谷歌Pixel 2人像模式代碼曝光，你看懂了嗎?

谷歌把他們所應(yīng)用的 AI 圖像分層算法 DeepLab-v3+ 變成開源代碼，讓第三方相機 app 都可以利用借此神經(jīng)網(wǎng)絡(luò)。

 開源代碼：

 import tensorflow as tf
 from deeplab.core import feature_extractor
 
 slim = tf.contrib.slim
 
 _LOGITS_SCOPE_NAME = 'logits'
 _MERGED_LOGITS_SCOPE = 'merged_logits'
 _IMAGE_POOLING_SCOPE = 'image_pooling'
 _ASPP_SCOPE = 'aspp'
 _CONCAT_PROJECTION_SCOPE = 'concat_projection'
 _DECODER_SCOPE = 'decoder'
 
 
 def get_extra_layer_scopes():
 """Gets the scopes for extra layers.
 
 Returns:
 A list of scopes for extra layers.
 """
 return [
 _LOGITS_SCOPE_NAME,
 _IMAGE_POOLING_SCOPE,
 _ASPP_SCOPE,
 _CONCAT_PROJECTION_SCOPE,
 _DECODER_SCOPE,
 ]
 
 
 def predict_labels_multi_scale(images,
 model_options,
 eval_scales=(1.0,),
 add_flipped_images=False):
 """Predicts segmentation labels.
 
 Args:
 images: A tensor of size [batch, height, width, channels].
 model_options: A ModelOptions instance to configure models.
 eval_scales: The scales to resize images for evaluation.
 add_flipped_images: Add flipped images for evaluation or not.
 
 Returns:
 A dictionary with keys specifying the output_type (e.g., semantic
 prediction) and values storing Tensors representing predictions (argmax
 over channels). Each prediction has size [batch, height, width].
 """
 outputs_to_predictions = {
 output: []
 for output in model_options.outputs_to_num_classes
 }
 
 for i, image_scale in enumerate(eval_scales):
 with tf.variable_scope(tf.get_variable_scope(), reuse=True if i else None):
 outputs_to_scales_to_logits = multi_scale_logits(
 images,
 model_options=model_options,
 image_pyramid=[image_scale],
 is_training=False,
 fine_tune_batch_norm=False)
 
 if add_flipped_images:
 with tf.variable_scope(tf.get_variable_scope(), reuse=True):
 outputs_to_scales_to_logits_reversed = multi_scale_logits(
 tf.reverse_v2(images, [2]),
 model_options=model_options,
 image_pyramid=[image_scale],
 is_training=False,
 fine_tune_batch_norm=False)
 
 for output in sorted(outputs_to_scales_to_logits):
 scales_to_logits = outputs_to_scales_to_logits[output]
 logits = tf.image.resize_bilinear(
 scales_to_logits[_MERGED_LOGITS_SCOPE],
 tf.shape(images)[1:3],
 align_corners=True)
 outputs_to_predictions[output].append(
 tf.expand_dims(tf.nn.softmax(logits), 4))
 
 if add_flipped_images:
 scales_to_logits_reversed = (
 outputs_to_scales_to_logits_reversed[output])
 logits_reversed = tf.image.resize_bilinear(
 tf.reverse_v2(scales_to_logits_reversed[_MERGED_LOGITS_SCOPE], [2]),
 tf.shape(images)[1:3],
 align_corners=True)
 outputs_to_predictions[output].append(
 tf.expand_dims(tf.nn.softmax(logits_reversed), 4))
 
 for output in sorted(outputs_to_predictions):
 predictions = outputs_to_predictions[output]
 # Compute average prediction across different scales and flipped images.
 predictions = tf.reduce_mean(tf.concat(predictions, 4), axis=4)
 outputs_to_predictions[output] = tf.argmax(predictions, 3)
 
 return outputs_to_predictions
 
 
 def predict_labels(images, model_options, image_pyramid=None):
 """Predicts segmentation labels.
 
 Args:
 images: A tensor of size [batch, height, width, channels].
 model_options: A ModelOptions instance to configure models.
 image_pyramid: Input image scales for multi-scale feature extraction.
 
 Returns:
 A dictionary with keys specifying the output_type (e.g., semantic
 prediction) and values storing Tensors representing predictions (argmax
 over channels). Each prediction has size [batch, height, width].
 """
 outputs_to_scales_to_logits = multi_scale_logits(
 images,
 model_options=model_options,
 image_pyramid=image_pyramid,
 is_training=False,
 fine_tune_batch_norm=False)
 
 predictions = {}
 for output in sorted(outputs_to_scales_to_logits):
 scales_to_logits = outputs_to_scales_to_logits[output]
 logits = tf.image.resize_bilinear(
 scales_to_logits[_MERGED_LOGITS_SCOPE],
 tf.shape(images)[1:3],
 align_corners=True)
 predictions[output] = tf.argmax(logits, 3)
 
 return predictions
 
 
 def scale_dimension(dim, scale):
 """Scales the input dimension.
 
 Args:
 dim: Input dimension (a scalar or a scalar Tensor).
 scale: The amount of scaling applied to the input.
 
 Returns:
 Scaled dimension.
 """
 if isinstance(dim, tf.Tensor):
 return tf.cast((tf.to_float(dim) - 1.0) scale + 1.0, dtype=tf.int32)
 else:
 return int((float(dim) - 1.0) scale + 1.0)
 
 
 def multi_scale_logits(images,
 model_options,
 image_pyramid,
 weight_decay=0.0001,
 is_training=False,
 fine_tune_batch_norm=False):
 """Gets the logits for multi-scale inputs.
 
 The returned logits are all downsampled (due to max-pooling layers)
 for both training and evaluation.
 更多查看：https://github.com/tensorflow/models/tree/master/research/deeplab