2018-09-02 deeplab + edge summary

[yzbx]

todo

• [x] ignore label for edge
• [x] graph summary
• [ ] dynamic edge width
• [x] dynamic edge loss weight

ignore label for edge (orange line for psp_edge, green line for pspnet)

[yzbx]

ignore label for edge (orange line for psp_edge, green line for pspnet)

[yzbx]

graph summary

graph viz

• in device view: purple means GPU and the green is CPU
• in structure view: when two high-level nodes have the same structure, they appear in the same color of the rainbow. Uniquely structured nodes are gray
• runtime statistics: such as total memory usage, total compute time, and tensor shapes for nodes

  summary, _ = sess.run([merged, train_step],
                            feed_dict=feed_dict(True),
                            options=run_options,
                            run_metadata=run_metadata)
      train_writer.add_run_metadata(run_metadata, 'step%d' % i)
      train_writer.add_summary(summary, i)

code demo with feed_dict

summary ops

• tensor_summary
• scalar
• histogram
• audio
• image
• merge
• merge_all

  Operations in TensorFlow don't do anything until you run them, or an op that depends on their output. And the summary nodes that we've just created are peripheral to your graph: none of the ops you are currently running depend on them. So, to generate summaries, we need to run all of these summary nodes. Managing them by hand would be tedious, so use tf.summary.merge_all to combine them into a single op that generates all the summary data.

How TensorBoard loads the data

Why does it read the whole directory, rather than an individual file? You might have been using supervisor.py to run your model, in which case if TensorFlow crashes, the supervisor will restart it from a checkpoint. When it restarts, it will start writing to a new events file, and TensorBoard will stitch the various event files together to produce a consistent history of what happened.

Tensorboard Text: Markdown Support

The Text Dashboard displays text snippets saved via tf.summary.text. Markdown features including hyperlinks, lists, and tables are all supported.

[yzbx]

slim and tensorboard

• mnist code example

  # use the accuracy as a metric
      accuracy = tf.metrics.accuracy(
          labels=labels,
          predictions=tf.argmax(logits, axis=1))
  
      # create an estimator spec with the loss and accuracy
      estimator_spec = tf.estimator.EstimatorSpec(
          mode=tf.estimator.ModeKeys.EVAL,
          loss=loss,
          eval_metric_ops={
              'accuracy': accuracy
          })

• the right way to use metrics, note for eval_metric_ops['accuracy']!=accuracy[0] if not run accuracy[1], the accuracy[0] will always be 0 and will not be updated!

  example from test/tensorboard_summary.py

      metric_map = {}
      metric_map['miou'] = tf.metrics.mean_iou(
          predictions, labels, num_classes, weights=weights)
      metric_map['acc'] = tf.metrics.accuracy(
          labels=labels, predictions=predictions, weights=tf.reshape(weights, shape=[-1]))
  
      metrics = {}
      metrics['miou']=tf.identity(metric_map['miou'][0],name='hello')
      metrics['acc']=tf.identity(metric_map['acc'][0],name='world')
      for x in ['miou', 'acc']:
          tf.summary.scalar('metrics/%s' % x, metrics[x])
  
      summary_op = tf.summary.merge_all()
      sess = tf.Session()
      sess.run(tf.global_variables_initializer())
      sess.run(tf.local_variables_initializer())
      for i in trange(5):
          np_img, np_seg, np_edge, img_name, np_height, np_width, summary, metrics_np = sess.run(
              [img, seg, edge, img_filename, height, width, summary_op, metric_map])
          print(img_name, np_height, np_width)
          print(metrics_np['miou'][0],metrics_np['acc'][0])
          writer.add_summary(summary, i)

[yzbx]

slim reference

• eval_image_classifier.py
• train_image_classifier.py

slim

• dataset_factory
• preprocessing_factory
• nets_factory: can export interface for placeholder

  with tf.Graph().as_default() as graph:
  dataset = dataset_factory.get_dataset(FLAGS.dataset_name, 'train',
                                        FLAGS.dataset_dir)
  network_fn = nets_factory.get_network_fn(
      FLAGS.model_name,
      num_classes=(dataset.num_classes - FLAGS.labels_offset),
      is_training=FLAGS.is_training)
  image_size = FLAGS.image_size or network_fn.default_image_size
  placeholder = tf.placeholder(name='input', dtype=tf.float32,
                               shape=[FLAGS.batch_size, image_size,
                                      image_size, 3])
  network_fn(placeholder)
  graph_def = graph.as_graph_def()
  with gfile.GFile(FLAGS.output_file, 'wb') as f:
    f.write(graph_def.SerializeToString())

data flow

1. dataset_factory->[height,width,channel]
2. slim.dataset_data_provider.DatasetDataProvider->[height,width,channel]
3. preprocessing_factory->[height,width,channel]
4. tf.train.batch->[batch,height,width,channel]
5. nets_factory->[batch,height,width,channel]

slim metrics

• tf.summary.scalar(summary_name, value, collections=[])
  • tf.add_to_collection(tf.GraphKeys.SUMMARIES,op)

    
    # Define the metrics:
    names_to_values, names_to_updates = slim.metrics.aggregate_metric_map({
    'Accuracy': slim.metrics.streaming_accuracy(predictions, labels),
    'Recall_5': slim.metrics.streaming_recall_at_k(
    logits, labels, 5),
    })

Print the summaries to screen.

for name, value in names_to_values.items(): summary_name = 'eval/%s' % name op = tf.summary.scalar(summary_name, value, collections=[]) op = tf.Print(op, [value], summary_name) tf.add_to_collection(tf.GraphKeys.SUMMARIES, op)

# slim init weight from checkpoint path

variables_to_restore = [] for var in slim.get_model_variables(): for exclusion in exclusions: if var.op.name.startswith(exclusion): break else: variables_to_restore.append(var)

init_fn=slim.assign_from_checkpoint_fn( checkpoint_path, variables_to_restore, ignore_missing_vars=FLAGS.ignore_missing_vars)