thirdparty/learning3d/models/pointconv.py

import torch
import torch.nn as nn
import torch.nn.functional as F
from .. utils import PointConvDensitySetAbstraction

class PointConvDensityClsSsg(torch.nn.Module):
	def __init__(self, emb_dims=1024, input_shape="bnc", input_channel_dim=3, classifier=False, num_classes=40, pretrained=None):
		super(PointConvDensityClsSsg, self).__init__()
		if input_shape not in ["bnc", "bcn"]:
			raise ValueError("Allowed shapes are 'bcn' (batch * channels * num_in_points), 'bnc' ")
		self.input_shape = input_shape
		self.emb_dims = emb_dims
		self.classifier = classifier
		self.input_channel_dim = input_channel_dim
		self.create_structure()
		if self.classifier: self.create_classifier(num_classes)

	def create_structure(self):
		# Arguments to define PointConv network using PointConvDensitySetAbstraction class.
			# npoint:			number of points sampled from input.
			# nsample:			number of neighbours chosen for each point in sampled point cloud.
			# in_channel:		number of channels in input.
			# mlp:				sizes of multi-layer perceptrons.
			# bandwidth:		used to compute gaussian density.
			# group_all:		group all points from input to a single point if set to True.
		self.sa1 = PointConvDensitySetAbstraction(npoint=512, nsample=32, in_channel=self.input_channel_dim, 
													mlp=[64, 64, 128], bandwidth = 0.1, group_all=False)
		self.sa2 = PointConvDensitySetAbstraction(npoint=128, nsample=64, in_channel=128 + 3, 
													mlp=[128, 128, 256], bandwidth = 0.2, group_all=False)
		self.sa3 = PointConvDensitySetAbstraction(npoint=1, nsample=None, in_channel=256 + 3, 
													mlp=[256, 512, self.emb_dims], bandwidth = 0.4, group_all=True)

	def create_classifier(self, num_classes):
		# These are simple fully-connected layers with batch-norm and dropouts.
		# This architecture is given by PointConv paper. Hence, I used it here as a default version.
		# This can be easily modified by overwriting this function or by using classifier.py class.
		self.fc1 = nn.Linear(self.emb_dims, 512)
		self.bn1 = nn.BatchNorm1d(512)
		self.drop1 = nn.Dropout(0.7)
		self.fc2 = nn.Linear(512, 256)
		self.bn2 = nn.BatchNorm1d(256)
		self.drop2 = nn.Dropout(0.7)
		self.fc3 = nn.Linear(256, num_classes)		

	def forward(self, input_data):
		if self.input_shape == "bnc":
			input_data = input_data.permute(0, 2, 1)
		batch_size = input_data.shape[0]

		# Convert point clouds to latent features using PointConv network.
		l1_points, l1_features = self.sa1(input_data[:, :3, :], input_data[:, 3:, :])
		l2_points, l2_features = self.sa2(l1_points, l1_features)
		l3_points, l3_features = self.sa3(l2_points, l2_features)
		features = l3_features.view(batch_size, self.emb_dims)

		if self.classifier:
			# Use these features to classify the input point cloud.
			features = self.drop1(F.relu(self.bn1(self.fc1(features))))
			features = self.drop2(F.relu(self.bn2(self.fc2(features))))
			features = self.fc3(features)
			output = F.log_softmax(features, -1)
		else:
			# Return the PointConv features for the use of other higher level tasks.
			output = features

		return output

def create_pointconv(classifier=False, pretrained=None):
	if classifier and pretrained is not None:
		class Network(torch.nn.Module):
			def __init__(self, emb_dims=1024, input_shape="bnc", input_channel_dim=3, classifier=False, num_classes=40, pretrained=None):
				# Arguments:
					# emb_dims:				Size of embeddings.
					# input_shape:			Shape of input point cloud.
					# input_channel_dim:	Number of channels in point cloud. [eg. Nx3 (only points) or Nx6 (points + normals)]
					# classifier:			Do you want to use default classifier layers or just the embedding layers.
					# num_classes:			If you use classifier then decide the number of classes in your dataset.
					# use_pretrained:		Use pretrained classification network.
				super(PointConv, self).__init__()
				self.pointconv = PointConvDensityClsSsg(emb_dims, input_shape, input_channel_dim, classifier, num_classes)
				# super().__init__(emb_dims, input_shape, input_channel_dim, classifier, num_classes)
				if classifier and pretrained is not None:
					self.use_pretrained(pretrained)

			def use_pretrained(self, pretrained):
				checkpoint = torch.load(pretrained, map_location='cpu')
				self.pointconv.load_state_dict(checkpoint['model_state_dict'])

			def forward(self, input_data):
				return self.pointconv(input_data)
		return Network
	else:
		class Network(PointConvDensityClsSsg):
			def __init__(self, emb_dims=1024, input_shape="bnc", input_channel_dim=3, classifier=False, num_classes=40, pretrained=None):
				super().__init__(emb_dims=emb_dims, input_shape=input_shape, input_channel_dim=input_channel_dim, classifier=classifier, num_classes=num_classes, pretrained=pretrained)
		return Network


if __name__ == '__main__':
	# Test the code.
	x = torch.rand((2,1024,3))

	PointConv = create_pointconv(classifier=False, pretrained='checkpoint.pth')
	pc = PointConv(input_channel_dim=3, classifier=False, pretrained='checkpoint.pth')
	y = pc(x)
	print("Network Architecture: ")
	print(pc)
	print("Input Shape of PointNet: ", x.shape, "\nOutput Shape of PointNet: ", y.shape)