refactor(praca): fix ruff violations in visualize scripts

This commit is contained in:
Krzysztof kuhy Rudnicki 2026-03-14 14:29:18 +01:00
parent be31e9abd7
commit 47c7679222
3 changed files with 773 additions and 660 deletions

View File

@ -6,6 +6,8 @@ on a small example graph, rendering each algorithm step by step.
from __future__ import annotations from __future__ import annotations
from dataclasses import dataclass
import logging
import os import os
from pathlib import Path from pathlib import Path
@ -33,6 +35,9 @@ OUTPUT_DIR = Path(__file__).resolve().parent / "videos"
OUTPUT_DIR.mkdir(parents=True, exist_ok=True) OUTPUT_DIR.mkdir(parents=True, exist_ok=True)
OUTPUT = str(OUTPUT_DIR / "q02_shortest_path.mp4") OUTPUT = str(OUTPUT_DIR / "q02_shortest_path.mp4")
logging.basicConfig(level=logging.INFO)
_logger = logging.getLogger(__name__)
# Graph definition # Graph definition
NODE_POS = {"S": (250, 280), "A": (550, 180), "B": (550, 450), "C": (850, 320)} NODE_POS = {"S": (250, 280), "A": (550, 180), "B": (550, 450), "C": (850, 320)}
EDGES_DIJKSTRA = [ EDGES_DIJKSTRA = [
@ -101,13 +106,13 @@ def _draw_circle(
def _draw_line( def _draw_line(
frame: np.ndarray, frame: np.ndarray,
x1: int, start: tuple[int, int],
y1: int, end: tuple[int, int],
x2: int,
y2: int,
color: tuple[int, ...], color: tuple[int, ...],
thickness: int = 2, thickness: int = 2,
) -> None: ) -> None:
x1, y1 = start
x2, y2 = end
length = max(int(np.sqrt((x2 - x1) ** 2 + (y2 - y1) ** 2)), 1) length = max(int(np.sqrt((x2 - x1) ** 2 + (y2 - y1) ** 2)), 1)
for i in range(length): for i in range(length):
frac = i / length frac = i / length
@ -122,13 +127,13 @@ def _draw_line(
def _draw_arrow( def _draw_arrow(
frame: np.ndarray, frame: np.ndarray,
x1: int, start: tuple[int, int],
y1: int, end: tuple[int, int],
x2: int,
y2: int,
color: tuple[int, ...], color: tuple[int, ...],
thickness: int = 2, thickness: int = 2,
) -> None: ) -> None:
x1, y1 = start
x2, y2 = end
r = 32 r = 32
length = max(np.sqrt((x2 - x1) ** 2 + (y2 - y1) ** 2), 1) length = max(np.sqrt((x2 - x1) ** 2 + (y2 - y1) ** 2), 1)
ddx = (x2 - x1) / length ddx = (x2 - x1) / length
@ -137,14 +142,14 @@ def _draw_arrow(
sy = int(y1 + ddy * r) sy = int(y1 + ddy * r)
ex = int(x2 - ddx * r) ex = int(x2 - ddx * r)
ey = int(y2 - ddy * r) ey = int(y2 - ddy * r)
_draw_line(frame, sx, sy, ex, ey, color, thickness) _draw_line(frame, (sx, sy), (ex, ey), color, thickness)
angle = np.arctan2(ey - sy, ex - sx) angle = np.arctan2(ey - sy, ex - sx)
arrow_len = 12 arrow_len = 12
for side in [-1, 1]: for side in [-1, 1]:
a = angle + np.pi + side * 0.4 a = angle + np.pi + side * 0.4
ax = int(ex + arrow_len * np.cos(a)) ax = int(ex + arrow_len * np.cos(a))
ay = int(ey + arrow_len * np.sin(a)) ay = int(ey + arrow_len * np.sin(a))
_draw_line(frame, ex, ey, ax, ay, color, thickness) _draw_line(frame, (ex, ey), (ax, ay), color, thickness)
def _render_graph( def _render_graph(
@ -163,7 +168,7 @@ def _render_graph(
sx, sy = nodes[src] sx, sy = nodes[src]
dx, dy = nodes[dst] dx, dy = nodes[dst]
ec = COL_EDGE_ACT if active_edge == (src, dst) else COL_EDGE ec = COL_EDGE_ACT if active_edge == (src, dst) else COL_EDGE
_draw_arrow(frame, sx, sy, dx, dy, ec, thickness=2) _draw_arrow(frame, (sx, sy), (dx, dy), ec, thickness=2)
for name, (x, y) in nodes.items(): for name, (x, y) in nodes.items():
if name == current: if name == current:
@ -184,19 +189,32 @@ def _render_graph(
return frame return frame
@dataclass
class _StepConfig:
"""Configuration for a single algorithm visualization step."""
nodes: dict[str, tuple[int, int]]
edges: list[tuple[str, str, int]]
distances: dict[str, str]
current: str | None = None
visited: set[str] | None = None
active_edge: tuple[str, str] | None = None
step_text: str = ""
algo_name: str = ""
def _make_step( def _make_step(
nodes: dict[str, tuple[int, int]], cfg: _StepConfig,
edges: list[tuple[str, str, int]],
distances: dict[str, str],
current: str | None = None,
visited: set[str] | None = None,
active_edge: tuple[str, str] | None = None,
step_text: str = "",
algo_name: str = "",
duration: float = STEP_DUR, duration: float = STEP_DUR,
) -> CompositeVideoClip: ) -> CompositeVideoClip:
if visited is None: nodes = cfg.nodes
visited = set() edges = cfg.edges
distances = cfg.distances
current = cfg.current
visited = cfg.visited if cfg.visited is not None else set()
active_edge = cfg.active_edge
step_text = cfg.step_text
algo_name = cfg.algo_name
graph_frame = _render_graph(nodes, edges, distances, current, visited, active_edge) graph_frame = _render_graph(nodes, edges, distances, current, visited, active_edge)
@ -305,50 +323,66 @@ def _dijkstra_steps() -> list[CompositeVideoClip]:
e = EDGES_DIJKSTRA e = EDGES_DIJKSTRA
return [ return [
_make_step( _make_step(
n, _StepConfig(
e, n,
{"S": "0", "A": INF, "B": INF, "C": INF}, e,
current="S", {"S": "0", "A": INF, "B": INF, "C": INF},
step_text="Inicjalizacja: d[S]=0, reszta=∞. Wybierz S (min d).", current="S",
algo_name="Algorytm Dijkstry", step_text="Inicjalizacja: d[S]=0, reszta=∞. Wybierz S (min d).",
algo_name="Algorytm Dijkstry",
),
), ),
_make_step( _make_step(
n, _StepConfig(
e, n,
{"S": "0", "A": "2", "B": "5", "C": INF}, e,
current="S", {"S": "0", "A": "2", "B": "5", "C": INF},
active_edge=("S", "A"), current="S",
step_text="Relaksacja S→A: d[A]=0+2=2. S→B: d[B]=0+5=5.", active_edge=("S", "A"),
algo_name="Algorytm Dijkstry", step_text="Relaksacja S→A: d[A]=0+2=2. S→B: d[B]=0+5=5.",
algo_name="Algorytm Dijkstry",
),
), ),
_make_step( _make_step(
n, _StepConfig(
e, n,
{"S": "0", "A": "2", "B": "5", "C": "5"}, e,
current="A", {"S": "0", "A": "2", "B": "5", "C": "5"},
visited={"S"}, current="A",
active_edge=("A", "C"), visited={"S"},
step_text="Zamknij S. Min=A(2). Relaksacja A→C: d[C]=2+3=5.", active_edge=("A", "C"),
algo_name="Algorytm Dijkstry", step_text="Zamknij S. Min=A(2). Relaksacja A→C: d[C]=2+3=5.",
algo_name="Algorytm Dijkstry",
),
), ),
_make_step( _make_step(
n, _StepConfig(
e, n,
{"S": "0", "A": "2", "B": "5", "C": "5"}, e,
current="B", {"S": "0", "A": "2", "B": "5", "C": "5"},
visited={"S", "A"}, current="B",
active_edge=("B", "A"), visited={"S", "A"},
step_text="Zamknij A. Min=B(5). B→A: 5+1=6>2, nie zmieniaj. B→C: 5+6=11>5.", active_edge=("B", "A"),
algo_name="Algorytm Dijkstry", step_text=(
"Zamknij A. Min=B(5). B→A: 5+1=6>2, "
"nie zmieniaj. B→C: 5+6=11>5."
),
algo_name="Algorytm Dijkstry",
),
), ),
_make_step( _make_step(
n, _StepConfig(
e, n,
{"S": "0", "A": "2", "B": "5", "C": "5"}, e,
current="C", {"S": "0", "A": "2", "B": "5", "C": "5"},
visited={"S", "A", "B"}, current="C",
step_text="Zamknij B. Min=C(5). Koniec! Wynik: d={S:0, A:2, B:5, C:5}.", visited={"S", "A", "B"},
algo_name="Dijkstra -- WYNIK", step_text=(
"Zamknij B. Min=C(5). Koniec! "
"Wynik: d={S:0, A:2, B:5, C:5}."
),
algo_name="Dijkstra -- WYNIK",
),
), ),
] ]
@ -358,42 +392,67 @@ def _bellman_ford_steps() -> list[CompositeVideoClip]:
e = EDGES_BF e = EDGES_BF
return [ return [
_make_step( _make_step(
n, _StepConfig(
e, n,
{"S": "0", "A": INF, "B": INF, "C": INF}, e,
step_text="Bellman-Ford: relaksuj WSZYSTKIE krawędzie V-1=3 razy. Ujemne wagi OK!", {"S": "0", "A": INF, "B": INF, "C": INF},
algo_name="Algorytm Bellmana-Forda", step_text=(
"Bellman-Ford: relaksuj WSZYSTKIE "
"krawędzie V-1=3 razy. Ujemne wagi OK!"
),
algo_name="Algorytm Bellmana-Forda",
),
), ),
_make_step( _make_step(
n, _StepConfig(
e, n,
{"S": "0", "A": "2", "B": "5", "C": "5"}, e,
active_edge=("S", "A"), {"S": "0", "A": "2", "B": "5", "C": "5"},
step_text="Iteracja 1: S→A:2, A→C:5, S→B:5. Potem B→A: 5+(-4)=1 < 2 → A=1!", active_edge=("S", "A"),
algo_name="Bellman-Ford -- iteracja 1", step_text=(
"Iteracja 1: S→A:2, A→C:5, S→B:5. "
"Potem B→A: 5+(-4)=1 < 2 → A=1!"
),
algo_name="Bellman-Ford -- iteracja 1",
),
), ),
_make_step( _make_step(
n, _StepConfig(
e, n,
{"S": "0", "A": "1", "B": "5", "C": "5"}, e,
active_edge=("B", "A"), {"S": "0", "A": "1", "B": "5", "C": "5"},
step_text="B→A z ujemną wagą -4: d[A] poprawione z 2 na 1! (Dijkstra by to pominął!)", active_edge=("B", "A"),
algo_name="Bellman-Ford -- ujemna waga", step_text=(
"B→A z ujemną wagą -4: d[A] poprawione "
"z 2 na 1! (Dijkstra by to pominął!)"
),
algo_name="Bellman-Ford -- ujemna waga",
),
), ),
_make_step( _make_step(
n, _StepConfig(
e, n,
{"S": "0", "A": "1", "B": "5", "C": "4"}, e,
active_edge=("A", "C"), {"S": "0", "A": "1", "B": "5", "C": "4"},
step_text="Iteracja 2: A→C: 1+3=4 < 5 → C=4. Propagacja poprawionego A.", active_edge=("A", "C"),
algo_name="Bellman-Ford -- iteracja 2", step_text=(
"Iteracja 2: A→C: 1+3=4 < 5 → C=4. "
"Propagacja poprawionego A."
),
algo_name="Bellman-Ford -- iteracja 2",
),
), ),
_make_step( _make_step(
n, _StepConfig(
e, n,
{"S": "0", "A": "1", "B": "5", "C": "4"}, e,
step_text="Iteracja 3: brak zmian. V-ta iteracja: brak popraw → brak cyklu ujemnego.", {"S": "0", "A": "1", "B": "5", "C": "4"},
algo_name="Bellman-Ford -- WYNIK, O(V*E)", step_text=(
"Iteracja 3: brak zmian. V-ta iteracja: "
"brak popraw → brak cyklu ujemnego."
),
algo_name="Bellman-Ford -- WYNIK, O(V*E)",
),
), ),
] ]
@ -403,40 +462,60 @@ def _astar_steps() -> list[CompositeVideoClip]:
e = EDGES_DIJKSTRA e = EDGES_DIJKSTRA
return [ return [
_make_step( _make_step(
n, _StepConfig(
e, n,
{"S": "0", "A": INF, "B": INF, "C": INF}, e,
current="S", {"S": "0", "A": INF, "B": INF, "C": INF},
step_text="A*: f(n)=g(n)+h(n). Cel=C. h(S)=5, h(A)=3, h(B)=4, h(C)=0. f(S)=0+5=5.", current="S",
algo_name="Algorytm A*", step_text=(
"A*: f(n)=g(n)+h(n). Cel=C. "
"h(S)=5, h(A)=3, h(B)=4, h(C)=0. f(S)=0+5=5."
),
algo_name="Algorytm A*",
),
), ),
_make_step( _make_step(
n, _StepConfig(
e, n,
{"S": "0", "A": "2", "B": "5", "C": INF}, e,
current="S", {"S": "0", "A": "2", "B": "5", "C": INF},
active_edge=("S", "A"), current="S",
step_text="Relaksuj S: A(g=2,f=2+3=5), B(g=5,f=5+4=9). Min f → A(5).", active_edge=("S", "A"),
algo_name="A* -- rozwijanie S", step_text=(
"Relaksuj S: A(g=2,f=2+3=5), "
"B(g=5,f=5+4=9). Min f → A(5)."
),
algo_name="A* -- rozwijanie S",
),
), ),
_make_step( _make_step(
n, _StepConfig(
e, n,
{"S": "0", "A": "2", "B": "5", "C": "5"}, e,
current="A", {"S": "0", "A": "2", "B": "5", "C": "5"},
visited={"S"}, current="A",
active_edge=("A", "C"), visited={"S"},
step_text="Rozwiń A(f=5): A→C: g=2+3=5, f=5+0=5. Min f → C(5) = CEL!", active_edge=("A", "C"),
algo_name="A* -- rozwijanie A", step_text=(
"Rozwiń A(f=5): A→C: g=2+3=5, "
"f=5+0=5. Min f → C(5) = CEL!"
),
algo_name="A* -- rozwijanie A",
),
), ),
_make_step( _make_step(
n, _StepConfig(
e, n,
{"S": "0", "A": "2", "B": "5", "C": "5"}, e,
current="C", {"S": "0", "A": "2", "B": "5", "C": "5"},
visited={"S", "A"}, current="C",
step_text="Dotarliśmy do C! Koszt=5. A* NIE przetwarza B (3 vs 4 w Dijkstrze).", visited={"S", "A"},
algo_name="A* -- cel osiągnięty!", step_text=(
"Dotarliśmy do C! Koszt=5. "
"A* NIE przetwarza B (3 vs 4 w Dijkstrze)."
),
algo_name="A* -- cel osiągnięty!",
),
), ),
] ]
@ -523,7 +602,7 @@ def main() -> None:
final.write_videofile( final.write_videofile(
OUTPUT, fps=FPS, codec="libx264", audio=False, preset="medium", threads=4 OUTPUT, fps=FPS, codec="libx264", audio=False, preset="medium", threads=4
) )
print(f"Video saved to: {OUTPUT}") _logger.info("Video saved to: %s", OUTPUT)
if __name__ == "__main__": if __name__ == "__main__":

File diff suppressed because it is too large Load Diff

View File

@ -11,6 +11,7 @@ Creates animated video demonstrating:
from __future__ import annotations from __future__ import annotations
import logging
import os import os
from pathlib import Path from pathlib import Path
@ -40,6 +41,8 @@ OUTPUT = str(OUTPUT_DIR / "q24_object_detection.mp4")
BG_COLOR = (15, 20, 35) BG_COLOR = (15, 20, 35)
_logger = logging.getLogger(__name__)
def _tc(**kwargs: object) -> TextClip: def _tc(**kwargs: object) -> TextClip:
"""TextClip wrapper that adds enough bottom margin to prevent clipping.""" """TextClip wrapper that adds enough bottom margin to prevent clipping."""
@ -203,7 +206,8 @@ def _hog_svm_demo() -> list[CompositeVideoClip]:
frame[ay - 1 : ay + 2, ax : ax + 20] = (150, 150, 170) frame[ay - 1 : ay + 2, ax : ax + 20] = (150, 150, 170)
# Show gradient computation example at bottom # Show gradient computation example at bottom
if progress > 0.2: gradient_phase = 0.2
if progress > gradient_phase:
# Mini pixel grid showing gradient computation # Mini pixel grid showing gradient computation
gx, gy = 100, 430 gx, gy = 100, 430
pixels = [50, 50, 200] pixels = [50, 50, 200]
@ -366,7 +370,8 @@ def _viola_jones_demo() -> list[CompositeVideoClip]:
(80, 620), (80, 620),
), ),
( (
"Haar: kontrast jasna/ciemna | Integral Image: suma prostokąta O(1) = 4 odczyty", "Haar: kontrast jasna/ciemna | Integral Image: "
"suma prostokąta O(1) = 4 odczyty",
14, 14,
"#78909C", "#78909C",
FONT_R, FONT_R,
@ -474,7 +479,8 @@ def _rcnn_evolution() -> list[CompositeVideoClip]:
("Faster R-CNN (2015)", 20, "#A5D6A7", FONT_B, (50, 580)), ("Faster R-CNN (2015)", 20, "#A5D6A7", FONT_B, (50, 580)),
("0.2 sec → 5 fps (RPN w sieci!)", 14, "#A5D6A7", FONT_R, (720, 600)), ("0.2 sec → 5 fps (RPN w sieci!)", 14, "#A5D6A7", FONT_R, (720, 600)),
( (
"Kluczowe innowacje: ROI Pooling → stały rozmiar | RPN → propozycje w sieci", "Kluczowe innowacje: ROI Pooling → stały rozmiar "
"| RPN → propozycje w sieci",
14, 14,
"#78909C", "#78909C",
FONT_R, FONT_R,
@ -527,13 +533,15 @@ def _rcnn_detailed() -> list[CompositeVideoClip]:
min(c + 50, 255) for c in color min(c + 50, 255) for c in color
) )
# Arrow down # Arrow down
if i < 4: arrow_limit = 4
if i < arrow_limit:
ax = bx + bw // 2 ax = bx + bw // 2
ay = by + bh + 5 ay = by + bh + 5
frame[ay : ay + 20, ax - 1 : ax + 2] = (150, 150, 170) frame[ay : ay + 20, ax - 1 : ax + 2] = (150, 150, 170)
# Illustration: many overlapping regions from Selective Search # Illustration: many overlapping regions from Selective Search
if progress > 0.2: overlay_phase = 0.2
if progress > overlay_phase:
rng_local = np.random.default_rng(42) rng_local = np.random.default_rng(42)
n_boxes = min(int((progress - 0.2) * 15), 8) n_boxes = min(int((progress - 0.2) * 15), 8)
for i in range(n_boxes): for i in range(n_boxes):
@ -599,94 +607,108 @@ def _rcnn_detailed() -> list[CompositeVideoClip]:
# ── ROI Pooling ────────────────────────────────────────────────── # ── ROI Pooling ──────────────────────────────────────────────────
def _draw_roi_pool_grid(frame: np.ndarray) -> None:
"""Draw the 3x3 ROI pool grid with max-pooled feature values."""
out_x, out_y = 400, 220
out_cell = 50
out_n = 3
roi_r1, roi_c1 = 2, 1
roi_r2, roi_c2 = 6, 5
roi_h = roi_r2 - roi_r1
roi_w = roi_c2 - roi_c1
for r in range(out_n):
for c in range(out_n):
x = out_x + c * out_cell
y = out_y + r * out_cell
# Compute the max from corresponding region
src_r1 = roi_r1 + r * roi_h // out_n
src_r2 = roi_r1 + (r + 1) * roi_h // out_n
src_c1 = roi_c1 + c * roi_w // out_n
src_c2 = roi_c1 + (c + 1) * roi_w // out_n
max_val = 0
for sr in range(src_r1, src_r2):
for sc in range(src_c1, src_c2):
v = 30 + ((sr * 7 + sc * 13 + 42) % 40)
max_val = max(max_val, v)
frame[y : y + out_cell - 2, x : x + out_cell - 2] = (
max_val,
max_val + 20,
max_val + 40,
)
frame[y : y + 2, x : x + out_cell - 2] = (80, 200, 120)
frame[y + out_cell - 4 : y + out_cell - 2, x : x + out_cell - 2] = (
80,
200,
120,
)
def _make_roi_frame(t: float) -> np.ndarray:
"""Render a single frame for the ROI pooling animation."""
frame = np.zeros((H, W, 3), dtype=np.uint8)
frame[:] = BG_COLOR
progress = min(t / (STEP_DUR * 0.7), 1.0)
# Left: feature map with ROI highlighted
fm_x, fm_y = 60, 180
fm_cell = 30
fm_grid = 8
for r in range(fm_grid):
for c in range(fm_grid):
x = fm_x + c * fm_cell
y = fm_y + r * fm_cell
# Random-looking feature values
val = 30 + ((r * 7 + c * 13 + 42) % 40)
frame[y : y + fm_cell - 1, x : x + fm_cell - 1] = (
val,
val + 10,
val + 20,
)
# ROI region highlighted
roi_r1, roi_c1 = 2, 1
roi_r2, roi_c2 = 6, 5
for tt in range(3):
ry1 = fm_y + roi_r1 * fm_cell - tt
ry2 = fm_y + roi_r2 * fm_cell + tt
rx1 = fm_x + roi_c1 * fm_cell - tt
rx2 = fm_x + roi_c2 * fm_cell + tt
frame[ry1:ry2, rx1 : rx1 + 2] = (255, 200, 50)
frame[ry1:ry2, rx2 - 2 : rx2] = (255, 200, 50)
frame[ry1 : ry1 + 2, rx1:rx2] = (255, 200, 50)
frame[ry2 - 2 : ry2, rx1:rx2] = (255, 200, 50)
# Arrow
arrow_phase = 0.3
if progress > arrow_phase:
frame[300:303, 310:380] = (150, 150, 170)
# Middle: ROI divided into 3x3 grid (output_size)
grid_phase = 0.3
if progress > grid_phase:
_draw_roi_pool_grid(frame)
# Arrow to FC
fc_phase = 0.6
if progress > fc_phase:
frame[300:303, 560:630] = (150, 150, 170)
# FC box
frame[270:340, 650:730] = (200, 100, 80)
frame[270:272, 650:730] = (240, 140, 120)
frame[338:340, 650:730] = (240, 140, 120)
return frame
def _roi_pooling_demo() -> list[CompositeVideoClip]: def _roi_pooling_demo() -> list[CompositeVideoClip]:
"""Animate ROI Pooling: key Fast R-CNN innovation.""" """Animate ROI Pooling: key Fast R-CNN innovation."""
slides = [] slides = []
def make_roi_frame(t: float) -> np.ndarray: roi_clip = VideoClip(_make_roi_frame, duration=STEP_DUR + 1).with_fps(FPS)
frame = np.zeros((H, W, 3), dtype=np.uint8)
frame[:] = BG_COLOR
progress = min(t / (STEP_DUR * 0.7), 1.0)
# Left: feature map with ROI highlighted
fm_x, fm_y = 60, 180
fm_cell = 30
fm_grid = 8
for r in range(fm_grid):
for c in range(fm_grid):
x = fm_x + c * fm_cell
y = fm_y + r * fm_cell
# Random-looking feature values
val = 30 + ((r * 7 + c * 13 + 42) % 40)
frame[y : y + fm_cell - 1, x : x + fm_cell - 1] = (
val,
val + 10,
val + 20,
)
# ROI region highlighted
roi_r1, roi_c1 = 2, 1
roi_r2, roi_c2 = 6, 5
for tt in range(3):
ry1 = fm_y + roi_r1 * fm_cell - tt
ry2 = fm_y + roi_r2 * fm_cell + tt
rx1 = fm_x + roi_c1 * fm_cell - tt
rx2 = fm_x + roi_c2 * fm_cell + tt
frame[ry1:ry2, rx1 : rx1 + 2] = (255, 200, 50)
frame[ry1:ry2, rx2 - 2 : rx2] = (255, 200, 50)
frame[ry1 : ry1 + 2, rx1:rx2] = (255, 200, 50)
frame[ry2 - 2 : ry2, rx1:rx2] = (255, 200, 50)
# Arrow
if progress > 0.3:
frame[300:303, 310:380] = (150, 150, 170)
# Middle: ROI divided into 3x3 grid (output_size)
if progress > 0.3:
out_x, out_y = 400, 220
out_cell = 50
out_n = 3
roi_h = roi_r2 - roi_r1
roi_w = roi_c2 - roi_c1
for r in range(out_n):
for c in range(out_n):
x = out_x + c * out_cell
y = out_y + r * out_cell
# Compute the max from corresponding region
src_r1 = roi_r1 + r * roi_h // out_n
src_r2 = roi_r1 + (r + 1) * roi_h // out_n
src_c1 = roi_c1 + c * roi_w // out_n
src_c2 = roi_c1 + (c + 1) * roi_w // out_n
max_val = 0
for sr in range(src_r1, src_r2):
for sc in range(src_c1, src_c2):
v = 30 + ((sr * 7 + sc * 13 + 42) % 40)
max_val = max(max_val, v)
frame[y : y + out_cell - 2, x : x + out_cell - 2] = (
max_val,
max_val + 20,
max_val + 40,
)
frame[y : y + 2, x : x + out_cell - 2] = (80, 200, 120)
frame[y + out_cell - 4 : y + out_cell - 2, x : x + out_cell - 2] = (
80,
200,
120,
)
# Arrow to FC
if progress > 0.6:
frame[300:303, 560:630] = (150, 150, 170)
# FC box
frame[270:340, 650:730] = (200, 100, 80)
frame[270:272, 650:730] = (240, 140, 120)
frame[338:340, 650:730] = (240, 140, 120)
return frame
roi_clip = VideoClip(make_roi_frame, duration=STEP_DUR + 1).with_fps(FPS)
dur = STEP_DUR + 1 dur = STEP_DUR + 1
labels = [ labels = [
("ROI Pooling: kluczowa innowacja Fast R-CNN", 26, "#FFE082", FONT_B, (80, 20)), ("ROI Pooling: kluczowa innowacja Fast R-CNN", 26, "#FFE082", FONT_B, (80, 20)),
@ -731,7 +753,8 @@ def _roi_pooling_demo() -> list[CompositeVideoClip]:
(80, 535), (80, 535),
), ),
( (
"Fast R-CNN: CNN raz → 1 feature mapa → ROI Pool 2000 regionów → 25x szybciej!", "Fast R-CNN: CNN raz → 1 feature mapa → "
"ROI Pool 2000 regionów → 25x szybciej!",
16, 16,
"#A5D6A7", "#A5D6A7",
FONT_R, FONT_R,
@ -788,7 +811,6 @@ def _rpn_anchors_demo() -> list[CompositeVideoClip]:
# Draw anchors around center: 3 sizes x 3 ratios = 9 # Draw anchors around center: 3 sizes x 3 ratios = 9
anchor_specs = [ anchor_specs = [
# (half_w, half_h, color)
(30, 30, (200, 80, 80)), # small 1:1 (30, 30, (200, 80, 80)), # small 1:1
(20, 40, (200, 60, 60)), # small 1:2 (20, 40, (200, 60, 60)), # small 1:2
(40, 20, (180, 60, 60)), # small 2:1 (40, 20, (180, 60, 60)), # small 2:1
@ -1014,7 +1036,8 @@ def _yolo_demo() -> list[CompositeVideoClip]:
frame[y : y + 1, img_x : img_x + img_size] = (100, 100, 120) frame[y : y + 1, img_x : img_x + img_size] = (100, 100, 120)
# Highlight cells containing object centers # Highlight cells containing object centers
if progress > 0.3: car_phase = 0.3
if progress > car_phase:
# Car center ~ cell (1, 1) # Car center ~ cell (1, 1)
cx, cy = 1, 2 cx, cy = 1, 2
hx = img_x + cx * cell hx = img_x + cx * cell
@ -1023,7 +1046,8 @@ def _yolo_demo() -> list[CompositeVideoClip]:
frame[hy : hy + cell, hx : hx + cell].astype(int) + 40, 0, 255 frame[hy : hy + cell, hx : hx + cell].astype(int) + 40, 0, 255
).astype(np.uint8) ).astype(np.uint8)
if progress > 0.5: person_phase = 0.5
if progress > person_phase:
# Person center ~ cell (4, 4) # Person center ~ cell (4, 4)
cx, cy = 4, 4 cx, cy = 4, 4
hx = img_x + cx * cell hx = img_x + cx * cell
@ -1033,7 +1057,8 @@ def _yolo_demo() -> list[CompositeVideoClip]:
).astype(np.uint8) ).astype(np.uint8)
# Bounding boxes predictions from cells # Bounding boxes predictions from cells
if progress > 0.6: bbox_phase = 0.6
if progress > bbox_phase:
# Car bbox # Car bbox
for tt in range(2): for tt in range(2):
frame[ frame[
@ -1100,7 +1125,8 @@ def _yolo_demo() -> list[CompositeVideoClip]:
(80, 620), (80, 620),
), ),
( (
"Two-stage (R-CNN): propozycje+klasyfikacja | One-stage (YOLO): bez propozycji!", "Two-stage (R-CNN): propozycje+klasyfikacja "
"| One-stage (YOLO): bez propozycji!",
14, 14,
"#90CAF9", "#90CAF9",
FONT_R, FONT_R,
@ -1152,13 +1178,15 @@ def _yolo_architecture() -> list[CompositeVideoClip]:
frame[by + bh - 2 : by + bh, bx : bx + bw] = tuple( frame[by + bh - 2 : by + bh, bx : bx + bw] = tuple(
min(c + 50, 255) for c in color min(c + 50, 255) for c in color
) )
if i < 4: arrow_limit = 4
if i < arrow_limit:
ax = bx + bw + 5 ax = bx + bw + 5
ay = by + bh // 2 ay = by + bh // 2
frame[ay - 1 : ay + 2, ax : ax + 25] = (150, 150, 170) frame[ay - 1 : ay + 2, ax : ax + 25] = (150, 150, 170)
# Output tensor breakdown (right side) # Output tensor breakdown (right side)
if progress > 0.6: tensor_phase = 0.6
if progress > tensor_phase:
# Show SxS grid # Show SxS grid
gx, gy = 850, 180 gx, gy = 850, 180
gs = 120 gs = 120
@ -1282,18 +1310,21 @@ def _detr_demo() -> list[CompositeVideoClip]:
frame[by + bh - 2 : by + bh, bx : bx + bw] = tuple( frame[by + bh - 2 : by + bh, bx : bx + bw] = tuple(
min(c + 50, 255) for c in color min(c + 50, 255) for c in color
) )
if i < 4: arrow_limit = 4
if i < arrow_limit:
ax = bx + bw + 5 ax = bx + bw + 5
ay = by + bh // 2 ay = by + bh // 2
frame[ay - 1 : ay + 2, ax : ax + 25] = (150, 150, 170) frame[ay - 1 : ay + 2, ax : ax + 25] = (150, 150, 170)
# Object queries illustration (right side) # Object queries illustration (right side)
if progress > 0.5: query_phase = 0.5
if progress > query_phase:
qx, qy = 800, 140 qx, qy = 800, 140
for i in range(6): for i in range(6):
y = qy + i * 50 y = qy + i * 50
w = 130 w = 130
active = i < 3 active_limit = 3
active = i < active_limit
color = (80, 180, 120) if active else (60, 50, 50) color = (80, 180, 120) if active else (60, 50, 50)
frame[y : y + 35, qx : qx + w] = color frame[y : y + 35, qx : qx + w] = color
frame[y : y + 1, qx : qx + w] = tuple(min(c + 40, 255) for c in color) frame[y : y + 1, qx : qx + w] = tuple(min(c + 40, 255) for c in color)
@ -1528,7 +1559,8 @@ def _detr_demo() -> list[CompositeVideoClip]:
(80, 540), (80, 540),
), ),
( (
" R-CNN (SS+CNN+SVM+NMS) → YOLO (backbone+head+NMS) → DETR (backbone+transformer)", " R-CNN (SS+CNN+SVM+NMS) → YOLO "
"(backbone+head+NMS) → DETR (backbone+transformer)",
14, 14,
"#90CAF9", "#90CAF9",
FONT_R, FONT_R,
@ -1572,15 +1604,18 @@ def _nms_iou_demo() -> list[CompositeVideoClip]:
boxes.append((ox + 350, oy + 50, 100, 100, 0.40, (80, 180, 255))) boxes.append((ox + 350, oy + 50, 100, 100, 0.40, (80, 180, 255)))
for i, (bx, by, bw, bh, _conf, color) in enumerate(boxes): for i, (bx, by, bw, bh, _conf, color) in enumerate(boxes):
if progress > 0.4 and i > 0 and i < 3: dc = color
nms_phase = 0.4
nms_limit = 3
if progress > nms_phase and i > 0 and i < nms_limit:
# After NMS, these get removed (shown as faded/crossed) # After NMS, these get removed (shown as faded/crossed)
color = (60, 40, 40) dc = (60, 40, 40)
for tt in range(2): for tt in range(2):
frame[by - tt : by + bh + tt, bx - tt : bx - tt + 2] = color frame[by - tt : by + bh + tt, bx - tt : bx - tt + 2] = dc
frame[by - tt : by + bh + tt, bx + bw + tt - 2 : bx + bw + tt] = color frame[by - tt : by + bh + tt, bx + bw + tt - 2 : bx + bw + tt] = dc
frame[by - tt : by - tt + 2, bx - tt : bx + bw + tt] = color frame[by - tt : by - tt + 2, bx - tt : bx + bw + tt] = dc
frame[by + bh + tt - 2 : by + bh + tt, bx - tt : bx + bw + tt] = color frame[by + bh + tt - 2 : by + bh + tt, bx - tt : bx + bw + tt] = dc
# IoU visualization on right side # IoU visualization on right side
iou_x, iou_y = 700, 200 iou_x, iou_y = 700, 200
@ -1884,7 +1919,7 @@ def main() -> None:
final.write_videofile( final.write_videofile(
OUTPUT, fps=FPS, codec="libx264", audio=False, preset="medium", threads=4 OUTPUT, fps=FPS, codec="libx264", audio=False, preset="medium", threads=4
) )
print(f"Video saved to: {OUTPUT}") _logger.info("Video saved to: %s", OUTPUT)
if __name__ == "__main__": if __name__ == "__main__":