = dsp.settings.lm
lm = lm.copy(temperature=0.7 + 0.001 * round_idx) if round_idx > 0 else lm
new_settings = dict(lm=lm) if round_idx > 0 else {}
with dsp.settings.context(**new_settings):
for name, predictor in teacher.named_predictors():
predictor_cache[name] = predictor.demos
predictor.demos = [x for x in predictor.demos if x != example]
prediction = teacher(**example.inputs())
trace = dsp.settings.trace
for name, predictor in teacher.named_predictors():
predictor.demos = predictor_cache[name]
success = (self.metric is None) or self.metric(example, prediction, trace)
# print(success, example, prediction)
except Exception as e:
success = False
with self.error_lock:
self.error_count += 1
current_error_count = self.error_count
if current_error_count >= self.max_errors:
raise e
print(f'Failed to run or to evaluate example {example} with {self.metric} due to {e}.')
if success:
for step in trace:
predictor, inputs, outputs = step
if 'dspy_uuid' in example:
demo = Example(augmented=True, dspy_uuid=example.dspy_uuid, **inputs, **outputs)
else:
# TODO: FIXME: This is a hack. RandomSearch will complain for now in this edge case.
demo = Example(augmented=True, **inputs, **outputs)
try:
predictor_name = self.predictor2name[id(predictor)]
except KeyError as e:
continue # FIXME: !
# TODO: Look closer into this. It's a bit tricky to reproduce.
print(f'Failed to find predictor {predictor} in {self.predictor2name}.')
print('Are you doing this in a notebook (Jupyter)? This might be caused by redefining values by rerunning cells.')
print('Try restarting the notebook, or open an issue.')
raise KeyError(f'Failed to find predictor {id(predictor)} {predictor} in {self.predictor2name}.') from e
name2traces[predictor_name].append(demo)
return success
def _train(self):
rng = random.Random(0)
raw_demos = self.validation
for name, predictor in self.student.named_predictors():
augmented_demos = self.name2traces[name][:self.max_bootstrapped_demos]
sample_size = min(self.max_labeled_demos - len(augmented_demos), len(raw_demos))
sample_size = max(0, sample_size)
raw_demos = rng.sample(raw_demos, sample_size)
import dspy
if dspy.settings.release >= 20230928:
predictor.demos = raw_demos + augmented_demos
else:
predictor.demos = augmented_demos + raw_demos
return self.student
|