Upload Kiritan.py

configs/Kiritan.py

@@ -0,0 +1,56 @@
+from pytorch_lightning.callbacks import LearningRateMonitor, ModelCheckpoint
+
+_base_ = [
+    "./_base_/archs/hifi_svc.py",
+    "./_base_/trainers/base.py",
+    "./_base_/schedulers/exponential.py",
+    "./_base_/datasets/hifi_svc.py",
+]
+
+speaker_mapping = {
+    "kiritan": 0,
+}
+
+model = dict(
+    type="HiFiSVC",
+    speaker_encoder=dict(
+        input_size=len(speaker_mapping),
+    ),
+)
+
+preprocessing = dict(
+    text_features_extractor=dict(
+        type="ContentVec",
+    ),
+    pitch_extractor=dict(
+        type="ParselMouthPitchExtractor",
+        keep_zeros=False,
+        f0_min=40.0,
+        f0_max=1600.0,
+    ),
+    energy_extractor=dict(
+        type="RMSEnergyExtractor",
+    ),
+    augmentations=[
+        dict(
+            type="FixedPitchShifting",
+            key_shifts=[-5.0, 5.0],
+            probability=0.75,
+        ),
+    ],
+)
+
+trainer = dict(
+    # Disable gradient clipping, which is not supported by custom optimization
+    gradient_clip_val=None,
+    val_check_interval=1000,
+    check_val_every_n_epoch=None,
+    callbacks=[
+        ModelCheckpoint(
+            filename="{epoch}-{step}-{valid_loss:.2f}",
+            every_n_train_steps=1000,
+            save_top_k=-1,
+        ),
+        LearningRateMonitor(logging_interval="step"),
+    ],
+)