Skip to content

D
DeepSpeech

PaddlePaddle / DeepSpeech
大约 2 年 前同步成功

通知 210
Star 8425
Fork 1598
D
DeepSpeech
提交 231499a3 编写于 作者: H Hui Zhang
浏览文件
操作

VarBase.__getitem__ work for np.int64, np.longlong; but __setitem_varbase__... 

VarBase.__getitem__ work for np.int64, np.longlong; but __setitem_varbase__ not support paddle.int16/set_value op not support
上级 c7508229
隐藏空白更改
内联 并排
Showing with 31 addition and 33 deletion
deepspeech/exps/u2/model.py
浏览文件 @ 231499a3
...@@ -568,26 +568,25 @@ class U2Tester(U2Trainer): ...@@ -568,26 +568,25 @@ class U2Tester(U2Trainer):
ctc_probs = ctc_probs.squeeze(0) ctc_probs = ctc_probs.squeeze(0)
target = target.squeeze(0) target = target.squeeze(0)
alignment = ctc_utils.forced_align(ctc_probs, target) alignment = ctc_utils.forced_align(ctc_probs, target)
logger.info("align ids", key[0], alignment) logger.info(f"align ids: {key[0]} {alignment}")
fout.write('{} {}\n'.format(key[0], alignment)) fout.write('{} {}\n'.format(key[0], alignment))
# 3. gen praat # 3. gen praat
# segment alignment # segment alignment
align_segs = text_grid.segment_alignment(alignment) align_segs = text_grid.segment_alignment(alignment)
logger.info("align tokens", key[0], align_segs) logger.info(f"align tokens: {key[0]}, {align_segs}")
# IntervalTier, List["start end token\n"] # IntervalTier, List["start end token\n"]
subsample = utility.get_subsample(self.config) subsample = utility.get_subsample(self.config)
tierformat = text_grid.align_to_tierformat( tierformat = text_grid.align_to_tierformat(
align_segs, subsample, token_dict) align_segs, subsample, token_dict)
# write tier # write tier
align_output_path = os.path.join( align_output_path = Path(self.args.result_file).parent / "align"
os.path.dirname(self.args.result_file), "align") align_output_path.mkdir(parents=True, exist_ok=True)
tier_path = os.path.join(align_output_path, key[0] + ".tier") tier_path = align_output_path / (key[0] + ".tier")
with open(tier_path, 'w') as f: with tier_path.open('w') as f:
f.writelines(tierformat) f.writelines(tierformat)
# write textgrid # write textgrid
textgrid_path = os.path.join(align_output_path, textgrid_path = align_output_path / (key[0] + ".TextGrid")
key[0] + ".TextGrid")
second_per_frame = 1. / (1000. / second_per_frame = 1. / (1000. /
stride_ms) # 25ms window, 10ms stride stride_ms) # 25ms window, 10ms stride
second_per_example = ( second_per_example = (
...@@ -595,7 +594,7 @@ class U2Tester(U2Trainer): ...@@ -595,7 +594,7 @@ class U2Tester(U2Trainer):
text_grid.generate_textgrid( text_grid.generate_textgrid(
maxtime=second_per_example, maxtime=second_per_example,
intervals=tierformat, intervals=tierformat,
output=textgrid_path) output=str(textgrid_path))
def run_align(self): def run_align(self):
self.resume_or_scratch() self.resume_or_scratch()
......
deepspeech/exps/u2_kaldi/model.py
浏览文件 @ 231499a3
...@@ -545,9 +545,8 @@ class U2Tester(U2Trainer): ...@@ -545,9 +545,8 @@ class U2Tester(U2Trainer):
self.model.eval() self.model.eval()
logger.info(f"Align Total Examples: {len(self.align_loader.dataset)}") logger.info(f"Align Total Examples: {len(self.align_loader.dataset)}")
stride_ms = self.config.collater.stride_ms stride_ms = self.align_loader.collate_fn.stride_ms
token_dict = self.args.char_list token_dict = self.align_loader.collate_fn.vocab_list
with open(self.args.result_file, 'w') as fout: with open(self.args.result_file, 'w') as fout:
# one example in batch # one example in batch
for i, batch in enumerate(self.align_loader): for i, batch in enumerate(self.align_loader):
...@@ -564,26 +563,25 @@ class U2Tester(U2Trainer): ...@@ -564,26 +563,25 @@ class U2Tester(U2Trainer):
ctc_probs = ctc_probs.squeeze(0) ctc_probs = ctc_probs.squeeze(0)
target = target.squeeze(0) target = target.squeeze(0)
alignment = ctc_utils.forced_align(ctc_probs, target) alignment = ctc_utils.forced_align(ctc_probs, target)
logger.info("align ids", key[0], alignment) logger.info(f"align ids: {key[0]} {alignment}")
fout.write('{} {}\n'.format(key[0], alignment)) fout.write('{} {}\n'.format(key[0], alignment))
# 3. gen praat # 3. gen praat
# segment alignment # segment alignment
align_segs = text_grid.segment_alignment(alignment) align_segs = text_grid.segment_alignment(alignment)
logger.info("align tokens", key[0], align_segs) logger.info(f"align tokens: {key[0]}, {align_segs}")
# IntervalTier, List["start end token\n"] # IntervalTier, List["start end token\n"]
subsample = utility.get_subsample(self.config) subsample = utility.get_subsample(self.config)
tierformat = text_grid.align_to_tierformat( tierformat = text_grid.align_to_tierformat(
align_segs, subsample, token_dict) align_segs, subsample, token_dict)
# write tier # write tier
align_output_path = os.path.join( align_output_path = Path(self.args.result_file).parent / "align"
os.path.dirname(self.args.result_file), "align") align_output_path.mkdir(parents=True, exist_ok=True)
tier_path = os.path.join(align_output_path, key[0] + ".tier") tier_path = align_output_path / (key[0] + ".tier")
with open(tier_path, 'w') as f: with tier_path.open('w') as f:
f.writelines(tierformat) f.writelines(tierformat)
# write textgrid # write textgrid
textgrid_path = os.path.join(align_output_path, textgrid_path = align_output_path / (key[0] + ".TextGrid")
key[0] + ".TextGrid")
second_per_frame = 1. / (1000. / second_per_frame = 1. / (1000. /
stride_ms) # 25ms window, 10ms stride stride_ms) # 25ms window, 10ms stride
second_per_example = ( second_per_example = (
...@@ -591,7 +589,7 @@ class U2Tester(U2Trainer): ...@@ -591,7 +589,7 @@ class U2Tester(U2Trainer):
text_grid.generate_textgrid( text_grid.generate_textgrid(
maxtime=second_per_example, maxtime=second_per_example,
intervals=tierformat, intervals=tierformat,
output=textgrid_path) output=str(textgrid_path))
def run_align(self): def run_align(self):
self.resume_or_scratch() self.resume_or_scratch()
......
deepspeech/exps/u2_st/model.py
浏览文件 @ 231499a3
...@@ -595,26 +595,25 @@ class U2STTester(U2STTrainer): ...@@ -595,26 +595,25 @@ class U2STTester(U2STTrainer):
ctc_probs = ctc_probs.squeeze(0) ctc_probs = ctc_probs.squeeze(0)
target = target.squeeze(0) target = target.squeeze(0)
alignment = ctc_utils.forced_align(ctc_probs, target) alignment = ctc_utils.forced_align(ctc_probs, target)
logger.info("align ids", key[0], alignment) logger.info(f"align ids: {key[0]} {alignment}")
fout.write('{} {}\n'.format(key[0], alignment)) fout.write('{} {}\n'.format(key[0], alignment))
# 3. gen praat # 3. gen praat
# segment alignment # segment alignment
align_segs = text_grid.segment_alignment(alignment) align_segs = text_grid.segment_alignment(alignment)
logger.info("align tokens", key[0], align_segs) logger.info(f"align tokens: {key[0]}, {align_segs}")
# IntervalTier, List["start end token\n"] # IntervalTier, List["start end token\n"]
subsample = utility.get_subsample(self.config) subsample = utility.get_subsample(self.config)
tierformat = text_grid.align_to_tierformat( tierformat = text_grid.align_to_tierformat(
align_segs, subsample, token_dict) align_segs, subsample, token_dict)
# write tier # write tier
align_output_path = os.path.join( align_output_path = Path(self.args.result_file).parent / "align"
os.path.dirname(self.args.result_file), "align") align_output_path.mkdir(parents=True, exist_ok=True)
tier_path = os.path.join(align_output_path, key[0] + ".tier") tier_path = align_output_path / (key[0] + ".tier")
with open(tier_path, 'w') as f: with tier_path.open('w') as f:
f.writelines(tierformat) f.writelines(tierformat)
# write textgrid # write textgrid
textgrid_path = os.path.join(align_output_path, textgrid_path = align_output_path / (key[0] + ".TextGrid")
key[0] + ".TextGrid")
second_per_frame = 1. / (1000. / second_per_frame = 1. / (1000. /
stride_ms) # 25ms window, 10ms stride stride_ms) # 25ms window, 10ms stride
second_per_example = ( second_per_example = (
...@@ -622,7 +621,7 @@ class U2STTester(U2STTrainer): ...@@ -622,7 +621,7 @@ class U2STTester(U2STTrainer):
text_grid.generate_textgrid( text_grid.generate_textgrid(
maxtime=second_per_example, maxtime=second_per_example,
intervals=tierformat, intervals=tierformat,
output=textgrid_path) output=str(textgrid_path))
def run_align(self): def run_align(self):
self.resume_or_scratch() self.resume_or_scratch()
......
deepspeech/utils/ctc_utils.py
浏览文件 @ 231499a3
...@@ -86,8 +86,10 @@ def forced_align(ctc_probs: paddle.Tensor, y: paddle.Tensor, ...@@ -86,8 +86,10 @@ def forced_align(ctc_probs: paddle.Tensor, y: paddle.Tensor,
log_alpha = paddle.zeros( log_alpha = paddle.zeros(
(ctc_probs.size(0), len(y_insert_blank))) #(T, 2L+1) (ctc_probs.size(0), len(y_insert_blank))) #(T, 2L+1)
log_alpha = log_alpha - float('inf') # log of zero log_alpha = log_alpha - float('inf') # log of zero
# self.__setitem_varbase__(item, value) When assign a value to a paddle.Tensor, the data type of the paddle.Tensor not support int16
state_path = (paddle.zeros( state_path = (paddle.zeros(
(ctc_probs.size(0), len(y_insert_blank)), dtype=paddle.int16) - 1 (ctc_probs.size(0), len(y_insert_blank)), dtype=paddle.int32) - 1
) # state path, Tuple((T, 2L+1)) ) # state path, Tuple((T, 2L+1))
# init start state # init start state
...@@ -111,8 +113,8 @@ def forced_align(ctc_probs: paddle.Tensor, y: paddle.Tensor, ...@@ -111,8 +113,8 @@ def forced_align(ctc_probs: paddle.Tensor, y: paddle.Tensor,
log_alpha[t, s] = paddle.max(candidates) + ctc_probs[t][ log_alpha[t, s] = paddle.max(candidates) + ctc_probs[t][
y_insert_blank[s]] y_insert_blank[s]]
state_path[t, s] = prev_state[paddle.argmax(candidates)] state_path[t, s] = prev_state[paddle.argmax(candidates)]
# self.__setitem_varbase__(item, value) When assign a value to a paddle.Tensor, the data type of the paddle.Tensor not support int16
state_seq = -1 * paddle.ones((ctc_probs.size(0), 1), dtype=paddle.int16) state_seq = -1 * paddle.ones((ctc_probs.size(0), 1), dtype=paddle.int32)
candidates = paddle.to_tensor([ candidates = paddle.to_tensor([
log_alpha[-1, len(y_insert_blank) - 1], # Sb log_alpha[-1, len(y_insert_blank) - 1], # Sb
......
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或