def decode_best_path(probs, ref=None, blank=0):
"""
解码最佳路径
"""
# 计算最可能的path
best_path = np.argmax(probs,axis=0).tolist()
# 折叠phone
hyp = []
for i,b in enumerate(best_path):
# 忽略blank
if b == blank:
continue
# 忽略重复label
elif i != 0 and b == best_path[i-1]:
continue
else:
hyp.append(b)
# 计算错误率
dist = 0
if ref is not None:
ref = ref.tolist()
dist,_,_,_,_ = ed.edit_distance(ref,hyp)
return hyp,dist
def decode_best_path(probs, ref=None, blank=0):
"""
Computes best path given sequence of probability distributions per frame.
Simply chooses most likely label at each timestep then collapses result to
remove blanks and repeats.
Optionally computes edit distance between reference transcription
and best path if reference provided.
Returns hypothesis transcription and edit distance if requested.
"""
# Compute best path
best_path = np.argmax(probs,axis=0).tolist()
# Collapse phone string
hyp = []
for i,b in enumerate(best_path):
# ignore blanks
if b == blank:
continue
# ignore repeats
elif i != 0 and b == best_path[i-1]:
continue
else:
hyp.append(b)
# Optionally compute phone error rate to ground truth
dist = 0
if ref is not None:
ref = ref.tolist()
dist,_,_,_,_ = ed.edit_distance(ref,hyp)
return hyp,dist
def main(args):
with open(args.pk_file1, 'rb') as fin:
hyps1 = pickle.load(fin)
refs = pickle.load(fin)
pickle.load(fin) # hypscores
pickle.load(fin) # refscores
numphones = pickle.load(fin)
with open(args.pk_file2, 'rb') as fin:
hyps2 = pickle.load(fin)
assert len(hyps1) == len(hyps2), 'hyps have different lengths'
differ = difflib.Differ()
num_diff = 0
hyp1_better = 0
hyp2_better = 0
for (hyp1, hyp2, ref) in zip(hyps1, hyps2, refs):
if hyp1 == hyp2:
continue
num_diff += 1
label1 = 'hyp1:'
label2 = 'hyp2:'
if args.score:
dist1, _, _, _, _ = ed.edit_distance(ref, hyp1)
dist2, _, _, _, _ = ed.edit_distance(ref, hyp2)
if dist1 < dist2:
hyp1_better += 1
label1 = blue(label1)
label2 = red(label2)
else:
hyp2_better += 1
label1 = red(label1)
label2 = blue(label2)
print label1, collapse_seq(hyp1)
print label2, collapse_seq(hyp2)
pprint(list(differ.compare([collapse_seq(hyp1)],
[collapse_seq(hyp2)])))
print green(' ref:'), collapse_seq(ref)
print '-' * 80
if args.score:
print 'hyp1 better: %d' % hyp1_better
print 'hyp2 better: %d' % hyp2_better
print 'Differ on %d/%d utts' % (num_diff, len(refs))
def main(args):
with open(args.pk_file1, 'rb') as fin:
hyps1 = pickle.load(fin)
refs = pickle.load(fin)
pickle.load(fin) # hypscores
pickle.load(fin) # refscores
numphones = pickle.load(fin)
with open(args.pk_file2, 'rb') as fin:
hyps2 = pickle.load(fin)
assert len(hyps1) == len(hyps2), 'hyps have different lengths'
differ = difflib.Differ()
num_diff = 0
hyp1_better = 0
hyp2_better = 0
for (hyp1, hyp2, ref) in zip(hyps1, hyps2, refs):
if hyp1 == hyp2:
continue
num_diff += 1
label1 = 'hyp1:'
label2 = 'hyp2:'
if args.score:
dist1, _, _, _, _ = ed.edit_distance(ref, hyp1)
dist2, _, _, _, _ = ed.edit_distance(ref, hyp2)
if dist1 < dist2:
hyp1_better += 1
label1 = blue(label1)
label2 = red(label2)
else:
hyp2_better += 1
label1 = red(label1)
label2 = blue(label2)
print label1, collapse_seq(hyp1)
print label2, collapse_seq(hyp2)
pprint(list(differ.compare([collapse_seq(hyp1)], [collapse_seq(hyp2)])))
print green(' ref:'), collapse_seq(ref)
print '-' * 80
if args.score:
print 'hyp1 better: %d' % hyp1_better
print 'hyp2 better: %d' % hyp2_better
print 'Differ on %d/%d utts' % (num_diff, len(refs))
def test(opts):
import editDistance as ed
print "Testing model %s" % opts.inFile
phone_map = get_phone_map_swbd()
with open(opts.inFile, 'r') as fid:
old_opts = pickle.load(fid)
_ = pickle.load(fid)
_ = pickle.load(fid)
loader = dl.DataLoader(opts.dataDir, old_opts.rawDim,
old_opts.inputDim)
if 'layers' not in dir(old_opts):
old_opts.layers = [old_opts.layerSize] * old_opts.numLayers
nn = nnet.NNet(old_opts.inputDim,
old_opts.outputDim,
old_opts.layers,
train=False)
nn.initParams()
nn.fromFile(fid)
totdist = numphones = 0
fid = open('hyp.txt', 'w')
for i in range(1, opts.numFiles + 1):
data_dict, alis, keys, sizes = loader.loadDataFileDict(i)
for k in keys:
gp.free_reuse_cache()
hyp = nn.costAndGrad(data_dict[k])
hyp = [phone_map[h] for h in hyp]
ref = [phone_map[int(r)] for r in alis[k]]
dist, ins, dels, subs, corr = ed.edit_distance(ref, hyp)
print "Distance %d/%d" % (dist, len(ref))
fid.write(k + ' ' + ' '.join(hyp) + '\n')
totdist += dist
numphones += len(alis[k])
fid.close()
print "PER : %f" % (100 * totdist / float(numphones))
def test(opts):
import editDistance as ed
print "Testing model %s"%opts.inFile
phone_map = get_phone_map_swbd()
with open(opts.inFile,'r') as fid:
old_opts = pickle.load(fid)
_ = pickle.load(fid)
_ = pickle.load(fid)
loader = dl.DataLoader(opts.dataDir,old_opts.rawDim,old_opts.inputDim)
if 'layers' not in dir(old_opts):
old_opts.layers = [old_opts.layerSize]*old_opts.numLayers
nn = nnet.NNet(old_opts.inputDim,old_opts.outputDim,old_opts.layers,train=False)
nn.initParams()
nn.fromFile(fid)
totdist = numphones = 0
fid = open('hyp.txt','w')
for i in range(1,opts.numFiles+1):
data_dict,alis,keys,sizes = loader.loadDataFileDict(i)
for k in keys:
gp.free_reuse_cache()
hyp = nn.costAndGrad(data_dict[k])
hyp = [phone_map[h] for h in hyp]
ref = [phone_map[int(r)] for r in alis[k]]
dist,ins,dels,subs,corr = ed.edit_distance(ref,hyp)
print "Distance %d/%d"%(dist,len(ref))
fid.write(k+' '+' '.join(hyp)+'\n')
totdist += dist
numphones += len(alis[k])
fid.close()
print "PER : %f"%(100*totdist/float(numphones))
