# Copyright (c) 2024 PaddlePaddle Authors. 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. import os from ...base import BaseRunner from ...base.utils.arg import gather_opts_args from ...base.utils.subprocess import CompletedProcess class SegRunner(BaseRunner): """Semantic Segmentation Runner""" def train( self, config_path: str, cli_args: list, device: str, ips: str, save_dir: str, do_eval=True, ) -> CompletedProcess: """train model Args: config_path (str): the config file path used to train. cli_args (list): the additional parameters. device (str): the training device. ips (str): the ip addresses of nodes when using distribution. save_dir (str): the directory path to save training output. do_eval (bool, optional): whether or not to evaluate model during training. Defaults to True. Returns: CompletedProcess: the result of training subprocess execution. """ args, env = self.distributed(device, ips, log_dir=save_dir) cli_args = self._gather_opts_args(cli_args) cmd = [*args, "tools/train.py"] if do_eval: cmd.append("--do_eval") cmd.extend(["--config", config_path, *cli_args]) return self.run_cmd( cmd, env=env, switch_wdir=True, echo=True, silent=False, capture_output=True, log_path=self._get_train_log_path(save_dir), ) def evaluate( self, config_path: str, cli_args: list, device: str, ips: str ) -> CompletedProcess: """run model evaluating Args: config_path (str): the config file path used to evaluate. cli_args (list): the additional parameters. device (str): the evaluating device. ips (str): the ip addresses of nodes when using distribution. Returns: CompletedProcess: the result of evaluating subprocess execution. """ args, env = self.distributed(device, ips) cli_args = self._gather_opts_args(cli_args) cmd = [*args, "tools/val.py", "--config", config_path, *cli_args] cp = self.run_cmd( cmd, env=env, switch_wdir=True, echo=True, silent=False, capture_output=True ) if cp.returncode == 0: metric_dict = _extract_eval_metrics(cp.stdout) cp.metrics = metric_dict return cp def predict( self, config_path: str, cli_args: list, device: str ) -> CompletedProcess: """run predicting using dynamic mode Args: config_path (str): the config file path used to predict. cli_args (list): the additional parameters. device (str): unused. Returns: CompletedProcess: the result of predicting subprocess execution. """ # `device` unused cli_args = self._gather_opts_args(cli_args) cmd = [self.python, "tools/predict.py", "--config", config_path, *cli_args] return self.run_cmd(cmd, switch_wdir=True, echo=True, silent=False) def analyse(self, config_path, cli_args, device, ips): """analyse""" args, env = self.distributed(device, ips) cli_args = self._gather_opts_args(cli_args) cmd = [*args, "tools/analyse.py", "--config", config_path, *cli_args] cp = self.run_cmd( cmd, env=env, switch_wdir=True, echo=True, silent=False, capture_output=True ) return cp def export(self, config_path: str, cli_args: list, device: str) -> CompletedProcess: """run exporting Args: config_path (str): the path of config file used to export. cli_args (list): the additional parameters. device (str): unused. Returns: CompletedProcess: the result of exporting subprocess execution. """ # `device` unused cli_args = self._gather_opts_args(cli_args) cmd = [ self.python, "tools/export.py", "--for_fd", "--config", config_path, *cli_args, ] cp = self.run_cmd(cmd, switch_wdir=True, echo=True, silent=False) return cp def infer(self, config_path: str, cli_args: list, device: str) -> CompletedProcess: """run predicting using inference model Args: config_path (str): the path of config file used to predict. cli_args (list): the additional parameters. device (str): unused. Returns: CompletedProcess: the result of inferring subprocess execution. """ # `device` unused cli_args = self._gather_opts_args(cli_args) cmd = [ self.python, "deploy/python/infer.py", "--config", config_path, *cli_args, ] return self.run_cmd(cmd, switch_wdir=True, echo=True, silent=False) def compression( self, config_path: str, train_cli_args: list, export_cli_args: list, device: str, train_save_dir: str, ) -> CompletedProcess: """run compression model Args: config_path (str): the path of config file used to predict. train_cli_args (list): the additional training parameters. export_cli_args (list): the additional exporting parameters. device (str): the running device. train_save_dir (str): the directory path to save output. Returns: CompletedProcess: the result of compression subprocess execution. """ # Step 1: Train model args, env = self.distributed(device, log_dir=train_save_dir) train_cli_args = self._gather_opts_args(train_cli_args) # Note that we add `--do_eval` here so we can have `train_save_dir/best_model/model.pdparams` saved cmd = [ *args, "deploy/slim/quant/qat_train.py", "--do_eval", "--config", config_path, *train_cli_args, ] cp_train = self.run_cmd( cmd, env=env, switch_wdir=True, echo=True, silent=False, capture_output=True, log_path=self._get_train_log_path(train_save_dir), ) # Step 2: Export model export_cli_args = self._gather_opts_args(export_cli_args) # We export the best model on the validation dataset weight_path = os.path.join(train_save_dir, "best_model", "model.pdparams") cmd = [ self.python, "deploy/slim/quant/qat_export.py", "--for_fd", "--config", config_path, "--model_path", weight_path, *export_cli_args, ] cp_export = self.run_cmd(cmd, switch_wdir=True, echo=True, silent=False) return cp_train, cp_export def _gather_opts_args(self, args): # Since `--opts` in PaddleSeg does not use `action='append'` # We collect and arrange all opts args here # e.g.: python tools/train.py --config xxx --opts a=1 c=3 --opts b=2 # => python tools/train.py --config xxx c=3 --opts a=1 b=2 return gather_opts_args(args, "--opts") def _extract_eval_metrics(stdout: str) -> dict: """extract evaluation metrics from training log Args: stdout (str): the training log Returns: dict: the training metric """ import re _DP = r"[-+]?[0-9]*\.?[0-9]+(?:[eE][-+]?[0-9]+)?" pattern = r"Images: \d+ mIoU: (_dp) Acc: (_dp) Kappa: (_dp) Dice: (_dp)".replace( "_dp", _DP ) keys = ["mIoU", "Acc", "Kappa", "Dice"] metric_dict = dict() pattern = re.compile(pattern) # TODO: Use lazy version to make it more efficient lines = stdout.splitlines() for line in lines: match = pattern.search(line) if match: for k, v in zip(keys, map(float, match.groups())): metric_dict[k] = v return metric_dict