How to use the code¶
To install the ganpdfs package, just run the following:
pip install --upgrade ganpdfs
The ganpdfs pogram takes as an input argument a run card file in which all the input parameters such as the name of the PDF set and the architecture of the network models are defined.
ganpdfs runcard/reference.yml [-t TOT_REPLICAS_SIZE]
In addition, it takes optional arguments that are defined by the following flags:
-o/--output: name of the output folder
-f/--force: overwrite the previous folder
-t/--totrep: total number of replicas (Np+Ns)
-s/--hyperopt: number of hyper-parameter scan trials
-c/--cluster: enable hyper-parameter scan in cluster
In the run card file, the following keys are required:
q- float : initial scale at which the PDF grid will be computed
x_grid- str : format of the x-grid (options: custom/lhapdf)
architecture- str: network models (options: dnn/dcnn)
gen_parameters- dict: architecture of the Generator
disc_parameters- dict: architecture of the Discriminator
gan_parameters- dict: parameters that defines the adversarial network
batch_size- int : size of the batches per training (in prercentage)
ng_steps- int : number of steps to train the Generator per epoch (optional, default=3)
nd_steps- int : number of steps to train the Discriminator per epoch (optional, default=4)
An example of standard run card is shown below:
#############################################################################################
# Input PDF #
#############################################################################################
pdf: 210127-n3fit-002
#############################################################################################
# PDF Grids: #
# --------- #
# * Inittial scale q (in GeV) #
# * Options for x-grid: #
# - custom: Custom GANs xgrid as defined in the Module #
# - lhapdf: Use the same xgrid as in the input PDF #
#############################################################################################
q : 1.65
x_grid : standard
#############################################################################################
# GAN setup: #
# --------- #
# * Options for architecture: #
# - dnn : Deep Neural Network #
# - dcnn: Deep Convolutional Neural Network #
#############################################################################################
use_saved_model : False
architecture : cnn
gan_parameters:
optimizer:
optimizer_name : RMSprop
learning_rate : 0.00005
loss : wasserstein
gen_parameters:
size_networks : 2
kernel_initializer : glorot_uniform
disc_parameters:
size_networks : 3
number_nodes : 450
use_bias : False
bias_initializer : zeros
kernel_initializer : glorot_uniform
weights_constraints : 0.01
optimizer:
optimizer_name : RMSprop
learning_rate : 0.00005
loss : wasserstein
activation : leakyrelu
ConvoluteOutput : False
#############################################################################################
# Training Setup: #
# -------------- #
# * batch size #
# * {i}_steps: number of steps to train a {i}={generator, discriminator/critic} at each #
# iteration. #
#############################################################################################
nd_steps : 4
ng_steps : 1
batch_size : 70
epochs : 1000
In order a hyper-parameter scan, the run card has to be slightly modified:
#############################################################################################
# Input PDF #
#############################################################################################
pdf: NNPDF40_nnlo_as_0118_1000rep
#############################################################################################
# PDF Grids: #
# --------- #
# * Inittial scale q (in GeV) #
# * Options for x-grid: #
# - custom: Custom GANs xgrid as defined in the Module #
# - lhapdf: Use the same xgrid as in the input PDF #
#############################################################################################
q : 1.65
x_grid : standard
use_saved_model : False
ConvoluteOutput : False
architecture : cnn
hyperopt:
#############################################################################################
# GAN setup: #
# --------- #
# * Options for architecture: #
# - dnn : Deep Neural Network #
# - dcnn: Deep Convolutional Neural Network #
#############################################################################################
gan_parameters:
optimizer:
optimizer_name : hp.choice('gan_opt', ['RMSprop', 'Adadelta'])
learning_rate : hp.choice('gan_lr', [0.00005, 0.0005])
gen_parameters:
size_networks : hp.choice('g_nn', [1, 2])
kernel_initializer : hp.choice('g_kini', ['GlorotUniform', 'RandomUniform'])
disc_parameters:
size_networks : hp.choice('d_nn', [1, 2])
number_nodes : hp.choice('d_nodes', [250, 450, 650, 1000])
kernel_initializer : hp.choice('d_kini', ['GlorotUniform', 'RandomUniform'])
weights_constraints : hp.choice('d_wc', [0.01, 0.1, 1])
optimizer:
optimizer_name : hp.choice('d_opt', ['RMSprop', 'Adadelta'])
learning_rate : hp.choice('d_lr', [0.00005, 0.0005])
activation : hp.choice('d_act', ['relu', 'leakyrelu'])
trainable : hp.choice('d_train', [True, False])
#############################################################################################
# Training Setup: #
# -------------- #
# * batch size #
# * {i}_steps: number of steps to train a {i}={generator, discriminator/critic} at each #
# iteration. #
#############################################################################################
nd_steps : hp.choice('nd_steps', [2, 3, 4])
ng_steps : hp.choice('ng_steps', [1, 2, 3])
batch_size : hp.choice('batch_size', [30, 50, 70])
epochs : hp.choice('epochs', [1000, 1500, 2000])
During the hyper-parameter scan, the models are optimized w.t.r to a similarity metric measue known as the Fréchet Inception Distance which measure the quality of the generated PDF.
How to generate PDF grid¶
The above will generate a folder named after the prior PDF name, appended with _enhanced. In order to evolve the PDFs, it suffices to run:
evolven3fit <PRIOR_PDF_NAME>_enhanced <TOT_REPLICAS_SIZE>
Then, to link the enhanced PDF set to the LHAPDF data directory, run the following:
postgans --pdf <PRIOR_PDF_NAME> --nenhanced <TOT_REPLICAS_SIZE>