sgbaird's picture
add c3_2 categorical option
5b742f8 verified
raw
history blame
11.2 kB
import numpy as np
import gradio as gr
import pandas as pd
from sklearn.preprocessing import MinMaxScaler
from surrogate import CrabNetSurrogateModel, PARAM_BOUNDS
from pydantic import (
BaseModel,
ValidationError,
ValidationInfo,
field_validator,
model_validator,
)
model = CrabNetSurrogateModel()
# Define the input parameters
example_parameterization = {
"N": 3,
"alpha": 0.5,
"d_model": 512,
"dim_feedforward": 2048,
"dropout": 0.1,
"emb_scaler": 0.5,
"epochs_step": 10,
"eps": 0.000001,
"fudge": 0.02,
"heads": 4,
"k": 6,
"lr": 0.001,
"pe_resolution": 5000,
"ple_resolution": 5000,
"pos_scaler": 0.5,
"weight_decay": 0,
"batch_size": 32,
"out_hidden4": 128,
"betas1": 0.9,
"betas2": 0.999,
"bias": False,
"criterion": "RobustL1",
"elem_prop": "mat2vec",
"train_frac": 0.5,
}
example_results = model.surrogate_evaluate([example_parameterization])
example_result = example_results[0]
# Initialize and fit scalers for each parameter
scalers = {}
for param_info in PARAM_BOUNDS:
if param_info["type"] == "range":
scaler = MinMaxScaler()
# Fit the scaler using the parameter bounds
scaler.fit([[bound] for bound in param_info["bounds"]])
scalers[param_info["name"]] = scaler
# HACK: Hardcoded
BLINDED_PARAM_BOUNDS = [
{"name": "x1", "type": "range", "bounds": [0.0, 1.0]},
{"name": "x2", "type": "range", "bounds": [0.0, 1.0]},
{"name": "x3", "type": "range", "bounds": [0.0, 1.0]},
{"name": "x4", "type": "range", "bounds": [0.0, 1.0]},
{"name": "x5", "type": "range", "bounds": [0.0, 1.0]},
{"name": "x6", "type": "range", "bounds": [0.0, 1.0]},
{"name": "x7", "type": "range", "bounds": [0.0, 1.0]},
{"name": "x8", "type": "range", "bounds": [0.0, 1.0]},
{"name": "x9", "type": "range", "bounds": [0.0, 1.0]},
{"name": "x10", "type": "range", "bounds": [0.0, 1.0]},
{"name": "x11", "type": "range", "bounds": [0.0, 1.0]},
{"name": "x12", "type": "range", "bounds": [0.0, 1.0]},
{"name": "x13", "type": "range", "bounds": [0.0, 1.0]},
{"name": "x14", "type": "range", "bounds": [0.0, 1.0]},
{"name": "x15", "type": "range", "bounds": [0.0, 1.0]},
{"name": "x16", "type": "range", "bounds": [0.0, 1.0]},
{"name": "x17", "type": "range", "bounds": [0.0, 1.0]},
{"name": "x18", "type": "range", "bounds": [0.0, 1.0]},
{"name": "x19", "type": "range", "bounds": [0.0, 1.0]},
{"name": "x20", "type": "range", "bounds": [0.0, 1.0]},
{"name": "c1", "type": "choice", "values": ["c1_0", "c1_1"]},
{"name": "c2", "type": "choice", "values": ["c2_0", "c2_1"]},
{"name": "c3", "type": "choice", "values": ["c3_0", "c3_1", "c3_2"]},
{"name": "fidelity1", "type": "range", "bounds": [0.0, 1.0]},
]
class BlindedParameterization(BaseModel):
x1: float # int
x2: float
x3: float # int
x4: float # int
x5: float
x6: float
x7: float # int
x8: float
x9: float
x10: float # int
x11: float # int
x12: float
x13: float # int
x14: float # int
x15: float
x16: float # int
x17: float # int
x18: float # int
x19: float
x20: float
c1: str # bool
c2: str
c3: str
fidelity1: float
@field_validator("*")
def check_bounds(cls, v: int, info: ValidationInfo) -> int:
param = next(
(item for item in BLINDED_PARAM_BOUNDS if item["name"] == info.field_name),
None,
)
if param is None:
return v
if param["type"] == "range":
min_val, max_val = param["bounds"]
if not min_val <= v <= max_val:
raise ValueError(
f"{info.field_name} must be between {min_val} and {max_val}"
)
elif param["type"] == "choice":
if v not in param["values"]:
raise ValueError(f"{info.field_name} must be one of {param['values']}")
return v
@model_validator(mode="after")
def check_constraints(self) -> "BlindedParameterization":
if self.x19 > self.x20:
raise ValueError(
f"Received x19={self.x19} which should be less than x20={self.x20}"
)
if self.x6 + self.x15 > 1.0:
raise ValueError(
f"Received x6={self.x6} and x15={self.x15} which should sum to less than or equal to 1.0" # noqa: E501
)
# Conversion from original to blinded representation
def convert_to_blinded(params):
blinded_params = {}
numeric_index = 1
choice_index = 1
for param in PARAM_BOUNDS:
if param["type"] == "range":
key = f"x{numeric_index}" if param["name"] != "train_frac" else "fidelity1"
blinded_params[key] = scalers[param["name"]].transform(
[[params[param["name"]]]]
)[0][0]
numeric_index += 1 if param["name"] != "train_frac" else 0
elif param["type"] == "choice":
key = f"c{choice_index}"
choice_index = param["values"].index(params[param["name"]])
blinded_params[key] = f"{key}_{choice_index}"
choice_index += 1
return blinded_params
# Conversion from blinded to original representation
def convert_from_blinded(blinded_params):
original_params = {}
numeric_index = 1
choice_index = 1
for param in PARAM_BOUNDS:
if param["type"] == "range":
key = f"x{numeric_index}" if param["name"] != "train_frac" else "fidelity1"
original_params[param["name"]] = scalers[param["name"]].inverse_transform(
[[blinded_params[key]]]
)[0][0]
numeric_index += 1 if param["name"] != "train_frac" else 0
elif param["type"] == "choice":
key = f"c{choice_index}"
choice_value = blinded_params[key].split("_")[-1]
original_params[param["name"]] = param["values"][int(choice_value)]
choice_index += 1
return original_params
def evaluate(*args):
# Assume args are in the order of BLINDED_PARAM_BOUNDS
blinded_params = dict(zip([param["name"] for param in BLINDED_PARAM_BOUNDS], args))
original_params = convert_from_blinded(blinded_params)
BlindedParameterization(**blinded_params) # Validation
params_list = [original_params]
results = model.surrogate_evaluate(params_list)
results_list = [list(result.values()) for result in results]
return results_list
def get_interface(param_info, numeric_index, choice_index):
key = param_info["name"]
default_value = example_parameterization[key]
if param_info["type"] == "range":
# Rescale the parameter to be between 0 and 1
scaler = scalers[key]
scaler.fit([[bound] for bound in param_info["bounds"]])
scaled_value = scaler.transform([[default_value]])[0][0]
scaled_bounds = scaler.transform([[bound] for bound in param_info["bounds"]])
label = f"fidelity1" if key == "train_frac" else f"x{numeric_index}"
return (
gr.Slider( # Change this line
value=scaled_value,
minimum=scaled_bounds[0][0],
maximum=scaled_bounds[1][0],
label=label,
step=(scaled_bounds[1][0] - scaled_bounds[0][0]) / 100,
),
numeric_index + 1,
choice_index,
)
elif param_info["type"] == "choice":
return (
gr.Radio(
choices=[
f"c{choice_index}_{i}" for i in range(len(param_info["values"]))
],
label=f"c{choice_index}",
value=f"c{choice_index}_{param_info['values'].index(default_value)}",
),
numeric_index,
choice_index + 1,
)
# test the evaluate function
blinded_results = evaluate(*[0.5] * 20, "c1_0", "c2_0", "c3_0", 0.5)
numeric_index = 1
choice_index = 1
inputs = []
for param in PARAM_BOUNDS:
input, numeric_index, choice_index = get_interface(
param, numeric_index, choice_index
)
inputs.append(input)
iface = gr.Interface(
title="Advanced Optimization",
fn=evaluate,
inputs=inputs,
outputs=gr.Numpy(
value=np.array([list(example_result.values())]),
headers=[f"y{i+1}" for i in range(len(example_result))],
col_count=(len(example_result), "fixed"),
datatype=["number"] * len(example_result),
),
description="""
## Objectives
**Minimize `y1`, `y2`, `y3`, and `y4`**
### Correlations
- `y1` and `y2` are correlated
- `y1` is anticorrelated with `y3`
- `y2` is anticorrelated with `y3`
### Noise
`y1`, `y2`, and `y3` are stochastic with heteroskedastic, parameter-free
noise, whereas `y4` is deterministic, but still considered 'black-box'. In
other words, repeat calls with the same input arguments will result in
different values for `y1`, `y2`, and `y3`, but the same value for `y4`.
### Objective thresholds
If `y1` is greater than 0.2, the result is considered "bad" no matter how
good the other values are. If `y2` is greater than 0.7, the result is
considered "bad" no matter how good the other values are. If `y3` is greater
than 1800, the result is considered "bad" no matter how good the other
values are. If `y4` is greater than 40e6, the result is considered "bad" no
matter how good the other values are.
## Search Space
### Fidelity
`fidelity1` is a fidelity parameter. The lowest fidelity is 0, and the
highest fidelity is 1. The higher the fidelity, the more expensive the
evaluation, and the higher the quality.
NOTE: `fidelity1` and `y3` are correlated.
### Constraints
- x<sub>19</sub> < x<sub>20</sub>
- x<sub>6</sub> + x<sub>15</sub> ≀ 1.0
### Parameter bounds
- 0 ≀ x<sub>i</sub> ≀ 1 for i ∈ {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13,
14, 15, 16, 17, 18, 19, 20}
- c<sub>1</sub> ∈ {c1_0, c1_1}
- c<sub>2</sub> ∈ {c2_0, c2_1}
- c<sub>3</sub> ∈ {c3_0, c3_1, c3_2}
- 0 ≀ fidelity1 ≀ 1
## Notion of best
Thresholded Pareto front hypervolume vs. running cost for three different
budgets, and averaged over 10 search campaigns.
## References:
1. Baird, S. G.; Liu, M.; Sparks, T. D. High-Dimensional Bayesian
Optimization of 23 Hyperparameters over 100 Iterations for an
Attention-Based Network to Predict Materials Property: A Case Study on
CrabNet Using Ax Platform and SAASBO. Computational Materials Science
2022, 211, 111505. https://doi.org/10.1016/j.commatsci.2022.111505.
2. Baird, S. G.; Parikh, J. N.; Sparks, T. D. Materials Science
Optimization Benchmark Dataset for High-Dimensional, Multi-Objective,
Multi-Fidelity Optimization of CrabNet Hyperparameters. ChemRxiv March
7, 2023. https://doi.org/10.26434/chemrxiv-2023-9s6r7.
""",
)
iface.launch(show_error=True)