Dev gzf exp (#1654)

* sensevoice finetune * sensevoice finetune * sensevoice finetune * sensevoice finetune * sensevoice finetune * sensevoice finetune * sensevoice finetune * sensevoice finetune * sensevoice finetune * sensevoice finetune * bugfix * update with main (#1631) * update seaco finetune * v1.0.24 --------- Co-authored-by: 维石 <[email protected]> * sensevoice * sensevoice * sensevoice * update with main (#1638) * update seaco finetune * v1.0.24 * update rwkv template --------- Co-authored-by: 维石 <[email protected]> * sensevoice * sensevoice * sensevoice * sensevoice * sensevoice * sensevoice * sensevoice * sensevoice * sensevoice * sensevoice * sensevoice * sensevoice * sensevoice * sensevoice * sensevoice * sense voice * sense voice * sense voice * sense voice * sense voice * sense voice * sense voice * sense voice * sense voice * sense voice * sense voice * sense voice * sense voice * sense voice * sense voice * sense voice * sense voice * sense voice * sense voice * sense voice * whisper * whisper * update style * update style --------- Co-authored-by: 维石 <[email protected]>
2026-10-12 02:19:18 +00:00 · 2024-04-24 16:03:38 +08:00
parent 7c3ba91f67
commit 861147c730
1015 changed files with 25663 additions and 30644 deletions
@@ -2,66 +2,94 @@ import os
 import numpy as np
 import sys

-def compute_wer(ref_file,
-                hyp_file,
-                cer_detail_file):
+
+def compute_wer(ref_file, hyp_file, cer_detail_file):
    rst = {
-        'Wrd': 0,
-        'Corr': 0,
-        'Ins': 0,
-        'Del': 0,
-        'Sub': 0,
-        'Snt': 0,
-        'Err': 0.0,
-        'S.Err': 0.0,
-        'wrong_words': 0,
-        'wrong_sentences': 0
+        "Wrd": 0,
+        "Corr": 0,
+        "Ins": 0,
+        "Del": 0,
+        "Sub": 0,
+        "Snt": 0,
+        "Err": 0.0,
+        "S.Err": 0.0,
+        "wrong_words": 0,
+        "wrong_sentences": 0,
    }

    hyp_dict = {}
    ref_dict = {}
-    with open(hyp_file, 'r') as hyp_reader:
+    with open(hyp_file, "r") as hyp_reader:
        for line in hyp_reader:
            key = line.strip().split()[0]
            value = line.strip().split()[1:]
            hyp_dict[key] = value
-    with open(ref_file, 'r') as ref_reader:
+    with open(ref_file, "r") as ref_reader:
        for line in ref_reader:
            key = line.strip().split()[0]
            value = line.strip().split()[1:]
            ref_dict[key] = value

-    cer_detail_writer = open(cer_detail_file, 'w')
+    cer_detail_writer = open(cer_detail_file, "w")
    for hyp_key in hyp_dict:
        if hyp_key in ref_dict:
-           out_item = compute_wer_by_line(hyp_dict[hyp_key], ref_dict[hyp_key])
-           rst['Wrd'] += out_item['nwords']
-           rst['Corr'] += out_item['cor']
-           rst['wrong_words'] += out_item['wrong']
-           rst['Ins'] += out_item['ins']
-           rst['Del'] += out_item['del']
-           rst['Sub'] += out_item['sub']
-           rst['Snt'] += 1
-           if out_item['wrong'] > 0:
-               rst['wrong_sentences'] += 1
-           cer_detail_writer.write(hyp_key + print_cer_detail(out_item) + '\n')
-           cer_detail_writer.write("ref:" + '\t' + "".join(ref_dict[hyp_key]) + '\n')
-           cer_detail_writer.write("hyp:" + '\t' + "".join(hyp_dict[hyp_key]) + '\n')
+            out_item = compute_wer_by_line(hyp_dict[hyp_key], ref_dict[hyp_key])
+            rst["Wrd"] += out_item["nwords"]
+            rst["Corr"] += out_item["cor"]
+            rst["wrong_words"] += out_item["wrong"]
+            rst["Ins"] += out_item["ins"]
+            rst["Del"] += out_item["del"]
+            rst["Sub"] += out_item["sub"]
+            rst["Snt"] += 1
+            if out_item["wrong"] > 0:
+                rst["wrong_sentences"] += 1
+            cer_detail_writer.write(hyp_key + print_cer_detail(out_item) + "\n")
+            cer_detail_writer.write("ref:" + "\t" + "".join(ref_dict[hyp_key]) + "\n")
+            cer_detail_writer.write("hyp:" + "\t" + "".join(hyp_dict[hyp_key]) + "\n")

-    if rst['Wrd'] > 0:
-        rst['Err'] = round(rst['wrong_words'] * 100 / rst['Wrd'], 2)
-    if rst['Snt'] > 0:
-        rst['S.Err'] = round(rst['wrong_sentences'] * 100 / rst['Snt'], 2)
+    if rst["Wrd"] > 0:
+        rst["Err"] = round(rst["wrong_words"] * 100 / rst["Wrd"], 2)
+    if rst["Snt"] > 0:
+        rst["S.Err"] = round(rst["wrong_sentences"] * 100 / rst["Snt"], 2)

-    cer_detail_writer.write('\n')
-    cer_detail_writer.write("%WER " + str(rst['Err']) + " [ " + str(rst['wrong_words'])+ " / " + str(rst['Wrd']) +
-                            ", " + str(rst['Ins']) + " ins, " + str(rst['Del']) + " del, " + str(rst['Sub']) + " sub ]" + '\n')
-    cer_detail_writer.write("%SER " + str(rst['S.Err']) + " [ " + str(rst['wrong_sentences']) + " / " + str(rst['Snt']) + " ]" + '\n')
-    cer_detail_writer.write("Scored " + str(len(hyp_dict)) + " sentences, " + str(len(hyp_dict) - rst['Snt']) + " not present in hyp." + '\n')
+    cer_detail_writer.write("\n")
+    cer_detail_writer.write(
+        "%WER "
+        + str(rst["Err"])
+        + " [ "
+        + str(rst["wrong_words"])
+        + " / "
+        + str(rst["Wrd"])
+        + ", "
+        + str(rst["Ins"])
+        + " ins, "
+        + str(rst["Del"])
+        + " del, "
+        + str(rst["Sub"])
+        + " sub ]"
+        + "\n"
+    )
+    cer_detail_writer.write(
+        "%SER "
+        + str(rst["S.Err"])
+        + " [ "
+        + str(rst["wrong_sentences"])
+        + " / "
+        + str(rst["Snt"])
+        + " ]"
+        + "\n"
+    )
+    cer_detail_writer.write(
+        "Scored "
+        + str(len(hyp_dict))
+        + " sentences, "
+        + str(len(hyp_dict) - rst["Snt"])
+        + " not present in hyp."
+        + "\n"
+    )

-     
-def compute_wer_by_line(hyp,
-                        ref):
+
+def compute_wer_by_line(hyp, ref):
    hyp = list(map(lambda x: x.lower(), hyp))
    ref = list(map(lambda x: x.lower(), ref))

@@ -96,14 +124,7 @@ def compute_wer_by_line(hyp,
    match_idx = []
    i = len_hyp
    j = len_ref
-    rst = {
-        'nwords': len_ref,
-        'cor': 0,
-        'wrong': 0,
-        'ins': 0,
-        'del': 0,
-        'sub': 0
-    }
+    rst = {"nwords": len_ref, "cor": 0, "wrong": 0, "ins": 0, "del": 0, "sub": 0}
    while i >= 0 or j >= 0:
        i_idx = max(0, i)
        j_idx = max(0, j)
@@ -111,42 +132,57 @@ def compute_wer_by_line(hyp,
        if ops_matrix[i_idx][j_idx] == 0:  # correct
            if i - 1 >= 0 and j - 1 >= 0:
                match_idx.append((j - 1, i - 1))
-                rst['cor'] += 1
+                rst["cor"] += 1

            i -= 1
            j -= 1

        elif ops_matrix[i_idx][j_idx] == 2:  # insert
            i -= 1
-            rst['ins'] += 1
+            rst["ins"] += 1

        elif ops_matrix[i_idx][j_idx] == 3:  # delete
            j -= 1
-            rst['del'] += 1
+            rst["del"] += 1

        elif ops_matrix[i_idx][j_idx] == 1:  # substitute
            i -= 1
            j -= 1
-            rst['sub'] += 1
+            rst["sub"] += 1

        if i < 0 and j >= 0:
-            rst['del'] += 1
+            rst["del"] += 1
        elif j < 0 and i >= 0:
-            rst['ins'] += 1
+            rst["ins"] += 1

    match_idx.reverse()
    wrong_cnt = cost_matrix[len_hyp][len_ref]
-    rst['wrong'] = wrong_cnt
+    rst["wrong"] = wrong_cnt

    return rst

-def print_cer_detail(rst):
-    return ("(" + "nwords=" + str(rst['nwords']) + ",cor=" + str(rst['cor'])
-            + ",ins=" + str(rst['ins']) + ",del=" + str(rst['del']) + ",sub="
-            + str(rst['sub']) + ") corr:" + '{:.2%}'.format(rst['cor']/rst['nwords'])
-            + ",cer:" + '{:.2%}'.format(rst['wrong']/rst['nwords']))

-if __name__ == '__main__':
+def print_cer_detail(rst):
+    return (
+        "("
+        + "nwords="
+        + str(rst["nwords"])
+        + ",cor="
+        + str(rst["cor"])
+        + ",ins="
+        + str(rst["ins"])
+        + ",del="
+        + str(rst["del"])
+        + ",sub="
+        + str(rst["sub"])
+        + ") corr:"
+        + "{:.2%}".format(rst["cor"] / rst["nwords"])
+        + ",cer:"
+        + "{:.2%}".format(rst["wrong"] / rst["nwords"])
+    )
+
+
+if __name__ == "__main__":
    if len(sys.argv) != 4:
        print("usage : python compute-wer.py test.ref test.hyp test.wer")
        sys.exit(0)
@@ -1,4 +1,3 @@
-
 import sys
 import re

@@ -7,25 +6,25 @@ out_f = sys.argv[2]


 with open(in_f, "r", encoding="utf-8") as f:
-  lines = f.readlines()
+    lines = f.readlines()

 with open(out_f, "w", encoding="utf-8") as f:
-  for line in lines:
-    outs = line.strip().split(" ", 1)
-    if len(outs) == 2:
-      idx, text = outs
-      text = re.sub("</s>", "", text)
-      text = re.sub("<s>", "", text)
-      text = re.sub("@@", "", text)
-      text = re.sub("@", "", text)
-      text = re.sub("<unk>", "", text)
-      text = re.sub(" ", "", text)
-      text = text.lower()
-    else:
-      idx = outs[0]
-      text = " "
+    for line in lines:
+        outs = line.strip().split(" ", 1)
+        if len(outs) == 2:
+            idx, text = outs
+            text = re.sub("</s>", "", text)
+            text = re.sub("<s>", "", text)
+            text = re.sub("@@", "", text)
+            text = re.sub("@", "", text)
+            text = re.sub("<unk>", "", text)
+            text = re.sub(" ", "", text)
+            text = text.lower()
+        else:
+            idx = outs[0]
+            text = " "

-    text = [x for x in text]
-    text = " ".join(text)
-    out = "{} {}\n".format(idx, text)
-    f.write(out)
+        text = [x for x in text]
+        text = " ".join(text)
+        out = "{} {}\n".format(idx, text)
+        f.write(out)
@@ -5,44 +5,52 @@ import librosa
 from funasr.utils.types import str2bool

 import argparse
+
 parser = argparse.ArgumentParser()
-parser.add_argument('--model_dir', type=str, required=True)
-parser.add_argument('--backend', type=str, default='onnx', help='["onnx", "torch"]')
-parser.add_argument('--wav_file', type=str, default=None, help='amp fallback number')
-parser.add_argument('--quantize', type=str2bool, default=False, help='quantized model')
-parser.add_argument('--intra_op_num_threads', type=int, default=1, help='intra_op_num_threads for onnx')
-parser.add_argument('--output_dir', type=str, default=None, help='amp fallback number')
+parser.add_argument("--model_dir", type=str, required=True)
+parser.add_argument("--backend", type=str, default="onnx", help='["onnx", "torch"]')
+parser.add_argument("--wav_file", type=str, default=None, help="amp fallback number")
+parser.add_argument("--quantize", type=str2bool, default=False, help="quantized model")
+parser.add_argument(
+    "--intra_op_num_threads", type=int, default=1, help="intra_op_num_threads for onnx"
+)
+parser.add_argument("--output_dir", type=str, default=None, help="amp fallback number")
 args = parser.parse_args()


 from funasr.runtime.python.libtorch.funasr_torch import Paraformer
+
 if args.backend == "onnx":
    from funasr.runtime.python.onnxruntime.funasr_onnx import Paraformer
-    
-model = Paraformer(args.model_dir, batch_size=1, quantize=args.quantize, intra_op_num_threads=args.intra_op_num_threads)

-wav_file_f = open(args.wav_file, 'r')
+model = Paraformer(
+    args.model_dir,
+    batch_size=1,
+    quantize=args.quantize,
+    intra_op_num_threads=args.intra_op_num_threads,
+)
+
+wav_file_f = open(args.wav_file, "r")
 wav_files = wav_file_f.readlines()

 output_dir = args.output_dir
 if not os.path.exists(output_dir):
    os.makedirs(output_dir)
-if os.name == 'nt':   # Windows
-    newline = '\r\n'
-else:   # Linux Mac
-    newline = '\n'
+if os.name == "nt":  # Windows
+    newline = "\r\n"
+else:  # Linux Mac
+    newline = "\n"
 text_f = open(os.path.join(output_dir, "text"), "w", newline=newline)
 token_f = open(os.path.join(output_dir, "token"), "w", newline=newline)

 for i, wav_path_i in enumerate(wav_files):
    wav_name, wav_path = wav_path_i.strip().split()
    result = model(wav_path)
-    text_i = "{} {}\n".format(wav_name, result[0]['preds'][0])
-    token_i = "{} {}\n".format(wav_name, result[0]['preds'][1])
+    text_i = "{} {}\n".format(wav_name, result[0]["preds"][0])
+    token_i = "{} {}\n".format(wav_name, result[0]["preds"][1])
    text_f.write(text_i)
    text_f.flush()
    token_f.write(token_i)
    token_f.flush()
 text_f.close()
 token_f.close()
-    
@@ -1,55 +1,79 @@
-
 import time
 import sys
 import librosa
 from funasr.utils.types import str2bool

 import argparse
+
 parser = argparse.ArgumentParser()
-parser.add_argument('--model_dir', type=str, required=True)
-parser.add_argument('--backend', type=str, default='onnx', help='["onnx", "torch"]')
-parser.add_argument('--wav_file', type=str, default=None, help='amp fallback number')
-parser.add_argument('--quantize', type=str2bool, default=False, help='quantized model')
-parser.add_argument('--intra_op_num_threads', type=int, default=1, help='intra_op_num_threads for onnx')
+parser.add_argument("--model_dir", type=str, required=True)
+parser.add_argument("--backend", type=str, default="onnx", help='["onnx", "torch"]')
+parser.add_argument("--wav_file", type=str, default=None, help="amp fallback number")
+parser.add_argument("--quantize", type=str2bool, default=False, help="quantized model")
+parser.add_argument(
+    "--intra_op_num_threads", type=int, default=1, help="intra_op_num_threads for onnx"
+)
 args = parser.parse_args()


 from funasr.runtime.python.libtorch.funasr_torch import Paraformer
+
 if args.backend == "onnx":
    from funasr.runtime.python.onnxruntime.funasr_onnx import Paraformer
-    
-model = Paraformer(args.model_dir, batch_size=1, quantize=args.quantize, intra_op_num_threads=args.intra_op_num_threads)

-wav_file_f = open(args.wav_file, 'r')
+model = Paraformer(
+    args.model_dir,
+    batch_size=1,
+    quantize=args.quantize,
+    intra_op_num_threads=args.intra_op_num_threads,
+)
+
+wav_file_f = open(args.wav_file, "r")
 wav_files = wav_file_f.readlines()

 # warm-up
 total = 0.0
 num = 30
-wav_path = wav_files[0].split("\t")[1].strip() if "\t" in wav_files[0] else wav_files[0].split(" ")[1].strip()
+wav_path = (
+    wav_files[0].split("\t")[1].strip()
+    if "\t" in wav_files[0]
+    else wav_files[0].split(" ")[1].strip()
+)
 for i in range(num):
    beg_time = time.time()
    result = model(wav_path)
    end_time = time.time()
-    duration = end_time-beg_time
+    duration = end_time - beg_time
    total += duration
    print(result)
-    print("num: {}, time, {}, avg: {}, rtf: {}".format(len(wav_path), duration, total/(i+1), (total/(i+1))/5.53))
+    print(
+        "num: {}, time, {}, avg: {}, rtf: {}".format(
+            len(wav_path), duration, total / (i + 1), (total / (i + 1)) / 5.53
+        )
+    )

 # infer time
 beg_time = time.time()
 for i, wav_path_i in enumerate(wav_files):
-    wav_path = wav_path_i.split("\t")[1].strip() if "\t" in wav_path_i else wav_path_i.split(" ")[1].strip()
+    wav_path = (
+        wav_path_i.split("\t")[1].strip()
+        if "\t" in wav_path_i
+        else wav_path_i.split(" ")[1].strip()
+    )
    result = model(wav_path)
 end_time = time.time()
-duration = (end_time-beg_time)*1000
+duration = (end_time - beg_time) * 1000
 print("total_time_comput_ms: {}".format(int(duration)))

 duration_time = 0.0
 for i, wav_path_i in enumerate(wav_files):
-    wav_path = wav_path_i.split("\t")[1].strip() if "\t" in wav_path_i else wav_path_i.split(" ")[1].strip()
+    wav_path = (
+        wav_path_i.split("\t")[1].strip()
+        if "\t" in wav_path_i
+        else wav_path_i.split(" ")[1].strip()
+    )
    waveform, _ = librosa.load(wav_path, sr=16000)
-    duration_time += len(waveform)/16.0
+    duration_time += len(waveform) / 16.0
 print("total_time_wav_ms: {}".format(int(duration_time)))

-print("total_rtf: {:.5}".format(duration/duration_time))
+print("total_rtf: {:.5}".format(duration / duration_time))
@@ -1,58 +1,82 @@
-
 import time
 import sys
 import librosa
 from funasr.utils.types import str2bool

 import argparse
+
 parser = argparse.ArgumentParser()
-parser.add_argument('--model_dir', type=str, required=True)
-parser.add_argument('--backend', type=str, default='onnx', help='["onnx", "torch"]')
-parser.add_argument('--wav_file', type=str, default=None, help='amp fallback number')
-parser.add_argument('--quantize', type=str2bool, default=False, help='quantized model')
-parser.add_argument('--intra_op_num_threads', type=int, default=1, help='intra_op_num_threads for onnx')
-parser.add_argument('--batch_size', type=int, default=1, help='batch_size for onnx')
+parser.add_argument("--model_dir", type=str, required=True)
+parser.add_argument("--backend", type=str, default="onnx", help='["onnx", "torch"]')
+parser.add_argument("--wav_file", type=str, default=None, help="amp fallback number")
+parser.add_argument("--quantize", type=str2bool, default=False, help="quantized model")
+parser.add_argument(
+    "--intra_op_num_threads", type=int, default=1, help="intra_op_num_threads for onnx"
+)
+parser.add_argument("--batch_size", type=int, default=1, help="batch_size for onnx")
 args = parser.parse_args()


 from funasr.runtime.python.libtorch.funasr_torch import Paraformer
+
 if args.backend == "onnx":
    from funasr.runtime.python.onnxruntime.funasr_onnx import Paraformer
-    
-model = Paraformer(args.model_dir, batch_size=args.batch_size, quantize=args.quantize, intra_op_num_threads=args.intra_op_num_threads)

-wav_file_f = open(args.wav_file, 'r')
+model = Paraformer(
+    args.model_dir,
+    batch_size=args.batch_size,
+    quantize=args.quantize,
+    intra_op_num_threads=args.intra_op_num_threads,
+)
+
+wav_file_f = open(args.wav_file, "r")
 wav_files = wav_file_f.readlines()

 # warm-up
 total = 0.0
 num = 30
-wav_path = wav_files[0].split("\t")[1].strip() if "\t" in wav_files[0] else wav_files[0].split(" ")[1].strip()
+wav_path = (
+    wav_files[0].split("\t")[1].strip()
+    if "\t" in wav_files[0]
+    else wav_files[0].split(" ")[1].strip()
+)
 for i in range(num):
    beg_time = time.time()
    result = model(wav_path)
    end_time = time.time()
-    duration = end_time-beg_time
+    duration = end_time - beg_time
    total += duration
    print(result)
-    print("num: {}, time, {}, avg: {}, rtf: {}".format(len(wav_path), duration, total/(i+1), (total/(i+1))/5.53))
+    print(
+        "num: {}, time, {}, avg: {}, rtf: {}".format(
+            len(wav_path), duration, total / (i + 1), (total / (i + 1)) / 5.53
+        )
+    )

 # infer time
 wav_path = []
 beg_time = time.time()
 for i, wav_path_i in enumerate(wav_files):
-    wav_path_i = wav_path_i.split("\t")[1].strip() if "\t" in wav_path_i else wav_path_i.split(" ")[1].strip()
+    wav_path_i = (
+        wav_path_i.split("\t")[1].strip()
+        if "\t" in wav_path_i
+        else wav_path_i.split(" ")[1].strip()
+    )
    wav_path += [wav_path_i]
 result = model(wav_path)
 end_time = time.time()
-duration = (end_time-beg_time)*1000
+duration = (end_time - beg_time) * 1000
 print("total_time_comput_ms: {}".format(int(duration)))

 duration_time = 0.0
 for i, wav_path_i in enumerate(wav_files):
-    wav_path = wav_path_i.split("\t")[1].strip() if "\t" in wav_path_i else wav_path_i.split(" ")[1].strip()
+    wav_path = (
+        wav_path_i.split("\t")[1].strip()
+        if "\t" in wav_path_i
+        else wav_path_i.split(" ")[1].strip()
+    )
    waveform, _ = librosa.load(wav_path, sr=16000)
-    duration_time += len(waveform)/16.0
+    duration_time += len(waveform) / 16.0
 print("total_time_wav_ms: {}".format(int(duration_time)))

-print("total_rtf: {:.5}".format(duration/duration_time))
+print("total_rtf: {:.5}".format(duration / duration_time))