Add grid AVSR task results

egs/grid/AVSR/audionet_ctc_asr/decode.py

@@ -30,8 +30,8 @@ import torch
 import torch.nn as nn
 
 from torch.utils.data import DataLoader
-from local.dataset_audio import MyDataset
+from local.dataset_audio import dataset_audio
-from model import TdnnLstm
+from model import AudioNet
 
 from icefall.checkpoint import average_checkpoints, load_checkpoint
 from icefall.decode import (
@@ -143,8 +143,9 @@ def get_params() -> AttributeDict:
             "video_path": Path("download/GRID/lip/"),
             "anno_path": Path("download/GRID/GRID_align_txt"),
             "val_list": Path("download/GRID/unseen_val.txt"),
-            "aud_padding": 200,
+            "aud_padding": 480,
-            "num_workers": 1,
+            "sample_rate": 16000,
+            "num_workers": 16,
             "batch_size": 120,
         }
     )
@@ -440,7 +441,7 @@ def main():
     else:
         G = None
 
-    model = TdnnLstm(
+    model = AudioNet(
         num_features=params.feature_dim,
         num_classes=max_token_id + 1,  # +1 for the blank symbol
         subsampling_factor=params.subsampling_factor,
@@ -466,14 +467,14 @@ def main():
     model.to(device)
     model.eval()
 
-    grid = MyDataset(
+    grid = dataset_audio(
         params.video_path,
         params.anno_path,
         params.val_list,
         params.aud_padding,
-        "test",
+        params.sample_rate,
-        16000,
         params.feature_dim,
+        "test",
     )
     test_dl = DataLoader(
         grid,

egs/grid/AVSR/audionet_ctc_asr/model.py

@@ -19,7 +19,7 @@ import torch
 import torch.nn as nn
 
 
-class TdnnLstm(nn.Module):
+class AudioNet(nn.Module):
     def __init__(
         self, num_features: int, num_classes: int, subsampling_factor: int = 3
     ) -> None:

egs/grid/AVSR/audionet_ctc_asr/train.py

@@ -32,9 +32,9 @@ import torch.nn as nn
 import torch.optim as optim
 from torch.utils.data import DataLoader
 
-from local.dataset_audio import MyDataset
+from local.dataset_audio import dataset_audio
 from lhotse.utils import fix_random_seed
-from model import TdnnLstm
+from model import AudioNet
 from torch import Tensor
 from torch.nn.parallel import DistributedDataParallel as DDP
 from torch.nn.utils import clip_grad_norm_
@@ -179,8 +179,9 @@ def get_params() -> AttributeDict:
             "video_path": Path("download/GRID/lip/"),
             "anno_path": Path("download/GRID/GRID_align_txt"),
             "train_list": Path("download/GRID/unseen_train.txt"),
-            "aud_padding": 200,
+            "aud_padding": 480,
-            "num_workers": 1,
+            "sample_rate": 16000,
+            "num_workers": 16,
             "batch_size": 120,
         }
     )
@@ -510,7 +511,7 @@ def run(rank, world_size, args):
 
     graph_compiler = CtcTrainingGraphCompiler(lexicon=lexicon, device=device)
 
-    model = TdnnLstm(
+    model = AudioNet(
         num_features=params.feature_dim,
         num_classes=max_token_id + 1,  # +1 for the blank symbol
         subsampling_factor=params.subsampling_factor,
@@ -532,14 +533,14 @@ def run(rank, world_size, args):
         optimizer.load_state_dict(checkpoints["optimizer"])
         scheduler.load_state_dict(checkpoints["scheduler"])
 
-    grid = MyDataset(
+    grid = dataset_audio(
         params.video_path,
         params.anno_path,
         params.train_list,
-        params.vid_padding,
+        params.aud_padding,
-        "train",
+        params.sample_rate,
-        16000,
         params.feature_dim,
+        "train",
     )
 
     train_dl = DataLoader(

egs/grid/AVSR/combinenet_ctc_avsr/decode.py

@@ -0,0 +1,514 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang
+#                                                    Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+from utils import encode_supervisions
+
+import k2
+import torch
+import torch.nn as nn
+
+from torch.utils.data import DataLoader
+from local.dataset_av import dataset_av
+from model import CombineNet
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.decode import (
+    get_lattice,
+    nbest_decoding,
+    one_best_decoding,
+    rescore_with_n_best_list,
+    rescore_with_whole_lattice,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    get_texts,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=19,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=5,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="whole-lattice-rescoring",
+        help="""Decoding method.
+        Supported values are:
+            - (1) 1best. Extract the best path from the decoding lattice as the
+              decoding result.
+            - (2) nbest. Extract n paths from the decoding lattice; the path
+              with the highest score is the decoding result.
+            - (3) nbest-rescoring. Extract n paths from the decoding lattice,
+              rescore them with an n-gram LM (e.g., a 4-gram LM), the path with
+              the highest score is the decoding result.
+            - (4) whole-lattice-rescoring. Rescore the decoding lattice with an
+              n-gram LM (e.g., a 4-gram LM), the best path of rescored lattice
+              is the decoding result.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""Number of paths for n-best based decoding method.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""The scale to be applied to `lattice.scores`.
+        It's needed if you use any kinds of n-best based rescoring.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring
+        A smaller value results in more unique paths.
+        """,
+    )
+
+    parser.add_argument(
+        "--export",
+        type=str2bool,
+        default=False,
+        help="""When enabled, the averaged model is saved to
+        tdnn/exp/pretrained.pt. Note: only model.state_dict() is saved.
+        pretrained.pt contains a dict {"model": model.state_dict()},
+        which can be loaded by `icefall.checkpoint.load_checkpoint()`.
+        """,
+    )
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            "exp_dir": Path("combinenet_ctc_avsr/exp"),
+            "lang_dir": Path("data/lang_character"),
+            "lm_dir": Path("data/lm"),
+            "feature_dim": 80,
+            "subsampling_factor": 3,
+            "search_beam": 20,
+            "output_beam": 5,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+            # parameters for dataset
+            "video_path": Path("download/GRID/lip/"),
+            "anno_path": Path("download/GRID/GRID_align_txt"),
+            "val_list": Path("download/GRID/unseen_val.txt"),
+            "vid_padding": 75,
+            "aud_padding": 450,
+            "sample_rate": 16000,
+            "num_workers": 16,
+            "batch_size": 120,
+        }
+    )
+    return params
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: k2.Fsa,
+    batch: dict,
+    lexicon: Lexicon,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if no rescoring is used, the key is the string `no_rescore`.
+               If LM rescoring is used, the key is the string `lm_scale_xxx`,
+               where `xxx` is the value of `lm_scale`. An example key is
+               `lm_scale_0.7`
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+
+        - params.method is "1best", it uses 1best decoding without LM rescoring.
+        - params.method is "nbest", it uses nbest decoding without LM rescoring.
+        - params.method is "nbest-rescoring", it uses nbest LM rescoring.
+        - params.method is "whole-lattice-rescoring", it uses whole lattice LM
+          rescoring.
+
+      model:
+        The neural model.
+      HLG:
+        The decoding graph.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      lexicon:
+        It contains word symbol table.
+      G:
+        An LM. It is not None when params.method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = HLG.device
+    audio_feature = batch["aud"]
+    video_feature = batch["vid"]
+
+    audio_feature = audio_feature.permute(0, 2, 1)
+    assert audio_feature.ndim == 3
+    audio_feature = audio_feature.to(device)
+
+    assert video_feature.ndim == 5
+    video_feature = video_feature.to(device)
+
+    nnet_output = model(video_feature, audio_feature)
+
+    nnet_output_shape = nnet_output.size()
+    supervision_segments, text = encode_supervisions(nnet_output_shape, batch)
+
+    lattice = get_lattice(
+        nnet_output=nnet_output,
+        decoding_graph=HLG,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+    )
+
+    if params.method in ["1best", "nbest"]:
+        if params.method == "1best":
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+            key = "no_rescore"
+        else:
+            best_path = nbest_decoding(
+                lattice=lattice,
+                num_paths=params.num_paths,
+                use_double_scores=params.use_double_scores,
+                nbest_scale=params.nbest_scale,
+            )
+            key = f"no_rescore-{params.num_paths}"
+        hyps = get_texts(best_path)
+        hyps = [[lexicon.word_table[i] for i in ids] for ids in hyps]
+        return {key: hyps}
+
+    assert params.method in ["nbest-rescoring", "whole-lattice-rescoring"]
+
+    lm_scale_list = [0.0, 0.01, 0.02, 0.03, 0.04, 0.05, 0.06, 0.07, 0.08, 0.09]
+    lm_scale_list += [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7]
+    lm_scale_list += [0.8, 0.9, 1.0, 1.1, 1.2, 1.3]
+    lm_scale_list += [1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 2.0]
+
+    if params.method == "nbest-rescoring":
+        best_path_dict = rescore_with_n_best_list(
+            lattice=lattice,
+            G=G,
+            num_paths=params.num_paths,
+            lm_scale_list=lm_scale_list,
+            nbest_scale=params.nbest_scale,
+        )
+    else:
+        best_path_dict = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=lm_scale_list,
+        )
+
+    ans = dict()
+    for lm_scale_str, best_path in best_path_dict.items():
+        hyps = get_texts(best_path)
+        hyps = [[lexicon.word_table[i] for i in ids] for ids in hyps]
+        ans[lm_scale_str] = hyps
+    return ans
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: k2.Fsa,
+    lexicon: Lexicon,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      HLG:
+        The decoding graph.
+      lexicon:
+        It contains word symbol table.
+      G:
+        An LM. It is not None when params.method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return a dict, whose key may be "no-rescore" if no LM rescoring
+      is used, or it may be "lm_scale_0.7" if LM rescoring is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    results = []
+
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["txt"]
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            HLG=HLG,
+            batch=batch,
+            lexicon=lexicon,
+            G=G,
+        )
+
+        for lm_scale, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for hyp_words, ref_text in zip(hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((ref_words, hyp_words))
+
+            results[lm_scale].extend(this_batch)
+
+        num_cuts += len(batch["txt"])
+
+        if batch_idx % 10 == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(
+                f"batch {batch_str}, cuts processed until now is {num_cuts}"
+            )
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[List[int], List[int]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.exp_dir / f"recogs-{test_set_name}-{key}.txt"
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out PERs, per-phone error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.exp_dir / f"errs-{test_set_name}-{key}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(f, f"{test_set_name}-{key}", results)
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.exp_dir / f"per-summary-{test_set_name}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tPER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, PER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+    params.update(vars(args))
+
+    setup_logger(f"{params.exp_dir}/log/log-decode")
+    logging.info("Decoding started")
+    logging.info(params)
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    HLG = k2.Fsa.from_dict(
+        torch.load(f"{params.lang_dir}/HLG.pt", map_location="cpu")
+    )
+
+    HLG = HLG.to(device)
+    assert HLG.requires_grad is False
+
+    if not hasattr(HLG, "lm_scores"):
+        HLG.lm_scores = HLG.scores.clone()
+
+    if params.method in ["nbest-rescoring", "whole-lattice-rescoring"]:
+        if not (params.lm_dir / "G_4_gram.pt").is_file():
+            logging.info("Loading G_4_gram.fst.txt")
+            logging.warning("It may take 8 minutes.")
+            with open(params.lm_dir / "G_4_gram.fst.txt") as f:
+                first_word_disambig_id = lexicon.word_table["#0"]
+
+                G = k2.Fsa.from_openfst(f.read(), acceptor=False)
+                # G.aux_labels is not needed in later computations, so
+                # remove it here.
+                del G.aux_labels
+                # CAUTION: The following line is crucial.
+                # Arcs entering the back-off state have label equal to #0.
+                # We have to change it to 0 here.
+                G.labels[G.labels >= first_word_disambig_id] = 0
+                G = k2.Fsa.from_fsas([G]).to(device)
+                G = k2.arc_sort(G)
+                torch.save(G.as_dict(), params.lm_dir / "G_4_gram.pt")
+        else:
+            logging.info("Loading pre-compiled G_4_gram.pt")
+            d = torch.load(params.lm_dir / "G_4_gram.pt", map_location="cpu")
+            G = k2.Fsa.from_dict(d).to(device)
+
+        if params.method == "whole-lattice-rescoring":
+            # Add epsilon self-loops to G as we will compose
+            # it with the whole lattice later
+            G = k2.add_epsilon_self_loops(G)
+            G = k2.arc_sort(G)
+            G = G.to(device)
+
+        # G.lm_scores is used to replace HLG.lm_scores during
+        # LM rescoring.
+        G.lm_scores = G.scores.clone()
+    else:
+        G = None
+
+    model = CombineNet(
+        num_features=params.feature_dim,
+        num_classes=max_token_id + 1,  # +1 for the blank symbol
+        subsampling_factor=params.subsampling_factor,
+    )
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.load_state_dict(average_checkpoints(filenames))
+
+    if params.export:
+        logging.info(f"Export averaged model to {params.exp_dir}/pretrained.pt")
+        torch.save(
+            {"model": model.state_dict()}, f"{params.exp_dir}/pretrained.pt"
+        )
+        return
+
+    model.to(device)
+    model.eval()
+
+    grid = dataset_av(
+        params.video_path,
+        params.anno_path,
+        params.val_list,
+        params.feature_dim,
+        params.vid_padding,
+        params.aud_padding,
+        params.sample_rate,
+        "test",
+    )
+    test_dl = DataLoader(
+        grid,
+        batch_size=params.batch_size,
+        shuffle=False,
+        num_workers=params.num_workers,
+        drop_last=False,
+    )
+    test_set = "test"
+
+    results_dict = decode_dataset(
+        dl=test_dl,
+        params=params,
+        model=model,
+        HLG=HLG,
+        lexicon=lexicon,
+        G=G,
+    )
+
+    save_results(
+        params=params, test_set_name=test_set, results_dict=results_dict
+    )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()

egs/grid/AVSR/combinenet_ctc_avsr/model.py

@@ -0,0 +1,176 @@
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import torch
+import torch.nn as nn
+
+
+class CombineNet(nn.Module):
+    def __init__(
+        self, num_features: int, num_classes: int, subsampling_factor: int = 3
+    ) -> None:
+        """
+        Args:
+          num_features:
+            The input dimension of the audio encoder.
+          num_classes:
+            The output dimension of the combinenet model.
+          subsampling_factor:
+            It reduces the number of output frames by this factor.
+        """
+        super().__init__()
+        self.num_features = num_features
+        self.num_classes = num_classes
+        self.subsampling_factor = subsampling_factor
+
+        # the audio encoder
+        self.audio_encoder = nn.Sequential(
+            nn.Conv1d(
+                in_channels=num_features,
+                out_channels=512,
+                kernel_size=3,
+                stride=1,
+                padding=1,
+            ),
+            nn.ReLU(inplace=True),
+            nn.BatchNorm1d(num_features=512, affine=False),
+            nn.Conv1d(
+                in_channels=512,
+                out_channels=512,
+                kernel_size=3,
+                stride=1,
+                padding=1,
+            ),
+            nn.ReLU(inplace=True),
+            nn.BatchNorm1d(num_features=512, affine=False),
+            nn.Conv1d(
+                in_channels=512,
+                out_channels=512,
+                kernel_size=3,
+                # stride=self.subsampling_factor,  # stride: subsampling_factor!
+                stride=1,
+                padding=1,
+            ),
+            nn.ReLU(inplace=True),
+            nn.BatchNorm1d(num_features=512, affine=False),
+            nn.Conv1d(
+                in_channels=512,
+                out_channels=512,
+                kernel_size=3,
+                stride=self.subsampling_factor,  # stride: subsampling_factor!
+            ),
+            nn.ReLU(inplace=True),
+            nn.BatchNorm1d(num_features=512, affine=False),
+        )
+
+        # the video encoder
+        self.video_encoder = nn.Sequential(
+            nn.Conv3d(
+                in_channels=3,
+                out_channels=32,
+                kernel_size=(3, 5, 5),
+                stride=(1, 2, 2),
+                padding=(1, 2, 2),
+            ),
+            nn.ReLU(inplace=True),
+            nn.Dropout3d(p=0.1),
+            nn.MaxPool3d(kernel_size=(1, 2, 2), stride=(1, 2, 2)),
+            nn.Conv3d(
+                in_channels=32,
+                out_channels=64,
+                kernel_size=(3, 5, 5),
+                stride=(1, 1, 1),
+                padding=(1, 2, 2),
+            ),
+            nn.ReLU(inplace=True),
+            nn.Dropout3d(p=0.1),
+            nn.MaxPool3d(kernel_size=(1, 2, 2), stride=(1, 2, 2)),
+            nn.Conv3d(
+                in_channels=64,
+                out_channels=96,
+                kernel_size=(3, 3, 3),
+                stride=(1, 1, 1),
+                padding=(1, 1, 1),
+            ),
+            nn.ReLU(inplace=True),
+            nn.Dropout3d(p=0.1),
+            nn.MaxPool3d(kernel_size=(1, 2, 2), stride=(1, 2, 2)),
+        )
+        self.linear_visual = nn.Linear(96 * 4 * 8, 512)
+
+        # the audio-visual combining encoder based on GRU
+        self.grus = nn.ModuleList(
+            [
+                nn.GRU(
+                    input_size=512 * 2,
+                    hidden_size=512,
+                    num_layers=1,
+                    bidirectional=True,
+                )
+                for _ in range(4)
+            ]
+        )
+        self.gru_bnorms = nn.ModuleList(
+            [nn.BatchNorm1d(num_features=1024, affine=False) for _ in range(4)]
+        )
+        self.dropout = nn.Dropout(0.2)
+        self.linear = nn.Linear(
+            in_features=512 * 2, out_features=self.num_classes
+        )
+
+    def forward(self, x_v, x_a):
+        """
+        Args:
+          x_v:
+            Its shape is [N, 3, H, W]
+          x_a:
+            Its shape is [N, C, T]
+        Returns:
+          The output tensor has shape [N, T, C]
+        """
+        x_v = self.video_encoder(x_v)
+        x_v = x_v.permute(2, 0, 1, 3, 4).contiguous()
+        x_v = x_v.view(x_v.size(0), x_v.size(1), -1)
+        x_v = self.linear_visual(x_v)
+        x_a = self.audio_encoder(x_a)
+
+        x_v = x_v.permute(1, 0, 2)
+        x_a = x_a.permute(0, 2, 1)
+
+        # Repeat the visual features
+        # to cat with the audio features in time axis.
+        x_v_copy = x_v
+        x_v_stack = torch.stack((x_v, x_v_copy), dim=2)
+        x_v = x_v_stack.view(
+            x_v_stack.size(0), 2 * x_v_stack.size(1), x_v_stack.size(3)
+        )
+
+        x = torch.cat((x_v, x_a), dim=2)
+        x = x.permute(1, 0, 2)  # (N, C, T) -> (T, N, C) -> how GRU expects it
+        for gru, bnorm in zip(self.grus, self.gru_bnorms):
+            x_new, _ = gru(x)
+            x_new = bnorm(x_new.permute(1, 2, 0)).permute(
+                2, 0, 1
+            )  # (T, N, C) -> (N, C, T) -> (T, N, C)
+            x_new = self.dropout(x_new)
+            x = x_new + x  # skip connections
+        x = x.transpose(
+            1, 0
+        )  # (T, N, C) -> (N, T, C) -> linear expects "features" in the last dim
+        x = self.linear(x)
+        x = nn.functional.log_softmax(x, dim=-1)
+        return x

egs/grid/AVSR/combinenet_ctc_avsr/train.py

@@ -0,0 +1,625 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang
+#                                                    Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+from pathlib import Path
+from shutil import copyfile
+from typing import Optional, Tuple
+
+from utils import encode_supervisions
+
+import k2
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+import torch.optim as optim
+from torch.utils.data import DataLoader
+
+from local.dataset_av import dataset_av
+from lhotse.utils import fix_random_seed
+from model import CombineNet
+from torch import Tensor
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.nn.utils import clip_grad_norm_
+from torch.optim.lr_scheduler import StepLR
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall.checkpoint import load_checkpoint
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.graph_compiler import CtcTrainingGraphCompiler
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    get_env_info,
+    setup_logger,
+    str2bool,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        audionet_ctc_asr/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    is saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - exp_dir: It specifies the directory where all training related
+                   files, e.g., checkpoints, log, etc, are saved
+
+        - lang_dir: It contains language related input files such as
+                    "lexicon.txt"
+
+        - lr: It specifies the initial learning rate
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - weight_decay:  The weight_decay for the optimizer.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval` is 0
+
+        - beam_size: It is used in k2.ctc_loss
+
+        - reduction: It is used in k2.ctc_loss
+
+        - use_double_scores: It is used in k2.ctc_loss
+    """
+    params = AttributeDict(
+        {
+            "exp_dir": Path("combinenet_ctc_avsr/exp"),
+            "lang_dir": Path("data/lang_character"),
+            "lr": 4e-4,
+            "feature_dim": 80,
+            "weight_decay": 5e-4,
+            "subsampling_factor": 3,
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 1,
+            "reset_interval": 200,
+            "valid_interval": 1000,
+            "beam_size": 10,
+            "reduction": "sum",
+            "use_double_scores": True,
+            "env_info": get_env_info(),
+            # parameters for dataset
+            "video_path": Path("download/GRID/lip/"),
+            "anno_path": Path("download/GRID/GRID_align_txt"),
+            "train_list": Path("download/GRID/unseen_train.txt"),
+            "vid_padding": 80,
+            "aud_padding": 480,
+            "sample_rate": 16000,
+            "num_workers": 16,
+            "batch_size": 120,
+        }
+    )
+
+    return params
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+) -> None:
+    """Load checkpoint from file.
+
+    If params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`. Otherwise, this function does nothing.
+
+    Apart from loading state dict for `model`, `optimizer` and `scheduler`,
+    it also updates `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The learning rate scheduler we are using.
+    Returns:
+      Return None.
+    """
+    if params.start_epoch <= 0:
+        return
+
+    filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    scheduler: torch.optim.lr_scheduler._LRScheduler,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    batch: dict,
+    graph_compiler: CtcTrainingGraphCompiler,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute CTC loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of TdnnLstm in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      graph_compiler:
+        It is used to build a decoding graph from a ctc topo and training
+        transcript. The training transcript is contained in the given `batch`,
+        while the ctc topo is built when this compiler is instantiated.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+    """
+    device = graph_compiler.device
+    audio_feature = batch["aud"]
+    video_feature = batch["vid"]
+
+    audio_feature = audio_feature.permute(
+        0, 2, 1
+    )  # now feature size is (N, C, T)
+    assert audio_feature.ndim == 3
+    audio_feature = audio_feature.to(device)
+
+    assert video_feature.ndim == 5
+    video_feature = video_feature.to(device)
+
+    with torch.set_grad_enabled(is_training):
+        nnet_output = model(
+            video_feature, audio_feature
+        )  # nnet_output size is (N, T, C)
+
+    # NOTE: We need `encode_supervisions` to sort sequences with
+    # different duration in decreasing order, required by
+    # `k2.intersect_dense` called in `k2.ctc_loss`
+    supervision_segments, texts = encode_supervisions(nnet_output.size(), batch)
+
+    decoding_graph = graph_compiler.compile(texts)
+    dense_fsa_vec = k2.DenseFsaVec(
+        nnet_output,
+        supervision_segments,
+        allow_truncate=params.subsampling_factor - 1,
+    )
+    loss = k2.ctc_loss(
+        decoding_graph=decoding_graph,
+        dense_fsa_vec=dense_fsa_vec,
+        output_beam=params.beam_size,
+        reduction=params.reduction,
+        use_double_scores=params.use_double_scores,
+    )
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    info["frames"] = supervision_segments[:, 2].sum().item()
+    info["loss"] = loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    graph_compiler: CtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process. The validation loss
+    is saved in `params.valid_loss`.
+    """
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            batch=batch,
+            graph_compiler=graph_compiler,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: CtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      graph_compiler:
+        It is used to convert transcripts to FSAs.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["txt"])
+
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            batch=batch,
+            graph_compiler=graph_compiler,
+            is_training=True,
+        )
+        # summary stats.
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        optimizer.zero_grad()
+        loss.backward()
+        clip_grad_norm_(model.parameters(), 5.0, 2.0)
+        optimizer.step()
+
+        if batch_idx % params.log_interval == 0:
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}"
+            )
+        if batch_idx % params.log_interval == 0:
+
+            if tb_writer is not None:
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(
+                    tb_writer, "train/tot_", params.batch_idx_train
+                )
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer,
+                    "train/valid_",
+                    params.batch_idx_train,
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(42)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+    logging.info(params)
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+
+    graph_compiler = CtcTrainingGraphCompiler(lexicon=lexicon, device=device)
+
+    model = CombineNet(
+        num_features=params.feature_dim,
+        num_classes=max_token_id + 1,  # +1 for the blank symbol
+        subsampling_factor=params.subsampling_factor,
+    )
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        model = DDP(model, device_ids=[rank])
+
+    optimizer = optim.AdamW(
+        model.parameters(),
+        lr=params.lr,
+        weight_decay=params.weight_decay,
+    )
+    scheduler = StepLR(optimizer, step_size=10, gamma=0.8)
+
+    if checkpoints:
+        optimizer.load_state_dict(checkpoints["optimizer"])
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    grid = dataset_av(
+        params.video_path,
+        params.anno_path,
+        params.train_list,
+        params.feature_dim,
+        params.vid_padding,
+        params.aud_padding,
+        params.sample_rate,
+        "train",
+    )
+
+    train_dl = DataLoader(
+        grid,
+        batch_size=params.batch_size,
+        shuffle=True,
+        num_workers=params.num_workers,
+        drop_last=False,
+    )
+    valid_dl = train_dl
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        # train_dl.sampler.set_epoch(epoch)
+
+        if epoch > params.start_epoch:
+            logging.info(f"epoch {epoch}, lr: {scheduler.get_last_lr()[0]}")
+
+        if tb_writer is not None:
+            tb_writer.add_scalar(
+                "train/lr",
+                scheduler.get_last_lr()[0],
+                params.batch_idx_train,
+            )
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            tb_writer=tb_writer,
+            world_size=world_size,
+        )
+
+        scheduler.step()
+
+        if epoch % 1 == 0:
+            save_checkpoint(
+                params=params,
+                model=model,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                rank=rank,
+            )
+
+    logging.info("Done!")
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def main():
+    parser = get_parser()
+    # TimitAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+if __name__ == "__main__":
+    main()

egs/grid/AVSR/combinenet_ctc_avsr/utils.py

@@ -0,0 +1,51 @@
+# Copyright      2021  Xiaomi Corp.        (authors: Mingshuang Luo)
+#
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+
+
+def encode_supervisions(nnet_output_shape, batch):
+    """
+    Encodes Lhotse's ``batch["supervisions"]`` dict into
+    a pair of torch Tensor, and a list of transcription strings.
+
+    The supervision tensor has shape ``(batch_size, 3)``.
+    Its second dimension contains information about sequence index [0],
+    start frames [1] and num frames [2].
+
+    The batch items might become re-ordered during this operation -- the
+    returned tensor and list of strings are guaranteed to be consistent with
+    each other.
+    """
+    N, T, D = nnet_output_shape
+
+    supervisions_idx = torch.arange(0, N).to(torch.int32)
+    start_frames = [0 for _ in range(N)]
+    supervisions_start_frame = torch.tensor(start_frames).to(torch.int32)
+    num_frames = [T for _ in range(N)]
+    supervisions_num_frames = torch.tensor(num_frames).to(torch.int32)
+
+    supervision_segments = torch.stack(
+        (
+            supervisions_idx,
+            supervisions_start_frame,
+            supervisions_num_frames,
+        ),
+        1,
+    ).to(torch.int32)
+    texts = batch["txt"]
+
+    return supervision_segments, texts

egs/grid/AVSR/lipnet_ctc_vsr/decode.py

@@ -30,7 +30,7 @@ import torch
 import torch.nn as nn
 
 from torch.utils.data import DataLoader
-from local.dataset import dataset_GRID
+from local.dataset_visual import dataset_visual
 from model import LipNet
 
 from icefall.checkpoint import average_checkpoints, load_checkpoint
@@ -461,7 +461,7 @@ def main():
     model.to(device)
     model.eval()
 
-    grid = dataset_GRID(
+    grid = dataset_visual(
         params.video_path,
         params.anno_path,
         params.val_list,

egs/grid/AVSR/lipnet_ctc_vsr/train.py

@@ -32,7 +32,7 @@ import torch.nn as nn
 import torch.optim as optim
 from torch.utils.data import DataLoader
 
-from local.dataset import dataset_GRID
+from local.dataset_visual import dataset_visual
 from lhotse.utils import fix_random_seed
 from model import LipNet
 from torch import Tensor
@@ -528,7 +528,7 @@ def run(rank, world_size, args):
         optimizer.load_state_dict(checkpoints["optimizer"])
         scheduler.load_state_dict(checkpoints["scheduler"])
 
-    grid = dataset_GRID(
+    grid = dataset_visual(
         params.video_path,
         params.anno_path,
         params.train_list,

egs/grid/AVSR/local/dataset_audio.py

@@ -8,22 +8,22 @@ import torchaudio
 from torch.utils.data import Dataset
 
 
-class MyDataset(Dataset):
+class dataset_audio(Dataset):
     def __init__(
         self,
         video_path,
         anno_path,
         file_list,
         aud_padding,
-        phase,
         sample_rate,
         feature_dim,
+        phase,
     ):
         self.anno_path = anno_path
         self.aud_padding = aud_padding
-        self.phase = phase
         self.sample_rate = sample_rate
         self.feature_dim = feature_dim
+        self.phase = phase
         with open(file_list, "r") as f:
             self.videos = [
                 os.path.join(video_path, line.strip()) for line in f.readlines()

egs/grid/AVSR/local/dataset_av.py

@@ -0,0 +1,111 @@
+# encoding: utf-8
+import cv2
+import os
+import kaldifeat
+import numpy as np
+
+import torch
+import torchaudio
+from torch.utils.data import Dataset
+
+from cvtransforms import HorizontalFlip, ColorNormalize
+
+
+class dataset_av(Dataset):
+    def __init__(
+        self,
+        video_path,
+        anno_path,
+        file_list,
+        feature_dim,
+        vid_pad,
+        aud_pad,
+        sample_rate,
+        phase,
+    ):
+        self.anno_path = anno_path
+        self.vid_pad = vid_pad
+        self.aud_pad = aud_pad
+        self.feature_dim = feature_dim
+        self.sample_rate = sample_rate
+        self.phase = phase
+        with open(file_list, "r") as f:
+            self.videos = [
+                os.path.join(video_path, line.strip()) for line in f.readlines()
+            ]
+
+        self.data = []
+        for vid in self.videos:
+            items = vid.split(os.path.sep)
+            aud = (
+                vid.replace("lip", "audio_25k").replace("/video/mpg_6000", "")
+                + ".wav"
+            )
+            self.data.append((vid, aud, items[-4], items[-1]))
+
+    def __getitem__(self, idx):
+        (vid, aud, spk, name) = self.data[idx]
+
+        vid = self._load_vid(vid)
+        aud = self._load_aud(aud)
+        vid = self._padding(vid, self.vid_pad)
+        aud = self._padding(aud, self.aud_pad)
+        anno = self._load_anno(
+            os.path.join(self.anno_path, spk, "align", name + ".align")
+        )
+
+        if self.phase == "train":
+            vid = HorizontalFlip(vid)
+        vid = ColorNormalize(vid)
+
+        vid = self._padding(vid, self.vid_pad)
+        aud = self._padding(aud, self.aud_pad)
+
+        return {
+            "vid": torch.FloatTensor(vid.transpose(3, 0, 1, 2)),
+            "aud": torch.FloatTensor(aud),
+            "txt": anno.upper(),
+        }
+
+    def __len__(self):
+        return len(self.data)
+
+    def _load_vid(self, p):
+        files = os.listdir(p)
+        files = list(filter(lambda file: file.find(".jpg") != -1, files))
+        files = sorted(files, key=lambda file: int(os.path.splitext(file)[0]))
+        array = [cv2.imread(os.path.join(p, file)) for file in files]
+        array = list(filter(lambda im: im is not None, array))
+        array = [
+            cv2.resize(im, (128, 64), interpolation=cv2.INTER_LANCZOS4)
+            for im in array
+        ]
+        array = np.stack(array, axis=0).astype(np.float32)
+        return array
+
+    def _load_aud(self, filename):
+        opts = kaldifeat.FbankOptions()
+        opts.frame_opts.dither = 0
+        opts.frame_opts.snip_edges = False
+        opts.frame_opts.samp_freq = self.sample_rate
+        opts.mel_opts.num_bins = self.feature_dim
+        fbank = kaldifeat.Fbank(opts)
+        wave, sample_rate = torchaudio.load(filename)
+        features = fbank(wave[0])
+
+        return features
+
+    def _load_anno(self, name):
+        with open(name, "r") as f:
+            lines = [line.strip().split(" ") for line in f.readlines()]
+            txt = [line[2] for line in lines]
+            txt = list(filter(lambda s: not s.upper() in ["SIL", "SP"], txt))
+            txt = " ".join(txt)
+        return txt
+
+    def _padding(self, array, length):
+        array = [array[_] for _ in range(array.shape[0])]
+        size = array[0].shape
+        for i in range(length - len(array)):
+            array.append(np.zeros(size))
+        return np.stack(array, axis=0)

egs/grid/AVSR/local/dataset_visual.py

@@ -7,17 +7,17 @@ from torch.utils.data import Dataset
 from cvtransforms import HorizontalFlip, ColorNormalize
 
 
-class dataset_GRID(Dataset):
+class dataset_visual(Dataset):
     def __init__(
         self,
         video_path,
         anno_path,
         file_list,
-        vid_pad,
+        vid_padding,
         phase,
     ):
         self.anno_path = anno_path
-        self.vid_pad = vid_pad
+        self.vid_padding = vid_padding
         self.phase = phase
         with open(file_list, "r") as f:
             self.videos = [
@@ -44,7 +44,7 @@ class dataset_GRID(Dataset):
             vid = HorizontalFlip(vid)
         vid = ColorNormalize(vid)
 
-        vid = self._padding(vid, self.vid_pad)
+        vid = self._padding(vid, self.vid_padding)
 
         return {
             "vid": torch.FloatTensor(vid.transpose(3, 0, 1, 2)),