File size: 7,420 Bytes
a01ef8c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
#!/usr/bin/env python
# -*- coding: utf-8 -*-
#
# Copyright (c) 2023 Intel Corporation
#
# 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.
#
# SPDX-License-Identifier: Apache-2.0
#

import os
import argparse
import tempfile

from transformers import AutoTokenizer

from filelock import FileLock

from downloader.datasets import DataDownloader
from downloader.models import ModelDownloader

from tlt.distributed.pytorch.utils.pyt_distributed_utils import (
    DistributedTorch,
    DistributedTrainingArguments,
    HorovodTrainer
)


if __name__ == "__main__":

    default_data_dir = os.path.join(tempfile.gettempdir(), 'data')
    default_output_dir = os.path.join(tempfile.gettempdir(), 'output')

    for d in [default_data_dir, default_output_dir]:
        if not os.path.exists(d):
            os.makedirs(d)

    def directory_path(path):
        if os.path.isdir(path):
            return path
        else:
            raise argparse.ArgumentTypeError("'{}' is not a valid directory path.".format(path))

    print("******Distributed Training*****")

    description = 'Distributed training with PyTorch.'

    parser = argparse.ArgumentParser(description=description)

    parser.add_argument('--master_addr', type=str, required=False, help="Master node to run this script")
    parser.add_argument('--master_port', type=str, required=False, default='29500', help='Master port')
    parser.add_argument('--backend', type=str, required=False, default='ccl', help='Type of backend to use '
                        '(default: ccl)')
    parser.add_argument('--use-case', '--use_case', type=str, required=True,
                        help='Use case (image_classification|text_classification)')
    parser.add_argument('--epochs', type=int, required=False, default=1, help='Total epochs to train the model')
    parser.add_argument('--batch_size', type=int, required=False, default=128,
                        help='Global batch size to distribute data (default: 128)')
    parser.add_argument('--disable_ipex', action='store_true', required=False, help="Disables IPEX optimization to "
                        "the model. No effect when given --use-horovod as horovod with IPEX isn't supported.")
    parser.add_argument('--tlt_saved_objects_dir', type=directory_path, required=False, help='Path to TLT saved '
                        'distributed objects. The path must be accessible to all the nodes. For example: mounted '
                        'NFS drive. This arg is helpful when using TLT API/CLI. '
                        'See DistributedTorch.load_saved_objects() for more information.')
    parser.add_argument('--use-horovod', '--use_horovod', action='store_true', help='Use horovod for distributed '
                        'training.')
    parser.add_argument('--cuda', action='store_true', help='Use cuda device for distributed training')
    parser.add_argument('--dataset-dir', '--dataset_dir', type=directory_path, default=default_data_dir,
                        help="Path to dataset directory to save/load tfds dataset. This arg is helpful if you "
                        "plan to use this as a stand-alone script. Custom dataset is not supported yet!")
    parser.add_argument('--output-dir', '--output_dir', type=directory_path, default=default_output_dir,
                        help="Path to save the trained model and store logs. This arg is helpful if you "
                        "plan to use this as a stand-alone script")
    parser.add_argument('--dataset-name', '--dataset_name', type=str, default=None,
                        help="Dataset name to load from torchvision/Huggingface. This arg is helpful if you "
                        "plan to use this as a stand-alone script. Custom dataset is not supported yet!")
    parser.add_argument('--model-name', '--model_name', type=str, default=None,
                        help="Torchvision image classification model name "
                        "(or) Huggingface hub name for text classification models. This arg is helpful if you "
                        "plan to use this as a stand-alone script.")
    parser.add_argument('--max_seq_length', type=int, default=128,
                        help='Maximum sequence length that the model will be used with for text classification')

    args = parser.parse_args()

    train_data = None
    model = None
    optimizer, loss = None, None
    data_kwargs = {}

    if args.tlt_saved_objects_dir is not None:
        # Load the saved dataset and model objects
        loaded_objects = DistributedTorch.load_saved_objects(args.tlt_saved_objects_dir)

        train_data = loaded_objects.get('train_data')
        model = loaded_objects['model']
        loss = loaded_objects['loss']
        optimizer = loaded_objects['optimizer']
        data_kwargs['is_preprocessed'] = True
    else:
        if args.dataset_name is None:
            raise argparse.ArgumentError(args.dataset_name, "Please provide a dataset name to load from torchvision "
                                         "(or) datasets using --dataset-name")
        if args.model_name is None:
            raise argparse.ArgumentError(args.model_name, "Please provide torchvision model name (or) "
                                         "Huggingface hub name using --model-name")

        catalog = 'torchvision' if args.use_case == 'image_classification' else 'hugging_face'
        with FileLock(os.path.expanduser('~/.horovod_lock')):
            train_data = DataDownloader(args.dataset_name, args.dataset_dir, catalog).download(split='train')
            model = ModelDownloader(args.model_name, catalog, args.output_dir).download()
            if args.use_case == 'text_classification':
                data_kwargs['hf_tokenizer'] = AutoTokenizer.from_pretrained(args.model_name)
                data_kwargs['max_seq_length'] = args.max_seq_length
                data_kwargs['text_column_names'] = [c for c in train_data.column_names if c != 'label']
        data_kwargs['is_preprocessed'] = False

    if args.use_horovod:
        hvd_trainer = HorovodTrainer(args.cuda)

        train_loader, train_sampler = hvd_trainer.prepare_data(train_data, args.use_case, args.batch_size,
                                                               **data_kwargs)
        hvd_trainer.prepare_model(model, args.use_case, optimizer, loss)

        hvd_trainer.fit(train_loader, train_sampler, args.use_case, args.epochs)
    else:
        # Launch distributed job
        training_args = DistributedTrainingArguments(
            dataset=train_data,
            model=model,
            criterion=loss,
            optimizer=optimizer,
            epochs=args.epochs,
            batch_size=args.batch_size,
            disable_ipex=args.disable_ipex
        )

        dt = DistributedTorch(use_case=args.use_case)
        dt.launch_distributed_job(training_args, args.master_addr, args.master_port, args.backend)