Spaces:
Running
Running
deepsource-autofix[bot]
commited on
Format code with black
Browse filesThis commit fixes the style issues introduced in 2b07f83 according to the output
from black.
Details: https://deepsource.io/gh/MilesCranmer/PySR/transform/e437d614-031f-4537-a160-7682b887fb55/
- benchmarks/hyperparamopt.py +72 -65
- example.py +15 -8
- pysr/export_jax.py +8 -4
- pysr/export_torch.py +32 -16
- pysr/feynman_problems.py +68 -35
- pysr/sr.py +508 -326
- setup.py +3 -9
- test/test.py +75 -54
- test/test_jax.py +25 -14
- test/test_torch.py +26 -15
benchmarks/hyperparamopt.py
CHANGED
|
@@ -10,6 +10,7 @@ import time
|
|
| 10 |
import contextlib
|
| 11 |
import numpy as np
|
| 12 |
|
|
|
|
| 13 |
@contextlib.contextmanager
|
| 14 |
def temp_seed(seed):
|
| 15 |
state = np.random.get_state()
|
|
@@ -20,11 +21,12 @@ def temp_seed(seed):
|
|
| 20 |
np.random.set_state(state)
|
| 21 |
|
| 22 |
|
| 23 |
-
#Change the following code to your file
|
| 24 |
################################################################################
|
| 25 |
-
TRIALS_FOLDER =
|
| 26 |
NUMBER_TRIALS_PER_RUN = 1
|
| 27 |
|
|
|
|
| 28 |
def run_trial(args):
|
| 29 |
"""Evaluate the model loss using the hyperparams in args
|
| 30 |
|
|
@@ -34,29 +36,29 @@ def run_trial(args):
|
|
| 34 |
"""
|
| 35 |
|
| 36 |
print("Running on", args)
|
| 37 |
-
args[
|
| 38 |
-
args[
|
| 39 |
-
args[
|
| 40 |
-
args[
|
| 41 |
-
args[
|
| 42 |
-
args[
|
| 43 |
-
args[
|
| 44 |
-
|
| 45 |
-
if args[
|
| 46 |
print("Bad parameters")
|
| 47 |
-
return {
|
| 48 |
|
| 49 |
-
args[
|
| 50 |
ntrials = 3
|
| 51 |
|
| 52 |
with temp_seed(0):
|
| 53 |
-
X = np.random.randn(100, 10)*3
|
| 54 |
|
| 55 |
eval_str = [
|
| 56 |
-
|
| 57 |
-
|
| 58 |
-
|
| 59 |
-
|
| 60 |
]
|
| 61 |
|
| 62 |
print(f"Starting", str(args))
|
|
@@ -67,51 +69,50 @@ def run_trial(args):
|
|
| 67 |
for j in range(ntrials):
|
| 68 |
print(f"Starting trial {j}")
|
| 69 |
y = eval(eval_str[i])
|
| 70 |
-
trial = pysr.pysr(
|
|
|
|
|
|
|
| 71 |
procs=4,
|
| 72 |
populations=20,
|
| 73 |
binary_operators=["plus", "mult", "pow", "div"],
|
| 74 |
unary_operators=["cos", "exp", "sin", "logm", "abs"],
|
| 75 |
maxsize=25,
|
| 76 |
-
constraints={
|
| 77 |
-
**args
|
| 78 |
-
|
|
|
|
|
|
|
| 79 |
trials.append(
|
| 80 |
-
|
| 81 |
)
|
| 82 |
print(f"Test {i} trial {j} with", str(args), f"got {trials[-1]}")
|
| 83 |
|
| 84 |
except ValueError:
|
| 85 |
print(f"Broken", str(args))
|
| 86 |
-
return {
|
| 87 |
-
'status': 'ok', # or 'fail' if nan loss
|
| 88 |
-
'loss': np.inf
|
| 89 |
-
}
|
| 90 |
loss = np.average(trials)
|
| 91 |
print(f"Finished with {loss}", str(args))
|
| 92 |
|
| 93 |
-
return {
|
| 94 |
-
'status': 'ok', # or 'fail' if nan loss
|
| 95 |
-
'loss': loss
|
| 96 |
-
}
|
| 97 |
|
| 98 |
|
| 99 |
space = {
|
| 100 |
-
|
| 101 |
-
|
| 102 |
-
|
| 103 |
-
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
|
| 107 |
-
|
| 108 |
-
|
| 109 |
-
|
| 110 |
-
|
| 111 |
}
|
| 112 |
|
| 113 |
################################################################################
|
| 114 |
|
|
|
|
| 115 |
def merge_trials(trials1, trials2_slice):
|
| 116 |
"""Merge two hyperopt trials objects
|
| 117 |
|
|
@@ -123,24 +124,23 @@ def merge_trials(trials1, trials2_slice):
|
|
| 123 |
"""
|
| 124 |
max_tid = 0
|
| 125 |
if len(trials1.trials) > 0:
|
| 126 |
-
max_tid = max([trial[
|
| 127 |
|
| 128 |
for trial in trials2_slice:
|
| 129 |
-
tid = trial[
|
| 130 |
hyperopt_trial = Trials().new_trial_docs(
|
| 131 |
-
|
| 132 |
-
|
| 133 |
-
results=[None],
|
| 134 |
-
miscs=[None])
|
| 135 |
hyperopt_trial[0] = trial
|
| 136 |
-
hyperopt_trial[0][
|
| 137 |
-
hyperopt_trial[0][
|
| 138 |
-
for key in hyperopt_trial[0][
|
| 139 |
-
hyperopt_trial[0][
|
| 140 |
-
trials1.insert_trial_docs(hyperopt_trial)
|
| 141 |
trials1.refresh()
|
| 142 |
return trials1
|
| 143 |
|
|
|
|
| 144 |
loaded_fnames = []
|
| 145 |
trials = None
|
| 146 |
# Run new hyperparameter trials until killed
|
|
@@ -149,15 +149,16 @@ while True:
|
|
| 149 |
|
| 150 |
# Load up all runs:
|
| 151 |
import glob
|
| 152 |
-
|
|
|
|
| 153 |
for fname in glob.glob(path):
|
| 154 |
if fname in loaded_fnames:
|
| 155 |
continue
|
| 156 |
|
| 157 |
-
trials_obj = pkl.load(open(fname,
|
| 158 |
-
n_trials = trials_obj[
|
| 159 |
-
trials_obj = trials_obj[
|
| 160 |
-
if len(loaded_fnames) == 0:
|
| 161 |
trials = trials_obj
|
| 162 |
else:
|
| 163 |
print("Merging trials")
|
|
@@ -171,23 +172,29 @@ while True:
|
|
| 171 |
|
| 172 |
n = NUMBER_TRIALS_PER_RUN
|
| 173 |
try:
|
| 174 |
-
best = fmin(
|
|
|
|
| 175 |
space=space,
|
| 176 |
algo=tpe.suggest,
|
| 177 |
max_evals=n + len(trials.trials),
|
| 178 |
trials=trials,
|
| 179 |
verbose=1,
|
| 180 |
-
rstate=np.random.RandomState(np.random.randint(1,10**6))
|
| 181 |
-
|
| 182 |
except hyperopt.exceptions.AllTrialsFailed:
|
| 183 |
continue
|
| 184 |
|
| 185 |
-
print(
|
| 186 |
hyperopt_trial = Trials()
|
| 187 |
|
| 188 |
# Merge with empty trials dataset:
|
| 189 |
save_trials = merge_trials(hyperopt_trial, trials.trials[-n:])
|
| 190 |
-
new_fname =
|
| 191 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 192 |
loaded_fnames.append(new_fname)
|
| 193 |
-
|
|
|
|
| 10 |
import contextlib
|
| 11 |
import numpy as np
|
| 12 |
|
| 13 |
+
|
| 14 |
@contextlib.contextmanager
|
| 15 |
def temp_seed(seed):
|
| 16 |
state = np.random.get_state()
|
|
|
|
| 21 |
np.random.set_state(state)
|
| 22 |
|
| 23 |
|
| 24 |
+
# Change the following code to your file
|
| 25 |
################################################################################
|
| 26 |
+
TRIALS_FOLDER = "trials"
|
| 27 |
NUMBER_TRIALS_PER_RUN = 1
|
| 28 |
|
| 29 |
+
|
| 30 |
def run_trial(args):
|
| 31 |
"""Evaluate the model loss using the hyperparams in args
|
| 32 |
|
|
|
|
| 36 |
"""
|
| 37 |
|
| 38 |
print("Running on", args)
|
| 39 |
+
args["niterations"] = 100
|
| 40 |
+
args["npop"] = 100
|
| 41 |
+
args["ncyclesperiteration"] = 1000
|
| 42 |
+
args["topn"] = 10
|
| 43 |
+
args["parsimony"] = 0.0
|
| 44 |
+
args["useFrequency"] = True
|
| 45 |
+
args["annealing"] = True
|
| 46 |
+
|
| 47 |
+
if args["npop"] < 20 or args["ncyclesperiteration"] < 3:
|
| 48 |
print("Bad parameters")
|
| 49 |
+
return {"status": "ok", "loss": np.inf}
|
| 50 |
|
| 51 |
+
args["weightDoNothing"] = 1.0
|
| 52 |
ntrials = 3
|
| 53 |
|
| 54 |
with temp_seed(0):
|
| 55 |
+
X = np.random.randn(100, 10) * 3
|
| 56 |
|
| 57 |
eval_str = [
|
| 58 |
+
"np.sign(X[:, 2])*np.abs(X[:, 2])**2.5 + 5*np.cos(X[:, 3]) - 5",
|
| 59 |
+
"np.exp(X[:, 0]/2) + 12.0 + np.log(np.abs(X[:, 0])*10 + 1)",
|
| 60 |
+
"(np.exp(X[:, 3]) + 3)/(np.abs(X[:, 1]) + np.cos(X[:, 0]) + 1.1)",
|
| 61 |
+
"X[:, 0] * np.sin(2*np.pi * (X[:, 1] * X[:, 2] - X[:, 3] / X[:, 4])) + 3.0",
|
| 62 |
]
|
| 63 |
|
| 64 |
print(f"Starting", str(args))
|
|
|
|
| 69 |
for j in range(ntrials):
|
| 70 |
print(f"Starting trial {j}")
|
| 71 |
y = eval(eval_str[i])
|
| 72 |
+
trial = pysr.pysr(
|
| 73 |
+
X,
|
| 74 |
+
y,
|
| 75 |
procs=4,
|
| 76 |
populations=20,
|
| 77 |
binary_operators=["plus", "mult", "pow", "div"],
|
| 78 |
unary_operators=["cos", "exp", "sin", "logm", "abs"],
|
| 79 |
maxsize=25,
|
| 80 |
+
constraints={"pow": (-1, 1)},
|
| 81 |
+
**args,
|
| 82 |
+
)
|
| 83 |
+
if len(trial) == 0:
|
| 84 |
+
raise ValueError
|
| 85 |
trials.append(
|
| 86 |
+
np.min(trial["MSE"]) ** 0.5 / np.std(eval(eval_str[i - 1]))
|
| 87 |
)
|
| 88 |
print(f"Test {i} trial {j} with", str(args), f"got {trials[-1]}")
|
| 89 |
|
| 90 |
except ValueError:
|
| 91 |
print(f"Broken", str(args))
|
| 92 |
+
return {"status": "ok", "loss": np.inf} # or 'fail' if nan loss
|
|
|
|
|
|
|
|
|
|
| 93 |
loss = np.average(trials)
|
| 94 |
print(f"Finished with {loss}", str(args))
|
| 95 |
|
| 96 |
+
return {"status": "ok", "loss": loss} # or 'fail' if nan loss
|
|
|
|
|
|
|
|
|
|
| 97 |
|
| 98 |
|
| 99 |
space = {
|
| 100 |
+
"alpha": hp.lognormal("alpha", np.log(10.0), 1.0),
|
| 101 |
+
"fractionReplacedHof": hp.lognormal("fractionReplacedHof", np.log(0.1), 1.0),
|
| 102 |
+
"fractionReplaced": hp.lognormal("fractionReplaced", np.log(0.1), 1.0),
|
| 103 |
+
"perturbationFactor": hp.lognormal("perturbationFactor", np.log(1.0), 1.0),
|
| 104 |
+
"weightMutateConstant": hp.lognormal("weightMutateConstant", np.log(4.0), 1.0),
|
| 105 |
+
"weightMutateOperator": hp.lognormal("weightMutateOperator", np.log(0.5), 1.0),
|
| 106 |
+
"weightAddNode": hp.lognormal("weightAddNode", np.log(0.5), 1.0),
|
| 107 |
+
"weightInsertNode": hp.lognormal("weightInsertNode", np.log(0.5), 1.0),
|
| 108 |
+
"weightDeleteNode": hp.lognormal("weightDeleteNode", np.log(0.5), 1.0),
|
| 109 |
+
"weightSimplify": hp.lognormal("weightSimplify", np.log(0.05), 1.0),
|
| 110 |
+
"weightRandomize": hp.lognormal("weightRandomize", np.log(0.25), 1.0),
|
| 111 |
}
|
| 112 |
|
| 113 |
################################################################################
|
| 114 |
|
| 115 |
+
|
| 116 |
def merge_trials(trials1, trials2_slice):
|
| 117 |
"""Merge two hyperopt trials objects
|
| 118 |
|
|
|
|
| 124 |
"""
|
| 125 |
max_tid = 0
|
| 126 |
if len(trials1.trials) > 0:
|
| 127 |
+
max_tid = max([trial["tid"] for trial in trials1.trials])
|
| 128 |
|
| 129 |
for trial in trials2_slice:
|
| 130 |
+
tid = trial["tid"] + max_tid + 1
|
| 131 |
hyperopt_trial = Trials().new_trial_docs(
|
| 132 |
+
tids=[None], specs=[None], results=[None], miscs=[None]
|
| 133 |
+
)
|
|
|
|
|
|
|
| 134 |
hyperopt_trial[0] = trial
|
| 135 |
+
hyperopt_trial[0]["tid"] = tid
|
| 136 |
+
hyperopt_trial[0]["misc"]["tid"] = tid
|
| 137 |
+
for key in hyperopt_trial[0]["misc"]["idxs"].keys():
|
| 138 |
+
hyperopt_trial[0]["misc"]["idxs"][key] = [tid]
|
| 139 |
+
trials1.insert_trial_docs(hyperopt_trial)
|
| 140 |
trials1.refresh()
|
| 141 |
return trials1
|
| 142 |
|
| 143 |
+
|
| 144 |
loaded_fnames = []
|
| 145 |
trials = None
|
| 146 |
# Run new hyperparameter trials until killed
|
|
|
|
| 149 |
|
| 150 |
# Load up all runs:
|
| 151 |
import glob
|
| 152 |
+
|
| 153 |
+
path = TRIALS_FOLDER + "/*.pkl"
|
| 154 |
for fname in glob.glob(path):
|
| 155 |
if fname in loaded_fnames:
|
| 156 |
continue
|
| 157 |
|
| 158 |
+
trials_obj = pkl.load(open(fname, "rb"))
|
| 159 |
+
n_trials = trials_obj["n"]
|
| 160 |
+
trials_obj = trials_obj["trials"]
|
| 161 |
+
if len(loaded_fnames) == 0:
|
| 162 |
trials = trials_obj
|
| 163 |
else:
|
| 164 |
print("Merging trials")
|
|
|
|
| 172 |
|
| 173 |
n = NUMBER_TRIALS_PER_RUN
|
| 174 |
try:
|
| 175 |
+
best = fmin(
|
| 176 |
+
run_trial,
|
| 177 |
space=space,
|
| 178 |
algo=tpe.suggest,
|
| 179 |
max_evals=n + len(trials.trials),
|
| 180 |
trials=trials,
|
| 181 |
verbose=1,
|
| 182 |
+
rstate=np.random.RandomState(np.random.randint(1, 10 ** 6)),
|
| 183 |
+
)
|
| 184 |
except hyperopt.exceptions.AllTrialsFailed:
|
| 185 |
continue
|
| 186 |
|
| 187 |
+
print("current best", best)
|
| 188 |
hyperopt_trial = Trials()
|
| 189 |
|
| 190 |
# Merge with empty trials dataset:
|
| 191 |
save_trials = merge_trials(hyperopt_trial, trials.trials[-n:])
|
| 192 |
+
new_fname = (
|
| 193 |
+
TRIALS_FOLDER
|
| 194 |
+
+ "/"
|
| 195 |
+
+ str(np.random.randint(0, sys.maxsize))
|
| 196 |
+
+ str(time.time())
|
| 197 |
+
+ ".pkl"
|
| 198 |
+
)
|
| 199 |
+
pkl.dump({"trials": save_trials, "n": n}, open(new_fname, "wb"))
|
| 200 |
loaded_fnames.append(new_fname)
|
|
|
example.py
CHANGED
|
@@ -2,18 +2,25 @@ import numpy as np
|
|
| 2 |
from pysr import pysr, best
|
| 3 |
|
| 4 |
# Dataset
|
| 5 |
-
X = 2*np.random.randn(100, 5)
|
| 6 |
-
y = 2*np.cos(X[:, 3]) + X[:, 0]**2 - 2
|
| 7 |
|
| 8 |
# Learn equations
|
| 9 |
-
equations = pysr(
|
|
|
|
|
|
|
|
|
|
| 10 |
binary_operators=["plus", "mult"],
|
| 11 |
unary_operators=[
|
| 12 |
-
|
| 13 |
-
|
| 14 |
-
|
| 15 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 16 |
|
| 17 |
-
|
| 18 |
|
| 19 |
print(best(equations))
|
|
|
|
| 2 |
from pysr import pysr, best
|
| 3 |
|
| 4 |
# Dataset
|
| 5 |
+
X = 2 * np.random.randn(100, 5)
|
| 6 |
+
y = 2 * np.cos(X[:, 3]) + X[:, 0] ** 2 - 2
|
| 7 |
|
| 8 |
# Learn equations
|
| 9 |
+
equations = pysr(
|
| 10 |
+
X,
|
| 11 |
+
y,
|
| 12 |
+
niterations=5,
|
| 13 |
binary_operators=["plus", "mult"],
|
| 14 |
unary_operators=[
|
| 15 |
+
"cos",
|
| 16 |
+
"exp",
|
| 17 |
+
"sin", # Pre-defined library of operators (see https://pysr.readthedocs.io/en/latest/docs/operators/)
|
| 18 |
+
"inv(x) = 1/x",
|
| 19 |
+
],
|
| 20 |
+
loss="loss(x, y) = abs(x - y)", # Custom loss function
|
| 21 |
+
julia_project="../SymbolicRegression.jl",
|
| 22 |
+
) # Define your own operator! (Julia syntax)
|
| 23 |
|
| 24 |
+
... # (you can use ctl-c to exit early)
|
| 25 |
|
| 26 |
print(best(equations))
|
pysr/export_jax.py
CHANGED
|
@@ -58,14 +58,16 @@ def sympy2jaxtext(expr, parameters, symbols_in):
|
|
| 58 |
elif issubclass(expr.func, sympy.Integer):
|
| 59 |
return f"{int(expr)}"
|
| 60 |
elif issubclass(expr.func, sympy.Symbol):
|
| 61 |
-
return
|
|
|
|
|
|
|
| 62 |
else:
|
| 63 |
_func = _jnp_func_lookup[expr.func]
|
| 64 |
args = [sympy2jaxtext(arg, parameters, symbols_in) for arg in expr.args]
|
| 65 |
if _func == MUL:
|
| 66 |
-
return
|
| 67 |
elif _func == ADD:
|
| 68 |
-
return
|
| 69 |
else:
|
| 70 |
return f'{_func}({", ".join(args)})'
|
| 71 |
|
|
@@ -75,6 +77,7 @@ jax = None
|
|
| 75 |
jnp = None
|
| 76 |
jsp = None
|
| 77 |
|
|
|
|
| 78 |
def _initialize_jax():
|
| 79 |
global jax_initialized
|
| 80 |
global jax
|
|
@@ -85,6 +88,7 @@ def _initialize_jax():
|
|
| 85 |
import jax as _jax
|
| 86 |
from jax import numpy as _jnp
|
| 87 |
from jax.scipy import special as _jsp
|
|
|
|
| 88 |
jax = _jax
|
| 89 |
jnp = _jnp
|
| 90 |
jsp = _jsp
|
|
@@ -169,7 +173,7 @@ def sympy2jax(expression, symbols_in, selection=None):
|
|
| 169 |
|
| 170 |
parameters = []
|
| 171 |
functional_form_text = sympy2jaxtext(expression, parameters, symbols_in)
|
| 172 |
-
hash_string =
|
| 173 |
text = f"def {hash_string}(X, parameters):\n"
|
| 174 |
if selection is not None:
|
| 175 |
# Impose the feature selection:
|
|
|
|
| 58 |
elif issubclass(expr.func, sympy.Integer):
|
| 59 |
return f"{int(expr)}"
|
| 60 |
elif issubclass(expr.func, sympy.Symbol):
|
| 61 |
+
return (
|
| 62 |
+
f"X[:, {[i for i in range(len(symbols_in)) if symbols_in[i] == expr][0]}]"
|
| 63 |
+
)
|
| 64 |
else:
|
| 65 |
_func = _jnp_func_lookup[expr.func]
|
| 66 |
args = [sympy2jaxtext(arg, parameters, symbols_in) for arg in expr.args]
|
| 67 |
if _func == MUL:
|
| 68 |
+
return " * ".join(["(" + arg + ")" for arg in args])
|
| 69 |
elif _func == ADD:
|
| 70 |
+
return " + ".join(["(" + arg + ")" for arg in args])
|
| 71 |
else:
|
| 72 |
return f'{_func}({", ".join(args)})'
|
| 73 |
|
|
|
|
| 77 |
jnp = None
|
| 78 |
jsp = None
|
| 79 |
|
| 80 |
+
|
| 81 |
def _initialize_jax():
|
| 82 |
global jax_initialized
|
| 83 |
global jax
|
|
|
|
| 88 |
import jax as _jax
|
| 89 |
from jax import numpy as _jnp
|
| 90 |
from jax.scipy import special as _jsp
|
| 91 |
+
|
| 92 |
jax = _jax
|
| 93 |
jnp = _jnp
|
| 94 |
jsp = _jsp
|
|
|
|
| 173 |
|
| 174 |
parameters = []
|
| 175 |
functional_form_text = sympy2jaxtext(expression, parameters, symbols_in)
|
| 176 |
+
hash_string = "A_" + str(abs(hash(str(expression) + str(symbols_in))))
|
| 177 |
text = f"def {hash_string}(X, parameters):\n"
|
| 178 |
if selection is not None:
|
| 179 |
# Impose the feature selection:
|
pysr/export_torch.py
CHANGED
|
@@ -7,17 +7,21 @@ import collections as co
|
|
| 7 |
import functools as ft
|
| 8 |
import sympy
|
| 9 |
|
|
|
|
| 10 |
def _reduce(fn):
|
| 11 |
def fn_(*args):
|
| 12 |
return ft.reduce(fn, args)
|
|
|
|
| 13 |
return fn_
|
| 14 |
|
|
|
|
| 15 |
torch_initialized = False
|
| 16 |
torch = None
|
| 17 |
_global_func_lookup = None
|
| 18 |
_Node = None
|
| 19 |
SingleSymPyModule = None
|
| 20 |
|
|
|
|
| 21 |
def _initialize_torch():
|
| 22 |
global torch_initialized
|
| 23 |
global torch
|
|
@@ -29,6 +33,7 @@ def _initialize_torch():
|
|
| 29 |
# but still allow this module to be loaded in __init__
|
| 30 |
if not torch_initialized:
|
| 31 |
import torch as _torch
|
|
|
|
| 32 |
torch = _torch
|
| 33 |
|
| 34 |
_global_func_lookup = {
|
|
@@ -85,6 +90,7 @@ def _initialize_torch():
|
|
| 85 |
|
| 86 |
class _Node(torch.nn.Module):
|
| 87 |
"""SympyTorch code from https://github.com/patrick-kidger/sympytorch"""
|
|
|
|
| 88 |
def __init__(self, *, expr, _memodict, _func_lookup, **kwargs):
|
| 89 |
super().__init__(**kwargs)
|
| 90 |
|
|
@@ -95,9 +101,13 @@ def _initialize_torch():
|
|
| 95 |
self._torch_func = lambda: self._value
|
| 96 |
self._args = ()
|
| 97 |
elif issubclass(expr.func, sympy.UnevaluatedExpr):
|
| 98 |
-
if len(expr.args) != 1 or not issubclass(
|
| 99 |
-
|
| 100 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 101 |
self._torch_func = lambda: self._value
|
| 102 |
self._args = ()
|
| 103 |
elif issubclass(expr.func, sympy.Integer):
|
|
@@ -117,7 +127,12 @@ def _initialize_torch():
|
|
| 117 |
try:
|
| 118 |
arg_ = _memodict[arg]
|
| 119 |
except KeyError:
|
| 120 |
-
arg_ = type(self)(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 121 |
_memodict[arg] = arg_
|
| 122 |
args.append(arg_)
|
| 123 |
self._args = torch.nn.ModuleList(args)
|
|
@@ -133,19 +148,22 @@ def _initialize_torch():
|
|
| 133 |
args.append(arg_)
|
| 134 |
return self._torch_func(*args)
|
| 135 |
|
| 136 |
-
|
| 137 |
class SingleSymPyModule(torch.nn.Module):
|
| 138 |
"""SympyTorch code from https://github.com/patrick-kidger/sympytorch"""
|
| 139 |
-
|
| 140 |
-
|
|
|
|
|
|
|
| 141 |
super().__init__(**kwargs)
|
| 142 |
-
|
| 143 |
if extra_funcs is None:
|
| 144 |
extra_funcs = {}
|
| 145 |
_func_lookup = co.ChainMap(_global_func_lookup, extra_funcs)
|
| 146 |
|
| 147 |
_memodict = {}
|
| 148 |
-
self._node = _Node(
|
|
|
|
|
|
|
| 149 |
self._expression_string = str(expression)
|
| 150 |
self._selection = selection
|
| 151 |
self.symbols_in = [str(symbol) for symbol in symbols_in]
|
|
@@ -156,13 +174,11 @@ def _initialize_torch():
|
|
| 156 |
def forward(self, X):
|
| 157 |
if self._selection is not None:
|
| 158 |
X = X[:, self._selection]
|
| 159 |
-
symbols = {symbol: X[:, i]
|
| 160 |
-
for i, symbol in enumerate(self.symbols_in)}
|
| 161 |
return self._node(symbols)
|
| 162 |
|
| 163 |
|
| 164 |
-
def sympy2torch(expression, symbols_in,
|
| 165 |
-
selection=None, extra_torch_mappings=None):
|
| 166 |
"""Returns a module for a given sympy expression with trainable parameters;
|
| 167 |
|
| 168 |
This function will assume the input to the module is a matrix X, where
|
|
@@ -172,6 +188,6 @@ def sympy2torch(expression, symbols_in,
|
|
| 172 |
|
| 173 |
_initialize_torch()
|
| 174 |
|
| 175 |
-
return SingleSymPyModule(
|
| 176 |
-
|
| 177 |
-
|
|
|
|
| 7 |
import functools as ft
|
| 8 |
import sympy
|
| 9 |
|
| 10 |
+
|
| 11 |
def _reduce(fn):
|
| 12 |
def fn_(*args):
|
| 13 |
return ft.reduce(fn, args)
|
| 14 |
+
|
| 15 |
return fn_
|
| 16 |
|
| 17 |
+
|
| 18 |
torch_initialized = False
|
| 19 |
torch = None
|
| 20 |
_global_func_lookup = None
|
| 21 |
_Node = None
|
| 22 |
SingleSymPyModule = None
|
| 23 |
|
| 24 |
+
|
| 25 |
def _initialize_torch():
|
| 26 |
global torch_initialized
|
| 27 |
global torch
|
|
|
|
| 33 |
# but still allow this module to be loaded in __init__
|
| 34 |
if not torch_initialized:
|
| 35 |
import torch as _torch
|
| 36 |
+
|
| 37 |
torch = _torch
|
| 38 |
|
| 39 |
_global_func_lookup = {
|
|
|
|
| 90 |
|
| 91 |
class _Node(torch.nn.Module):
|
| 92 |
"""SympyTorch code from https://github.com/patrick-kidger/sympytorch"""
|
| 93 |
+
|
| 94 |
def __init__(self, *, expr, _memodict, _func_lookup, **kwargs):
|
| 95 |
super().__init__(**kwargs)
|
| 96 |
|
|
|
|
| 101 |
self._torch_func = lambda: self._value
|
| 102 |
self._args = ()
|
| 103 |
elif issubclass(expr.func, sympy.UnevaluatedExpr):
|
| 104 |
+
if len(expr.args) != 1 or not issubclass(
|
| 105 |
+
expr.args[0].func, sympy.Float
|
| 106 |
+
):
|
| 107 |
+
raise ValueError(
|
| 108 |
+
"UnevaluatedExpr should only be used to wrap floats."
|
| 109 |
+
)
|
| 110 |
+
self.register_buffer("_value", torch.tensor(float(expr.args[0])))
|
| 111 |
self._torch_func = lambda: self._value
|
| 112 |
self._args = ()
|
| 113 |
elif issubclass(expr.func, sympy.Integer):
|
|
|
|
| 127 |
try:
|
| 128 |
arg_ = _memodict[arg]
|
| 129 |
except KeyError:
|
| 130 |
+
arg_ = type(self)(
|
| 131 |
+
expr=arg,
|
| 132 |
+
_memodict=_memodict,
|
| 133 |
+
_func_lookup=_func_lookup,
|
| 134 |
+
**kwargs,
|
| 135 |
+
)
|
| 136 |
_memodict[arg] = arg_
|
| 137 |
args.append(arg_)
|
| 138 |
self._args = torch.nn.ModuleList(args)
|
|
|
|
| 148 |
args.append(arg_)
|
| 149 |
return self._torch_func(*args)
|
| 150 |
|
|
|
|
| 151 |
class SingleSymPyModule(torch.nn.Module):
|
| 152 |
"""SympyTorch code from https://github.com/patrick-kidger/sympytorch"""
|
| 153 |
+
|
| 154 |
+
def __init__(
|
| 155 |
+
self, expression, symbols_in, selection=None, extra_funcs=None, **kwargs
|
| 156 |
+
):
|
| 157 |
super().__init__(**kwargs)
|
| 158 |
+
|
| 159 |
if extra_funcs is None:
|
| 160 |
extra_funcs = {}
|
| 161 |
_func_lookup = co.ChainMap(_global_func_lookup, extra_funcs)
|
| 162 |
|
| 163 |
_memodict = {}
|
| 164 |
+
self._node = _Node(
|
| 165 |
+
expr=expression, _memodict=_memodict, _func_lookup=_func_lookup
|
| 166 |
+
)
|
| 167 |
self._expression_string = str(expression)
|
| 168 |
self._selection = selection
|
| 169 |
self.symbols_in = [str(symbol) for symbol in symbols_in]
|
|
|
|
| 174 |
def forward(self, X):
|
| 175 |
if self._selection is not None:
|
| 176 |
X = X[:, self._selection]
|
| 177 |
+
symbols = {symbol: X[:, i] for i, symbol in enumerate(self.symbols_in)}
|
|
|
|
| 178 |
return self._node(symbols)
|
| 179 |
|
| 180 |
|
| 181 |
+
def sympy2torch(expression, symbols_in, selection=None, extra_torch_mappings=None):
|
|
|
|
| 182 |
"""Returns a module for a given sympy expression with trainable parameters;
|
| 183 |
|
| 184 |
This function will assume the input to the module is a matrix X, where
|
|
|
|
| 188 |
|
| 189 |
_initialize_torch()
|
| 190 |
|
| 191 |
+
return SingleSymPyModule(
|
| 192 |
+
expression, symbols_in, selection=selection, extra_funcs=extra_torch_mappings
|
| 193 |
+
)
|
pysr/feynman_problems.py
CHANGED
|
@@ -7,6 +7,7 @@ from pathlib import Path
|
|
| 7 |
PKG_DIR = Path(__file__).parents[1]
|
| 8 |
FEYNMAN_DATASET = PKG_DIR / "datasets" / "FeynmanEquations.csv"
|
| 9 |
|
|
|
|
| 10 |
class Problem:
|
| 11 |
"""
|
| 12 |
Problem API to work with PySR.
|
|
@@ -15,6 +16,7 @@ class Problem:
|
|
| 15 |
|
| 16 |
Should be able to call pysr(problem.X, problem.y, var_names=problem.var_names) and have it work
|
| 17 |
"""
|
|
|
|
| 18 |
def __init__(self, X, y, form=None, variable_names=None):
|
| 19 |
self.X = X
|
| 20 |
self.y = y
|
|
@@ -27,34 +29,39 @@ class FeynmanProblem(Problem):
|
|
| 27 |
Stores the data for the problems from the 100 Feynman Equations on Physics.
|
| 28 |
This is the benchmark used in the AI Feynman Paper
|
| 29 |
"""
|
|
|
|
| 30 |
def __init__(self, row, gen=False, dp=500):
|
| 31 |
"""
|
| 32 |
row: a row read as a dict from the FeynmanEquations dataset provided in the datasets folder of the repo
|
| 33 |
gen: If true the problem will have dp X and y values randomly generated else they will be None
|
| 34 |
"""
|
| 35 |
-
self.eq_id
|
| 36 |
-
self.n_vars
|
| 37 |
-
super(FeynmanProblem, self).__init__(
|
| 38 |
-
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 42 |
if gen:
|
| 43 |
self.X = np.random.uniform(0.01, 25, size=(self.dp, self.n_vars))
|
| 44 |
d = {}
|
| 45 |
for var in range(len(self.variable_names)):
|
| 46 |
d[self.variable_names[var]] = self.X[:, var]
|
| 47 |
-
d[
|
| 48 |
-
d[
|
| 49 |
-
d[
|
| 50 |
-
d[
|
| 51 |
-
d[
|
| 52 |
-
d[
|
| 53 |
-
d[
|
| 54 |
-
d[
|
| 55 |
-
d[
|
| 56 |
-
d[
|
| 57 |
-
self.y = eval(self.form,d)
|
| 58 |
return
|
| 59 |
|
| 60 |
def __str__(self):
|
|
@@ -77,7 +84,8 @@ class FeynmanProblem(Problem):
|
|
| 77 |
for i, row in enumerate(reader):
|
| 78 |
if ind > first:
|
| 79 |
break
|
| 80 |
-
if row[
|
|
|
|
| 81 |
try:
|
| 82 |
p = FeynmanProblem(row, gen=gen, dp=dp)
|
| 83 |
ret.append(p)
|
|
@@ -93,18 +101,34 @@ def run_on_problem(problem, verbosity=0, multiprocessing=True):
|
|
| 93 |
Takes in a problem and returns a tuple: (equations, best predicted equation, actual equation)
|
| 94 |
"""
|
| 95 |
from time import time
|
|
|
|
| 96 |
starting = time()
|
| 97 |
-
equations = pysr(
|
| 98 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 99 |
others = {"time": timing, "problem": problem}
|
| 100 |
if not multiprocessing:
|
| 101 |
-
others[
|
| 102 |
return str(best(equations)), problem.form, others
|
| 103 |
|
| 104 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 105 |
import multiprocessing as mp
|
| 106 |
from tqdm import tqdm
|
| 107 |
-
|
|
|
|
|
|
|
|
|
|
| 108 |
ids = []
|
| 109 |
predictions = []
|
| 110 |
true_equations = []
|
|
@@ -117,22 +141,31 @@ def do_feynman_experiments_parallel(first=100, verbosity=0, dp=500, output_file_
|
|
| 117 |
pbar.update()
|
| 118 |
for res in results:
|
| 119 |
prediction, true_equation, others = res
|
| 120 |
-
problem = others[
|
| 121 |
ids.append(problem.eq_id)
|
| 122 |
predictions.append(prediction)
|
| 123 |
true_equations.append(true_equation)
|
| 124 |
-
time_takens.append(others[
|
| 125 |
-
with open(output_file_path,
|
| 126 |
-
writer = csv.writer(f, delimiter=
|
| 127 |
-
writer.writerow([
|
| 128 |
for i in range(len(ids)):
|
| 129 |
writer.writerow([ids[i], predictions[i], true_equations[i], time_takens[i]])
|
| 130 |
return
|
| 131 |
|
| 132 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 133 |
from tqdm import tqdm
|
| 134 |
|
| 135 |
-
problems = FeynmanProblem.mk_problems(
|
|
|
|
|
|
|
| 136 |
indx = range(len(problems))
|
| 137 |
ids = []
|
| 138 |
predictions = []
|
|
@@ -143,10 +176,10 @@ def do_feynman_experiments(first=100, verbosity=0, dp=500, output_file_path="Fey
|
|
| 143 |
ids.append(problem.eq_id)
|
| 144 |
predictions.append(prediction)
|
| 145 |
true_equations.append(true_equation)
|
| 146 |
-
time_takens.append(others[
|
| 147 |
-
with open(output_file_path,
|
| 148 |
-
writer = csv.writer(f, delimiter=
|
| 149 |
-
writer.writerow([
|
| 150 |
for i in range(len(ids)):
|
| 151 |
writer.writerow([ids[i], predictions[i], true_equations[i], time_takens[i]])
|
| 152 |
return
|
|
|
|
| 7 |
PKG_DIR = Path(__file__).parents[1]
|
| 8 |
FEYNMAN_DATASET = PKG_DIR / "datasets" / "FeynmanEquations.csv"
|
| 9 |
|
| 10 |
+
|
| 11 |
class Problem:
|
| 12 |
"""
|
| 13 |
Problem API to work with PySR.
|
|
|
|
| 16 |
|
| 17 |
Should be able to call pysr(problem.X, problem.y, var_names=problem.var_names) and have it work
|
| 18 |
"""
|
| 19 |
+
|
| 20 |
def __init__(self, X, y, form=None, variable_names=None):
|
| 21 |
self.X = X
|
| 22 |
self.y = y
|
|
|
|
| 29 |
Stores the data for the problems from the 100 Feynman Equations on Physics.
|
| 30 |
This is the benchmark used in the AI Feynman Paper
|
| 31 |
"""
|
| 32 |
+
|
| 33 |
def __init__(self, row, gen=False, dp=500):
|
| 34 |
"""
|
| 35 |
row: a row read as a dict from the FeynmanEquations dataset provided in the datasets folder of the repo
|
| 36 |
gen: If true the problem will have dp X and y values randomly generated else they will be None
|
| 37 |
"""
|
| 38 |
+
self.eq_id = row["Filename"]
|
| 39 |
+
self.n_vars = int(row["# variables"])
|
| 40 |
+
super(FeynmanProblem, self).__init__(
|
| 41 |
+
None,
|
| 42 |
+
None,
|
| 43 |
+
form=row["Formula"],
|
| 44 |
+
variable_names=[row[f"v{i + 1}_name"] for i in range(self.n_vars)],
|
| 45 |
+
)
|
| 46 |
+
self.low = [float(row[f"v{i+1}_low"]) for i in range(self.n_vars)]
|
| 47 |
+
self.high = [float(row[f"v{i+1}_high"]) for i in range(self.n_vars)]
|
| 48 |
+
self.dp = dp
|
| 49 |
if gen:
|
| 50 |
self.X = np.random.uniform(0.01, 25, size=(self.dp, self.n_vars))
|
| 51 |
d = {}
|
| 52 |
for var in range(len(self.variable_names)):
|
| 53 |
d[self.variable_names[var]] = self.X[:, var]
|
| 54 |
+
d["exp"] = np.exp
|
| 55 |
+
d["sqrt"] = np.sqrt
|
| 56 |
+
d["pi"] = np.pi
|
| 57 |
+
d["cos"] = np.cos
|
| 58 |
+
d["sin"] = np.sin
|
| 59 |
+
d["tan"] = np.tan
|
| 60 |
+
d["tanh"] = np.tanh
|
| 61 |
+
d["ln"] = np.log
|
| 62 |
+
d["log"] = np.log # Quite sure the Feynman dataset has no base 10 logs
|
| 63 |
+
d["arcsin"] = np.arcsin
|
| 64 |
+
self.y = eval(self.form, d)
|
| 65 |
return
|
| 66 |
|
| 67 |
def __str__(self):
|
|
|
|
| 84 |
for i, row in enumerate(reader):
|
| 85 |
if ind > first:
|
| 86 |
break
|
| 87 |
+
if row["Filename"] == "":
|
| 88 |
+
continue
|
| 89 |
try:
|
| 90 |
p = FeynmanProblem(row, gen=gen, dp=dp)
|
| 91 |
ret.append(p)
|
|
|
|
| 101 |
Takes in a problem and returns a tuple: (equations, best predicted equation, actual equation)
|
| 102 |
"""
|
| 103 |
from time import time
|
| 104 |
+
|
| 105 |
starting = time()
|
| 106 |
+
equations = pysr(
|
| 107 |
+
problem.X,
|
| 108 |
+
problem.y,
|
| 109 |
+
variable_names=problem.variable_names,
|
| 110 |
+
verbosity=verbosity,
|
| 111 |
+
)
|
| 112 |
+
timing = time() - starting
|
| 113 |
others = {"time": timing, "problem": problem}
|
| 114 |
if not multiprocessing:
|
| 115 |
+
others["equations"] = equations
|
| 116 |
return str(best(equations)), problem.form, others
|
| 117 |
|
| 118 |
+
|
| 119 |
+
def do_feynman_experiments_parallel(
|
| 120 |
+
first=100,
|
| 121 |
+
verbosity=0,
|
| 122 |
+
dp=500,
|
| 123 |
+
output_file_path="FeynmanExperiment.csv",
|
| 124 |
+
data_dir=FEYNMAN_DATASET,
|
| 125 |
+
):
|
| 126 |
import multiprocessing as mp
|
| 127 |
from tqdm import tqdm
|
| 128 |
+
|
| 129 |
+
problems = FeynmanProblem.mk_problems(
|
| 130 |
+
first=first, gen=True, dp=dp, data_dir=data_dir
|
| 131 |
+
)
|
| 132 |
ids = []
|
| 133 |
predictions = []
|
| 134 |
true_equations = []
|
|
|
|
| 141 |
pbar.update()
|
| 142 |
for res in results:
|
| 143 |
prediction, true_equation, others = res
|
| 144 |
+
problem = others["problem"]
|
| 145 |
ids.append(problem.eq_id)
|
| 146 |
predictions.append(prediction)
|
| 147 |
true_equations.append(true_equation)
|
| 148 |
+
time_takens.append(others["time"])
|
| 149 |
+
with open(output_file_path, "a") as f:
|
| 150 |
+
writer = csv.writer(f, delimiter=",")
|
| 151 |
+
writer.writerow(["ID", "Predicted", "True", "Time"])
|
| 152 |
for i in range(len(ids)):
|
| 153 |
writer.writerow([ids[i], predictions[i], true_equations[i], time_takens[i]])
|
| 154 |
return
|
| 155 |
|
| 156 |
+
|
| 157 |
+
def do_feynman_experiments(
|
| 158 |
+
first=100,
|
| 159 |
+
verbosity=0,
|
| 160 |
+
dp=500,
|
| 161 |
+
output_file_path="FeynmanExperiment.csv",
|
| 162 |
+
data_dir=FEYNMAN_DATASET,
|
| 163 |
+
):
|
| 164 |
from tqdm import tqdm
|
| 165 |
|
| 166 |
+
problems = FeynmanProblem.mk_problems(
|
| 167 |
+
first=first, gen=True, dp=dp, data_dir=data_dir
|
| 168 |
+
)
|
| 169 |
indx = range(len(problems))
|
| 170 |
ids = []
|
| 171 |
predictions = []
|
|
|
|
| 176 |
ids.append(problem.eq_id)
|
| 177 |
predictions.append(prediction)
|
| 178 |
true_equations.append(true_equation)
|
| 179 |
+
time_takens.append(others["time"])
|
| 180 |
+
with open(output_file_path, "a") as f:
|
| 181 |
+
writer = csv.writer(f, delimiter=",")
|
| 182 |
+
writer.writerow(["ID", "Predicted", "True", "Time"])
|
| 183 |
for i in range(len(ids)):
|
| 184 |
writer.writerow([ids[i], predictions[i], true_equations[i], time_takens[i]])
|
| 185 |
return
|
pysr/sr.py
CHANGED
|
@@ -15,7 +15,7 @@ from datetime import datetime
|
|
| 15 |
import warnings
|
| 16 |
|
| 17 |
global_state = dict(
|
| 18 |
-
equation_file=
|
| 19 |
n_features=None,
|
| 20 |
variable_names=[],
|
| 21 |
extra_sympy_mappings={},
|
|
@@ -25,108 +25,112 @@ global_state = dict(
|
|
| 25 |
output_torch_format=False,
|
| 26 |
multioutput=False,
|
| 27 |
nout=1,
|
| 28 |
-
selection=None
|
| 29 |
)
|
| 30 |
|
| 31 |
sympy_mappings = {
|
| 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 |
"""Run symbolic regression to fit f(X[i, :]) ~ y[i] for all i.
|
| 131 |
Note: most default parameters have been tuned over several example
|
| 132 |
equations, but you should adjust `niterations`,
|
|
@@ -244,7 +248,7 @@ def pysr(X, y, weights=None,
|
|
| 244 |
:type: pd.DataFrame/list
|
| 245 |
"""
|
| 246 |
if binary_operators is None:
|
| 247 |
-
binary_operators =
|
| 248 |
if unary_operators is None:
|
| 249 |
unary_operators = []
|
| 250 |
if extra_sympy_mappings is None:
|
|
@@ -255,16 +259,18 @@ def pysr(X, y, weights=None,
|
|
| 255 |
constraints = {}
|
| 256 |
|
| 257 |
if progress is not None:
|
| 258 |
-
if progress and (
|
| 259 |
-
warnings.warn(
|
|
|
|
|
|
|
| 260 |
progress = False
|
| 261 |
else:
|
| 262 |
-
if
|
| 263 |
progress = True
|
| 264 |
else:
|
| 265 |
progress = False
|
| 266 |
|
| 267 |
-
assert optimizer_algorithm in [
|
| 268 |
assert tournament_selection_n < npop
|
| 269 |
|
| 270 |
if isinstance(X, pd.DataFrame):
|
|
@@ -275,25 +281,34 @@ def pysr(X, y, weights=None,
|
|
| 275 |
X = X[:, None]
|
| 276 |
|
| 277 |
if len(variable_names) == 0:
|
| 278 |
-
variable_names = [f
|
| 279 |
-
|
| 280 |
-
use_custom_variable_names =
|
| 281 |
-
|
| 282 |
-
_check_assertions(
|
| 283 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 284 |
_check_for_julia_installation()
|
| 285 |
|
| 286 |
-
|
| 287 |
if len(X) > 10000 and not batching:
|
| 288 |
-
warnings.warn(
|
|
|
|
|
|
|
| 289 |
|
| 290 |
if maxsize > 40:
|
| 291 |
-
warnings.warn(
|
|
|
|
|
|
|
| 292 |
|
| 293 |
X, variable_names, selection = _handle_feature_selection(
|
| 294 |
-
|
| 295 |
-
|
| 296 |
-
)
|
| 297 |
|
| 298 |
if maxdepth is None:
|
| 299 |
maxdepth = maxsize
|
|
@@ -312,81 +327,102 @@ def pysr(X, y, weights=None,
|
|
| 312 |
else:
|
| 313 |
raise NotImplementedError("y shape not supported!")
|
| 314 |
|
| 315 |
-
kwargs = dict(
|
| 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 |
kwargs = {**_set_paths(tempdir), **kwargs}
|
| 358 |
|
| 359 |
if temp_equation_file:
|
| 360 |
-
equation_file = kwargs[
|
| 361 |
elif equation_file is None:
|
| 362 |
date_time = datetime.now().strftime("%Y-%m-%d_%H%M%S.%f")[:-3]
|
| 363 |
-
equation_file =
|
| 364 |
|
| 365 |
kwargs = {**dict(equation_file=equation_file), **kwargs}
|
| 366 |
|
| 367 |
-
|
| 368 |
-
pkg_directory = kwargs['pkg_directory']
|
| 369 |
manifest_file = None
|
| 370 |
-
if kwargs[
|
| 371 |
-
manifest_filepath = Path(kwargs[
|
| 372 |
else:
|
| 373 |
-
manifest_filepath = pkg_directory /
|
| 374 |
|
| 375 |
-
kwargs[
|
| 376 |
|
| 377 |
if not (manifest_filepath).is_file():
|
| 378 |
-
kwargs[
|
| 379 |
-
|
|
|
|
|
|
|
| 380 |
print("OK. I will install at launch.")
|
| 381 |
assert update
|
| 382 |
|
| 383 |
-
kwargs[
|
| 384 |
|
| 385 |
_handle_constraints(**kwargs)
|
| 386 |
|
| 387 |
-
kwargs[
|
| 388 |
-
kwargs[
|
| 389 |
-
kwargs[
|
| 390 |
|
| 391 |
_create_julia_files(**kwargs)
|
| 392 |
_final_pysr_process(**kwargs)
|
|
@@ -395,7 +431,7 @@ def pysr(X, y, weights=None,
|
|
| 395 |
equations = get_hof(**kwargs)
|
| 396 |
|
| 397 |
if delete_tempfiles:
|
| 398 |
-
shutil.rmtree(kwargs[
|
| 399 |
|
| 400 |
return equations
|
| 401 |
|
|
@@ -403,7 +439,7 @@ def pysr(X, y, weights=None,
|
|
| 403 |
def _set_globals(X, **kwargs):
|
| 404 |
global global_state
|
| 405 |
|
| 406 |
-
global_state[
|
| 407 |
for key, value in kwargs.items():
|
| 408 |
if key in global_state:
|
| 409 |
global_state[key] = value
|
|
@@ -411,34 +447,37 @@ def _set_globals(X, **kwargs):
|
|
| 411 |
|
| 412 |
def _final_pysr_process(julia_optimization, runfile_filename, timeout, **kwargs):
|
| 413 |
command = [
|
| 414 |
-
f
|
|
|
|
| 415 |
str(runfile_filename),
|
| 416 |
]
|
| 417 |
if timeout is not None:
|
| 418 |
-
command = [f
|
| 419 |
_cmd_runner(command, **kwargs)
|
| 420 |
|
|
|
|
| 421 |
def _cmd_runner(command, progress, **kwargs):
|
| 422 |
-
if kwargs[
|
| 423 |
-
print("Running on",
|
| 424 |
process = subprocess.Popen(command, stdout=subprocess.PIPE, bufsize=-1)
|
| 425 |
try:
|
| 426 |
while True:
|
| 427 |
line = process.stdout.readline()
|
| 428 |
-
if not line:
|
| 429 |
-
|
|
|
|
| 430 |
if progress:
|
| 431 |
-
decoded_line = (
|
| 432 |
-
|
| 433 |
-
|
| 434 |
-
|
| 435 |
-
|
| 436 |
-
|
| 437 |
-
|
| 438 |
sys.stdout.buffer.write(decoded_line)
|
| 439 |
sys.stdout.flush()
|
| 440 |
else:
|
| 441 |
-
print(decoded_line, end=
|
| 442 |
|
| 443 |
process.stdout.close()
|
| 444 |
process.wait()
|
|
@@ -446,62 +485,94 @@ def _cmd_runner(command, progress, **kwargs):
|
|
| 446 |
print("Killing process... will return when done.")
|
| 447 |
process.kill()
|
| 448 |
|
| 449 |
-
|
| 450 |
-
|
| 451 |
-
|
| 452 |
-
|
| 453 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 454 |
print(def_hyperparams, file=f)
|
| 455 |
-
with open(dataset_filename,
|
| 456 |
print(def_datasets, file=f)
|
| 457 |
-
with open(runfile_filename,
|
| 458 |
if julia_project is None:
|
| 459 |
julia_project = pkg_directory
|
| 460 |
else:
|
| 461 |
julia_project = Path(julia_project)
|
| 462 |
-
print(f
|
| 463 |
print(f'Pkg.activate("{_escape_filename(julia_project)}")', file=f)
|
| 464 |
if need_install:
|
| 465 |
-
print(f
|
| 466 |
-
print(f
|
| 467 |
-
print(f
|
| 468 |
elif update:
|
| 469 |
-
print(f
|
| 470 |
-
print(f
|
| 471 |
print(f'include("{_escape_filename(hyperparam_filename)}")', file=f)
|
| 472 |
print(f'include("{_escape_filename(dataset_filename)}")', file=f)
|
| 473 |
if len(variable_names) == 0:
|
| 474 |
varMap = "[" + ",".join([f'"x{i}"' for i in range(X.shape[1])]) + "]"
|
| 475 |
else:
|
| 476 |
-
varMap =
|
|
|
|
|
|
|
| 477 |
|
| 478 |
if weights is not None:
|
| 479 |
-
print(
|
|
|
|
|
|
|
|
|
|
| 480 |
else:
|
| 481 |
-
print(
|
|
|
|
|
|
|
|
|
|
| 482 |
|
| 483 |
|
| 484 |
-
def _make_datasets_julia_str(
|
| 485 |
-
|
|
|
|
| 486 |
def_datasets = """using DelimitedFiles"""
|
| 487 |
-
np.savetxt(X_filename, X.astype(np.float32), delimiter=
|
| 488 |
if multioutput:
|
| 489 |
-
np.savetxt(y_filename, y.astype(np.float32), delimiter=
|
| 490 |
else:
|
| 491 |
-
np.savetxt(y_filename, y.reshape(-1, 1).astype(np.float32), delimiter=
|
| 492 |
if weights is not None:
|
| 493 |
if multioutput:
|
| 494 |
-
np.savetxt(weights_filename, weights.astype(np.float32), delimiter=
|
| 495 |
else:
|
| 496 |
-
np.savetxt(
|
|
|
|
|
|
|
|
|
|
|
|
|
| 497 |
def_datasets += f"""
|
| 498 |
X = copy(transpose(readdlm("{_escape_filename(X_filename)}", ',', Float32, '\\n')))"""
|
| 499 |
|
| 500 |
if multioutput:
|
| 501 |
-
def_datasets+= f"""
|
| 502 |
y = copy(transpose(readdlm("{_escape_filename(y_filename)}", ',', Float32, '\\n')))"""
|
| 503 |
else:
|
| 504 |
-
def_datasets+= f"""
|
| 505 |
y = readdlm("{_escape_filename(y_filename)}", ',', Float32, '\\n')[:, 1]"""
|
| 506 |
|
| 507 |
if weights is not None:
|
|
@@ -513,30 +584,69 @@ weights = copy(transpose(readdlm("{_escape_filename(weights_filename)}", ',', Fl
|
|
| 513 |
weights = readdlm("{_escape_filename(weights_filename)}", ',', Float32, '\\n')[:, 1]"""
|
| 514 |
return def_datasets
|
| 515 |
|
| 516 |
-
|
| 517 |
-
|
| 518 |
-
|
| 519 |
-
|
| 520 |
-
|
| 521 |
-
|
| 522 |
-
|
| 523 |
-
|
| 524 |
-
|
| 525 |
-
|
| 526 |
-
|
| 527 |
-
|
| 528 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 529 |
try:
|
| 530 |
term_width = shutil.get_terminal_size().columns
|
| 531 |
except:
|
| 532 |
-
_, term_width = subprocess.check_output([
|
|
|
|
| 533 |
def tuple_fix(ops):
|
| 534 |
if len(ops) > 1:
|
| 535 |
-
return
|
| 536 |
elif len(ops) == 0:
|
| 537 |
-
return
|
| 538 |
else:
|
| 539 |
-
return ops[0] +
|
| 540 |
|
| 541 |
def_hyperparams += f"""\n
|
| 542 |
plus=(+)
|
|
@@ -606,7 +716,7 @@ progress={'true' if progress else 'false'},
|
|
| 606 |
terminal_width={term_width:d}
|
| 607 |
"""
|
| 608 |
|
| 609 |
-
def_hyperparams +=
|
| 610 |
return def_hyperparams
|
| 611 |
|
| 612 |
|
|
@@ -639,16 +749,20 @@ def _handle_constraints(binary_operators, constraints, unary_operators, **kwargs
|
|
| 639 |
for op in binary_operators:
|
| 640 |
if op not in constraints:
|
| 641 |
constraints[op] = (-1, -1)
|
| 642 |
-
if op in [
|
| 643 |
if constraints[op][0] != constraints[op][1]:
|
| 644 |
raise NotImplementedError(
|
| 645 |
-
"You need equal constraints on both sides for - and *, due to simplification strategies."
|
| 646 |
-
|
|
|
|
| 647 |
# Make sure the complex expression is in the left side.
|
| 648 |
if constraints[op][0] == -1:
|
| 649 |
continue
|
| 650 |
elif constraints[op][1] == -1 or constraints[op][0] < constraints[op][1]:
|
| 651 |
-
constraints[op][0], constraints[op][1] =
|
|
|
|
|
|
|
|
|
|
| 652 |
|
| 653 |
|
| 654 |
def _create_inline_operators(binary_operators, unary_operators, **kwargs):
|
|
@@ -656,27 +770,33 @@ def _create_inline_operators(binary_operators, unary_operators, **kwargs):
|
|
| 656 |
for op_list in [binary_operators, unary_operators]:
|
| 657 |
for i in range(len(op_list)):
|
| 658 |
op = op_list[i]
|
| 659 |
-
is_user_defined_operator =
|
| 660 |
|
| 661 |
if is_user_defined_operator:
|
| 662 |
def_hyperparams += op + "\n"
|
| 663 |
# Cut off from the first non-alphanumeric char:
|
| 664 |
first_non_char = [
|
| 665 |
-
j
|
| 666 |
-
|
|
|
|
|
|
|
| 667 |
function_name = op[:first_non_char]
|
| 668 |
op_list[i] = function_name
|
| 669 |
return def_hyperparams
|
| 670 |
|
| 671 |
|
| 672 |
-
def _handle_feature_selection(
|
|
|
|
|
|
|
| 673 |
if select_k_features is not None:
|
| 674 |
selection = run_feature_selection(X, y, select_k_features)
|
| 675 |
print(f"Using features {selection}")
|
| 676 |
X = X[:, selection]
|
| 677 |
|
| 678 |
if use_custom_variable_names:
|
| 679 |
-
variable_names = [
|
|
|
|
|
|
|
| 680 |
else:
|
| 681 |
selection = None
|
| 682 |
return X, variable_names, selection
|
|
@@ -687,22 +807,34 @@ def _set_paths(tempdir):
|
|
| 687 |
pkg_directory = Path(__file__).parents[1]
|
| 688 |
default_project_file = pkg_directory / "Project.toml"
|
| 689 |
tmpdir = Path(tempfile.mkdtemp(dir=tempdir))
|
| 690 |
-
hyperparam_filename = tmpdir / f
|
| 691 |
-
dataset_filename = tmpdir / f
|
| 692 |
-
runfile_filename = tmpdir / f
|
| 693 |
X_filename = tmpdir / "X.csv"
|
| 694 |
y_filename = tmpdir / "y.csv"
|
| 695 |
weights_filename = tmpdir / "weights.csv"
|
| 696 |
-
return dict(
|
| 697 |
-
|
| 698 |
-
|
| 699 |
-
|
| 700 |
-
|
| 701 |
-
|
| 702 |
-
|
| 703 |
-
|
| 704 |
-
|
| 705 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 706 |
# Check for potential errors before they happen
|
| 707 |
assert len(unary_operators) + len(binary_operators) > 0
|
| 708 |
assert len(X.shape) == 2
|
|
@@ -714,76 +846,108 @@ def _check_assertions(X, binary_operators, unary_operators, use_custom_variable_
|
|
| 714 |
if use_custom_variable_names:
|
| 715 |
assert len(variable_names) == X.shape[1]
|
| 716 |
|
|
|
|
| 717 |
def _check_for_julia_installation():
|
| 718 |
try:
|
| 719 |
process = subprocess.Popen(["julia", "-v"], stdout=subprocess.PIPE, bufsize=-1)
|
| 720 |
while True:
|
| 721 |
line = process.stdout.readline()
|
| 722 |
-
if not line:
|
|
|
|
| 723 |
process.stdout.close()
|
| 724 |
process.wait()
|
| 725 |
except FileNotFoundError:
|
| 726 |
import os
|
| 727 |
-
|
|
|
|
|
|
|
|
|
|
| 728 |
process.kill()
|
| 729 |
|
| 730 |
|
| 731 |
def run_feature_selection(X, y, select_k_features):
|
| 732 |
"""Use a gradient boosting tree regressor as a proxy for finding
|
| 733 |
-
|
| 734 |
-
|
| 735 |
|
| 736 |
from sklearn.ensemble import RandomForestRegressor
|
| 737 |
from sklearn.feature_selection import SelectFromModel, SelectKBest
|
| 738 |
|
| 739 |
clf = RandomForestRegressor(n_estimators=100, max_depth=3, random_state=0)
|
| 740 |
clf.fit(X, y)
|
| 741 |
-
selector = SelectFromModel(
|
| 742 |
-
|
|
|
|
| 743 |
return selector.get_support(indices=True)
|
| 744 |
|
| 745 |
-
|
| 746 |
-
|
| 747 |
-
|
| 748 |
-
|
| 749 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 750 |
"""Get the equations from a hall of fame file. If no arguments
|
| 751 |
entered, the ones used previously from a call to PySR will be used."""
|
| 752 |
|
| 753 |
global global_state
|
| 754 |
|
| 755 |
-
if equation_file is None:
|
| 756 |
-
|
| 757 |
-
if
|
| 758 |
-
|
| 759 |
-
if
|
| 760 |
-
|
| 761 |
-
if
|
| 762 |
-
|
| 763 |
-
if
|
| 764 |
-
|
| 765 |
-
if
|
| 766 |
-
|
| 767 |
-
|
| 768 |
-
|
| 769 |
-
|
| 770 |
-
|
| 771 |
-
|
| 772 |
-
|
| 773 |
-
|
| 774 |
-
|
| 775 |
-
|
| 776 |
-
|
| 777 |
-
|
| 778 |
-
global_state[
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 779 |
|
| 780 |
try:
|
| 781 |
if multioutput:
|
| 782 |
-
all_outputs = [
|
|
|
|
|
|
|
|
|
|
| 783 |
else:
|
| 784 |
-
all_outputs = [pd.read_csv(str(equation_file) +
|
| 785 |
except FileNotFoundError:
|
| 786 |
-
raise RuntimeError(
|
|
|
|
|
|
|
| 787 |
|
| 788 |
ret_outputs = []
|
| 789 |
|
|
@@ -798,19 +962,16 @@ def get_hof(equation_file=None, n_features=None, variable_names=None,
|
|
| 798 |
jax_format = []
|
| 799 |
if output_torch_format:
|
| 800 |
torch_format = []
|
| 801 |
-
use_custom_variable_names =
|
| 802 |
-
local_sympy_mappings = {
|
| 803 |
-
**extra_sympy_mappings,
|
| 804 |
-
**sympy_mappings
|
| 805 |
-
}
|
| 806 |
|
| 807 |
if use_custom_variable_names:
|
| 808 |
sympy_symbols = [sympy.Symbol(variable_names[i]) for i in range(n_features)]
|
| 809 |
else:
|
| 810 |
-
sympy_symbols = [sympy.Symbol(
|
| 811 |
|
| 812 |
for i in range(len(output)):
|
| 813 |
-
eqn = sympify(output.loc[i,
|
| 814 |
sympy_format.append(eqn)
|
| 815 |
|
| 816 |
# Numpy:
|
|
@@ -819,37 +980,46 @@ def get_hof(equation_file=None, n_features=None, variable_names=None,
|
|
| 819 |
# JAX:
|
| 820 |
if output_jax_format:
|
| 821 |
from .export_jax import sympy2jax
|
|
|
|
| 822 |
func, params = sympy2jax(eqn, sympy_symbols, selection)
|
| 823 |
-
jax_format.append({
|
| 824 |
|
| 825 |
# Torch:
|
| 826 |
if output_torch_format:
|
| 827 |
from .export_torch import sympy2torch
|
|
|
|
| 828 |
module = sympy2torch(eqn, sympy_symbols, selection=selection)
|
| 829 |
torch_format.append(module)
|
| 830 |
|
| 831 |
-
curMSE = output.loc[i,
|
| 832 |
-
curComplexity = output.loc[i,
|
| 833 |
|
| 834 |
if lastMSE is None:
|
| 835 |
cur_score = 0.0
|
| 836 |
else:
|
| 837 |
-
cur_score = -
|
| 838 |
|
| 839 |
scores.append(cur_score)
|
| 840 |
lastMSE = curMSE
|
| 841 |
lastComplexity = curComplexity
|
| 842 |
|
| 843 |
-
output[
|
| 844 |
-
output[
|
| 845 |
-
output[
|
| 846 |
-
output_cols = [
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 847 |
if output_jax_format:
|
| 848 |
-
output_cols += [
|
| 849 |
-
output[
|
| 850 |
if output_torch_format:
|
| 851 |
-
output_cols += [
|
| 852 |
-
output[
|
| 853 |
|
| 854 |
ret_outputs.append(output[output_cols])
|
| 855 |
|
|
@@ -858,67 +1028,80 @@ def get_hof(equation_file=None, n_features=None, variable_names=None,
|
|
| 858 |
else:
|
| 859 |
return ret_outputs[0]
|
| 860 |
|
|
|
|
| 861 |
def best_row(equations=None):
|
| 862 |
"""Return the best row of a hall of fame file using the score column.
|
| 863 |
By default this uses the last equation file.
|
| 864 |
"""
|
| 865 |
-
if equations is None:
|
|
|
|
| 866 |
if isinstance(equations, list):
|
| 867 |
-
return [eq.iloc[np.argmax(eq[
|
| 868 |
else:
|
| 869 |
-
return equations.iloc[np.argmax(equations[
|
|
|
|
| 870 |
|
| 871 |
def best_tex(equations=None):
|
| 872 |
"""Return the equation with the best score, in latex format
|
| 873 |
By default this uses the last equation file.
|
| 874 |
"""
|
| 875 |
-
if equations is None:
|
|
|
|
| 876 |
if isinstance(equations, list):
|
| 877 |
-
return [
|
|
|
|
|
|
|
| 878 |
else:
|
| 879 |
-
return sympy.latex(best_row(equations)[
|
|
|
|
| 880 |
|
| 881 |
def best(equations=None):
|
| 882 |
"""Return the equation with the best score, in sympy format.
|
| 883 |
By default this uses the last equation file.
|
| 884 |
"""
|
| 885 |
-
if equations is None:
|
|
|
|
| 886 |
if isinstance(equations, list):
|
| 887 |
-
return [best_row(eq)[
|
| 888 |
else:
|
| 889 |
-
return best_row(equations)[
|
|
|
|
| 890 |
|
| 891 |
def best_callable(equations=None):
|
| 892 |
"""Return the equation with the best score, in callable format.
|
| 893 |
By default this uses the last equation file.
|
| 894 |
"""
|
| 895 |
-
if equations is None:
|
|
|
|
| 896 |
if isinstance(equations, list):
|
| 897 |
-
return [best_row(eq)[
|
| 898 |
else:
|
| 899 |
-
return best_row(equations)[
|
|
|
|
| 900 |
|
| 901 |
def _escape_filename(filename):
|
| 902 |
"""Turns a file into a string representation with correctly escaped backslashes"""
|
| 903 |
repr = str(filename)
|
| 904 |
-
repr = repr.replace(
|
| 905 |
return repr
|
| 906 |
|
|
|
|
| 907 |
# https://gist.github.com/garrettdreyfus/8153571
|
| 908 |
def _yesno(question):
|
| 909 |
"""Simple Yes/No Function."""
|
| 910 |
-
prompt = f
|
| 911 |
ans = input(prompt).strip().lower()
|
| 912 |
-
if ans not in [
|
| 913 |
-
print(f
|
| 914 |
return _yesno(question)
|
| 915 |
-
if ans ==
|
| 916 |
return True
|
| 917 |
return False
|
| 918 |
|
| 919 |
|
| 920 |
class CallableEquation(object):
|
| 921 |
"""Simple wrapper for numpy lambda functions built with sympy"""
|
|
|
|
| 922 |
def __init__(self, sympy_symbols, eqn, selection=None):
|
| 923 |
self._sympy = eqn
|
| 924 |
self._sympy_symbols = sympy_symbols
|
|
@@ -933,4 +1116,3 @@ class CallableEquation(object):
|
|
| 933 |
return self._lambda(*X[:, self._selection].T)
|
| 934 |
else:
|
| 935 |
return self._lambda(*X.T)
|
| 936 |
-
|
|
|
|
| 15 |
import warnings
|
| 16 |
|
| 17 |
global_state = dict(
|
| 18 |
+
equation_file="hall_of_fame.csv",
|
| 19 |
n_features=None,
|
| 20 |
variable_names=[],
|
| 21 |
extra_sympy_mappings={},
|
|
|
|
| 25 |
output_torch_format=False,
|
| 26 |
multioutput=False,
|
| 27 |
nout=1,
|
| 28 |
+
selection=None,
|
| 29 |
)
|
| 30 |
|
| 31 |
sympy_mappings = {
|
| 32 |
+
"div": lambda x, y: x / y,
|
| 33 |
+
"mult": lambda x, y: x * y,
|
| 34 |
+
"sqrt_abs": lambda x: sympy.sqrt(abs(x)),
|
| 35 |
+
"square": lambda x: x ** 2,
|
| 36 |
+
"cube": lambda x: x ** 3,
|
| 37 |
+
"plus": lambda x, y: x + y,
|
| 38 |
+
"sub": lambda x, y: x - y,
|
| 39 |
+
"neg": lambda x: -x,
|
| 40 |
+
"pow": lambda x, y: abs(x) ** y,
|
| 41 |
+
"cos": lambda x: sympy.cos(x),
|
| 42 |
+
"sin": lambda x: sympy.sin(x),
|
| 43 |
+
"tan": lambda x: sympy.tan(x),
|
| 44 |
+
"cosh": lambda x: sympy.cosh(x),
|
| 45 |
+
"sinh": lambda x: sympy.sinh(x),
|
| 46 |
+
"tanh": lambda x: sympy.tanh(x),
|
| 47 |
+
"exp": lambda x: sympy.exp(x),
|
| 48 |
+
"acos": lambda x: sympy.acos(x),
|
| 49 |
+
"asin": lambda x: sympy.asin(x),
|
| 50 |
+
"atan": lambda x: sympy.atan(x),
|
| 51 |
+
"acosh": lambda x: sympy.acosh(abs(x) + 1),
|
| 52 |
+
"acosh_abs": lambda x: sympy.acosh(abs(x) + 1),
|
| 53 |
+
"asinh": lambda x: sympy.asinh(x),
|
| 54 |
+
"atanh": lambda x: sympy.atanh(sympy.Mod(x + 1, 2) - 1),
|
| 55 |
+
"atanh_clip": lambda x: sympy.atanh(sympy.Mod(x + 1, 2) - 1),
|
| 56 |
+
"abs": lambda x: abs(x),
|
| 57 |
+
"mod": lambda x, y: sympy.Mod(x, y),
|
| 58 |
+
"erf": lambda x: sympy.erf(x),
|
| 59 |
+
"erfc": lambda x: sympy.erfc(x),
|
| 60 |
+
"log_abs": lambda x: sympy.log(abs(x)),
|
| 61 |
+
"log10_abs": lambda x: sympy.log(abs(x), 10),
|
| 62 |
+
"log2_abs": lambda x: sympy.log(abs(x), 2),
|
| 63 |
+
"log1p_abs": lambda x: sympy.log(abs(x) + 1),
|
| 64 |
+
"floor": lambda x: sympy.floor(x),
|
| 65 |
+
"ceil": lambda x: sympy.ceil(x),
|
| 66 |
+
"sign": lambda x: sympy.sign(x),
|
| 67 |
+
"gamma": lambda x: sympy.gamma(x),
|
| 68 |
}
|
| 69 |
|
| 70 |
+
|
| 71 |
+
def pysr(
|
| 72 |
+
X,
|
| 73 |
+
y,
|
| 74 |
+
weights=None,
|
| 75 |
+
binary_operators=None,
|
| 76 |
+
unary_operators=None,
|
| 77 |
+
procs=4,
|
| 78 |
+
loss="L2DistLoss()",
|
| 79 |
+
populations=20,
|
| 80 |
+
niterations=100,
|
| 81 |
+
ncyclesperiteration=300,
|
| 82 |
+
alpha=0.1,
|
| 83 |
+
annealing=False,
|
| 84 |
+
fractionReplaced=0.10,
|
| 85 |
+
fractionReplacedHof=0.10,
|
| 86 |
+
npop=1000,
|
| 87 |
+
parsimony=1e-4,
|
| 88 |
+
migration=True,
|
| 89 |
+
hofMigration=True,
|
| 90 |
+
shouldOptimizeConstants=True,
|
| 91 |
+
topn=10,
|
| 92 |
+
weightAddNode=1,
|
| 93 |
+
weightInsertNode=3,
|
| 94 |
+
weightDeleteNode=3,
|
| 95 |
+
weightDoNothing=1,
|
| 96 |
+
weightMutateConstant=10,
|
| 97 |
+
weightMutateOperator=1,
|
| 98 |
+
weightRandomize=1,
|
| 99 |
+
weightSimplify=0.01,
|
| 100 |
+
perturbationFactor=1.0,
|
| 101 |
+
timeout=None,
|
| 102 |
+
extra_sympy_mappings=None,
|
| 103 |
+
extra_torch_mappings=None,
|
| 104 |
+
extra_jax_mappings=None,
|
| 105 |
+
equation_file=None,
|
| 106 |
+
verbosity=1e9,
|
| 107 |
+
progress=None,
|
| 108 |
+
maxsize=20,
|
| 109 |
+
fast_cycle=False,
|
| 110 |
+
maxdepth=None,
|
| 111 |
+
variable_names=None,
|
| 112 |
+
batching=False,
|
| 113 |
+
batchSize=50,
|
| 114 |
+
select_k_features=None,
|
| 115 |
+
warmupMaxsizeBy=0.0,
|
| 116 |
+
constraints=None,
|
| 117 |
+
useFrequency=True,
|
| 118 |
+
tempdir=None,
|
| 119 |
+
delete_tempfiles=True,
|
| 120 |
+
julia_optimization=3,
|
| 121 |
+
julia_project=None,
|
| 122 |
+
user_input=True,
|
| 123 |
+
update=True,
|
| 124 |
+
temp_equation_file=False,
|
| 125 |
+
output_jax_format=False,
|
| 126 |
+
output_torch_format=False,
|
| 127 |
+
optimizer_algorithm="BFGS",
|
| 128 |
+
optimizer_nrestarts=3,
|
| 129 |
+
optimize_probability=1.0,
|
| 130 |
+
optimizer_iterations=10,
|
| 131 |
+
tournament_selection_n=10,
|
| 132 |
+
tournament_selection_p=1.0,
|
| 133 |
+
):
|
| 134 |
"""Run symbolic regression to fit f(X[i, :]) ~ y[i] for all i.
|
| 135 |
Note: most default parameters have been tuned over several example
|
| 136 |
equations, but you should adjust `niterations`,
|
|
|
|
| 248 |
:type: pd.DataFrame/list
|
| 249 |
"""
|
| 250 |
if binary_operators is None:
|
| 251 |
+
binary_operators = "+ * - /".split(" ")
|
| 252 |
if unary_operators is None:
|
| 253 |
unary_operators = []
|
| 254 |
if extra_sympy_mappings is None:
|
|
|
|
| 259 |
constraints = {}
|
| 260 |
|
| 261 |
if progress is not None:
|
| 262 |
+
if progress and ("buffer" not in sys.stdout.__dir__()):
|
| 263 |
+
warnings.warn(
|
| 264 |
+
"Note: it looks like you are running in Jupyter. The progress bar will be turned off."
|
| 265 |
+
)
|
| 266 |
progress = False
|
| 267 |
else:
|
| 268 |
+
if "buffer" in sys.stdout.__dir__():
|
| 269 |
progress = True
|
| 270 |
else:
|
| 271 |
progress = False
|
| 272 |
|
| 273 |
+
assert optimizer_algorithm in ["NelderMead", "BFGS"]
|
| 274 |
assert tournament_selection_n < npop
|
| 275 |
|
| 276 |
if isinstance(X, pd.DataFrame):
|
|
|
|
| 281 |
X = X[:, None]
|
| 282 |
|
| 283 |
if len(variable_names) == 0:
|
| 284 |
+
variable_names = [f"x{i}" for i in range(X.shape[1])]
|
| 285 |
+
|
| 286 |
+
use_custom_variable_names = len(variable_names) != 0
|
| 287 |
+
|
| 288 |
+
_check_assertions(
|
| 289 |
+
X,
|
| 290 |
+
binary_operators,
|
| 291 |
+
unary_operators,
|
| 292 |
+
use_custom_variable_names,
|
| 293 |
+
variable_names,
|
| 294 |
+
weights,
|
| 295 |
+
y,
|
| 296 |
+
)
|
| 297 |
_check_for_julia_installation()
|
| 298 |
|
|
|
|
| 299 |
if len(X) > 10000 and not batching:
|
| 300 |
+
warnings.warn(
|
| 301 |
+
"Note: you are running with more than 10,000 datapoints. You should consider turning on batching (https://pysr.readthedocs.io/en/latest/docs/options/#batching). You should also reconsider if you need that many datapoints. Unless you have a large amount of noise (in which case you should smooth your dataset first), generally < 10,000 datapoints is enough to find a functional form with symbolic regression. More datapoints will lower the search speed."
|
| 302 |
+
)
|
| 303 |
|
| 304 |
if maxsize > 40:
|
| 305 |
+
warnings.warn(
|
| 306 |
+
"Note: Using a large maxsize for the equation search will be exponentially slower and use significant memory. You should consider turning `useFrequency` to False, and perhaps use `warmupMaxsizeBy`."
|
| 307 |
+
)
|
| 308 |
|
| 309 |
X, variable_names, selection = _handle_feature_selection(
|
| 310 |
+
X, select_k_features, use_custom_variable_names, variable_names, y
|
| 311 |
+
)
|
|
|
|
| 312 |
|
| 313 |
if maxdepth is None:
|
| 314 |
maxdepth = maxsize
|
|
|
|
| 327 |
else:
|
| 328 |
raise NotImplementedError("y shape not supported!")
|
| 329 |
|
| 330 |
+
kwargs = dict(
|
| 331 |
+
X=X,
|
| 332 |
+
y=y,
|
| 333 |
+
weights=weights,
|
| 334 |
+
alpha=alpha,
|
| 335 |
+
annealing=annealing,
|
| 336 |
+
batchSize=batchSize,
|
| 337 |
+
batching=batching,
|
| 338 |
+
binary_operators=binary_operators,
|
| 339 |
+
fast_cycle=fast_cycle,
|
| 340 |
+
fractionReplaced=fractionReplaced,
|
| 341 |
+
ncyclesperiteration=ncyclesperiteration,
|
| 342 |
+
niterations=niterations,
|
| 343 |
+
npop=npop,
|
| 344 |
+
topn=topn,
|
| 345 |
+
verbosity=verbosity,
|
| 346 |
+
progress=progress,
|
| 347 |
+
update=update,
|
| 348 |
+
julia_optimization=julia_optimization,
|
| 349 |
+
timeout=timeout,
|
| 350 |
+
fractionReplacedHof=fractionReplacedHof,
|
| 351 |
+
hofMigration=hofMigration,
|
| 352 |
+
maxdepth=maxdepth,
|
| 353 |
+
maxsize=maxsize,
|
| 354 |
+
migration=migration,
|
| 355 |
+
optimizer_algorithm=optimizer_algorithm,
|
| 356 |
+
optimizer_nrestarts=optimizer_nrestarts,
|
| 357 |
+
optimize_probability=optimize_probability,
|
| 358 |
+
optimizer_iterations=optimizer_iterations,
|
| 359 |
+
parsimony=parsimony,
|
| 360 |
+
perturbationFactor=perturbationFactor,
|
| 361 |
+
populations=populations,
|
| 362 |
+
procs=procs,
|
| 363 |
+
shouldOptimizeConstants=shouldOptimizeConstants,
|
| 364 |
+
unary_operators=unary_operators,
|
| 365 |
+
useFrequency=useFrequency,
|
| 366 |
+
use_custom_variable_names=use_custom_variable_names,
|
| 367 |
+
variable_names=variable_names,
|
| 368 |
+
warmupMaxsizeBy=warmupMaxsizeBy,
|
| 369 |
+
weightAddNode=weightAddNode,
|
| 370 |
+
weightDeleteNode=weightDeleteNode,
|
| 371 |
+
weightDoNothing=weightDoNothing,
|
| 372 |
+
weightInsertNode=weightInsertNode,
|
| 373 |
+
weightMutateConstant=weightMutateConstant,
|
| 374 |
+
weightMutateOperator=weightMutateOperator,
|
| 375 |
+
weightRandomize=weightRandomize,
|
| 376 |
+
weightSimplify=weightSimplify,
|
| 377 |
+
constraints=constraints,
|
| 378 |
+
extra_sympy_mappings=extra_sympy_mappings,
|
| 379 |
+
extra_jax_mappings=extra_jax_mappings,
|
| 380 |
+
extra_torch_mappings=extra_torch_mappings,
|
| 381 |
+
julia_project=julia_project,
|
| 382 |
+
loss=loss,
|
| 383 |
+
output_jax_format=output_jax_format,
|
| 384 |
+
output_torch_format=output_torch_format,
|
| 385 |
+
selection=selection,
|
| 386 |
+
multioutput=multioutput,
|
| 387 |
+
nout=nout,
|
| 388 |
+
tournament_selection_n=tournament_selection_n,
|
| 389 |
+
tournament_selection_p=tournament_selection_p,
|
| 390 |
+
)
|
| 391 |
|
| 392 |
kwargs = {**_set_paths(tempdir), **kwargs}
|
| 393 |
|
| 394 |
if temp_equation_file:
|
| 395 |
+
equation_file = kwargs["tmpdir"] / f"hall_of_fame.csv"
|
| 396 |
elif equation_file is None:
|
| 397 |
date_time = datetime.now().strftime("%Y-%m-%d_%H%M%S.%f")[:-3]
|
| 398 |
+
equation_file = "hall_of_fame_" + date_time + ".csv"
|
| 399 |
|
| 400 |
kwargs = {**dict(equation_file=equation_file), **kwargs}
|
| 401 |
|
| 402 |
+
pkg_directory = kwargs["pkg_directory"]
|
|
|
|
| 403 |
manifest_file = None
|
| 404 |
+
if kwargs["julia_project"] is not None:
|
| 405 |
+
manifest_filepath = Path(kwargs["julia_project"]) / "Manifest.toml"
|
| 406 |
else:
|
| 407 |
+
manifest_filepath = pkg_directory / "Manifest.toml"
|
| 408 |
|
| 409 |
+
kwargs["need_install"] = False
|
| 410 |
|
| 411 |
if not (manifest_filepath).is_file():
|
| 412 |
+
kwargs["need_install"] = (not user_input) or _yesno(
|
| 413 |
+
"I will install Julia packages using PySR's Project.toml file. OK?"
|
| 414 |
+
)
|
| 415 |
+
if kwargs["need_install"]:
|
| 416 |
print("OK. I will install at launch.")
|
| 417 |
assert update
|
| 418 |
|
| 419 |
+
kwargs["def_hyperparams"] = _create_inline_operators(**kwargs)
|
| 420 |
|
| 421 |
_handle_constraints(**kwargs)
|
| 422 |
|
| 423 |
+
kwargs["constraints_str"] = _make_constraints_str(**kwargs)
|
| 424 |
+
kwargs["def_hyperparams"] = _make_hyperparams_julia_str(**kwargs)
|
| 425 |
+
kwargs["def_datasets"] = _make_datasets_julia_str(**kwargs)
|
| 426 |
|
| 427 |
_create_julia_files(**kwargs)
|
| 428 |
_final_pysr_process(**kwargs)
|
|
|
|
| 431 |
equations = get_hof(**kwargs)
|
| 432 |
|
| 433 |
if delete_tempfiles:
|
| 434 |
+
shutil.rmtree(kwargs["tmpdir"])
|
| 435 |
|
| 436 |
return equations
|
| 437 |
|
|
|
|
| 439 |
def _set_globals(X, **kwargs):
|
| 440 |
global global_state
|
| 441 |
|
| 442 |
+
global_state["n_features"] = X.shape[1]
|
| 443 |
for key, value in kwargs.items():
|
| 444 |
if key in global_state:
|
| 445 |
global_state[key] = value
|
|
|
|
| 447 |
|
| 448 |
def _final_pysr_process(julia_optimization, runfile_filename, timeout, **kwargs):
|
| 449 |
command = [
|
| 450 |
+
f"julia",
|
| 451 |
+
f"-O{julia_optimization:d}",
|
| 452 |
str(runfile_filename),
|
| 453 |
]
|
| 454 |
if timeout is not None:
|
| 455 |
+
command = [f"timeout", f"{timeout}"] + command
|
| 456 |
_cmd_runner(command, **kwargs)
|
| 457 |
|
| 458 |
+
|
| 459 |
def _cmd_runner(command, progress, **kwargs):
|
| 460 |
+
if kwargs["verbosity"] > 0:
|
| 461 |
+
print("Running on", " ".join(command))
|
| 462 |
process = subprocess.Popen(command, stdout=subprocess.PIPE, bufsize=-1)
|
| 463 |
try:
|
| 464 |
while True:
|
| 465 |
line = process.stdout.readline()
|
| 466 |
+
if not line:
|
| 467 |
+
break
|
| 468 |
+
decoded_line = line.decode("utf-8")
|
| 469 |
if progress:
|
| 470 |
+
decoded_line = (
|
| 471 |
+
decoded_line.replace("\\033[K", "\033[K")
|
| 472 |
+
.replace("\\033[1A", "\033[1A")
|
| 473 |
+
.replace("\\033[1B", "\033[1B")
|
| 474 |
+
.replace("\\r", "\r")
|
| 475 |
+
.encode(sys.stdout.encoding, errors="replace")
|
| 476 |
+
)
|
| 477 |
sys.stdout.buffer.write(decoded_line)
|
| 478 |
sys.stdout.flush()
|
| 479 |
else:
|
| 480 |
+
print(decoded_line, end="")
|
| 481 |
|
| 482 |
process.stdout.close()
|
| 483 |
process.wait()
|
|
|
|
| 485 |
print("Killing process... will return when done.")
|
| 486 |
process.kill()
|
| 487 |
|
| 488 |
+
|
| 489 |
+
def _create_julia_files(
|
| 490 |
+
dataset_filename,
|
| 491 |
+
def_datasets,
|
| 492 |
+
hyperparam_filename,
|
| 493 |
+
def_hyperparams,
|
| 494 |
+
fractionReplaced,
|
| 495 |
+
ncyclesperiteration,
|
| 496 |
+
niterations,
|
| 497 |
+
npop,
|
| 498 |
+
runfile_filename,
|
| 499 |
+
topn,
|
| 500 |
+
verbosity,
|
| 501 |
+
julia_project,
|
| 502 |
+
procs,
|
| 503 |
+
weights,
|
| 504 |
+
X,
|
| 505 |
+
variable_names,
|
| 506 |
+
pkg_directory,
|
| 507 |
+
need_install,
|
| 508 |
+
update,
|
| 509 |
+
**kwargs,
|
| 510 |
+
):
|
| 511 |
+
with open(hyperparam_filename, "w") as f:
|
| 512 |
print(def_hyperparams, file=f)
|
| 513 |
+
with open(dataset_filename, "w") as f:
|
| 514 |
print(def_datasets, file=f)
|
| 515 |
+
with open(runfile_filename, "w") as f:
|
| 516 |
if julia_project is None:
|
| 517 |
julia_project = pkg_directory
|
| 518 |
else:
|
| 519 |
julia_project = Path(julia_project)
|
| 520 |
+
print(f"import Pkg", file=f)
|
| 521 |
print(f'Pkg.activate("{_escape_filename(julia_project)}")', file=f)
|
| 522 |
if need_install:
|
| 523 |
+
print(f"Pkg.instantiate()", file=f)
|
| 524 |
+
print(f"Pkg.update()", file=f)
|
| 525 |
+
print(f"Pkg.precompile()", file=f)
|
| 526 |
elif update:
|
| 527 |
+
print(f"Pkg.update()", file=f)
|
| 528 |
+
print(f"using SymbolicRegression", file=f)
|
| 529 |
print(f'include("{_escape_filename(hyperparam_filename)}")', file=f)
|
| 530 |
print(f'include("{_escape_filename(dataset_filename)}")', file=f)
|
| 531 |
if len(variable_names) == 0:
|
| 532 |
varMap = "[" + ",".join([f'"x{i}"' for i in range(X.shape[1])]) + "]"
|
| 533 |
else:
|
| 534 |
+
varMap = (
|
| 535 |
+
"[" + ",".join(['"' + vname + '"' for vname in variable_names]) + "]"
|
| 536 |
+
)
|
| 537 |
|
| 538 |
if weights is not None:
|
| 539 |
+
print(
|
| 540 |
+
f"EquationSearch(X, y, weights=weights, niterations={niterations:d}, varMap={varMap}, options=options, numprocs={procs})",
|
| 541 |
+
file=f,
|
| 542 |
+
)
|
| 543 |
else:
|
| 544 |
+
print(
|
| 545 |
+
f"EquationSearch(X, y, niterations={niterations:d}, varMap={varMap}, options=options, numprocs={procs})",
|
| 546 |
+
file=f,
|
| 547 |
+
)
|
| 548 |
|
| 549 |
|
| 550 |
+
def _make_datasets_julia_str(
|
| 551 |
+
X, X_filename, weights, weights_filename, y, y_filename, multioutput, **kwargs
|
| 552 |
+
):
|
| 553 |
def_datasets = """using DelimitedFiles"""
|
| 554 |
+
np.savetxt(X_filename, X.astype(np.float32), delimiter=",")
|
| 555 |
if multioutput:
|
| 556 |
+
np.savetxt(y_filename, y.astype(np.float32), delimiter=",")
|
| 557 |
else:
|
| 558 |
+
np.savetxt(y_filename, y.reshape(-1, 1).astype(np.float32), delimiter=",")
|
| 559 |
if weights is not None:
|
| 560 |
if multioutput:
|
| 561 |
+
np.savetxt(weights_filename, weights.astype(np.float32), delimiter=",")
|
| 562 |
else:
|
| 563 |
+
np.savetxt(
|
| 564 |
+
weights_filename,
|
| 565 |
+
weights.reshape(-1, 1).astype(np.float32),
|
| 566 |
+
delimiter=",",
|
| 567 |
+
)
|
| 568 |
def_datasets += f"""
|
| 569 |
X = copy(transpose(readdlm("{_escape_filename(X_filename)}", ',', Float32, '\\n')))"""
|
| 570 |
|
| 571 |
if multioutput:
|
| 572 |
+
def_datasets += f"""
|
| 573 |
y = copy(transpose(readdlm("{_escape_filename(y_filename)}", ',', Float32, '\\n')))"""
|
| 574 |
else:
|
| 575 |
+
def_datasets += f"""
|
| 576 |
y = readdlm("{_escape_filename(y_filename)}", ',', Float32, '\\n')[:, 1]"""
|
| 577 |
|
| 578 |
if weights is not None:
|
|
|
|
| 584 |
weights = readdlm("{_escape_filename(weights_filename)}", ',', Float32, '\\n')[:, 1]"""
|
| 585 |
return def_datasets
|
| 586 |
|
| 587 |
+
|
| 588 |
+
def _make_hyperparams_julia_str(
|
| 589 |
+
X,
|
| 590 |
+
alpha,
|
| 591 |
+
annealing,
|
| 592 |
+
batchSize,
|
| 593 |
+
batching,
|
| 594 |
+
binary_operators,
|
| 595 |
+
constraints_str,
|
| 596 |
+
def_hyperparams,
|
| 597 |
+
equation_file,
|
| 598 |
+
fast_cycle,
|
| 599 |
+
fractionReplacedHof,
|
| 600 |
+
hofMigration,
|
| 601 |
+
maxdepth,
|
| 602 |
+
maxsize,
|
| 603 |
+
migration,
|
| 604 |
+
optimizer_algorithm,
|
| 605 |
+
optimizer_nrestarts,
|
| 606 |
+
optimize_probability,
|
| 607 |
+
optimizer_iterations,
|
| 608 |
+
npop,
|
| 609 |
+
parsimony,
|
| 610 |
+
perturbationFactor,
|
| 611 |
+
populations,
|
| 612 |
+
procs,
|
| 613 |
+
shouldOptimizeConstants,
|
| 614 |
+
unary_operators,
|
| 615 |
+
useFrequency,
|
| 616 |
+
use_custom_variable_names,
|
| 617 |
+
variable_names,
|
| 618 |
+
warmupMaxsizeBy,
|
| 619 |
+
weightAddNode,
|
| 620 |
+
ncyclesperiteration,
|
| 621 |
+
fractionReplaced,
|
| 622 |
+
topn,
|
| 623 |
+
verbosity,
|
| 624 |
+
progress,
|
| 625 |
+
loss,
|
| 626 |
+
weightDeleteNode,
|
| 627 |
+
weightDoNothing,
|
| 628 |
+
weightInsertNode,
|
| 629 |
+
weightMutateConstant,
|
| 630 |
+
weightMutateOperator,
|
| 631 |
+
weightRandomize,
|
| 632 |
+
weightSimplify,
|
| 633 |
+
weights,
|
| 634 |
+
tournament_selection_n,
|
| 635 |
+
tournament_selection_p,
|
| 636 |
+
**kwargs,
|
| 637 |
+
):
|
| 638 |
try:
|
| 639 |
term_width = shutil.get_terminal_size().columns
|
| 640 |
except:
|
| 641 |
+
_, term_width = subprocess.check_output(["stty", "size"]).split()
|
| 642 |
+
|
| 643 |
def tuple_fix(ops):
|
| 644 |
if len(ops) > 1:
|
| 645 |
+
return ", ".join(ops)
|
| 646 |
elif len(ops) == 0:
|
| 647 |
+
return ""
|
| 648 |
else:
|
| 649 |
+
return ops[0] + ","
|
| 650 |
|
| 651 |
def_hyperparams += f"""\n
|
| 652 |
plus=(+)
|
|
|
|
| 716 |
terminal_width={term_width:d}
|
| 717 |
"""
|
| 718 |
|
| 719 |
+
def_hyperparams += "\n)"
|
| 720 |
return def_hyperparams
|
| 721 |
|
| 722 |
|
|
|
|
| 749 |
for op in binary_operators:
|
| 750 |
if op not in constraints:
|
| 751 |
constraints[op] = (-1, -1)
|
| 752 |
+
if op in ["plus", "sub"]:
|
| 753 |
if constraints[op][0] != constraints[op][1]:
|
| 754 |
raise NotImplementedError(
|
| 755 |
+
"You need equal constraints on both sides for - and *, due to simplification strategies."
|
| 756 |
+
)
|
| 757 |
+
elif op == "mult":
|
| 758 |
# Make sure the complex expression is in the left side.
|
| 759 |
if constraints[op][0] == -1:
|
| 760 |
continue
|
| 761 |
elif constraints[op][1] == -1 or constraints[op][0] < constraints[op][1]:
|
| 762 |
+
constraints[op][0], constraints[op][1] = (
|
| 763 |
+
constraints[op][1],
|
| 764 |
+
constraints[op][0],
|
| 765 |
+
)
|
| 766 |
|
| 767 |
|
| 768 |
def _create_inline_operators(binary_operators, unary_operators, **kwargs):
|
|
|
|
| 770 |
for op_list in [binary_operators, unary_operators]:
|
| 771 |
for i in range(len(op_list)):
|
| 772 |
op = op_list[i]
|
| 773 |
+
is_user_defined_operator = "(" in op
|
| 774 |
|
| 775 |
if is_user_defined_operator:
|
| 776 |
def_hyperparams += op + "\n"
|
| 777 |
# Cut off from the first non-alphanumeric char:
|
| 778 |
first_non_char = [
|
| 779 |
+
j
|
| 780 |
+
for j in range(len(op))
|
| 781 |
+
if not (op[j].isalpha() or op[j].isdigit())
|
| 782 |
+
][0]
|
| 783 |
function_name = op[:first_non_char]
|
| 784 |
op_list[i] = function_name
|
| 785 |
return def_hyperparams
|
| 786 |
|
| 787 |
|
| 788 |
+
def _handle_feature_selection(
|
| 789 |
+
X, select_k_features, use_custom_variable_names, variable_names, y
|
| 790 |
+
):
|
| 791 |
if select_k_features is not None:
|
| 792 |
selection = run_feature_selection(X, y, select_k_features)
|
| 793 |
print(f"Using features {selection}")
|
| 794 |
X = X[:, selection]
|
| 795 |
|
| 796 |
if use_custom_variable_names:
|
| 797 |
+
variable_names = [
|
| 798 |
+
variable_names[selection[i]] for i in range(len(selection))
|
| 799 |
+
]
|
| 800 |
else:
|
| 801 |
selection = None
|
| 802 |
return X, variable_names, selection
|
|
|
|
| 807 |
pkg_directory = Path(__file__).parents[1]
|
| 808 |
default_project_file = pkg_directory / "Project.toml"
|
| 809 |
tmpdir = Path(tempfile.mkdtemp(dir=tempdir))
|
| 810 |
+
hyperparam_filename = tmpdir / f"hyperparams.jl"
|
| 811 |
+
dataset_filename = tmpdir / f"dataset.jl"
|
| 812 |
+
runfile_filename = tmpdir / f"runfile.jl"
|
| 813 |
X_filename = tmpdir / "X.csv"
|
| 814 |
y_filename = tmpdir / "y.csv"
|
| 815 |
weights_filename = tmpdir / "weights.csv"
|
| 816 |
+
return dict(
|
| 817 |
+
pkg_directory=pkg_directory,
|
| 818 |
+
default_project_file=default_project_file,
|
| 819 |
+
X_filename=X_filename,
|
| 820 |
+
dataset_filename=dataset_filename,
|
| 821 |
+
hyperparam_filename=hyperparam_filename,
|
| 822 |
+
runfile_filename=runfile_filename,
|
| 823 |
+
tmpdir=tmpdir,
|
| 824 |
+
weights_filename=weights_filename,
|
| 825 |
+
y_filename=y_filename,
|
| 826 |
+
)
|
| 827 |
+
|
| 828 |
+
|
| 829 |
+
def _check_assertions(
|
| 830 |
+
X,
|
| 831 |
+
binary_operators,
|
| 832 |
+
unary_operators,
|
| 833 |
+
use_custom_variable_names,
|
| 834 |
+
variable_names,
|
| 835 |
+
weights,
|
| 836 |
+
y,
|
| 837 |
+
):
|
| 838 |
# Check for potential errors before they happen
|
| 839 |
assert len(unary_operators) + len(binary_operators) > 0
|
| 840 |
assert len(X.shape) == 2
|
|
|
|
| 846 |
if use_custom_variable_names:
|
| 847 |
assert len(variable_names) == X.shape[1]
|
| 848 |
|
| 849 |
+
|
| 850 |
def _check_for_julia_installation():
|
| 851 |
try:
|
| 852 |
process = subprocess.Popen(["julia", "-v"], stdout=subprocess.PIPE, bufsize=-1)
|
| 853 |
while True:
|
| 854 |
line = process.stdout.readline()
|
| 855 |
+
if not line:
|
| 856 |
+
break
|
| 857 |
process.stdout.close()
|
| 858 |
process.wait()
|
| 859 |
except FileNotFoundError:
|
| 860 |
import os
|
| 861 |
+
|
| 862 |
+
raise RuntimeError(
|
| 863 |
+
f"Your current $PATH is: {os.environ['PATH']}\nPySR could not start julia. Make sure julia is installed and on your $PATH."
|
| 864 |
+
)
|
| 865 |
process.kill()
|
| 866 |
|
| 867 |
|
| 868 |
def run_feature_selection(X, y, select_k_features):
|
| 869 |
"""Use a gradient boosting tree regressor as a proxy for finding
|
| 870 |
+
the k most important features in X, returning indices for those
|
| 871 |
+
features as output."""
|
| 872 |
|
| 873 |
from sklearn.ensemble import RandomForestRegressor
|
| 874 |
from sklearn.feature_selection import SelectFromModel, SelectKBest
|
| 875 |
|
| 876 |
clf = RandomForestRegressor(n_estimators=100, max_depth=3, random_state=0)
|
| 877 |
clf.fit(X, y)
|
| 878 |
+
selector = SelectFromModel(
|
| 879 |
+
clf, threshold=-np.inf, max_features=select_k_features, prefit=True
|
| 880 |
+
)
|
| 881 |
return selector.get_support(indices=True)
|
| 882 |
|
| 883 |
+
|
| 884 |
+
def get_hof(
|
| 885 |
+
equation_file=None,
|
| 886 |
+
n_features=None,
|
| 887 |
+
variable_names=None,
|
| 888 |
+
output_jax_format=None,
|
| 889 |
+
output_torch_format=None,
|
| 890 |
+
selection=None,
|
| 891 |
+
extra_sympy_mappings=None,
|
| 892 |
+
extra_jax_mappings=None,
|
| 893 |
+
extra_torch_mappings=None,
|
| 894 |
+
multioutput=None,
|
| 895 |
+
nout=None,
|
| 896 |
+
**kwargs,
|
| 897 |
+
):
|
| 898 |
"""Get the equations from a hall of fame file. If no arguments
|
| 899 |
entered, the ones used previously from a call to PySR will be used."""
|
| 900 |
|
| 901 |
global global_state
|
| 902 |
|
| 903 |
+
if equation_file is None:
|
| 904 |
+
equation_file = global_state["equation_file"]
|
| 905 |
+
if n_features is None:
|
| 906 |
+
n_features = global_state["n_features"]
|
| 907 |
+
if variable_names is None:
|
| 908 |
+
variable_names = global_state["variable_names"]
|
| 909 |
+
if extra_sympy_mappings is None:
|
| 910 |
+
extra_sympy_mappings = global_state["extra_sympy_mappings"]
|
| 911 |
+
if extra_jax_mappings is None:
|
| 912 |
+
extra_jax_mappings = global_state["extra_jax_mappings"]
|
| 913 |
+
if extra_torch_mappings is None:
|
| 914 |
+
extra_torch_mappings = global_state["extra_torch_mappings"]
|
| 915 |
+
if output_torch_format is None:
|
| 916 |
+
output_torch_format = global_state["output_torch_format"]
|
| 917 |
+
if output_jax_format is None:
|
| 918 |
+
output_jax_format = global_state["output_jax_format"]
|
| 919 |
+
if multioutput is None:
|
| 920 |
+
multioutput = global_state["multioutput"]
|
| 921 |
+
if nout is None:
|
| 922 |
+
nout = global_state["nout"]
|
| 923 |
+
if selection is None:
|
| 924 |
+
selection = global_state["selection"]
|
| 925 |
+
|
| 926 |
+
global_state["selection"] = selection
|
| 927 |
+
global_state["equation_file"] = equation_file
|
| 928 |
+
global_state["n_features"] = n_features
|
| 929 |
+
global_state["variable_names"] = variable_names
|
| 930 |
+
global_state["extra_sympy_mappings"] = extra_sympy_mappings
|
| 931 |
+
global_state["extra_jax_mappings"] = extra_jax_mappings
|
| 932 |
+
global_state["extra_torch_mappings"] = extra_torch_mappings
|
| 933 |
+
global_state["output_torch_format"] = output_torch_format
|
| 934 |
+
global_state["output_jax_format"] = output_jax_format
|
| 935 |
+
global_state["multioutput"] = multioutput
|
| 936 |
+
global_state["nout"] = nout
|
| 937 |
+
global_state["selection"] = selection
|
| 938 |
|
| 939 |
try:
|
| 940 |
if multioutput:
|
| 941 |
+
all_outputs = [
|
| 942 |
+
pd.read_csv(str(equation_file) + f".out{i}" + ".bkup", sep="|")
|
| 943 |
+
for i in range(1, nout + 1)
|
| 944 |
+
]
|
| 945 |
else:
|
| 946 |
+
all_outputs = [pd.read_csv(str(equation_file) + ".bkup", sep="|")]
|
| 947 |
except FileNotFoundError:
|
| 948 |
+
raise RuntimeError(
|
| 949 |
+
"Couldn't find equation file! The equation search likely exited before a single iteration completed."
|
| 950 |
+
)
|
| 951 |
|
| 952 |
ret_outputs = []
|
| 953 |
|
|
|
|
| 962 |
jax_format = []
|
| 963 |
if output_torch_format:
|
| 964 |
torch_format = []
|
| 965 |
+
use_custom_variable_names = len(variable_names) != 0
|
| 966 |
+
local_sympy_mappings = {**extra_sympy_mappings, **sympy_mappings}
|
|
|
|
|
|
|
|
|
|
| 967 |
|
| 968 |
if use_custom_variable_names:
|
| 969 |
sympy_symbols = [sympy.Symbol(variable_names[i]) for i in range(n_features)]
|
| 970 |
else:
|
| 971 |
+
sympy_symbols = [sympy.Symbol("x%d" % i) for i in range(n_features)]
|
| 972 |
|
| 973 |
for i in range(len(output)):
|
| 974 |
+
eqn = sympify(output.loc[i, "Equation"], locals=local_sympy_mappings)
|
| 975 |
sympy_format.append(eqn)
|
| 976 |
|
| 977 |
# Numpy:
|
|
|
|
| 980 |
# JAX:
|
| 981 |
if output_jax_format:
|
| 982 |
from .export_jax import sympy2jax
|
| 983 |
+
|
| 984 |
func, params = sympy2jax(eqn, sympy_symbols, selection)
|
| 985 |
+
jax_format.append({"callable": func, "parameters": params})
|
| 986 |
|
| 987 |
# Torch:
|
| 988 |
if output_torch_format:
|
| 989 |
from .export_torch import sympy2torch
|
| 990 |
+
|
| 991 |
module = sympy2torch(eqn, sympy_symbols, selection=selection)
|
| 992 |
torch_format.append(module)
|
| 993 |
|
| 994 |
+
curMSE = output.loc[i, "MSE"]
|
| 995 |
+
curComplexity = output.loc[i, "Complexity"]
|
| 996 |
|
| 997 |
if lastMSE is None:
|
| 998 |
cur_score = 0.0
|
| 999 |
else:
|
| 1000 |
+
cur_score = -np.log(curMSE / lastMSE) / (curComplexity - lastComplexity)
|
| 1001 |
|
| 1002 |
scores.append(cur_score)
|
| 1003 |
lastMSE = curMSE
|
| 1004 |
lastComplexity = curComplexity
|
| 1005 |
|
| 1006 |
+
output["score"] = np.array(scores)
|
| 1007 |
+
output["sympy_format"] = sympy_format
|
| 1008 |
+
output["lambda_format"] = lambda_format
|
| 1009 |
+
output_cols = [
|
| 1010 |
+
"Complexity",
|
| 1011 |
+
"MSE",
|
| 1012 |
+
"score",
|
| 1013 |
+
"Equation",
|
| 1014 |
+
"sympy_format",
|
| 1015 |
+
"lambda_format",
|
| 1016 |
+
]
|
| 1017 |
if output_jax_format:
|
| 1018 |
+
output_cols += ["jax_format"]
|
| 1019 |
+
output["jax_format"] = jax_format
|
| 1020 |
if output_torch_format:
|
| 1021 |
+
output_cols += ["torch_format"]
|
| 1022 |
+
output["torch_format"] = torch_format
|
| 1023 |
|
| 1024 |
ret_outputs.append(output[output_cols])
|
| 1025 |
|
|
|
|
| 1028 |
else:
|
| 1029 |
return ret_outputs[0]
|
| 1030 |
|
| 1031 |
+
|
| 1032 |
def best_row(equations=None):
|
| 1033 |
"""Return the best row of a hall of fame file using the score column.
|
| 1034 |
By default this uses the last equation file.
|
| 1035 |
"""
|
| 1036 |
+
if equations is None:
|
| 1037 |
+
equations = get_hof()
|
| 1038 |
if isinstance(equations, list):
|
| 1039 |
+
return [eq.iloc[np.argmax(eq["score"])] for eq in equations]
|
| 1040 |
else:
|
| 1041 |
+
return equations.iloc[np.argmax(equations["score"])]
|
| 1042 |
+
|
| 1043 |
|
| 1044 |
def best_tex(equations=None):
|
| 1045 |
"""Return the equation with the best score, in latex format
|
| 1046 |
By default this uses the last equation file.
|
| 1047 |
"""
|
| 1048 |
+
if equations is None:
|
| 1049 |
+
equations = get_hof()
|
| 1050 |
if isinstance(equations, list):
|
| 1051 |
+
return [
|
| 1052 |
+
sympy.latex(best_row(eq)["sympy_format"].simplify()) for eq in equations
|
| 1053 |
+
]
|
| 1054 |
else:
|
| 1055 |
+
return sympy.latex(best_row(equations)["sympy_format"].simplify())
|
| 1056 |
+
|
| 1057 |
|
| 1058 |
def best(equations=None):
|
| 1059 |
"""Return the equation with the best score, in sympy format.
|
| 1060 |
By default this uses the last equation file.
|
| 1061 |
"""
|
| 1062 |
+
if equations is None:
|
| 1063 |
+
equations = get_hof()
|
| 1064 |
if isinstance(equations, list):
|
| 1065 |
+
return [best_row(eq)["sympy_format"].simplify() for eq in equations]
|
| 1066 |
else:
|
| 1067 |
+
return best_row(equations)["sympy_format"].simplify()
|
| 1068 |
+
|
| 1069 |
|
| 1070 |
def best_callable(equations=None):
|
| 1071 |
"""Return the equation with the best score, in callable format.
|
| 1072 |
By default this uses the last equation file.
|
| 1073 |
"""
|
| 1074 |
+
if equations is None:
|
| 1075 |
+
equations = get_hof()
|
| 1076 |
if isinstance(equations, list):
|
| 1077 |
+
return [best_row(eq)["lambda_format"] for eq in equations]
|
| 1078 |
else:
|
| 1079 |
+
return best_row(equations)["lambda_format"]
|
| 1080 |
+
|
| 1081 |
|
| 1082 |
def _escape_filename(filename):
|
| 1083 |
"""Turns a file into a string representation with correctly escaped backslashes"""
|
| 1084 |
repr = str(filename)
|
| 1085 |
+
repr = repr.replace("\\", "\\\\")
|
| 1086 |
return repr
|
| 1087 |
|
| 1088 |
+
|
| 1089 |
# https://gist.github.com/garrettdreyfus/8153571
|
| 1090 |
def _yesno(question):
|
| 1091 |
"""Simple Yes/No Function."""
|
| 1092 |
+
prompt = f"{question} (y/n): "
|
| 1093 |
ans = input(prompt).strip().lower()
|
| 1094 |
+
if ans not in ["y", "n"]:
|
| 1095 |
+
print(f"{ans} is invalid, please try again...")
|
| 1096 |
return _yesno(question)
|
| 1097 |
+
if ans == "y":
|
| 1098 |
return True
|
| 1099 |
return False
|
| 1100 |
|
| 1101 |
|
| 1102 |
class CallableEquation(object):
|
| 1103 |
"""Simple wrapper for numpy lambda functions built with sympy"""
|
| 1104 |
+
|
| 1105 |
def __init__(self, sympy_symbols, eqn, selection=None):
|
| 1106 |
self._sympy = eqn
|
| 1107 |
self._sympy_symbols = sympy_symbols
|
|
|
|
| 1116 |
return self._lambda(*X[:, self._selection].T)
|
| 1117 |
else:
|
| 1118 |
return self._lambda(*X.T)
|
|
|
setup.py
CHANGED
|
@@ -12,19 +12,13 @@ setuptools.setup(
|
|
| 12 |
long_description=long_description,
|
| 13 |
long_description_content_type="text/markdown",
|
| 14 |
url="https://github.com/MilesCranmer/pysr",
|
| 15 |
-
install_requires=[
|
| 16 |
-
"numpy",
|
| 17 |
-
"pandas",
|
| 18 |
-
"sympy"
|
| 19 |
-
],
|
| 20 |
packages=setuptools.find_packages(),
|
| 21 |
-
package_data={
|
| 22 |
-
'pysr': ['../Project.toml', '../datasets/*']
|
| 23 |
-
},
|
| 24 |
include_package_data=False,
|
| 25 |
classifiers=[
|
| 26 |
"Programming Language :: Python :: 3",
|
| 27 |
"Operating System :: OS Independent",
|
| 28 |
],
|
| 29 |
-
python_requires=
|
| 30 |
)
|
|
|
|
| 12 |
long_description=long_description,
|
| 13 |
long_description_content_type="text/markdown",
|
| 14 |
url="https://github.com/MilesCranmer/pysr",
|
| 15 |
+
install_requires=["numpy", "pandas", "sympy"],
|
|
|
|
|
|
|
|
|
|
|
|
|
| 16 |
packages=setuptools.find_packages(),
|
| 17 |
+
package_data={"pysr": ["../Project.toml", "../datasets/*"]},
|
|
|
|
|
|
|
| 18 |
include_package_data=False,
|
| 19 |
classifiers=[
|
| 20 |
"Programming Language :: Python :: 3",
|
| 21 |
"Operating System :: OS Independent",
|
| 22 |
],
|
| 23 |
+
python_requires=">=3.7",
|
| 24 |
)
|
test/test.py
CHANGED
|
@@ -6,6 +6,7 @@ import sympy
|
|
| 6 |
from sympy import lambdify
|
| 7 |
import pandas as pd
|
| 8 |
|
|
|
|
| 9 |
class TestPipeline(unittest.TestCase):
|
| 10 |
def setUp(self):
|
| 11 |
self.default_test_kwargs = dict(
|
|
@@ -17,86 +18,105 @@ class TestPipeline(unittest.TestCase):
|
|
| 17 |
)
|
| 18 |
np.random.seed(0)
|
| 19 |
self.X = np.random.randn(100, 5)
|
| 20 |
-
|
| 21 |
def test_linear_relation(self):
|
| 22 |
y = self.X[:, 0]
|
| 23 |
equations = pysr(self.X, y, **self.default_test_kwargs)
|
| 24 |
print(equations)
|
| 25 |
-
self.assertLessEqual(equations.iloc[-1][
|
| 26 |
|
| 27 |
def test_multioutput_custom_operator(self):
|
| 28 |
-
y = self.X[:, [0, 1]]**2
|
| 29 |
-
equations = pysr(
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 34 |
print(equations)
|
| 35 |
-
self.assertLessEqual(equations[0].iloc[-1][
|
| 36 |
-
self.assertLessEqual(equations[1].iloc[-1][
|
| 37 |
|
| 38 |
def test_multioutput_weighted_with_callable(self):
|
| 39 |
-
y = self.X[:, [0, 1]]**2
|
| 40 |
w = np.random.rand(*y.shape)
|
| 41 |
w[w < 0.5] = 0.0
|
| 42 |
w[w >= 0.5] = 1.0
|
| 43 |
|
| 44 |
# Double equation when weights are 0:
|
| 45 |
-
y += (1-w) * y
|
| 46 |
# Thus, pysr needs to use the weights to find the right equation!
|
| 47 |
|
| 48 |
-
equations = pysr(
|
| 49 |
-
|
| 50 |
-
|
| 51 |
-
|
| 52 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 53 |
|
| 54 |
np.testing.assert_almost_equal(
|
| 55 |
-
|
| 56 |
-
|
| 57 |
-
decimal=4)
|
| 58 |
np.testing.assert_almost_equal(
|
| 59 |
-
|
| 60 |
-
|
| 61 |
-
decimal=4)
|
| 62 |
|
| 63 |
def test_empty_operators_single_input(self):
|
| 64 |
X = np.random.randn(100, 1)
|
| 65 |
y = X[:, 0] + 3.0
|
| 66 |
-
equations = pysr(
|
| 67 |
-
|
| 68 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 69 |
|
| 70 |
-
self.assertLessEqual(equations.iloc[-1]['MSE'], 1e-4)
|
| 71 |
|
| 72 |
class TestBest(unittest.TestCase):
|
| 73 |
def setUp(self):
|
| 74 |
-
equations = pd.DataFrame(
|
| 75 |
-
|
| 76 |
-
|
| 77 |
-
|
| 78 |
-
|
|
|
|
|
|
|
| 79 |
|
| 80 |
-
equations[
|
| 81 |
-
|
|
|
|
| 82 |
|
| 83 |
self.equations = get_hof(
|
| 84 |
-
|
| 85 |
-
|
| 86 |
-
|
| 87 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 88 |
|
| 89 |
def test_best(self):
|
| 90 |
-
self.assertEqual(best(self.equations), sympy.cos(sympy.Symbol(
|
| 91 |
-
self.assertEqual(best(), sympy.cos(sympy.Symbol(
|
| 92 |
|
| 93 |
def test_best_tex(self):
|
| 94 |
-
self.assertEqual(best_tex(self.equations),
|
| 95 |
-
self.assertEqual(best_tex(),
|
| 96 |
|
| 97 |
def test_best_lambda(self):
|
| 98 |
X = np.random.randn(10, 2)
|
| 99 |
-
y = np.cos(X[:, 0])**2
|
| 100 |
for f in [best_callable(), best_callable(self.equations)]:
|
| 101 |
np.testing.assert_almost_equal(f(X), y, decimal=4)
|
| 102 |
|
|
@@ -107,22 +127,23 @@ class TestFeatureSelection(unittest.TestCase):
|
|
| 107 |
|
| 108 |
def test_feature_selection(self):
|
| 109 |
X = np.random.randn(20000, 5)
|
| 110 |
-
y = X[:, 2]**2 + X[:, 3]**2
|
| 111 |
selected = run_feature_selection(X, y, select_k_features=2)
|
| 112 |
self.assertEqual(sorted(selected), [2, 3])
|
| 113 |
|
| 114 |
def test_feature_selection_handler(self):
|
| 115 |
X = np.random.randn(20000, 5)
|
| 116 |
-
y = X[:, 2]**2 + X[:, 3]**2
|
| 117 |
-
var_names = [f
|
| 118 |
selected_X, selected_var_names, selection = _handle_feature_selection(
|
| 119 |
-
|
| 120 |
-
|
| 121 |
-
|
| 122 |
-
|
|
|
|
|
|
|
| 123 |
self.assertTrue((2 in selection) and (3 in selection))
|
| 124 |
-
self.assertEqual(set(selected_var_names), set(
|
| 125 |
np.testing.assert_array_equal(
|
| 126 |
-
|
| 127 |
-
|
| 128 |
-
)
|
|
|
|
| 6 |
from sympy import lambdify
|
| 7 |
import pandas as pd
|
| 8 |
|
| 9 |
+
|
| 10 |
class TestPipeline(unittest.TestCase):
|
| 11 |
def setUp(self):
|
| 12 |
self.default_test_kwargs = dict(
|
|
|
|
| 18 |
)
|
| 19 |
np.random.seed(0)
|
| 20 |
self.X = np.random.randn(100, 5)
|
| 21 |
+
|
| 22 |
def test_linear_relation(self):
|
| 23 |
y = self.X[:, 0]
|
| 24 |
equations = pysr(self.X, y, **self.default_test_kwargs)
|
| 25 |
print(equations)
|
| 26 |
+
self.assertLessEqual(equations.iloc[-1]["MSE"], 1e-4)
|
| 27 |
|
| 28 |
def test_multioutput_custom_operator(self):
|
| 29 |
+
y = self.X[:, [0, 1]] ** 2
|
| 30 |
+
equations = pysr(
|
| 31 |
+
self.X,
|
| 32 |
+
y,
|
| 33 |
+
unary_operators=["sq(x) = x^2"],
|
| 34 |
+
binary_operators=["plus"],
|
| 35 |
+
extra_sympy_mappings={"sq": lambda x: x ** 2},
|
| 36 |
+
**self.default_test_kwargs,
|
| 37 |
+
procs=0,
|
| 38 |
+
)
|
| 39 |
print(equations)
|
| 40 |
+
self.assertLessEqual(equations[0].iloc[-1]["MSE"], 1e-4)
|
| 41 |
+
self.assertLessEqual(equations[1].iloc[-1]["MSE"], 1e-4)
|
| 42 |
|
| 43 |
def test_multioutput_weighted_with_callable(self):
|
| 44 |
+
y = self.X[:, [0, 1]] ** 2
|
| 45 |
w = np.random.rand(*y.shape)
|
| 46 |
w[w < 0.5] = 0.0
|
| 47 |
w[w >= 0.5] = 1.0
|
| 48 |
|
| 49 |
# Double equation when weights are 0:
|
| 50 |
+
y += (1 - w) * y
|
| 51 |
# Thus, pysr needs to use the weights to find the right equation!
|
| 52 |
|
| 53 |
+
equations = pysr(
|
| 54 |
+
self.X,
|
| 55 |
+
y,
|
| 56 |
+
weights=w,
|
| 57 |
+
unary_operators=["sq(x) = x^2"],
|
| 58 |
+
binary_operators=["plus"],
|
| 59 |
+
extra_sympy_mappings={"sq": lambda x: x ** 2},
|
| 60 |
+
**self.default_test_kwargs,
|
| 61 |
+
procs=0,
|
| 62 |
+
)
|
| 63 |
|
| 64 |
np.testing.assert_almost_equal(
|
| 65 |
+
best_callable()[0](self.X), self.X[:, 0] ** 2, decimal=4
|
| 66 |
+
)
|
|
|
|
| 67 |
np.testing.assert_almost_equal(
|
| 68 |
+
best_callable()[1](self.X), self.X[:, 1] ** 2, decimal=4
|
| 69 |
+
)
|
|
|
|
| 70 |
|
| 71 |
def test_empty_operators_single_input(self):
|
| 72 |
X = np.random.randn(100, 1)
|
| 73 |
y = X[:, 0] + 3.0
|
| 74 |
+
equations = pysr(
|
| 75 |
+
X,
|
| 76 |
+
y,
|
| 77 |
+
unary_operators=[],
|
| 78 |
+
binary_operators=["plus"],
|
| 79 |
+
**self.default_test_kwargs,
|
| 80 |
+
)
|
| 81 |
+
|
| 82 |
+
self.assertLessEqual(equations.iloc[-1]["MSE"], 1e-4)
|
| 83 |
|
|
|
|
| 84 |
|
| 85 |
class TestBest(unittest.TestCase):
|
| 86 |
def setUp(self):
|
| 87 |
+
equations = pd.DataFrame(
|
| 88 |
+
{
|
| 89 |
+
"Equation": ["1.0", "cos(x0)", "square(cos(x0))"],
|
| 90 |
+
"MSE": [1.0, 0.1, 1e-5],
|
| 91 |
+
"Complexity": [1, 2, 3],
|
| 92 |
+
}
|
| 93 |
+
)
|
| 94 |
|
| 95 |
+
equations["Complexity MSE Equation".split(" ")].to_csv(
|
| 96 |
+
"equation_file.csv.bkup", sep="|"
|
| 97 |
+
)
|
| 98 |
|
| 99 |
self.equations = get_hof(
|
| 100 |
+
"equation_file.csv",
|
| 101 |
+
n_features=2,
|
| 102 |
+
variables_names="x0 x1".split(" "),
|
| 103 |
+
extra_sympy_mappings={},
|
| 104 |
+
output_jax_format=False,
|
| 105 |
+
multioutput=False,
|
| 106 |
+
nout=1,
|
| 107 |
+
)
|
| 108 |
|
| 109 |
def test_best(self):
|
| 110 |
+
self.assertEqual(best(self.equations), sympy.cos(sympy.Symbol("x0")) ** 2)
|
| 111 |
+
self.assertEqual(best(), sympy.cos(sympy.Symbol("x0")) ** 2)
|
| 112 |
|
| 113 |
def test_best_tex(self):
|
| 114 |
+
self.assertEqual(best_tex(self.equations), "\\cos^{2}{\\left(x_{0} \\right)}")
|
| 115 |
+
self.assertEqual(best_tex(), "\\cos^{2}{\\left(x_{0} \\right)}")
|
| 116 |
|
| 117 |
def test_best_lambda(self):
|
| 118 |
X = np.random.randn(10, 2)
|
| 119 |
+
y = np.cos(X[:, 0]) ** 2
|
| 120 |
for f in [best_callable(), best_callable(self.equations)]:
|
| 121 |
np.testing.assert_almost_equal(f(X), y, decimal=4)
|
| 122 |
|
|
|
|
| 127 |
|
| 128 |
def test_feature_selection(self):
|
| 129 |
X = np.random.randn(20000, 5)
|
| 130 |
+
y = X[:, 2] ** 2 + X[:, 3] ** 2
|
| 131 |
selected = run_feature_selection(X, y, select_k_features=2)
|
| 132 |
self.assertEqual(sorted(selected), [2, 3])
|
| 133 |
|
| 134 |
def test_feature_selection_handler(self):
|
| 135 |
X = np.random.randn(20000, 5)
|
| 136 |
+
y = X[:, 2] ** 2 + X[:, 3] ** 2
|
| 137 |
+
var_names = [f"x{i}" for i in range(5)]
|
| 138 |
selected_X, selected_var_names, selection = _handle_feature_selection(
|
| 139 |
+
X,
|
| 140 |
+
select_k_features=2,
|
| 141 |
+
use_custom_variable_names=True,
|
| 142 |
+
variable_names=[f"x{i}" for i in range(5)],
|
| 143 |
+
y=y,
|
| 144 |
+
)
|
| 145 |
self.assertTrue((2 in selection) and (3 in selection))
|
| 146 |
+
self.assertEqual(set(selected_var_names), set("x2 x3".split(" ")))
|
| 147 |
np.testing.assert_array_equal(
|
| 148 |
+
np.sort(selected_X, axis=1), np.sort(X[:, [2, 3]], axis=1)
|
| 149 |
+
)
|
|
|
test/test_jax.py
CHANGED
|
@@ -7,37 +7,48 @@ from jax import random
|
|
| 7 |
from jax import grad
|
| 8 |
import sympy
|
| 9 |
|
|
|
|
| 10 |
class TestJAX(unittest.TestCase):
|
| 11 |
def setUp(self):
|
| 12 |
np.random.seed(0)
|
| 13 |
|
| 14 |
def test_sympy2jax(self):
|
| 15 |
-
x, y, z = sympy.symbols(
|
| 16 |
cosx = 1.0 * sympy.cos(x) + y
|
| 17 |
key = random.PRNGKey(0)
|
| 18 |
X = random.normal(key, (1000, 2))
|
| 19 |
true = 1.0 * jnp.cos(X[:, 0]) + X[:, 1]
|
| 20 |
f, params = sympy2jax(cosx, [x, y, z])
|
| 21 |
self.assertTrue(jnp.all(jnp.isclose(f(X, params), true)).item())
|
|
|
|
| 22 |
def test_pipeline(self):
|
| 23 |
X = np.random.randn(100, 10)
|
| 24 |
-
equations = pd.DataFrame(
|
| 25 |
-
|
| 26 |
-
|
| 27 |
-
|
| 28 |
-
|
|
|
|
|
|
|
| 29 |
|
| 30 |
-
equations[
|
| 31 |
-
|
|
|
|
| 32 |
|
| 33 |
equations = get_hof(
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 37 |
|
| 38 |
jformat = equations.iloc[-1].jax_format
|
| 39 |
np.testing.assert_almost_equal(
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
|
| 43 |
)
|
|
|
|
| 7 |
from jax import grad
|
| 8 |
import sympy
|
| 9 |
|
| 10 |
+
|
| 11 |
class TestJAX(unittest.TestCase):
|
| 12 |
def setUp(self):
|
| 13 |
np.random.seed(0)
|
| 14 |
|
| 15 |
def test_sympy2jax(self):
|
| 16 |
+
x, y, z = sympy.symbols("x y z")
|
| 17 |
cosx = 1.0 * sympy.cos(x) + y
|
| 18 |
key = random.PRNGKey(0)
|
| 19 |
X = random.normal(key, (1000, 2))
|
| 20 |
true = 1.0 * jnp.cos(X[:, 0]) + X[:, 1]
|
| 21 |
f, params = sympy2jax(cosx, [x, y, z])
|
| 22 |
self.assertTrue(jnp.all(jnp.isclose(f(X, params), true)).item())
|
| 23 |
+
|
| 24 |
def test_pipeline(self):
|
| 25 |
X = np.random.randn(100, 10)
|
| 26 |
+
equations = pd.DataFrame(
|
| 27 |
+
{
|
| 28 |
+
"Equation": ["1.0", "cos(x0)", "square(cos(x0))"],
|
| 29 |
+
"MSE": [1.0, 0.1, 1e-5],
|
| 30 |
+
"Complexity": [1, 2, 3],
|
| 31 |
+
}
|
| 32 |
+
)
|
| 33 |
|
| 34 |
+
equations["Complexity MSE Equation".split(" ")].to_csv(
|
| 35 |
+
"equation_file.csv.bkup", sep="|"
|
| 36 |
+
)
|
| 37 |
|
| 38 |
equations = get_hof(
|
| 39 |
+
"equation_file.csv",
|
| 40 |
+
n_features=2,
|
| 41 |
+
variables_names="x1 x2 x3".split(" "),
|
| 42 |
+
extra_sympy_mappings={},
|
| 43 |
+
output_jax_format=True,
|
| 44 |
+
multioutput=False,
|
| 45 |
+
nout=1,
|
| 46 |
+
selection=[1, 2, 3],
|
| 47 |
+
)
|
| 48 |
|
| 49 |
jformat = equations.iloc[-1].jax_format
|
| 50 |
np.testing.assert_almost_equal(
|
| 51 |
+
np.array(jformat["callable"](jnp.array(X), jformat["parameters"])),
|
| 52 |
+
np.square(np.cos(X[:, 1])), # Select feature 1
|
| 53 |
+
decimal=4,
|
| 54 |
)
|
test/test_torch.py
CHANGED
|
@@ -5,38 +5,49 @@ from pysr import sympy2torch, get_hof
|
|
| 5 |
import torch
|
| 6 |
import sympy
|
| 7 |
|
|
|
|
| 8 |
class TestTorch(unittest.TestCase):
|
| 9 |
def setUp(self):
|
| 10 |
np.random.seed(0)
|
| 11 |
|
| 12 |
def test_sympy2torch(self):
|
| 13 |
-
x, y, z = sympy.symbols(
|
| 14 |
cosx = 1.0 * sympy.cos(x) + y
|
| 15 |
X = torch.tensor(np.random.randn(1000, 3))
|
| 16 |
true = 1.0 * torch.cos(X[:, 0]) + X[:, 1]
|
| 17 |
torch_module = sympy2torch(cosx, [x, y, z])
|
| 18 |
self.assertTrue(
|
| 19 |
-
|
| 20 |
)
|
|
|
|
| 21 |
def test_pipeline(self):
|
| 22 |
X = np.random.randn(100, 10)
|
| 23 |
-
equations = pd.DataFrame(
|
| 24 |
-
|
| 25 |
-
|
| 26 |
-
|
| 27 |
-
|
|
|
|
|
|
|
| 28 |
|
| 29 |
-
equations[
|
| 30 |
-
|
|
|
|
| 31 |
|
| 32 |
equations = get_hof(
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 36 |
|
| 37 |
tformat = equations.iloc[-1].torch_format
|
| 38 |
np.testing.assert_almost_equal(
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
|
| 42 |
)
|
|
|
|
| 5 |
import torch
|
| 6 |
import sympy
|
| 7 |
|
| 8 |
+
|
| 9 |
class TestTorch(unittest.TestCase):
|
| 10 |
def setUp(self):
|
| 11 |
np.random.seed(0)
|
| 12 |
|
| 13 |
def test_sympy2torch(self):
|
| 14 |
+
x, y, z = sympy.symbols("x y z")
|
| 15 |
cosx = 1.0 * sympy.cos(x) + y
|
| 16 |
X = torch.tensor(np.random.randn(1000, 3))
|
| 17 |
true = 1.0 * torch.cos(X[:, 0]) + X[:, 1]
|
| 18 |
torch_module = sympy2torch(cosx, [x, y, z])
|
| 19 |
self.assertTrue(
|
| 20 |
+
np.all(np.isclose(torch_module(X).detach().numpy(), true.detach().numpy()))
|
| 21 |
)
|
| 22 |
+
|
| 23 |
def test_pipeline(self):
|
| 24 |
X = np.random.randn(100, 10)
|
| 25 |
+
equations = pd.DataFrame(
|
| 26 |
+
{
|
| 27 |
+
"Equation": ["1.0", "cos(x0)", "square(cos(x0))"],
|
| 28 |
+
"MSE": [1.0, 0.1, 1e-5],
|
| 29 |
+
"Complexity": [1, 2, 3],
|
| 30 |
+
}
|
| 31 |
+
)
|
| 32 |
|
| 33 |
+
equations["Complexity MSE Equation".split(" ")].to_csv(
|
| 34 |
+
"equation_file.csv.bkup", sep="|"
|
| 35 |
+
)
|
| 36 |
|
| 37 |
equations = get_hof(
|
| 38 |
+
"equation_file.csv",
|
| 39 |
+
n_features=2,
|
| 40 |
+
variables_names="x1 x2 x3".split(" "),
|
| 41 |
+
extra_sympy_mappings={},
|
| 42 |
+
output_torch_format=True,
|
| 43 |
+
multioutput=False,
|
| 44 |
+
nout=1,
|
| 45 |
+
selection=[1, 2, 3],
|
| 46 |
+
)
|
| 47 |
|
| 48 |
tformat = equations.iloc[-1].torch_format
|
| 49 |
np.testing.assert_almost_equal(
|
| 50 |
+
tformat(torch.tensor(X)).detach().numpy(),
|
| 51 |
+
np.square(np.cos(X[:, 1])), # Selection 1st feature
|
| 52 |
+
decimal=4,
|
| 53 |
)
|