def init(args):
if not exists(args.dir_path + '/particlenet/'):
mkdir(args.dir_path + '/particlenet/')
args_dict = vars(args)
dirs = ['cmodels', 'closses', 'cargs', 'couts']
for dir in dirs:
args_dict[dir + '_path'] = args.dir_path + '/particlenet/' + dir + '/'
if not exists(args_dict[dir + '_path']):
mkdir(args_dict[dir + '_path'])
args = utils.objectview(args_dict)
args.datasets_path = args.dir_path + '/datasets/'
args.outs_path = args.dir_path + '/outs/'
setup.init_logging(args)
prev_models = [f[:-4]
for f in listdir(args.cargs_path)] # removing txt part
if (args.name in prev_models):
logging.info("name already used")
# if(not args.load_model):
# sys.exit()
else:
try:
mkdir(args.closses_path + args.name)
except FileExistsError:
logging.debug("losses dir exists")
try:
mkdir(args.cmodels_path + args.name)
except FileExistsError:
logging.debug("models dir exists")
if (not args.load_model):
f = open(args.cargs_path + args.name + ".txt", "w+")
f.write(str(vars(args)))
f.close()
else:
temp = args.start_epoch, args.num_epochs
f = open(args.cargs_path + args.name + ".txt", "r")
args_dict = vars(args)
load_args_dict = eval(f.read())
for key in load_args_dict:
args_dict[key] = load_args_dict[key]
args = utils.objectview(args_dict)
f.close()
args.load_model = True
args.start_epoch, args.num_epochs = temp
args.device = device
return args
def load_args(args):
if args.load_model:
if args.start_epoch == -1:
prev_models = [
int(f[:-3].split('_')[-1])
for f in listdir(args.models_path + args.name + '/')
]
if len(prev_models):
args.start_epoch = max(prev_models)
else:
logging.debug("No model to load from")
args.start_epoch = 0
args.load_model = False
if args.start_epoch == 0: args.load_model = False
else:
args.start_epoch = 0
if (not args.load_model):
f = open(args.args_path + args.name + ".txt", "w+")
f.write(str(vars(args)))
f.close()
elif (not args.override_args):
temp = args.start_epoch, args.num_epochs
f = open(args.args_path + args.name + ".txt", "r")
args_dict = vars(args)
load_args_dict = eval(f.read())
for key in load_args_dict:
args_dict[key] = load_args_dict[key]
args = utils.objectview(args_dict)
f.close()
args.load_model = True
args.start_epoch, args.num_epochs = temp
return args
def init_project_dirs(args):
if args.dir_path == "":
if args.n: args.dir_path = "/graphganvol/mnist_graph_gan/jets"
elif args.lx:
args.dir_path = "/eos/user/r/rkansal/mnist_graph_gan/jets"
else:
args.dir_path = dirname(realpath(__file__))
args_dict = vars(args)
dirs = [
'models', 'losses', 'args', 'figs', 'datasets', 'err', 'evaluation',
'outs', 'noise'
]
for dir in dirs:
args_dict[dir + '_path'] = args.dir_path + '/' + dir + '/'
if not exists(args_dict[dir + '_path']):
mkdir(args_dict[dir + '_path'])
args = utils.objectview(args_dict)
return args
#
# realw1m = [0.00584264, 0.00556786, 0.0014096]
# realw1std = [0.00214083, 0.00204827, 0.00051136]
batch_size = 128
normal_dist = Normal(torch.tensor(0.).to(device), torch.tensor(0.2).to(device))
dir = './'
# dir = '/graphganvol/mnist_graph_gan/jets/'
args = utils.objectview({
'dataset_path': dir + 'datasets/',
'num_hits': 30,
'coords': 'polarrel',
'latent_node_size': 32,
'clabels': 0,
'jets': 'g',
'norm': 1,
'mask': False
})
X = JetsDataset(args)
labels = X[:][1]
# X_loaded = DataLoader(X, shuffle=True, batch_size=32, pin_memory=True)
X = X[:][0]
N = len(X)
rng = np.random.default_rng()
num_samples = 100000
args_txt = {
'g': 'args/218_g30_mask_c_dea_no_pos_diffs.txt',
't': 'args/206_t30_mask_c_lrx2_dea_no_pos_diffs.txt',
'q': 'args/230_q30_mask_c_lrx05_dea_no_pos_diffs.txt'
}
samples_dict = {'g': {}, 't': {}, 'q': {}}
for dataset in samples_dict.keys():
print(dataset)
args = eval(open(args_txt[dataset]).read())
args['device'] = torch.device('cuda')
args['datasets_path'] = './datasets/'
args['fpnd_batch_size'] = 512
args['evaluation_path'] = './evaluation/'
args = utils.objectview(args)
C, mu2, sigma2 = evaluation.load(args)
X = JetsDataset(args, train=False)
rng = np.random.default_rng()
X_loaded = DataLoader(TensorDataset(torch.tensor(X[:50000][0])),
batch_size=256)
C.eval()
for i, gen_jets in tqdm(enumerate(X_loaded), total=len(X_loaded)):
gen_jets = gen_jets[0]
if args.mask:
mask = gen_jets[:, :, 3:4] >= 0
gen_jets = (gen_jets * mask)[:, :, :3]
if (i == 0):
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
batch_size = 128
normal_dist = Normal(torch.tensor(0.).to(device), torch.tensor(0.2).to(device))
# dir = './'
dir = '/graphganvol/mnist_graph_gan/jets/'
args = {
'dataset_path': dir + 'datasets/',
'num_hits': 30,
'coords': 'polarrel',
'latent_node_size': 32
}
X = JetsDataset(utils.objectview(args))
N = len(X)
name = '7_batch_size_128_coords_polarrel'
full_path = dir + 'models/' + name + '/'
rng = np.random.default_rng()
num_samples = np.array([100, 1000, 10000])
num_batches = np.array(100000 / num_samples, dtype=int)
num_batches
epochs = 980
s1 = s1[:-2] + "]"
s2 = s2[:-2] + "]"
print(s1)
print(s2)
np.argsort(np.linalg.norm(w1m[:, :3], axis=1))[:20] * 5
# realw1m = [0.00584264, 0.00556786, 0.0014096]
# realw1std = [0.00214083, 0.00204827, 0.00051136]
batch_size = 128
normal_dist = Normal(torch.tensor(0.).to(device), torch.tensor(0.2).to(device))
dir = './'
args = utils.objectview({'datasets_path': dir + 'datasets/', 'figs_path': dir + 'figs/' + str(model), 'node_feat_size': 3, 'num_hits': 30, 'coords': 'polarrel', 'latent_node_size': 32, 'clabels': 1, 'jets': 'g', 'norm': 1, 'mask': False, 'mask_manual': False, 'real_only': False, 'mask_feat': False})
args = eval(open("./args/" + "179_t30_graphcnngan_knn_20.txt").read())
args['device'] = device
args['datasets_path'] = dir + 'datasets/'
# args['mask_feat'] = False
# args['mask_learn'] = False
# args['mask_c'] = False
args['figs_path'] = dir + 'figs/' + str(model) + '/' + str(epoch)
args = utils.objectview(args)
args
X = JetsDataset(args)
loadX = DataLoader(X, batch_size=128)
def init(args):
torch.manual_seed(4)
torch.autograd.set_detect_anomaly(True)
args.model_path = args.dir_path + '/models/'
args.losses_path = args.dir_path + '/losses/'
args.args_path = args.dir_path + '/args/'
args.figs_path = args.dir_path + '/figs/'
args.dataset_path = args.dir_path + '/datasets/'
args.err_path = args.dir_path + '/err/'
args.eval_path = args.dir_path + '/evaluation/'
args.noise_path = args.dir_path + '/noise/'
if (not exists(args.model_path)):
mkdir(args.model_path)
if (not exists(args.losses_path)):
mkdir(args.losses_path)
if (not exists(args.args_path)):
mkdir(args.args_path)
if (not exists(args.figs_path)):
mkdir(args.figs_path)
if (not exists(args.err_path)):
mkdir(args.err_path)
# if(not exists(args.noise_path)):
# mkdir(args.noise_path)
if (not exists(args.dataset_path)):
mkdir(args.dataset_path)
prev_models = [f[:-4] for f in listdir(args.args_path)] # removing .txt
if (args.name in prev_models):
print("name already used")
# if(not args.load_model):
# sys.exit()
else:
mkdir(args.losses_path + args.name)
mkdir(args.model_path + args.name)
mkdir(args.figs_path + args.name)
if args.load_model:
if args.start_epoch == -1:
prev_models = [
int(f[:-3].split('_')[-1])
for f in listdir(args.model_path + args.name + '/')
]
if len(prev_models):
args.start_epoch = max(prev_models)
else:
print("No model to load from")
args.start_epoch = 0
args.load_model = False
if args.start_epoch == 0: args.load_model = False
else:
args.start_epoch = 0
if (not args.load_model):
f = open(args.args_path + args.name + ".txt", "w+")
f.write(str(vars(args)))
f.close()
elif (not args.override_args):
temp = args.start_epoch, args.num_epochs
f = open(args.args_path + args.name + ".txt", "r")
args_dict = vars(args)
load_args_dict = eval(f.read())
for key in load_args_dict:
args_dict[key] = load_args_dict[key]
args = utils.objectview(args_dict)
f.close()
args.load_model = True
args.start_epoch, args.num_epochs = temp
args.device = device
return args
import evaluation
import torch
from torch.utils.data import DataLoader
num_samples = 50000
samples_dict = {}
for dataset in ['g', 't', 'q']:
args = utils.objectview({
'datasets_path': 'datasets/',
'ttsplit': 0.7,
'node_feat_size': 3,
'num_hits': 30,
'coords': 'polarrel',
'dataset': 'jets',
'clabels': 0,
'jets': dataset,
'norm': 1,
'mask': True,
'real_only': False
})
X = JetsDataset(args, train=False)
X = X[:][0]
X_rn, mask_real = utils.unnorm_data(args,
X[:num_samples].cpu().detach().numpy(),
real=True)
samples_dict[dataset] = (X_rn, mask_real)
efps = {}
for dataset in samples_dict.keys():
def init(args):
torch.manual_seed(4)
torch.autograd.set_detect_anomaly(True)
args.model_path = args.dir_path + '/models/'
args.losses_path = args.dir_path + '/losses/'
args.args_path = args.dir_path + '/args/'
args.figs_path = args.dir_path + '/figs/'
args.dataset_path = args.dir_path + '/raw/' if not args.sparse_mnist else args.dir_path + '/mnist_dataset/'
args.err_path = args.dir_path + '/err/'
args.eval_path = args.dir_path + '/evaluation/'
args.noise_path = args.dir_path + '/noise/'
if (not exists(args.model_path)):
mkdir(args.model_path)
if (not exists(args.losses_path)):
mkdir(args.losses_path)
if (not exists(args.args_path)):
mkdir(args.args_path)
if (not exists(args.figs_path)):
mkdir(args.figs_path)
if (not exists(args.err_path)):
mkdir(args.err_path)
if (not exists(args.noise_path)):
mkdir(args.noise_path)
if (not exists(args.dataset_path)):
mkdir(args.dataset_path)
print("Downloading dataset")
if (not args.sparse_mnist):
import tarfile, urllib
# url = 'http://ls7-www.cs.uni-dortmund.de/cvpr_geometric_dl/mnist_superpixels.tar.gz'
url = 'https://ls7-www.cs.tu-dortmund.de/fileadmin/ls7-www/misc/cvpr/mnist_superpixels.tar.gz'
try:
# python2
file_tmp = urllib.urlretrieve(url)[0]
except:
# python3
file_tmp = urllib.request.urlretrieve(url)[0]
tar = tarfile.open(file_tmp)
tar.extractall(args.dataset_path)
else:
import requests
r = requests.get(
'https://pjreddie.com/media/files/mnist_train.csv',
allow_redirects=True)
open(args.dataset_path + 'mnist_train.csv', 'wb').write(r.content)
r = requests.get('https://pjreddie.com/media/files/mnist_test.csv',
allow_redirects=True)
open(args.dataset_path + 'mnist_test.csv', 'wb').write(r.content)
print("Downloaded dataset")
prev_models = [f[:-4] for f in listdir(args.args_path)] # removing .txt
if (args.name in prev_models):
print("name already used")
# if(not args.load_model):
# sys.exit()
else:
mkdir(args.losses_path + args.name)
mkdir(args.model_path + args.name)
mkdir(args.figs_path + args.name)
if args.load_model:
if args.start_epoch == -1:
prev_models = [
int(f[:-3].split('_')[-1])
for f in listdir(args.model_path + args.name + '/')
]
if len(prev_models):
args.start_epoch = max(prev_models)
else:
print("No model to load from")
args.start_epoch = 0
args.load_model = False
else:
args.start_epoch = 0
if (not args.load_model):
f = open(args.args_path + args.name + ".txt", "w+")
f.write(str(vars(args)))
f.close()
elif (not args.override_args):
temp = args.start_epoch, args.num_epochs
f = open(args.args_path + args.name + ".txt", "r")
args_dict = vars(args)
load_args_dict = eval(f.read())
for key in load_args_dict:
args_dict[key] = load_args_dict[key]
args = utils.objectview(args_dict)
f.close()
args.load_model = True
args.start_epoch, args.num_epochs = temp
args.device = device
return args
plt.ylabel('True label')
plt.xlim(-1, len(target_names))
plt.ylim(-1, len(target_names))
plt.xlabel('Predicted label\naccuracy={:0.4f}; misclass={:0.4f}'.format(
accuracy, misclass))
plt.tight_layout()
plt.savefig(fname + '.pdf')
plt.close(fig)
return fig, ax
args = utils.objectview({
'mask': False,
'datasets_path': dir_path + 'datasets/',
'node_feat_size': 3,
'num_hits': 30
})
test_dataset = JetsClassifierDataset(args, train=False)
print("dataset loaded")
test_loaded = DataLoader(test_dataset)
C = ParticleNet(args.num_hits,
args.node_feat_size,
num_classes=5,
device=device).to(device)
C.load_state_dict(
torch.load(dir_path + 'particlenet/cmodels/c5_pnet_adam/C_18.pt',
