def train_selected_simulations(model,
typ,
epochs,
dropout_prob,
simtype,
q,
featuresize,
index=5000,
splits=5,
change_architecture=None,
save=True,
eval_set=None,
data_dir=r"/home/fr/fr_fr/fr_mw263"):
X, Y = preprocessing.get_simulations(data_dir=os.path.join(
data_dir, f"scripts/data/simulations/{simtype}"),
drop_parameters=True)
results = pd.read_csv(
os.path.join(data_dir,
f"output/grid_search/grid_search_results_{model}1.csv"))
best_model = results.iloc[results['mae_val'].idxmin()].to_dict()
if not change_architecture is None:
for item in change_architecture.items():
best_model[item[0]] = item[1]
dev = __import__(f"setup.dev_{model}", fromlist=["selected"])
X, Y = X[:index, :], Y[:index, :]
print(f"Fitting {model} {typ} with dropout prob {dropout_prob}")
print(f"Simulation type {simtype}")
dev.selected(X, Y, model, typ, best_model, epochs, splits,
simtype, featuresize, dropout_prob,
os.path.join(data_dir, "output"), save, eval_set)
import setup.preprocessing as preprocessing
import pandas as pd
import numpy as np
import time
from setup.dev_cnn import _selection_parallel
import multiprocessing as mp
import itertools
import torch.nn as nn
#%% Load Data
data_dir = r"/home/fr/fr_fr/fr_mw263"
X, Y = preprocessing.get_simulations(data_dir=os.path.join(
data_dir, r"scripts/data/simulations/uniform_params"),
drop_parameters=False)
# Use only first 10 percent of data.
ind = int(np.floor(X.shape[0] / 100 * 10))
X, Y = X[:ind], Y[:ind]
#%% Grid search of hparams
#%% Grid search of hparams
hiddensize = [8, 16, 32, 64, 128, 256]
batchsize = [8, 16, 32, 64, 128, 256]
learningrate = [1e-4, 1e-3, 5e-3, 1e-2]
history = [10, 14]
channels = [[10], [14], [28], [10, 20], [14, 28], [28, 52]]
kernelsize = [2, 3, 4]
activation = [nn.ReLU]
def pretraining(model, typ, epochs, dropout_prob, sims_fraction, q, simtype = None, featuresize = 7,
save = True, eval_set = None, data_dir = r"/home/fr/fr_fr/fr_mw263"):
"""
Args:
simtype: One out of None (for models 7 and 8) or ParamsFix, normal_params, uniform_params
"""
# Load the simulations
if typ == 7:
X, Y = preprocessing.get_simulations(data_dir = os.path.join(data_dir, r"scripts/data/simulations/normal_params"), drop_parameters=False)
if typ == 8:
X, Y = preprocessing.get_simulations(data_dir = os.path.join(data_dir, r"scripts/data/simulations/uniform_params"), drop_parameters=False)
if typ ==6:
X, Y = preprocessing.get_simulations(data_dir = os.path.join(data_dir, f"scripts/data/simulations/paramsFix"), drop_parameters=True)
# Architectures selected for Observations.
if typ == 5:
X, Y = preprocessing.get_simulations(data_dir = os.path.join(data_dir, f"scripts/data/simulations/paramsFix"), drop_parameters=True)
if typ == 9:
X, Y = preprocessing.get_simulations(data_dir = os.path.join(data_dir, f"scripts/data/simulations/paramsFix"), drop_parameters=True)
if typ == 10:
X, Y = preprocessing.get_simulations(data_dir = os.path.join(data_dir, f"scripts/data/simulations/uniform_params"), drop_parameters=True)
# Use full simulations or only parts of it?
if not sims_fraction is None:
print("Using only", sims_fraction, "% of the Simulations")
ind = int(np.floor(X.shape[0]/100*sims_fraction))
X, Y = X[:ind], Y[:ind]
# Set the hyperparameters and architecture of the network from preselection.
hparams, model_design = set_model_parameters(model, typ, epochs, featuresize, X, Y)
# Directory where fitted network is saved to
if dropout_prob == 0.0:
data_dir = os.path.join(data_dir, f"output/models/{model}{typ}/nodropout")
else:
data_dir = os.path.join(data_dir, f"output/models/{model}{typ}/dropout")
if not sims_fraction is None:
data_dir = os.path.join(data_dir, f"sims_frac{sims_fraction}")
else:
data_dir = os.path.join(data_dir, f"sims_frac100")
# Load script from which to use train_model_CV
dev = __import__(f"setup.dev_{model}", fromlist=["selected"])
# Train model
start = time.time()
running_losses,performance, y_tests, y_preds = dev.train_model_CV(hparams, model_design,
X, Y,
eval_set, dropout_prob,
data_dir, save)
end = time.time()
# Save: Results
if not simtype is None:
data_dir = os.path.join(data_dir, f"{simtype}")
# performance returns: rmse_train, rmse_test, mae_train, mae_test in this order.
performance = np.mean(np.array(performance), axis=0)
rets = [(end-start),
hparams["hiddensize"], hparams["batchsize"], hparams["learningrate"], hparams["history"], model_design["activation"],
performance[0], performance[1], performance[2], performance[3]]
results = pd.DataFrame([rets],
columns=["execution_time", "hiddensize", "batchsize", "learningrate", "history", "activation", "rmse_train", "rmse_val", "mae_train", "mae_val"])
results.to_csv(os.path.join(data_dir, r"selected_results.csv"), index = False)
# Save: Running losses, ytests and ypreds.
np.save(os.path.join(data_dir, "running_losses.npy"), running_losses)
np.save(os.path.join(data_dir, "y_tests.npy"), y_tests)
np.save(os.path.join(data_dir, "y_preds.npy"), y_preds)
import torch.nn as nn
import setup.preprocessing as preprocessing
import visualizations
import numpy as np
#%%
data_dir = "OneDrive\Dokumente\Sc_Master\Masterthesis\Project\DomAdapt"
#%%
X, Y = preprocessing.get_splits(sites=['le_bray'],
years=[2001, 2003],
datadir=os.path.join(data_dir, "data"),
dataset="profound",
simulations=None)
#%%
X_sims, Y_sims = preprocessing.get_simulations(data_dir=os.path.join(
data_dir, r"data\simulations\uniform_params"),
drop_parameters=True)
#%%
hparams = {
"batchsize": 256,
"epochs": 1000,
"history": 7,
"hiddensize": 128,
"learningrate": 0.01
}
model_design = {
"dimensions": [X.shape[1], 128, Y.shape[1]],
"activation": nn.ReLU,
"channels": [14, 28],
"kernelsize": 3
import setup.preprocessing as preprocessing
import visualizations
#%%
data_dir = "OneDrive\Dokumente\Sc_Master\Masterthesis\Project\DomAdapt"
X, Y = preprocessing.get_splits(sites = ['bily_kriz'],
years = [2001,2003, 2004, 2005, 2006],
datadir = os.path.join(data_dir, "data"),
dataset = "profound",
simulations = None)
X_test, Y_test = preprocessing.get_splits(sites = ['bily_kriz'],
years = [2008],
datadir = os.path.join(data_dir, "data"),
dataset = "profound",
simulations = None)
X_sims, Y_sims = preprocessing.get_simulations(os.path.join(data_dir, r"data\simulations\uniform_params"),
to_numpy = True,
DOY=False,
standardized=True)
X_sims_ss, Y_sims_ss = X_sims[:2000,], Y_sims[:2000,]
X_sims_ss_test, Y_sims_ss_test = X_sims[3000:3300,], Y_sims[3000:3300,]
#%%
hparams = {"batchsize": 256,
"epochs":3000,
"history":2,
"hiddensize":[32],
"learningrate":0.01}
model_design = {"dimensions": [X.shape[1], 32, Y.shape[1]],
"activation": nn.ReLU,
"featuresize": 7}
#eval_set = {"X_test":X_test, "Y_test":Y_test}
eval_set = {"X_test":X_sims_ss_test, "Y_test":Y_sims_ss_test}
