easy_vision.python.model.deeplab

easy_vision.python.model.deeplab.deeplab

class easy_vision.python.model.deeplab.deeplab.DeepLab(model_config, feature_dict, label_dict=None, mode='predict', categories=None)

Bases: easy_vision.python.model.cv_model.CVModel

implement deeplab v3+ model, which include an aspp head and a decoder head

__init__(model_config, feature_dict, label_dict=None, mode='predict', categories=None)

x.__init__(…) initializes x; see help(type(x)) for signature

build_loss_graph()

build_metric_graph(eval_config)

build metric graph for deeplab the following metrics are calculated:

mean_iou: TP / (TP + FP + FN) accuracy: TP / (TP + FP + FN + TN) positive_accuracy: TP / (TP + FN) negative_accuracy: TN / (TN + FP) positive_pixel_percentage: (TP + TN) / (TP + FP + TN + FN)

TP: true positive FP: false positive TN: true negative FN: false negative

Parameters:eval_config – protos.eval_config_pb2.EvalConfig instance Returns:a dict of tf.metrics ops

build_predict_graph()

create the forward graph

Returns: a dict of tensors

logits, batch_size * width * height * num_classes, tf.float probs, batch_size * width * height * num_classes, softmax result, tf.float preds, batch_size * width * height, tf.int32, predict label of each pixel

classmethod create_class(name)

get_outputs()

return a list of output key, which can be used to index output tensor result in prediction_dict

easy_vision.python.model.deeplab.deeplab_helper

easy_vision.python.model.deeplab.deeplab_helper.aspp_block(inputs, is_training, aspp_config)

build aspp block on top of inputs

Parameters:

• inputs – tensor of tf.float32, batch_size x height x width x channels
• is_training – bool, whether the training phase
• batchnorm_trainable – bool, whether to train the batchnorm parameters
• aspp_config – protos.aspp_block_pb2.ASPPBlock

Returns:

tensor of tf.float32

easy_vision.python.model.deeplab.deeplab_helper.split_separable_conv2d(inputs, num_outputs, kernel_size, rate=1, weight_decay=4e-05, depthwise_weights_initializer_stddev=0.33, pointwise_weights_initializer_stddev=0.06, scope=None)

Splits a separable conv2d into depthwise and pointwise conv2d. This operation differs from slim.separable_conv2d as this operation applies activation function between depthwise and pointwise conv2d.

Parameters:

• inputs – Input tensor with shape [batch, height, width, channels].
• num_outputs – Number of filters in the 1x1 pointwise convolution.
• kernel_size – A list of length 2: [kernel_height, kernel_width] of of the filters. Can be an int if both values are the same.
• rate – Atrous convolution rate for the depthwise convolution.
• weight_decay – The weight decay to use for regularizing the model.
• depthwise_weights_initializer_stddev – The standard deviation of the truncated normal weight initializer for depthwise convolution.
• pointwise_weights_initializer_stddev – The standard deviation of the truncated normal weight initializer for pointwise convolution.
• scope – Optional scope for the operation.

Returns:

Computed features after split separable conv2d.

easy_vision.python.model.deeplab.seg_decode_head

class easy_vision.python.model.deeplab.seg_decode_head.SegDecoderHead(feature_dict, head_config, label_dict=None, mode='predict')

Bases: easy_vision.python.model.cv_head.CVHead

implement the decoder head for semantic segmentation the decoder will refine the edge of raw segmentation results

__init__(feature_dict, head_config, label_dict=None, mode='predict')

x.__init__(…) initializes x; see help(type(x)) for signature

build_loss_graph()

create the loss graph, label values < 0 or >= 255 are ignored

build_postprocess_graph()

currently only resize to the original image shape TODO: add crf module

build_predict_graph()

fuse backbone features, then generate outputs

Returns:

tensor dict, which includes only logits of shape

[batch_size * height * width * num_classes], and not normalized by softmax