Parametric curve in space fitting with PyTorch

This post is part of a series of posts on the fitting of mathematical objects (functions, curves and surfaces) through a MLP (Multi-Layer Perceptron) neural network; for an introduction on the subject please see the post Fitting with highly configurable multi layer perceptrons.

The topic of this post is the fitting of a parametric curve on a 3-D reference system $Oxyz$ with parameter $t$ within a closed interval of reals is defined with a triple of real functions $x(t) \colon [a,b] \to \rm I\!R$, $y(t) \colon [a,b] \to \rm I\!R$ and $z(t) \colon [a,b] \to \rm I\!R$ that return respectively the values of $x$, $y$ and $z$ coordinates as the parameter $t$ changes; equivalently a parametric curve in space is also defined with a vector function $f(t) \colon [a,b] \to {\rm I\!R x \rm I\!R x \rm I\!R}$ in this way: $$f(t) = \begin{bmatrix} x(t) \\ y(t) \\ z(t) \\ \end{bmatrix}$$ where the three components of the vectors of the function image are respectively the $x$, $y$ and $z$ coordinates of the curve on 3-D reference system.

The second definition of parametric curve suggests an architecture of single MLP where the input layer contains only one neuron because the dimension of domain is 1 instead the output layer contains 3 neurons because the dimension of codomain is 3.
As part of the fitting is to allow the user to test different combinations of MLP architectures, their own activation functions, training algorithm and loss function without writing code but working only on the command line of the four Python scripts (plus two variants) which separately implement the following features:

  • Dataset generation
  • MLP architecture definition + Training
  • Prediction
  • Visualization of the result
The related code requires the version 3 of Python and it uses PyTorch technology; it requires also NumPy and MatPlotLib libraries.
To get the code please see the paragraph Download of the complete code at the end of this post.

The exact same mechanism was created using TensorFlow 2 (with Keras) technology; see the post Parametric curve in space fitting with TensorFlow always published on this website.

Dataset generation

Goal of the pmc3t_gen.py Python program is to generate datasets (both training and test ones) to be used in later phases; it takes in command line the two component functions to be approximated (in lambda body syntax), the interval of independent parameter $t$ (begin, end and discretization step) and it generates the dataset in an output csv file applying the two functions to the passed interval of the parameter $t$.
In fact the output csv file has three columns (without header): first column contains the sorted values of independent parameter $t$ within the passed interval discretized by discretization step; second, third and fourth columns contain respectively the values of the components $x$, $y$ and $z$ ie the values of functions $x(t)$, $y(t)$ and $z(t)$ correspondent to values of $t$ of first column.

To get the program usage you can run this following command: 

$ python pmc3t_gen.py --help
and the output got is: 

usage: pmc3t_gen.py [-h] 
  -h, --help                  show this help message and exit
  --dsout DS_OUTPUT_FILENAME  dataset output file (csv format)
  --xt FUNCX_T_BODY           x=x(t) body (lamba format)
  --yt FUNCY_T_BODY           y=y(t) body (lamba format)
  --zt FUNCY_T_BODY           z=z(t) body (lamba format)
  --rbegin RANGE_BEGIN        begin range (default:-5.0)
  --rend RANGE_END            end range (default:+5.0)
  --rstep RANGE_STEP          step range (default: 0.01)
Please read the file README.md for a complete detail of the semantics of the parameters supported on the command line.

An example of using the program pmc3t_gen.py

Suppose you want to approximate in the range $[0,2\pi]$ the disrupted circle described by the following function: $$f(t) = \begin{bmatrix} x(t)= \cos t \\ y(t)= \sin t \\ z(t)= \frac{\cos 4t}{4} \\ \end{bmatrix}$$ Keeping in mind that np is the alias of NumPy library, the translation of this function in lambda body Python syntax is: 

np.cos(t)
np.sin(t)
np.cos(4 * t) / 4.0
To generate the training dataset, run the following command: 

$ python pmc3t_gen.py \
  --dsout mytrain.csv \
  --xt "np.cos(t)" \
  --yt "np.sin(t)" \
  --zt "np.cos(4 * t) / 4.0" \
  --rbegin 0.0 \
  --rend 6.28 \
  --rstep 0.01
instead to generate the test dataset, run the following command: 

$ python pmc3t_gen.py \
  --dsout mytest.csv \
  --xt "np.cos(t)" \
  --yt "np.sin(t)" \
  --zt "np.cos(4 * t) / 4.0" \
  --rbegin 0.0 \
  --rend 6.28 \
  --rstep 0.0475
Note that the discretization step of the test dataset is larger than that of training and it is a normal fact because the training, to be accurate, it must be run on more data. Also note that it is appropriate for the discretization step of the test dataset is not a multiple of the training one in order to ensure that the test dataset contains most of the data that is not present in training dataset, and this makes prediction more interesting.

MLP architecture definition + Training

This feature has two different implementations according to the 2nd mathematical definitions of parametric curve in space, namely the one that fits the vector function using a single MLP
The official version implemented by Python program pmc3t_fit.py creates dynamically an MLP and performs its training according to the passed parameters through the command line in order to fit the vector function that defines the curve on plane.
To get the official program usage you can run this following command: 

$ python pmc3t_fit.py --help
and the output got is: 

usage: pmc3t_fit.py [-h]
  --trainds TRAIN_DATASET_FILENAME
  --modelout MODEL_PATH
  [--epochs EPOCHS]
  [--batch_size BATCH_SIZE]
  [--hlayers HIDDEN_LAYERS_LAYOUT [HIDDEN_LAYERS_LAYOUT ...]]
  [--hactivations ACTIVATION_FUNCTIONS [ACTIVATION_FUNCTIONS ...]]
  [--optimizer OPTIMIZER]
  [--loss LOSS]
  [--device DEVICE]
Please read the file README.md for a complete detail of the semantics of the parameters supported on the command line of both programs.

An example of using the program pmc3t_fit.py

Suppose you have a training dataset available (for example generated through pmc3t_gen.py program as shown in the previous paragraph) and you want the MLP to have three hidden layers with respectively 150, 300 and 150 neurons each and you want to use the Tanh activation function exiting the three layers; moreover you want to perform 500 training epochs with a 200 items batch size using the Adamax with learning rate equal to 0.005 and loss function equal to MSELoss. To put all this into action, run the following command: 

$ python pmc3t_fit.py \
  --trainds mytrain.csv \
  --modelout mymodel.pth \
  --hlayers 150 300 150 \
  --hactivation 'Tanh()' 'Tanh()' 'Tanh()' \
  --epochs 500 \
  --batch_size 200 \
  --optimizer 'Adamax(lr=0.005)' \
  --loss 'MSELoss'
at the end of which the file mymodel.pth will contain the MLP model trained on mytrain.csv dataset according to the parameters passed on the command line.

Prediction

As with training, this feature also has two different implementations according to the two mathematical definitions of parametric curve on plane: the official implementation that makes the prediction using a single pre-trained MLP that fits the curve defined by a vector function.

Goal of the official version implemented by pmc3t_predict.py Python program is to apply the MLP model generated through pmc3t_fit.py to an input dataset (for example generated through pmc3t_gen.py program as shown in a previous paragraph); the execution of the program produces in output a csv file with three columns (without header): the first column contains the values of indepedent parameter $t$ taken from input dataset and the second and third columns contain the predicted values of the two components, ie the values of the prediction that fits the parametric curve $f(t)$ correspondent to values of $t$ of first column.

To get the official program usage you can run this following command: 

$ python pmc3t_predict.py --help
and the output got is: 

usage: pmc3t_predict.py [-h]
  --model MODEL_PATH
  --ds DATASET_FILENAME
  --predictionout PREDICTION_DATA_FILENAME
  [--device DEVICE]
Please read the file README.md for a complete detail of the semantics of the parameters supported on the command line.

An example of using the program pmc3t_predict.py

Suppose you have the test dataset mytest.csv available (for example generated through pmc3t_gen.py program as shown in a previous paragraph) and the trained model of MLP in the file mymodel.pth (generated through pmc3t_fit.py program as shown in the example of previous paragraph); run the following command: 

$ python pmc3t_predict.py \
  --model mymodel.pth \
  --ds mytest.csv \
  --predictionout myprediction.csv
at the end of which the file myprediction.csv will contain the fitting of the initial function.

Visualization of the result

Goal of the pmc3t_plot.py Python program is to visualize the prediction curve superimposed on initial dataset curve (test or training, as you prefer) and it allows the visual comparison of the two curves.
To get the program usage you can run this following command: 

$ python pmc3t_plot.py --help
and the output got is: 

usage: pmc3t_plot.py [-h]
  --ds DATASET_FILENAME
  --prediction PREDICTION_DATA_FILENAME
  [--savefig SAVE_FIGURE_FILENAME]
Please read the file README.md for a complete detail of the semantics of the parameters supported on the command line.

An example of using the program pmc3t_plot.py

Having the test dataset mytest.csv available (for example generated through pmc3t_gen.py program as shown in a previous paragraph) and the prediction csv file (generated through pmc3t_predict.py program as shown in the previous paragraph), to generate the two xy-scatter charts, execute the following command: 

$ python pmc3t_plot.py \
  --ds mytest.csv \
  --prediction myprediction.csv
which shows the two xy-scatter charts superimposed: in blue the input dataset, in red the prediction one.
Note: Given the stochastic nature of the training phase, your specific results may vary. Consider running the example a few times.

Chart generated by the program pmc3t_plot.py that shows the fitting done by the MLP of the curve $f(t) = \begin{bmatrix} x(t)= \cos t \\ y(t)= \sin t \\ z(t)= \frac{\cos 4t}{4} \\ \end{bmatrix}$

Examples of cascade use of the four programs

In the folder parametric-curve-in-space-fitting/examples there are five shell scripts that show the use in cascade of the four programs, with training and prediction in the official version, in various combinations of parameters (MLP architecture, activation functions, optimization algorithm, loss function, training procedure parameters). To run the five examples for official version, run the following commands: 

$ cd parametric-curve-in-space-fitting/examples
$ sh example1.sh
$ sh example2.sh
$ sh example3.sh
$ sh example4.sh
$ sh example5.sh
Note: Given the stochastic nature of these examples (relative to the training phase), your specific results may vary. Consider running the example a few times.

Download of the complete code

The complete code is available at GitHub.
These materials are distributed under MIT license; feel free to use, share, fork and adapt these materials as you see fit.
Also please feel free to submit pull-requests and bug-reports to this GitHub repository or contact me on my social media channels available on the top right corner of this page.