File size: 6,277 Bytes
33355bf 8f34820 33355bf 8f34820 33355bf | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 | from __future__ import annotations
import json
import os
import random
import shutil
import sys
from importlib.metadata import version
from pathlib import Path
os.environ.setdefault("TF_CPP_MIN_LOG_LEVEL", "2")
os.environ.setdefault("TF_ENABLE_ONEDNN_OPTS", "0")
import numpy as np
import torch
from torch import nn
ROOT = Path(__file__).resolve().parent
sys.path.insert(0, str(ROOT / "src"))
from forge_lab.models.numpy_runtime import ( # noqa: E402
DEFAULT_CENTER,
DEFAULT_SCALE,
WINDOW_SIZE,
)
from forge_lab.models.pytorch_model import TinyDriftNet # noqa: E402
ARTIFACTS = ROOT
MODEL_VERSION = "forge-tiny-drift-v0.1"
def synthetic_windows(count: int = 2048, seed: int = 17) -> tuple[np.ndarray, np.ndarray]:
rng = np.random.default_rng(seed)
windows = np.empty((count, WINDOW_SIZE, 2), dtype=np.float32)
labels = np.empty(count, dtype=np.float32)
x = np.arange(WINDOW_SIZE, dtype=np.float32)
for index in range(count):
slope = rng.uniform(-0.01, 0.13)
force = (
rng.uniform(96.0, 104.0)
+ slope * x
+ rng.normal(0.0, rng.uniform(0.03, 0.18), WINDOW_SIZE)
)
deviation = (
0.07 + np.maximum(0.0, force - 101.0) * 0.011 + rng.normal(0.0, 0.002, WINDOW_SIZE)
)
windows[index, :, 0] = force
windows[index, :, 1] = deviation
shift = force[-10:].mean() - force[:10].mean()
labels[index] = float(slope >= 0.07 and shift >= 2.0)
return windows, labels
def train_model() -> tuple[TinyDriftNet, dict]:
random.seed(17)
np.random.seed(17)
torch.manual_seed(17)
x, y = synthetic_windows()
model = TinyDriftNet()
optimizer = torch.optim.AdamW(model.classifier.parameters(), lr=0.04, weight_decay=0.001)
loss_fn = nn.BCELoss()
x_tensor = torch.from_numpy(x)
y_tensor = torch.from_numpy(y)
model.train()
for _ in range(240):
optimizer.zero_grad(set_to_none=True)
loss = loss_fn(model(x_tensor), y_tensor)
loss.backward()
optimizer.step()
model.eval()
with torch.no_grad():
probabilities = model(x_tensor)
predictions = probabilities >= 0.5
labels = y_tensor.bool()
accuracy = float((predictions == labels).float().mean())
false_positives = ((predictions == 1) & (labels == 0)).sum()
false_negatives = ((predictions == 0) & (labels == 1)).sum()
false_positive_rate = float(false_positives / (labels == 0).sum())
false_negative_rate = float(false_negatives / (labels == 1).sum())
metrics = {
"synthetic_examples": len(x),
"accuracy": round(accuracy, 6),
"false_positive_rate": round(false_positive_rate, 6),
"false_negative_rate": round(false_negative_rate, 6),
"training_seed": 17,
}
return model, metrics
def export_pytorch_and_onnx(model: TinyDriftNet) -> None:
torch.save(
{
"model_version": MODEL_VERSION,
"state_dict": model.state_dict(),
},
ARTIFACTS / "tiny_drift_pytorch.pt",
)
example = torch.zeros(2, WINDOW_SIZE, 2, dtype=torch.float32)
batch = torch.export.Dim("batch", min=1, max=512)
onnx_program = torch.onnx.export(
model,
(example,),
input_names=["telemetry"],
output_names=["probability"],
dynamic_shapes=({0: batch},),
dynamo=True,
verify=True,
external_data=False,
)
onnx_program.save(ARTIFACTS / "tiny_drift.onnx")
def export_tensorflow_and_litert(weight: np.ndarray, bias: float) -> None:
import tensorflow as tf
from forge_lab.models.tensorflow_model import TensorFlowDriftModel
model = TensorFlowDriftModel(weight, bias)
saved_model = ARTIFACTS / "tensorflow_saved_model"
if saved_model.exists():
shutil.rmtree(saved_model)
concrete = model.__call__.get_concrete_function()
tf.saved_model.save(model, saved_model, signatures={"serving_default": concrete})
converter = tf.lite.TFLiteConverter.from_concrete_functions([concrete], model)
converter.optimizations = [tf.lite.Optimize.DEFAULT]
(ARTIFACTS / "tiny_drift.tflite").write_bytes(converter.convert())
def write_metadata(model: TinyDriftNet, metrics: dict) -> None:
weight = model.classifier.weight.detach().cpu().numpy().reshape(-1)
bias = float(model.classifier.bias.detach().cpu().item())
weights = {
"model_version": MODEL_VERSION,
"feature_center": DEFAULT_CENTER.tolist(),
"feature_scale": DEFAULT_SCALE.tolist(),
"linear_weight": weight.tolist(),
"linear_bias": bias,
}
(ARTIFACTS / "weights.json").write_text(json.dumps(weights, indent=2) + "\n")
manifest = {
"model_version": MODEL_VERSION,
"input": {"name": "telemetry", "shape": ["batch", WINDOW_SIZE, 2], "dtype": "float32"},
"channels": ["peak_force_kn", "part_deviation_mm"],
"outputs": ["probability"],
"runtimes": ["numpy", "pytorch", "onnxruntime", "tensorflow", "litert"],
"framework_versions": {
"python": sys.version.split()[0],
"numpy": np.__version__,
"pytorch": torch.__version__,
"pytorch_cuda_build": torch.version.cuda,
"onnx": version("onnx"),
"onnxruntime": version("onnxruntime"),
"tensorflow": version("tensorflow-cpu"),
"litert": version("ai-edge-litert"),
},
"metrics": metrics,
"data": "deterministic synthetic telemetry; no plant data",
"safety": "L0 decision support only; output cannot actuate equipment",
"cuda_validation": "not executed; no GPU or nvcc in the build workspace",
}
(ARTIFACTS / "model_manifest.json").write_text(json.dumps(manifest, indent=2) + "\n")
def main() -> None:
ARTIFACTS.mkdir(parents=True, exist_ok=True)
model, metrics = train_model()
write_metadata(model, metrics)
weight = model.classifier.weight.detach().cpu().numpy().reshape(-1)
bias = float(model.classifier.bias.detach().cpu().item())
export_pytorch_and_onnx(model)
export_tensorflow_and_litert(weight, bias)
print(json.dumps(metrics, indent=2))
if __name__ == "__main__":
main()
|