File size: 13,855 Bytes
b5d6b93 2d24284 b5d6b93 74cfe42 b5d6b93 74cfe42 b5d6b93 2d24284 b5d6b93 2d24284 b4bd4a7 b5d6b93 b4bd4a7 b5d6b93 b4bd4a7 b5d6b93 b4bd4a7 b5d6b93 b4bd4a7 b5d6b93 b4bd4a7 ba694f9 b4bd4a7 ba694f9 b4bd4a7 6783013 b4bd4a7 6783013 b4bd4a7 ba694f9 b4bd4a7 6783013 b4bd4a7 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 | """
game/ui.py β "Beat the Machine" duel tab.
A card from the curated deck is dealt; you call its safety tier (SAFE / CAUTION /
DEADLY) and the same on-device pipeline calls it at the same instant. Score
accrues per session; post it to the leaderboard. The machine's edge is that it
refuses when unsure β so the deadly cards punish overconfidence in both players.
Stay sharper than the machine you built.
"""
from functools import partial
import gradio as gr
import pandas as pd
from PIL import Image
from pipeline.convergence import build_result
from . import datastore
from . import deck as deck_mod
TIER_COLOR = {"SAFE": "#2f6b2b", "CAUTION": "#87671c", "DEADLY": "#8c1d14", "UNKNOWN": "#57544c"}
_LB_COLS = ["Player", "You %", "Machine %", "Rounds"]
MIN_RANK_ROUNDS = 5 # ranked only after 5 rounds, so a 1-round 100% can't top the board
def _machine_call(pipe, path: str):
res = build_result(pipe.identify(Image.open(path)))
tier = "UNKNOWN" if res.abstained else res.safety
return tier, res
def _scoreboard(score: dict) -> str:
return (
f"<div class='gm-score'>"
f"<span class='gm-you'>YOU <b>{score['you']}</b></span>"
f"<span class='gm-vs'>/ {score['total']} Β· vs Β·</span>"
f"<span class='gm-mach'>MACHINE <b>{score['machine']}</b></span>"
f"</div>"
)
def _pretty(species: str) -> str:
return species.replace("_toxic", "").replace("_deadly", "").replace("_", " ").title()
def _chip(label: str, tier: str, right: bool) -> str:
c = TIER_COLOR.get(tier, "#57544c")
mark = "β" if right else "β"
# color set inline with !important on the spans: Gradio's theme paints spans in
# the display panel a light color that overrides inherited/class color (the big
# truth div is fine, but these spans aren't) β inline !important wins the cascade.
return (
f"<div class='gm-chip' style='border-color:{c}'>"
f"<span class='gm-chip-h' style='color:{c} !important'>{label}</span>"
f"<span class='gm-chip-t' style='color:{c} !important'>{tier} {mark}</span></div>"
)
def _reveal_html(card: dict, you_tier: str, you_right: bool, mtier: str, mach_right: bool) -> str:
tc = TIER_COLOR[card["tier"]]
if you_right and not mach_right:
flav = "π You beat the machine on this one."
elif mach_right and not you_right:
flav = "The machine got you. Study the look-alike."
elif you_right and mach_right:
flav = "Dead heat β both correct."
else:
flav = "Neither nailed it. The woods don't grade on a curve."
abst = ""
if mtier == "UNKNOWN":
abst = ("<div class='gm-abst'>The machine refused to commit β that's its whole "
"point. A refusal beats a confident wrong call.</div>")
return (
f"<div class='gm-reveal'>"
f"<div class='gm-truth-h'>TRUTH</div>"
f"<div class='gm-truth' style='color:{tc}'>{card['tier']}</div>"
f"<div class='gm-species'>{_pretty(card['species'])} Β· <i>{card['scientific']}</i></div>"
f"<div class='gm-chips'>{_chip('YOUR CALL', you_tier, you_right)}"
f"{_chip('MACHINE', mtier, mach_right)}</div>"
f"<div class='gm-flav'>{flav}</div>{abst}</div>"
)
def _lb_df() -> pd.DataFrame:
data = []
for r in datastore.load_leaderboard():
t = int(r.get("skill_total", 0) or 0)
if t < MIN_RANK_ROUNDS:
continue
data.append([
r.get("contributor", "?"),
round(100 * r.get("skill_correct", 0) / t),
round(100 * r.get("machine_correct", 0) / t),
t,
])
data.sort(key=lambda x: (-x[1], -x[3]))
return pd.DataFrame(data, columns=_LB_COLS)
def build_game_tab(pipe) -> None:
"""Build the duel tab inside the current gr.Blocks context."""
gr.HTML(
"<div class='gm-intro'>"
"<div class='gm-intro-h'>CAN YOU BEAT THE MACHINE?</div>"
"<div class='gm-intro-b'>We built the machine. Now try to out-forage it. "
"A card is dealt β call its tier before it does. Remember: the machine's "
"trick is knowing when <i>not</i> to guess. Every round you play sharpens you "
"and sharpens the open dataset behind it.</div></div>"
)
score_state = gr.State({"you": 0, "machine": 0, "total": 0})
card_state = gr.State(None)
with gr.Row():
with gr.Column(scale=1):
card_img = gr.Image(type="filepath", show_label=False, interactive=False,
elem_classes="eink-input", height=300)
deal_btn = gr.Button("βΈ DEAL A CARD", variant="primary", elem_classes="eink-scan")
with gr.Row():
safe_b = gr.Button("SAFE", elem_classes="gm-btn gm-safe")
caut_b = gr.Button("CAUTION", elem_classes="gm-btn gm-caut")
dead_b = gr.Button("DEADLY", elem_classes="gm-btn gm-dead")
with gr.Column(scale=1, elem_classes="eink-screen"):
reveal = gr.HTML("<div class='gm-idle'>Deal a card to start the duel.</div>")
score_html = gr.HTML(_scoreboard({"you": 0, "machine": 0, "total": 0}))
gr.HTML("<div class='gm-divider'>β LEADERBOARD Β· ranked at 5+ rounds</div>")
post_btn = gr.Button("POST MY SCORE", elem_classes="eink-scan")
lb_status = gr.HTML("")
lb = gr.Dataframe(value=_lb_df(), headers=_LB_COLS, interactive=False,
elem_classes="gm-lb", wrap=True)
refresh_btn = gr.Button("β³ refresh", elem_classes="gm-refresh")
# ββ handlers (closures capture `pipe`) βββββββββββββββββββββββββββββββββββ
def _deal(score, card):
prev = card["file"] if card else None
c = deck_mod.random_card(exclude_file=prev)
new = {**c, "answered": False}
return c["path"], new, "<div class='gm-idle'>Your call?</div>", _scoreboard(score)
def _guess(tier, card, score):
if not card or card.get("answered"):
return gr.update(), card, score # ignore double-taps / no card yet
mtier, _ = _machine_call(pipe, card["path"])
you_right = tier == card["tier"]
mach_right = mtier == card["tier"]
score = {
"you": score["you"] + int(you_right),
"machine": score["machine"] + int(mach_right),
"total": score["total"] + 1,
}
card = {**card, "answered": True}
return _reveal_html(card, tier, you_right, mtier, mach_right), card, score
def _post(score, profile: gr.OAuthProfile | None = None):
if profile is None:
return "<div class='gm-warn'>Log in (top of the page) to post your score.</div>", gr.update()
if score["total"] == 0:
return "<div class='gm-warn'>Play at least one round first.</div>", gr.update()
datastore.post_score(profile.username, score["you"], score["total"], score["machine"])
note = "" if datastore.persistence_enabled() else \
" <span class='gm-note'>(session-only until the dataset token is set)</span>"
return f"<div class='gm-ok'>Posted as {profile.username}.{note}</div>", _lb_df()
deal_btn.click(_deal, [score_state, card_state], [card_img, card_state, reveal, score_html])
for b, t in [(safe_b, "SAFE"), (caut_b, "CAUTION"), (dead_b, "DEADLY")]:
b.click(partial(_guess, t), [card_state, score_state], [reveal, card_state, score_state]) \
.then(_scoreboard, [score_state], [score_html])
post_btn.click(_post, [score_state], [lb_status, lb])
refresh_btn.click(lambda: _lb_df(), None, [lb])
# ββ Stump the Machine: upload your own find -> the data flywheel ββββββββββββββ
_CONTRIB_COLS = ["Contributor", "Sightings"]
def _contrib_df() -> pd.DataFrame:
rows = datastore.load_contributors()
return pd.DataFrame([[r["contributor"], r["count"]] for r in rows], columns=_CONTRIB_COLS)
def _stump_unrouted_html(status: str) -> str:
body = {
"stored": "<div class='gm-ok'>Saved for review. If it's a real find, you just handed us a "
"training example the model is missing.</div>",
"duplicate": "<div class='gm-warn'>Already flagged for review β this one's logged.</div>",
"disabled": "<div class='gm-warn'>Couldn't save for review (review-queue token scope not set "
"on the Space).</div>",
}.get(status, "")
return (
"<div class='gm-reveal'>"
"<div class='gm-truth-h'>YOU FOUND A GAP</div>"
"<div class='gm-truth' style='color:#57544c'>OFF THE MAP</div>"
"<div class='gm-species'>The machine couldn't place this in berry, mushroom, or plant β "
"the ultimate stump.</div>"
f"{body}"
"<div class='gm-abst'>Genuinely off-topic shots (not a wild plant, mushroom, or berry) get "
"filtered out in review.</div>"
"</div>"
)
def _stump_result_html(label: str, res, status: str) -> str:
mtier = "UNKNOWN" if res.abstained else res.safety
c = TIER_COLOR.get(mtier, "#57544c")
mcall = "refused to commit" if res.abstained else f"{_pretty(res.species)} Β· {mtier}"
head = "π You stumped the machine!" if res.abstained else "The machine made its call."
store_line = {
"stored": "<div class='gm-ok'>β Added to the open dataset β thank you. This is exactly "
"the data that makes the next model better.</div>",
"duplicate": "<div class='gm-warn'>Already in the dataset β this photo (or a near-match) "
"was logged before, so it wasn't added again.</div>",
"disabled": "<div class='gm-warn'>Couldn't persist (dataset token not set on the Space).</div>",
}.get(status, "")
return (
f"<div class='gm-reveal'>"
f"<div class='gm-truth-h'>YOUR FIND</div>"
f"<div class='gm-flav'>{head}</div>"
f"<div class='gm-species'>you said: <b>{label}</b></div>"
f"<div class='gm-chips'><div class='gm-chip' style='border-color:{c}'>"
f"<span class='gm-chip-h' style='color:{c} !important'>MACHINE</span>"
f"<span class='gm-chip-t' style='color:{c} !important'>{mcall}</span></div></div>"
f"{store_line}"
f"<div class='gm-abst'>An ID here is never permission to eat β verify with an expert.</div>"
f"</div>"
)
def build_stump_tab(pipe) -> None:
"""Upload a real find: the router gates it, the model calls it, and (consented) it
joins the public dataset. The flywheel β every submission trains the next model."""
gr.HTML(
"<div class='gm-intro'>"
"<div class='gm-intro-h'>STUMP THE MACHINE</div>"
"<div class='gm-intro-b'>Upload your own find. The machine calls it on the spot β and "
"when it can't, you've stumped it. Every accepted photo joins the open "
"<b>CC-BY-4.0</b> dataset that trains the next model. You're not just playing; you're "
"building the thing.</div></div>"
)
with gr.Row():
with gr.Column(scale=1):
up = gr.Image(type="pil", sources=["upload", "webcam"], label="YOUR FIND",
elem_classes="eink-input", height=300)
guess = gr.Textbox(label="What do you think it is?", placeholder="e.g. chanterelle",
elem_classes="eink-input")
consent = gr.Checkbox(
value=False,
label="Contribute this photo to the public CC-BY-4.0 dataset (HomesteaderLabs/forager-sightings).")
submit = gr.Button("βΈ SUBMIT FIND", variant="primary", elem_classes="eink-scan")
with gr.Column(scale=1, elem_classes="eink-screen"):
sresult = gr.HTML("<div class='gm-idle'>Upload a find, add your guess, and submit.</div>")
gr.HTML("<div class='gm-divider'>β TOP CONTRIBUTORS</div>")
contrib = gr.Dataframe(value=_contrib_df(), headers=_CONTRIB_COLS, interactive=False,
elem_classes="gm-lb", wrap=True)
crefresh = gr.Button("β³ refresh", elem_classes="gm-refresh")
def _submit(image, user_label, consented, profile: gr.OAuthProfile | None = None):
if profile is None:
return "<div class='gm-warn'>Log in (top of the page) to contribute.</div>", gr.update()
if image is None:
return "<div class='gm-warn'>Upload a photo first.</div>", gr.update()
if not (user_label or "").strip():
return "<div class='gm-warn'>Add your guess first.</div>", gr.update()
if not consented:
return "<div class='gm-warn'>Check the consent box to contribute.</div>", gr.update()
call = pipe.identify(image)
res = build_result(call)
in_domain = (not res.abstained) or call.get("reason", "") == "low_confidence"
if not in_domain:
router = {"domain": call.get("domain", "unknown"),
"domain_confidence": call.get("domain_confidence", 0.0),
"reason": call.get("reason", "")}
status = datastore.append_unrouted(image, user_label.strip(), router, profile.username)
return _stump_unrouted_html(status), gr.update() # contributor board unchanged
machine = {"species": res.species, "confidence": res.confidence,
"abstained": res.abstained, "safety": res.safety, "domain": res.domain}
status = datastore.append_sighting(image, user_label.strip(), machine, profile.username)
board = _contrib_df() if status == "stored" else gr.update()
return _stump_result_html(user_label.strip(), res, status), board
submit.click(_submit, [up, guess, consent], [sresult, contrib])
crefresh.click(lambda: _contrib_df(), None, [contrib])
|