输出结果

设计思路

https://yunyaniu.blog.csdn.net/article/details/80330637

核心代码

1、train.py

#!/usr/bin/env python

"""

Copyright 2017-2018 Fizyr (https://fizyr.com)

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 argparse

import os

import sys

import keras

import keras.preprocessing.image

import tensorflow as tf

import keras_retinanet.losses

from keras_retinanet.callbacks import RedirectModel

from keras_retinanet.utils.config import read_config_file, parse_anchor_parameters

from keras_retinanet.utils.transform import random_transform_generator

from keras_retinanet.utils.keras_version import check_keras_version

from keras_retinanet.utils.model import freeze as freeze_model

# Allow relative imports when being executed as script.

if __name__ == "__main__" and __package__ is None:

   sys.path.insert(0, os.path.join(os.path.dirname(__file__), '..', '..'))

   import keras_maskrcnn.bin

   __package__ = "keras_maskrcnn.bin"

# Change these to absolute imports if you copy this script outside the keras_retinanet package.

from .. import losses

from .. import models

from ..callbacks.eval import Evaluate

def get_session():

   config = tf.ConfigProto()

   config.gpu_options.allow_growth = True

   return tf.Session(config=config)

def model_with_weights(model, weights, skip_mismatch):

   if weights is not None:

       model.load_weights(weights, by_name=True, skip_mismatch=skip_mismatch)

   return model

def create_models(backbone_retinanet, num_classes, weights, freeze_backbone=False, class_specific_filter=True, anchor_params=None):

   modifier = freeze_model if freeze_backbone else None

   model            = model_with_weights(

       backbone_retinanet(

           num_classes,

           nms=True,

           class_specific_filter=class_specific_filter,

           modifier=modifier,

           anchor_params=anchor_params

       ), weights=weights, skip_mismatch=True)

   training_model   = model

   prediction_model = model

   # compile model

   training_model.compile(

       loss={

           'regression'    : keras_retinanet.losses.smooth_l1(),

           'classification': keras_retinanet.losses.focal(),

           'masks'         : losses.mask(),

       },

       optimizer=keras.optimizers.adam(lr=1e-5, clipnorm=0.001)

   )

   return model, training_model, prediction_model

def create_callbacks(model, training_model, prediction_model, validation_generator, args):

   callbacks = []

   # save the prediction model

   if args.snapshots:

       # ensure directory created first; otherwise h5py will error after epoch.

       os.makedirs(args.snapshot_path, exist_ok=True)

       checkpoint = keras.callbacks.ModelCheckpoint(

           os.path.join(

               args.snapshot_path,

               '{backbone}_{dataset_type}_{{epoch:02d}}.h5'.format(backbone=args.backbone, dataset_type=args.dataset_type)

           ),

           verbose=1

       )

       checkpoint = RedirectModel(checkpoint, prediction_model)

       callbacks.append(checkpoint)

   tensorboard_callback = None

   if args.tensorboard_dir:

       tensorboard_callback = keras.callbacks.TensorBoard(

           log_dir                = args.tensorboard_dir,

           histogram_freq         = 0,

           batch_size             = args.batch_size,

           write_graph            = True,

           write_grads            = False,

           write_images           = False,

           embeddings_freq        = 0,

           embeddings_layer_names = None,

           embeddings_metadata    = None

       )

       callbacks.append(tensorboard_callback)

   if args.evaluation and validation_generator:

       if args.dataset_type == 'coco':

           from ..callbacks.coco import CocoEval

           # use prediction model for evaluation

           evaluation = CocoEval(validation_generator)

       else:

           evaluation = Evaluate(validation_generator, tensorboard=tensorboard_callback, weighted_average=args.weighted_average)

       evaluation = RedirectModel(evaluation, prediction_model)

       callbacks.append(evaluation)

   callbacks.append(keras.callbacks.ReduceLROnPlateau(

       monitor  = 'loss',

       factor   = 0.1,

       patience = 2,

       verbose  = 1,

       mode     = 'auto',

       epsilon  = 0.0001,

       cooldown = 0,

       min_lr   = 0

   ))

   return callbacks

def create_generators(args):

   # create random transform generator for augmenting training data

   transform_generator = random_transform_generator(flip_x_chance=0.5)

   if args.dataset_type == 'coco':

       # import here to prevent unnecessary dependency on cocoapi

       from ..preprocessing.coco import CocoGenerator

       train_generator = CocoGenerator(

           args.coco_path,

           'train2017',

           transform_generator=transform_generator,

           batch_size=args.batch_size,

           config=args.config

       )

       validation_generator = CocoGenerator(

           args.coco_path,

           'val2017',

           batch_size=args.batch_size,

           config=args.config

       )

   elif args.dataset_type == 'csv':

       from ..preprocessing.csv_generator import CSVGenerator

       train_generator = CSVGenerator(

           args.annotations,

           args.classes,

           transform_generator=transform_generator,

           batch_size=args.batch_size,

           config=args.config

       )

       if args.val_annotations:

           validation_generator = CSVGenerator(

               args.val_annotations,

               args.classes,

               batch_size=args.batch_size,

               config=args.config

           )

       else:

           validation_generator = None

   else:

       raise ValueError('Invalid data type received: {}'.format(args.dataset_type))

   return train_generator, validation_generator

def check_args(parsed_args):

   """

   Function to check for inherent contradictions within parsed arguments.

   For example, batch_size < num_gpus

   Intended to raise errors prior to backend initialisation.

   :param parsed_args: parser.parse_args()

   :return: parsed_args

   """

   return parsed_args

def parse_args(args):

   parser     = argparse.ArgumentParser(description='Simple training script for training a RetinaNet mask network.')

   subparsers = parser.add_subparsers(help='Arguments for specific dataset types.', dest='dataset_type')

   subparsers.required = True

   coco_parser = subparsers.add_parser('coco')

   coco_parser.add_argument('coco_path', help='Path to dataset directory (ie. /tmp/COCO).')

   csv_parser = subparsers.add_parser('csv')

   csv_parser.add_argument('annotations', help='Path to CSV file containing annotations for training.')

   csv_parser.add_argument('classes', help='Path to a CSV file containing class label mapping.')

   csv_parser.add_argument('--val-annotations', help='Path to CSV file containing annotations for validation (optional).')

   group = parser.add_mutually_exclusive_group()

   group.add_argument('--snapshot',          help='Resume training from a snapshot.')

   group.add_argument('--imagenet-weights',  help='Initialize the model with pretrained imagenet weights. This is the default behaviour.', action='store_const', const=True, default=True)

   group.add_argument('--weights',           help='Initialize the model with weights from a file.')

   group.add_argument('--no-weights',        help='Don\'t initialize the model with any weights.', dest='imagenet_weights', action='store_const', const=False)

   parser.add_argument('--backbone',         help='Backbone model used by retinanet.', default='resnet50', type=str)

   parser.add_argument('--batch-size',       help='Size of the batches.', default=1, type=int)

   parser.add_argument('--gpu',              help='Id of the GPU to use (as reported by nvidia-smi).')

   parser.add_argument('--epochs',           help='Number of epochs to train.', type=int, default=50)

   parser.add_argument('--steps',            help='Number of steps per epoch.', type=int, default=10000)

   parser.add_argument('--snapshot-path',    help='Path to store snapshots of models during training (defaults to \'./snapshots\')', default='./snapshots')

   parser.add_argument('--tensorboard-dir',  help='Log directory for Tensorboard output', default='./logs')

   parser.add_argument('--no-snapshots',     help='Disable saving snapshots.', dest='snapshots', action='store_false')

   parser.add_argument('--no-evaluation',    help='Disable per epoch evaluation.', dest='evaluation', action='store_false')

   parser.add_argument('--freeze-backbone',  help='Freeze training of backbone layers.', action='store_true')

   parser.add_argument('--no-class-specific-filter', help='Disables class specific filtering.', dest='class_specific_filter', action='store_false')

   parser.add_argument('--config',           help='Path to a configuration parameters .ini file.')

   parser.add_argument('--weighted-average', help='Compute the mAP using the weighted average of precisions among classes.', action='store_true')

   return check_args(parser.parse_args(args))

def main(args=None):

   # parse arguments

   if args is None:

       args = sys.argv[1:]

   args = parse_args(args)

   # make sure keras is the minimum required version

   check_keras_version()

   # create object that stores backbone information

   backbone = models.backbone(args.backbone)

   # optionally choose specific GPU

   if args.gpu:

       os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu

   keras.backend.tensorflow_backend.set_session(get_session())

   # optionally load config parameters

   if args.config:

       args.config = read_config_file(args.config)

   # create the generators

   train_generator, validation_generator = create_generators(args)

   # create the model

   if args.snapshot is not None:

       print('Loading model, this may take a second...')

       model            = models.load_model(args.snapshot, backbone_name=args.backbone)

       training_model   = model

       prediction_model = model

   else:

       weights = args.weights

       # default to imagenet if nothing else is specified

       if weights is None and args.imagenet_weights:

           weights = backbone.download_imagenet()

       anchor_params = None

       if args.config and 'anchor_parameters' in args.config:

           anchor_params = parse_anchor_parameters(args.config)

       print('Creating model, this may take a second...')

       model, training_model, prediction_model = create_models(

           backbone_retinanet=backbone.maskrcnn,

           num_classes=train_generator.num_classes(),

           weights=weights,

           freeze_backbone=args.freeze_backbone,

           class_specific_filter=args.class_specific_filter,

           anchor_params=anchor_params

       )

   # print model summary

   print(model.summary())

   # create the callbacks

   callbacks = create_callbacks(

       model,

       training_model,

       prediction_model,

       validation_generator,

       args,

   )

   # start training

   training_model.fit_generator(

       generator=train_generator,

       steps_per_epoch=args.steps,

       epochs=args.epochs,

       verbose=1,

       callbacks=callbacks,

       max_queue_size=1,

   )

if __name__ == '__main__':

   main()