Fix tests (#168)

Author: sjkoelle

analytics/tests/tests/models/test_tacotron2.py

@@ -20,11 +20,9 @@ def test_tacotron2_model(self):
         with open("analytics/tests/fixtures/ljtest/taco2_lj2lj.json") as f:
             config.update(json.load(f))
         hparams = HParams(**config)
-        if torch.cuda.is_available():
-            hparams.cudnn_enabled = True
+
+        hparams.cudnn_enabled = False
         model = Tacotron2(hparams)
-        if torch.cuda.is_available():
-            model.cuda()
 
         trainer = Tacotron2Trainer(hparams, rank=0, world_size=0)
         (
@@ -43,16 +41,16 @@ def test_tacotron2_model(self):
                 "speaker_ids",
                 "gst",
                 "mel_padded",
-                "output_lengths",
+                "mel_lengths",
             ]
         )
         model_output = model(
             input_text=model_input["text_int_padded"],
             input_lengths=model_input["input_lengths"],
             speaker_ids=model_input["speaker_ids"],
-            embedded_gst=model_input["gst"],
+            embedded_gst=model_input.get("gst", None),
             targets=model_input["mel_padded"],
-            output_lengths=model_input["output_lengths"],
+            output_lengths=model_input["mel_lengths"],
         )
 
         # 'mel_outputs', 'mel_outputs_postnet', 'gate_predicted', 'output_lengths', 'alignments'

analytics/tests/tests/test_data_loader.py

@@ -31,7 +31,7 @@ def test_batch_structure(self):
         collate_fn = Collate()
         dl = DataLoader(ds, 12, collate_fn=collate_fn)
         for i, batch in enumerate(dl):
-            assert len(batch) == 9
+            assert len(batch) == 6
 
     def test_batch_dimensions(self):
         ds = Data(
@@ -44,13 +44,13 @@ def test_batch_dimensions(self):
         collate_fn = Collate()
         dl = DataLoader(ds, 12, collate_fn=collate_fn)
         for i, batch in enumerate(dl):
-            output_lengths = batch["output_lengths"]
-            gate_target = batch["gate_target"]
+            output_lengths = batch["mel_lengths"]
+            gate_target = batch["gate_padded"]
             mel_padded = batch["mel_padded"]
             assert output_lengths.item() == 566
             assert gate_target.size(1) == 566
             assert mel_padded.size(2) == 566
-            assert len(batch) == 9
+            assert len(batch) == 6
 
     def test_batch_dimensions_partial(self):
         ds = Data(
@@ -63,9 +63,9 @@ def test_batch_dimensions_partial(self):
         collate_fn = Collate(n_frames_per_step=5)
         dl = DataLoader(ds, 12, collate_fn=collate_fn)
         for i, batch in enumerate(dl):
-            assert batch["output_lengths"].item() == 566
+            assert batch["mel_lengths"].item() == 566
             assert (
                 batch["mel_padded"].size(2) == 566
             )  # I'm not sure why this was 570 - maybe 566 + 5 (i.e. the n_frames_per_step)
-            assert batch["gate_target"].size(1) == 566
-            assert len(batch) == 9
+            assert batch["gate_padded"].size(1) == 566
+            assert len(batch) == 6

analytics/tests/tests/trainer/test_trainer.py

@@ -1,8 +1,10 @@
-from uberduck_ml_dev.trainer.base import TTSTrainer
 import torch
 import math
+
 from uberduck_ml_dev.vendor.tfcompat.hparam import HParams
 from uberduck_ml_dev.trainer.base import DEFAULTS as TRAINER_DEFAULTS
+from uberduck_ml_dev.trainer.base import TTSTrainer
+from uberduck_ml_dev.models.common import MelSTFT
 
 
 class TestTrainer:
@@ -22,7 +24,8 @@ def test_trainer_base(self):
 
         assert trainer.cudnn_enabled == True
         mel = torch.load("analytics/tests/fixtures/stevejobs-1.pt")
-        audio = trainer.sample(mel)
+        mel_stft = MelSTFT()
+        audio = mel_stft.griffin_lim(mel)
         assert audio.size(0) == 1
 
 

uberduck_ml_dev/data/data.py

@@ -51,9 +51,7 @@
 F0_MAX = 640
 
 
-# NOTE (Sam): generic dataset class for all purposes avoids writing redundant methods (e.g. get pitch when text isn't available).
-# However, functional factorization of this dataloader (e.g. get_mels) and merging classes as needed would be preferable.
-# NOTE (Sam): "load" means load from file. "return" means return to collate. "get" is a functional element.  "has" and "with" are tbd equivalent in trainer/model.
+# TODO (Sam): replace with Dataset.
 class Data(Dataset):
     def __init__(
         self,

uberduck_ml_dev/losses.py

@@ -19,7 +19,7 @@ def __init__(self, pos_weight):
 
     # NOTE (Sam): making function inputs explicit makes less sense in situations like this with obvious subcategories.
     def forward(self, model_output: Batch, target: Batch):
-        mel_target, gate_target = target["mel_padded"], target["gate_target"]
+        mel_target, gate_target = target["mel_padded"], target["gate_padded"]
         mel_target.requires_grad = False
         gate_target.requires_grad = False
         mel_out, mel_out_postnet, gate_out = (