 在 TensorFlow.org 上查看
|
 在 Google Colab 中运行
|
 在 GitHub 上查看
|
 下载笔记本
|
本教程演示了如何对来自 TensorFlow Model Garden 软件包 (tensorflow-models) 的以 ResNet-50 为骨干网的 RetinaNet 进行微调，以检测 BCCD 数据集中的三种不同血细胞。该 RetinaNet 在 COCO train2017 上进行了预训练，并在 COCO val2017 上进行了评估。
Model Garden 包含一系列使用 TensorFlow 高级 API 实现的最先进模型。这些实现展示了建模的最佳实践，让用户能够充分利用 TensorFlow 进行研究和产品开发。
本教程将演示如何
- 使用来自 Tensorflow Model Garden (TFM) 软件包的模型。
- 微调一个以 ResNet-50 为骨干网的预训练 RetinaNet 模型，用于目标检测。
- 导出已调优的 RetinaNet 模型
安装必要的依赖项
pip install -U -q "tf-models-official"导入所需的库
import os
import io
import pprint
import tempfile
import matplotlib
import numpy as np
import tensorflow as tf
import matplotlib.pyplot as plt
from PIL import Image
from six import BytesIO
from IPython import display
from urllib.request import urlopen
2023-11-09 12:15:18.455434: E tensorflow/compiler/xla/stream_executor/cuda/cuda_dnn.cc:9342] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered 2023-11-09 12:15:18.455488: E tensorflow/compiler/xla/stream_executor/cuda/cuda_fft.cc:609] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered 2023-11-09 12:15:18.455537: E tensorflow/compiler/xla/stream_executor/cuda/cuda_blas.cc:1518] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered
从 tensorflow models 中导入所需的库
import orbit
import tensorflow_models as tfm
from official.core import exp_factory
from official.core import config_definitions as cfg
from official.vision.serving import export_saved_model_lib
from official.vision.ops.preprocess_ops import normalize_image
from official.vision.ops.preprocess_ops import resize_and_crop_image
from official.vision.utils.object_detection import visualization_utils
from official.vision.dataloaders.tf_example_decoder import TfExampleDecoder
pp = pprint.PrettyPrinter(indent=4) # Set Pretty Print Indentation
print(tf.__version__) # Check the version of tensorflow used
%matplotlib inline
2.14.0
为目标检测准备自定义数据集
官方仓库（model-garden）中的模型需要 TFRecords 格式的数据。
请查阅此资源以了解更多关于 TFRecords 数据格式的信息。
将您的自定义数据上传到 Notebook 的云端硬盘或本地磁盘并解压数据
curl -L 'https://public.roboflow.com/ds/ZpYLqHeT0W?key=ZXfZLRnhsc' > './BCCD.v1-bccd.coco.zip'unzip -q -o './BCCD.v1-bccd.coco.zip' -d './BCC.v1-bccd.coco/'rm './BCCD.v1-bccd.coco.zip'
% Total % Received % Xferd Average Speed Time Time Time Current
Dload Upload Total Spent Left Speed
100 892 100 892 0 0 3391 0 --:--:-- --:--:-- --:--:-- 3391
100 15.2M 100 15.2M 0 0 28.7M 0 --:--:-- --:--:-- --:--:-- 288M
用于转换数据（训练数据）的 CLI 命令。
TRAIN_DATA_DIR='./BCC.v1-bccd.coco/train'TRAIN_ANNOTATION_FILE_DIR='./BCC.v1-bccd.coco/train/_annotations.coco.json'OUTPUT_TFRECORD_TRAIN='./bccd_coco_tfrecords/train'# Need to provide# 1. image_dir: where images are present# 2. object_annotations_file: where annotations are listed in json format# 3. output_file_prefix: where to write output convered TFRecords filespython -m official.vision.data.create_coco_tf_record --logtostderr \--image_dir=${TRAIN_DATA_DIR} \--object_annotations_file=${TRAIN_ANNOTATION_FILE_DIR} \--output_file_prefix=$OUTPUT_TFRECORD_TRAIN \--num_shards=1
2023-11-09 12:15:24.203902: E tensorflow/compiler/xla/stream_executor/cuda/cuda_dnn.cc:9342] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered 2023-11-09 12:15:24.203959: E tensorflow/compiler/xla/stream_executor/cuda/cuda_fft.cc:609] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered 2023-11-09 12:15:24.203987: E tensorflow/compiler/xla/stream_executor/cuda/cuda_blas.cc:1518] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered 2023-11-09 12:15:27.899086: W tensorflow/core/common_runtime/gpu/gpu_device.cc:2211] Cannot dlopen some GPU libraries. Please make sure the missing libraries mentioned above are installed properly if you would like to use GPU. Follow the guide at https://tensorflowcn.cn/install/gpu for how to download and setup the required libraries for your platform. Skipping registering GPU devices... I1109 12:15:27.900500 140558803519296 create_coco_tf_record.py:502] writing to output path: ./bccd_coco_tfrecords/train I1109 12:15:28.110808 140558803519296 create_coco_tf_record.py:374] Building bounding box index. I1109 12:15:28.112704 140558803519296 create_coco_tf_record.py:385] 0 images are missing bboxes. I1109 12:15:28.376975 140558803519296 tfrecord_lib.py:168] On image 0 I1109 12:15:28.383184 140558803519296 tfrecord_lib.py:168] On image 100 I1109 12:15:28.388326 140558803519296 tfrecord_lib.py:168] On image 200 I1109 12:15:28.393295 140558803519296 tfrecord_lib.py:168] On image 300 I1109 12:15:28.398152 140558803519296 tfrecord_lib.py:168] On image 400 I1109 12:15:28.403067 140558803519296 tfrecord_lib.py:168] On image 500 I1109 12:15:28.407919 140558803519296 tfrecord_lib.py:168] On image 600 I1109 12:15:28.412595 140558803519296 tfrecord_lib.py:168] On image 700 I1109 12:15:28.438800 140558803519296 tfrecord_lib.py:180] Finished writing, skipped 6 annotations. I1109 12:15:28.445898 140558803519296 create_coco_tf_record.py:537] Finished writing, skipped 6 annotations.
用于转换数据（验证数据）的 CLI 命令。
VALID_DATA_DIR='./BCC.v1-bccd.coco/valid'VALID_ANNOTATION_FILE_DIR='./BCC.v1-bccd.coco/valid/_annotations.coco.json'OUTPUT_TFRECORD_VALID='./bccd_coco_tfrecords/valid'python -m official.vision.data.create_coco_tf_record --logtostderr \--image_dir=$VALID_DATA_DIR \--object_annotations_file=$VALID_ANNOTATION_FILE_DIR \--output_file_prefix=$OUTPUT_TFRECORD_VALID \--num_shards=1
2023-11-09 12:15:29.695864: E tensorflow/compiler/xla/stream_executor/cuda/cuda_dnn.cc:9342] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered 2023-11-09 12:15:29.695909: E tensorflow/compiler/xla/stream_executor/cuda/cuda_fft.cc:609] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered 2023-11-09 12:15:29.695940: E tensorflow/compiler/xla/stream_executor/cuda/cuda_blas.cc:1518] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered 2023-11-09 12:15:33.418966: W tensorflow/core/common_runtime/gpu/gpu_device.cc:2211] Cannot dlopen some GPU libraries. Please make sure the missing libraries mentioned above are installed properly if you would like to use GPU. Follow the guide at https://tensorflowcn.cn/install/gpu for how to download and setup the required libraries for your platform. Skipping registering GPU devices... I1109 12:15:33.419316 140466682771264 create_coco_tf_record.py:502] writing to output path: ./bccd_coco_tfrecords/valid I1109 12:15:33.426927 140466682771264 create_coco_tf_record.py:374] Building bounding box index. I1109 12:15:33.427194 140466682771264 create_coco_tf_record.py:385] 0 images are missing bboxes. I1109 12:15:33.632256 140466682771264 tfrecord_lib.py:168] On image 0 I1109 12:15:33.659518 140466682771264 tfrecord_lib.py:180] Finished writing, skipped 0 annotations. I1109 12:15:33.660955 140466682771264 create_coco_tf_record.py:537] Finished writing, skipped 0 annotations.
用于转换数据（测试数据）的 CLI 命令。
TEST_DATA_DIR='./BCC.v1-bccd.coco/test'TEST_ANNOTATION_FILE_DIR='./BCC.v1-bccd.coco/test/_annotations.coco.json'OUTPUT_TFRECORD_TEST='./bccd_coco_tfrecords/test'python -m official.vision.data.create_coco_tf_record --logtostderr \--image_dir=$TEST_DATA_DIR \--object_annotations_file=$TEST_ANNOTATION_FILE_DIR \--output_file_prefix=$OUTPUT_TFRECORD_TEST \--num_shards=1
2023-11-09 12:15:34.949263: E tensorflow/compiler/xla/stream_executor/cuda/cuda_dnn.cc:9342] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered 2023-11-09 12:15:34.949311: E tensorflow/compiler/xla/stream_executor/cuda/cuda_fft.cc:609] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered 2023-11-09 12:15:34.949341: E tensorflow/compiler/xla/stream_executor/cuda/cuda_blas.cc:1518] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered 2023-11-09 12:15:38.592568: W tensorflow/core/common_runtime/gpu/gpu_device.cc:2211] Cannot dlopen some GPU libraries. Please make sure the missing libraries mentioned above are installed properly if you would like to use GPU. Follow the guide at https://tensorflowcn.cn/install/gpu for how to download and setup the required libraries for your platform. Skipping registering GPU devices... I1109 12:15:38.592885 140186454914880 create_coco_tf_record.py:502] writing to output path: ./bccd_coco_tfrecords/test I1109 12:15:38.596868 140186454914880 create_coco_tf_record.py:374] Building bounding box index. I1109 12:15:38.597061 140186454914880 create_coco_tf_record.py:385] 0 images are missing bboxes. I1109 12:15:38.788634 140186454914880 tfrecord_lib.py:168] On image 0 I1109 12:15:38.814095 140186454914880 tfrecord_lib.py:180] Finished writing, skipped 0 annotations. I1109 12:15:38.815249 140186454914880 create_coco_tf_record.py:537] Finished writing, skipped 0 annotations.
为自定义数据集配置 Retinanet Resnet FPN COCO 模型。
用于微调检查点的数据集是血细胞检测 (BCCD) 数据集。
train_data_input_path = './bccd_coco_tfrecords/train-00000-of-00001.tfrecord'
valid_data_input_path = './bccd_coco_tfrecords/valid-00000-of-00001.tfrecord'
test_data_input_path = './bccd_coco_tfrecords/test-00000-of-00001.tfrecord'
model_dir = './trained_model/'
export_dir ='./exported_model/'
在 Model Garden 中，定义模型的参数集合称为*配置（configs）*。Model Garden 可以通过工厂模式根据一组已知参数创建配置。
使用由 tfm.vision.configs.retinanet.retinanet_resnetfpn_coco 定义的 retinanet_resnetfpn_coco 实验配置。
该配置定义了一个实验，旨在训练一个以 Resnet-50 为骨干网、FPN 为解码器的 Retinanet。默认配置是在 COCO train2017 上训练，并在 COCO val2017 上评估的。
还有其他可选实验，例如 retinanet_resnetfpn_coco、retinanet_spinenet_coco、fasterrcnn_resnetfpn_coco 等。可以通过更改 get_exp_config 函数中的实验名称参数来切换这些实验。
我们将对默认配置中已包含的 Resnet-50 骨干网检查点进行微调。
exp_config = exp_factory.get_exp_config('retinanet_resnetfpn_coco')
调整模型和数据集配置，使其适用于自定义数据集（在本例中为 BCCD）。
batch_size = 8
num_classes = 3
HEIGHT, WIDTH = 256, 256
IMG_SIZE = [HEIGHT, WIDTH, 3]
# Backbone config.
exp_config.task.freeze_backbone = False
exp_config.task.annotation_file = ''
# Model config.
exp_config.task.model.input_size = IMG_SIZE
exp_config.task.model.num_classes = num_classes + 1
exp_config.task.model.detection_generator.tflite_post_processing.max_classes_per_detection = exp_config.task.model.num_classes
# Training data config.
exp_config.task.train_data.input_path = train_data_input_path
exp_config.task.train_data.dtype = 'float32'
exp_config.task.train_data.global_batch_size = batch_size
exp_config.task.train_data.parser.aug_scale_max = 1.0
exp_config.task.train_data.parser.aug_scale_min = 1.0
# Validation data config.
exp_config.task.validation_data.input_path = valid_data_input_path
exp_config.task.validation_data.dtype = 'float32'
exp_config.task.validation_data.global_batch_size = batch_size
调整训练器配置。
logical_device_names = [logical_device.name for logical_device in tf.config.list_logical_devices()]
if 'GPU' in ''.join(logical_device_names):
print('This may be broken in Colab.')
device = 'GPU'
elif 'TPU' in ''.join(logical_device_names):
print('This may be broken in Colab.')
device = 'TPU'
else:
print('Running on CPU is slow, so only train for a few steps.')
device = 'CPU'
train_steps = 1000
exp_config.trainer.steps_per_loop = 100 # steps_per_loop = num_of_training_examples // train_batch_size
exp_config.trainer.summary_interval = 100
exp_config.trainer.checkpoint_interval = 100
exp_config.trainer.validation_interval = 100
exp_config.trainer.validation_steps = 100 # validation_steps = num_of_validation_examples // eval_batch_size
exp_config.trainer.train_steps = train_steps
exp_config.trainer.optimizer_config.warmup.linear.warmup_steps = 100
exp_config.trainer.optimizer_config.learning_rate.type = 'cosine'
exp_config.trainer.optimizer_config.learning_rate.cosine.decay_steps = train_steps
exp_config.trainer.optimizer_config.learning_rate.cosine.initial_learning_rate = 0.1
exp_config.trainer.optimizer_config.warmup.linear.warmup_learning_rate = 0.05
Running on CPU is slow, so only train for a few steps. 2023-11-09 12:15:40.194897: W tensorflow/core/common_runtime/gpu/gpu_device.cc:2211] Cannot dlopen some GPU libraries. Please make sure the missing libraries mentioned above are installed properly if you would like to use GPU. Follow the guide at https://tensorflowcn.cn/install/gpu for how to download and setup the required libraries for your platform. Skipping registering GPU devices...
打印修改后的配置。
pp.pprint(exp_config.as_dict())
display.Javascript('google.colab.output.setIframeHeight("500px");')
{ 'runtime': { 'all_reduce_alg': None,
'batchnorm_spatial_persistent': False,
'dataset_num_private_threads': None,
'default_shard_dim': -1,
'distribution_strategy': 'mirrored',
'enable_xla': False,
'gpu_thread_mode': None,
'loss_scale': None,
'mixed_precision_dtype': 'bfloat16',
'num_cores_per_replica': 1,
'num_gpus': 0,
'num_packs': 1,
'per_gpu_thread_count': 0,
'run_eagerly': False,
'task_index': -1,
'tpu': None,
'tpu_enable_xla_dynamic_padder': None,
'use_tpu_mp_strategy': False,
'worker_hosts': None},
'task': { 'allow_image_summary': False,
'annotation_file': '',
'differential_privacy_config': None,
'export_config': { 'cast_detection_classes_to_float': False,
'cast_num_detections_to_float': False,
'output_intermediate_features': False,
'output_normalized_coordinates': False},
'freeze_backbone': False,
'init_checkpoint': 'gs://cloud-tpu-checkpoints/vision-2.0/resnet50_imagenet/ckpt-28080',
'init_checkpoint_modules': 'backbone',
'losses': { 'box_loss_weight': 50,
'focal_loss_alpha': 0.25,
'focal_loss_gamma': 1.5,
'huber_loss_delta': 0.1,
'l2_weight_decay': 0.0001,
'loss_weight': 1.0},
'max_num_eval_detections': 100,
'model': { 'anchor': { 'anchor_size': 4.0,
'aspect_ratios': [0.5, 1.0, 2.0],
'num_scales': 3},
'backbone': { 'resnet': { 'bn_trainable': True,
'depth_multiplier': 1.0,
'model_id': 50,
'replace_stem_max_pool': False,
'resnetd_shortcut': False,
'scale_stem': True,
'se_ratio': 0.0,
'stem_type': 'v0',
'stochastic_depth_drop_rate': 0.0},
'type': 'resnet'},
'decoder': { 'fpn': { 'fusion_type': 'sum',
'num_filters': 256,
'use_keras_layer': False,
'use_separable_conv': False},
'type': 'fpn'},
'detection_generator': { 'apply_nms': True,
'box_coder_weights': None,
'max_num_detections': 100,
'nms_iou_threshold': 0.5,
'nms_version': 'v2',
'pre_nms_score_threshold': 0.05,
'pre_nms_top_k': 5000,
'return_decoded': None,
'soft_nms_sigma': None,
'tflite_post_processing': { 'max_classes_per_detection': 4,
'max_detections': 200,
'nms_iou_threshold': 0.5,
'nms_score_threshold': 0.1,
'normalize_anchor_coordinates': False,
'omit_nms': False,
'use_regular_nms': False},
'use_class_agnostic_nms': False,
'use_cpu_nms': False},
'head': { 'attribute_heads': [],
'num_convs': 4,
'num_filters': 256,
'share_classification_heads': False,
'share_level_convs': True,
'use_separable_conv': False},
'input_size': [256, 256, 3],
'max_level': 7,
'min_level': 3,
'norm_activation': { 'activation': 'relu',
'norm_epsilon': 0.001,
'norm_momentum': 0.99,
'use_sync_bn': False},
'num_classes': 4},
'name': None,
'per_category_metrics': False,
'train_data': { 'apply_tf_data_service_before_batching': False,
'autotune_algorithm': None,
'block_length': 1,
'cache': False,
'cycle_length': None,
'decoder': { 'simple_decoder': { 'attribute_names': [ ],
'mask_binarize_threshold': None,
'regenerate_source_id': False},
'type': 'simple_decoder'},
'deterministic': None,
'drop_remainder': True,
'dtype': 'float32',
'enable_shared_tf_data_service_between_parallel_trainers': False,
'enable_tf_data_service': False,
'file_type': 'tfrecord',
'global_batch_size': 8,
'input_path': './bccd_coco_tfrecords/train-00000-of-00001.tfrecord',
'is_training': True,
'parser': { 'aug_policy': None,
'aug_rand_hflip': True,
'aug_scale_max': 1.0,
'aug_scale_min': 1.0,
'aug_type': None,
'match_threshold': 0.5,
'max_num_instances': 100,
'num_channels': 3,
'skip_crowd_during_training': True,
'unmatched_threshold': 0.5},
'prefetch_buffer_size': None,
'seed': None,
'sharding': True,
'shuffle_buffer_size': 10000,
'tf_data_service_address': None,
'tf_data_service_job_name': None,
'tfds_as_supervised': False,
'tfds_data_dir': '',
'tfds_name': '',
'tfds_skip_decoding_feature': '',
'tfds_split': '',
'trainer_id': None,
'weights': None},
'use_coco_metrics': True,
'use_wod_metrics': False,
'validation_data': { 'apply_tf_data_service_before_batching': False,
'autotune_algorithm': None,
'block_length': 1,
'cache': False,
'cycle_length': None,
'decoder': { 'simple_decoder': { 'attribute_names': [ ],
'mask_binarize_threshold': None,
'regenerate_source_id': False},
'type': 'simple_decoder'},
'deterministic': None,
'drop_remainder': True,
'dtype': 'float32',
'enable_shared_tf_data_service_between_parallel_trainers': False,
'enable_tf_data_service': False,
'file_type': 'tfrecord',
'global_batch_size': 8,
'input_path': './bccd_coco_tfrecords/valid-00000-of-00001.tfrecord',
'is_training': False,
'parser': { 'aug_policy': None,
'aug_rand_hflip': False,
'aug_scale_max': 1.0,
'aug_scale_min': 1.0,
'aug_type': None,
'match_threshold': 0.5,
'max_num_instances': 100,
'num_channels': 3,
'skip_crowd_during_training': True,
'unmatched_threshold': 0.5},
'prefetch_buffer_size': None,
'seed': None,
'sharding': True,
'shuffle_buffer_size': 10000,
'tf_data_service_address': None,
'tf_data_service_job_name': None,
'tfds_as_supervised': False,
'tfds_data_dir': '',
'tfds_name': '',
'tfds_skip_decoding_feature': '',
'tfds_split': '',
'trainer_id': None,
'weights': None} },
'trainer': { 'allow_tpu_summary': False,
'best_checkpoint_eval_metric': '',
'best_checkpoint_export_subdir': '',
'best_checkpoint_metric_comp': 'higher',
'checkpoint_interval': 100,
'continuous_eval_timeout': 3600,
'eval_tf_function': True,
'eval_tf_while_loop': False,
'loss_upper_bound': 1000000.0,
'max_to_keep': 5,
'optimizer_config': { 'ema': None,
'learning_rate': { 'cosine': { 'alpha': 0.0,
'decay_steps': 1000,
'initial_learning_rate': 0.1,
'name': 'CosineDecay',
'offset': 0},
'type': 'cosine'},
'optimizer': { 'sgd': { 'clipnorm': None,
'clipvalue': None,
'decay': 0.0,
'global_clipnorm': None,
'momentum': 0.9,
'name': 'SGD',
'nesterov': False},
'type': 'sgd'},
'warmup': { 'linear': { 'name': 'linear',
'warmup_learning_rate': 0.05,
'warmup_steps': 100},
'type': 'linear'} },
'preemption_on_demand_checkpoint': True,
'recovery_begin_steps': 0,
'recovery_max_trials': 0,
'steps_per_loop': 100,
'summary_interval': 100,
'train_steps': 1000,
'train_tf_function': True,
'train_tf_while_loop': True,
'validation_interval': 100,
'validation_steps': 100,
'validation_summary_subdir': 'validation'} }
<IPython.core.display.Javascript object>
设置分发策略。
if exp_config.runtime.mixed_precision_dtype == tf.float16:
tf.keras.mixed_precision.set_global_policy('mixed_float16')
if 'GPU' in ''.join(logical_device_names):
distribution_strategy = tf.distribute.MirroredStrategy()
elif 'TPU' in ''.join(logical_device_names):
tf.tpu.experimental.initialize_tpu_system()
tpu = tf.distribute.cluster_resolver.TPUClusterResolver(tpu='/device:TPU_SYSTEM:0')
distribution_strategy = tf.distribute.experimental.TPUStrategy(tpu)
else:
print('Warning: this will be really slow.')
distribution_strategy = tf.distribute.OneDeviceStrategy(logical_device_names[0])
print('Done')
Warning: this will be really slow. Done
从 config_definitions.TaskConfig 创建 Task 对象（tfm.core.base_task.Task）。
Task 对象拥有构建数据集、构建模型以及运行训练和评估所需的所有方法。这些方法由 tfm.core.train_lib.run_experiment 驱动。
with distribution_strategy.scope():
task = tfm.core.task_factory.get_task(exp_config.task, logging_dir=model_dir)
可视化一批数据。
for images, labels in task.build_inputs(exp_config.task.train_data).take(1):
print()
print(f'images.shape: {str(images.shape):16} images.dtype: {images.dtype!r}')
print(f'labels.keys: {labels.keys()}')
images.shape: (8, 256, 256, 3) images.dtype: tf.float32 labels.keys: dict_keys(['cls_targets', 'box_targets', 'anchor_boxes', 'cls_weights', 'box_weights', 'image_info'])
创建类别索引字典，将标签映射到相应的标签名称。
category_index={
1: {
'id': 1,
'name': 'Platelets'
},
2: {
'id': 2,
'name': 'RBC'
},
3: {
'id': 3,
'name': 'WBC'
}
}
tf_ex_decoder = TfExampleDecoder()
用于可视化 TFRecords 结果的辅助函数。
使用 visualization_utils 中的 visualize_boxes_and_labels_on_image_array 在图像上绘制边界框。
def show_batch(raw_records, num_of_examples):
plt.figure(figsize=(20, 20))
use_normalized_coordinates=True
min_score_thresh = 0.30
for i, serialized_example in enumerate(raw_records):
plt.subplot(1, 3, i + 1)
decoded_tensors = tf_ex_decoder.decode(serialized_example)
image = decoded_tensors['image'].numpy().astype('uint8')
scores = np.ones(shape=(len(decoded_tensors['groundtruth_boxes'])))
visualization_utils.visualize_boxes_and_labels_on_image_array(
image,
decoded_tensors['groundtruth_boxes'].numpy(),
decoded_tensors['groundtruth_classes'].numpy().astype('int'),
scores,
category_index=category_index,
use_normalized_coordinates=use_normalized_coordinates,
max_boxes_to_draw=200,
min_score_thresh=min_score_thresh,
agnostic_mode=False,
instance_masks=None,
line_thickness=4)
plt.imshow(image)
plt.axis('off')
plt.title(f'Image-{i+1}')
plt.show()
训练数据的可视化
边界框检测包含两个部分
- 检测到的对象的类别标签（例如 RBC）
- 预测边界框与标注（Ground Truth）边界框之间的匹配百分比。
buffer_size = 20
num_of_examples = 3
raw_records = tf.data.TFRecordDataset(
exp_config.task.train_data.input_path).shuffle(
buffer_size=buffer_size).take(num_of_examples)
show_batch(raw_records, num_of_examples)
训练和评估。
我们遵循 COCO 挑战赛的惯例，基于 mAP（平均精度均值）评估目标检测的准确性。请点击此处查看关于如何完成检测任务评估指标的详细说明。
IoU（交并比）：定义为预测边界框与标注边界框的交集面积除以并集面积。
model, eval_logs = tfm.core.train_lib.run_experiment(
distribution_strategy=distribution_strategy,
task=task,
mode='train_and_eval',
params=exp_config,
model_dir=model_dir,
run_post_eval=True)
restoring or initializing model...
INFO:tensorflow:Customized initialization is done through the passed `init_fn`.
INFO:tensorflow:Customized initialization is done through the passed `init_fn`.
train | step: 0 | training until step 100...
2023-11-09 12:15:53.688817: W tensorflow/core/framework/dataset.cc:959] Input of GeneratorDatasetOp::Dataset will not be optimized because the dataset does not implement the AsGraphDefInternal() method needed to apply optimizations.
train | step: 100 | steps/sec: 0.7 | output:
{'box_loss': 0.02193972,
'cls_loss': 0.6757131,
'learning_rate': 0.09755283,
'model_loss': 1.7726997,
'total_loss': 2.8906868,
'training_loss': 2.8906868}
saved checkpoint to ./trained_model/ckpt-100.
eval | step: 100 | running 100 steps of evaluation...
2023-11-09 12:18:11.406759: W tensorflow/core/framework/dataset.cc:959] Input of GeneratorDatasetOp::Dataset will not be optimized because the dataset does not implement the AsGraphDefInternal() method needed to apply optimizations.
creating index...
index created!
creating index...
index created!
Running per image evaluation...
Evaluate annotation type *bbox*
DONE (t=0.42s).
Accumulating evaluation results...
DONE (t=0.06s).
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.000
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.000
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.000
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.000
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.000
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.000
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.000
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.000
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.000
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.000
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.000
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.000
eval | step: 100 | steps/sec: 10.1 | eval time: 9.9 sec | output:
{'AP': 0.0,
'AP50': 0.0,
'AP75': 0.0,
'APl': 0.0,
'APm': 0.0,
'APs': 0.0,
'ARl': 0.0,
'ARm': 0.0,
'ARmax1': 0.0,
'ARmax10': 0.0,
'ARmax100': 0.0,
'ARs': 0.0,
'box_loss': 18.058144,
'cls_loss': 11249.595,
'model_loss': 12152.503,
'steps_per_second': 10.0835173990815,
'total_loss': 12153.81,
'validation_loss': 12153.81}
train | step: 100 | training until step 200...
train | step: 200 | steps/sec: 0.7 | output:
{'box_loss': 0.008522138,
'cls_loss': 0.50075334,
'learning_rate': 0.090450846,
'model_loss': 0.92686033,
'total_loss': 2.2234242,
'training_loss': 2.2234242}
saved checkpoint to ./trained_model/ckpt-200.
eval | step: 200 | running 100 steps of evaluation...
2023-11-09 12:20:26.584188: W tensorflow/core/framework/dataset.cc:959] Input of GeneratorDatasetOp::Dataset will not be optimized because the dataset does not implement the AsGraphDefInternal() method needed to apply optimizations.
creating index...
index created!
creating index...
index created!
Running per image evaluation...
Evaluate annotation type *bbox*
DONE (t=0.61s).
Accumulating evaluation results...
DONE (t=0.06s).
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.011
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.034
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.003
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.000
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.010
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.059
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.004
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.029
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.157
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.000
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.049
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.177
eval | step: 200 | steps/sec: 13.2 | eval time: 7.6 sec | output:
{'AP': 0.010944664,
'AP50': 0.03427069,
'AP75': 0.0026895092,
'APl': 0.05921977,
'APm': 0.01015669,
'APs': 0.0,
'ARl': 0.17723133,
'ARm': 0.048520368,
'ARmax1': 0.0035891088,
'ARmax10': 0.029232906,
'ARmax100': 0.15689641,
'ARs': 0.0,
'box_loss': 0.009609308,
'cls_loss': 4.135805,
'model_loss': 4.61627,
'steps_per_second': 13.19492688372793,
'total_loss': 5.900796,
'validation_loss': 5.900796}
train | step: 200 | training until step 300...
train | step: 300 | steps/sec: 0.8 | output:
{'box_loss': 0.00830541,
'cls_loss': 0.4176439,
'learning_rate': 0.07938927,
'model_loss': 0.8329147,
'total_loss': 2.1065934,
'training_loss': 2.1065934}
saved checkpoint to ./trained_model/ckpt-300.
eval | step: 300 | running 100 steps of evaluation...
2023-11-09 12:22:38.377585: W tensorflow/core/framework/dataset.cc:959] Input of GeneratorDatasetOp::Dataset will not be optimized because the dataset does not implement the AsGraphDefInternal() method needed to apply optimizations.
creating index...
index created!
creating index...
index created!
Running per image evaluation...
Evaluate annotation type *bbox*
DONE (t=1.35s).
Accumulating evaluation results...
DONE (t=0.06s).
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.016
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.069
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.002
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.000
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.013
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.009
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.032
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.063
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.141
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.000
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.100
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.124
eval | step: 300 | steps/sec: 12.5 | eval time: 8.0 sec | output:
{'AP': 0.016204117,
'AP50': 0.06889212,
'AP75': 0.0020602986,
'APl': 0.009428242,
'APm': 0.013351731,
'APs': 0.0,
'ARl': 0.123588346,
'ARm': 0.10017456,
'ARmax1': 0.03215091,
'ARmax10': 0.063206546,
'ARmax100': 0.14125936,
'ARs': 0.0,
'box_loss': 0.00870863,
'cls_loss': 0.5394852,
'model_loss': 0.97491676,
'steps_per_second': 12.503431188892941,
'total_loss': 2.2379956,
'validation_loss': 2.2379956}
train | step: 300 | training until step 400...
train | step: 400 | steps/sec: 0.8 | output:
{'box_loss': 0.008178472,
'cls_loss': 0.39109412,
'learning_rate': 0.06545085,
'model_loss': 0.8000177,
'total_loss': 2.05375,
'training_loss': 2.05375}
saved checkpoint to ./trained_model/ckpt-400.
eval | step: 400 | running 100 steps of evaluation...
2023-11-09 12:24:51.305428: W tensorflow/core/framework/dataset.cc:959] Input of GeneratorDatasetOp::Dataset will not be optimized because the dataset does not implement the AsGraphDefInternal() method needed to apply optimizations.
creating index...
index created!
creating index...
index created!
Running per image evaluation...
Evaluate annotation type *bbox*
DONE (t=1.32s).
Accumulating evaluation results...
DONE (t=0.06s).
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.170
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.417
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.080
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.000
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.126
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.164
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.139
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.211
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.289
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.000
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.223
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.265
eval | step: 400 | steps/sec: 12.5 | eval time: 8.0 sec | output:
{'AP': 0.16998358,
'AP50': 0.41680792,
'AP75': 0.07954546,
'APl': 0.16424511,
'APm': 0.12631872,
'APs': 0.0,
'ARl': 0.26493624,
'ARm': 0.22271821,
'ARmax1': 0.13854705,
'ARmax10': 0.2114396,
'ARmax100': 0.2887498,
'ARs': 0.0,
'box_loss': 0.007682202,
'cls_loss': 0.44814157,
'model_loss': 0.8322517,
'steps_per_second': 12.476437706321125,
'total_loss': 2.0770469,
'validation_loss': 2.0770469}
train | step: 400 | training until step 500...
train | step: 500 | steps/sec: 0.8 | output:
{'box_loss': 0.008003844,
'cls_loss': 0.37224656,
'learning_rate': 0.049999997,
'model_loss': 0.77243865,
'total_loss': 2.0097175,
'training_loss': 2.0097175}
saved checkpoint to ./trained_model/ckpt-500.
eval | step: 500 | running 100 steps of evaluation...
2023-11-09 12:27:03.978075: W tensorflow/core/framework/dataset.cc:959] Input of GeneratorDatasetOp::Dataset will not be optimized because the dataset does not implement the AsGraphDefInternal() method needed to apply optimizations.
creating index...
index created!
creating index...
index created!
Running per image evaluation...
Evaluate annotation type *bbox*
DONE (t=1.31s).
Accumulating evaluation results...
DONE (t=0.06s).
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.144
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.409
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.018
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.000
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.068
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.146
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.141
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.177
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.249
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.000
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.163
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.228
eval | step: 500 | steps/sec: 12.6 | eval time: 8.0 sec | output:
{'AP': 0.14389119,
'AP50': 0.40903544,
'AP75': 0.01820877,
'APl': 0.14554419,
'APm': 0.06774697,
'APs': 0.0,
'ARl': 0.22750455,
'ARm': 0.16317539,
'ARmax1': 0.14100254,
'ARmax10': 0.17656359,
'ARmax100': 0.24859329,
'ARs': 0.0,
'box_loss': 0.008371367,
'cls_loss': 0.46203443,
'model_loss': 0.8806029,
'steps_per_second': 12.550521613819056,
'total_loss': 2.110864,
'validation_loss': 2.110864}
train | step: 500 | training until step 600...
train | step: 600 | steps/sec: 0.8 | output:
{'box_loss': 0.0075427825,
'cls_loss': 0.36228406,
'learning_rate': 0.034549143,
'model_loss': 0.739423,
'total_loss': 1.9641409,
'training_loss': 1.9641409}
saved checkpoint to ./trained_model/ckpt-600.
eval | step: 600 | running 100 steps of evaluation...
2023-11-09 12:29:16.661897: W tensorflow/core/framework/dataset.cc:959] Input of GeneratorDatasetOp::Dataset will not be optimized because the dataset does not implement the AsGraphDefInternal() method needed to apply optimizations.
creating index...
index created!
creating index...
index created!
Running per image evaluation...
Evaluate annotation type *bbox*
DONE (t=1.26s).
Accumulating evaluation results...
DONE (t=0.06s).
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.205
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.494
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.074
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.000
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.165
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.210
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.178
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.249
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.304
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.000
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.288
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.281
eval | step: 600 | steps/sec: 12.3 | eval time: 8.1 sec | output:
{'AP': 0.20549679,
'AP50': 0.49406928,
'AP75': 0.074281454,
'APl': 0.20956378,
'APm': 0.16475147,
'APs': 0.0,
'ARl': 0.2809654,
'ARm': 0.28829592,
'ARmax1': 0.17803712,
'ARmax10': 0.24863629,
'ARmax100': 0.3042056,
'ARs': 0.0,
'box_loss': 0.007942874,
'cls_loss': 0.3965486,
'model_loss': 0.79369235,
'steps_per_second': 12.345320504588557,
'total_loss': 2.013407,
'validation_loss': 2.013407}
train | step: 600 | training until step 700...
train | step: 700 | steps/sec: 0.8 | output:
{'box_loss': 0.007070377,
'cls_loss': 0.34681392,
'learning_rate': 0.02061074,
'model_loss': 0.7003328,
'total_loss': 1.91634,
'training_loss': 1.91634}
saved checkpoint to ./trained_model/ckpt-700.
eval | step: 700 | running 100 steps of evaluation...
2023-11-09 12:31:29.396952: W tensorflow/core/framework/dataset.cc:959] Input of GeneratorDatasetOp::Dataset will not be optimized because the dataset does not implement the AsGraphDefInternal() method needed to apply optimizations.
creating index...
index created!
creating index...
index created!
Running per image evaluation...
Evaluate annotation type *bbox*
DONE (t=1.25s).
Accumulating evaluation results...
DONE (t=0.06s).
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.266
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.548
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.195
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.000
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.211
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.320
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.199
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.293
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.354
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.000
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.314
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.339
eval | step: 700 | steps/sec: 12.4 | eval time: 8.0 sec | output:
{'AP': 0.26574683,
'AP50': 0.54845077,
'AP75': 0.19461285,
'APl': 0.3196148,
'APm': 0.21140255,
'APs': 0.0,
'ARl': 0.33888888,
'ARm': 0.31413966,
'ARmax1': 0.19884779,
'ARmax10': 0.29322097,
'ARmax100': 0.35407078,
'ARs': 0.0,
'box_loss': 0.006610607,
'cls_loss': 0.35634473,
'model_loss': 0.68687505,
'steps_per_second': 12.437609994369083,
'total_loss': 1.8996844,
'validation_loss': 1.8996844}
train | step: 700 | training until step 800...
train | step: 800 | steps/sec: 0.8 | output:
{'box_loss': 0.006680321,
'cls_loss': 0.3391388,
'learning_rate': 0.009549147,
'model_loss': 0.673155,
'total_loss': 1.8838484,
'training_loss': 1.8838484}
saved checkpoint to ./trained_model/ckpt-800.
eval | step: 800 | running 100 steps of evaluation...
2023-11-09 12:33:41.871412: W tensorflow/core/framework/dataset.cc:959] Input of GeneratorDatasetOp::Dataset will not be optimized because the dataset does not implement the AsGraphDefInternal() method needed to apply optimizations.
creating index...
index created!
creating index...
index created!
Running per image evaluation...
Evaluate annotation type *bbox*
DONE (t=1.21s).
Accumulating evaluation results...
DONE (t=0.06s).
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.270
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.551
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.228
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.000
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.238
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.293
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.203
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.292
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.355
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.000
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.330
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.310
eval | step: 800 | steps/sec: 12.2 | eval time: 8.2 sec | output:
{'AP': 0.27034393,
'AP50': 0.55116975,
'AP75': 0.22773828,
'APl': 0.29330468,
'APm': 0.23779006,
'APs': 0.0,
'ARl': 0.30956283,
'ARm': 0.33016625,
'ARmax1': 0.2025409,
'ARmax10': 0.29244527,
'ARmax100': 0.35519278,
'ARs': 0.0,
'box_loss': 0.006337466,
'cls_loss': 0.3529123,
'model_loss': 0.6697856,
'steps_per_second': 12.213763576089187,
'total_loss': 1.8787861,
'validation_loss': 1.8787861}
train | step: 800 | training until step 900...
train | step: 900 | steps/sec: 0.8 | output:
{'box_loss': 0.006552434,
'cls_loss': 0.3347254,
'learning_rate': 0.002447176,
'model_loss': 0.6623471,
'total_loss': 1.8704419,
'training_loss': 1.8704419}
saved checkpoint to ./trained_model/ckpt-900.
eval | step: 900 | running 100 steps of evaluation...
2023-11-09 12:35:54.334538: W tensorflow/core/framework/dataset.cc:959] Input of GeneratorDatasetOp::Dataset will not be optimized because the dataset does not implement the AsGraphDefInternal() method needed to apply optimizations.
creating index...
index created!
creating index...
index created!
Running per image evaluation...
Evaluate annotation type *bbox*
DONE (t=1.22s).
Accumulating evaluation results...
DONE (t=0.06s).
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.265
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.563
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.146
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.000
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.261
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.306
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.197
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.290
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.348
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.000
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.344
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.326
eval | step: 900 | steps/sec: 12.2 | eval time: 8.2 sec | output:
{'AP': 0.26503274,
'AP50': 0.5632806,
'AP75': 0.1462794,
'APl': 0.30633608,
'APm': 0.26065546,
'APs': 0.0,
'ARl': 0.3263206,
'ARm': 0.34426433,
'ARmax1': 0.19695416,
'ARmax10': 0.28974226,
'ARmax100': 0.3476218,
'ARs': 0.0,
'box_loss': 0.006341949,
'cls_loss': 0.3344063,
'model_loss': 0.6515038,
'steps_per_second': 12.173977857621107,
'total_loss': 1.8589785,
'validation_loss': 1.8589785}
train | step: 900 | training until step 1000...
train | step: 1000 | steps/sec: 0.8 | output:
{'box_loss': 0.0064173574,
'cls_loss': 0.33288643,
'learning_rate': 0.0,
'model_loss': 0.6537543,
'total_loss': 1.8610404,
'training_loss': 1.8610404}
saved checkpoint to ./trained_model/ckpt-1000.
eval | step: 1000 | running 100 steps of evaluation...
2023-11-09 12:38:06.900922: W tensorflow/core/framework/dataset.cc:959] Input of GeneratorDatasetOp::Dataset will not be optimized because the dataset does not implement the AsGraphDefInternal() method needed to apply optimizations.
creating index...
index created!
creating index...
index created!
Running per image evaluation...
Evaluate annotation type *bbox*
DONE (t=1.22s).
Accumulating evaluation results...
DONE (t=0.06s).
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.258
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.554
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.137
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.000
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.247
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.298
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.193
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.287
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.344
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.000
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.329
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.324
eval | step: 1000 | steps/sec: 11.6 | eval time: 8.6 sec | output:
{'AP': 0.25785798,
'AP50': 0.5541018,
'AP75': 0.13742597,
'APl': 0.29831147,
'APm': 0.24709457,
'APs': 0.0,
'ARl': 0.32358834,
'ARm': 0.3290108,
'ARmax1': 0.19289383,
'ARmax10': 0.28744015,
'ARmax100': 0.3443296,
'ARs': 0.0,
'box_loss': 0.006123655,
'cls_loss': 0.3268734,
'model_loss': 0.6330561,
'steps_per_second': 11.647036187140033,
'total_loss': 1.8402674,
'validation_loss': 1.8402674}
eval | step: 1000 | running 100 steps of evaluation...
creating index...
index created!
creating index...
index created!
Running per image evaluation...
Evaluate annotation type *bbox*
DONE (t=1.21s).
Accumulating evaluation results...
DONE (t=0.06s).
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.258
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.554
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.137
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.000
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.247
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.298
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.193
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.287
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.344
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.000
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.329
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.324
eval | step: 1000 | steps/sec: 12.1 | eval time: 8.2 sec | output:
{'AP': 0.25785798,
'AP50': 0.5541018,
'AP75': 0.13742597,
'APl': 0.29831147,
'APm': 0.24709457,
'APs': 0.0,
'ARl': 0.32358834,
'ARm': 0.3290108,
'ARmax1': 0.19289383,
'ARmax10': 0.28744015,
'ARmax100': 0.3443296,
'ARs': 0.0,
'box_loss': 0.006123655,
'cls_loss': 0.3268734,
'model_loss': 0.6330561,
'steps_per_second': 12.140613310097448,
'total_loss': 1.8402674,
'validation_loss': 1.8402674}
在 tensorboard 中加载日志。
%load_ext tensorboard
%tensorboard --logdir './trained_model/'
保存并导出训练好的模型。
由 train_lib.run_experiment 返回的 keras.Model 对象要求数据由数据集加载器使用与 preprocess_ops.normalize_image(image, offset=MEAN_RGB, scale=STDDEV_RGB) 中相同的均值和方差统计信息进行标准化。此导出函数处理了这些细节，因此您可以传入 tf.uint8 格式的图像并获得正确的结果。
export_saved_model_lib.export_inference_graph(
input_type='image_tensor',
batch_size=1,
input_image_size=[HEIGHT, WIDTH],
params=exp_config,
checkpoint_path=tf.train.latest_checkpoint(model_dir),
export_dir=export_dir)
WARNING:tensorflow:Skipping full serialization of Keras layer <official.vision.modeling.retinanet_model.RetinaNetModel object at 0x7f29ec799a90>, because it is not built. WARNING:tensorflow:Skipping full serialization of Keras layer <official.vision.modeling.retinanet_model.RetinaNetModel object at 0x7f29ec799a90>, because it is not built. WARNING:tensorflow:Skipping full serialization of Keras layer <official.vision.modeling.layers.detection_generator.MultilevelDetectionGenerator object at 0x7f2a30060940>, because it is not built. WARNING:tensorflow:Skipping full serialization of Keras layer <official.vision.modeling.layers.detection_generator.MultilevelDetectionGenerator object at 0x7f2a30060940>, because it is not built. INFO:tensorflow:Assets written to: ./exported_model/assets INFO:tensorflow:Assets written to: ./exported_model/assets
使用训练好的模型进行推理
def load_image_into_numpy_array(path):
"""Load an image from file into a numpy array.
Puts image into numpy array to feed into tensorflow graph.
Note that by convention we put it into a numpy array with shape
(height, width, channels), where channels=3 for RGB.
Args:
path: the file path to the image
Returns:
uint8 numpy array with shape (img_height, img_width, 3)
"""
image = None
if(path.startswith('http')):
response = urlopen(path)
image_data = response.read()
image_data = BytesIO(image_data)
image = Image.open(image_data)
else:
image_data = tf.io.gfile.GFile(path, 'rb').read()
image = Image.open(BytesIO(image_data))
(im_width, im_height) = image.size
return np.array(image.getdata()).reshape(
(1, im_height, im_width, 3)).astype(np.uint8)
def build_inputs_for_object_detection(image, input_image_size):
"""Builds Object Detection model inputs for serving."""
image, _ = resize_and_crop_image(
image,
input_image_size,
padded_size=input_image_size,
aug_scale_min=1.0,
aug_scale_max=1.0)
return image
可视化测试数据。
num_of_examples = 3
test_ds = tf.data.TFRecordDataset(
'./bccd_coco_tfrecords/test-00000-of-00001.tfrecord').take(
num_of_examples)
show_batch(test_ds, num_of_examples)
导入 SavedModel。
imported = tf.saved_model.load(export_dir)
model_fn = imported.signatures['serving_default']
可视化预测结果。
input_image_size = (HEIGHT, WIDTH)
plt.figure(figsize=(20, 20))
min_score_thresh = 0.30 # Change minimum score for threshold to see all bounding boxes confidences.
for i, serialized_example in enumerate(test_ds):
plt.subplot(1, 3, i+1)
decoded_tensors = tf_ex_decoder.decode(serialized_example)
image = build_inputs_for_object_detection(decoded_tensors['image'], input_image_size)
image = tf.expand_dims(image, axis=0)
image = tf.cast(image, dtype = tf.uint8)
image_np = image[0].numpy()
result = model_fn(image)
visualization_utils.visualize_boxes_and_labels_on_image_array(
image_np,
result['detection_boxes'][0].numpy(),
result['detection_classes'][0].numpy().astype(int),
result['detection_scores'][0].numpy(),
category_index=category_index,
use_normalized_coordinates=False,
max_boxes_to_draw=200,
min_score_thresh=min_score_thresh,
agnostic_mode=False,
instance_masks=None,
line_thickness=4)
plt.imshow(image_np)
plt.axis('off')
plt.show()
