A default parameter configuration is available in default.yaml. The parameters are categorized after model components where e.g. asr.opt contains parameters relvant for the ASR optimizer, asr.mdl contains ASR model parameters and asr contains other parameters in general.
A couple of other hyperparameters that are related to preprocessing and more technial details of training are found in ./src/preprocess.py.
These parameters should be configured for each model component indvidually.
| Parameter | Description |
|---|---|
| opt.type | The optimizer to use for training, see PyTorch documentation for more information |
| opt.learning_rate | The initial learning rate to use for the optimizer. Depending on the optimizer used, this might not be applicable but is sufficient for e.g. ADAM or ADADELTA. |
| train_index * | Path to training index produced by ./src/preprocess.py. |
| valid_index * | Path to validation index |
| test_index * | Path to test index |
(*) Currently, the baseline ASR is the only model that supports testing. It is important that these indexes are correctly sorted since we use torch.pack_padded_sequences which speeds up training. The sequences to train on must be sorted in decreasing order of original length on the padded axis. So, for e.g. the baseline ASR the index must be sorted by unpadded_num_frames and `.src.preprocess.sort_index() can be used exactly for this.
In most cases the index should be sorted by unpadded_num_frames but if training is performed on padded text tokens, then sort the index by s_len using the same function.
Some additional parameters can be set for each model component seperately and have default values if not set.
| Parameter | Description | Default value |
|---|---|---|
| valid_step | Number of steps between validation measurements | 500 |
| save_step | Number of steps between saving the most current version of the model | 1000 |
| logging_step | Number of steps between verbose logging. Type of training determines what type of logging is output | 250 |
| train_batch_size | Training batch size | 32 |
| valid_batch_size | Validation batch size | 32 |
| test_batch_size | Test batch size | 1 |
| n_epohcs | Number of training epochs | 5 |
| Parameter | Description |
|---|---|
| mdl.encoder_state_size | The state size of each pBLSTM unit in the ASR encoder |
| mdl.mlp_out_size | Output dimensions of the Φ and Ψ attention projection networks |
| mdl.decoder_state_size | The state sie of each LSTM unit in the ASR decoder |
| mdl.tf_rate | Teacher forcing rate of the decoder |
| mdl.feature_dim | The feature dimension of the input. This should match with .src.preprocess.N_DIMS |
| decode_beam_size | Number of hypothesis to consider at each level of beam search |
| decode_lm_weight | The weight that determines the influence of the language model during decoding. See thesis for more information |
| wer_step | The number of steps between each measure of WER on the training set. |
The speech autoencoder contains a CNN encoder. Here we only consider CNNs with 3 layers of convolutional -> batch norm -> RELU -> max pool.
| Parameter | Description |
|---|---|
| mdl.kernel_sizes | The 2D kernel sizes of each convolutional layer. |
| mdl.num_filters | The number of filters in each convolutional layer |
| mdl.pool_kernel_sizes | The 2D pooling kernels in each pooling layer. We aim to pool over the entire feature in the last pooling layer so the last pool kernel size has to be chosen with that in mind. In the default configuration we choose [2000, 40] since it approximately the size of the largest Malromur utterance. |
| Parameter | Description |
|---|---|
| mdl.state_size | The state size of each BLSTM unit in the text encoder |
| mdl.emb_dim | The character embedding dimensionality |
| mdl.num_layers | Number of BLSTM layers in the encoder |
For adversarial training we configure the optimizers of both the discriminator, D_optimizer, and the generator, G_optimizer.
| Parameter | Description |
|---|---|
| mdl.hidden_dim | The hidden dimension of the simple discriminator. |
| Parameter | Description |
|---|---|
| mdl.hidden_size | The state size of the RNN used in the language model. This is also used as the output dimension of the character embeddings used in the language model. |
| mdl.tf_rate | The teacher forcing rate used during decoding. |
| chunk_size | The size of each training sample. |
| Parameter | Description |
|---|---|
| its | The number of iterations of the main training loop to perform |
This is a short description of the parameters that are found in ./src/preprocess.py.
| Parameter | Description |
|---|---|
| CHARS | The latin alphabet and digits |
| ICE_CHARS | Special Icelandic characters |
| SPECIAL_CHARS | Some tokens that are likely to affect pronounciation. |
| SOS_TKN | Appended to the start of each sentence |
| EOS_TKN | Appended to the end of each sentence |
| UNK_TKN | Tokens in the unprocessed text that are not covered by CHARS/ICE_CHARS/SPECIAL_CHARS are replaced with this token |
| N_JOBS | no. jobs to run in parallel when writing the index |
| N_DIMS | no. frequency coefficients in the spectrograms |
| WIN_SIZE | size of window in STFT |
| STRIDE | stride of the window |
| TEXT_XTSN | extension of token files (if applicable) |