def main():
###########################################################################
### SPECIFY DATA PATHS
P = '/local/meliao/projects/fourier_neural_operator/'
DATA_DIR = os.path.join(P, 'data/')
MODEL_DIR = os.path.join(P, 'experiments/31_different_activations/models')
PLOTS_DIR = os.path.join(P, 'experiments/31_different_activations/plots/')
RESULTS_DIR = os.path.join(P,
'experiments/31_different_activations/results')
if not os.path.isdir(PLOTS_DIR):
os.mkdir(PLOTS_DIR)
###########################################################################
### LOAD DATA
FP_FMT = "2021-09-29_NLS_data_00_{}_test.mat"
# DSET_KEYS = ['00', '01', '02', '03', '04']
DSET_NAME_DD = {
'00': 'Flat DFT Coeffs on [1, ..., 5]',
'01': 'GRF Original',
'02': 'GRF on [1, ..., 5]',
'03': 'GRF high coefficient decay',
'04': 'GRF low coefficient decay'
}
data_fp_dd = {
i: os.path.join(DATA_DIR, FP_FMT.format(i))
for i in DSET_NAME_DD.keys()
}
raw_data_dd = {k: sio.loadmat(v) for k, v in data_fp_dd.items()}
data_dd = {
k: OneStepDataSetComplex(v['output'], v['t'], v['x'])
for k, v in raw_data_dd.items()
}
# data_dd = sio.loadmat(DATA_FP)
# dset = OneStepDataSetComplex(data_dd['output'], data_dd['t'], data_dd['x'])
###########################################################################
### SET UP EXPERIMENT
experiment = Experiment.MultiDataExperiment(name="31")
experiment.register_data_variable(data_dd)
ACTIVATIONS = ['tanh', 'sigmoid', 'relu', 'sin']
# ACTIVATIONS = ['relu', 'sin']
K_ACTIVATION = 'activation'
experiment.register_new_variable(K_ACTIVATION, ACTIVATIONS)
###########################################################################
### LOAD MODELS
MODEL_FMT = "time_10_dset_{dataset}_activation_{activation}_ep1000"
experiment.load_models(MODEL_DIR, MODEL_FMT)
###########################################################################
### MAKE PREDICTIONS
with torch.no_grad():
experiment.run_prediction_and_errors(
prediction_fn=data_management.prediction_fn_composed,
data_spec_fn=data_management.data_spec_fn_fno,
error_fn=data_management.error_fn_l2_normalized)
sin_relu_set = set()
sin_relu_set.add('sin')
sin_relu_set.add('relu')
sin_relu_dd = {'activation': sin_relu_set}
for dset_key, dset_name in DSET_NAME_DD.items():
logging.info(f"Plotting errors on dset {dset_key}")
# make plot for all activation types
plot_fp = os.path.join(PLOTS_DIR,
f"dset_{dset_key}_composed_errors_all.png")
dset_vals = set()
dset_vals.add(dset_key)
filter_dd = {'dataset': dset_vals}
errors_dd = experiment.get_error_dd("{activation}",
remove_first_row=True,
**filter_dd)
plotting_utils.plot_time_errors(errors_dd=errors_dd,
title=f"Errors on dataset {dset_name}",
fp=plot_fp)
# make plot for sin and relu
plot_fp_i = os.path.join(
PLOTS_DIR, f"dset_{dset_key}_composed_errors_sin-relu.png")
sin_relu_dd['dataset'] = dset_vals
errors_dd_i = experiment.get_error_dd("{activation}",
remove_first_row=True,
**sin_relu_dd)
plotting_utils.plot_time_errors(errors_dd=errors_dd_i,
title=f"Errors on dataset {dset_name}",
fp=plot_fp_i)
# make predictions plot for sin and relu
preds_dd = experiment.get_preds_dd("{activation}", **sin_relu_dd)
for test_case in range(5):
plot_fp_i = os.path.join(
PLOTS_DIR,
f'test_case_{test_case}_dset_{dset_key}_sin-relu.png')
preds_dd_for_plt = {k: v[test_case] for k, v in preds_dd.items()}
solns = experiment.dataset_dd[dset_key].X[test_case]
plotting_utils.plot_one_testcase_panels(
preds_dd=preds_dd_for_plt,
solns=solns,
show_n_timesteps=5,
title=f"Testcase {test_case} on dataset {dset_name}",
fp=plot_fp_i)
def main():
###########################################################################
### SPECIFY DATA PATHS
logging.info("Beginning plotting")
P = '/local/meliao/projects/fourier_neural_operator/'
DATA_DIR = os.path.join(P, 'data/')
MODEL_DIR = os.path.join(P, 'experiments/31_different_activations/models')
PLOTS_DIR = os.path.join(P, 'experiments/32_scaling_for_freq/plots/')
if not os.path.isdir(PLOTS_DIR):
os.mkdir(PLOTS_DIR)
###########################################################################
### LOAD DATA
FP_FMT = "2021-09-29_NLS_data_00_{}_test.mat"
DSET_NAME_DD = {
# '00': 'Flat DFT Coeffs on [1, ..., 5]',
'01': 'GRF Original',
# '02': 'GRF on [1, ..., 5]',
# '03': 'GRF high coefficient decay',
'04': 'GRF low coefficient decay'
}
data_fp_dd = {
i: os.path.join(DATA_DIR, FP_FMT.format(i))
for i in DSET_NAME_DD.keys()
}
raw_data_dd = {k: sio.loadmat(v) for k, v in data_fp_dd.items()}
data_dd = {
k: OneStepDataSetComplex(v['output'], v['t'], v['x'])
for k, v in raw_data_dd.items()
}
scaled_data_dd = {
k: FreqScalingDataSet(raw_data_dd['04']['output'],
raw_data_dd['04']['t'],
raw_data_dd['04']['x'],
scale_param=1 / k)
for k in [2, 3]
}
###########################################################################
### SET UP COMPOSED PREDS EXPERIMENT AND MAKE PREDICTIONS
experiment1 = Experiment.MultiDataExperiment(name="Standard data dd")
experiment1.register_data_variable(data_dd)
ACTIVATIONS = ['relu', 'sin']
K_ACTIVATION = 'activation'
experiment1.register_new_variable(K_ACTIVATION, ACTIVATIONS)
MODEL_FMT = "time_10_dset_{dataset}_activation_{activation}_ep1000"
experiment1.load_models(model_dir=MODEL_DIR, model_fmt=MODEL_FMT)
with torch.no_grad():
experiment1.run_prediction_and_errors(
prediction_fn=data_management.prediction_fn_composed,
data_spec_fn=data_management.data_spec_fn_fno,
error_fn=data_management.error_fn_l2_normalized)
###########################################################################
### SET UP SCALING BY 2 EXPERIMENT
experiment2 = Experiment.MultiDataExperiment(name='Scaling dd')
experiment2.register_data_variable(scaled_data_dd)
experiment2.register_new_variable(K_ACTIVATION, ['relu', 'sin'])
experiment2.register_new_variable('train_dset', ['01', '03', '04'])
MODEL_FMT = "time_10_dset_{train_dset}_activation_{activation}_ep1000"
experiment2.load_models(model_dir=MODEL_DIR, model_fmt=MODEL_FMT)
with torch.no_grad():
experiment2.run_prediction_and_errors(
prediction_fn=data_management.prediction_fn_scale_for_freq,
data_spec_fn=data_management.data_spec_fn_fno,
error_fn=data_management.error_fn_l2_normalized)