from typing import Optional, Union
from segmentation_models_pytorch.encoders import get_encoder
from segmentation_models_pytorch.base import (
SegmentationModel,
SegmentationHead,
ClassificationHead,
)
from .decoder import PSPDecoder
[docs]class PSPNet(SegmentationModel):
"""PSPNet_ is a fully convolution neural network for image semantic segmentation. Consist of
*encoder* and *Spatial Pyramid* (decoder). Spatial Pyramid build on top of encoder and does not
use "fine-features" (features of high spatial resolution). PSPNet can be used for multiclass segmentation
of high resolution images, however it is not good for detecting small objects and producing accurate,
pixel-level mask.
Args:
encoder_name: Name of the classification model that will be used as an encoder (a.k.a backbone)
to extract features of different spatial resolution
encoder_depth: A number of stages used in encoder in range [3, 5]. Each stage generate features
two times smaller in spatial dimensions than previous one (e.g. for depth 0 we will have features
with shapes [(N, C, H, W),], for depth 1 - [(N, C, H, W), (N, C, H // 2, W // 2)] and so on).
Default is 5
encoder_weights: One of **None** (random initialization), **"imagenet"** (pre-training on ImageNet) and
other pretrained weights (see table with available weights for each encoder_name)
psp_out_channels: A number of filters in Spatial Pyramid
psp_use_batchnorm: If **True**, BatchNorm2d layer between Conv2D and Activation layers
is used. If **"inplace"** InplaceABN will be used, allows to decrease memory consumption.
Available options are **True, False, "inplace"**
psp_dropout: Spatial dropout rate in [0, 1) used in Spatial Pyramid
in_channels: A number of input channels for the model, default is 3 (RGB images)
classes: A number of classes for output mask (or you can think as a number of channels of output mask)
activation: An activation function to apply after the final convolution layer.
Available options are **"sigmoid"**, **"softmax"**, **"logsoftmax"**, **"tanh"**, **"identity"**,
**callable** and **None**.
Default is **None**
upsampling: Final upsampling factor. Default is 8 to preserve input-output spatial shape identity
aux_params: Dictionary with parameters of the auxiliary output (classification head). Auxiliary output is build
on top of encoder if **aux_params** is not **None** (default). Supported params:
- classes (int): A number of classes
- pooling (str): One of "max", "avg". Default is "avg"
- dropout (float): Dropout factor in [0, 1)
- activation (str): An activation function to apply "sigmoid"/"softmax"
(could be **None** to return logits)
Returns:
``torch.nn.Module``: **PSPNet**
.. _PSPNet:
https://arxiv.org/abs/1612.01105
"""
def __init__(
self,
encoder_name: str = "resnet34",
encoder_weights: Optional[str] = "imagenet",
encoder_depth: int = 3,
psp_out_channels: int = 512,
psp_use_batchnorm: bool = True,
psp_dropout: float = 0.2,
in_channels: int = 3,
classes: int = 1,
activation: Optional[Union[str, callable]] = None,
upsampling: int = 8,
aux_params: Optional[dict] = None,
):
super().__init__()
self.encoder = get_encoder(
encoder_name,
in_channels=in_channels,
depth=encoder_depth,
weights=encoder_weights,
)
self.decoder = PSPDecoder(
encoder_channels=self.encoder.out_channels,
use_batchnorm=psp_use_batchnorm,
out_channels=psp_out_channels,
dropout=psp_dropout,
)
self.segmentation_head = SegmentationHead(
in_channels=psp_out_channels,
out_channels=classes,
kernel_size=3,
activation=activation,
upsampling=upsampling,
)
if aux_params:
self.classification_head = ClassificationHead(in_channels=self.encoder.out_channels[-1], **aux_params)
else:
self.classification_head = None
self.name = "psp-{}".format(encoder_name)
self.initialize()