-
Notifications
You must be signed in to change notification settings - Fork 0
/
recipe.py
377 lines (317 loc) · 15.1 KB
/
recipe.py
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
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
from cook import create_task, Task
import json
from pathlib import Path
import shutil
from summaries.scripts.infer_posterior import INFERENCE_CONFIGS
from summaries.scripts.train_transformer import TRAIN_CONFIGS
from summaries.util import load_pickle
from typing import Dict
create_task("requirements", action="pip-compile -v", targets=["requirements.txt"],
dependencies=["requirements.in", "setup.py"])
create_task("tests", action="pytest -v --cov=summaries --cov-report=html --cov-report=term-missing "
"--cov-fail-under=100")
ROOT = Path("workspace")
BENCHMARK_ROOT = ROOT / "benchmark"
COALESCENT_ROOT = ROOT / "coalescent"
TREE_ROOT = ROOT / "tree"
# Random number generator seeds generated, at some point, by np.random.randint(10_000).
SEEDS = [6389, 9074, 7627]
def prepare_coalescent_data() -> Dict[str, Path]:
"""
Download, extract, and preprocess the coalescent dataset.
"""
data_root = COALESCENT_ROOT / "data"
url = "https://github.com/tillahoffmann/coaloracle/releases/download/0.2/csv.zip"
archive = data_root / "coal.zip"
create_task("coalescent:download", action=f"curl -Lo {archive} {url}", targets=[archive])
coaloracle = data_root / "coaloracle.csv"
create_task("coalescent:extract", dependencies=[archive], targets=[coaloracle],
action=f"unzip -ojd {data_root} {archive}")
# Preprocess the dataset by splitting it train, test, and validation sets.
splits = {"test": 1_000, "validation": 10_000, "train": 989_000}
split_targets = {split: data_root / f"{split}.pkl" for split in splits}
split_args = ' '.join(f"{split}:{size}" for split, size in splits.items())
action = f"python -m summaries.scripts.preprocess_coalescent --seed={21} {coaloracle} " \
f"{data_root} {split_args}"
create_task("coalescent:preprocess", dependencies=[coaloracle], targets=split_targets.values(),
action=action)
return split_targets
def _pick_best_transformer(task: Task) -> None:
# Find the best one and copy it; a symlink would be nice but that breaks all sorts of stuff,
# e.g., evaluating digests.
best = min(task.dependencies, key=lambda path: load_pickle(path)["last_validation_loss"])
shutil.copy(best, task.targets[0])
def train_transformer(experiment: str, splits: Dict[str, Path], config: str) -> Path:
"""
Train a single transformer multiple times using different seeds. We create a symlink to the
"best" transformer as evaluated by the last validation loss of the training run.
Args:
experiment: Parent folder of the experiment.
splits: Datasets for training and validation.
config: Training configuration supported by `summaries.scripts.train_transformer`.
Returns:
Path to trained transformer.
"""
dependencies = [splits["train"], splits["validation"]]
name = f"{experiment}:train:{config}"
# Train a bunch of transformers with different seeds.
transformer_targets = []
for seed in SEEDS:
transformer_target = ROOT / f"{experiment}/transformers/{config}-{seed}.pkl"
action = ["python", "-m", "summaries.scripts.train_transformer", f"--seed={seed}", config,
*dependencies, transformer_target]
create_task(f"{name}-{seed}", dependencies=dependencies, action=action,
targets=[transformer_target])
transformer_targets.append(transformer_target)
# Pick the best one using the validation loss.
transformer_target = ROOT / f"{experiment}/transformers/{config}.pkl"
create_task(name, action=_pick_best_transformer, targets=[transformer_target],
dependencies=transformer_targets)
return transformer_target
def train_coalescent_transformers(splits: Dict[str, Path]) -> Dict[str, Path]:
"""
Train transformers for the coalescent dataset.
"""
return {config: train_transformer("coalescent", splits, config) for config in TRAIN_CONFIGS
if config.startswith("Coalescent")}
def infer_posterior(experiment: str, splits: Dict[str, Path], config: str,
transformer: Path | None = None, suffix: str | None = None) -> Path:
"""
Draw posterior samples using approximate Bayesian computation.
Args:
experiment: Parent folder of the experiment.
splits: Datasets for training and validation.
config: Training configuration supported by `summaries.scripts.infer_posterior`.
transformer: Optional path to pickled transformer.
suffix: Suffix for file comprising samples (defaults to transformer name if available).
Returns:
Path to file comprising posterior samples.
"""
dependencies = [splits["train"], splits["test"]]
if transformer:
dependencies.append(transformer)
kwargs = {"transformer": str(transformer)}
suffix = suffix or transformer.with_suffix('').name
else:
kwargs = {}
name = f"{config}-{suffix}" if suffix else config
posterior_target = ROOT / f"{experiment}/samples/{name}.pkl"
action = [
"python", "-m", "summaries.scripts.infer_posterior", "--transformer-kwargs",
json.dumps(kwargs), config, *dependencies[:2], posterior_target,
]
create_task(f"{experiment}:infer:{name}", dependencies=dependencies, targets=[posterior_target],
action=action)
return posterior_target
def infer_mdn_posterior(experiment: str, splits: Dict[str, Path], transformer: Path,
loader: str | None = None) -> Path:
"""
Infer posterior samples by sampling from a mixture density network.
Args:
experiment: Parent folder of the experiment.
splits: Datasets for training and validation.
transformer: Path to pickled mixture density network.
loader: Name of the data loader (required for loading graph data).
"""
dependencies = [transformer, splits["test"]]
name = f"{experiment}:infer:{transformer.with_suffix('').name}"
target = ROOT / f"{experiment}/samples/mdn-{transformer.name}"
action = ["python", "-m", "summaries.scripts.infer_mdn", transformer, splits["test"], target]
if loader:
action.append(f"--loader={loader}")
create_task(name, dependencies=dependencies, targets=[target], action=action)
return target
def create_coalescent_tasks() -> Dict[str, Path]:
"""
Create all tasks for the coalescent experiment.
Returns:
Map from names to paths comprising samples.
"""
splits = prepare_coalescent_data()
transformers = train_coalescent_transformers(splits)
sample_targets = {
f"CoalescentNeuralConfig-{config}":
infer_posterior("coalescent", splits, "CoalescentNeuralConfig", transformer) for
config, transformer in transformers.items()
} | {
config: infer_posterior("coalescent", splits, config) for config in INFERENCE_CONFIGS if
config.startswith("Coalescent") and config != "CoalescentNeuralConfig"
} | {
"CoalescentMixtureDensityConfig": infer_mdn_posterior(
"coalescent", splits, transformers["CoalescentMixtureDensityConfig"]
),
"PriorConfig": infer_posterior("coalescent", splits, "PriorConfig")
}
return {
"samples": sample_targets,
"experiment": "coalescent",
"transformers": transformers,
}
def simulate_tree_data(experiment: str, n_observations: int) -> Dict[str, Path]:
"""
Simulate data from a growing tree.
Args:
experiment: Parent folder of the experiment.
n_observations: Number of nodes per graph.
Returns:
Dataset splits.
"""
data_root = ROOT / experiment / "data"
split_paths = {}
splits = {"train": (100_000, 0), "validation": (1_000, 1), "test": (1_000, 2), "debug": (10, 3)}
for split, (n_samples, seed) in splits.items():
target = data_root / f"{split}.pkl"
action = [
"python", "-m", "summaries.scripts.simulate_data", f"--n-samples={n_samples}",
f"--seed={seed}", f"--n-observations={n_observations}", "TreeSimulationConfig", target
]
create_task(f"{experiment}:data:{split}", targets=[target], action=action)
split_paths[split] = target
return split_paths
def infer_tree_posterior_with_history_sampler(experiment: str, splits: Dict[str, Path]) -> Path:
"""
Infer tree posteriors using the history sampler of https://github.com/gstonge/fasttr.
"""
config = "TreeKernelHistorySamplerConfig"
posterior_target = ROOT / f"{experiment}/samples/{config}.pkl"
action = [
"python", "-m", "summaries.scripts.infer_tree_posterior", "--n-samples=1000",
splits["test"], posterior_target,
]
create_task(f"{experiment}:infer:{config}", dependencies=[splits["test"]],
targets=[posterior_target], action=action)
return posterior_target
def train_tree_transformers(experiment: str, splits: Dict[str, Path]) -> Dict[str, Path]:
"""
Train transformers for the tree dataset.
"""
targets = {}
for config in TRAIN_CONFIGS:
if not config.startswith("Tree"):
continue
targets[config] = train_transformer(experiment, splits, config)
return targets
def compute_tree_summaries(experiment: str, splits: Dict[str, Path]) -> Dict[str, Path]:
"""
Precompute summary statistics for inferring the kernel parameter of trees.
"""
data_root = ROOT / experiment / "data"
summary_paths = {}
for split, path in splits.items():
target = data_root / f"{split}-summaries.pkl"
action = ["python", "-m", "summaries.scripts.compute_tree_summaries", path, target]
create_task(f"{experiment}:data:{split}-summaries", targets=[target], action=action)
summary_paths[split] = target
return summary_paths
def create_tree_tasks(experiment: str, n_observations: int) -> None:
splits = simulate_tree_data(experiment, n_observations)
summaries_splits = compute_tree_summaries(experiment, splits)
transformers = train_tree_transformers(experiment, splits)
samples = {
f"TreeKernelNeuralConfig-{config}":
infer_posterior(experiment, splits, "TreeKernelNeuralConfig", transformer) for
config, transformer in transformers.items()
} | {
"TreeKernelHistorySamplerConfig":
infer_tree_posterior_with_history_sampler(experiment, splits),
"PriorConfig": infer_posterior(experiment, splits, "PriorConfig"),
"TreeMixtureDensityConfig": infer_mdn_posterior(
experiment, splits, transformers["TreeMixtureDensityConfig"], loader="tree",
),
} | {
config: infer_posterior(experiment, summaries_splits, config) for config in
INFERENCE_CONFIGS if config.startswith("Tree") and config != "TreeKernelNeuralConfig"
}
return {
"splits": splits,
"transformers": transformers,
"samples": samples,
"experiment": experiment,
}
def simulate_benchmark_data(experiment: str, n_observations: int) -> Dict[str, Path]:
data_root = ROOT / experiment / "data"
split_paths = {}
splits = {"train": (1_000_000, 0), "validation": (10_000, 1), "test": (1_000, 2),
"debug": (10, 3)}
for split, (n_samples, seed) in splits.items():
target = data_root / f"{split}.pkl"
action = [
"python", "-m", "summaries.scripts.simulate_data", f"--n-samples={n_samples}",
f"--seed={seed}", f"--n-observations={n_observations}", "BenchmarkSimulationConfig",
target
]
create_task(f"{experiment}:data:{split}", targets=[target], action=action)
split_paths[split] = target
return split_paths
def train_benchmark_transformers(experiment: str, splits: Dict[str, Path]) -> Dict[str, Path]:
"""
Train transformers for the benchmark dataset.
"""
targets = {}
for config in TRAIN_CONFIGS:
if not config.startswith("Benchmark"):
continue
targets[config] = train_transformer(experiment, splits, config)
return targets
def create_benchmark_tasks(experiment: str, n_observations: int) -> None:
splits = simulate_benchmark_data(experiment, n_observations)
transformers = train_benchmark_transformers(experiment, splits)
samples = {
f"BenchmarkNeuralConfig-{config}":
infer_posterior(experiment, splits, "BenchmarkNeuralConfig", transformer) for
config, transformer in transformers.items()
} | {
config: infer_posterior(experiment, splits, config) for config in
INFERENCE_CONFIGS if config.startswith("Benchmark") and config != "BenchmarkNeuralConfig"
} | {
"PriorConfig": infer_posterior(experiment, splits, "PriorConfig"),
"BenchmarkMixtureDensityConfig": infer_mdn_posterior(
experiment, splits, transformers["BenchmarkMixtureDensityConfig"]
),
"BenchmarkMixtureDensityConfigReduced": infer_mdn_posterior(
experiment, splits, transformers["BenchmarkMixtureDensityConfigReduced"]
),
}
# Add the Stan likelihood-based sampler.
target = ROOT / experiment / "samples" / "BenchmarkStanConfig.pkl"
action = ["python", "-m", "summaries.scripts.infer_benchmark", splits["test"], target]
create_task(
f"{experiment}:infer:stan", dependencies=[splits["test"]], targets=[target],
action=action
)
samples["BenchmarkStanConfig"] = target
return {
"splits": splits,
"transformers": transformers,
"samples": samples,
"experiment": experiment,
}
coalescent_tasks = create_coalescent_tasks()
benchmark_tasks_small = create_benchmark_tasks("benchmark-small", 10)
benchmark_tasks_large = create_benchmark_tasks("benchmark-large", 100)
tree_tasks_small = create_tree_tasks("tree-small", 100)
tree_tasks_large = create_tree_tasks("tree-large", 748)
# Create the transfer learning tasks: using the networks trained on smaller datasets for inference
# on the larger ones.
tree_tasks_large["samples"]["TreeKernelNeuralConfig-TreeMixtureDensityConfig-small"] = \
infer_posterior("tree-large", tree_tasks_large["splits"], "TreeKernelNeuralConfig",
tree_tasks_small["transformers"]["TreeMixtureDensityConfig"],
suffix="TreeMixtureDensityConfig-small")
benchmark_tasks_large["samples"]["BenchmarkNeuralConfig-BenchmarkMixtureDensityConfig-small"] = \
infer_posterior("benchmark-large", benchmark_tasks_large["splits"], "BenchmarkNeuralConfig",
benchmark_tasks_small["transformers"]["BenchmarkMixtureDensityConfig"],
suffix="BenchmarkMixtureDensityConfig-small")
# Add evaluation for each batch of tasks.
for tasks in [coalescent_tasks, benchmark_tasks_large, benchmark_tasks_small, tree_tasks_large,
tree_tasks_small]:
experiment: str = tasks["experiment"]
target = ROOT / experiment / "evaluation.csv"
paths = list(tasks["samples"].values())
bounds = "none"
if experiment.startswith("tree"):
bounds = "0,2"
elif experiment == "coalescent":
bounds = "0,10,0,10"
action = f"python -m summaries.scripts.evaluate --bounds={bounds} --csv={target} " \
+ " ".join(map(str, paths))
create_task(f"{experiment}:evaluation", targets=[target], action=action, dependencies=paths)