Add DistilHuBERT (huggingface#14174)

* Add conversion

* Rename

* Add an integration test and remove layer_norm

* Remove layer_norm from the converter

* wording

* Fix imports

src/transformers/models/hubert/configuration_hubert.py

@@ -70,6 +70,8 @@ class HubertConfig(PretrainedConfig):
             convolutional layers.
         feat_proj_dropout (:obj:`float`, `optional`, defaults to 0.0):
             The dropout probability for output of the feature extractor.
+        feat_proj_layer_norm (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Whether to apply LayerNorm to the output of the feature extractor.
         feat_extract_activation (:obj:`str, `optional`, defaults to :obj:`"gelu"`):
             The non-linear activation function (function or string) in the 1D convolutional layers of the feature
             extractor. If string, :obj:`"gelu"`, :obj:`"relu"`, :obj:`"selu"` and :obj:`"gelu_new"` are supported.
@@ -149,6 +151,7 @@ def __init__(
         hidden_dropout=0.1,
         activation_dropout=0.1,
         attention_dropout=0.1,
+        feat_proj_layer_norm=True,
         feat_proj_dropout=0.0,
         final_dropout=0.1,
         layerdrop=0.1,
@@ -195,6 +198,7 @@ def __init__(
         self.hidden_dropout = hidden_dropout
         self.attention_dropout = attention_dropout
         self.activation_dropout = activation_dropout
+        self.feat_proj_layer_norm = feat_proj_layer_norm
         self.feat_proj_dropout = feat_proj_dropout
         self.final_dropout = final_dropout
         self.layerdrop = layerdrop

src/transformers/models/hubert/convert_distilhubert_original_s3prl_checkpoint_to_pytorch.py

@@ -0,0 +1,218 @@
+# coding=utf-8
+# Copyright 2021 The HuggingFace Inc. team.
+#
+# 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.
+"""Convert Hubert checkpoint."""
+
+
+import argparse
+
+import torch
+
+from s3prl.hub import distilhubert
+from transformers import HubertConfig, HubertModel, Wav2Vec2FeatureExtractor, logging
+
+
+logging.set_verbosity_info()
+logger = logging.get_logger(__name__)
+
+MAPPING = {
+    "post_extract_proj": "feature_projection.projection",
+    "encoder.pos_conv.0": "encoder.pos_conv_embed.conv",
+    "self_attn.k_proj": "encoder.layers.*.attention.k_proj",
+    "self_attn.v_proj": "encoder.layers.*.attention.v_proj",
+    "self_attn.q_proj": "encoder.layers.*.attention.q_proj",
+    "self_attn.out_proj": "encoder.layers.*.attention.out_proj",
+    "self_attn_layer_norm": "encoder.layers.*.layer_norm",
+    "fc1": "encoder.layers.*.feed_forward.intermediate_dense",
+    "fc2": "encoder.layers.*.feed_forward.output_dense",
+    "final_layer_norm": "encoder.layers.*.final_layer_norm",
+    "encoder.layer_norm": "encoder.layer_norm",
+    "mask_emb": "masked_spec_embed",
+}
+
+
+def set_recursively(hf_pointer, key, value, full_name, weight_type):
+    for attribute in key.split("."):
+        hf_pointer = getattr(hf_pointer, attribute)
+
+    if weight_type is not None:
+        hf_shape = getattr(hf_pointer, weight_type).shape
+    else:
+        hf_shape = hf_pointer.shape
+
+    assert (
+        hf_shape == value.shape
+    ), f"Shape of hf {key + '.' + weight_type if weight_type is not None else ''} is {hf_shape}, but should be {value.shape} for {full_name}"
+
+    if weight_type == "weight":
+        hf_pointer.weight.data = value
+    elif weight_type == "weight_g":
+        hf_pointer.weight_g.data = value
+    elif weight_type == "weight_v":
+        hf_pointer.weight_v.data = value
+    elif weight_type == "bias":
+        hf_pointer.bias.data = value
+    else:
+        hf_pointer.data = value
+
+    logger.info(f"{key + '.' + weight_type if weight_type is not None else ''} was initialized from {full_name}.")
+
+
+def recursively_load_weights(fairseq_model, hf_model):
+    unused_weights = []
+    fairseq_dict = fairseq_model.state_dict()
+
+    feature_extractor = hf_model.feature_extractor
+
+    for name, value in fairseq_dict.items():
+        is_used = False
+        if "conv_layers" in name:
+            load_conv_layer(
+                name,
+                value,
+                feature_extractor,
+                unused_weights,
+                hf_model.config.feat_extract_norm == "group",
+            )
+            is_used = True
+        else:
+            for key, mapped_key in MAPPING.items():
+                mapped_key = mapped_key
+
+                if key in name:
+                    is_used = True
+                    if "*" in mapped_key:
+                        layer_index = name.split(key)[0].split(".")[-2]
+                        mapped_key = mapped_key.replace("*", layer_index)
+                    if "weight_g" in name:
+                        weight_type = "weight_g"
+                    elif "weight_v" in name:
+                        weight_type = "weight_v"
+                    elif "weight" in name:
+                        weight_type = "weight"
+                    elif "bias" in name:
+                        weight_type = "bias"
+                    else:
+                        weight_type = None
+                    set_recursively(hf_model, mapped_key, value, name, weight_type)
+                continue
+        if not is_used:
+            unused_weights.append(name)
+
+    logger.warning(f"Unused weights: {unused_weights}")
+
+
+def load_conv_layer(full_name, value, feature_extractor, unused_weights, use_group_norm):
+    name = full_name.split("conv_layers.")[-1]
+    items = name.split(".")
+    layer_id = int(items[0])
+    type_id = int(items[1])
+
+    if type_id == 0:
+        if "bias" in name:
+            assert (
+                value.shape == feature_extractor.conv_layers[layer_id].conv.bias.data.shape
+            ), f"{full_name} has size {value.shape}, but {feature_extractor.conv_layers[layer_id].conv.bias.data.shape} was found."
+            feature_extractor.conv_layers[layer_id].conv.bias.data = value
+            logger.info(f"Feat extract conv layer {layer_id} was initialized from {full_name}.")
+        elif "weight" in name:
+            assert (
+                value.shape == feature_extractor.conv_layers[layer_id].conv.weight.data.shape
+            ), f"{full_name} has size {value.shape}, but {feature_extractor.conv_layers[layer_id].conv.weight.data.shape} was found."
+            feature_extractor.conv_layers[layer_id].conv.weight.data = value
+            logger.info(f"Feat extract conv layer {layer_id} was initialized from {full_name}.")
+    elif (type_id == 2 and not use_group_norm) or (type_id == 2 and layer_id == 0 and use_group_norm):
+        if "bias" in name:
+            assert (
+                value.shape == feature_extractor.conv_layers[layer_id].layer_norm.bias.data.shape
+            ), f"{full_name} has size {value.shape}, but {feature_extractor[layer_id].layer_norm.bias.data.shape} was found."
+            feature_extractor.conv_layers[layer_id].layer_norm.bias.data = value
+            logger.info(f"Feat extract layer norm weight of layer {layer_id} was initialized from {full_name}.")
+        elif "weight" in name:
+            assert (
+                value.shape == feature_extractor.conv_layers[layer_id].layer_norm.weight.data.shape
+            ), f"{full_name} has size {value.shape}, but {feature_extractor[layer_id].layer_norm.weight.data.shape} was found."
+            feature_extractor.conv_layers[layer_id].layer_norm.weight.data = value
+            logger.info(f"Feat extract layer norm weight of layer {layer_id} was initialized from {full_name}.")
+    else:
+        unused_weights.append(full_name)
+
+
+def convert_config(model):
+    config = HubertConfig()
+    fs_config = model.config
+
+    config.activation_dropout = fs_config.activation_dropout
+    config.apply_spec_augment = False
+    config.attention_dropout = fs_config.attention_dropout
+    config.conv_bias = False
+    conv_layers = eval(fs_config.extractor_conv_feature_layers)
+    config.conv_dim = [x[0] for x in conv_layers]
+    config.conv_kernel = [x[1] for x in conv_layers]
+    config.conv_stride = [x[2] for x in conv_layers]
+    config.feat_extract_activation = "gelu"
+    config.feat_extract_norm = "layer" if fs_config.extractor_mode == "layer_norm" else "group"
+    config.feat_proj_layer_norm = False
+    config.feat_proj_dropout = 0.0
+    config.final_dropout = 0.0
+    config.hidden_act = fs_config.activation_fn
+    config.hidden_dropout = fs_config.dropout
+    config.hidden_size = fs_config.encoder_embed_dim
+    config.initializer_range = 0.02
+    config.intermediate_size = fs_config.encoder_ffn_embed_dim
+    config.layer_norm_eps = 1e-5
+    config.layerdrop = 0.0
+    config.num_attention_heads = fs_config.encoder_attention_heads
+    config.num_conv_pos_embedding_groups = fs_config.conv_pos_groups
+    config.num_conv_pos_embeddings = fs_config.conv_pos
+    config.num_feat_extract_layers = len(conv_layers)
+    config.num_hidden_layers = fs_config.encoder_layers
+
+    return config
+
+
+@torch.no_grad()
+def convert_hubert_checkpoint(pytorch_dump_folder_path, config_path=None):
+    """
+    Copy/paste/tweak model's weights to transformers design.
+    """
+    model = distilhubert().model.model
+
+    if config_path is not None:
+        config = HubertConfig.from_pretrained(config_path)
+    else:
+        config = convert_config(model)
+    model = model.eval()
+
+    feature_extractor = Wav2Vec2FeatureExtractor(
+        feature_size=1,
+        sampling_rate=16000,
+        padding_value=0,
+        do_normalize=False,
+        return_attention_mask=False,
+    )
+    hf_model = HubertModel(config)
+
+    recursively_load_weights(model, hf_model)
+
+    feature_extractor.save_pretrained(pytorch_dump_folder_path)
+    hf_model.save_pretrained(pytorch_dump_folder_path)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--pytorch_dump_folder_path", default=None, type=str, help="Path to the output PyTorch model.")
+    parser.add_argument("--config_path", default=None, type=str, help="Path to hf config.json of model to convert")
+    args = parser.parse_args()
+    convert_hubert_checkpoint(args.pytorch_dump_folder_path, args.config_path)

src/transformers/models/hubert/modeling_hubert.py

@@ -325,13 +325,16 @@ def custom_forward(*inputs):
 class HubertFeatureProjection(nn.Module):
     def __init__(self, config):
         super().__init__()
-        self.layer_norm = nn.LayerNorm(config.conv_dim[-1], eps=config.layer_norm_eps)
+        self.feat_proj_layer_norm = config.feat_proj_layer_norm
+        if self.feat_proj_layer_norm:
+            self.layer_norm = nn.LayerNorm(config.conv_dim[-1], eps=config.layer_norm_eps)
         self.projection = nn.Linear(config.conv_dim[-1], config.hidden_size)
         self.dropout = nn.Dropout(config.feat_proj_dropout)
 
     def forward(self, hidden_states):
         # non-projected hidden states are needed for quantization
-        hidden_states = self.layer_norm(hidden_states)
+        if self.feat_proj_layer_norm:
+            hidden_states = self.layer_norm(hidden_states)
         hidden_states = self.projection(hidden_states)
         hidden_states = self.dropout(hidden_states)
         return hidden_states

tests/test_modeling_hubert.py

@@ -760,3 +760,46 @@ def test_inference_emotion_recognition(self):
         self.assertListEqual(predicted_ids.tolist(), expected_labels)
         # TODO: lower the tolerance after merging the padding fix https://github.com/pytorch/fairseq/pull/3572
         self.assertTrue(torch.allclose(predicted_logits, expected_logits, atol=1e-1))
+
+    def test_inference_distilhubert(self):
+        model = HubertModel.from_pretrained("anton-l/distilhubert").to(torch_device)
+        processor = Wav2Vec2FeatureExtractor.from_pretrained("anton-l/distilhubert")
+
+        # TODO: can't test on batched inputs due to incompatible padding https://github.com/pytorch/fairseq/pull/3572
+        input_speech = self._load_datasamples(1)
+
+        inputs = processor(input_speech, return_tensors="pt", padding=True)
+
+        input_values = inputs.input_values.to(torch_device)
+
+        with torch.no_grad():
+            outputs = model(input_values).last_hidden_state
+
+        # expected outputs taken from the original SEW implementation
+        expected_outputs_first = torch.tensor(
+            [
+                [
+                    [-0.3505, 0.1167, 0.0608, 0.1294],
+                    [-0.3085, 0.0481, 0.1106, 0.0955],
+                    [-0.3107, -0.0391, 0.0739, 0.1360],
+                    [-0.2385, -0.1795, -0.0928, 0.2389],
+                ]
+            ],
+            device=torch_device,
+        )
+        expected_outputs_last = torch.tensor(
+            [
+                [
+                    [-0.0732, 0.0255, 0.0529, -0.1372],
+                    [-0.0812, 0.1259, 0.0564, -0.0438],
+                    [-0.0054, 0.0758, -0.0002, -0.1617],
+                    [0.0133, -0.0320, -0.0687, 0.0062],
+                ]
+            ],
+            device=torch_device,
+        )
+        expected_output_sum = -3776.0730
+
+        self.assertTrue(torch.allclose(outputs[:, :4, :4], expected_outputs_first, atol=5e-3))
+        self.assertTrue(torch.allclose(outputs[:, -4:, -4:], expected_outputs_last, atol=5e-3))
+        self.assertTrue(abs(outputs.sum() - expected_output_sum) < 0.1)