From 8dab6d184a034ca86eafa644ea0d2100aadfe27d Mon Sep 17 00:00:00 2001
From: jmwang66 <wangjiaming.wjm@alibaba-inc.com>
Date: 星期二, 09 五月 2023 10:58:33 +0800
Subject: [PATCH] Merge pull request #473 from alibaba-damo-academy/dev_smohan

---
 funasr/tasks/abs_task.py |   52 +++++++++++++++++++++++++++++++++++++++++++---------
 1 files changed, 43 insertions(+), 9 deletions(-)

diff --git a/funasr/tasks/abs_task.py b/funasr/tasks/abs_task.py
index 3f20b4f..55a5d79 100644
--- a/funasr/tasks/abs_task.py
+++ b/funasr/tasks/abs_task.py
@@ -445,6 +445,12 @@
             help='Perform on "collect stats" mode',
         )
         group.add_argument(
+            "--mc",
+            type=bool,
+            default=False,
+            help="MultiChannel input",
+        )
+        group.add_argument(
             "--write_collected_feats",
             type=str2bool,
             default=False,
@@ -463,6 +469,12 @@
             type=int,
             default=sys.maxsize,
             help="The maximum number update step to train",
+        )
+        parser.add_argument(
+            "--batch_interval",
+            type=int,
+            default=-1,
+            help="The batch interval for saving model.",
         )
         group.add_argument(
             "--patience",
@@ -541,6 +553,12 @@
             type=int,
             default=1,
             help="The number of gradient accumulation",
+        )
+        group.add_argument(
+            "--bias_grad_times",
+            type=float,
+            default=1.0,
+            help="To scale the gradient of contextual related params",
         )
         group.add_argument(
             "--no_forward_run",
@@ -629,8 +647,8 @@
         group.add_argument(
             "--init_param",
             type=str,
+            action="append",
             default=[],
-            nargs="*",
             help="Specify the file path used for initialization of parameters. "
                  "The format is '<file_path>:<src_key>:<dst_key>:<exclude_keys>', "
                  "where file_path is the model file path, "
@@ -656,7 +674,7 @@
             "--freeze_param",
             type=str,
             default=[],
-            nargs="*",
+            action="append",
             help="Freeze parameters",
         )
 
@@ -1147,10 +1165,10 @@
         elif args.distributed and args.simple_ddp:
             distributed_option.init_torch_distributed_pai(args)
             args.ngpu = dist.get_world_size()
-            if args.dataset_type == "small":
+            if args.dataset_type == "small" and args.ngpu > 0:
                 if args.batch_size is not None:
                     args.batch_size = args.batch_size * args.ngpu
-                if args.batch_bins is not None:
+                if args.batch_bins is not None and args.ngpu > 0:
                     args.batch_bins = args.batch_bins * args.ngpu
 
         # filter samples if wav.scp and text are mismatch
@@ -1193,12 +1211,18 @@
             # logging.basicConfig() is invoked in main_worker() instead of main()
             # because it can be invoked only once in a process.
             # FIXME(kamo): Should we use logging.getLogger()?
+            # BUGFIX: Remove previous handlers and reset log level
+            for handler in logging.root.handlers[:]:
+                logging.root.removeHandler(handler)
             logging.basicConfig(
                 level=args.log_level,
                 format=f"[{os.uname()[1].split('.')[0]}]"
                        f" %(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s",
             )
         else:
+            # BUGFIX: Remove previous handlers and reset log level
+            for handler in logging.root.handlers[:]:
+                logging.root.removeHandler(handler)
             # Suppress logging if RANK != 0
             logging.basicConfig(
                 level="ERROR",
@@ -1304,6 +1328,7 @@
                     data_path_and_name_and_type=args.train_data_path_and_name_and_type,
                     key_file=train_key_file,
                     batch_size=args.batch_size,
+                    mc=args.mc,
                     dtype=args.train_dtype,
                     num_workers=args.num_workers,
                     allow_variable_data_keys=args.allow_variable_data_keys,
@@ -1315,6 +1340,7 @@
                     data_path_and_name_and_type=args.valid_data_path_and_name_and_type,
                     key_file=valid_key_file,
                     batch_size=args.valid_batch_size,
+                    mc=args.mc,
                     dtype=args.train_dtype,
                     num_workers=args.num_workers,
                     allow_variable_data_keys=args.allow_variable_data_keys,
@@ -1349,15 +1375,15 @@
                 from funasr.datasets.large_datasets.build_dataloader import ArkDataLoader
                 train_iter_factory = ArkDataLoader(args.train_data_file, args.token_list, args.dataset_conf,
                                                    frontend_conf=args.frontend_conf if hasattr(args, "frontend_conf") else None,
-                                                   seg_dict_file=args.seg_dict_file if hasattr(args,
-                                                                                               "seg_dict_file") else None,
+                                                   seg_dict_file=args.seg_dict_file if hasattr(args, "seg_dict_file") else None,
                                                    punc_dict_file=args.punc_list if hasattr(args, "punc_list") else None,
+                                                   bpemodel_file=args.bpemodel if hasattr(args, "bpemodel") else None,
                                                    mode="train")
                 valid_iter_factory = ArkDataLoader(args.valid_data_file, args.token_list, args.dataset_conf, 
                                                    frontend_conf=args.frontend_conf if hasattr(args, "frontend_conf") else None,
-                                                   seg_dict_file=args.seg_dict_file if hasattr(args,
-                                                                                               "seg_dict_file") else None,
+                                                   seg_dict_file=args.seg_dict_file if hasattr(args, "seg_dict_file") else None,
                                                    punc_dict_file=args.punc_list if hasattr(args, "punc_list") else None,
+                                                   bpemodel_file=args.bpemodel if hasattr(args, "bpemodel") else None,
                                                    mode="eval")
             elif args.dataset_type == "small":
                 train_iter_factory = cls.build_iter_factory(
@@ -1570,13 +1596,21 @@
     ) -> AbsIterFactory:
         assert check_argument_types()
 
+        if hasattr(args, "frontend_conf"):
+            if args.frontend_conf is not None and "fs" in args.frontend_conf:
+                dest_sample_rate = args.frontend_conf["fs"]
+            else:
+                dest_sample_rate = 16000
+        else:
+            dest_sample_rate = 16000
+
         dataset = ESPnetDataset(
             iter_options.data_path_and_name_and_type,
             float_dtype=args.train_dtype,
             preprocess=iter_options.preprocess_fn,
             max_cache_size=iter_options.max_cache_size,
             max_cache_fd=iter_options.max_cache_fd,
-            dest_sample_rate=args.frontend_conf["fs"],
+            dest_sample_rate=dest_sample_rate,
         )
         cls.check_task_requirements(
             dataset, args.allow_variable_data_keys, train=iter_options.train

--
Gitblit v1.9.1