Paste some text and click Analyze.
"
tokens = tokenizer.encode(text)
if len(tokens) > 512:
tokens = tokens[:512]
with torch.no_grad():
input_ids = torch.tensor([tokens]).to(device)
all_logits = model(input_ids).logits[0].cpu()
css = """
"""
parts = [css, '']
for i in range(len(tokens)):
tok = html_lib.escape(tokenizer.decode([tokens[i]]))
if i == 0:
parts.append(f'
{tok}
')
continue
probs = torch.softmax(all_logits[i - 1], dim=-1)
actual_p = probs[tokens[i]].item()
top_p, top_idx = probs.topk(top_k)
color, bg = get_color(actual_p)
rank = None
alts = ""
for j in range(top_k):
a_text = html_lib.escape(tokenizer.decode([top_idx[j].item()]))
a_p = top_p[j].item()
hit = top_idx[j].item() == tokens[i]
if hit: rank = j + 1
cls = ' class="w hi"' if hit else ' class="w"'
pcls = ' class="p hi"' if hit else ' class="p"'
alts += f'
{a_text}{a_p:.4f}
'
rank_s = f"rank #{rank}" if rank else f"rank >{top_k}"
tooltip = f'''
“{tok}”
P = {actual_p:.4f} ({rank_s})
Top {top_k} alternatives
{alts}
'''
parts.append(f'
{tok}{tooltip}
')
parts.append('
Enter text and click Predict Next.
"
tokens = tokenizer.encode(text)
if len(tokens) > 512:
tokens = tokens[:512]
with torch.no_grad():
input_ids = torch.tensor([tokens]).to(device)
logits = model(input_ids).logits[0, -1].cpu()
probs = torch.softmax(logits, dim=-1)
log_probs = torch.log(probs)
top_p, top_idx = probs.topk(num_candidates)
top_lp = log_probs[top_idx]
rows = ""
for j in range(num_candidates):
tok_text = html_lib.escape(tokenizer.decode([top_idx[j].item()]))
p = top_p[j].item()
lp = top_lp[j].item()
bar_width = max(1, int(p * 100))
hue = min(p * 120, 120)
rows += f"""