先拆开两个容易混淆的问题 · 实现代码

# Detached readonly runs write only to an isolated ephemeral directory.
import os, tempfile
_LAB_WORKDIR = tempfile.mkdtemp(prefix="lecture03-lab13-")
os.chdir(_LAB_WORKDIR)
print("ephemeral workdir:", _LAB_WORKDIR)

seqs=[[1,4,5,2],[1,6,2,0]]
print('warm-up input',[s[:-1] for s in seqs])
print('warm-up target',[s[1:] for s in seqs])
print('warm-up loss mask',[[int(t!=0) for t in s[1:]] for s in seqs])
# Every outgoing distribution sums to one. Both paths terminate after two tokens.
tree={():{'A':.55,'B':.45},('A',):{'X':.45,'Z':.55},('B',):{'Y':.80,'Z':.20}}
assert all(abs(sum(p.values())-1)<1e-9 for p in tree.values())
greedy_path=();greedy_probability=1.
for _ in range(2):
    token=max(tree[greedy_path],key=tree[greedy_path].get)
    greedy_probability*=tree[greedy_path][token];greedy_path+= (token,)
beams=[((),1.)]
for _ in range(2):
    beams=sorted([(seq+(word,),p*conditional) for seq,p in beams for word,conditional in tree[seq].items()],key=lambda b:b[1],reverse=True)[:2]
print('actual tree greedy:',greedy_path,greedy_probability,'beam2:',beams[0])