sm-riti16's picture
Update app.py
9bc3558 verified
# -*- coding: utf-8 -*-
"""natural_language_gradio.ipynb
Automatically generated by Colab.
Original file is located at
https://colab.research.google.com/drive/135ewhMpX2YZA9ysE_QNN0yFQVWYaeW34
"""
import json, pandas as pd
from dataclasses import dataclass, asdict
import math
from dataclasses import dataclass, asdict
from typing import Dict, Any
@dataclass
class PIDInputs:
tau_s: float # plant time constant [s], tau > 0
Kp: float # proportional gain
Ki: float # integral gain (>0 for step tracking)
Kd: float # derivative gain
step_amplitude: float = 1.0 # unit step default
settling_pct: float = 0.02 # 2% criterion for settling time
def validate_inputs(x: PIDInputs) -> Dict[str, Any]:
issues = []
# Reasonable ranges for a compact, safe demo. Adjust as needed.
if not (0.01 <= x.tau_s <= 10.0):
issues.append(f"tau must be in [0.01, 10] s, got {x.tau_s:.4g}")
if not (-0.9 <= x.Kp <= 200.0):
issues.append(f"Kp should be in [-0.9, 200], got {x.Kp:.4g}")
if not (1e-6 <= x.Ki <= 1e4):
issues.append(f"Ki should be in [1e-6, 1e4], got {x.Ki:.4g}")
if not (-0.009 <= x.Kd <= 100.0):
issues.append(f"Kd should be in [-0.009, 100], got {x.Kd:.4g}")
if x.tau_s + x.Kd <= 0:
issues.append("tau + Kd must be > 0 for a proper 2nd-order form.")
if x.step_amplitude == 0:
issues.append("step amplitude should be non-zero for meaningful metrics.")
if not (0.005 <= x.settling_pct <= 0.1):
issues.append("settling_pct should be within [0.005, 0.1] (i.e., 0.5% to 10%).")
return {"ok": len(issues) == 0, "issues": issues}
def compute_pid(x: PIDInputs) -> Dict[str, Any]:
val = validate_inputs(x)
status = "ok" if val["ok"] else "invalid"
wn = None
zeta = None
if (x.tau_s + x.Kd) > 0 and x.Ki > 0:
wn = math.sqrt(x.Ki / (x.tau_s + x.Kd))
denom = 2.0 * math.sqrt(x.Ki * (x.tau_s + x.Kd))
zeta = (x.Kp + 1.0) / denom
# --- NEW: poles & damping classification ---
poles = None
damping_class = None
if wn is not None and wn > 0 and zeta is not None:
# Standard 2nd-order characteristic: s^2 + 2ζωn s + ωn^2 = 0
# Poles: s = -ζωn ± ωn*sqrt(ζ^2 - 1)
re = -zeta * wn
disc = zeta**2 - 1.0
if disc < 0:
# complex-conjugate poles
im = wn * math.sqrt(1.0 - zeta**2)
poles = [complex(re, im), complex(re, -im)]
damping_class = "underdamped (ζ<1): complex-conjugate poles"
elif abs(disc) < 1e-12:
poles = [complex(re, 0.0), complex(re, 0.0)]
damping_class = "critically damped (ζ≈1): repeated real pole"
else:
# distinct real poles
root = wn * math.sqrt(disc)
poles = [complex(re + root, 0.0), complex(re - root, 0.0)]
damping_class = "overdamped (ζ>1): two distinct real poles"
metrics = {}
if wn is not None and zeta is not None and wn > 0 and zeta > 0:
if zeta < 1.0:
wd = wn * math.sqrt(1.0 - zeta**2)
Tp = math.pi / wd
Mp = math.exp(-math.pi * zeta / math.sqrt(1.0 - zeta**2)) # ratio
else:
wd = None
Tp = None
Mp = 0.0
Ts = 4.0 / (zeta * wn) * (0.02 / x.settling_pct)
if zeta < 1.0:
theta = math.acos(zeta)
Tr = (math.pi - theta) / (wn * math.sqrt(1.0 - zeta**2))
else:
Tr = 2.0 / wn
ess = 0.0
metrics = {
"wn_rad_s": wn,
"zeta": zeta,
"wd_rad_s": wd,
"Mp_pct": 100.0 * Mp,
"Tp_s": Tp,
"Ts_s": Ts,
"Tr_s": Tr,
"ess": ess,
}
structured = {
"meta": {
"model": "PID_on_1stOrder_v1",
"deterministic": True,
"assumptions": [
"Unity feedback.",
"1st-order plant G(s) = 1/(tau s + 1).",
"Linear time-invariant dynamics.",
"PID controller C(s) = Kp + Ki/s + Kd s.",
"Small-signal step response analysis."
],
"units": {
"tau_s": "s",
"wn_rad_s": "rad/s",
"wd_rad_s": "rad/s",
"Tp_s": "s",
"Ts_s": "s",
"Tr_s": "s",
"Mp_pct": "%"
},
"valid_ranges": {
"tau_s": "[0.01, 10] s",
"Kp": "[-0.9, 200]",
"Ki": "[1e-6, 1e4]",
"Kd": "[-0.009, 100]",
"tau+Kd": "> 0",
"Ki_positive": "> 0",
"settling_pct": "[0.005, 0.1]"
}
},
"inputs": asdict(x),
"validation": val,
"normalized_second_order": {
"a2": x.tau_s + x.Kd,
"a1": 1.0 + x.Kp,
"a0": x.Ki,
"wn": wn,
"zeta": zeta
},
# --- NEW: add poles & classification in the payload ---
"poles": [complex(p).real if abs(p.imag) < 1e-15 else p for p in (poles or [])],
"damping_class": damping_class,
"metrics": metrics,
"status": status
}
return structured
import gradio as gr
import pandas as pd
from transformers import pipeline
from typing import Dict, Any
# from core import PIDInputs, compute_pid
from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
MODEL_ID = "HuggingFaceTB/SmolLM2-135M-Instruct"
_tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
_model = AutoModelForCausalLM.from_pretrained(MODEL_ID, device_map="auto")
explainer = pipeline(task="text-generation", model=_model, tokenizer=_tokenizer)
def explain_structured(d: dict) -> str:
"""
Explain what the OUTPUT means (stability class, ωn, ζ, poles, overshoot, Tr/Tp/Ts, ess).
Uses the SmolLM explainer with deterministic decoding, then falls back to a
deterministic Markdown explanation if the model returns too little text.
"""
meta = d.get("meta", {})
m = d.get("metrics", {})
norm = d.get("normalized_second_order", {})
poles = d.get("poles", [])
dampc = d.get("damping_class", None)
val = d.get("validation", {})
status = d.get("status")
issues = val.get("issues", [])
# ---------- helpers ----------
def r(v, n=4, na="N/A"):
try:
return f"{float(v):.{n}g}"
except Exception:
return na if v is None else str(v)
def pstr(p):
try:
# p may already be complex or a float
if isinstance(p, complex) or (hasattr(p, "imag") and p.imag != 0):
return f"{p.real:+.4g} {'+' if p.imag>=0 else '-'} j{abs(p.imag):.4g}"
return f"{float(p):+.4g}"
except Exception:
return str(p)
def dedup_lines(md: str) -> str:
seen, out = set(), []
for line in md.splitlines():
key = line.strip()
# never dedup headers; only de-dup plain bullet/paragraph lines
if key and not key.startswith("#") and key in seen:
continue
seen.add(key)
out.append(line)
return "\n".join(out)
# ---------- invalid → deterministic, no LLM ----------
if status != "ok" or issues:
bullets = "\n".join([f"- {iss}" for iss in issues]) if issues else "- Check inputs."
return f"""# Results Explanation
**Status:** ❌ Invalid inputs
Fix these first:
{bullets}
**Why it matters**
- τ+Kd must be > 0 to form a valid 2nd-order model.
- Ki > 0 (type-1) gives zero steady-state error to a step.
"""
# ---------- numeric snapshot for prompt & fallback ----------
wn = norm.get("wn")
zeta = norm.get("zeta")
Mp = m.get("Mp_pct")
Tp = m.get("Tp_s")
Ts = m.get("Ts_s")
Tr = m.get("Tr_s")
ess = m.get("ess")
poles_text = ", ".join(pstr(p) for p in poles) if poles else "N/A"
snapshot = (
f"- ωₙ (natural frequency): {r(wn)} rad/s\n"
f"- ζ (damping ratio): {r(zeta)}{dampc or 'N/A'}\n"
f"- Poles: {poles_text}\n"
f"- Overshoot: ≈ {r(Mp,3)} %\n"
f"- Rise time Tr: ≈ {r(Tr)} s\n"
f"- Peak time Tp: ≈ {r(Tp)} s\n"
f"- Settling time Ts: ≈ {r(Ts)} s\n"
f"- Steady-state error (step): {r(ess)}"
)
# ---------- LLM prompt (deterministic, stability-focused) ----------
prompt = (
"You are a controls engineer. Explain what the OUTPUT VALUES MEAN.\n"
"Write CLEAR MARKDOWN with short, specific bullets. No repetition.\n\n"
"## Stability classification (what ζ and the poles tell you)\n"
"- State whether the system is underdamped, critically damped, or overdamped based on ζ and the pole pattern.\n"
"- Explain what complex vs real poles imply for oscillations and smoothness.\n\n"
"## What ωₙ means (speed)\n"
"- Explain that ωₙ sets the overall speed scale of the response (larger ωₙ → shorter Tr and Ts).\n\n"
"## What ζ means (smoothness vs overshoot)\n"
"- Interpret ζ ranges (<1, ≈1, >1) in terms of oscillation and overshoot.\n\n"
"## What each time/percent metric means\n"
"- Overshoot: how much the peak exceeds final value.\n"
"- Tr: time to go from low to near-final (e.g., 10–90%).\n"
"- Tp: time to first peak.\n"
"- Ts: time to settle within the chosen band.\n"
"- ess: final error for a step; with Ki>0 it is 0.\n\n"
"## How the poles relate to that behavior\n"
"- Connect pole real part to decay speed; imaginary part to oscillation frequency.\n\n"
"## Numeric snapshot\n"
f"{snapshot}\n"
)
# ---------- deterministic generation with anti-repetition ----------
gen = explainer(
prompt,
max_new_tokens=220,
do_sample=False,
temperature=0.0,
top_p=1.0,
top_k=0,
repetition_penalty=1.15,
no_repeat_ngram_size=4,
eos_token_id=_tokenizer.eos_token_id,
pad_token_id=_tokenizer.eos_token_id,
return_full_text=False
)[0]["generated_text"]
# ---------- SHORT-OUTPUT FALLBACK (your requested addition) ----------
MIN_WORDS = 30
if not gen or len(gen.split()) < MIN_WORDS:
gen = f"""## Stability classification
- ζ = {r(zeta)}{dampc or 'N/A'}.
## Meaning of ωₙ and ζ
- ωₙ = {r(wn)} rad/s sets the speed scale (larger ωₙ → faster rise/settle).
- ζ controls smoothness/overshoot: ζ<1 underdamped; ζ≈1 critically damped; ζ>1 overdamped.
## Poles and behavior
- Poles: {poles_text}
- More negative real part → faster decay; nonzero imaginary part → oscillations.
## Time-domain metrics
- Overshoot ≈ {r(Mp,3)} % | Tr ≈ {r(Tr)} s | Tp ≈ {r(Tp)} s | Ts ≈ {r(Ts)} s | ess = {r(ess)}
## Tuning tip
- Raise Ki to increase ωₙ (speed). If overshoot or oscillation appears (ζ too low), add Kd or increase Kp to raise damping.
"""
return dedup_lines(gen)
def run_calc(tau_s, Kp, Ki, Kd, step_amp, settling_pct):
x = PIDInputs(
tau_s=float(tau_s), Kp=float(Kp), Ki=float(Ki), Kd=float(Kd),
step_amplitude=float(step_amp), settling_pct=float(settling_pct) / 100.0 # slider in %, convert to fraction
)
structured = compute_pid(x)
# Display normalized form + metrics in a compact table
rows = []
for k, v in structured.get("normalized_second_order", {}).items():
rows.append(["2nd-order", k, v])
for k, v in structured.get("metrics", {}).items():
rows.append(["metrics", k, v])
df = pd.DataFrame(rows, columns=["section", "key", "value"])
explanation = explain_structured(structured)
return df, explanation, structured
with gr.Blocks(title="PID Controls Calculator (1st-Order Plant)", theme=gr.themes.Soft()) as demo:
gr.Markdown("# PID Feedback Controls — Deterministic Calculator")
gr.Markdown(
"Unity-feedback PID on a first-order plant G(s)=1/(τs+1). "
"We derive the equivalent 2nd-order parameters (ωₙ, ζ) and step-response metrics (overshoot, rise, peak, settling)."
)
with gr.Row():
with gr.Column():
tau_s = gr.Slider(0.01, 10.0, value=0.5, step=0.01, label="Plant time constant τ [s]")
Kp = gr.Slider(-0.9, 200.0, value=1.0, step=0.1, label="Kp")
Ki = gr.Slider(1e-6, 1e4, value=1.0, step=0.1, label="Ki")
Kd = gr.Slider(-0.009, 100.0, value=0.0, step=0.001, label="Kd")
step_amp = gr.Slider(0.1, 10.0, value=1.0, step=0.1, label="Step amplitude")
settling_pct = gr.Slider(0.5, 10.0, value=2.0, step=0.1, label="Settling band [%]")
go = gr.Button("Compute", variant="primary")
with gr.Column():
gr.Markdown("### Numerical Results")
table = gr.Dataframe(headers=["section", "key", "value"], interactive=False)
gr.Markdown("### Explain the Results")
explanation = gr.Markdown()
gr.Markdown("### Raw Structured Output")
json_out = gr.JSON(label="Structured JSON")
go.click(run_calc, inputs=[tau_s, Kp, Ki, Kd, step_amp, settling_pct],
outputs=[table, explanation, json_out])
if __name__ == "__main__":
demo.launch()