Merge branch 'master' of upstream

docs/data.md

@@ -6,7 +6,7 @@ The most important concepts of `parakeet.data` are `DatasetMixin`, `DataCargo`,
 
 ## Dataset
 
-Dataset, as we assume here, is a list of examples. You can get its length by `len(dataset)`(which means it length is known, and we have to implement `__len__()` method for it). And you can access its items randomly by `dataset[i]`(which means we have to implement `__getitem__()` method for it). Furthermore,  you can iterate over it by `iter(dataset)` or `for example in dataset`, which means we have to implement `__iter__()` method for it.
+Dataset, as we assume here, is a list of examples. You can get its length by `len(dataset)`(which means its length is known, and we have to implement `__len__()` method for it). And you can access its items randomly by `dataset[i]`(which means we have to implement `__getitem__()` method for it). Furthermore,  you can iterate over it by `iter(dataset)` or `for example in dataset`, which means we have to implement `__iter__()` method for it.
 
 ### DatasetMixin
 
@@ -16,11 +16,11 @@ We also define several high-order Dataset classes, the obejcts of which can be b
 
 ### TupleDataset
 
-Dataset that is a combination of sevral datasets of the same length. An example of a `Tupledataset` is a tuple of examples of its constituent datasets.
+Dataset that is a combination of several datasets of the same length. An example of a `Tupledataset` is a tuple of examples of its constituent datasets.
 
 ### DictDataset
 
-Dataset that is a combination of sevral datasets of the same length. An example of the `Dictdataset` is a dict of examples of its constituent datasets.
+Dataset that is a combination of several datasets of the same length. An example of the `Dictdataset` is a dict of examples of its constituent datasets.
 
 ### SliceDataset
 
@@ -36,11 +36,11 @@ Dataset that is a combination of sevral datasets of the same length. An example
 
 ### TransformDataset
 
-A `TransformeDataset` is created by applying a `transform` to the base dataset. The `transform` is a callable object which takes an `example` of the base dataset as parameter and returns an `example` of the `TransformDataset`. The transformation is lazy, which means it is applied to an example only when requested.
+A `TransformeDataset` is created by applying a `transform` to the examples of the base dataset. The `transform` is a callable object which takes an example of the base dataset as parameter and returns an example of the `TransformDataset`. The transformation is lazy, which means it is applied to an example only when requested.
 
 ### FilterDataset
 
-A `FilterDataset` is created by applying a `filter` to the base dataset. A `filter` is a predicate that takes an `example` of the base dataset as parameter and returns a boolean. Only those examples that pass the filter are included in the `FilterDataset`.
+A `FilterDataset` is created by applying a `filter` to the base dataset. A `filter` is a predicate that takes an example of the base dataset as parameter and returns a boolean. Only those examples that pass the filter are included in the `FilterDataset`.
 
 Note that the filter is applied to all the examples in the base dataset when initializing a `FilterDataset`.
 
@@ -52,11 +52,11 @@ Finally, if preprocessing the dataset is slow and the processed dataset is too l
 
 ## DataCargo
 
-`DataCargo`, like `Dataset`, is an iterable object, but it is an iterable oject of batches. We need `Datacargo` because in deep learning, batching examples into batches exploits the computational resources of modern hardwares. You can iterate it by `iter(datacargo)` or `for batch in datacargo`. `DataCargo` is an iterable object but not an iterator, in that in can be iterated more than once.
+`DataCargo`, like `Dataset`, is an iterable object, but it is an iterable oject of batches. We need `Datacargo` because in deep learning, batching examples into batches exploits the computational resources of modern hardwares. You can iterate over it by `iter(datacargo)` or `for batch in datacargo`. `DataCargo` is an iterable object but not an iterator, in that in can be iterated over more than once.
 
 ### batch function
 
-The concept of `batch` is something transformed from a list of examples. Assume that an example is a structure(tuple in python, or struct in C and C++) consists of several fields, then a list of examples is an array of structures(AOS, e.g. a dataset is an AOS). Then a batch here is a structure of arrays (SOA). Here is an example:
+The concept of a `batch` is something transformed from a list of examples. Assume that an example is a structure(tuple in python, or struct in C and C++) consists of several fields, then a list of examples is an array of structures(AOS, e.g. a dataset is an AOS). Then a batch here is a structure of arrays (SOA). Here is an example:
 
 The table below represents 2 examples, each of which contains 5 fields.
 
@@ -93,7 +93,7 @@ Equipped with a batch function(we have known __how to batch__), here comes the n
 
 A `Sampler` is represented as an iterable object of integers. Assume the dataset has `N` examples, then an iterable object of intergers in the range`[0, N)` is an appropriate sampler for this dataset to build a `DataCargo`.
 
-We provide several samplers that is ready to use. The `SequentialSampler`, `RandomSampler` and so on.
+We provide several samplers that are ready to use, for example, `SequentialSampler` and `RandomSampler`.
 
 ## DataIterator
 
@@ -309,7 +309,7 @@ valid_cargo = DataCargo(
   sampler=SequentialSampler(ljspeech_valid))
 ```
 
-Here comes the next question, how to bring batches into Paddle's computation. Do we need some adaptor to transform numpy.ndarray into Paddle's native Variable type? Yes.
+Here comes the next question, how to bring batches into Paddle's computation. Do we need some adapter to transform numpy.ndarray into Paddle's native Variable type? Yes.
 
 First we can use `var = dg.to_variable(array)` to transform ndarray into Variable.
 
@@ -326,16 +326,16 @@ for batch in train_cargo:
 In the code above, processing of the data and training of the model run in the same process. So the next batch starts to load after the training of the current batch has finished. There is actually better solutions for this. Data processing and model training can be run asynchronously. To accomplish this, we would use `DataLoader` from Paddle. This serves as an adapter to transform an iterable object of batches into another iterable object of batches, which runs asynchronously and transform each ndarray into `Variable`.
 
 ```python
-# connects our data cargos with corresponding DataLoader
+# connect our data cargos with corresponding DataLoader
 # now the data cargo is connected with paddle
 with dg.guard(place):
-  train_loader = fluid.io.DataLoader.from_generator(
-   capacity=10,return_list=True).set_batch_generator(train_cargo, place)
-  valid_loader = fluid.io.DataLoader.from_generator(
-   capacity=10, return_list=True).set_batch_generator(valid_cargo, place)
-
-  # iterate over the dataloader
-  for batch in train_loader:
-    audios, mels, audio_starts = batch
-    # your trains cript here
+    train_loader = fluid.io.DataLoader.from_generator(
+        capacity=10,return_list=True).set_batch_generator(train_cargo, place)
+   valid_loader = fluid.io.DataLoader.from_generator(
+        capacity=10, return_list=True).set_batch_generator(valid_cargo, place)
+
+    # iterate over the dataloader
+    for batch in train_loader:
+        audios, mels, audio_starts = batch
+        # your trains cript here
 ```

docs/experiment_guide.md

@@ -2,7 +2,7 @@
 
 For a general deep learning experiment, there are 4 parts to care for.
 
-1. Preprocess dataset to meet the needs for model training and iterate them in batches;
+1. Preprocess dataset to meet the needs for model training and iterate over them in batches;
 2. Define the model and the optimizer;
 3. Write the training process (including forward-backward computation, parameter update, logging, evaluation, etc.)
 4. Configure and launch the experiment.
@@ -13,13 +13,13 @@ For processing data, `parakeet.data` provides `DatasetMixin`, `DataCargo` and `D
 
 Dataset is an iterable object of examples. `DatasetMixin` provides the standard indexing interface, and other classes in [parakeet.data.dataset](../parakeet/data/dataset.py) provide flexible interfaces for building customized datasets.
 
-`DataCargo` is an iterable object of batches. It differs from a dataset in that it can be iterated in batches. In addition to a dataset, a `Sampler` and a `batch function` are required to build a `DataCargo`. `Sampler` specifies which examples to pick, and `batch function` specifies how to create a batch from them. Commonly used `Samplers` are provides by [parakeet.data](../parakeet/data/). Users should define a `batch function` for a datasets, in order to batch its examples.
+`DataCargo` is an iterable object of batches. It differs from a dataset in that it can be iterated over in batches. In addition to a dataset, a `Sampler` and a `batch function` are required to build a `DataCargo`. `Sampler` specifies which examples to pick, and `batch function` specifies how to create a batch from them. Commonly used `Samplers` are provided by [parakeet.data](../parakeet/data/). Users should define a `batch function` for a datasets, in order to batch its examples.
 
- `DataIterator` is an iterator class for `DataCargo`. It is create when explicitly creating an iterator of a `DataCargo` by `iter(DataCargo)`, or iterating a `DataCargo` with `for` loop.
+ `DataIterator` is an iterator class for `DataCargo`. It is create when explicitly creating an iterator of a `DataCargo` by `iter(DataCargo)`, or iterating over a `DataCargo` with `for` loop.
 
 Data processing is splited into two phases: sample-level processing and batching.
 
-1. Sample-level processing. This process is transforming an example into another example. This process can be defined as `get_example()` method of a dataset, or as a `transform` (callable object) and build a `TransformDataset` with it.
+1. Sample-level processing. This process is transforming an example into another. This process can be defined as `get_example()` method of a dataset, or as a `transform` (callable object) and build a `TransformDataset` with it.
 
 2. Batching. It is the process of transforming a list of examples into a batch. The rationale is to transform an array of structures into a structure of arrays. We generally define a batch function (or a callable object) to do this.
 
@@ -37,7 +37,7 @@ The user need to define a customized transform and a batch function to accomplis
 
 ## Model
 
-Parakeet provides commonly used functions, modules and models for the users to define their own models. Functions contains no trainable `Parameter`s, and are used in modules and models. Modules and modes are subclasses of `fluid.dygraph.Layer`. The distinction is that `module`s tend to be generic, simple and highly reusable, while `model`s tend to be task-sepcific, complicated and not that reusable. Some models are so complicated that we extract building blocks from it as separate classes but if these building blocks are not common and reusable enough, they are considered as submodels.
+Parakeet provides commonly used functions, modules and models for the users to define their own models. Functions contain no trainable `Parameter`s, and are used in modules and models. Modules and modes are subclasses of `fluid.dygraph.Layer`. The distinction is that `module`s tend to be generic, simple and highly reusable, while `model`s tend to be task-sepcific, complicated and not that reusable. Some models are so complicated that we extract building blocks from it as separate classes but if these building blocks are not common and reusable enough, they are considered as submodels.
 
 In the structure of the project, modules are placed in [parakeet.modules](../parakeet/modules/), while models are in [parakeet.models](../parakeet/models) and grouped into folders like `waveflow` and `wavenet`, which include the whole model and their submodels.
 
@@ -55,9 +55,9 @@ Training process is basically running a training loop for multiple times. A typi
 4. Updating Parameters;
 5. Evaluating the model on validation dataset;
 6. Logging or saving intermediate results;
-7. Saving checkpoint of the model and the optimizer.
+7. Saving checkpoints of the model and the optimizer.
 
-In section `DataProcrssing` we have cover 1 and 2.
+In section `DataProcessing` we have cover 1 and 2.
 
 `Model` and `Optimizer` cover 3 and 4.
 
@@ -66,7 +66,7 @@ To keep the training loop clear, it's a good idea to define functions for saving
 Code is typically organized in this way:
 
 ```text
-├── configs          (example configuration)
+├── configs/         (example configuration)
 ├── data.py          (definition of custom Dataset, transform and batch function)
 ├── README.md        (README for the experiment)
 ├── synthesis.py     (code for inference)
@@ -76,11 +76,11 @@ Code is typically organized in this way:
 
 ## Configuration
 
-Deep learning experiments have many options to configure. These configurations can be roughly grouped into different types: configurations about path of the dataset and path to save results, configurations about how to process data, configuration about the model and configurations about the training process.
+Deep learning experiments have many options to configure. These configurations can be roughly grouped into different types: configurations about path of the dataset and path to save results, configurations about how to process data, configurations about the model and configurations about the training process.
 
 Some configurations tend to change when running the code at different times, for example, path of the data and path to save results and whether to load model before training, etc. For these configurations, it's better to define them as command line arguments. We use `argparse` to handle them.
 
-Other groups of configuration may overlap with others. For example, data processing and model may have some common options. The recommended way is to save them as configuration files, for example, `yaml` or `json`. We prefer `yaml`, for it is more human-reabable.
+Other groups of configurations may overlap with others. For example, data processing and model may have some common options. The recommended way is to save them as configuration files, for example, `yaml` or `json`. We prefer `yaml`, for it is more human-reabable.
 
 
 

examples/fastspeech/synthesis.sh

@@ -6,6 +6,7 @@ python -u synthesis.py \
 --checkpoint_path='checkpoint/' \
 --fastspeech_step=71000 \
 --log_dir='./log' \
+--config_path='configs/synthesis.yaml' \
 
 if [ $? -ne 0 ]; then
     echo "Failed in synthesis!"

examples/fastspeech/train.sh

@@ -13,7 +13,7 @@ python -u train.py \
 --transformer_step=160000 \
 --save_path='./checkpoint' \
 --log_dir='./log' \
---config_path='config/fastspeech.yaml' \
+--config_path='configs/fastspeech.yaml' \
 #--checkpoint_path='./checkpoint' \
 #--fastspeech_step=97000 \
 

examples/transformer_tts/synthesis.py

@@ -84,7 +84,7 @@ def synthesis(text_input, args):
             dec_slf_mask = get_triu_tensor(
                 mel_input.numpy(), mel_input.numpy()).astype(np.float32)
             dec_slf_mask = fluid.layers.cast(
-                dg.to_variable(dec_slf_mask == 0), np.float32)
+                dg.to_variable(dec_slf_mask != 0), np.float32) * (-2**32 + 1)
             pos_mel = np.arange(1, mel_input.shape[1] + 1)
             pos_mel = fluid.layers.unsqueeze(dg.to_variable(pos_mel), [0])
             mel_pred, postnet_pred, attn_probs, stop_preds, attn_enc, attn_dec = model(
@@ -157,6 +157,5 @@ if __name__ == '__main__':
     parser = argparse.ArgumentParser(description="Synthesis model")
     add_config_options_to_parser(parser)
     args = parser.parse_args()
-    synthesis(
-        "They emphasized the necessity that the information now being furnished be handled with judgment and care.",
-        args)
+    synthesis("Parakeet stands for Paddle PARAllel text-to-speech toolkit.",
+              args)

examples/transformer_tts/synthesis.sh

@@ -2,14 +2,14 @@
 # train model
 CUDA_VISIBLE_DEVICES=0 \
 python -u synthesis.py \
---max_len=600 \
---transformer_step=160000 \
---vocoder_step=90000 \
+--max_len=300 \
+--transformer_step=120000 \
+--vocoder_step=100000 \
 --use_gpu=1 \
 --checkpoint_path='./checkpoint' \
 --log_dir='./log' \
 --sample_path='./sample' \
---config_path='config/synthesis.yaml' \
+--config_path='configs/synthesis.yaml' \
 
 if [ $? -ne 0 ]; then
     echo "Failed in training!"

examples/transformer_tts/train_transformer.sh

@@ -14,7 +14,7 @@ python -u train_transformer.py \
 --data_path='../../dataset/LJSpeech-1.1' \
 --save_path='./checkpoint' \
 --log_dir='./log' \
---config_path='config/train_transformer.yaml' \
+--config_path='configs/train_transformer.yaml' \
 #--checkpoint_path='./checkpoint' \
 #--transformer_step=160000 \
 

examples/transformer_tts/train_vocoder.sh

@@ -12,7 +12,7 @@ python -u train_vocoder.py \
 --data_path='../../dataset/LJSpeech-1.1' \
 --save_path='./checkpoint' \
 --log_dir='./log' \
---config_path='config/train_vocoder.yaml' \
+--config_path='configs/train_vocoder.yaml' \
 #--checkpoint_path='./checkpoint' \
 #--vocoder_step=27000 \
 

parakeet/models/fastspeech/decoder.py

@@ -23,12 +23,26 @@ class Decoder(dg.Layer):
                  n_layers,
                  n_head,
                  d_k,
-                 d_v,
+                 d_q,
                  d_model,
                  d_inner,
                  fft_conv1d_kernel,
                  fft_conv1d_padding,
                  dropout=0.1):
+        """Decoder layer of FastSpeech.
+
+        Args:
+            len_max_seq (int): the max mel len of sequence.
+            n_layers (int): the layers number of FFTBlock.
+            n_head (int): the head number of multihead attention.
+            d_k (int): the dim of key in multihead attention.
+            d_q (int): the dim of query in multihead attention.
+            d_model (int): the dim of hidden layer in multihead attention.
+            d_inner (int): the dim of hidden layer in ffn.
+            fft_conv1d_kernel (int): the conv kernel size in FFTBlock.
+            fft_conv1d_padding (int): the conv padding size in FFTBlock.
+            dropout (float, optional): dropout probability of FFTBlock. Defaults to 0.1.
+        """
         super(Decoder, self).__init__()
 
         n_position = len_max_seq + 1
@@ -48,7 +62,7 @@ class Decoder(dg.Layer):
                 d_inner,
                 n_head,
                 d_k,
-                d_v,
+                d_q,
                 fft_conv1d_kernel,
                 fft_conv1d_padding,
                 dropout=dropout) for _ in range(n_layers)
@@ -58,16 +72,20 @@ class Decoder(dg.Layer):
 
     def forward(self, enc_seq, enc_pos, non_pad_mask, slf_attn_mask=None):
         """
-        Decoder layer of FastSpeech.
+        Compute decoder outputs.
         
         Args:
-            enc_seq (Variable), Shape(B, text_T, C), dtype: float32. 
-                The output of length regulator.
-            enc_pos (Variable, optional): Shape(B, T_mel),
-                dtype: int64. The spectrum position. T_mel means the timesteps of input spectrum.
+            enc_seq (Variable): shape(B, T_text, C), dtype float32,
+                the output of length regulator, where T_text means the timesteps of input text, 
+            enc_pos (Variable): shape(B, T_mel), dtype int64, 
+                the spectrum position, where T_mel means the timesteps of input spectrum, 
+            non_pad_mask (Variable): shape(B, T_mel, 1), dtype int64, the mask with non pad.
+            slf_attn_mask (Variable, optional): shape(B, T_mel, T_mel), dtype int64, 
+                the mask of mel spectrum. Defaults to None.
+
         Returns:
-            dec_output (Variable), Shape(B, mel_T, C), the decoder output.
-            dec_slf_attn_list (Variable), Shape(B, mel_T, mel_T), the decoder self attention list.
+            dec_output (Variable): shape(B, T_mel, C), the decoder output.
+            dec_slf_attn_list (list[Variable]): len(n_layers), the decoder self attention list.
         """
         dec_slf_attn_list = []
         slf_attn_mask = layers.expand(slf_attn_mask, [self.n_head, 1, 1])

parakeet/models/fastspeech/encoder.py

@@ -24,12 +24,27 @@ class Encoder(dg.Layer):
                  n_layers,
                  n_head,
                  d_k,
-                 d_v,
+                 d_q,
                  d_model,
                  d_inner,
                  fft_conv1d_kernel,
                  fft_conv1d_padding,
                  dropout=0.1):
+        """Encoder layer of FastSpeech.
+
+        Args:
+            n_src_vocab (int): the number of source vocabulary.
+            len_max_seq (int): the max mel len of sequence.
+            n_layers (int): the layers number of FFTBlock.
+            n_head (int): the head number of multihead attention.
+            d_k (int): the dim of key in multihead attention.
+            d_q (int): the dim of query in multihead attention.
+            d_model (int): the dim of hidden layer in multihead attention.
+            d_inner (int): the dim of hidden layer in ffn.
+            fft_conv1d_kernel (int): the conv kernel size in FFTBlock.
+            fft_conv1d_padding (int): the conv padding size in FFTBlock.
+            dropout (float, optional): dropout probability of FFTBlock. Defaults to 0.1.
+        """
         super(Encoder, self).__init__()
         n_position = len_max_seq + 1
         self.n_head = n_head
@@ -53,7 +68,7 @@ class Encoder(dg.Layer):
                 d_inner,
                 n_head,
                 d_k,
-                d_v,
+                d_q,
                 fft_conv1d_kernel,
                 fft_conv1d_padding,
                 dropout=dropout) for _ in range(n_layers)
@@ -63,18 +78,20 @@ class Encoder(dg.Layer):
 
     def forward(self, character, text_pos, non_pad_mask, slf_attn_mask=None):
         """
-        Encoder layer of FastSpeech.
-        
-        Args:
-            character (Variable): Shape(B, T_text), dtype: float32. The input text
-                characters. T_text means the timesteps of input characters.
-            text_pos (Variable): Shape(B, T_text), dtype: int64. The input text
-                position. T_text means the timesteps of input characters.
+        Encode text sequence.
 
+        Args:
+            character (Variable): shape(B, T_text), dtype float32, the input text characters, 
+                where T_text means the timesteps of input characters,
+            text_pos (Variable): shape(B, T_text), dtype int64, the input text position. 
+            non_pad_mask (Variable): shape(B, T_text, 1), dtype int64, the mask with non pad.
+            slf_attn_mask (Variable, optional): shape(B, T_text, T_text), dtype int64, 
+                the mask of input characters. Defaults to None.
+        
         Returns:
-            enc_output (Variable), Shape(B, text_T, C), the encoder output.
-            non_pad_mask (Variable), Shape(B, T_text, 1), the mask with non pad.
-            enc_slf_attn_list (list<Variable>), Len(n_layers), Shape(B * n_head, text_T, text_T), the encoder self attention list.
+            enc_output (Variable): shape(B, T_text, C), the encoder output. 
+            non_pad_mask (Variable): shape(B, T_text, 1), the mask with non pad.
+            enc_slf_attn_list (list[Variable]): len(n_layers), the encoder self attention list.
         """
         enc_slf_attn_list = []
         slf_attn_mask = layers.expand(slf_attn_mask, [self.n_head, 1, 1])

parakeet/models/fastspeech/fastspeech.py

@@ -25,7 +25,11 @@ from parakeet.models.fastspeech.decoder import Decoder
 
 class FastSpeech(dg.Layer):
     def __init__(self, cfg):
-        " FastSpeech"
+        """FastSpeech model.
+
+        Args:
+            cfg: the yaml configs used in FastSpeech model.
+        """
         super(FastSpeech, self).__init__()
 
         self.encoder = Encoder(
@@ -34,7 +38,7 @@ class FastSpeech(dg.Layer):
             n_layers=cfg['encoder_n_layer'],
             n_head=cfg['encoder_head'],
             d_k=cfg['fs_hidden_size'] // cfg['encoder_head'],
-            d_v=cfg['fs_hidden_size'] // cfg['encoder_head'],
+            d_q=cfg['fs_hidden_size'] // cfg['encoder_head'],
             d_model=cfg['fs_hidden_size'],
             d_inner=cfg['encoder_conv1d_filter_size'],
             fft_conv1d_kernel=cfg['fft_conv1d_filter'],
@@ -50,7 +54,7 @@ class FastSpeech(dg.Layer):
             n_layers=cfg['decoder_n_layer'],
             n_head=cfg['decoder_head'],
             d_k=cfg['fs_hidden_size'] // cfg['decoder_head'],
-            d_v=cfg['fs_hidden_size'] // cfg['decoder_head'],
+            d_q=cfg['fs_hidden_size'] // cfg['decoder_head'],
             d_model=cfg['fs_hidden_size'],
             d_inner=cfg['decoder_conv1d_filter_size'],
             fft_conv1d_kernel=cfg['fft_conv1d_filter'],
@@ -82,34 +86,37 @@ class FastSpeech(dg.Layer):
                 text_pos,
                 enc_non_pad_mask,
                 dec_non_pad_mask,
+                mel_pos=None,
                 enc_slf_attn_mask=None,
                 dec_slf_attn_mask=None,
-                mel_pos=None,
                 length_target=None,
                 alpha=1.0):
         """
-        FastSpeech model.
+        Compute mel output from text character.
         
         Args:
-            character (Variable): Shape(B, T_text), dtype: float32. The input text
-                characters. T_text means the timesteps of input characters.
-            text_pos (Variable): Shape(B, T_text), dtype: int64. The input text
-                position. T_text means the timesteps of input characters.
-            mel_pos (Variable, optional): Shape(B, T_mel),
-                dtype: int64. The spectrum position. T_mel means the timesteps of input spectrum.
-            length_target (Variable, optional): Shape(B, T_text),
-                dtype: int64. The duration of phoneme compute from pretrained transformerTTS.
-            alpha (Constant): 
-                dtype: float32. The hyperparameter to determine the length of the expanded sequence 
-                mel, thereby controlling the voice speed.
+            character (Variable): shape(B, T_text), dtype float32, the input text characters, 
+                where T_text means the timesteps of input characters, 
+            text_pos (Variable): shape(B, T_text), dtype int64, the input text position. 
+            mel_pos (Variable, optional): shape(B, T_mel), dtype int64, the spectrum position, 
+                where T_mel means the timesteps of input spectrum,  
+            enc_non_pad_mask (Variable): shape(B, T_text, 1), dtype int64, the mask with non pad.
+            dec_non_pad_mask (Variable): shape(B, T_mel, 1), dtype int64, the mask with non pad.
+            enc_slf_attn_mask (Variable, optional): shape(B, T_text, T_text), dtype int64, 
+                the mask of input characters. Defaults to None.
+            slf_attn_mask (Variable, optional): shape(B, T_mel, T_mel), dtype int64,
+                the mask of mel spectrum. Defaults to None.
+            length_target (Variable, optional): shape(B, T_text), dtype int64, 
+                the duration of phoneme compute from pretrained transformerTTS. Defaults to None. 
+            alpha (float32, optional): The hyperparameter to determine the length of the expanded sequence 
+                mel, thereby controlling the voice speed. Defaults to 1.0.
 
         Returns:
-            mel_output (Variable), Shape(B, mel_T, C), the mel output before postnet.
-            mel_output_postnet (Variable), Shape(B, mel_T, C), the mel output after postnet.
-            duration_predictor_output (Variable), Shape(B, text_T), the duration of phoneme compute 
-            with duration predictor.
-            enc_slf_attn_list (Variable), Shape(B, text_T, text_T), the encoder self attention list.
-            dec_slf_attn_list (Variable), Shape(B, mel_T, mel_T), the decoder self attention list.
+            mel_output (Variable): shape(B, T_mel, C), the mel output before postnet.
+            mel_output_postnet (Variable): shape(B, T_mel, C), the mel output after postnet.
+            duration_predictor_output (Variable): shape(B, T_text), the duration of phoneme compute with duration predictor. 
+            enc_slf_attn_list (List[Variable]): len(enc_n_layers), the encoder self attention list. 
+            dec_slf_attn_list (List[Variable]): len(dec_n_layers), the decoder self attention list.
         """
 
         encoder_output, enc_slf_attn_list = self.encoder(
@@ -118,7 +125,6 @@ class FastSpeech(dg.Layer):
             enc_non_pad_mask,
             slf_attn_mask=enc_slf_attn_mask)
         if fluid.framework._dygraph_tracer()._train_mode:
-
             length_regulator_output, duration_predictor_output = self.length_regulator(
                 encoder_output, target=length_target, alpha=alpha)
             decoder_output, dec_slf_attn_list = self.decoder(

parakeet/models/fastspeech/fft_block.py

@@ -26,15 +26,27 @@ class FFTBlock(dg.Layer):
                  d_inner,
                  n_head,
                  d_k,
-                 d_v,
+                 d_q,
                  filter_size,
                  padding,
                  dropout=0.2):
+        """Feed forward structure based on self-attention.
+
+        Args:
+            d_model (int): the dim of hidden layer in multihead attention.
+            d_inner (int): the dim of hidden layer in ffn.
+            n_head (int): the head number of multihead attention.
+            d_k (int): the dim of key in multihead attention.
+            d_q (int): the dim of query in multihead attention.
+            filter_size (int): the conv kernel size.
+            padding (int): the conv padding size.
+            dropout (float, optional): dropout probability. Defaults to 0.2.
+        """
         super(FFTBlock, self).__init__()
         self.slf_attn = MultiheadAttention(
             d_model,
             d_k,
-            d_v,
+            d_q,
             num_head=n_head,
             is_bias=True,
             dropout=dropout,
@@ -48,18 +60,18 @@ class FFTBlock(dg.Layer):
 
     def forward(self, enc_input, non_pad_mask, slf_attn_mask=None):
         """
-        Feed Forward Transformer block in FastSpeech.
+        Feed forward block of FastSpeech
         
         Args:
-            enc_input (Variable): Shape(B, T, C), dtype: float32. The embedding characters input. 
-                T means the timesteps of input.
-            non_pad_mask (Variable): Shape(B, T, 1), dtype: int64. The mask of sequence.
-            slf_attn_mask (Variable): Shape(B, len_q, len_k), dtype: int64. The mask of self attention. 
-                len_q means the sequence length of query, len_k means the sequence length of key.
-
+            enc_input (Variable): shape(B, T, C), dtype float32, the embedding characters input, 
+                where T means the timesteps of input.   
+            non_pad_mask (Variable): shape(B, T, 1), dtype int64, the mask of sequence.
+            slf_attn_mask (Variable, optional): shape(B, len_q, len_k), dtype int64, the mask of self attention,
+                where len_q means the sequence length of query and len_k means the sequence length of key. Defaults to None. 
+                     
         Returns:
-            output (Variable), Shape(B, T, C), the output after self-attention & ffn.
-            slf_attn (Variable), Shape(B * n_head, T, T), the self attention.
+            output (Variable): shape(B, T, C), the output after self-attention & ffn. 
+            slf_attn (Variable): shape(B * n_head, T, T), the self attention.
         """
         output, slf_attn = self.slf_attn(
             enc_input, enc_input, enc_input, mask=slf_attn_mask)

parakeet/models/fastspeech/length_regulator.py

@@ -22,6 +22,14 @@ from parakeet.modules.customized import Conv1D
 
 class LengthRegulator(dg.Layer):
     def __init__(self, input_size, out_channels, filter_size, dropout=0.1):
+        """Length Regulator block in FastSpeech.
+
+        Args:
+            input_size (int): the channel number of input.
+            out_channels (int): the output channel number.
+            filter_size (int): the filter size of duration predictor.
+            dropout (float, optional): dropout probability. Defaults to 0.1.
+        """
         super(LengthRegulator, self).__init__()
         self.duration_predictor = DurationPredictor(
             input_size=input_size,
@@ -66,18 +74,18 @@ class LengthRegulator(dg.Layer):
 
     def forward(self, x, alpha=1.0, target=None):
         """
-        Length Regulator block in FastSpeech.
+        Compute length of mel from encoder output use TransformerTTS attention
         
         Args:
-            x (Variable): Shape(B, T, C), dtype: float32. The encoder output.
-            alpha (Constant): dtype: float32. The hyperparameter to determine the length of 
-                the expanded sequence mel, thereby controlling the voice speed.
-            target (Variable): (Variable, optional): Shape(B, T_text),
-                dtype: int64. The duration of phoneme compute from pretrained transformerTTS.
+            x (Variable): shape(B, T, C), dtype float32, the encoder output.
+            alpha (float32, optional): the hyperparameter to determine the length of 
+                the expanded sequence mel, thereby controlling the voice speed. Defaults to 1.0.
+            target (Variable, optional): shape(B, T_text), dtype int64, the duration of phoneme compute from pretrained transformerTTS. 
+                Defaults to None. 
 
         Returns:
-            output (Variable), Shape(B, T, C), the output after exppand.
-            duration_predictor_output (Variable), Shape(B, T, C), the output of duration predictor.
+            output (Variable): shape(B, T, C), the output after exppand.
+            duration_predictor_output (Variable): shape(B, T, C), the output of duration predictor.
         """
         duration_predictor_output = self.duration_predictor(x)
         if fluid.framework._dygraph_tracer()._train_mode:
@@ -93,6 +101,14 @@ class LengthRegulator(dg.Layer):
 
 class DurationPredictor(dg.Layer):
     def __init__(self, input_size, out_channels, filter_size, dropout=0.1):
+        """Duration Predictor block in FastSpeech.
+
+        Args:
+            input_size (int): the channel number of input.
+            out_channels (int): the output channel number.
+            filter_size (int): the filter size.
+            dropout (float, optional): dropout probability. Defaults to 0.1.
+        """
         super(DurationPredictor, self).__init__()
         self.input_size = input_size
         self.out_channels = out_channels
@@ -135,12 +151,13 @@ class DurationPredictor(dg.Layer):
 
     def forward(self, encoder_output):
         """
-        Duration Predictor block in FastSpeech.
+        Predict the duration of each character.
         
         Args:
-            encoder_output (Variable): Shape(B, T, C), dtype: float32. The encoder output.
+            encoder_output (Variable): shape(B, T, C), dtype float32, the encoder output.
+        
         Returns:
-            out (Variable), Shape(B, T, C), the output of duration predictor.
+            out (Variable): shape(B, T, C), the output of duration predictor.
         """
         # encoder_output.shape(N, T, C)
         out = layers.transpose(encoder_output, [0, 2, 1])

parakeet/models/transformer_tts/cbhg.py

@@ -30,16 +30,19 @@ class CBHG(dg.Layer):
                  num_gru_layers=2,
                  max_pool_kernel_size=2,
                  is_post=False):
-        super(CBHG, self).__init__()
-        """
-        :param hidden_size: dimension of hidden unit
-        :param batch_size: batch size
-        :param K: # of convolution banks
-        :param projection_size: dimension of projection unit
-        :param num_gru_layers: # of layers of GRUcell
-        :param max_pool_kernel_size: max pooling kernel size
-        :param is_post: whether post processing or not
+        """CBHG Module
+
+        Args:
+            hidden_size (int): dimension of hidden unit.
+            batch_size (int): batch size of input.
+            K (int, optional): number of convolution banks. Defaults to 16.
+            projection_size (int, optional): dimension of projection unit. Defaults to 256.
+            num_gru_layers (int, optional): number of layers of GRUcell. Defaults to 2.
+            max_pool_kernel_size (int, optional): max pooling kernel size. Defaults to 2
+            is_post (bool, optional): whether post processing or not. Defaults to False.
         """
+        super(CBHG, self).__init__()
+
         self.hidden_size = hidden_size
         self.projection_size = projection_size
         self.conv_list = []
@@ -176,7 +179,15 @@ class CBHG(dg.Layer):
             return x
 
     def forward(self, input_):
-        # input_.shape = [N, C, T]
+        """
+        Convert linear spectrum to Mel spectrum.
+
+        Args:
+            input_ (Variable): shape(B, C, T), dtype float32, the sequentially input.  
+
+        Returns:
+            out (Variable): shape(B, C, T), the CBHG output.
+        """
 
         conv_list = []
         conv_input = input_
@@ -217,6 +228,12 @@ class CBHG(dg.Layer):
 
 class Highwaynet(dg.Layer):
     def __init__(self, num_units, num_layers=4):
+        """Highway network
+
+        Args:
+            num_units (int): dimension of hidden unit.
+            num_layers (int, optional): number of highway layers. Defaults to 4.
+        """
         super(Highwaynet, self).__init__()
         self.num_units = num_units
         self.num_layers = num_layers
@@ -249,6 +266,15 @@ class Highwaynet(dg.Layer):
             self.add_sublayer("gates_{}".format(i), gate)
 
     def forward(self, input_):
+        """
+        Compute result of Highway network.
+
+        Args:
+            input_(Variable): shape(B, T, C), dtype float32, the sequentially input.
+            
+        Returns:
+            out(Variable): the Highway output.
+        """
         out = input_
 
         for linear, gate in zip(self.linears, self.gates):

parakeet/models/transformer_tts/decoder.py

@@ -22,7 +22,15 @@ from parakeet.models.transformer_tts.post_convnet import PostConvNet
 
 
 class Decoder(dg.Layer):
-    def __init__(self, num_hidden, config, num_head=4):
+    def __init__(self, num_hidden, config, num_head=4, n_layers=3):
+        """Decoder layer of TransformerTTS.
+
+        Args:
+            num_hidden (int): the number of source vocabulary.
+            config: the yaml configs used in decoder.
+            n_layers (int, optional): the layers number of multihead attention. Defaults to 4.
+            num_head (int, optional): the head number of multihead attention. Defaults to 3.
+        """
         super(Decoder, self).__init__()
         self.num_hidden = num_hidden
         self.num_head = num_head
@@ -58,20 +66,20 @@ class Decoder(dg.Layer):
 
         self.selfattn_layers = [
             MultiheadAttention(num_hidden, num_hidden // num_head,
-                               num_hidden // num_head) for _ in range(3)
+                               num_hidden // num_head) for _ in range(n_layers)
         ]
         for i, layer in enumerate(self.selfattn_layers):
             self.add_sublayer("self_attn_{}".format(i), layer)
         self.attn_layers = [
             MultiheadAttention(num_hidden, num_hidden // num_head,
-                               num_hidden // num_head) for _ in range(3)
+                               num_hidden // num_head) for _ in range(n_layers)
         ]
         for i, layer in enumerate(self.attn_layers):
             self.add_sublayer("attn_{}".format(i), layer)
         self.ffns = [
             PositionwiseFeedForward(
                 num_hidden, num_hidden * num_head, filter_size=1)
-            for _ in range(3)
+            for _ in range(n_layers)
         ]
         for i, layer in enumerate(self.ffns):
             self.add_sublayer("ffns_{}".format(i), layer)
@@ -108,6 +116,28 @@ class Decoder(dg.Layer):
                 m_mask=None,
                 m_self_mask=None,
                 zero_mask=None):
+        """
+        Compute decoder outputs.
+        
+        Args:
+            key (Variable): shape(B, T_text, C), dtype float32, the input key of decoder,
+                where T_text means the timesteps of input text,
+            value (Variable): shape(B, T_text, C), dtype float32, the input value of decoder.
+            query (Variable): shape(B, T_mel, C), dtype float32, the input query of decoder,
+                where T_mel means the timesteps of input spectrum,
+            positional (Variable): shape(B, T_mel), dtype int64, the spectrum position. 
+            mask (Variable): shape(B, T_mel, T_mel), dtype int64, the mask of decoder self attention.
+            m_mask (Variable, optional): shape(B, T_mel, 1), dtype int64, the query mask of encoder-decoder attention. Defaults to None.
+            m_self_mask (Variable, optional): shape(B, T_mel, 1), dtype int64, the query mask of decoder self attention. Defaults to None.
+            zero_mask (Variable, optional): shape(B, T_mel, T_text), dtype int64, query mask of encoder-decoder attention. Defaults to None.
+                
+        Returns:
+            mel_out (Variable): shape(B, T_mel, C), the decoder output after mel linear projection.
+            out (Variable): shape(B, T_mel, C), the decoder output after post mel network.
+            stop_tokens (Variable): shape(B, T_mel, 1), the stop tokens of output.
+            attn_list (list[Variable]): len(n_layers), the encoder-decoder attention list.
+            selfattn_list (list[Variable]): len(n_layers), the decoder self attention list.
+        """
 
         # get decoder mask with triangular matrix
 
@@ -121,7 +151,7 @@ class Decoder(dg.Layer):
         else:
             m_mask, m_self_mask, zero_mask = None, None, None
 
-# Decoder pre-network
+        # Decoder pre-network
         query = self.decoder_prenet(query)
 
         # Centered position

parakeet/models/transformer_tts/encoder.py

@@ -20,7 +20,15 @@ from parakeet.models.transformer_tts.encoderprenet import EncoderPrenet
 
 
 class Encoder(dg.Layer):
-    def __init__(self, embedding_size, num_hidden, num_head=4):
+    def __init__(self, embedding_size, num_hidden, num_head=4, n_layers=3):
+        """Encoder layer of TransformerTTS.
+
+        Args:
+            embedding_size (int): the size of position embedding.
+            num_hidden (int): the size of hidden layer in network.
+            n_layers (int, optional): the layers number of multihead attention. Defaults to 4.
+            num_head (int, optional): the head number of multihead attention. Defaults to 3.
+        """
         super(Encoder, self).__init__()
         self.num_hidden = num_hidden
         self.num_head = num_head
@@ -42,7 +50,7 @@ class Encoder(dg.Layer):
             use_cudnn=True)
         self.layers = [
             MultiheadAttention(num_hidden, num_hidden // num_head,
-                               num_hidden // num_head) for _ in range(3)
+                               num_hidden // num_head) for _ in range(n_layers)
         ]
         for i, layer in enumerate(self.layers):
             self.add_sublayer("self_attn_{}".format(i), layer)
@@ -51,12 +59,26 @@ class Encoder(dg.Layer):
                 num_hidden,
                 num_hidden * num_head,
                 filter_size=1,
-                use_cudnn=True) for _ in range(3)
+                use_cudnn=True) for _ in range(n_layers)
         ]
         for i, layer in enumerate(self.ffns):
             self.add_sublayer("ffns_{}".format(i), layer)
 
     def forward(self, x, positional, mask=None, query_mask=None):
+        """
+        Encode text sequence.
+        
+        Args:
+            x (Variable): shape(B, T_text), dtype float32, the input character,
+                where T_text means the timesteps of input text,
+            positional (Variable): shape(B, T_text), dtype int64, the characters position. 
+            mask (Variable, optional): shape(B, T_text, T_text), dtype int64, the mask of encoder self attention. Defaults to None.
+            query_mask (Variable, optional): shape(B, T_text, 1), dtype int64, the query mask of encoder self attention. Defaults to None.
+                
+        Returns:
+            x (Variable): shape(B, T_text, C), the encoder output.
+            attentions (list[Variable]): len(n_layers), the encoder self attention list.
+        """
 
         if fluid.framework._dygraph_tracer()._train_mode:
             seq_len_key = x.shape[1]
@@ -66,12 +88,12 @@ class Encoder(dg.Layer):
         else:
             query_mask, mask = None, None
         # Encoder pre_network
-        x = self.encoder_prenet(x)  #(N,T,C)
+        x = self.encoder_prenet(x)
 
         # Get positional encoding
         positional = self.pos_emb(positional)
 
-        x = positional * self.alpha + x  #(N, T, C)
+        x = positional * self.alpha + x
 
         # Positional dropout
         x = layers.dropout(x, 0.1, dropout_implementation='upscale_in_train')

parakeet/models/transformer_tts/encoderprenet.py

@@ -22,6 +22,13 @@ import numpy as np
 
 class EncoderPrenet(dg.Layer):
     def __init__(self, embedding_size, num_hidden, use_cudnn=True):
+        """ Encoder prenet layer of TransformerTTS.
+
+        Args:
+            embedding_size (int): the size of embedding.
+            num_hidden (int): the size of hidden layer in network.
+            use_cudnn (bool, optional): use cudnn or not. Defaults to True.
+        """
         super(EncoderPrenet, self).__init__()
         self.embedding_size = embedding_size
         self.num_hidden = num_hidden
@@ -81,8 +88,17 @@ class EncoderPrenet(dg.Layer):
                 low=-k, high=k)))
 
     def forward(self, x):
+        """
+        Prepare encoder input.
+        
+        Args:
+            x (Variable): shape(B, T_text), dtype float32, the input character, where T_text means the timesteps of input text.
+                
+        Returns:
+            (Variable): shape(B, T_text, C), the encoder prenet output.
+        """
 
-        x = self.embedding(x)  #(batch_size, seq_len, embending_size)
+        x = self.embedding(x)
         x = layers.transpose(x, [0, 2, 1])
         for batch_norm, conv in zip(self.batch_norm_list, self.conv_list):
             x = layers.dropout(

parakeet/models/transformer_tts/post_convnet.py

@@ -29,6 +29,19 @@ class PostConvNet(dg.Layer):
                  use_cudnn=True,
                  dropout=0.1,
                  batchnorm_last=False):
+        """Decocder post conv net of TransformerTTS.
+
+        Args:
+            n_mels (int, optional): the number of mel bands when calculating mel spectrograms. Defaults to 80.
+            num_hidden (int, optional): the size of hidden layer in network. Defaults to 512.
+            filter_size (int, optional): the filter size of Conv.  Defaults to 5.
+            padding (int, optional): the padding size of Conv. Defaults to 0.
+            num_conv (int, optional): the num of Conv layers in network. Defaults to 5.
+            outputs_per_step (int, optional): the num of output frames per step . Defaults to 1.
+            use_cudnn (bool, optional): use cudnn in Conv or not. Defaults to True.
+            dropout (float, optional): dropout probability. Defaults to 0.1.
+            batchnorm_last (bool, optional): if batchnorm at last layer or not. Defaults to False.
+        """
         super(PostConvNet, self).__init__()
 
         self.dropout = dropout
@@ -93,12 +106,13 @@ class PostConvNet(dg.Layer):
 
     def forward(self, input):
         """
-        Post Conv Net.
+        Compute the mel spectrum.
         
         Args:
-            input (Variable): Shape(B, T, C), dtype: float32. The input value.
+            input (Variable): shape(B, T, C), dtype float32, the result of mel linear projection. 
+               
         Returns:
-            output (Variable), Shape(B, T, C), the result after postconvnet.
+           output (Variable): shape(B, T, C), the result after postconvnet.
         """
 
         input = layers.transpose(input, [0, 2, 1])

parakeet/models/transformer_tts/prenet.py

@@ -19,10 +19,13 @@ import paddle.fluid.layers as layers
 
 class PreNet(dg.Layer):
     def __init__(self, input_size, hidden_size, output_size, dropout_rate=0.2):
-        """
-        :param input_size: dimension of input
-        :param hidden_size: dimension of hidden unit
-        :param output_size: dimension of output
+        """Prenet before passing through the network.
+
+        Args:
+            input_size (int): the input channel size.
+            hidden_size (int): the size of hidden layer in network.
+            output_size (int): the output channel size.
+            dropout_rate (float, optional): dropout probability. Defaults to 0.2.
         """
         super(PreNet, self).__init__()
         self.input_size = input_size
@@ -49,19 +52,20 @@ class PreNet(dg.Layer):
 
     def forward(self, x):
         """
-        Pre Net before passing through the network.
+        Prepare network input.
         
         Args:
-            x (Variable): Shape(B, T, C), dtype: float32. The input value.
+            x (Variable): shape(B, T, C), dtype float32, the input value.
+                
         Returns:
-            x (Variable), Shape(B, T, C), the result after pernet.
+            output (Variable): shape(B, T, C), the result after pernet.
         """
         x = layers.dropout(
             layers.relu(self.linear1(x)),
             self.dropout_rate,
             dropout_implementation='upscale_in_train')
-        x = layers.dropout(
+        output = layers.dropout(
             layers.relu(self.linear2(x)),
             self.dropout_rate,
             dropout_implementation='upscale_in_train')
-        return x
+        return output

parakeet/models/transformer_tts/transformer_tts.py

@@ -19,6 +19,11 @@ from parakeet.models.transformer_tts.decoder import Decoder
 
 class TransformerTTS(dg.Layer):
     def __init__(self, config):
+        """TransformerTTS model.
+
+        Args:
+            config: the yaml configs used in TransformerTTS model.
+        """
         super(TransformerTTS, self).__init__()
         self.encoder = Encoder(config['embedding_size'], config['hidden_size'])
         self.decoder = Decoder(config['hidden_size'], config)
@@ -35,6 +40,31 @@ class TransformerTTS(dg.Layer):
                 enc_dec_mask=None,
                 dec_query_slf_mask=None,
                 dec_query_mask=None):
+        """
+        TransformerTTS network.
+        
+        Args:
+            characters (Variable): shape(B, T_text), dtype float32, the input character,
+                where T_text means the timesteps of input text,
+            mel_input (Variable): shape(B, T_mel, C), dtype float32, the input query of decoder,
+                where T_mel means the timesteps of input spectrum,
+            pos_text (Variable): shape(B, T_text), dtype int64, the characters position. 
+            dec_slf_mask (Variable): shape(B, T_mel), dtype int64, the spectrum position. 
+            mask (Variable): shape(B, T_mel, T_mel), dtype int64, the mask of decoder self attention.
+            enc_slf_mask (Variable, optional): shape(B, T_text, T_text), dtype int64, the mask of encoder self attention. Defaults to None.
+            enc_query_mask (Variable, optional): shape(B, T_text, 1), dtype int64, the query mask of encoder self attention. Defaults to None.
+            dec_query_mask (Variable, optional): shape(B, T_mel, 1), dtype int64, the query mask of encoder-decoder attention. Defaults to None.
+            dec_query_slf_mask (Variable, optional): shape(B, T_mel, 1), dtype int64, the query mask of decoder self attention. Defaults to None.
+            enc_dec_mask (Variable, optional): shape(B, T_mel, T_text), dtype int64, query mask of encoder-decoder attention. Defaults to None.
+                
+        Returns:
+            mel_output (Variable): shape(B, T_mel, C), the decoder output after mel linear projection.
+            postnet_output (Variable): shape(B, T_mel, C), the decoder output after post mel network.
+            stop_preds (Variable): shape(B, T_mel, 1), the stop tokens of output.
+            attn_probs (list[Variable]): len(n_layers), the encoder-decoder attention list.
+            attns_enc (list[Variable]): len(n_layers), the encoder self attention list.
+            attns_dec (list[Variable]): len(n_layers), the decoder self attention list.
+        """
         key, attns_enc = self.encoder(
             characters, pos_text, mask=enc_slf_mask, query_mask=enc_query_mask)
 
@@ -48,5 +78,3 @@ class TransformerTTS(dg.Layer):
             m_self_mask=dec_query_slf_mask,
             m_mask=dec_query_mask)
         return mel_output, postnet_output, attn_probs, stop_preds, attns_enc, attns_dec
-
-        return mel_output, postnet_output, attn_probs, stop_preds, attns_enc, attns_dec

parakeet/models/transformer_tts/vocoder.py

@@ -19,11 +19,13 @@ from parakeet.models.transformer_tts.cbhg import CBHG
 
 
 class Vocoder(dg.Layer):
-    """
-    CBHG Network (mel -> linear)
-    """
-
     def __init__(self, config, batch_size):
+        """CBHG Network (mel -> linear)
+
+        Args:
+            config: the yaml configs used in Vocoder model.
+            batch_size (int): the batch size of input.
+        """
         super(Vocoder, self).__init__()
         self.pre_proj = Conv1D(
             num_channels=config['audio']['num_mels'],
@@ -36,6 +38,15 @@ class Vocoder(dg.Layer):
             filter_size=1)
 
     def forward(self, mel):
+        """
+        Compute mel spectrum to linear spectrum.
+        
+        Args:
+            mel (Variable): shape(B, C, T), dtype float32, the input mel spectrum.
+                
+        Returns:
+            mag_pred (Variable): shape(B, T, C), the linear output.
+        """
         mel = layers.transpose(mel, [0, 2, 1])
         mel = self.pre_proj(mel)
         mel = self.cbhg(mel)

parakeet/modules/dynamic_gru.py

@@ -43,9 +43,10 @@ class DynamicGRU(dg.Layer):
         Dynamic GRU block.
         
         Args:
-            input (Variable): Shape(B, T, C), dtype: float32. The input value.
+            input (Variable): shape(B, T, C), dtype float32, the input value.
+                 
         Returns:
-            output (Variable), Shape(B, T, C), the result compute by GRU.
+            output (Variable): shape(B, T, C), the result compute by GRU.
         """
         hidden = self.h_0
         res = []

parakeet/modules/ffn.py

@@ -19,8 +19,6 @@ from parakeet.modules.customized import Conv1D
 
 
 class PositionwiseFeedForward(dg.Layer):
-    ''' A two-feed-forward-layer module '''
-
     def __init__(self,
                  d_in,
                  num_hidden,
@@ -28,6 +26,16 @@ class PositionwiseFeedForward(dg.Layer):
                  padding=0,
                  use_cudnn=True,
                  dropout=0.1):
+        """A two-feed-forward-layer module.
+
+        Args:
+            d_in (int): the size of input channel.
+            num_hidden (int): the size of hidden layer in network.
+            filter_size (int): the filter size of Conv
+            padding (int, optional): the padding size of Conv. Defaults to 0.
+            use_cudnn (bool, optional): use cudnn in Conv or not. Defaults to True.
+            dropout (float, optional): dropout probability. Defaults to 0.1.
+        """
         super(PositionwiseFeedForward, self).__init__()
         self.num_hidden = num_hidden
         self.use_cudnn = use_cudnn
@@ -59,12 +67,13 @@ class PositionwiseFeedForward(dg.Layer):
 
     def forward(self, input):
         """
-        Feed Forward Network.
+        Compute feed forward network result.
         
         Args:
-            input (Variable): Shape(B, T, C), dtype: float32. The input value.
+            input (Variable): shape(B, T, C), dtype float32, the input value. 
+                
         Returns:
-            output (Variable), Shape(B, T, C), the result after FFN.
+            output (Variable): shape(B, T, C), the result after FFN. 
         """
         x = layers.transpose(input, [0, 2, 1])
         #FFN Networt

parakeet/modules/multihead_attention.py

@@ -50,6 +50,11 @@ class Linear(dg.Layer):
 
 class ScaledDotProductAttention(dg.Layer):
     def __init__(self, d_key):
+        """Scaled dot product attention module.
+
+        Args:
+            d_key (int): the dim of key in multihead attention.
+        """
         super(ScaledDotProductAttention, self).__init__()
 
         self.d_key = d_key
@@ -63,18 +68,18 @@ class ScaledDotProductAttention(dg.Layer):
                 query_mask=None,
                 dropout=0.1):
         """
-        Scaled Dot Product Attention.
+        Compute scaled dot product attention.
         
         Args:
-            key (Variable): Shape(B, T, C), dtype: float32. The input key of attention.
-            value (Variable): Shape(B, T, C), dtype: float32. The input value of attention.
-            query (Variable): Shape(B, T, C), dtype: float32. The input query of attention.
-            mask (Variable): Shape(B, len_q, len_k), dtype: float32. The mask of key.
-            query_mask (Variable): Shape(B, len_q, 1), dtype: float32. The mask of query.
-            dropout (Constant): dtype: float32. The probability of dropout.
+            key (Variable): shape(B, T, C), dtype float32, the input key of scaled dot product attention.
+            value (Variable): shape(B, T, C), dtype float32, the input value of scaled dot product attention.
+            query (Variable): shape(B, T, C), dtype float32, the input query of scaled dot product attention.
+            mask (Variable, optional): shape(B, T_q, T_k), dtype float32, the mask of key.  Defaults to None.
+            query_mask (Variable, optional): shape(B, T_q, T_q), dtype float32, the mask of query.  Defaults to None.
+            dropout (float32, optional): the probability of dropout. Defaults to 0.1.
         Returns:
-            result (Variable), Shape(B, T, C), the result of mutihead attention.
-            attention (Variable), Shape(n_head * B, T, C), the attention of key.
+            result (Variable): shape(B, T, C), the result of mutihead attention.
+            attention (Variable): shape(n_head * B, T, C), the attention of key.
         """
         # Compute attention score
         attention = layers.matmul(
@@ -105,6 +110,17 @@ class MultiheadAttention(dg.Layer):
                  is_bias=False,
                  dropout=0.1,
                  is_concat=True):
+        """Multihead Attention.
+
+        Args:
+            num_hidden (int): the number of hidden layer in network.
+            d_k (int): the dim of key in multihead attention.
+            d_q (int): the dim of query in multihead attention.
+            num_head (int, optional): the head number of multihead attention. Defaults to 4.
+            is_bias (bool, optional): whether have bias in linear layers. Default to False.
+            dropout (float, optional): dropout probability of FFTBlock. Defaults to 0.1.
+            is_concat (bool, optional): whether concat query and result. Default to True.
+        """
         super(MultiheadAttention, self).__init__()
         self.num_hidden = num_hidden
         self.num_head = num_head
@@ -128,17 +144,18 @@ class MultiheadAttention(dg.Layer):
 
     def forward(self, key, value, query_input, mask=None, query_mask=None):
         """
-        Multihead Attention.
+        Compute attention.
         
         Args:
-            key (Variable): Shape(B, T, C), dtype: float32. The input key of attention.
-            value (Variable): Shape(B, T, C), dtype: float32. The input value of attention.
-            query_input (Variable): Shape(B, T, C), dtype: float32. The input query of attention.
-            mask (Variable): Shape(B, len_q, len_k), dtype: float32. The mask of key.
-            query_mask (Variable): Shape(B, len_q, 1), dtype: float32. The mask of query.
+            key (Variable): shape(B, T, C), dtype float32, the input key of attention.
+            value (Variable): shape(B, T, C), dtype float32, the input value of attention.
+            query_input (Variable): shape(B, T, C), dtype float32, the input query of attention.
+            mask (Variable, optional): shape(B, T_query, T_key), dtype float32, the mask of key. Defaults to None.
+            query_mask (Variable, optional): shape(B, T_query, T_key), dtype float32, the mask of query. Defaults to None.
+                
         Returns:
-            result (Variable), Shape(B, T, C), the result of mutihead attention.
-            attention (Variable), Shape(n_head * B, T, C), the attention of key.
+            result (Variable): shape(B, T, C), the result of mutihead attention. 
+            attention (Variable): shape(num_head * B, T, C), the attention of key and query. 
         """
 
         batch_size = key.shape[0]
@@ -146,7 +163,6 @@ class MultiheadAttention(dg.Layer):
         seq_len_query = query_input.shape[1]
 
         # Make multihead attention
-        # key & value.shape = (batch_size, seq_len, feature)(feature = num_head * num_hidden_per_attn)
         key = layers.reshape(
             self.key(key), [batch_size, seq_len_key, self.num_head, self.d_k])
         value = layers.reshape(
@@ -168,18 +184,6 @@ class MultiheadAttention(dg.Layer):
         result, attention = self.scal_attn(
             key, value, query, mask=mask, query_mask=query_mask)
 
-        key = layers.reshape(
-            layers.transpose(key, [2, 0, 1, 3]), [-1, seq_len_key, self.d_k])
-        value = layers.reshape(
-            layers.transpose(value, [2, 0, 1, 3]),
-            [-1, seq_len_key, self.d_k])
-        query = layers.reshape(
-            layers.transpose(query, [2, 0, 1, 3]),
-            [-1, seq_len_query, self.d_q])
-
-        result, attention = self.scal_attn(
-            key, value, query, mask=mask, query_mask=query_mask)
-
         # concat all multihead result
         result = layers.reshape(
             result, [self.num_head, batch_size, seq_len_query, self.d_q])