thuanan's picture
Create app.py
01b145c verified
import gradio as gr
from typing import Optional, Tuple, Any
import numpy as np
from PIL import Image, ImageOps
def _to_pil(img: Any) -> Optional[Any]:
if img is None:
return None
if Image is None:
raise RuntimeError("Pillow not available. Please install 'pillow'.")
if isinstance(img, Image.Image):
return img
arr = np.asarray(img)
if not (arr.ndim == 2 or (arr.ndim == 3 and arr.shape[2] in (3, 4))):
raise ValueError("Unsupported image array shape")
return Image.fromarray(arr.astype(np.uint8))
def preprocess_image(img: Any, max_side: int = 512, progress: Optional[gr.Progress] = None) -> Optional[Any]:
if img is None:
gr.Warning("Please upload an image first.")
return None
if progress:
progress(0, desc="Loading image…")
pil = _to_pil(img)
if pil is None:
return None
if progress:
progress(0.3, desc="Resizing…")
# Keep aspect ratio, cap the longest side
w, h = pil.size
scale = min(1.0, max_side / max(w, h))
if scale < 1.0:
pil = pil.resize((int(w * scale), int(h * scale)))
if progress:
progress(0.7, desc="Auto-contrast…")
pil = ImageOps.autocontrast(pil)
if progress:
progress(1.0, desc="Done")
return pil
def detect_edges(img: Any, strength: float = 1.0, progress: Optional[gr.Progress] = None) -> Optional[Any]:
if img is None:
gr.Warning("Please run Preprocess first or upload an image.")
return None
pil = _to_pil(img).convert("L") # grayscale
if progress:
progress(0.2, desc="Computing gradients…")
arr = np.asarray(pil, dtype=np.float32)
# Use numpy gradient as a simple edge detector (fast and dependency-free)
gy, gx = np.gradient(arr)
mag = np.hypot(gx, gy)
mag *= (255.0 / (mag.max() + 1e-6))
if progress:
progress(0.7, desc="Applying strength…")
mag = np.clip(mag * float(max(0.1, strength)), 0, 255).astype(np.uint8)
out = Image.fromarray(mag)
if progress:
progress(1.0, desc="Done")
return out
def enhance_image(img: Any, progress: Optional[gr.Progress] = None) -> Optional[Any]:
if img is None:
gr.Warning("Please run Detect Edges first.")
return None
pil = _to_pil(img)
if progress:
progress(0.5, desc="Enhancing…")
# Simple enhancement via auto-contrast again; could be extended
pil = ImageOps.autocontrast(pil)
if progress:
progress(1.0, desc="Done")
return pil
def run_all_image(image: Any, strength: float = 1.0, progress: Optional[gr.Progress] = None):
if image is None:
gr.Warning("Please upload an image.")
return None, None, None
# Use the same progress object for simplicity
p = preprocess_image(image, progress=progress)
e = detect_edges(p, strength=strength, progress=progress)
h = enhance_image(e, progress=progress)
return p, e, h
# -----------------------
# Text pipeline helpers
# -----------------------
def clean_text(text: str) -> str:
if not text:
gr.Warning("Please enter text.")
return ""
# Normalize whitespace and quotes
cleaned = " ".join(text.strip().split())
return cleaned
def summarize_text(text: str, max_sentences: int = 2) -> str:
if not text:
gr.Warning("Please clean the text first.")
return ""
# Naive sentence-based summarization: pick first N sentences
import re
sents = re.split(r"(?<=[.!?])\s+", text)
summary = " ".join(sents[: max(1, int(max_sentences))])
return summary
def sentiment(text: str) -> Tuple[str, float]:
if not text:
gr.Warning("Please provide text.")
return ("neutral", 0.0)
# Tiny lexicon-based scorer
pos = {"good", "great", "excellent", "amazing", "love", "like", "happy", "awesome", "fantastic"}
neg = {"bad", "terrible", "awful", "hate", "dislike", "sad", "poor", "horrible", "worse"}
words = [w.strip(".,!?;:").lower() for w in text.split()]
score = sum(1 for w in words if w in pos) - sum(1 for w in words if w in neg)
label = "positive" if score > 0 else ("negative" if score < 0 else "neutral")
# Normalize score into [-1, 1] by a simple squash
norm = max(1.0, len(words) / 10.0)
val = float(score / norm)
# Clamp to [-1, 1]
val = max(-1.0, min(1.0, val))
return (label, val)
with gr.Blocks(title="Complex Multi-step Workflows", theme=gr.themes.Soft()) as demo:
gr.Markdown("""
# Complex Apps with Gradio Blocks
Multi-step workflows across image and text pipelines. Each step updates state and UI.
""")
with gr.Tabs():
# ---------------- Image pipeline tab ----------------
with gr.TabItem("Image Pipeline"):
with gr.Row():
with gr.Column(scale=1):
image_in = gr.Image(label="Upload Image", type="pil")
strength = gr.Slider(0.1, 3.0, value=1.0, step=0.1, label="Edge Strength")
# Removed Demo Delay slider
with gr.Row():
btn_pre = gr.Button("Step 1: Preprocess")
btn_edge = gr.Button("Step 2: Detect Edges")
btn_enh = gr.Button("Step 3: Enhance")
with gr.Row():
btn_run_all = gr.Button("Run All", variant="primary")
btn_reset_img = gr.Button("Reset")
# Internal states to pass between steps
st_pre = gr.State()
st_edge = gr.State()
with gr.Column(scale=1):
out_pre = gr.Image(label="Preprocessed", interactive=False)
out_edge = gr.Image(label="Edges", interactive=False)
out_enh = gr.Image(label="Enhanced", interactive=False)
# Wiring events for image pipeline
def _preprocess_and_store(img, progress=gr.Progress(track_tqdm=True)):
p = preprocess_image(img, progress=progress)
return p, p
btn_pre.click(_preprocess_and_store, inputs=[image_in], outputs=[out_pre, st_pre])
def _edge_and_store(img_pre, k, progress=gr.Progress(track_tqdm=True)):
if img_pre is None:
gr.Warning("Run Step 1 first.")
return None, None
e = detect_edges(img_pre, strength=k, progress=progress)
return e, e
btn_edge.click(_edge_and_store, inputs=[st_pre, strength], outputs=[out_edge, st_edge])
def _enhance(img_edge, progress=gr.Progress(track_tqdm=True)):
if img_edge is None:
gr.Warning("Run Step 2 first.")
return None
return enhance_image(img_edge, progress=progress)
btn_enh.click(_enhance, inputs=[st_edge], outputs=out_enh)
def _run_all(img, k, progress=gr.Progress(track_tqdm=True)):
p, e, h = run_all_image(img, k, progress=progress)
# Also store states for continuity
return p, e, h, p, e
btn_run_all.click(_run_all, inputs=[image_in, strength], outputs=[out_pre, out_edge, out_enh, st_pre, st_edge])
def _reset_img():
return None, None, None, None, None
btn_reset_img.click(_reset_img, outputs=[image_in, out_pre, out_edge, out_enh, st_pre])
# ---------------- Text pipeline tab ----------------
with gr.TabItem("Text Pipeline"):
with gr.Row():
with gr.Column(scale=1):
text_in = gr.Textbox(label="Input Text", lines=8, placeholder="Paste or type some text…")
with gr.Accordion("Options", open=False):
max_sents = gr.Slider(1, 5, value=2, step=1, label="Summary Sentences")
with gr.Row():
btn_clean = gr.Button("Step 1: Clean")
btn_sum = gr.Button("Step 2: Summarize")
btn_sent = gr.Button("Step 3: Sentiment")
with gr.Row():
btn_run_all_txt = gr.Button("Run All", variant="primary")
btn_reset_txt = gr.Button("Reset")
st_clean = gr.State()
st_sum = gr.State()
with gr.Column(scale=1):
out_clean = gr.Textbox(label="Cleaned Text", lines=8)
out_sum = gr.Textbox(label="Summary", lines=6)
out_sent = gr.Label(label="Sentiment")
# Wiring events for text pipeline
def _clean_and_store(t):
c = clean_text(t)
return c, c
btn_clean.click(_clean_and_store, inputs=text_in, outputs=[out_clean, st_clean])
def _summarize_and_store(c, n):
if not c:
gr.Warning("Run Step 1 first.")
return "", ""
s = summarize_text(c, int(n))
return s, s
btn_sum.click(_summarize_and_store, inputs=[st_clean, max_sents], outputs=[out_sum, st_sum])
def _sentiment(s):
if not s:
gr.Warning("Run Step 2 first.")
return {"positive": 0.0, "neutral": 1.0, "negative": 0.0}
label, score = sentiment(s)
# Map score in [-1,1] to a 3-class distribution
p_pos = max(0.0, score)
p_neg = max(0.0, -score)
p_neu = 1.0 - abs(score)
return {"positive": round(p_pos, 3), "neutral": round(p_neu, 3), "negative": round(p_neg, 3)}
btn_sent.click(_sentiment, inputs=st_sum, outputs=out_sent)
def _run_all_txt(t, n):
c = clean_text(t)
s = summarize_text(c, int(n))
label, score = sentiment(s)
p_pos = max(0.0, score)
p_neg = max(0.0, -score)
p_neu = 1.0 - abs(score)
return c, s, {"positive": round(p_pos, 3), "neutral": round(p_neu, 3), "negative": round(p_neg, 3)}, c, s
btn_run_all_txt.click(_run_all_txt, inputs=[text_in, max_sents], outputs=[out_clean, out_sum, out_sent, st_clean, st_sum])
def _reset_txt():
return "", "", None, "", ""
btn_reset_txt.click(_reset_txt, outputs=[text_in, out_sum, out_sent, st_clean, st_sum])
demo.queue().launch()