Conform to latest fairseq

contentvec/data/__init__.py

@@ -40,6 +40,11 @@
 from .numel_dataset import NumelDataset
 from .num_samples_dataset import NumSamplesDataset
 from .offset_tokens_dataset import OffsetTokensDataset
+from .padding_mask_dataset import (
+    LeftPaddingMaskDataset,
+    PaddingMaskDataset,
+    RightPaddingMaskDataset,
+)
 from .pad_dataset import LeftPadDataset, PadDataset, RightPadDataset
 from .prepend_dataset import PrependDataset
 from .prepend_token_dataset import PrependTokenDataset
@@ -49,6 +54,7 @@
 from .roll_dataset import RollDataset
 from .round_robin_zip_datasets import RoundRobinZipDatasets
 from .sort_dataset import SortDataset
+from .speech_dlm_dataset import SpeechDLMDataset
 from .strip_token_dataset import StripTokenDataset
 from .subsample_dataset import SubsampleDataset
 from .token_block_dataset import TokenBlockDataset
@@ -58,6 +64,7 @@
 from .multilingual.sampled_multi_dataset import SampledMultiDataset
 from .multilingual.sampled_multi_epoch_dataset import SampledMultiEpochDataset
 from .fasta_dataset import FastaDataset, EncodedFastaDataset
+from .transform_eos_concat_langpair_dataset import TransformEosConcatLangPairDataset
 
 from .iterators import (
     CountingIterator,
@@ -120,11 +127,13 @@
     "SampledMultiEpochDataset",
     "ShardedIterator",
     "SortDataset",
+    "SpeechDLMDataset",
     "StripTokenDataset",
     "SubsampleDataset",
     "TokenBlockDataset",
     "TransformEosDataset",
     "TransformEosLangPairDataset",
+    "TransformEosConcatLangPairDataset",
     "TruncateDataset",
     "TruncatedDictionary",
 ]

contentvec/models/hubert/__init__.py

@@ -2,6 +2,7 @@
 #
 # This source code is licensed under the MIT license found in the
 # LICENSE file in the root directory of this source tree.
+
 from .hubert import *  # noqa
 from .contentvec import *  # noqa
 from .hubert_asr import *  # noqa

contentvec/models/wav2vec/__init__.py

@@ -6,4 +6,5 @@
 from .wav2vec import *  # noqa
 from .wav2vec2 import *  # noqa
 from .wav2vec2_asr import *  # noqa
+from .wav2vec2_laser import *  # noqa
 from .wav2vec2_1 import *

contentvec/modules/__init__.py

@@ -12,10 +12,13 @@
 from .conv_tbc import ConvTBC
 from .cross_entropy import cross_entropy
 from .downsampled_multihead_attention import DownsampledMultiHeadAttention
-from .dynamic_convolution import DynamicConv, DynamicConv1dTBC
+from .dynamic_convolution import DynamicConv, DynamicConv1dTBC, DynamicConv_scripatable
 from .dynamic_crf_layer import DynamicCRF
+from .ema_module import EMAModuleConfig, EMAModule
 from .fairseq_dropout import FairseqDropout
+from .fp32_batch_norm import Fp32BatchNorm
 from .fp32_group_norm import Fp32GroupNorm, GroupNormMasked
+from .fp32_instance_norm import Fp32InstanceNorm
 from .gelu import gelu, gelu_accurate
 from .grad_multiply import GradMultiply
 from .gumbel_vector_quantizer import GumbelVectorQuantizer
@@ -30,7 +33,7 @@
 from .lstm_cell_with_zoneout import LSTMCellWithZoneOut
 from .multihead_attention import MultiheadAttention
 from .positional_embedding import PositionalEmbedding
-from .same_pad import SamePad
+from .same_pad import SamePad, SamePad2d
 from .scalar_bias import ScalarBias
 from .sinusoidal_positional_embedding import SinusoidalPositionalEmbedding
 from .transformer_sentence_encoder_layer import TransformerSentenceEncoderLayer
@@ -39,6 +42,15 @@
 from .unfold import unfold1d
 from .transformer_layer import TransformerDecoderLayer, TransformerEncoderLayer
 from .vggblock import VGGBlock
+from .espnet_multihead_attention import (
+    ESPNETMultiHeadedAttention,
+    RelPositionMultiHeadedAttention,
+    RotaryPositionMultiHeadedAttention,
+)
+from .rotary_positional_embedding import RotaryPositionalEmbedding
+from .positional_encoding import (
+    RelPositionalEncoding,
+)
 
 __all__ = [
     "AdaptiveInput",
@@ -51,10 +63,15 @@
     "DownsampledMultiHeadAttention",
     "DynamicConv1dTBC",
     "DynamicConv",
+    "DynamicConv_scripatable",
     "DynamicCRF",
+    "EMAModule",
+    "EMAModuleConfig",
     "FairseqDropout",
+    "Fp32BatchNorm",
     "Fp32GroupNorm",
     "Fp32LayerNorm",
+    "Fp32InstanceNorm",
     "gelu",
     "gelu_accurate",
     "GradMultiply",
@@ -73,6 +90,7 @@
     "MultiheadAttention",
     "PositionalEmbedding",
     "SamePad",
+    "SamePad2d",
     "ScalarBias",
     "SinusoidalPositionalEmbedding",
     "TransformerSentenceEncoderLayer",
@@ -82,4 +100,10 @@
     "TransposeLast",
     "VGGBlock",
     "unfold1d",
-]
+    "ESPNETMultiheadedAttention",
+    "PositionalEmbedding",
+    "RelPositionMultiHeadedAttention",
+    "RelPositionalEncoding",
+    "RotaryPositionalEmbedding",
+    "RotaryPositionMultiHeadedAttention",
+]

setup.sh

@@ -28,7 +28,7 @@ if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
     fi
     cd ${cwd}/fairseq
     # checkout the fairseq version to use
-    git reset --hard 0b21875e45f332bedbcc0617dcf9379d3c03855f
+    git reset --hard 3f6ba43f07a6e9e2acf957fc24e57251a7a3f55c
 
     if [ $(pip freeze | grep fairseq | wc -l ) -gt 0 ]; then
         echo "Already installed fairseq. Skip..."