Final final updates to app.py

app.py

@@ -207,6 +207,20 @@ def read_input(file, text):
     return open(path, encoding="utf-8", errors="replace").read()
 
 
+def bar(k, n, colour):
+    #count + a thin bar showing what share of the segments got flagged
+    return (f"<td><div class='cell'><b>{k}</b><span class='bar'>"
+            f"<span style='width:{k / n * 100:.0f}%;background:{colour}'></span></span></div></td>")
+
+
+def meter(p, thr):
+    #track from 0 to 1, fill = top bert score, white tick = threshold
+    colour = "#818cf8" if p >= thr else "#4b5370"
+    return (f"<div class='cell'><span class='meter'><span class='fill' style='width:{p * 100:.0f}%;background:{colour}'></span>"
+            f"<span class='tick' style='left:{thr * 100:.0f}%'></span></span>"
+            f"<span class='small'>{p:.2f} / {thr:.2f}</span></div>")
+
+
 @GPU
 def analyse(file, text):
     raw = read_input(file, text)
@@ -232,6 +246,9 @@ def analyse(file, text):
         s3 = (f"FAILED: interviewer found ({', '.join(removed)}) but nothing was left after removing them, "
               f"because the astronaut's name tag was not recognised.")
     s3 += f" Words left: {len(narrative.split()):,}."
+    if method == "no_tags":
+        #no cover page and no NAME: tags --> probably not an oral history, the models still run but read with care
+        s3 += " Note: this does not look like a NASA oral history transcript, so the scores below are less meaningful."
 
     #step 4
     segs = segment(narrative)
@@ -243,6 +260,11 @@ def analyse(file, text):
         return s1, s2, s3, "<h4>4. Segmenting</h4>FAILED: no text left to segment.", "", None, empty
     segs = segs[:300]   #keeps a free cpu quick
     wc = [len(s.split()) for s in segs]
+    long = sum(w > MAX_WORDS for w in wc)
+    if long:
+        #one "sentence" over 300 words (tables, references, no full stops) can't be split, bert only reads the first 512 tokens
+        note += (f"<br><span class='small'>{long} segment{'s are' if long != 1 else ' is'} over {MAX_WORDS} words because the text "
+                 f"has very long runs with no full stop, BERT only reads the first 512 tokens of those.</span>")
     s4 = (f"<h4>4. Segmenting</h4><b>{len(segs)} segment{"s" if len(segs) != 1 else ""}</b>, {min(wc)} to {max(wc)} words each{note}<br><br>" +
           "".join(f"<div class='seg'><b>#{i + 1}</b> ({wc[i]} words) {html.escape(s[:220])}{'...' if len(s) > 220 else ''}</div>"
                   for i, s in enumerate(segs[:3])) +
@@ -252,14 +274,17 @@ def analyse(file, text):
     v = np.array([vader_flags(s) for s in segs])
     t = tfidf_probs(segs)
     b = bert_probs(segs)
+    n = len(segs)
     rows = ""
     for i, c in enumerate(CATEGORIES):
-        rows += (f"<tr><td>{NAMES[c]}</td><td>{int(v[:, i].sum())}</td><td>{int((t[:, i] >= 0.5).sum())}</td>"
-                 f"<td><b>{int((b[:, i] >= BERT_THRESHOLDS[c]).sum())}</b></td><td>{b[:, i].max():.2f} / {BERT_THRESHOLDS[c]:.2f}</td></tr>")
-    s5 = (f"<h4>5. Classification: all three models</h4><table><tr><th>Category</th><th>VADER</th><th>TF-IDF</th><th>BERT + LoRA</th>"
-          f"<th>BERT top prob / threshold</th></tr>{rows}</table>"
-          f"<p class='small'>Numbers are segments flagged out of {len(segs)}. BERT is my final model. "
-          f"A research prototype, it flags passages for a person to read and does not diagnose anyone.</p>")
+        counts = [int(v[:, i].sum()), int((t[:, i] >= 0.5).sum()), int((b[:, i] >= BERT_THRESHOLDS[c]).sum())]
+        cells = "".join(bar(k, n, col) for k, col in zip(counts, ["#7c86a8", "#38bdf8", "#818cf8"]))
+        rows += f"<tr><td>{NAMES[c]}</td>{cells}<td>{meter(b[:, i].max(), BERT_THRESHOLDS[c])}</td></tr>"
+    s5 = (f"<h4>5. Classification: all three models</h4>"
+          f"<table><tr><th>Category</th><th>VADER</th><th>TF-IDF</th><th>BERT + LoRA</th><th>BERT top score vs threshold</th></tr>{rows}</table>"
+          f"<p class='small'>Segments flagged out of {n}. White line = BERT threshold, the bar fills to the top segment score "
+          f"(indigo if it passes). BERT is my final model. A research prototype, it flags passages for a person to read "
+          f"and does not diagnose anyone.</p>")
 
     order = np.argsort(-b.max(axis=1))   #strongest segments first for shap
     choices = [f"#{k + 1}  (BERT max {b[k].max():.2f})  {segs[k][:60]}..." for k in order]
@@ -294,9 +319,15 @@ CSS = """
 .gradio-container {background: radial-gradient(ellipse at top, #141b33 0%, #05070f 70%) !important;}
 * {box-shadow: none !important;}
 .seg {border: 1px solid #ffffff44; border-radius: 12px; padding: 8px 12px; margin: 6px 0; color: #e6ebff;}
-table {width: 100%; border-collapse: collapse; color: #e6ebff;}
-th, td {border-bottom: 1px solid #ffffff33; padding: 6px 8px; text-align: left;}
+table {width: 100%; border-collapse: collapse; border: none !important; color: #e6ebff;}
+th, td {border: none !important; border-bottom: 1px solid #ffffff33 !important; padding: 6px 8px; text-align: left;}
 .small {color: #9aa4c4; font-size: 13px;}
+.cell {display: flex; align-items: center; gap: 8px;}
+.bar {flex: 1; min-width: 40px; height: 6px; background: #ffffff14; border-radius: 3px; overflow: hidden;}
+.bar span {display: block; height: 100%;}
+.meter {position: relative; flex: 1; min-width: 80px; height: 8px; background: #ffffff14; border-radius: 4px;}
+.meter .fill {position: absolute; left: 0; top: 0; bottom: 0; border-radius: 4px;}
+.meter .tick {position: absolute; top: -4px; bottom: -4px; width: 2px; background: #fff;}
 """
 
 with gr.Blocks(title="SPACEMIND") as demo: