|
| 1 | +import logging |
| 2 | +import re |
| 3 | +import tempfile |
| 4 | +from argparse import ArgumentParser |
| 5 | +from collections import OrderedDict |
| 6 | +from functools import partial |
| 7 | +from pathlib import Path |
| 8 | + |
| 9 | +import numpy as np |
| 10 | +import pandas as pd |
| 11 | +import torch |
| 12 | +from mmengine import Config, DictAction |
| 13 | +from mmengine.analysis import get_model_complexity_info |
| 14 | +from mmengine.analysis.print_helper import _format_size |
| 15 | +from mmengine.fileio import FileClient |
| 16 | +from mmengine.logging import MMLogger |
| 17 | +from mmengine.model import revert_sync_batchnorm |
| 18 | +from mmengine.registry import init_default_scope |
| 19 | +from mmengine.runner import Runner |
| 20 | +from modelindex.load_model_index import load |
| 21 | +from rich.console import Console |
| 22 | +from rich.table import Table |
| 23 | +from rich.text import Text |
| 24 | +from tqdm import tqdm |
| 25 | + |
| 26 | +from mmocr.registry import MODELS |
| 27 | + |
| 28 | +console = Console() |
| 29 | +MMOCR_ROOT = Path(__file__).absolute().parents[1] |
| 30 | + |
| 31 | + |
| 32 | +def parse_args(): |
| 33 | + parser = ArgumentParser(description='Valid all models in model-index.yml') |
| 34 | + parser.add_argument( |
| 35 | + '--shape', |
| 36 | + type=int, |
| 37 | + nargs='+', |
| 38 | + default=[1280, 800], |
| 39 | + help='input image size') |
| 40 | + parser.add_argument( |
| 41 | + '--checkpoint_root', |
| 42 | + help='Checkpoint file root path. If set, load checkpoint before test.') |
| 43 | + parser.add_argument('--img', default='demo/demo.jpg', help='Image file') |
| 44 | + parser.add_argument('--models', nargs='+', help='models name to inference') |
| 45 | + parser.add_argument( |
| 46 | + '--batch-size', |
| 47 | + type=int, |
| 48 | + default=1, |
| 49 | + help='The batch size during the inference.') |
| 50 | + parser.add_argument( |
| 51 | + '--flops', action='store_true', help='Get Flops and Params of models') |
| 52 | + parser.add_argument( |
| 53 | + '--flops-str', |
| 54 | + action='store_true', |
| 55 | + help='Output FLOPs and params counts in a string form.') |
| 56 | + parser.add_argument( |
| 57 | + '--cfg-options', |
| 58 | + nargs='+', |
| 59 | + action=DictAction, |
| 60 | + help='override some settings in the used config, the key-value pair ' |
| 61 | + 'in xxx=yyy format will be merged into config file. If the value to ' |
| 62 | + 'be overwritten is a list, it should be like key="[a,b]" or key=a,b ' |
| 63 | + 'It also allows nested list/tuple values, e.g. key="[(a,b),(c,d)]" ' |
| 64 | + 'Note that the quotation marks are necessary and that no white space ' |
| 65 | + 'is allowed.') |
| 66 | + parser.add_argument( |
| 67 | + '--size_divisor', |
| 68 | + type=int, |
| 69 | + default=32, |
| 70 | + help='Pad the input image, the minimum size that is divisible ' |
| 71 | + 'by size_divisor, -1 means do not pad the image.') |
| 72 | + args = parser.parse_args() |
| 73 | + return args |
| 74 | + |
| 75 | + |
| 76 | +def inference(config_file, checkpoint, work_dir, args, exp_name): |
| 77 | + logger = MMLogger.get_instance(name='MMLogger') |
| 78 | + logger.warning('if you want test flops, please make sure torch>=1.12') |
| 79 | + cfg = Config.fromfile(config_file) |
| 80 | + cfg.work_dir = work_dir |
| 81 | + cfg.load_from = checkpoint |
| 82 | + cfg.log_level = 'WARN' |
| 83 | + cfg.experiment_name = exp_name |
| 84 | + if args.cfg_options is not None: |
| 85 | + cfg.merge_from_dict(args.cfg_options) |
| 86 | + init_default_scope(cfg.get('default_scope', 'mmocr')) |
| 87 | + |
| 88 | + # forward the model |
| 89 | + result = {'model': config_file.stem} |
| 90 | + |
| 91 | + if args.flops: |
| 92 | + |
| 93 | + if len(args.shape) == 1: |
| 94 | + h = w = args.shape[0] |
| 95 | + elif len(args.shape) == 2: |
| 96 | + h, w = args.shape |
| 97 | + else: |
| 98 | + raise ValueError('invalid input shape') |
| 99 | + divisor = args.size_divisor |
| 100 | + if divisor > 0: |
| 101 | + h = int(np.ceil(h / divisor)) * divisor |
| 102 | + w = int(np.ceil(w / divisor)) * divisor |
| 103 | + |
| 104 | + input_shape = (3, h, w) |
| 105 | + result['resolution'] = input_shape |
| 106 | + |
| 107 | + try: |
| 108 | + cfg = Config.fromfile(config_file) |
| 109 | + if args.cfg_options is not None: |
| 110 | + cfg.merge_from_dict(args.cfg_options) |
| 111 | + |
| 112 | + model = MODELS.build(cfg.model) |
| 113 | + input = torch.rand(1, *input_shape) |
| 114 | + if torch.cuda.is_available(): |
| 115 | + model.cuda() |
| 116 | + input = input.cuda() |
| 117 | + model = revert_sync_batchnorm(model) |
| 118 | + inputs = (input, ) |
| 119 | + model.eval() |
| 120 | + outputs = get_model_complexity_info( |
| 121 | + model, input_shape, inputs, show_table=False, show_arch=False) |
| 122 | + flops = outputs['flops'] |
| 123 | + params = outputs['params'] |
| 124 | + activations = outputs['activations'] |
| 125 | + result['Get Types'] = 'Random input' |
| 126 | + except: # noqa 772 |
| 127 | + logger = MMLogger.get_instance(name='MMLogger') |
| 128 | + logger.warning( |
| 129 | + 'Direct get flops failed, try to get flops with data') |
| 130 | + cfg = Config.fromfile(config_file) |
| 131 | + data_loader = Runner.build_dataloader(cfg.val_dataloader) |
| 132 | + data_batch = next(iter(data_loader)) |
| 133 | + model = MODELS.build(cfg.model) |
| 134 | + if torch.cuda.is_available(): |
| 135 | + model = model.cuda() |
| 136 | + model = revert_sync_batchnorm(model) |
| 137 | + model.eval() |
| 138 | + _forward = model.forward |
| 139 | + data = model.data_preprocessor(data_batch) |
| 140 | + del data_loader |
| 141 | + model.forward = partial( |
| 142 | + _forward, data_samples=data['data_samples']) |
| 143 | + outputs = get_model_complexity_info( |
| 144 | + model, |
| 145 | + input_shape, |
| 146 | + data['inputs'], |
| 147 | + show_table=False, |
| 148 | + show_arch=False) |
| 149 | + flops = outputs['flops'] |
| 150 | + params = outputs['params'] |
| 151 | + activations = outputs['activations'] |
| 152 | + result['Get Types'] = 'Dataloader' |
| 153 | + |
| 154 | + if args.flops_str: |
| 155 | + flops = _format_size(flops) |
| 156 | + params = _format_size(params) |
| 157 | + activations = _format_size(activations) |
| 158 | + |
| 159 | + result['flops'] = flops |
| 160 | + result['params'] = params |
| 161 | + |
| 162 | + return result |
| 163 | + |
| 164 | + |
| 165 | +def show_summary(summary_data, args): |
| 166 | + table = Table(title='Validation Benchmark Regression Summary') |
| 167 | + table.add_column('Model') |
| 168 | + table.add_column('Validation') |
| 169 | + table.add_column('Resolution (c, h, w)') |
| 170 | + if args.flops: |
| 171 | + table.add_column('Flops', justify='right', width=11) |
| 172 | + table.add_column('Params', justify='right') |
| 173 | + |
| 174 | + for model_name, summary in summary_data.items(): |
| 175 | + row = [model_name] |
| 176 | + valid = summary['valid'] |
| 177 | + color = 'green' if valid == 'PASS' else 'red' |
| 178 | + row.append(f'[{color}]{valid}[/{color}]') |
| 179 | + if valid == 'PASS': |
| 180 | + row.append(str(summary['resolution'])) |
| 181 | + if args.flops: |
| 182 | + row.append(str(summary['flops'])) |
| 183 | + row.append(str(summary['params'])) |
| 184 | + table.add_row(*row) |
| 185 | + |
| 186 | + console.print(table) |
| 187 | + table_data = { |
| 188 | + x.header: [Text.from_markup(y).plain for y in x.cells] |
| 189 | + for x in table.columns |
| 190 | + } |
| 191 | + table_pd = pd.DataFrame(table_data) |
| 192 | + table_pd.to_csv('./mmocr_flops.csv') |
| 193 | + |
| 194 | + |
| 195 | +# Sample test whether the inference code is correct |
| 196 | +def main(args): |
| 197 | + model_index_file = MMOCR_ROOT / 'model-index.yml' |
| 198 | + model_index = load(str(model_index_file)) |
| 199 | + model_index.build_models_with_collections() |
| 200 | + models = OrderedDict({model.name: model for model in model_index.models}) |
| 201 | + |
| 202 | + logger = MMLogger( |
| 203 | + 'validation', |
| 204 | + logger_name='validation', |
| 205 | + log_file='benchmark_test_image.log', |
| 206 | + log_level=logging.INFO) |
| 207 | + |
| 208 | + if args.models: |
| 209 | + patterns = [ |
| 210 | + re.compile(pattern.replace('+', '_')) for pattern in args.models |
| 211 | + ] |
| 212 | + filter_models = {} |
| 213 | + for k, v in models.items(): |
| 214 | + k = k.replace('+', '_') |
| 215 | + if any([re.match(pattern, k) for pattern in patterns]): |
| 216 | + filter_models[k] = v |
| 217 | + if len(filter_models) == 0: |
| 218 | + print('No model found, please specify models in:') |
| 219 | + print('\n'.join(models.keys())) |
| 220 | + return |
| 221 | + models = filter_models |
| 222 | + |
| 223 | + summary_data = {} |
| 224 | + tmpdir = tempfile.TemporaryDirectory() |
| 225 | + for model_name, model_info in tqdm(models.items()): |
| 226 | + |
| 227 | + if model_info.config is None: |
| 228 | + continue |
| 229 | + |
| 230 | + model_info.config = model_info.config.replace('%2B', '+') |
| 231 | + config = Path(model_info.config) |
| 232 | + |
| 233 | + try: |
| 234 | + config.exists() |
| 235 | + except: # noqa 722 |
| 236 | + logger.error(f'{model_name}: {config} not found.') |
| 237 | + continue |
| 238 | + |
| 239 | + logger.info(f'Processing: {model_name}') |
| 240 | + |
| 241 | + http_prefix = 'https://download.openmmlab.com/mmocr/' |
| 242 | + if args.checkpoint_root is not None: |
| 243 | + root = args.checkpoint_root |
| 244 | + if 's3://' in args.checkpoint_root: |
| 245 | + from petrel_client.common.exception import AccessDeniedError |
| 246 | + file_client = FileClient.infer_client(uri=root) |
| 247 | + checkpoint = file_client.join_path( |
| 248 | + root, model_info.weights[len(http_prefix):]) |
| 249 | + try: |
| 250 | + exists = file_client.exists(checkpoint) |
| 251 | + except AccessDeniedError: |
| 252 | + exists = False |
| 253 | + else: |
| 254 | + checkpoint = Path(root) / model_info.weights[len(http_prefix):] |
| 255 | + exists = checkpoint.exists() |
| 256 | + if exists: |
| 257 | + checkpoint = str(checkpoint) |
| 258 | + else: |
| 259 | + print(f'WARNING: {model_name}: {checkpoint} not found.') |
| 260 | + checkpoint = None |
| 261 | + else: |
| 262 | + checkpoint = None |
| 263 | + |
| 264 | + try: |
| 265 | + # build the model from a config file and a checkpoint file |
| 266 | + result = inference(MMOCR_ROOT / config, checkpoint, tmpdir.name, |
| 267 | + args, model_name) |
| 268 | + result['valid'] = 'PASS' |
| 269 | + except Exception: # noqa 722 |
| 270 | + import traceback |
| 271 | + logger.error(f'"{config}" :\n{traceback.format_exc()}') |
| 272 | + result = {'valid': 'FAIL'} |
| 273 | + |
| 274 | + summary_data[model_name] = result |
| 275 | + |
| 276 | + tmpdir.cleanup() |
| 277 | + show_summary(summary_data, args) |
| 278 | + |
| 279 | + |
| 280 | +if __name__ == '__main__': |
| 281 | + args = parse_args() |
| 282 | + main(args) |
0 commit comments