Training in progress, step 6000

config.json

@@ -34,6 +34,7 @@
   "num_mel_bins": 80,
   "pad_token_id": 50257,
   "scale_embedding": false,
+  "suppress_tokens": [],
   "torch_dtype": "float32",
   "transformers_version": "4.26.0.dev0",
   "use_cache": false,

e5_interleaving-cl.ipynb

@@ -92,8 +92,8 @@
       "Building dependency tree       \n",
       "Reading state information... Done\n",
       "git-lfs is already the newest version (2.9.2-1).\n",
-      "0 upgraded, 0 newly installed, 0 to remove and 154 not upgraded.\n",
-      "Error: Failed to call git rev-parse --git-dir: exit status 128 \n",
+      "0 upgraded, 0 newly installed, 0 to remove and 155 not upgraded.\n",
+      "Updated git hooks.\n",
       "Git LFS initialized.\n"
      ]
     }
@@ -341,7 +341,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 3,
+   "execution_count": 2,
    "id": "d74b38c5-a1fb-4214-b4f4-b5bf0869f169",
    "metadata": {
     "colab": {
@@ -356,7 +356,7 @@
      "name": "stdout",
      "output_type": "stream",
      "text": [
-      "Sun Dec 11 12:49:44 2022       \n",
+      "Mon Dec 12 18:05:43 2022       \n",
       "+-----------------------------------------------------------------------------+\n",
       "| NVIDIA-SMI 515.65.01    Driver Version: 515.65.01    CUDA Version: 11.7     |\n",
       "|-------------------------------+----------------------+----------------------+\n",
@@ -365,7 +365,7 @@
       "|                               |                      |               MIG M. |\n",
       "|===============================+======================+======================|\n",
       "|   0  NVIDIA A100-SXM...  On   | 00000000:06:00.0 Off |                    0 |\n",
-      "| N/A   32C    P0    44W / 400W |      0MiB / 40960MiB |      0%      Default |\n",
+      "| N/A   74C    P0   350W / 400W |  36965MiB / 40960MiB |    100%      Default |\n",
       "|                               |                      |             Disabled |\n",
       "+-------------------------------+----------------------+----------------------+\n",
       "                                                                               \n",
@@ -374,7 +374,7 @@
       "|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |\n",
       "|        ID   ID                                                   Usage      |\n",
       "|=============================================================================|\n",
-      "|  No running processes found                                                 |\n",
+      "|    0   N/A  N/A   3651714      C   python                          36963MiB |\n",
       "+-----------------------------------------------------------------------------+\n"
      ]
     }
@@ -735,7 +735,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 6,
+   "execution_count": 1,
    "id": "dff27c76-575c-432b-8916-b1b810efef4a",
    "metadata": {
     "colab": {
@@ -768,7 +768,7 @@
     {
      "data": {
       "application/vnd.jupyter.widget-view+json": {
-       "model_id": "172290fa003f462aa40a9b03e56e91b1",
+       "model_id": "6a2bfa275b6f4cdba66b6abdab11e859",
        "version_major": 2,
        "version_minor": 0
       },
@@ -850,7 +850,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 7,
+   "execution_count": 1,
    "id": "065a8cf7-e54f-4ac3-900e-609c80714fca",
    "metadata": {
     "id": "065a8cf7-e54f-4ac3-900e-609c80714fca"
@@ -898,7 +898,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 8,
+   "execution_count": 2,
    "id": "a2787582-554f-44ce-9f38-4180a5ed6b44",
    "metadata": {
     "id": "a2787582-554f-44ce-9f38-4180a5ed6b44",
@@ -960,7 +960,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 9,
+   "execution_count": 3,
    "id": "qOwlctMhNmCG",
    "metadata": {
     "id": "qOwlctMhNmCG",
@@ -984,7 +984,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 10,
+   "execution_count": 4,
    "id": "imRHJOpm4V_j",
    "metadata": {
     "id": "imRHJOpm4V_j"
@@ -1031,7 +1031,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 11,
+   "execution_count": 5,
    "id": "77d9f0c5-8607-4642-a8ac-c3ab2e223ea6",
    "metadata": {
     "id": "77d9f0c5-8607-4642-a8ac-c3ab2e223ea6"
@@ -1066,7 +1066,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 12,
+   "execution_count": 6,
    "id": "ab5a13b4-9bd4-4aa0-aef2-b3de9b762988",
    "metadata": {
     "colab": {
@@ -1083,7 +1083,7 @@
        " 'sentence': Value(dtype='string', id=None)}"
       ]
      },
-     "execution_count": 12,
+     "execution_count": 6,
      "metadata": {},
      "output_type": "execute_result"
     }
@@ -1111,7 +1111,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 13,
+   "execution_count": 7,
    "id": "3ab6a724-3d1e-478b-a9e9-d2f85feb6c39",
    "metadata": {
     "id": "3ab6a724-3d1e-478b-a9e9-d2f85feb6c39"
@@ -1135,7 +1135,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 14,
+   "execution_count": 8,
    "id": "d041650e-1c48-4439-87b3-5b6f4a514107",
    "metadata": {
     "id": "d041650e-1c48-4439-87b3-5b6f4a514107"
@@ -1145,7 +1145,7 @@
     "import string\n",
     "import re\n",
     "\n",
-    "do_lower_case = True\n",
+    "do_lower_case = False\n",
     "do_remove_punctuation = False\n",
     "\n",
     "punctuation_to_remove = string.punctuation.replace(\"'\", \"\")  # don't remove apostrophes\n",
@@ -1171,7 +1171,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 15,
+   "execution_count": 9,
    "id": "c085911c-a10a-41ef-8874-306e0503e9bb",
    "metadata": {
     "id": "c085911c-a10a-41ef-8874-306e0503e9bb"
@@ -1212,7 +1212,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 16,
+   "execution_count": 10,
    "id": "a37a7cdb-9013-427f-8de9-6a8d0e9dc684",
    "metadata": {
     "id": "a37a7cdb-9013-427f-8de9-6a8d0e9dc684"
@@ -1234,7 +1234,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 17,
+   "execution_count": 11,
    "id": "1b145699-acfc-4b1d-93a2-a2ad3d62674c",
    "metadata": {
     "id": "1b145699-acfc-4b1d-93a2-a2ad3d62674c"
@@ -1259,7 +1259,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 18,
+   "execution_count": 12,
    "id": "01cb25ef-4bb0-4325-9461-f59198acadf6",
    "metadata": {
     "id": "01cb25ef-4bb0-4325-9461-f59198acadf6"
@@ -1284,7 +1284,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 19,
+   "execution_count": 13,
    "id": "333f7f6e-6053-4d3b-8924-c733c79b82ac",
    "metadata": {
     "id": "333f7f6e-6053-4d3b-8924-c733c79b82ac"
@@ -1364,7 +1364,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 20,
+   "execution_count": 14,
    "id": "8326221e-ec13-4731-bb4e-51e5fc1486c5",
    "metadata": {
     "id": "8326221e-ec13-4731-bb4e-51e5fc1486c5"
@@ -1416,7 +1416,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 21,
+   "execution_count": 15,
    "id": "fc834702-c0d3-4a96-b101-7b87be32bf42",
    "metadata": {
     "id": "fc834702-c0d3-4a96-b101-7b87be32bf42"
@@ -1449,7 +1449,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 22,
+   "execution_count": 16,
    "id": "b22b4011-f31f-4b57-b684-c52332f92890",
    "metadata": {
     "colab": {
@@ -1500,7 +1500,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 23,
+   "execution_count": 17,
    "id": "a11d1bfc-9e28-460f-a287-72d8f7bc1acb",
    "metadata": {
     "id": "a11d1bfc-9e28-460f-a287-72d8f7bc1acb"
@@ -1549,7 +1549,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 24,
+   "execution_count": 18,
    "id": "5a10cc4b-07ec-4ebd-ac1d-7c601023594f",
    "metadata": {
     "colab": {
@@ -1605,7 +1605,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 25,
+   "execution_count": 19,
    "id": "62038ba3-88ed-4fce-84db-338f50dcd04f",
    "metadata": {
     "id": "62038ba3-88ed-4fce-84db-338f50dcd04f"
@@ -1614,7 +1614,10 @@
    "source": [
     "model.config.forced_decoder_ids = None\n",
     "model.config.suppress_tokens = []\n",
-    "model.config.use_cache = False"
+    "model.config.use_cache = False\n",
+    "model.config.dropout=0.1\n",
+    "#model.config.dropout=0.05\n",
+    "#model.config.dropout = model.config.attention_dropout = 0.05"
    ]
   },
   {
@@ -1639,7 +1642,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 26,
+   "execution_count": 20,
    "id": "0ae3e9af-97b7-4aa0-ae85-20b23b5bcb3a",
    "metadata": {
     "colab": {
@@ -1659,8 +1662,11 @@
     "    gradient_accumulation_steps=1,  # increase by 2x for every 2x decrease in batch size\n",
     "    learning_rate=1e-5,\n",
     "    warmup_steps=500,\n",
-    "    max_steps=5000,\n",
+    "    max_steps=7000,\n",
     "    gradient_checkpointing=True,\n",
+    "    resume_from_checkpoint=5000,\n",
+    "    #resume_from_checkpoint=\"checkpoint-4000\"\n",
+    "    ignore_data_skip=True,\n",
     "    fp16=True,\n",
     "    evaluation_strategy=\"steps\",\n",
     "    per_device_eval_batch_size=16,\n",
@@ -1680,7 +1686,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 27,
+   "execution_count": 21,
    "id": "o72eOpGzD_sK",
    "metadata": {
     "colab": {
@@ -1694,7 +1700,7 @@
      "name": "stdout",
      "output_type": "stream",
      "text": [
-      "Sun Dec 11 12:50:47 2022       \n",
+      "Tue Dec 13 00:36:23 2022       \n",
       "+-----------------------------------------------------------------------------+\n",
       "| NVIDIA-SMI 515.65.01    Driver Version: 515.65.01    CUDA Version: 11.7     |\n",
       "|-------------------------------+----------------------+----------------------+\n",
@@ -1744,7 +1750,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 28,
+   "execution_count": 22,
    "id": "3ac16b62-b3c0-4c68-8f3d-9ecf471534b2",
    "metadata": {
     "id": "3ac16b62-b3c0-4c68-8f3d-9ecf471534b2"
@@ -1777,7 +1783,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 29,
+   "execution_count": 23,
    "id": "d546d7fe-0543-479a-b708-2ebabec19493",
    "metadata": {
     "colab": {
@@ -2355,7 +2361,7 @@
      "name": "stderr",
      "output_type": "stream",
      "text": [
-      "/home/ubuntu/./whisper-medium-el is already a clone of https://huggingface.co/emilios/whisper-medium-el. Make sure you pull the latest changes with `repo.git_pull()`.\n",
+      "/home/ubuntu/whisper-medium-el/./whisper-medium-el is already a clone of https://huggingface.co/emilios/whisper-medium-el. Make sure you pull the latest changes with `repo.git_pull()`.\n",
       "max_steps is given, it will override any value given in num_train_epochs\n",
       "Using cuda_amp half precision backend\n"
      ]
@@ -2391,7 +2397,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 30,
+   "execution_count": null,
    "id": "a1ccb9ed-cbc8-4419-91c0-651e9424b672",
    "metadata": {
     "id": "a1ccb9ed-cbc8-4419-91c0-651e9424b672"
@@ -2401,12 +2407,7 @@
      "name": "stderr",
      "output_type": "stream",
      "text": [
-      "Configuration saved in ./whisper-medium-el/config.json\n",
-      "Model weights saved in ./whisper-medium-el/pytorch_model.bin\n",
-      "Feature extractor saved in ./whisper-medium-el/preprocessor_config.json\n",
-      "tokenizer config file saved in ./whisper-medium-el/tokenizer_config.json\n",
-      "Special tokens file saved in ./whisper-medium-el/special_tokens_map.json\n",
-      "added tokens file saved in ./whisper-medium-el/added_tokens.json\n"
+      "Configuration saved in ./whisper-medium-el/config.json\n"
      ]
     }
    ],
@@ -2479,67 +2480,10 @@
    "metadata": {
     "id": "ee8b7b8e-1c9a-4d77-9137-1778a629e6de"
    },
-   "outputs": [
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "/home/ubuntu/.local/lib/python3.8/site-packages/transformers/optimization.py:306: FutureWarning: This implementation of AdamW is deprecated and will be removed in a future version. Use the PyTorch implementation torch.optim.AdamW instead, or set `no_deprecation_warning=True` to disable this warning\n",
-      "  warnings.warn(\n",
-      "***** Running training *****\n",
-      "  Num examples = 160000\n",
-      "  Num Epochs = 9223372036854775807\n",
-      "  Instantaneous batch size per device = 32\n",
-      "  Total train batch size (w. parallel, distributed & accumulation) = 32\n",
-      "  Gradient Accumulation steps = 1\n",
-      "  Total optimization steps = 5000\n",
-      "  Number of trainable parameters = 763857920\n",
-      "Reading metadata...: 1914it [00:00, 58899.17it/s]\n",
-      "The following columns in the training set don't have a corresponding argument in `WhisperForConditionalGeneration.forward` and have been ignored: input_length. If input_length are not expected by `WhisperForConditionalGeneration.forward`,  you can safely ignore this message.\n"
-     ]
-    },
-    {
-     "data": {
-      "text/html": [
-       "\n",
-       "    <div>\n",
-       "      \n",
-       "      <progress value='21' max='5000' style='width:300px; height:20px; vertical-align: middle;'></progress>\n",
-       "      [  21/5000 02:01 < 8:48:39, 0.16 it/s, Epoch 0.00/9223372036854775807]\n",
-       "    </div>\n",
-       "    <table border=\"1\" class=\"dataframe\">\n",
-       "  <thead>\n",
-       " <tr style=\"text-align: left;\">\n",
-       "      <th>Step</th>\n",
-       "      <th>Training Loss</th>\n",
-       "      <th>Validation Loss</th>\n",
-       "    </tr>\n",
-       "  </thead>\n",
-       "  <tbody>\n",
-       "  </tbody>\n",
-       "</table><p>"
-      ],
-      "text/plain": [
-       "<IPython.core.display.HTML object>"
-      ]
-     },
-     "metadata": {},
-     "output_type": "display_data"
-    },
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "***** Running Evaluation *****\n",
-      "  Num examples: Unknown\n",
-      "  Batch size = 16\n",
-      "Reading metadata...: 1696it [00:00, 51485.08it/s]\n",
-      "The following columns in the evaluation set don't have a corresponding argument in `WhisperForConditionalGeneration.forward` and have been ignored: path, accent, down_votes, age, gender, segment, up_votes, input_length, locale, client_id. If path, accent, down_votes, age, gender, segment, up_votes, input_length, locale, client_id are not expected by `WhisperForConditionalGeneration.forward`,  you can safely ignore this message.\n"
-     ]
-    }
-   ],
+   "outputs": [],
    "source": [
-    "trainer.train()"
+    "#trainer.train()\n",
+    "trainer.train(resume_from_checkpoint = True)\n"
    ]
   },
   {
@@ -2580,10 +2524,11 @@
     "    \"dataset\": \"Common Voice 11.0\",  # a 'pretty' name for the training dataset\n",
     "    #\"dataset\": \"Google FLEURS\",  # a 'pretty' name for the training dataset\n",
     "    \"language\": \"el\",\n",
-    "    \"model_name\": \"Whisper Medium El Greco Greek\",  # a 'pretty' name for your model\n",
+    "    \"model_name\": \"Whisper Medium El Greco\",  # a 'pretty' name for your model\n",
     "    \"finetuned_from\": \"openai/whisper-medium\",\n",
     "    \"tasks\": \"automatic-speech-recognition\",\n",
-    "    \"tags\": \"hf-asr-leaderboard, whisper-medium, mozilla-foundation/common_voice_11_0, greek, whisper-event\",\n",
+    "    \"tags\": \"hf-asr-leaderboard\",\n",
+    "    #\"tags\": \"hf-asr-leaderboard, whisper-medium, mozilla-foundation/common_voice_11_0, greek, whisper-event\",\n",
     "}"
    ]
   },
@@ -2608,6 +2553,14 @@
    "source": [
     "trainer.push_to_hub(**kwargs)"
    ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "afa8a29a-59ab-463f-b2bf-2fbf2bae9b82",
+   "metadata": {},
+   "outputs": [],
+   "source": []
   }
  ],
  "metadata": {

push.py

@@ -0,0 +1,758 @@
+#!/usr/bin/env python
+# coding=utf-8
+# Copyright 2022 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Fine-tuning the library models for sequence to sequence speech recognition
+with 🤗 Datasets' streaming mode.
+"""
+# This progam was modified by Michael Kamfonas (mkamfonas@infokarta.com) on Dec 11 2022
+# - added options for drpout, gradient_checkpointing, use_cache, stopping_strategy and streaming
+# - restructured it to enable both streaming and non-streaming modes 
+# - allows concatenation of mutiple datasets (single-string comma-separated) for interleaving
+#     The following params must have the same number of comma-separated (,) elements:
+#       dataset_name, 
+#       dataset_config_name, 
+#       train_split_name and eval_split_name (each element plus-separated (+) for multiple splits), 
+#       text_column_name and audio_column_name 
+ 
+
+import logging
+import os
+import sys
+from dataclasses import dataclass, field
+from typing import Any, Dict, List, Optional, Union
+
+import datasets
+import torch
+from datasets import Audio, DatasetDict, IterableDatasetDict, interleave_datasets, load_dataset
+from torch.utils.data import IterableDataset
+
+import evaluate
+import transformers
+from transformers import (
+    AutoConfig,
+    AutoFeatureExtractor,
+    AutoModelForSpeechSeq2Seq,
+    AutoProcessor,
+    AutoTokenizer,
+    HfArgumentParser,
+    Seq2SeqTrainer,
+    Seq2SeqTrainingArguments,
+    TrainerCallback,
+    set_seed,
+)
+from transformers.trainer_pt_utils import IterableDatasetShard
+from transformers.trainer_utils import get_last_checkpoint, is_main_process
+from transformers.utils import check_min_version, send_example_telemetry
+from transformers.utils.versions import require_version
+from transformers.models.whisper.english_normalizer import BasicTextNormalizer
+
+#TEXT_COL_NAME="text"
+TEXT_COL_NAME="sentence,transcription"
+AUDIO_COL_NAME="audio"
+
+# Will error if the minimal version of Transformers is not installed. Remove at your own risks.
+check_min_version("4.25.0.dev0")
+
+require_version("datasets>=1.18.2", "To fix: pip install -r examples/pytorch/speech-recognition/requirements.txt")
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class ModelArguments:
+    """
+    Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
+    """
+
+    model_name_or_path: str = field(
+        metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
+    )
+    config_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
+    )
+    tokenizer_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
+    )
+    feature_extractor_name: Optional[str] = field(
+        default=None, metadata={"help": "feature extractor name or path if not the same as model_name"}
+    )
+    cache_dir: Optional[str] = field(
+        default=None,
+        metadata={"help": "Where to store the pretrained models downloaded from huggingface.co"},
+    )
+    use_fast_tokenizer: bool = field(
+        default=True,
+        metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
+    )
+    model_revision: str = field(
+        default="main",
+        metadata={"help": "The specific model version to use (can be a branch name, tag name or commit id)."},
+    )
+    use_auth_token: bool = field(
+        default=False,
+        metadata={
+            "help": (
+                "Will use the token generated when running `huggingface-cli login` (necessary to use this script "
+                "with private models)."
+            )
+        },
+    )
+    freeze_feature_encoder: bool = field(
+        default=True, metadata={"help": "Deprecated - Whether to freeze the feature encoder layers of the model."}
+    )
+    freeze_encoder: bool = field(
+        default=False, metadata={"help": "Whether to freeze the entire encoder of the seq2seq model."}
+    )
+    forced_decoder_ids: List[List[int]] = field(
+        default=None,
+        metadata={
+            "help": (
+                "A list of pairs of integers which indicates a mapping from generation indices to token indices "
+                "that will be forced before sampling. For example, [[0, 123]] means the first generated token "
+                "will always be a token of index 123."
+            )
+        },
+    )
+    suppress_tokens: List[int] = field(
+        default=None, metadata={"help": "A list of tokens that will be suppressed at generation."}
+    )
+    model_index_name: str = field(default=None, metadata={"help": "Pretty name for the model card."})
+
+    ## added by Michael Kamfonas
+    use_cache: bool = field(
+        default=False, metadata={"help": "Whether to use cache."}
+    )
+
+    dropout: float = field(
+        default = 0.0, metadata = {"help": "dropout probability."}
+    )
+
+
+@dataclass
+class DataTrainingArguments:
+    """
+    Arguments pertaining to what data we are going to input our model for training and eval.
+    """
+
+    dataset_name: str = field(
+        default=None, 
+        metadata={"help": "The name of the dataset to use (via the datasets library)."}
+    )
+    dataset_config_name: Optional[str] = field(
+        default=None, 
+        metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
+    )
+    text_column: Optional[str] = field(
+        default=None,
+        metadata={"help": "The name of the column in the datasets containing the full texts (for summarization)."},
+    )
+    max_train_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": (
+                "For debugging purposes or quicker training, truncate the number of training examples to this "
+                "value if set."
+            )
+        },
+    )
+    max_eval_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": (
+                "For debugging purposes or quicker training, truncate the number of evaluation examples to this "
+                "value if set."
+            )
+        },
+    )
+    audio_column_name: str = field(
+        default="audio",
+        metadata={"help": "The name of the dataset column containing the audio data. Defaults to 'audio'"},
+    )
+    text_column_name: str = field(
+        default="text",
+        metadata={"help": "The name of the dataset column containing the text data. Defaults to 'text'"},
+    )
+    max_duration_in_seconds: float = field(
+        default=20.0,
+        metadata={
+            "help": (
+                "Truncate audio files that are longer than `max_duration_in_seconds` seconds to"
+                " 'max_duration_in_seconds`"
+            )
+        },
+    )
+    min_duration_in_seconds: float = field(
+        default=0.0, metadata={"help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"}
+    )
+    train_split_name: str = field(
+        default="train",
+        metadata={
+            "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
+        },
+    )
+    eval_split_name: str = field(
+        default="test",
+        metadata={
+            "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
+        },
+    )
+    do_lower_case: bool = field(
+        default=False,
+        metadata={"help": "Whether the target text should be lower cased."},
+    )
+    do_remove_punctuation: bool = field(
+        default=False,
+        metadata={"help": "Whether the target text should be striped of punctuation."},
+    )
+    do_normalize_eval: bool = field(
+        default=True,
+        metadata={"help": "Whether to normalise the references and predictions in the eval WER calculation."},
+    )
+    language: str = field(
+        default=None,
+        metadata={
+            "help": (
+                "Language for multilingual fine-tuning. This argument should be set for multilingual fine-tuning "
+                "only. For English speech recognition, it should be set to `None`."
+            )
+        },
+    )
+    task: str = field(
+        default="transcribe",
+        metadata={"help": "Task, either `transcribe` for speech recognition or `translate` for speech translation."},
+    )
+    shuffle_buffer_size: Optional[int] = field(
+        default=500,
+        metadata={
+            "help": (
+                "The number of streamed examples to download before shuffling them. The large the buffer, "
+                "the closer it is to real offline shuffling."
+            )
+        },
+    )
+    stopping_strategy: Optional[str] = field(
+        default="all_exhausted",
+        metadata={
+             "help": "Strategy used to consume interleaved data. Default = 'all_exhausted'"
+        }
+    )
+    streaming: bool = field(
+        default=True,
+        metadata={"help": "Whether to use streaming mode to load and pre-process the data."},
+    )
+
+@dataclass
+class DataCollatorSpeechSeq2SeqWithPadding:
+    """
+    Data collator that will dynamically pad the inputs received.
+    Args:
+        processor ([`WhisperProcessor`])
+            The processor used for processing the data.
+        decoder_start_token_id (`int`)
+            The begin-of-sentence of the decoder.
+    """
+
+    processor: Any
+    decoder_start_token_id: int
+
+    def __call__(self, features: List[Dict[str, Union[List[int], torch.Tensor]]]) -> Dict[str, torch.Tensor]:
+        # split inputs and labels since they have to be of different lengths and need
+        # different padding methods
+        model_input_name = self.processor.model_input_names[0]
+        input_features = [{model_input_name: feature[model_input_name]} for feature in features]
+        label_features = [{"input_ids": feature["labels"]} for feature in features]
+
+        batch = self.processor.feature_extractor.pad(input_features, return_tensors="pt")
+
+        labels_batch = self.processor.tokenizer.pad(label_features, return_tensors="pt")
+
+        # replace padding with -100 to ignore loss correctly
+        labels = labels_batch["input_ids"].masked_fill(labels_batch.attention_mask.ne(1), -100)
+
+        # if bos token is appended in previous tokenization step,
+        # cut bos token here as it's append later anyways
+        if (labels[:, 0] == self.decoder_start_token_id).all().cpu().item():
+            labels = labels[:, 1:]
+
+        batch["labels"] = labels
+
+        return batch
+
+
+def load_streaming_dataset(dataset_name, dataset_config_name, split="train", **kwargs):
+    """
+    Utility function to load a dataset in streaming mode. For datasets with multiple splits,
+    each split is loaded individually and then splits combined by taking alternating examples from
+    each (interleaving).
+    """
+    if "+" in split:
+        # load multiple splits separated by the `+` symbol with streaming mode
+        dataset_splits = [
+            load_dataset(dataset_name, dataset_config_name, split=split_name, streaming=True, **kwargs)
+            for split_name in split.split("+")
+        ]
+        # interleave multiple splits to form one dataset
+        interleaved_dataset = interleave_datasets(dataset_splits)
+        return interleaved_dataset
+    else:
+        # load a single split *with* streaming mode
+        dataset = load_dataset(dataset_name, dataset_config_name, split=split, streaming=True, **kwargs)
+        return dataset
+
+def load_multiple_streaming_datasets(
+    dataset_names: List,
+    dataset_config_names: List,
+    splits: Optional[List] = None,
+    text_column_names: Optional[List] = None,
+    audio_column_names: Optional[List] = None,
+    sampling_rate: Optional[int] = 16000,
+    stopping_strategy: Optional[str] = "all_exhausted",
+    streaming = True,
+    **kwargs
+    ):
+
+    if len(dataset_names) != len(dataset_config_names):
+        raise ValueError(
+            f"Ensure one config is passed for each dataset, got {len(dataset_names)} datasets and"
+            f" {len(dataset_config_names)} configs."
+        )
+
+    if splits is not None and len(splits) != len(dataset_names):
+        raise ValueError(
+            f"Ensure one train_split is passed for each dataset, got {len(dataset_names)} datasets and {len(splits)} splits."
+        )
+
+    if text_column_names is not None and len(text_column_names) != len(dataset_names):
+        raise ValueError(
+            f"Ensure one text column name is passed for each dataset, got {len(dataset_names)} datasets and"
+            f" {len(text_column_names)} text column names."
+        )
+
+    if audio_column_names is not None and len(audio_column_names) != len(dataset_names):
+        raise ValueError(
+            f"Ensure one text column name is passed for each dataset, got {len(dataset_names)} datasets and"
+            f" {len(audio_column_names)} text column names."
+        )
+
+    splits = splits if splits is not None \
+        else ["train" for i in range(len(dataset_names))]
+
+    text_column_names = (
+        text_column_names if text_column_names is not None \
+            else [TEXT_COL_NAME for i in range(len(dataset_names))]
+    )
+
+    audio_column_names = (
+        audio_column_names if audio_column_names is not None \
+            else [AUDIO_COL_NAME for i in range(len(dataset_names))]
+    )
+
+    all_data_splits = []
+    # iterate over the datasets we want to interleave
+    for dset, cfgNm, splt,  txtColNm, audColNm in zip(dataset_names,dataset_config_names,\
+                                                splits,text_column_names, audio_column_names):
+
+        dset_splits = [load_dataset(dset, cfgNm, split=c, streaming=streaming, **kwargs) \
+            for c in splt.split('+') if c != '-']
+
+        if streaming:
+            dset_splits = [ds if TEXT_COL_NAME  in ds.features else ds.rename_column(txtColNm, TEXT_COL_NAME) \
+                for ds in dset_splits ]
+            dset_splits = [ds if AUDIO_COL_NAME in ds.features else ds.rename_column(audColNm, AUDIO_COL_NAME) \
+                for ds in dset_splits]
+
+            if len(dset_splits)>0 and sampling_rate != next(iter(dset_splits[0]))[AUDIO_COL_NAME]['sampling_rate']:
+                dset_splits = [ds.cast_column(AUDIO_COL_NAME, Audio(sampling_rate)) for ds in dset_splits]
+        else:
+
+            dset_splits = [ds if TEXT_COL_NAME  in ds.column_names else ds.rename_column(txtColNm, TEXT_COL_NAME) \
+                for ds in dset_splits ]
+            dset_splits = [ds if AUDIO_COL_NAME in ds.column_names else ds.rename_column(audColNm, AUDIO_COL_NAME) \
+                for ds in dset_splits]
+        
+        if len(dset_splits)>0 and sampling_rate != next(iter(dset_splits[0]))[AUDIO_COL_NAME]['sampling_rate']:
+            dset_splits = [ds.cast_column(AUDIO_COL_NAME, Audio(sampling_rate)) for ds in dset_splits]
+
+        cols2keep = set([AUDIO_COL_NAME, TEXT_COL_NAME])
+
+        dset_splits = [ds.remove_columns(set(ds.features.keys()) - cols2keep) for ds in dset_splits]
+
+        all_data_splits +=   dset_splits
+
+    return interleave_datasets(all_data_splits, stopping_strategy=stopping_strategy)
+
+def main():
+    # 1. Parse input arguments
+    # See all possible arguments in src/transformers/training_args.py
+    # or by passing the --help flag to this script.
+    # We now keep distinct sets of args, for a cleaner separation of concerns.
+    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, Seq2SeqTrainingArguments))
+
+    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
+        # If we pass only one argument to the script and it's the path to a json file,
+        # let's parse it to get our arguments.
+        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
+    else:
+        model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    # Sending telemetry. Tracking the example usage helps us better allocate resources to maintain them. The
+    # information sent is the one passed as arguments along with your Python/PyTorch versions.
+    send_example_telemetry("run_speech_recognition_seq2seq_streaming", model_args, data_args)
+
+    # 2. Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        handlers=[logging.StreamHandler(sys.stdout)],
+    )
+    log_level = training_args.get_process_log_level()
+    logger.setLevel(log_level)
+    datasets.utils.logging.set_verbosity(log_level)
+    transformers.utils.logging.set_verbosity(log_level)
+    transformers.utils.logging.enable_default_handler()
+    transformers.utils.logging.enable_explicit_format()
+
+    logger.setLevel(logging.INFO if is_main_process(training_args.local_rank) else logging.WARN)
+
+    # Log on each process the small summary:
+    logger.warning(
+        f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+        f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
+    )
+    logger.info(f"Training/evaluation parameters {training_args}")
+
+    # Set the verbosity to info of the Transformers logger (on main process only):
+    if is_main_process(training_args.local_rank):
+        transformers.utils.logging.set_verbosity_info()
+    logger.info("Training/evaluation parameters %s", training_args)
+
+    # 3. Detecting last checkpoint and eventually continue from last checkpoint
+    last_checkpoint = None
+    if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
+        last_checkpoint = get_last_checkpoint(training_args.output_dir)
+        if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
+            raise ValueError(
+                f"Output directory ({training_args.output_dir}) already exists and is not empty. "
+                "Use --overwrite_output_dir to overcome."
+            )
+        elif last_checkpoint is not None and training_args.resume_from_checkpoint is None:
+            logger.info(
+                f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
+                "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
+            )
+
+    # Set seed before initializing model.
+    set_seed(training_args.seed)
+
+    # 5. Load pretrained model, tokenizer, and feature extractor
+    #
+    # Distributed training:
+    # The .from_pretrained methods guarantee that only one local process can concurrently
+    config = AutoConfig.from_pretrained(
+        model_args.config_name if model_args.config_name else model_args.model_name_or_path,
+        cache_dir=model_args.cache_dir,
+        revision=model_args.model_revision,
+        use_auth_token=True if model_args.use_auth_token else None,
+    )
+
+    config.update({ "forced_decoder_ids": model_args.forced_decoder_ids, 
+                    "suppress_tokens": model_args.suppress_tokens})
+
+    feature_extractor = AutoFeatureExtractor.from_pretrained(
+        model_args.feature_extractor_name if model_args.feature_extractor_name else model_args.model_name_or_path,
+        cache_dir=model_args.cache_dir,
+        revision=model_args.model_revision,
+        use_auth_token=True if model_args.use_auth_token else None,
+    )
+    tokenizer = AutoTokenizer.from_pretrained(
+        model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
+        cache_dir=model_args.cache_dir,
+        use_fast=model_args.use_fast_tokenizer,
+        revision=model_args.model_revision,
+        use_auth_token=True if model_args.use_auth_token else None,
+    )
+    model = AutoModelForSpeechSeq2Seq.from_pretrained(
+        model_args.model_name_or_path,
+        config=config,
+        cache_dir=model_args.cache_dir,
+        revision=model_args.model_revision,
+        use_auth_token=True if model_args.use_auth_token else None,
+    )
+
+    model.config.use_cache = model_args.use_cache
+    model.config.dropout = model_args.dropout
+    if training_args.gradient_checkpointing:
+        model.gradient_checkpointing_enable()
+
+    if model.config.decoder_start_token_id is None:
+        raise ValueError("Make sure that `config.decoder_start_token_id` is correctly defined")
+
+    # deprecated
+    #if model_args.freeze_feature_encoder:
+    #    model.freeze_feature_encoder()
+
+    if model_args.freeze_encoder:
+        model.freeze_encoder()
+        model.model.encoder.gradient_checkpointing = False
+
+    if data_args.language is not None:
+        # We only need to set the task id when the language is specified (i.e. in a multilingual setting)
+        tokenizer.set_prefix_tokens(language=data_args.language, task=data_args.task)
+
+
+    # 4. Load dataset
+    raw_datasets = IterableDatasetDict() if data_args.streaming else DatasetDict()
+
+    # if training_args.do_train:
+    #     raw_datasets["train"] = load_streaming_dataset(
+    #         data_args.dataset_name,
+    #         data_args.dataset_config_name,
+    #         split=data_args.train_split_name,
+    #         use_auth_token=True if model_args.use_auth_token else None,
+    #     )
+
+    # if training_args.do_eval:
+    #     raw_datasets["eval"] = load_streaming_dataset(
+    #         data_args.dataset_name,
+    #         data_args.dataset_config_name,
+    #         split=data_args.eval_split_name,
+    #         use_auth_token=True if model_args.use_auth_token else None,
+    #     )
+
+    if training_args.do_train:
+        raw_datasets["train"] = load_multiple_streaming_datasets(
+            dataset_names=data_args.dataset_name.split(","), 
+            dataset_config_names=data_args.dataset_config_name.split(","),
+            splits = data_args.train_split_name.split(","),
+            text_column_names = data_args.text_column_name.split(","),
+            sampling_rate  = feature_extractor.sampling_rate, 
+            streaming=data_args.streaming,
+            use_auth_token=True if model_args.use_auth_token else None,
+        )
+
+    if training_args.do_eval:
+        raw_datasets["eval"] = load_multiple_streaming_datasets(
+            dataset_names=data_args.dataset_name.split(","), 
+            dataset_config_names=data_args.dataset_config_name.split(","),
+            splits  = data_args.eval_split_name.split(","),
+            text_column_names = data_args.text_column_name.split(","),
+            sampling_rate  = feature_extractor.sampling_rate,
+            streaming=False,
+            use_auth_token=True if model_args.use_auth_token else None,
+        )
+
+    raw_datasets_features = list(next(iter(raw_datasets.values())).features.keys())
+
+    if AUDIO_COL_NAME not in raw_datasets_features:
+        raise ValueError(
+            f"--audio_column_name '{data_args.audio_column_name}' not found in dataset '{data_args.dataset_name}'. "
+            "Make sure to set `--audio_column_name` to the correct audio column - one of "
+            f"{', '.join(raw_datasets_features)}."
+        )
+
+    if TEXT_COL_NAME not in raw_datasets_features:
+        raise ValueError(
+            f"--text_column_name {TEXT_COL_NAME} not found in dataset. "
+            "Make sure to set `--text_column_name` to the the respective correct text columns."
+        )
+
+
+    # 6. Resample speech dataset if necessary
+    #dataset_sampling_rate = next(iter(raw_datasets.values())).features[data_args.audio_column_name].sampling_rate
+    #if dataset_sampling_rate != feature_extractor.sampling_rate:
+    #    raw_datasets = raw_datasets.cast_column(
+    #        data_args.audio_column_name, datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate)
+    #    )
+
+    # 7. Preprocessing the datasets.
+    # We need to read the audio files as arrays and tokenize the targets.
+    max_input_length = data_args.max_duration_in_seconds * feature_extractor.sampling_rate
+    min_input_length = data_args.min_duration_in_seconds * feature_extractor.sampling_rate
+    audio_column_name = AUDIO_COL_NAME
+    text_column_name = TEXT_COL_NAME
+    model_input_name = feature_extractor.model_input_names[0]
+    do_lower_case = data_args.do_lower_case
+    do_remove_punctuation = data_args.do_remove_punctuation
+    normalizer = BasicTextNormalizer()  # 'official' text normalizer from OpenAI
+
+    if data_args.max_train_samples is not None:
+        raw_datasets["train"] = (
+            raw_datasets["train"].take(data_args.max_train_samples)
+            if data_args.streaming
+            else raw_datasets["train"].select(range(data_args.max_train_samples))
+        )
+
+    if data_args.max_eval_samples is not None:
+        raw_datasets["eval"] = (
+            raw_datasets["eval"].take(data_args.max_eval_samples)
+            if data_args.streaming
+            else raw_datasets["eval"].select(range(data_args.max_eval_samples))
+        )
+
+    def prepare_dataset(batch):
+        # process audio
+        sample = batch[audio_column_name]
+        inputs = feature_extractor(sample["array"], sampling_rate=sample["sampling_rate"])
+        # process audio length
+        batch[model_input_name] = inputs.get(model_input_name)[0]
+        batch["input_length"] = len(sample["array"])
+
+        # process targets
+        input_str = batch[text_column_name].lower() if do_lower_case else batch[text_column_name]
+        if do_remove_punctuation:
+            input_str = normalizer(input_str).strip()
+        batch["labels"] = tokenizer(input_str).input_ids
+        return batch
+
+    with training_args.main_process_first(desc="dataset map pre-processing"):
+        vectorized_datasets = raw_datasets.map(
+            prepare_dataset,
+            remove_columns=raw_datasets_features
+        ).with_format("torch")
+
+        if training_args.do_train and data_args.streaming:
+            # manually shuffle if streaming (done by the trainer for non-streaming)
+            vectorized_datasets["train"] = vectorized_datasets["train"].shuffle(
+                buffer_size=data_args.shuffle_buffer_size,
+                seed=training_args.seed,
+            )
+
+    # filter training data that is shorter than min_input_length or longer than
+    # max_input_length
+    def is_audio_in_length_range(length):
+        return min_input_length < length < max_input_length
+
+    if training_args.do_train:
+        vectorized_datasets["train"] = vectorized_datasets["train"].filter(
+            is_audio_in_length_range,
+            input_columns=["input_length"],
+        )
+
+    # 8. Load Metric
+    metric = evaluate.load("wer")
+    do_normalize_eval = data_args.do_normalize_eval
+
+    def compute_metrics(pred):
+        pred_ids = pred.predictions
+
+        pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
+
+        pred_str = tokenizer.batch_decode(pred_ids, skip_special_tokens=True)
+        # we do not want to group tokens when computing the metrics
+        label_str = tokenizer.batch_decode(pred.label_ids, skip_special_tokens=True)
+
+        if do_normalize_eval:
+            pred_str = [normalizer(pred) for pred in pred_str]
+            label_str = [normalizer(label) for label in label_str]
+            # filtering step to only evaluate the samples that correspond to non-zero references:
+            pred_str = [pred_str[i] for i in range(len(pred_str)) if len(label_str[i]) > 0]
+            label_str = [label_str[i] for i in range(len(label_str)) if len(label_str[i]) > 0]
+
+        wer = 100 * metric.compute(predictions=pred_str, references=label_str)
+
+        return {"wer": wer}
+
+    # 9. Create a single speech processor
+    if is_main_process(training_args.local_rank):
+        # save feature extractor, tokenizer and config
+        feature_extractor.save_pretrained(training_args.output_dir)
+        tokenizer.save_pretrained(training_args.output_dir)
+        config.save_pretrained(training_args.output_dir)
+
+    processor = AutoProcessor.from_pretrained(training_args.output_dir)
+
+    # 10. Define data collator
+    data_collator = DataCollatorSpeechSeq2SeqWithPadding(
+        processor=processor,
+        decoder_start_token_id=model.config.decoder_start_token_id,
+    )
+
+    # 11. Configure Trainer
+    # Trainer callback to reinitialise and reshuffle the streamable datasets at the beginning of each epoch
+    # Only required for streaming: Trainer automatically shuffles non-streaming datasets
+    class ShuffleCallback(TrainerCallback):
+        def on_epoch_begin(self, args, state, control, train_dataloader, **kwargs):
+            if isinstance(train_dataloader.dataset, IterableDatasetShard):
+                pass  # set_epoch() is handled by the Trainer
+            elif isinstance(train_dataloader.dataset, IterableDataset):
+                train_dataloader.dataset.set_epoch(train_dataloader.dataset._epoch + 1)
+
+    # Initialize Trainer
+    trainer = Seq2SeqTrainer(
+        model=model,
+        args=training_args,
+        train_dataset=vectorized_datasets["train"] if training_args.do_train else None,
+        eval_dataset=vectorized_datasets["eval"] if training_args.do_eval else None,
+        tokenizer=feature_extractor,
+        data_collator=data_collator,
+        compute_metrics=compute_metrics if training_args.predict_with_generate else None,
+        callbacks=[ShuffleCallback()] if data_args.streaming else None,
+    )
+
+    # 12. Training
+    if training_args.do_train:
+        checkpoint = None
+        if training_args.resume_from_checkpoint is not None:
+            checkpoint = training_args.resume_from_checkpoint
+        elif last_checkpoint is not None:
+            checkpoint = last_checkpoint
+        train_result = trainer.train(resume_from_checkpoint=checkpoint)
+        trainer.save_model()  # Saves the feature extractor too for easy upload
+
+        metrics = train_result.metrics
+        if data_args.max_train_samples:
+            metrics["train_samples"] = data_args.max_train_samples
+        trainer.log_metrics("train", metrics)
+        trainer.save_metrics("train", metrics)
+        trainer.save_state()
+
+    # 13. Evaluation
+    results = {}
+    if training_args.do_eval:
+        logger.info("*** Evaluate ***")
+        metrics = trainer.evaluate(
+            metric_key_prefix="eval",
+            max_length=training_args.generation_max_length,
+            num_beams=training_args.generation_num_beams,
+        )
+        if data_args.max_eval_samples:
+            metrics["eval_samples"] = data_args.max_eval_samples
+
+        trainer.log_metrics("eval", metrics)
+        trainer.save_metrics("eval", metrics)
+
+    # 14. Write Training Stats
+    kwargs = {
+        "finetuned_from": model_args.model_name_or_path,
+        "tasks": "automatic-speech-recognition",
+        "tags": "whisper-event",
+    }
+    if data_args.dataset_name is not None:
+        kwargs["dataset_tags"] = data_args.dataset_name
+        if data_args.dataset_config_name is not None:
+            kwargs["dataset"] = f"{data_args.dataset_name} el"
+        else:
+            kwargs["dataset"] = data_args.dataset_name
+        if "common_voice" in data_args.dataset_name:
+            kwargs["language"] = data_args.dataset_config_name[:2]
+        if model_args.model_index_name is not None:
+            kwargs["model_name"] = model_args.model_index_name
+
+        trainer.push_to_hub(**kwargs)
+
+
+if __name__ == "__main__":
+    main()

pytorch_model.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:aa9c1cc306eaa444f99ef35b5e3ffb717f457ea531fd3a8d4727ffdbd4b4cccd
+oid sha256:9b29d215205456ef40d6ac76daabab944f90a03dcb2153898b46883c2733cf26
 size 3055754841

run.sh

@@ -0,0 +1,39 @@
+python run_speech_recognition_seq2seq_streaming.py \
+	--model_name_or_path="emilios/whisper-medium-el" \
+	--dataset_name="mozilla-foundation/common_voice_11_0" \
+	--dataset_name="mozilla-foundation/common_voice_11_0,google/fleurs" \
+	--dataset_config_name="el,el_gr" \
+	--language="greek" \
+	--train_split_name="train+validation,train+validation" \
+	--eval_split_name="test,-" \
+	--model_index_name="Whisper Medium El Greco" \
+	--text_column_name="sentence,transcription" \
+	--audio_column_name="audio,audio" \
+	--resume_from_checkpoint="4000" \
+	--streaming="False" \
+	--max_steps="5000" \
+	--max_steps="9000" \
+	--output_dir="./" \
+	--per_device_train_batch_size="32" \
+	--per_device_eval_batch_size="16" \
+	--logging_steps="25" \
+	--learning_rate="1e-5" \
+	--warmup_steps="500" \
+	--evaluation_strategy="steps" \
+	--eval_steps="1000" \
+	--save_strategy="steps" \
+	--save_steps="1000" \
+	--generation_max_length="225" \
+	--length_column_name="input_length" \
+	--max_duration_in_seconds="30" \
+	--freeze_feature_encoder="False" \
+	--report_to="tensorboard" \
+	--gradient_checkpointing \
+	--fp16 \
+	--overwrite_output_dir \
+	--do_train="True" \
+	--do_eval="True" \
+	--predict_with_generate \
+	--do_normalize_eval \
+	--use_auth_token \
+	--push_to_hub

run_eval_whisper_streaming.py

@@ -0,0 +1,150 @@
+import argparse
+
+from transformers import pipeline
+from transformers.models.whisper.english_normalizer import BasicTextNormalizer
+from datasets import load_dataset, Audio
+import evaluate
+
+wer_metric = evaluate.load("wer")
+
+
+def is_target_text_in_range(ref):
+    if ref.strip() == "ignore time segment in scoring":
+        return False
+    else:
+        return ref.strip() != ""
+
+
+def get_text(sample):
+    if "text" in sample:
+        return sample["text"]
+    elif "sentence" in sample:
+        return sample["sentence"]
+    elif "normalized_text" in sample:
+        return sample["normalized_text"]
+    elif "transcript" in sample:
+        return sample["transcript"]
+    elif "transcription" in sample:
+        return sample["transcription"]
+    else:
+        raise ValueError(
+            f"Expected transcript column of either 'text', 'sentence', 'normalized_text' or 'transcript'. Got sample of "
+            ".join{sample.keys()}. Ensure a text column name is present in the dataset."
+        )
+
+
+whisper_norm = BasicTextNormalizer()
+
+
+def normalise(batch):
+    batch["norm_text"] = whisper_norm(get_text(batch))
+    return batch
+
+
+def data(dataset):
+    for i, item in enumerate(dataset):
+        yield {**item["audio"], "reference": item["norm_text"]}
+
+
+def main(args):
+    batch_size = args.batch_size
+    whisper_asr = pipeline(
+        "automatic-speech-recognition", model=args.model_id, device=args.device
+    )
+
+    whisper_asr.model.config.forced_decoder_ids = (
+        whisper_asr.tokenizer.get_decoder_prompt_ids(
+            language=args.language, task="transcribe"
+        )
+    )
+
+    dataset = load_dataset(
+        args.dataset,
+        args.config,
+        split=args.split,
+        streaming=args.streaming,
+        use_auth_token=True,
+    )
+
+    # Only uncomment for debugging
+    dataset = dataset.take(args.max_eval_samples)
+
+    dataset = dataset.cast_column("audio", Audio(sampling_rate=16000))
+    dataset = dataset.map(normalise)
+    dataset = dataset.filter(is_target_text_in_range, input_columns=["norm_text"])
+
+    predictions = []
+    references = []
+
+    # run streamed inference
+    for out in whisper_asr(data(dataset), batch_size=batch_size):
+        predictions.append(whisper_norm(out["text"]))
+        references.append(out["reference"][0])
+
+    wer = wer_metric.compute(references=references, predictions=predictions)
+    wer = round(100 * wer, 2)
+
+    print("WER:", wer)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--model_id",
+        type=str,
+        required=True,
+        help="Model identifier. Should be loadable with 🤗 Transformers",
+    )
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        default="mozilla-foundation/common_voice_11_0",
+        help="Dataset name to evaluate the `model_id`. Should be loadable with 🤗 Datasets",
+    )
+    parser.add_argument(
+        "--config",
+        type=str,
+        required=True,
+        help="Config of the dataset. *E.g.* `'en'` for the English split of Common Voice",
+    )
+    parser.add_argument(
+        "--split",
+        type=str,
+        default="test",
+        help="Split of the dataset. *E.g.* `'test'`",
+    )
+
+    parser.add_argument(
+        "--device",
+        type=int,
+        default=-1,
+        help="The device to run the pipeline on. -1 for CPU (default), 0 for the first GPU and so on.",
+    )
+    parser.add_argument(
+        "--batch_size",
+        type=int,
+        default=16,
+        help="Number of samples to go through each streamed batch.",
+    )
+    parser.add_argument(
+        "--max_eval_samples",
+        type=int,
+        default=None,
+        help="Number of samples to be evaluated. Put a lower number e.g. 64 for testing this script.",
+    )
+    parser.add_argument(
+        "--streaming",
+        type=bool,
+        default=True,
+        help="Choose whether you'd like to download the entire dataset or stream it during the evaluation.",
+    )
+    parser.add_argument(
+        "--language",
+        type=str,
+        required=True,
+        help="Two letter language code for the transcription language, e.g. use 'en' for English.",
+    )
+    args = parser.parse_args()
+
+    main(args)

run_interleave.py

@@ -61,7 +61,7 @@ from transformers.models.whisper.english_normalizer import BasicTextNormalizer
 
 #TEXT_COL_NAME="text"
 TEXT_COL_NAME="sentence,transcription"
-AUDIO_COL_NAME="audio,audio"
+AUDIO_COL_NAME="audio"
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
 check_min_version("4.25.0.dev0")

run_speech_recognition_seq2seq_streaming.py

@@ -0,0 +1,629 @@
+#!/usr/bin/env python
+# coding=utf-8
+# Copyright 2022 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Fine-tuning the library models for sequence to sequence speech recognition
+with 🤗 Datasets' streaming mode.
+"""
+# You can also adapt this script for your own sequence to sequence speech
+# recognition task. Pointers for this are left as comments.
+
+import logging
+import os
+import sys
+from dataclasses import dataclass, field
+from typing import Any, Dict, List, Optional, Union
+
+import datasets
+import torch
+from datasets import DatasetDict, IterableDatasetDict, interleave_datasets, load_dataset
+from torch.utils.data import IterableDataset
+
+import evaluate
+import transformers
+from transformers import (
+    AutoConfig,
+    AutoFeatureExtractor,
+    AutoModelForSpeechSeq2Seq,
+    AutoProcessor,
+    AutoTokenizer,
+    HfArgumentParser,
+    Seq2SeqTrainer,
+    Seq2SeqTrainingArguments,
+    TrainerCallback,
+    set_seed,
+)
+from transformers.models.whisper.english_normalizer import BasicTextNormalizer
+from transformers.trainer_pt_utils import IterableDatasetShard
+from transformers.trainer_utils import get_last_checkpoint, is_main_process
+from transformers.utils import check_min_version, send_example_telemetry
+from transformers.utils.versions import require_version
+
+
+# Will error if the minimal version of Transformers is not installed. Remove at your own risks.
+check_min_version("4.25.0.dev0")
+
+require_version("datasets>=1.18.2", "To fix: pip install -r examples/pytorch/speech-recognition/requirements.txt")
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class ModelArguments:
+    """
+    Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
+    """
+
+    model_name_or_path: str = field(
+        metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
+    )
+    config_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
+    )
+    tokenizer_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
+    )
+    feature_extractor_name: Optional[str] = field(
+        default=None, metadata={"help": "feature extractor name or path if not the same as model_name"}
+    )
+    cache_dir: Optional[str] = field(
+        default=None,
+        metadata={"help": "Where to store the pretrained models downloaded from huggingface.co"},
+    )
+    use_fast_tokenizer: bool = field(
+        default=True,
+        metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
+    )
+    model_revision: str = field(
+        default="main",
+        metadata={"help": "The specific model version to use (can be a branch name, tag name or commit id)."},
+    )
+    use_auth_token: bool = field(
+        default=False,
+        metadata={
+            "help": (
+                "Will use the token generated when running `huggingface-cli login` (necessary to use this script "
+                "with private models)."
+            )
+        },
+    )
+    freeze_feature_encoder: bool = field(
+        default=True, metadata={"help": "Whether to freeze the feature encoder layers of the model."}
+    )
+    freeze_encoder: bool = field(
+        default=False, metadata={"help": "Whether to freeze the entire encoder of the seq2seq model."}
+    )
+    forced_decoder_ids: List[List[int]] = field(
+        default=None,
+        metadata={
+            "help": (
+                "A list of pairs of integers which indicates a mapping from generation indices to token indices "
+                "that will be forced before sampling. For example, [[0, 123]] means the first generated token "
+                "will always be a token of index 123."
+            )
+        },
+    )
+    suppress_tokens: List[int] = field(
+        default=None, metadata={"help": "A list of tokens that will be suppressed at generation."}
+    )
+    model_index_name: str = field(default=None, metadata={"help": "Pretty name for the model card."})
+
+
+@dataclass
+class DataTrainingArguments:
+    """
+    Arguments pertaining to what data we are going to input our model for training and eval.
+    """
+
+    dataset_name: str = field(
+        default=None, metadata={"help": "The name of the dataset to use (via the datasets library)."}
+    )
+    dataset_config_name: Optional[str] = field(
+        default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
+    )
+    text_column: Optional[str] = field(
+        default=None,
+        metadata={"help": "The name of the column in the datasets containing the full texts (for summarization)."},
+    )
+    max_train_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": (
+                "For debugging purposes or quicker training, truncate the number of training examples to this "
+                "value if set."
+            )
+        },
+    )
+    max_eval_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": (
+                "For debugging purposes or quicker training, truncate the number of evaluation examples to this "
+                "value if set."
+            )
+        },
+    )
+    audio_column_name: str = field(
+        default="audio",
+        metadata={"help": "The name of the dataset column containing the audio data. Defaults to 'audio'"},
+    )
+    text_column_name: str = field(
+        default="text",
+        metadata={"help": "The name of the dataset column containing the text data. Defaults to 'text'"},
+    )
+    max_duration_in_seconds: float = field(
+        default=20.0,
+        metadata={
+            "help": (
+                "Truncate audio files that are longer than `max_duration_in_seconds` seconds to"
+                " 'max_duration_in_seconds`"
+            )
+        },
+    )
+    min_duration_in_seconds: float = field(
+        default=0.0, metadata={"help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"}
+    )
+    train_split_name: str = field(
+        default="train",
+        metadata={
+            "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
+        },
+    )
+    eval_split_name: str = field(
+        default="test",
+        metadata={
+            "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
+        },
+    )
+    do_lower_case: bool = field(
+        default=False,
+        metadata={"help": "Whether the target text should be lower cased."},
+    )
+    do_remove_punctuation: bool = field(
+        default=False,
+        metadata={"help": "Whether the target text should be striped of punctuation."},
+    )
+    do_normalize_eval: bool = field(
+        default=True,
+        metadata={"help": "Whether to normalise the references and predictions in the eval WER calculation."},
+    )
+    language: str = field(
+        default=None,
+        metadata={
+            "help": (
+                "Language for multilingual fine-tuning. This argument should be set for multilingual fine-tuning "
+                "only. For English speech recognition, it should be set to `None`."
+            )
+        },
+    )
+    task: str = field(
+        default="transcribe",
+        metadata={"help": "Task, either `transcribe` for speech recognition or `translate` for speech translation."},
+    )
+    shuffle_buffer_size: Optional[int] = field(
+        default=500,
+        metadata={
+            "help": (
+                "The number of streamed examples to download before shuffling them. The large the buffer, "
+                "the closer it is to real offline shuffling."
+            )
+        },
+    )
+    streaming: bool = field(
+        default=True,
+        metadata={"help": "Whether to use streaming mode to load and pre-process the data."},
+    )
+
+
+@dataclass
+class DataCollatorSpeechSeq2SeqWithPadding:
+    """
+    Data collator that will dynamically pad the inputs received.
+    Args:
+        processor ([`WhisperProcessor`])
+            The processor used for processing the data.
+        decoder_start_token_id (`int`)
+            The begin-of-sentence of the decoder.
+    """
+
+    processor: Any
+    decoder_start_token_id: int
+
+    def __call__(self, features: List[Dict[str, Union[List[int], torch.Tensor]]]) -> Dict[str, torch.Tensor]:
+        # split inputs and labels since they have to be of different lengths and need
+        # different padding methods
+        model_input_name = self.processor.model_input_names[0]
+        input_features = [{model_input_name: feature[model_input_name]} for feature in features]
+        label_features = [{"input_ids": feature["labels"]} for feature in features]
+
+        batch = self.processor.feature_extractor.pad(input_features, return_tensors="pt")
+
+        labels_batch = self.processor.tokenizer.pad(label_features, return_tensors="pt")
+
+        # replace padding with -100 to ignore loss correctly
+        labels = labels_batch["input_ids"].masked_fill(labels_batch.attention_mask.ne(1), -100)
+
+        # if bos token is appended in previous tokenization step,
+        # cut bos token here as it's append later anyways
+        if (labels[:, 0] == self.decoder_start_token_id).all().cpu().item():
+            labels = labels[:, 1:]
+
+        batch["labels"] = labels
+
+        return batch
+
+
+def load_maybe_streaming_dataset(dataset_name, dataset_config_name, split="train", streaming=True, **kwargs):
+    """
+    Utility function to load a dataset in streaming mode. For datasets with multiple splits,
+    each split is loaded individually and then splits combined by taking alternating examples from
+    each (interleaving).
+    """
+    if "+" in split:
+        # load multiple splits separated by the `+` symbol with streaming mode
+        dataset_splits = [
+            load_dataset(dataset_name, dataset_config_name, split=split_name, streaming=streaming, **kwargs)
+            for split_name in split.split("+")
+        ]
+        # interleave multiple splits to form one dataset
+        interleaved_dataset = interleave_datasets(dataset_splits)
+        return interleaved_dataset
+    else:
+        # load a single split *with* streaming mode
+        dataset = load_dataset(dataset_name, dataset_config_name, split=split, streaming=streaming, **kwargs)
+        return dataset
+
+
+def main():
+    # 1. Parse input arguments
+    # See all possible arguments in src/transformers/training_args.py
+    # or by passing the --help flag to this script.
+    # We now keep distinct sets of args, for a cleaner separation of concerns.
+    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, Seq2SeqTrainingArguments))
+
+    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
+        # If we pass only one argument to the script and it's the path to a json file,
+        # let's parse it to get our arguments.
+        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
+    else:
+        model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    # Sending telemetry. Tracking the example usage helps us better allocate resources to maintain them. The
+    # information sent is the one passed as arguments along with your Python/PyTorch versions.
+    send_example_telemetry("run_speech_recognition_seq2seq_streaming", model_args, data_args)
+
+    # 2. Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        handlers=[logging.StreamHandler(sys.stdout)],
+    )
+    log_level = training_args.get_process_log_level()
+    logger.setLevel(log_level)
+    datasets.utils.logging.set_verbosity(log_level)
+    transformers.utils.logging.set_verbosity(log_level)
+    transformers.utils.logging.enable_default_handler()
+    transformers.utils.logging.enable_explicit_format()
+
+    logger.setLevel(logging.INFO if is_main_process(training_args.local_rank) else logging.WARN)
+
+    # Log on each process the small summary:
+    logger.warning(
+        f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+        f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
+    )
+    logger.info(f"Training/evaluation parameters {training_args}")
+
+    # Set the verbosity to info of the Transformers logger (on main process only):
+    if is_main_process(training_args.local_rank):
+        transformers.utils.logging.set_verbosity_info()
+    logger.info("Training/evaluation parameters %s", training_args)
+
+    # 3. Detecting last checkpoint and eventually continue from last checkpoint
+    last_checkpoint = None
+    if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
+        last_checkpoint = get_last_checkpoint(training_args.output_dir)
+        if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
+            raise ValueError(
+                f"Output directory ({training_args.output_dir}) already exists and is not empty. "
+                "Use --overwrite_output_dir to overcome."
+            )
+        elif last_checkpoint is not None and training_args.resume_from_checkpoint is None:
+            logger.info(
+                f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
+                "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
+            )
+
+    # Set seed before initializing model.
+    set_seed(training_args.seed)
+
+    # 4. Load dataset
+    raw_datasets = IterableDatasetDict() if data_args.streaming else DatasetDict()
+
+    if training_args.do_train:
+        raw_datasets["train"] = load_maybe_streaming_dataset(
+            data_args.dataset_name,
+            data_args.dataset_config_name,
+            split=data_args.train_split_name,
+            use_auth_token=True if model_args.use_auth_token else None,
+            streaming=data_args.streaming,
+        )
+
+    if training_args.do_eval:
+        raw_datasets["eval"] = load_maybe_streaming_dataset(
+            data_args.dataset_name,
+            data_args.dataset_config_name,
+            split=data_args.eval_split_name,
+            use_auth_token=True if model_args.use_auth_token else None,
+            streaming=data_args.streaming,
+        )
+
+    raw_datasets_features = list(next(iter(raw_datasets.values())).features.keys())
+
+    if data_args.audio_column_name not in raw_datasets_features:
+        raise ValueError(
+            f"--audio_column_name '{data_args.audio_column_name}' not found in dataset '{data_args.dataset_name}'. "
+            "Make sure to set `--audio_column_name` to the correct audio column - one of "
+            f"{', '.join(raw_datasets_features)}."
+        )
+
+    if data_args.text_column_name not in raw_datasets_features:
+        raise ValueError(
+            f"--text_column_name {data_args.text_column_name} not found in dataset '{data_args.dataset_name}'. "
+            "Make sure to set `--text_column_name` to the correct text column - one of "
+            f"{', '.join(raw_datasets_features)}."
+        )
+
+    # 5. Load pretrained model, tokenizer, and feature extractor
+    #
+    # Distributed training:
+    # The .from_pretrained methods guarantee that only one local process can concurrently
+    config = AutoConfig.from_pretrained(
+        model_args.config_name if model_args.config_name else model_args.model_name_or_path,
+        cache_dir=model_args.cache_dir,
+        revision=model_args.model_revision,
+        use_auth_token=True if model_args.use_auth_token else None,
+    )
+
+    config.update({"forced_decoder_ids": model_args.forced_decoder_ids, "suppress_tokens": model_args.suppress_tokens})
+
+    if training_args.gradient_checkpointing:
+        config.update({"use_cache": False})
+
+    feature_extractor = AutoFeatureExtractor.from_pretrained(
+        model_args.feature_extractor_name if model_args.feature_extractor_name else model_args.model_name_or_path,
+        cache_dir=model_args.cache_dir,
+        revision=model_args.model_revision,
+        use_auth_token=True if model_args.use_auth_token else None,
+    )
+    tokenizer = AutoTokenizer.from_pretrained(
+        model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
+        cache_dir=model_args.cache_dir,
+        use_fast=model_args.use_fast_tokenizer,
+        revision=model_args.model_revision,
+        use_auth_token=True if model_args.use_auth_token else None,
+    )
+    model = AutoModelForSpeechSeq2Seq.from_pretrained(
+        model_args.model_name_or_path,
+        config=config,
+        cache_dir=model_args.cache_dir,
+        revision=model_args.model_revision,
+        use_auth_token=True if model_args.use_auth_token else None,
+    )
+
+    if model.config.decoder_start_token_id is None:
+        raise ValueError("Make sure that `config.decoder_start_token_id` is correctly defined")
+
+    if model_args.freeze_feature_encoder:
+        model.freeze_feature_encoder()
+
+    if model_args.freeze_encoder:
+        model.freeze_encoder()
+
+    if data_args.language is not None:
+        # We only need to set the task id when the language is specified (i.e. in a multilingual setting)
+        tokenizer.set_prefix_tokens(language=data_args.language, task=data_args.task)
+
+    # 6. Resample speech dataset if necessary
+    dataset_sampling_rate = next(iter(raw_datasets.values())).features[data_args.audio_column_name].sampling_rate
+    if dataset_sampling_rate != feature_extractor.sampling_rate:
+        raw_datasets = raw_datasets.cast_column(
+            data_args.audio_column_name, datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate)
+        )
+
+    # 7. Preprocessing the datasets.
+    # We need to read the audio files as arrays and tokenize the targets.
+    max_input_length = data_args.max_duration_in_seconds * feature_extractor.sampling_rate
+    min_input_length = data_args.min_duration_in_seconds * feature_extractor.sampling_rate
+    audio_column_name = data_args.audio_column_name
+    text_column_name = data_args.text_column_name
+    model_input_name = feature_extractor.model_input_names[0]
+    do_lower_case = data_args.do_lower_case
+    do_remove_punctuation = data_args.do_remove_punctuation
+    normalizer = BasicTextNormalizer()  # 'official' text normalizer from OpenAI
+
+    if data_args.max_train_samples is not None:
+        raw_datasets["train"] = (
+            raw_datasets["train"].take(data_args.max_train_samples)
+            if data_args.streaming
+            else raw_datasets["train"].select(range(data_args.max_train_samples))
+        )
+
+    if data_args.max_eval_samples is not None:
+        raw_datasets["eval"] = (
+            raw_datasets["eval"].take(data_args.max_eval_samples)
+            if data_args.streaming
+            else raw_datasets["eval"].select(range(data_args.max_eval_samples))
+        )
+
+    def prepare_dataset(batch):
+        # process audio
+        sample = batch[audio_column_name]
+        inputs = feature_extractor(sample["array"], sampling_rate=sample["sampling_rate"])
+        # process audio length
+        batch[model_input_name] = inputs.get(model_input_name)[0]
+        batch["input_length"] = len(sample["array"])
+
+        # process targets
+        input_str = batch[text_column_name].lower() if do_lower_case else batch[text_column_name]
+        if do_remove_punctuation:
+            input_str = normalizer(input_str).strip()
+        batch["labels"] = tokenizer(input_str).input_ids
+        return batch
+
+    with training_args.main_process_first(desc="dataset map pre-processing"):
+        vectorized_datasets = raw_datasets.map(
+            prepare_dataset,
+            remove_columns=raw_datasets_features,
+        ).with_format("torch")
+
+        if training_args.do_train and data_args.streaming:
+            # manually shuffle if streaming (done by the trainer for non-streaming)
+            vectorized_datasets["train"] = vectorized_datasets["train"].shuffle(
+                buffer_size=data_args.shuffle_buffer_size,
+                seed=training_args.seed,
+            )
+
+    # filter training data that is shorter than min_input_length or longer than
+    # max_input_length
+    def is_audio_in_length_range(length):
+        return min_input_length < length < max_input_length
+
+    if training_args.do_train:
+        vectorized_datasets["train"] = vectorized_datasets["train"].filter(
+            is_audio_in_length_range,
+            input_columns=["input_length"],
+        )
+
+    # 8. Load Metric
+    metric = evaluate.load("wer")
+    do_normalize_eval = data_args.do_normalize_eval
+
+    def compute_metrics(pred):
+        pred_ids = pred.predictions
+
+        pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
+
+        pred_str = tokenizer.batch_decode(pred_ids, skip_special_tokens=True)
+        # we do not want to group tokens when computing the metrics
+        label_str = tokenizer.batch_decode(pred.label_ids, skip_special_tokens=True)
+
+        if do_normalize_eval:
+            pred_str = [normalizer(pred) for pred in pred_str]
+            label_str = [normalizer(label) for label in label_str]
+            # filtering step to only evaluate the samples that correspond to non-zero references:
+            pred_str = [pred_str[i] for i in range(len(pred_str)) if len(label_str[i]) > 0]
+            label_str = [label_str[i] for i in range(len(label_str)) if len(label_str[i]) > 0]
+
+        wer = 100 * metric.compute(predictions=pred_str, references=label_str)
+
+        return {"wer": wer}
+
+    # 9. Create a single speech processor
+    if is_main_process(training_args.local_rank):
+        # save feature extractor, tokenizer and config
+        feature_extractor.save_pretrained(training_args.output_dir)
+        tokenizer.save_pretrained(training_args.output_dir)
+        config.save_pretrained(training_args.output_dir)
+
+    processor = AutoProcessor.from_pretrained(training_args.output_dir)
+
+    # 10. Define data collator
+    data_collator = DataCollatorSpeechSeq2SeqWithPadding(
+        processor=processor,
+        decoder_start_token_id=model.config.decoder_start_token_id,
+    )
+
+    # 11. Configure Trainer
+    # Trainer callback to reinitialise and reshuffle the streamable datasets at the beginning of each epoch
+    # Only required for streaming: Trainer automatically shuffles non-streaming datasets
+    class ShuffleCallback(TrainerCallback):
+        def on_epoch_begin(self, args, state, control, train_dataloader, **kwargs):
+            if isinstance(train_dataloader.dataset, IterableDatasetShard):
+                pass  # set_epoch() is handled by the Trainer
+            elif isinstance(train_dataloader.dataset, IterableDataset):
+                train_dataloader.dataset.set_epoch(train_dataloader.dataset._epoch + 1)
+
+    # Initialize Trainer
+    trainer = Seq2SeqTrainer(
+        model=model,
+        args=training_args,
+        train_dataset=vectorized_datasets["train"] if training_args.do_train else None,
+        eval_dataset=vectorized_datasets["eval"] if training_args.do_eval else None,
+        tokenizer=feature_extractor,
+        data_collator=data_collator,
+        compute_metrics=compute_metrics if training_args.predict_with_generate else None,
+        callbacks=[ShuffleCallback()] if data_args.streaming else None,
+    )
+
+    # 12. Training
+    if training_args.do_train:
+        checkpoint = None
+        if training_args.resume_from_checkpoint is not None:
+            checkpoint = training_args.resume_from_checkpoint
+        elif last_checkpoint is not None:
+            checkpoint = last_checkpoint
+        train_result = trainer.train(resume_from_checkpoint=checkpoint)
+        trainer.save_model()  # Saves the feature extractor too for easy upload
+
+        metrics = train_result.metrics
+        if data_args.max_train_samples:
+            metrics["train_samples"] = data_args.max_train_samples
+        trainer.log_metrics("train", metrics)
+        trainer.save_metrics("train", metrics)
+        trainer.save_state()
+
+    # 13. Evaluation
+    results = {}
+    if training_args.do_eval:
+        logger.info("*** Evaluate ***")
+        metrics = trainer.evaluate(
+            metric_key_prefix="eval",
+            max_length=training_args.generation_max_length,
+            num_beams=training_args.generation_num_beams,
+        )
+        if data_args.max_eval_samples:
+            metrics["eval_samples"] = data_args.max_eval_samples
+
+        trainer.log_metrics("eval", metrics)
+        trainer.save_metrics("eval", metrics)
+
+    # 14. Write Training Stats
+    kwargs = {
+        "finetuned_from": model_args.model_name_or_path,
+        "tasks": "automatic-speech-recognition",
+        "tags": "whisper-event",
+    }
+    if data_args.dataset_name is not None:
+        kwargs["dataset_tags"] = data_args.dataset_name
+        if data_args.dataset_config_name is not None:
+            kwargs["dataset"] = f"{data_args.dataset_name} {data_args.dataset_config_name}"
+        else:
+            kwargs["dataset"] = data_args.dataset_name
+        if "common_voice" in data_args.dataset_name:
+            kwargs["language"] = data_args.dataset_config_name[:2]
+        if model_args.model_index_name is not None:
+            kwargs["model_name"] = model_args.model_index_name
+
+    if training_args.push_to_hub:
+        trainer.push_to_hub(**kwargs)
+    else:
+        trainer.create_model_card(**kwargs)
+
+    return results
+
+
+if __name__ == "__main__":
+    main()

run_whisper-md-el-intlv-xs.sh

@@ -29,6 +29,7 @@ python run_interleave.py \
                 --dropout         0.1 \
                 --warmup_steps   500 \
                 --max_steps   5000 \
+		--resume_from_checkpoint="4000" \
                 --eval_steps   1000 \
                 --gradient_checkpointing   True \
                 --cache_dir   '~/.cache' \
@@ -42,8 +43,7 @@ python run_interleave.py \
                 --load_best_model_at_end   True \
                 --metric_for_best_model   wer \
                 --greater_is_better   False \
-                --push_to_hub   False 
+                --push_to_hub   True 
 
 
-#
 

runs/Dec13_00-20-31_150-136-33-0/1670891059.409924/events.out.tfevents.1670891059.150-136-33-0.3897722.1

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dd75ccf3b1bd98778e44cae4793e895aeec95a0d406fd4842436a2557e542e15
+size 5917

runs/Dec13_00-20-31_150-136-33-0/events.out.tfevents.1670891059.150-136-33-0.3897722.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c9c9398f6ed427bd41ff5741f352098952e1b913961b916e2d365d493d692c78
+size 4344

runs/Dec13_00-26-09_150-136-33-0/1670891185.5348835/events.out.tfevents.1670891185.150-136-33-0.3897722.3

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:35265950f470f0bb0bc82f0f6bb1ea4a83bc795bf5d255645f4f880d140d506a
+size 5917

runs/Dec13_00-26-09_150-136-33-0/events.out.tfevents.1670891185.150-136-33-0.3897722.2

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:94d957e6f71dfd826ff305f81cafa3d73556930b286f440116360b825b11332f
+size 4343

runs/Dec13_00-29-21_150-136-33-0/1670891404.5181074/events.out.tfevents.1670891404.150-136-33-0.127865.1

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2de5049f3bf41fae21055257bab18c7324d889c6f89810468c3ba3c622d9eb3b
+size 5917

runs/Dec13_00-29-21_150-136-33-0/1670891530.7157552/events.out.tfevents.1670891530.150-136-33-0.127865.2

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:367be3341ff0617872ec5a6d9835404033aeb617bd43a6c309b3eaf37b96a8c1
+size 5917

runs/Dec13_00-29-21_150-136-33-0/events.out.tfevents.1670891404.150-136-33-0.127865.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:49d3593a4263cbf140645c09337c65104f6b6257df0e3fdff3f10a4027c867c4
+size 9354

runs/Dec13_00-33-53_150-136-33-0/1670891644.5885968/events.out.tfevents.1670891644.150-136-33-0.127865.4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:33cf05b8a4dd04ead085260d1bbb46d5e480d28366bcf7be1970bc9b5e4eced2
+size 5917

runs/Dec13_00-33-53_150-136-33-0/events.out.tfevents.1670891644.150-136-33-0.127865.3

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0417f0824bb1e400e36093b415ea327c75a275f99909bd52a68f2639b4a2d2db
+size 4343

runs/Dec13_00-36-22_150-136-33-0/1670891797.5398705/events.out.tfevents.1670891797.150-136-33-0.152213.1

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:707fb35872449414d5446b923f39ed46d3efaf89754df4c9cdff9d2ad7e27dc3
+size 5917

runs/Dec13_00-36-22_150-136-33-0/events.out.tfevents.1670891797.150-136-33-0.152213.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f972921d72e7ec59f8f92edf0ffa2b0b0a0b3e0f1000b294806893018aba7201
+size 10941

training_args.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a3ebd5ec6b9b2945948080568a87ff00951d03abd4956eab0351149cc37e37fe
-size 3579
+oid sha256:dccf5c02020cb4db0b8e5e421cc7fbb47dfb469165fb362c0937177d17244b57
+size 3643