Add files using upload-large-folder tool

Evaluations/.ipynb_checkpoints/split_generation-checkpoint.ipynb

@@ -0,0 +1,6 @@
+{
+ "cells": [],
+ "metadata": {},
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

data/openslr37/dataset/dataset_dict.json

@@ -0,0 +1 @@
+{"splits": ["train", "validation"]}

data/openslr37/dataset/train/dataset_info.json

@@ -0,0 +1,16 @@
+{
+  "citation": "",
+  "description": "",
+  "features": {
+    "audio": {
+      "dtype": "large_string",
+      "_type": "Value"
+    },
+    "text": {
+      "dtype": "large_string",
+      "_type": "Value"
+    }
+  },
+  "homepage": "",
+  "license": ""
+}

data/openslr37/dataset/train/state.json

@@ -0,0 +1,13 @@
+{
+  "_data_files": [
+    {
+      "filename": "data-00000-of-00001.arrow"
+    }
+  ],
+  "_fingerprint": "12671127c760c362",
+  "_format_columns": null,
+  "_format_kwargs": {},
+  "_format_type": null,
+  "_output_all_columns": false,
+  "_split": null
+}

data/openslr37/dataset/validation/dataset_info.json

@@ -0,0 +1,16 @@
+{
+  "citation": "",
+  "description": "",
+  "features": {
+    "audio": {
+      "dtype": "large_string",
+      "_type": "Value"
+    },
+    "text": {
+      "dtype": "large_string",
+      "_type": "Value"
+    }
+  },
+  "homepage": "",
+  "license": ""
+}

data/openslr37/dataset/validation/state.json

@@ -0,0 +1,13 @@
+{
+  "_data_files": [
+    {
+      "filename": "data-00000-of-00001.arrow"
+    }
+  ],
+  "_fingerprint": "767706cfdfc011bc",
+  "_format_columns": null,
+  "_format_kwargs": {},
+  "_format_type": null,
+  "_output_all_columns": false,
+  "_split": null
+}

data/openslr37/openslr37/openslr37/.ipynb_checkpoints/train_37-checkpoint.yaml

@@ -0,0 +1,29 @@
+pretrained_path: "/workspace/models/menstrual_bangla_tts"  # Path to the directory containing your pre-trained model
+train_manifest: "/workspace/data/openslr37/openslr37/openslr37/metadata.jsonl"  # Path to your training dataset JSONL file (already created)
+val_manifest: ""  # If you don't have a validation dataset, you can leave it empty or set it to ""
+sample_rate: 16000  # Ensure this matches the sample rate of your dataset (OpenSLR37)
+batch_size: 8  # Adjust based on your available hardware (increase if you have enough GPU memory)
+grad_accum_steps: 2  # Accumulate gradients for stability (adjust if necessary)
+num_workers: 6  # Number of workers for data loading (adjust based on your system)
+num_iters: 75000  # Number of training iterations
+log_interval: 50  # Log training stats every 50 steps
+valid_interval: 4000  # Validation interval (every 4000 steps)
+save_interval: 2000  # Save checkpoints every 2000 steps
+learning_rate: 0.0001  # Learning rate (tune if necessary)
+weight_decay: 0.01  # Regularization (tune if necessary)
+warmup_steps: 2000  # Warmup steps (set according to your needs)
+max_steps: 75000  # Maximum training steps (adjust to match num_iters)
+max_batch_tokens: 8192  # Maximum number of tokens per batch (adjust if needed)
+save_path: "/workspace/models/menstrual_bangla_tts/checkpoints"  # Path to save the fine-tuned model checkpoints
+tensorboard: "/workspace/models/menstrual_bangla_tts/tensorboard"  # Path to save tensorboard logs
+
+# LoRA config (optional)
+lora:
+  enable_lm: true  # Enable LoRA for the language model
+  enable_dit: true  # Enable LoRA for the DIT (Disentangled Intermediate Transformer)
+  enable_proj: true  # Enable LoRA for the projection layers
+  r: 64  # LoRA rank (adjust if necessary)
+  alpha: 32  # LoRA alpha (adjust if necessary)
+  dropout: 0.0  # Dropout rate (set to 0 for no dropout)
+  target_modules_lm: ["q_proj", "v_proj", "k_proj", "o_proj"]  # Target modules for LoRA in the LM
+  target_modules_dit: ["q_proj", "v_proj", "k_proj", "o_proj"]  # Target modules for LoRA in the DIT

data/openslr37/openslr37/openslr37/train_37.yaml

@@ -0,0 +1,29 @@
+pretrained_path: "/workspace/models/menstrual_bangla_tts"  # Path to the directory containing your pre-trained model
+train_manifest: "/workspace/data/openslr37/openslr37/openslr37/metadata.jsonl"  # Path to your training dataset JSONL file (already created)
+val_manifest: ""  # If you don't have a validation dataset, you can leave it empty or set it to ""
+sample_rate: 16000  # Ensure this matches the sample rate of your dataset (OpenSLR37)
+batch_size: 8  # Adjust based on your available hardware (increase if you have enough GPU memory)
+grad_accum_steps: 2  # Accumulate gradients for stability (adjust if necessary)
+num_workers: 6  # Number of workers for data loading (adjust based on your system)
+num_iters: 75000  # Number of training iterations
+log_interval: 50  # Log training stats every 50 steps
+valid_interval: 4000  # Validation interval (every 4000 steps)
+save_interval: 2000  # Save checkpoints every 2000 steps
+learning_rate: 0.0001  # Learning rate (tune if necessary)
+weight_decay: 0.01  # Regularization (tune if necessary)
+warmup_steps: 2000  # Warmup steps (set according to your needs)
+max_steps: 75000  # Maximum training steps (adjust to match num_iters)
+max_batch_tokens: 8192  # Maximum number of tokens per batch (adjust if needed)
+save_path: "/workspace/models/menstrual_bangla_tts/checkpoints"  # Path to save the fine-tuned model checkpoints
+tensorboard: "/workspace/models/menstrual_bangla_tts/tensorboard"  # Path to save tensorboard logs
+
+# LoRA config (optional)
+lora:
+  enable_lm: true  # Enable LoRA for the language model
+  enable_dit: true  # Enable LoRA for the DIT (Disentangled Intermediate Transformer)
+  enable_proj: true  # Enable LoRA for the projection layers
+  r: 64  # LoRA rank (adjust if necessary)
+  alpha: 32  # LoRA alpha (adjust if necessary)
+  dropout: 0.0  # Dropout rate (set to 0 for no dropout)
+  target_modules_lm: ["q_proj", "v_proj", "k_proj", "o_proj"]  # Target modules for LoRA in the LM
+  target_modules_dit: ["q_proj", "v_proj", "k_proj", "o_proj"]  # Target modules for LoRA in the DIT