Skip to content

Commit

Permalink
Before parallelization
Browse files Browse the repository at this point in the history
  • Loading branch information
Guillaume Levy committed Jan 10, 2023
1 parent 6d2975e commit 9ed9b65
Show file tree
Hide file tree
Showing 4 changed files with 161 additions and 62 deletions.
6 changes: 3 additions & 3 deletions core/create_dataset.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -178,7 +178,7 @@ def __init__(self, root= "GrandDataset",
]
if self.add_degree:
self.all_transforms.append(tg.transforms.OneHotDegree(max_degree=max_degree))
self.lst_pourcent = [0, 2, 5, 10, 15, 20, 30, 40]
self.lst_pourcent = [0, 2, 5, 10, 15, 20, 22, 25, 30, 35, 40]

super().__init__("./data/" + root)
self.root = "./data/" + root
Expand Down Expand Up @@ -396,5 +396,5 @@ def process(self):


if __name__ == '__main__':
dataset = GrandDataset("GrandDatasetAll", is_core_contained=False, max_degree=20, distance=1500)
print(dataset.num_classes)
dataset = GrandDataset("GrandDatasetCor", is_core_contained=False, max_degree=20, distance=1500)
print(dataset.train_datasets[(25, 25)])
141 changes: 104 additions & 37 deletions core/main.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -6,6 +6,7 @@

import torch_geometric as tg
import torch
from tqdm import tqdm
from torch.utils.data import DataLoader
import torch.nn.functional as F

Expand Down Expand Up @@ -117,7 +118,7 @@ def parser_to_config():

return vars(parser.parse_args())

def create_loader(config_cfg:Dict[str, Union[str, int, float]]):
def create_loader(config_cfg:Dict[str, Union[str, int, float]], train_dataset, test_dataset):

"""Create data loaders"""

Expand All @@ -134,6 +135,21 @@ def create_loader(config_cfg:Dict[str, Union[str, int, float]]):
)
return train_loader, test_loader

def load_dataset(config_cfg:Dict[str, Union[str, int ,float]]):
"""Loads dataset using config"""
if config_cfg["dataset"] == "Signal":
print("Warning depricated")
dataset = GrandDatasetSignal().shuffle()
train_dataset = dataset.train_dataset
test_dataset = dataset.test_dataset
elif config_cfg['dataset'] == "Classic":
dataset = GrandDataset(root=config_cfg["root"])
train_dataset = dataset.train_datasets[(config_cfg["infill_ratio_train"], config_cfg["coarse_ratio_train"])]
test_dataset = dataset.test_datasets[(config_cfg["infill_ratio_test"], config_cfg["coarse_ratio_test"])]
else:
raise ValueError("This dataset don't exist")

return dataset, train_dataset, test_dataset
def apply_transforms(config_cfg:Dict[str, Union[str, int, float]],
dataset: tg.data.InMemoryDataset):
"""Apply torch geometric transformations to add the edges and additional informations inside the dataset"""
Expand Down Expand Up @@ -196,13 +212,13 @@ def find_loss(loss_name:str):
raise ValueError("loss function not recognized")
return loss_fn

def load_models(model_dir:str, num_features:int, config_cfg:Dict, device='cpu'):
def load_models(model_dir:str, num_features:int, num_classes:int, config_cfg:Dict, device='cpu'):
"""Load the models"""
models = [0 for i in range(len(os.listdir(model_dir)))]
print(os.listdir(model_dir))
for model_path in os.listdir(model_dir):
checkpoint = torch.load(model_dir + "/" + model_path, map_location=device)
model = model_class(num_features, config_cfg["embed_size"], 1, config=config_cfg).to(device)
model = model_class(num_features, config_cfg["embed_size"], num_classes, config=config_cfg).to(device)
model.load_state_dict(checkpoint["model_state_dict"])
model.eval()

Expand Down Expand Up @@ -339,19 +355,21 @@ def train_model(model_id:int,
loss_fn:Callable,
config_cfg:Dict,
model_dir:str,
train_loader,
test_loader,
num_features:int,
num_classes:int,
device:torch.device=torch.device("cpu")):

"""Train the model"""
print(f"Model: {model_id}")
lst_train_perf = []
lst_test_perf = []
#Create loaders
train_loader, test_loader = create_loader(config_cfg)
# Create the model with given dimensions

###TODO:Change the num classes property of the class
model, cnn_embed, optimizer, lr_scheduler = create_model(config_cfg,
train_dataset.num_features,
num_classes=dataset.num_classes,
num_features,
num_classes=num_classes,
device=device)
model.train()
for epoch in range(config_cfg["epochs"]):
Expand Down Expand Up @@ -439,33 +457,26 @@ def train_model(model_id:int,
if not os.path.exists(model_dir_path):
os.mkdir(model_dir_path)

if config["fig_dir_name"] is None:
fig_name = config["model_name"]
else:
fig_name = config["fig_dir_name"]

fig_dir_path = "./Figures/" + fig_name
if not os.path.exists(fig_dir_path):
os.mkdir(fig_dir_path)
loss_fn = find_loss(config["loss_fn"])

if config["dataset"] == "Signal":
print("Warning depricated")
dataset = GrandDatasetSignal().shuffle()
elif config['dataset'] == "Classic":
dataset = GrandDataset(root=config["root"])
train_dataset = dataset.train_datasets[(config["infill_ratio_train"], config["coarse_ratio_train"])]
test_dataset = dataset.test_datasets[(config["infill_ratio_test"], config["coarse_ratio_test"])]
def train():
"""Training function"""
#Create dirs
if config["fig_dir_name"] is None:
fig_name = config["model_name"]
else:
fig_name = config["fig_dir_name"]

#train_dataset = apply_transforms(config, train_dataset)
#test_dataset = apply_transforms(config, test_dataset)
fig_dir_path = "./Figures/" + fig_name
if not os.path.exists(fig_dir_path):
os.mkdir(fig_dir_path)

else:
raise ValueError("This dataset don't exist")
#Load the dataset and loaders
dataset, train_dataset, test_dataset = load_dataset(config)

loss_fn = find_loss(config["loss_fn"])
train_loader, test_loader = create_loader(config, train_dataset, test_dataset)

def train():
"""Training function"""
device = torch.device(config["device"])
print("device: ", device)

Expand All @@ -478,6 +489,10 @@ def train():
lst_train_perf, lst_test_perf = train_model(model_id,
loss_fn=loss_fn,
config_cfg=config,
train_loader=train_loader,
test_loader=test_loader,
num_features=train_dataset.num_features,
num_classes=dataset.num_classes,
model_dir=model_dir_path,
device=device
)
Expand All @@ -486,16 +501,21 @@ def train():
plot_training_results(lst_train_perf, lst_test_perf, config, model_id,
fig_dir_path + "/" + str(model_id))

def test_process(distrib_infill:int, distrib_coarse:int, models, dataset:GrandDataset, config_cfg, device):
"""Test the models on a given dataset"""
train_dataset = dataset.train_datasets[(distrib_infill, distrib_coarse)]
test_dataset = dataset.test_datasets[(distrib_infill, distrib_coarse)]

train_loader, test_loader = create_loader(config_cfg, train_dataset, test_dataset)

def test():
"""Test function"""
# We don't want to shuffle to keep the same order in the data
print(f"config: {config}")
device = 'cpu'
train_loader, test_loader = create_loader(config)
if config_cfg["fig_dir_name"] is None:
fig_name = config_cfg["model_name"]
else:
fig_name = config_cfg["fig_dir_name"]

models = load_models(model_dir_path, train_dataset.num_features, config, device=device)
fig_dir_path = f"./Figures/{fig_name}_{distrib_infill}_{distrib_coarse}"
if not os.path.exists(fig_dir_path):
os.mkdir(fig_dir_path)

_, train_pred, train_energy = compute_preds_dataset(models, train_loader,
loss_fn=loss_fn, device=device)
Expand All @@ -504,14 +524,61 @@ def test():
# The same thing for test data #

_, test_pred, test_energy = compute_preds_dataset(models, test_loader,
loss_fn=loss_fn, device=device)
loss_fn=loss_fn, device=device)
plot_individual_preformance(test_pred, test_energy, mode="test", fig_dir=fig_dir_path)

# ## We computes the bins for all the models ## #
train_values = compute_bins(train_pred, train_energy)
test_values = compute_bins(test_pred, test_energy)

plot_bins_results(train_values, test_values, fig_dir=fig_dir_path)
#return (train_pred, train_energy), (test_pred, test_energy), (train_values, test_values)

def test():
"""Test function"""
# We don't want to shuffle to keep the same order in the data
print(f"config: {config}")
device = 'cpu'

#Load the dataset and loaders to intialize the models
dataset, train_dataset, test_dataset = load_dataset(config)

models = load_models(model_dir_path,
dataset.train_datasets[(0,0)].num_features,
num_classes=dataset.num_classes,
config_cfg=config,
device=device)

for (distrib_infill, distrib_coarse) in tqdm(dataset.train_datasets.keys()):
train_dataset = dataset.train_datasets[(distrib_infill, distrib_coarse)]
test_dataset = dataset.test_datasets[(distrib_infill, distrib_coarse)]

train_loader, test_loader = create_loader(config, train_dataset, test_dataset)

if config["fig_dir_name"] is None:
fig_name = config["model_name"]
else:
fig_name = config["fig_dir_name"]

fig_dir_path = f"./Figures/{fig_name}_{distrib_infill}_{distrib_coarse}"
if not os.path.exists(fig_dir_path):
os.mkdir(fig_dir_path)

_, train_pred, train_energy = compute_preds_dataset(models, train_loader,
loss_fn=loss_fn, device=device)
plot_individual_preformance(train_pred, train_energy, mode="train", fig_dir=fig_dir_path)

# The same thing for test data #

_, test_pred, test_energy = compute_preds_dataset(models, test_loader,
loss_fn=loss_fn, device=device)
plot_individual_preformance(test_pred, test_energy, mode="test", fig_dir=fig_dir_path)

# ## We computes the bins for all the models ## #
train_values = compute_bins(train_pred, train_energy)
test_values = compute_bins(test_pred, test_energy)

plot_bins_results(train_values, test_values, fig_dir=fig_dir_path)


if config["test"]:
Expand Down
56 changes: 42 additions & 14 deletions core/plot.ipynb
View file

Large diffs are not rendered by default.

20 changes: 12 additions & 8 deletions script_test.sh
View file
Original file line number Diff line number Diff line change
Expand Up @@ -4,16 +4,20 @@ source ../grand_env/bin/activate
# The arguments are saved under $1, $2, $3 depending on the order in the command

echo "algo: $1"
echo "num_layers: $2"
echo "embed_size: $3"
echo "num_layers_gnn: $2"
echo "num_layers_dense: $3"
echo "embed_size: $4"

echo "Model name: $4"
echo "Cuda and output: $5"
echo "dataset root: $6"
echo "Model name: $5"
echo "Cuda and output: $6"
echo "dataset root: $7"
echo "Topk ratio: $8"

for distrib in 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0
for distrib_infill in 0 5 10 15 20 25 30 35 40
do
nohup python -u core/main.py --batch_size=50 --seed=0 --loss_fn=scaled_l1 --root=$6 --embed_size=$3 --lr=1e-3 --num_layers=$2 --algo=$1 --model_name=$4 --device=cuda:$5 --ant_ratio_test=$distrib --fig_dir_name=$4$distrib --test > output$5.txt &
sleep 10
for distrib_coarse in 0 5 10 15 20 25 30 35 40
do
nohup python -u core/main.py --algo=$1 --infill_ratio_test=$distrib_infill --infill_ratio_train=$distrib_infill --coarse_ratio_test=$distrib_coarse --coarse_ratio_train=$distrib_coarse --batch_size=50 --seed=0 --loss_fn=scaled_l1 --root=$7 --embed_size=$4 --lr=1e-3 --num_layers_gnn=$2 --num_layers_dense=$3 --topkratio=$8 --model_name=$5 --device=cuda:$6 --fig_dir_name=${5}_${distrib_infill}_${distrib_coarse} --not_drop_nodes --test > output$6.txt &
done
done

0 comments on commit 9ed9b65

Please sign in to comment.