grseb9s/3d_deep_learning

Deep Learning in 3D data

3D Deep learning

Python module for making 3D Deep Learning easier.

threeDlearn content:

• load3D
• keras_generator
• transforms
• visualization
• models
• weights
• datasets

load3D

Functions for reading several 3D file formats and generic array formats:

• .mat

• .npy / .npz

• .obj

• .off

• .ply

keras_generator

For generating augmented voxelizations of 3D data on-the-fly.

You can use VoxelGridDataGenerator as you would use kera's ImageDataGenerator:

from threeDlearn.models import Voxnet
from threeDlearn.keras_generator import VoxelGridDataGenerator

model = Voxnet(10, weights="voxnet10.h5")

# Finetune last layer
for layer in model.layers[:-1]:
    layer.trainable = False
    
gen = VoxelGridDataGenerator(z_rotation_range=10)
train_batches = gen.flow_from_directory("3DMNIST/train")

model.fit_generator(train_batches, train_batches.samples // train_batches.batch_size)

Current supported augmentations:

x_rotation_range : float, optional (Default None)
    Rotation range in Degrees (0-180) along the x axis.
    Equivalent to 'Roll' in aircraft principal axis.
    
y_rotation_range : float, optional (Default None)
    Rotation range in Degrees (0-180) along the y axis.
    Equivalent to 'Pitch' in aircraft principal axis.
    
z_rotation_range : float, optional (Default None)
    Rotation range in Degrees (0-180) along the z axis.
    Equivalent to 'Yaw' in aircraft principal axis.

x_shift_voxel_range : uint, optional (Default None)
    Number of voxels to be shifted along x axis.
    
y_shift_voxel_range : uint, optional (Default None)
    Number of voxels to be shifted along y axis.
    
z_shift_voxel_range : uint, optional (Default None)
    Number of voxels to be shifted along z axis.

x_flip : bool, optional (Default False)
    Flip around x axis with random probability
    
y_flip : bool, optional (Default False)
    Flip around y axis with random probability

z_flip : bool, optional (Default False)
    Flip around z axis with random probability

transforms

Functions used to generate augmented voxelgrids (see keras_generator above).

visualization

Use plot_feature_vector to visualize a voxelgrid sliced along the "z" axis.

models

Pre-defined models.

Currently avaliable:

• VoxNet ('VoxNet: A 3D Convolutional Neural Network for Real-Time Object Recognition.')

datasets

• 3DMNIST