REBASED: Transform Speaker Encoder in a Generic Encoder and Implement Emotion Encoder training support (#1349)

* Rename Speaker encoder module to encoder

* Add a generic emotion dataset formatter

* Transform the Speaker Encoder dataset to a generic dataset and create emotion encoder config

* Add class map in emotion config

* Add Base encoder config

* Add evaluation encoder script

* Fix the bug in plot_embeddings

* Enable Weight decay for encoder training

* Add argumnet to disable storage

* Add Perfect Sampler and remove storage

* Add evaluation during encoder training

* Fix lint checks

* Remove useless config parameter

* Active evaluation in speaker encoder test and use multispeaker dataset for this test

* Unit tests fixs

* Remove useless tests for speedup the aux_tests

* Use get_optimizer in Encoder

* Add BaseEncoder Class

* Fix the unitests

* Add Perfect Batch Sampler unit test

* Add compute encoder accuracy in a function

tests/inputs/test_glow_tts.json

@@ -66,8 +66,8 @@
     "use_mas": false,       // use Monotonic Alignment Search if true. Otherwise use pre-computed attention alignments.
 
     // TRAINING
-    "batch_size": 2,       // Batch size for training. Lower values than 32 might cause hard to learn attention. It is overwritten by 'gradual_training'.
-    "eval_batch_size":1,
+    "batch_size": 8,       // Batch size for training. Lower values than 32 might cause hard to learn attention. It is overwritten by 'gradual_training'.
+    "eval_batch_size": 8,
     "r": 1,                 // Number of decoder frames to predict per iteration. Set the initial values if gradual training is enabled.
     "loss_masking": true,   // enable / disable loss masking against the sequence padding.
     "data_dep_init_iter": 1,

tests/inputs/test_speaker_encoder_config.json

@@ -36,8 +36,8 @@
     "warmup_steps": 4000, // Noam decay steps to increase the learning rate from 0 to "lr"
     "tb_model_param_stats": false, // true, plots param stats per layer on tensorboard. Might be memory consuming, but good for debugging.
     "steps_plot_stats": 10, // number of steps to plot embeddings.
-    "num_speakers_in_batch": 64, // Batch size for training. Lower values than 32 might cause hard to learn attention. It is overwritten by 'gradual_training'.
-    "num_utters_per_speaker": 10,  //
+    "num_classes_in_batch": 64, // Batch size for training. Lower values than 32 might cause hard to learn attention. It is overwritten by 'gradual_training'.
+    "num_utter_per_class": 10,  //
     "num_loader_workers": 8,        // number of training data loader processes. Don't set it too big. 4-8 are good values.
     "wd": 0.000001, // Weight decay weight.
     "checkpoint": true, // If true, it saves checkpoints per "save_step"

tests/inputs/test_tacotron2_config.json

@@ -61,8 +61,8 @@
     "reinit_layers": [],    // give a list of layer names to restore from the given checkpoint. If not defined, it reloads all heuristically matching layers.
 
     // TRAINING
-    "batch_size": 1,       // Batch size for training. Lower values than 32 might cause hard to learn attention. It is overwritten by 'gradual_training'.
-    "eval_batch_size":1,
+    "batch_size": 8,       // Batch size for training. Lower values than 32 might cause hard to learn attention. It is overwritten by 'gradual_training'.
+    "eval_batch_size": 8,
     "r": 7,                 // Number of decoder frames to predict per iteration. Set the initial values if gradual training is enabled.
     "gradual_training": [[0, 7, 4], [1, 5, 2]], //set gradual training steps [first_step, r, batch_size]. If it is null, gradual training is disabled. For Tacotron, you might need to reduce the 'batch_size' as you proceeed.
     "loss_masking": true,         // enable / disable loss masking against the sequence padding.

tests/inputs/test_tacotron_config.json

@@ -61,8 +61,8 @@
     "reinit_layers": [],    // give a list of layer names to restore from the given checkpoint. If not defined, it reloads all heuristically matching layers.
 
     // TRAINING
-    "batch_size": 1,       // Batch size for training. Lower values than 32 might cause hard to learn attention. It is overwritten by 'gradual_training'.
-    "eval_batch_size":1,
+    "batch_size": 8,       // Batch size for training. Lower values than 32 might cause hard to learn attention. It is overwritten by 'gradual_training'.
+    "eval_batch_size": 8,
     "r": 7,                 // Number of decoder frames to predict per iteration. Set the initial values if gradual training is enabled.
     "gradual_training": [[0, 7, 4], [1, 5, 2]], //set gradual training steps [first_step, r, batch_size]. If it is null, gradual training is disabled. For Tacotron, you might need to reduce the 'batch_size' as you proceeed.
     "loss_masking": true,         // enable / disable loss masking against the sequence padding.