A Keras implementation of Convolutional Neural Networks (CNNs) used for the super-resolution of classical Ising configurations sampled with standard Monte Carlo methods.
This repository is still under construction. I update this as I am cleaning up the code used for the actual project.
It is possible to change the default loading/saving directories in data/directories.py file. There is another directories file for plots in plots/plot_directories.py to set the directories of .npy files to be plotted.
os, numpy, Tensorflow, Keras, argparse
Optional for linear regressions: scipy.stats
1) train.py: Trains a model. Depending on the -CR options this can be on critical temperature data or data in different temperatures. The metrics during training are saved as an .npy file and the Keras model graph as an .h5 file.
- Basic information:
-CR: True if data at critical temperature are used. False for data in different temperatures.
-RGWD: Use well-defined block spin RG for tied blocks. This assigns the sign of the upper left square. If False the sign is assigned randomly.
-L: Linear size of the output configuration. The input is L/2.
-Tind: Temperature indices to train on. The default T_list is defined in directories.py. It is not required to train on all temperatures every time we run the training script. Used only if -CR is False. If [] all temperatures are used.
- Sample numbers:
-nTR: Number of samples in the training data set. This is the number that appears in the file name.
-nTE: Number of samples in the test data set. This is the number that appears in the file name.
-TRS: Number of samples that we finally use for training.
-VALS: Number of samples that we finally use for validation (from the test data set).
- Architecture Hyperparameters:
-VER: Version to appear in model names in order to save multiple models with the same hyperparameters. Used only when -CR is True.
-PBC: Use periodic boundary conditions for padding.
-ACT: Activation function to use in hidden layers. Give string: can use any keras activation. Activation in the last layer is by default sigmoid.
-HF: List of hidden filters. Last layer filter is by default 1.
-K: List of kernels. Must have length one more than -HF.
- Compiling Hyperparameters:
-OPT: Optimizer. Use any keras optimizer
-CE: If True it uses cross-entropy as loss. If False it uses mean squared error.
-magR: Coefficient of the magnetization regularization in loss.
-enR: Coefficient of the energy regularization in loss.
- Training Hyperparameters
-EP: Number of epochs for training. If -ES is True this is ignored and turned to 1000.
-BS: Batch size for training. Also included in model name.
-ES: If True use keras early stopping. If False train for -EP epochs.
-ESpat: Early stopping patience. Only when -ES is True.
-ESdelta: Early stopping delta. Only when -ES is True.
-VB: Keras verbose for messages during training.
2) test.py: Tests a trained network by calculating thermodynamic quantities from the predicted output. It loads a trained keras model and use it to predict and then calculate quantities. The results are saved in .npy file with the format: (temperatures, output interpretation, quantity).
-
temperatures: The different temperatures. If
-CRisFalsethis is ommited. -
output intepretation: Five values: 0=Original MC, 1=Decimated RG, 2=Continuous SR, 3=Rounded SR, 4=Sampled SR.
-
quantity: 0=Magnetization, 1=Energy, 2=Susceptibility, 3=Heat Capacity, 4=Mag^2, 5=Mag^4, 6=En^2, 7=TPF(L/4), 8=TPF(L/2), 9=S0, 10=S1, 11=S2. TPF = Two-point function, S=Fourier transform of TPF (see https://arxiv.org/abs/1101.3281).
Some settings are the same with train.py and are not repeated here.
-Mind: Model index in the listdir listing.
-OUT: If True it saves the predicted continuous output in .npy files.
-Tind: Do the calculation for specific temperatures. If [] the calculation is done in all temperatures.
3) list_models.py: Lists trained models using listdir. This can be used to find the index required when we run test.py.
4) train_multiple_exponents.py: Runs the training for critical configurations and calculates critical exponents. The calculation is run multiple times according to -C setting. The .h5 graph is saved for every trained network and an .npy file with predicted observables. The format of this file is ... .
Some settings are the same with train.py and are not repeated here.
-C: Number of times to run training and calculation.
-UP: Number of upsamplings for the critical exponent calculation.
-PRreg: If True print regression results after every calculation.
-TPF: If True calculate two-point function in observables.
-CORR: If True calculate correlation length.
Accessing sizes larger than 128x128 sometimes gives memory error either in the network or in the vectorized quantity calculation. A solution is to do the calculation in batches instead of the whole dataset of (say) 100,000 samples. The relevant settings are:
-CB: If True it enables the calculation with batches.
-CBS: How many configurations in each batch.
-NUP: Number of upsamplings before starting the batch use. If NUP=3 the first three upsamplings will be done normally and after that batches will be used.
In the plot folder. To work we must change the NAME to the corresponding .npy file to be read.
1) plot_directories.py: Set directories of .npy files to read/plot.
2) plot_quantities.py: Plots for comparison between MC, RG and SR. Reads the different temperature .npy produced by test.py.
3) output_hist.py: Plots configurations from the output .npy files produced by test.py at different temperatures with -OUT enabled. Also plots the block sum histograms using different sampling methods.
4) plot_multiple_exponents.py: Calculates critical exponents using the .npy produced by train_multiple_exponents.py.