Merge pull request #1947 from coqui-ai/dataset-split

Automatic dataset split/alphabet generation

training/coqui_stt_training/evaluate.py

@@ -27,8 +27,6 @@ from .util.evaluate_tools import calculate_and_print_report, save_samples_json
 from .util.feeding import create_dataset
 from .util.helpers import check_ctcdecoder_version
 
-check_ctcdecoder_version()
-
 
 def sparse_tensor_value_to_texts(value, alphabet):
     r"""
@@ -179,6 +177,7 @@ def test():
 
 def main():
     initialize_globals_from_cli()
+    check_ctcdecoder_version()
 
     if not Config.test_files:
         raise RuntimeError(

training/coqui_stt_training/train.py

@@ -15,6 +15,7 @@ import json
 import shutil
 import time
 from datetime import datetime
+from pathlib import Path
 
 import numpy as np
 import progressbar
@@ -64,7 +65,6 @@ from .util.io import (
     remove_remote,
 )
 
-check_ctcdecoder_version()
 
 # Accuracy and Loss
 # =================
@@ -240,55 +240,40 @@ def average_gradients(tower_gradients):
     return average_grads
 
 
-# Logging
+def early_training_checks():
-# =======
+    check_ctcdecoder_version()
+
+    # Check for proper scorer early
+    if Config.scorer_path:
+        scorer = Scorer(
+            Config.lm_alpha, Config.lm_beta, Config.scorer_path, Config.alphabet
+        )
+        del scorer
+
+    if (
+        Config.train_files
+        and Config.test_files
+        and Config.load_checkpoint_dir != Config.save_checkpoint_dir
+    ):
+        log_warn(
+            "WARNING: You specified different values for --load_checkpoint_dir "
+            "and --save_checkpoint_dir, but you are running training and testing "
+            "in a single invocation. The testing step will respect --load_checkpoint_dir, "
+            "and thus WILL NOT TEST THE CHECKPOINT CREATED BY THE TRAINING STEP. "
+            "Train and test in two separate invocations, specifying the correct "
+            "--load_checkpoint_dir in both cases, or use the same location "
+            "for loading and saving."
+        )
 
 
-def log_variable(variable, gradient=None):
+def create_training_datasets(
-    r"""
+    exception_box,
-    We introduce a function for logging a tensor variable's current state.
+) -> (tf.data.Dataset, [tf.data.Dataset], [tf.data.Dataset],):
-    It logs scalar values for the mean, standard deviation, minimum and maximum.
+    """Creates training datasets from input flags.
-    Furthermore it logs a histogram of its state and (if given) of an optimization gradient.
+
+    Returns a single training dataset and two lists of datasets for validation
+    and metrics tracking.
     """
-    name = variable.name.replace(":", "_")
-    mean = tf.reduce_mean(input_tensor=variable)
-    tfv1.summary.scalar(name="%s/mean" % name, tensor=mean)
-    tfv1.summary.scalar(
-        name="%s/sttdev" % name,
-        tensor=tf.sqrt(tf.reduce_mean(input_tensor=tf.square(variable - mean))),
-    )
-    tfv1.summary.scalar(
-        name="%s/max" % name, tensor=tf.reduce_max(input_tensor=variable)
-    )
-    tfv1.summary.scalar(
-        name="%s/min" % name, tensor=tf.reduce_min(input_tensor=variable)
-    )
-    tfv1.summary.histogram(name=name, values=variable)
-    if gradient is not None:
-        if isinstance(gradient, tf.IndexedSlices):
-            grad_values = gradient.values
-        else:
-            grad_values = gradient
-        if grad_values is not None:
-            tfv1.summary.histogram(name="%s/gradients" % name, values=grad_values)
-
-
-def log_grads_and_vars(grads_and_vars):
-    r"""
-    Let's also introduce a helper function for logging collections of gradient/variable tuples.
-    """
-    for gradient, variable in grads_and_vars:
-        log_variable(variable, gradient=gradient)
-
-
-def train():
-    early_training_checks()
-
-    tfv1.reset_default_graph()
-    tfv1.set_random_seed(Config.random_seed)
-
-    exception_box = ExceptionBox()
-
     # Create training and validation datasets
     train_set = create_dataset(
         Config.train_files,
@@ -304,17 +289,8 @@ def train():
         buffering=Config.read_buffer,
     )
 
-    iterator = tfv1.data.Iterator.from_structure(
+    dev_sets = []
-        tfv1.data.get_output_types(train_set),
-        tfv1.data.get_output_shapes(train_set),
-        output_classes=tfv1.data.get_output_classes(train_set),
-    )
-
-    # Make initialization ops for switching between the two sets
-    train_init_op = iterator.make_initializer(train_set)
-
     if Config.dev_files:
-        dev_sources = Config.dev_files
         dev_sets = [
             create_dataset(
                 [source],
@@ -327,12 +303,11 @@ def train():
                 limit=Config.limit_dev,
                 buffering=Config.read_buffer,
             )
-            for source in dev_sources
+            for source in Config.dev_files
         ]
-        dev_init_ops = [iterator.make_initializer(dev_set) for dev_set in dev_sets]
 
+    metrics_sets = []
     if Config.metrics_files:
-        metrics_sources = Config.metrics_files
         metrics_sets = [
             create_dataset(
                 [source],
@@ -345,8 +320,31 @@ def train():
                 limit=Config.limit_dev,
                 buffering=Config.read_buffer,
             )
-            for source in metrics_sources
+            for source in Config.metrics_files
         ]
+
+    return train_set, dev_sets, metrics_sets
+
+
+def train():
+    early_training_checks()
+
+    tfv1.reset_default_graph()
+    tfv1.set_random_seed(Config.random_seed)
+
+    exception_box = ExceptionBox()
+
+    train_set, dev_sets, metrics_sets = create_training_datasets(exception_box)
+
+    iterator = tfv1.data.Iterator.from_structure(
+        tfv1.data.get_output_types(train_set),
+        tfv1.data.get_output_shapes(train_set),
+        output_classes=tfv1.data.get_output_classes(train_set),
+    )
+
+    # Make initialization ops for switching between the two sets
+    train_init_op = iterator.make_initializer(train_set)
+    dev_init_ops = [iterator.make_initializer(dev_set) for dev_set in dev_sets]
     metrics_init_ops = [
         iterator.make_initializer(metrics_set) for metrics_set in metrics_sets
     ]
@@ -387,7 +385,6 @@ def train():
 
     # Average tower gradients across GPUs
     avg_tower_gradients = average_gradients(gradients)
-    log_grads_and_vars(avg_tower_gradients)
 
     # global_step is automagically incremented by the optimizer
     global_step = tfv1.train.get_or_create_global_step()
@@ -567,7 +564,7 @@ def train():
                     # Validation
                     dev_loss = 0.0
                     total_steps = 0
-                    for source, init_op in zip(dev_sources, dev_init_ops):
+                    for source, init_op in zip(Config.dev_files, dev_init_ops):
                         log_progress("Validating epoch %d on %s..." % (epoch, source))
                         set_loss, steps = run_set("dev", epoch, init_op, dataset=source)
                         dev_loss += set_loss * steps
@@ -647,7 +644,7 @@ def train():
 
                 if Config.metrics_files:
                     # Read only metrics, not affecting best validation loss tracking
-                    for source, init_op in zip(metrics_sources, metrics_init_ops):
+                    for source, init_op in zip(Config.metrics_files, metrics_init_ops):
                         log_progress("Metrics for epoch %d on %s..." % (epoch, source))
                         set_loss, _ = run_set("metrics", epoch, init_op, dataset=source)
                         log_progress(
@@ -665,30 +662,6 @@ def train():
     log_debug("Session closed.")
 
 
-def early_training_checks():
-    # Check for proper scorer early
-    if Config.scorer_path:
-        scorer = Scorer(
-            Config.lm_alpha, Config.lm_beta, Config.scorer_path, Config.alphabet
-        )
-        del scorer
-
-    if (
-        Config.train_files
-        and Config.test_files
-        and Config.load_checkpoint_dir != Config.save_checkpoint_dir
-    ):
-        log_warn(
-            "WARNING: You specified different values for --load_checkpoint_dir "
-            "and --save_checkpoint_dir, but you are running training and testing "
-            "in a single invocation. The testing step will respect --load_checkpoint_dir, "
-            "and thus WILL NOT TEST THE CHECKPOINT CREATED BY THE TRAINING STEP. "
-            "Train and test in two separate invocations, specifying the correct "
-            "--load_checkpoint_dir in both cases, or use the same location "
-            "for loading and saving."
-        )
-
-
 def main():
     initialize_globals_from_cli()
 

training/coqui_stt_training/util/auto_input.py

@@ -0,0 +1,194 @@
+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+from pathlib import Path
+from typing import Optional
+
+import pandas
+from tqdm import tqdm
+
+from .io import open_remote
+from .sample_collections import samples_from_sources
+from coqui_stt_ctcdecoder import Alphabet
+
+
+def create_alphabet_from_sources(sources: [str]) -> ([str], Alphabet):
+    """Generate an Alphabet from characters in given sources.
+
+    sources: List of paths to input sources (CSV, SDB).
+
+    Returns a 2-tuple with list of characters and Alphabet instance.
+    """
+    characters = set()
+    for sample in tqdm(samples_from_sources(sources)):
+        characters |= set(sample.transcript)
+    characters = list(sorted(characters))
+    alphabet = Alphabet()
+    alphabet.InitFromLabels(characters)
+    return characters, alphabet
+
+
+def _get_sample_size(population_size):
+    """calculates the sample size for a 99% confidence and 1% margin of error"""
+    margin_of_error = 0.01
+    fraction_picking = 0.50
+    z_score = 2.58  # Corresponds to confidence level 99%
+    numerator = (z_score ** 2 * fraction_picking * (1 - fraction_picking)) / (
+        margin_of_error ** 2
+    )
+    sample_size = 0
+    for train_size in range(population_size, 0, -1):
+        denominator = 1 + (z_score ** 2 * fraction_picking * (1 - fraction_picking)) / (
+            margin_of_error ** 2 * train_size
+        )
+        sample_size = int(numerator / denominator)
+        if 2 * sample_size + train_size <= population_size:
+            break
+    return sample_size
+
+
+def _split_sets(samples: pandas.DataFrame, sample_size):
+    """
+    randomply split the datasets into train, validation, and test sets where the size of the
+    validation and test sets are determined by the `get_sample_size` function.
+    """
+    samples = samples.sample(frac=1).reset_index(drop=True)
+
+    train_beg = 0
+    train_end = len(samples) - 2 * sample_size
+
+    dev_beg = train_end
+    dev_end = train_end + sample_size
+
+    test_beg = dev_end
+    test_end = len(samples)
+
+    return (
+        samples[train_beg:train_end],
+        samples[dev_beg:dev_end],
+        samples[test_beg:test_end],
+    )
+
+
+def create_datasets_from_auto_input(
+    auto_input_dataset: Path, alphabet_config_path: Optional[Path]
+) -> (Path, Path, Path, Path):
+    """Creates training datasets from --auto_input_dataset flag.
+
+    auto_input_dataset: Path to input CSV or folder containing CSV.
+
+    Returns paths to generated train set, dev set and test set, and the path
+    to the alphabet file, either generated from the data, existing alongside
+    data, or specified manually by the user.
+    """
+    if auto_input_dataset.is_dir():
+        auto_input_dir = auto_input_dataset
+        all_csvs = list(auto_input_dataset.glob("*.csv"))
+        if not all_csvs:
+            raise RuntimeError(
+                "--auto_input_dataset is a directory but no CSV file was found "
+                "inside of it. Either make sure a CSV file is in the directory "
+                "or specify the file it directly."
+            )
+
+        non_subsets = [f for f in all_csvs if f.stem not in ("train", "dev", "test")]
+        if len(non_subsets) == 1:
+            auto_input_csv = non_subsets[0]
+        elif len(non_subsets) > 1:
+            non_subsets_fmt = ", ".join(str(s) for s in non_subsets)
+            raise RuntimeError(
+                "--auto_input_dataset is a directory but there are multiple CSV "
+                f"files not matching a subset name (train/dev/test): {non_subsets_fmt}. "
+                "Either remove extraneous CSV files or specify the correct file "
+                "to use for dataset formatting directly instead of the directory."
+            )
+        # else (empty) -> fall through, sets already present and get picked up below
+    else:
+        auto_input_dir = auto_input_dataset.parent
+        auto_input_csv = auto_input_dataset
+
+    train_set_path = auto_input_dir / "train.csv"
+    dev_set_path = auto_input_dir / "dev.csv"
+    test_set_path = auto_input_dir / "test.csv"
+
+    if train_set_path.exists() != dev_set_path.exists() != test_set_path.exists():
+        raise RuntimeError(
+            "Specifying --auto_input_dataset with some generated files present "
+            "and some missing. Either all three sets (train.csv, dev.csv, test.csv) "
+            "should exist alongside {auto_input_csv} (in which case they will be used), "
+            "or none of those files should exist (in which case they will be generated.)"
+        )
+
+    print(f"I Processing --auto_input_dataset input: {auto_input_csv}...")
+    df = pandas.read_csv(auto_input_csv)
+
+    if set(df.columns) < set(("wav_filename", "wav_filesize", "transcript")):
+        raise RuntimeError(
+            "Missing columns in --auto_input_dataset CSV. STT training inputs "
+            "require wav_filename, wav_filesize, and transcript columns."
+        )
+
+    dev_test_size = _get_sample_size(len(df))
+    if dev_test_size == 0:
+        if len(df) >= 2:
+            dev_test_size = 1
+        else:
+            raise RuntimeError(
+                "--auto_input_dataset dataset is too small for automatic splitting "
+                "into sets. Specify a larger input dataset or split it manually."
+            )
+
+    data_characters = sorted(list(set("".join(df["transcript"].values))))
+    alphabet_alongside_data_path = auto_input_dir / "alphabet.txt"
+    if alphabet_config_path:
+        alphabet = Alphabet(str(alphabet_config_path))
+        if not alphabet.CanEncode("".join(data_characters)):
+            raise RuntimeError(
+                "--alphabet_config_path was specified alongside --auto_input_dataset, "
+                "but alphabet contents don't match dataset transcripts. Make sure the "
+                "alphabet covers all transcripts or leave --alphabet_config_path "
+                "unspecified so that one will be generated automatically."
+            )
+        print(f"I Using specified --alphabet_config_path: {alphabet_config_path}")
+        generated_alphabet_path = alphabet_config_path
+    elif alphabet_alongside_data_path.exists():
+        alphabet = Alphabet(str(alphabet_alongside_data_path))
+        if not alphabet.CanEncode("".join(data_characters)):
+            raise RuntimeError(
+                "alphabet.txt exists alongside --auto_input_dataset file, but "
+                "alphabet contents don't match dataset transcripts. Make sure the "
+                "alphabet covers all transcripts or remove alphabet.txt file "
+                "from the data folderso that one will be generated automatically."
+            )
+        generated_alphabet_path = alphabet_alongside_data_path
+        print(f"I Using existing alphabet file: {alphabet_alongside_data_path}")
+    else:
+        alphabet = Alphabet()
+        alphabet.InitFromLabels(data_characters)
+        generated_alphabet_path = auto_input_dir / "alphabet.txt"
+        print(
+            f"I Saved generated alphabet with characters ({data_characters}) into {generated_alphabet_path}"
+        )
+        with open_remote(str(generated_alphabet_path), "wb") as fout:
+            fout.write(alphabet.SerializeText())
+
+    # If splits don't already exist, generate and save them.
+    # We check above that all three splits either exist or don't exist together,
+    # so we can check a single one for existence here.
+    if not train_set_path.exists():
+        train_set, dev_set, test_set = _split_sets(df, dev_test_size)
+        print(f"I Generated train set size: {len(train_set)} samples.")
+        print(f"I Generated validation set size: {len(dev_set)} samples.")
+        print(f"I Generated test set size: {len(test_set)} samples.")
+
+        print(f"I Writing train set to {train_set_path}")
+        train_set.to_csv(train_set_path, index=False)
+
+        print(f"I Writing dev set to {dev_set_path}")
+        dev_set.to_csv(dev_set_path, index=False)
+
+        print(f"I Writing test set to {test_set_path}")
+        test_set.to_csv(test_set_path, index=False)
+    else:
+        print("I Generated splits found alongside --auto_input_dataset, using them.")
+
+    return train_set_path, dev_set_path, test_set_path, generated_alphabet_path

training/coqui_stt_training/util/config.py

@@ -3,6 +3,7 @@ from __future__ import absolute_import, division, print_function
 import os
 import sys
 from dataclasses import asdict, dataclass, field
+from pathlib import Path
 from typing import List
 
 import progressbar
@@ -11,13 +12,12 @@ from attrdict import AttrDict
 from coqpit import MISSING, Coqpit, check_argument
 from coqui_stt_ctcdecoder import Alphabet, UTF8Alphabet
 from xdg import BaseDirectory as xdg
-from tqdm import tqdm
 
 from .augmentations import NormalizeSampleRate, parse_augmentations
+from .auto_input import create_alphabet_from_sources, create_datasets_from_auto_input
 from .gpu import get_available_gpus
 from .helpers import parse_file_size
 from .io import path_exists_remote
-from .sample_collections import samples_from_sources
 
 
 class _ConfigSingleton:
@@ -129,6 +129,30 @@ class _SttConfig(Coqpit):
             self.save_checkpoint_dir, "alphabet.txt"
         )
 
+        if not (
+            bool(self.auto_input_dataset)
+            != (self.train_files or self.dev_files or self.test_files)
+        ):
+            raise RuntimeError(
+                "When using --auto_input_dataset, do not specify --train_files, "
+                "--dev_files, or --test_files."
+            )
+
+        if self.auto_input_dataset:
+            (
+                gen_train,
+                gen_dev,
+                gen_test,
+                gen_alphabet,
+            ) = create_datasets_from_auto_input(
+                Path(self.auto_input_dataset),
+                Path(self.alphabet_config_path) if self.alphabet_config_path else None,
+            )
+            self.train_files = [str(gen_train)]
+            self.dev_files = [str(gen_dev)]
+            self.test_files = [str(gen_test)]
+            self.alphabet_config_path = str(gen_alphabet)
+
         if self.bytes_output_mode and self.alphabet_config_path:
             raise RuntimeError(
                 "You cannot set --alphabet_config_path *and* --bytes_output_mode"
@@ -136,7 +160,7 @@ class _SttConfig(Coqpit):
         elif self.bytes_output_mode:
             self.alphabet = UTF8Alphabet()
         elif self.alphabet_config_path:
-            self.alphabet = Alphabet(os.path.abspath(self.alphabet_config_path))
+            self.alphabet = Alphabet(self.alphabet_config_path)
         elif os.path.exists(loaded_checkpoint_alphabet_file):
             print(
                 "I --alphabet_config_path not specified, but found an alphabet file "
@@ -145,6 +169,24 @@ class _SttConfig(Coqpit):
             )
             self.alphabet = Alphabet(loaded_checkpoint_alphabet_file)
         elif self.train_files and self.dev_files and self.test_files:
+            # If all subsets are in the same folder and there's an alphabet file
+            # alongside them, use it.
+            self.alphabet = None
+            sources = self.train_files + self.dev_files + self.test_files
+            parents = set(Path(p).parent for p in sources)
+            if len(parents) == 1:
+                possible_alphabet = list(parents)[0] / "alphabet.txt"
+                if possible_alphabet.exists():
+                    print(
+                        "I --alphabet_config_path not specified, but all input "
+                        "datasets are present and in the same folder (--train_files, "
+                        "--dev_files and --test_files), and an alphabet.txt file "
+                        f"was found alongside the sets ({possible_alphabet}). "
+                        "Will use this alphabet file for this run."
+                    )
+                    self.alphabet = Alphabet(str(possible_alphabet))
+
+            if not self.alphabet:
                 # Generate alphabet automatically from input dataset, but only if
                 # fully specified, to avoid confusion in case a missing set has extra
                 # characters.
@@ -154,17 +196,9 @@ class _SttConfig(Coqpit):
                     "will be generated automatically from the data and placed alongside "
                     f"the checkpoint ({saved_checkpoint_alphabet_file})."
                 )
-            characters = set()
+                characters, alphabet = create_alphabet_from_sources(sources)
-            for sample in tqdm(
-                samples_from_sources(
-                    self.train_files + self.dev_files + self.test_files
-                )
-            ):
-                characters |= set(sample.transcript)
-            characters = list(sorted(characters))
                 print(f"I Generated alphabet characters: {characters}.")
-            self.alphabet = Alphabet()
+                self.alphabet = alphabet
-            self.alphabet.InitFromLabels(characters)
         else:
             raise RuntimeError(
                 "Missing --alphabet_config_path flag. Couldn't find an alphabet file\n"
@@ -281,6 +315,12 @@ class _SttConfig(Coqpit):
             help="space-separated list of files specifying the datasets used for tracking of metrics (after validation step). Currently the only metric is the CTC loss but without affecting the tracking of best validation loss. Multiple files will get reported separately. If empty, metrics will not be computed."
         ),
     )
+    auto_input_dataset: str = field(
+        default="",
+        metadata=dict(
+            help="path to a single CSV file to use for training. Cannot be specified alongside --train_files, --dev_files, --test_files. Training/validation/testing subsets will be automatically generated from the input, alongside with an alphabet file, if not already present.",
+        ),
+    )
 
     read_buffer: str = field(
         default="1MB",