您可以从单个 anaconda 提示 运行 训练和评估过程吗?

Can you run training and evalutaion process from a single anaconda prompt?

我在训练 Tensorflow2 自定义对象检测器期间无法评估我的训练过程。在阅读了与此问题相关的几个问题后,我发现评估和培训应该被视为两个独立的过程,因此我应该使用新的 anaconda 提示来开始评估工作。 我正在训练 ssd_mobilenetv2 640x640 版本。我的管道配置:

model {
  ssd {
    num_classes: 6
    image_resizer {
      fixed_shape_resizer {
        height: 640
        width: 640
      }
    }
    feature_extractor {
      type: "ssd_mobilenet_v2_fpn_keras"
      depth_multiplier: 1.0
      min_depth: 16
      conv_hyperparams {
        regularizer {
          l2_regularizer {
            weight: 3.9999998989515007e-05
          }
        }
        initializer {
          random_normal_initializer {
            mean: 0.0
            stddev: 0.009999999776482582
          }
        }
        activation: RELU_6
        batch_norm {
          decay: 0.996999979019165
          scale: true
          epsilon: 0.0010000000474974513
        }
      }
      use_depthwise: true
      override_base_feature_extractor_hyperparams: true
      fpn {
        min_level: 3
        max_level: 7
        additional_layer_depth: 128
      }
    }
    box_coder {
      faster_rcnn_box_coder {
        y_scale: 10.0
        x_scale: 10.0
        height_scale: 5.0
        width_scale: 5.0
      }
    }
    matcher {
      argmax_matcher {
        matched_threshold: 0.5
        unmatched_threshold: 0.5
        ignore_thresholds: false
        negatives_lower_than_unmatched: true
        force_match_for_each_row: true
        use_matmul_gather: true
      }
    }
    similarity_calculator {
      iou_similarity {
      }
    }
    box_predictor {
      weight_shared_convolutional_box_predictor {
        conv_hyperparams {
          regularizer {
            l2_regularizer {
              weight: 3.9999998989515007e-05
            }
          }
          initializer {
            random_normal_initializer {
              mean: 0.0
              stddev: 0.009999999776482582
            }
          }
          activation: RELU_6
          batch_norm {
            decay: 0.996999979019165
            scale: true
            epsilon: 0.0010000000474974513
          }
        }
        depth: 128
        num_layers_before_predictor: 4
        kernel_size: 3
        class_prediction_bias_init: -4.599999904632568
        share_prediction_tower: true
        use_depthwise: true
      }
    }
    anchor_generator {
      multiscale_anchor_generator {
        min_level: 3
        max_level: 7
        anchor_scale: 4.0
        aspect_ratios: 1.0
        aspect_ratios: 2.0
        aspect_ratios: 0.5
        scales_per_octave: 2
      }
    }
    post_processing {
      batch_non_max_suppression {
        score_threshold: 9.99999993922529e-09
        iou_threshold: 0.6000000238418579
        max_detections_per_class: 100
        max_total_detections: 100
        use_static_shapes: false
      }
      score_converter: SIGMOID
    }
    normalize_loss_by_num_matches: true
    loss {
      localization_loss {
        weighted_smooth_l1 {
        }
      }
      classification_loss {
        weighted_sigmoid_focal {
          gamma: 2.0
          alpha: 0.25
        }
      }
      classification_weight: 1.0
      localization_weight: 1.0
    }
    encode_background_as_zeros: true
    normalize_loc_loss_by_codesize: true
    inplace_batchnorm_update: true
    freeze_batchnorm: false
  }
}
train_config {
  batch_size: 4
  data_augmentation_options {
    random_horizontal_flip {
    }
  }
  #data_augmentation_options {
    #random_crop_image {
      #min_object_covered: 0.0
      #min_aspect_ratio: 0.75
      #max_aspect_ratio: 3.0
      #min_area: 0.75
      #max_area: 1.0
      #overlap_thresh: 0.0
    #}
  #}
  optimizer {
    momentum_optimizer {
      learning_rate {
        cosine_decay_learning_rate {
          learning_rate_base: 0.04999999821186066
          total_steps: 50000
          warmup_learning_rate: 0.0026666000485420227
          warmup_steps: 600
        }
      }
      momentum_optimizer_value: 0.8999999761581421
    }
    use_moving_average: false
  }
  fine_tune_checkpoint: "pre-trained-models\ssd_mobilenet_v2_fpnlite_640x640_coco17_tpu-8\checkpoint\ckpt-0"
  num_steps: 50000
  startup_delay_steps: 0.0
  replicas_to_aggregate: 8
  max_number_of_boxes: 100
  unpad_groundtruth_tensors: false
  fine_tune_checkpoint_type: "detection"
  fine_tune_checkpoint_version: V2
  from_detection_checkpoint: true
}
train_input_reader {
  label_map_path: "annotations/label_map.pbtxt"
  tf_record_input_reader {
    input_path: "data/train.record"
  }
 
}
eval_config {
  metrics_set: "coco_detection_metrics"
  use_moving_averages: false
}
eval_input_reader {
  label_map_path: "annotations/label_map.pbtxt"
  shuffle: false
  num_epochs: 1
  tf_record_input_reader {
    input_path: "data/test.record"
  }
}

我已经用命令开始训练了:

python model_main_tf2.py --model_dir=models/my_ssd2_3/ --pipeline_config_path=models/my_ssd2_3/pipeline.config --sample_1_of_n_eval_examples 1 --logtostderr

我希望设置评估示例的数量会对启动评估工作产生影响。无论如何,我尝试 运行 在不同的终端 window 中进行评估:python model_main_tf2.py --model_dir=models/my_ssd2_3 --pipeline_config_path=models/my_ssd2_3/pipeline.config --checkpoint_dir=models/my_ssd2_3/ --alsologtostderr
一旦开始评估,训练作业就会崩溃并出现以下错误:error

我认为我没有足够的硬件的问题:

  1. 8GB 内存
  2. 英伟达 GTX960M(2GB 内存)

会不会是我使用的所有输入图像都是3000x3000的问题,因此预处理器必须加载太多信息?如果是这样,有什么办法可以解决吗?我不想在生成 TF 记录文件之前调整所有图像的大小,因为我必须重新标记所有图像。我显然缺乏对在训练过程开始时如何分配内存的洞察力,因此非常感谢一些细节。


第二个问题是,在监控张量板上的训练期间,图像以各种亮度显示我尝试将 model_lib_v2.py 文件中的 627 行更改为:

data= (features[fields.InputDataFields.image]-np.min(features[fields.InputDataFields.image]))/(np.max(features[fields.InputDataFields.image])-np.min(features[fields.InputDataFields.image]))
,

根据这个解决方案:https://github.com/tensorflow/models/issues/9115 没有任何运气。这个问题有解决办法吗?如果我可以在那里监视模型建议的边界框,那就太好了。谢谢。

通过对 model_lib.py 中的 train_loop 函数进行一些更改,您可以在同一应用程序中交替进行训练和评估。请参阅下面的示例。

据我了解,Tensorflow 对象检测 API 的开发重点是分布式学习,如果您使用的是多个 GPUs/TPUs,那么您可以让一些设备进行训练,而其他设备则进行训练评估。所以我怀疑model_lib.py目前的实现方式并不完全支持在同一设备上进行训练和评估。

我不确定您看到的错误的根本原因,通常我看到 Tensorflow 在出现内存问题时抛出 OOM 错误。可能是 Tensorflow 使用 CUDA 的方式不支持使用同一设备的两个应用程序。

关于你的第二个问题,我听从了建议 here on the same thread,这对我有用。复制下面第三个代码块中的代码。最初,这似乎对我不起作用,因为我天真地更新了我创建的对象检测存储库中的文件,但您的应用程序可能正在使用安装在您的站点库中的对象检测 API,所以我会建议确认您正在更改的文件与导入语句中加载的文件相同。

--

这在训练循环之外

  ##Set up evaluation data and writer
  eval_config = configs['eval_config']
  eval_input_configs = configs['eval_input_configs']
  eval_input_config = eval_input_configs[0]
  eval_input = strategy.experimental_distribute_dataset(
    inputs.eval_input(
        eval_config=eval_config,
        eval_input_config=eval_input_config,
        model_config=model_config,
        model=detection_model))
  
  summary_writer_eval = tf.compat.v2.summary.create_file_writer(os.path.join(model_dir, 'eval', eval_input_config.name))

这是修改后的 train/evaluation 循环。评估接近尾声。

for _ in range(global_step.value(), train_steps, num_steps_per_iteration):
  
  tf.logging.info('Performing Training')
  with summary_writer_train.as_default():
    with tf.compat.v2.summary.record_if(lambda: global_step % num_steps_per_iteration == 0):

      losses_dict = _dist_train_step(train_input_iter)

      time_taken = time.time() - last_step_time
      last_step_time = time.time()
      steps_per_sec = num_steps_per_iteration * 1.0 / time_taken

      tf.compat.v2.summary.scalar(
          'steps_per_sec', steps_per_sec, step=global_step)

      steps_per_sec_list.append(steps_per_sec)

      logged_dict = losses_dict.copy()
      logged_dict['learning_rate'] = learning_rate_fn()

      for key, val in logged_dict.items():
        tf.compat.v2.summary.scalar(key, val, step=global_step)

      if global_step.value() - logged_step >= 0:
        logged_dict_np = {name: value.numpy() for name, value in
                          logged_dict.items()}
        tf.logging.info(
            'Step {} per-step time {:.3f}s'.format(
                global_step.value(), time_taken / num_steps_per_iteration))
        tf.logging.info(pprint.pformat(logged_dict_np, width=40))
        print_gpu_memory_usage()
        logged_step = global_step.value()

      if ((int(global_step.value()) - checkpointed_step) >=
          checkpoint_every_n):
        manager.save()
        checkpointed_step = int(global_step.value())
        
  tf.logging.info('Performing Evaluation')
  with summary_writer_eval.as_default():
    eager_eval_loop(
        detection_model,
        configs,
        eval_input,
        use_tpu=use_tpu,
        global_step=global_step,
        )

修复 TensorBoard 中的图像渲染

if record_summaries:
    imgs = features[fields.InputDataFields.image][:3] 
    imgs = tf.div(tf.subtract(imgs, tf.reduce_min(imgs)), tf.subtract(tf.reduce_max(imgs), tf.reduce_min(imgs)))
    tf.compat.v2.summary.image(name='train_input_images', step=global_step, data=imgs, max_outputs=3)