def save_aucs(self, summary_path, AUCs_path, **kwargs):
summaries = summaryProcessor.get_summaries_from_path(summary_path)
if not os.path.exists(AUCs_path):
Path(AUCs_path).mkdir(parents=True, exist_ok=False)
for _, summary in summaries.df.iterrows():
filename = summary.training_output_path.split("/")[-1]
utils.set_random_seed(summary.seed)
data_processor = DataProcessor(summary=summary)
data_loader = self.__get_data_loader(summary)
auc_params = self.model_evaluator.get_aucs(
summary=summary,
AUCs_path=AUCs_path,
filename=filename,
data_processor=data_processor,
data_loader=data_loader,
**kwargs)
if auc_params is None:
continue
(aucs, auc_path, append, write_header) = auc_params
self.__save_aucs_to_csv(aucs=aucs,
path=auc_path,
append=append,
write_header=write_header)
def get_qcd_test_data(self, summary, normalize=False):
utils.set_random_seed(summary.seed)
data_processor = DataProcessor(summary=summary)
data_loader = self.__get_data_loader(summary)
return self.model_evaluator.get_qcd_test_data(summary,
data_processor,
data_loader,
normalize=normalize)
def __init__(
self,
# general settings of the training
model_trainer_path,
validation_data_fraction,
test_data_fraction,
include_hlf,
include_efp,
hlf_to_drop,
# arguments that will be passed to the specialized trainer class
**training_settings):
"""
Constructor of the general Trainer class, which will delegate architecture-specific tasks to
a specialized Trainer class.
"""
# Import correct specialized class
self.model_class = utils.import_class(model_trainer_path)
# Save general training arguments
self.validation_data_fraction = validation_data_fraction
self.test_data_fraction = test_data_fraction
self.include_hlf = include_hlf
self.include_efp = include_efp
self.hlf_to_drop = hlf_to_drop
# Draw, set and save random seed
self.seed = np.random.randint(0, 99999999)
utils.set_random_seed(self.seed)
# Save training output path (used to save the model later on)
self.training_output_path = training_settings["training_output_path"]
# Prepare data processor and data loader for the specialized class
data_processor = DataProcessor(
validation_fraction=validation_data_fraction,
test_fraction=test_data_fraction,
seed=self.seed)
data_loader = DataLoader()
data_loader.set_params(include_hlf=include_hlf,
include_eflow=include_efp,
hlf_to_drop=hlf_to_drop)
# Initialize specialized trainer object
self.model_trainer = self.model_class(data_processor=data_processor,
data_loader=data_loader,
**training_settings)
def __init__(
self,
args,
train_dataloader,
valid_dataloader,
model,
device=torch.device("cpu"),
patience=5,
):
self.args = args
self.device = device
# data
self.num_class = train_dataloader.dataset.num_classes
self.train_dataloader = train_dataloader
self.valid_dataloader = valid_dataloader
# model
self.model = model
self.model.to(self.device)
# optimizer
self.set_optimizer()
# metric
self.set_metrics()
self.best_val = 0.0
self.best_train = 0.0
# early stopping
self.patience = patience
self.es_count = self.patience
# save dir
self.save_path = os.path.join(self.args.model_dir, "model_best.pkl")
if not os.path.isdir(self.args.model_dir):
os.mkdir(self.args.model_dir)
# tensorboardX writer
self.writer = SummaryWriter(os.path.join(self.args.model_dir, "train_logs"))
# set random seed
set_random_seed(self.args.random_seed)
def draw_roc_curves(self, summary_path, summary_version, **kwargs):
summaries = summaryProcessor.get_summaries_from_path(summary_path)
plotting_args = {
k: v
for k, v in kwargs.items()
if k not in ["signals", "signals_base_path"]
}
fig, ax_begin, ax_end, plt_end, colors = self.__get_plot_params(
n_plots=1, **plotting_args)
ax = ax_begin(0)
for _, summary in summaries.df.iterrows():
version = summaryProcessor.get_version(summary.summary_path)
if version != summary_version:
continue
utils.set_random_seed(summary.seed)
kwargs["filename"] = summary.training_output_path.split("/")[-1]
data_processor = DataProcessor(summary=summary)
data_loader = self.__get_data_loader(summary)
self.model_evaluator.draw_roc_curves(summary=summary,
data_processor=data_processor,
data_loader=data_loader,
ax=ax,
colors=colors,
**kwargs)
x = [i for i in np.arange(0, 1.1, 0.1)]
ax.plot(x, x, '--', c='black')
ax_end("false positive rate", "true positive rate")
plt_end()
plt.show()
parser.add_argument("target_domain", type=str, help="Target domain.")
parser.add_argument("output_path",
type=str,
help="Output path to store prediction.")
parser.add_argument("--batch_size",
type=int,
default=32,
help="Batch size.")
parser.add_argument("--random_seed",
type=int,
default=42,
help="random seed.")
args = parser.parse_args()
set_random_seed(args.random_seed)
dataset = DigitTestDataset(args.images_dir)
dataloader = DataLoader(dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=8)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# Model dir
if args.target_domain == "svhn":
model_dir = os.path.join("models", "dsn_m2s", "model_best.pth.tar")
else:
model_dir = os.path.join("models", "dsn_s2m", "model_best.pth.tar")
# Models
import random
import argparse
import numpy as np
import torch
import torch.nn.functional as F
import matplotlib.pyplot as plt
from torch.utils.data import DataLoader
from sklearn.manifold import TSNE
from module.utils import set_random_seed
from module.dataset.digit import DigitDataset
from module.da.dann import DANN
from module.dsn.dsn import DSN
set_random_seed(42)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
def plot_tsne(
source_features_ls,
source_labels_ls,
source_domains_ls,
target_features_ls,
target_labels_ls,
target_domains_ls,
fname,
sample_size=250,
):
source_features_ls = np.concatenate(source_features_ls)
source_labels_ls = np.concatenate(source_labels_ls)
fig_1_path = os.path.join(args.output_dir, "fig1_2.jpg")
fig_2_path = os.path.join(args.output_dir, "fig2_2.jpg")
dcgen = DCGen(latent_dim=128)
dcgen.load_state_dict(torch.load(args.dcgan_model))
dcgen.to(device)
acgen = ACGen(latent_dim=120,
n_feature_maps=128,
embedding_dim=4,
n_attributes=1)
acgen.load_state_dict(torch.load(args.acgan_model))
acgen.to(device)
# prepare dc noise
set_random_seed(int(args.random_seed[0]))
# dc_noise = torch.randn(64, 128, 1, 1, device=device)
# dc_noise[1], dc_noise[2], dc_noise[5] = dc_noise[42], dc_noise[50], dc_noise[54]
# dc_noise[9], dc_noise[18], dc_noise[28] = dc_noise[60], dc_noise[32], dc_noise[34]
# dc_noise[27] = dc_noise[63]
# dc_noise = dc_noise[:32]
# tmp_dc_noise = dc_noise.detach().cpu().numpy()
# with open("models/dcgan/fixed_noise.pkl", "wb") as fout:
# pickle.dump(tmp_dc_noise, fout)
with open(os.path.join("models", "dcgan", "fixed_noise.pkl"), "rb") as fin:
dc_noise = pickle.load(fin)
dc_noise = torch.tensor(dc_noise).float()
dc_noise = dc_noise.to(device)
def get_error(self, input_data, summary, scaler=None):
utils.set_random_seed(summary.seed)
data_processor = DataProcessor(summary=summary)
return self.model_evaluator.get_error(input_data, summary,
data_processor, scaler)
def get_signal_test_data(self, name, path, summary):
utils.set_random_seed(summary.seed)
data_processor = DataProcessor(summary=summary)
data_loader = self.__get_data_loader(summary)
return self.model_evaluator.get_signal_test_data(
name, path, summary, data_processor, data_loader)
def __init__(
self,
qcd_path,
training_params,
bottleneck_size,
intermediate_architecture=(30, 30),
test_data_fraction=0.15,
validation_data_fraction=0.15,
norm_type="",
norm_args=None,
hlf_to_drop=None,
):
"""
Creates auto-encoder trainer with random seed, provided training parameters and architecture.
Loads specified data, splits them into training, validation and test samples according to
provided arguments. Normalizes the data as specified by norm_percentile.
High-level features specified in hlf_to_drop will not be used for training.
"""
if hlf_to_drop is None:
hlf_to_drop = ['Energy', 'Flavor']
self.seed = np.random.randint(0, 99999999)
utils.set_random_seed(self.seed)
self.qcd_path = qcd_path
self.hlf_to_drop = hlf_to_drop
self.training_params = training_params
self.test_data_fraction = test_data_fraction
self.validation_data_fraction = validation_data_fraction
data_loader = DataLoader()
# Load QCD samples
(self.qcd, qcd_jets, qcd_event,
qcd_flavor) = data_loader.load_all_data(qcd_path,
"qcd background",
include_hlf=True,
include_eflow=True,
hlf_to_drop=hlf_to_drop)
data_processor = DataProcessor(
validation_fraction=self.validation_data_fraction,
test_fraction=self.test_data_fraction,
seed=self.seed)
(train_data, validation_data, test_data, _,
_) = data_processor.split_to_train_validate_test(data_table=self.qcd)
train_data.output_file_prefix = "qcd training data"
validation_data.output_file_prefix = "qcd validation data"
# Normalize the input
self.norm_type = norm_type
self.norm_args = norm_args
self.data_ranges = np.asarray([])
self.means_train, self.stds_train = None, None
self.means_validation, self.stds_validation = None, None
if norm_type == "Custom":
self.data_ranges = utils.percentile_normalization_ranges(
train_data, norm_args["norm_percentile"])
elif norm_type == "CustomStandard":
self.means_train, self.stds_train = train_data.get_means_and_stds()
self.means_validation, self.stds_validation = validation_data.get_means_and_stds(
)
print("Trainer scaler args: ", self.norm_args)
self.train_data_normalized = data_processor.normalize(
data_table=train_data,
normalization_type=self.norm_type,
norm_args=self.norm_args,
data_ranges=self.data_ranges,
means=self.means_train,
stds=self.stds_train)
self.validation_data_normalized = data_processor.normalize(
data_table=validation_data,
normalization_type=self.norm_type,
norm_args=self.norm_args,
data_ranges=self.data_ranges,
means=self.means_validation,
stds=self.stds_validation)
# Build the model
self.input_size = len(self.qcd.columns)
self.intermediate_architecture = intermediate_architecture
self.bottleneck_size = bottleneck_size
self.model = self.get_auto_encoder_model()
