def __main__():
local_folder = "./data/kaggle_dessins/"
train_images_fname = "train_images.npy"
train_labels_fname = "train_labels.csv"
home_path = "/home/simon/"
destination_folder = "annleukemia"
data_folder = "data"
results_folder = "results"
extra_class = True # TODO change to put the number... curious to see if more than one is desirable
meta_destination_folder = "pandas_meta_df"
plots_folder_path = "/".join([home_path, destination_folder, results_folder, "plots/"])
dataset_name = "dessins"
activation = "relu"
early_stopping = 200
n_epochs = 1000
gt_input = 0
use_conv = False # Not applicable if not sequence (images, videos, sentences, DNA...)
lr = 1e-5
l1 = 1e-5
l2 = 1e-10
dropout = 0.5
batch_size = 16
is_pruning = False
# mc = 1
# iw = 1
# Neurons layers
h_dims = [1024, 1024, 1024]
from utils.utils import adapt_datasets
train_arrays = np.load(local_folder + train_images_fname, encoding="latin1")
train_dataset = np.vstack(train_arrays[:, 1])
train_labels = genfromtxt(local_folder + train_labels_fname, delimiter=",", dtype=str, skip_header=True)[:, 1]
test_dataset = np.vstack(np.load(local_folder + "test_images.npy", encoding="latin1")[:, 1])
meta_df = pd.DataFrame(train_dataset, columns=train_labels)
img_shape = [1, 100, 100]
labels = set(list(meta_df.columns))
mlp = MLP(input_size=meta_df.shape[0], input_shape=(meta_df.shape[0]),
indices_names=list(range(meta_df.shape[0])), num_classes=len(labels),
h_dims=h_dims, extra_class=extra_class, l1=l1, l2=l2, batch_norm=True)
mlp.labels = labels
mlp.labels_set = list(set(labels))
mlp.set_configs(home_path=home_path, results_folder=results_folder, data_folder=data_folder,
destination_folder=destination_folder, dataset_name=dataset_name, lr=lr,
meta_destination_folder="meta_pandas_dataframes", csv_filename="csv_loggers", is_unlabelled=False)
print("Labeled data shape (35371, 624)", meta_df.shape)
if meta_df is not None:
mlp.import_dataframe(meta_df, batch_size, labelled=True)
train_total_loss_histories = [[] for x in range(10)]
train_accuracy_histories = [[] for x in range(10)]
valid_total_loss_histories = [[] for x in range(10)]
valid_accuracy_histories = [[] for x in range(10)]
for i in range(10):
print("Random train/valid split", i)
mlp.set_data(labels_per_class=-1, is_example=False, extra_class=extra_class)
mlp.glorot_init()
mlp.run(n_epochs, verbose=3, show_progress=10, hist_epoch=20, is_balanced_relu=False, all0=False)
def main():
from data_preparation.GeoParser import GeoParser
from models.discriminative.artificial_neural_networks.MultiLayerPerceptron import MLP
load_from_disk = True
load_merge = False
geo_ids = ["GSE33000"]
# files_destinations
home_path = "/home/simon/"
destination_folder = "annleukemia"
data_folder = "data"
results_folder = "results"
meta_destination_folder = "pandas_meta_df"
plots_folder_path = "/".join(
[home_path, destination_folder, results_folder, "plots/"])
dataset_name = "gse33000_no_huntington"
activation = "relu"
# nrep = 3
early_stopping = 200
n_epochs = 1000
gt_input = 0
extra_class = False
dataset_name = dataset_name + "extra_class" + str(extra_class)
# if ladder is yes builds a ladder vae. Do not combine with auxiliary (yet; might be possible and relatively
# not too hard to implement, but might be overkill. Might be interesting too)
translate = "n"
use_conv = False # Not applicable if not sequence (images, videos, sentences, DNA...)
lr = 1e-4
l1 = 0.
l2 = 0.
batch_size = 32
# mc = 1
# iw = 1
# Neurons layers
h_dims = [128, 128]
from utils.utils import adapt_datasets
g = GeoParser(home_path=home_path, geo_ids=geo_ids)
g.get_geo(load_from_disk=load_from_disk, automatic_attribute_list=None)
meta_df = g.merge_datasets(load_from_disk=load_merge, labelled=True)
if translate is "y":
for geo_id in geo_ids:
g.translate_indices_df(geo_id, labelled=True)
labels = set(list(meta_df.columns))
print(labels)
mlp = MLP(input_size=meta_df.shape[0],
input_shape=(meta_df.shape[0]),
indices_names=list(range(meta_df.shape[0])),
num_classes=len(labels),
h_dims=h_dims,
extra_class=extra_class,
l1=l1,
l2=l2,
batch_norm=True)
mlp.set_configs(home_path=home_path,
results_folder=results_folder,
data_folder=data_folder,
destination_folder=destination_folder,
dataset_name=dataset_name,
lr=lr,
meta_destination_folder="meta_pandas_dataframes",
csv_filename="csv_loggers",
is_unlabelled=False)
print("Labeled data shape (35371, 624)", meta_df.shape)
if meta_df is not None:
mlp.import_dataframe(meta_df, batch_size, labelled=True)
train_total_loss_histories = [[] for x in range(10)]
train_accuracy_histories = [[] for x in range(10)]
valid_total_loss_histories = [[] for x in range(10)]
valid_accuracy_histories = [[] for x in range(10)]
for i in range(100):
print("Random train/valid split", i)
mlp.set_data(labels_per_class=-1,
is_example=False,
extra_class=extra_class,
ignore_training_inputs=1)
mlp.glorot_init()
mlp.run(n_epochs,
verbose=0,
show_progress=10,
hist_epoch=20,
is_balanced_relu=True,
all0=True,
overall_mean=True)
def __main__():
import os
os.chdir("..") # To return at the root of the project
from models.discriminative.artificial_neural_networks.MultiLayerPerceptron import MLP
geo_ids = ["GSE33000"]
unlabelled_geo_ids = ["GSE33000"]
load_from_disk = True
load_merge = False
home_path = "/home/simon/"
destination_folder = "annleukemia"
data_folder = "data"
results_folder = "results"
translate = "f"
extra_class = True # TODO change to put the number... curious to see if more than one is desirable
meta_destination_folder = "pandas_meta_df"
plots_folder_path = "/".join(
[home_path, destination_folder, results_folder, "plots/"])
dataset_name = "gse33000"
activation = "relu"
early_stopping = 200
n_epochs = 1000
gt_input = 0
use_conv = False # Not applicable if not sequence (images, videos, sentences, DNA...)
lr = 1e-3
l1 = 0.
l2 = 0.
dropout = 0.5
batch_size = 32
is_pruning = False
# mc = 1
# iw = 1
# Neurons layers
h_dims = [128, 128]
from utils.utils import adapt_datasets
g = GeoParser(home_path=home_path, geo_ids=geo_ids)
g.get_geo(load_from_disk=load_from_disk, automatic_attribute_list=None)
meta_df = g.merge_datasets(load_from_disk=load_merge, labelled=True)
labels = set(list(meta_df.columns))
mlp = MLP(input_size=meta_df.shape[0],
input_shape=(meta_df.shape[0]),
indices_names=list(range(meta_df.shape[0])),
num_classes=len(labels),
h_dims=h_dims,
extra_class=extra_class,
l1=l1,
l2=l2,
batch_norm=True)
mlp.labels = labels
mlp.labels_set = list(set(labels))
mlp.set_configs(home_path=home_path,
results_folder=results_folder,
data_folder=data_folder,
destination_folder=destination_folder,
dataset_name=dataset_name,
lr=lr,
meta_destination_folder="meta_pandas_dataframes",
csv_filename="csv_loggers",
is_unlabelled=False)
print("Labeled data shape (35371, 624)", meta_df.shape)
if meta_df is not None:
mlp.import_dataframe(meta_df, batch_size, labelled=True)
train_total_loss_histories = [[] for x in range(10)]
train_accuracy_histories = [[] for x in range(10)]
valid_total_loss_histories = [[] for x in range(10)]
valid_accuracy_histories = [[] for x in range(10)]
for i in range(10):
print("Random train/valid split", i)
mlp.set_data(labels_per_class=-1,
is_example=False,
extra_class=extra_class)
mlp.glorot_init()
mlp.run(n_epochs,
verbose=2,
show_progress=10,
hist_epoch=20,
is_balanced_relu=False,
all0=False)
def __main__():
train_labels_fname = "train_labels.csv"
home_path = "/home/simon/"
destination_folder = "annleukemia"
data_folder = "data"
results_folder = "results"
extra_class = True # TODO change to put the number... curious to see if more than one is desirable
dataset_name = "dessins"
n_epochs = 1000
lr = 1e-3
l1 = 0.
l2 = 0.
batch_size = 256
resized_shape = 100
# Neurons layers
h_dims = [1024, 512, 256, 128]
input_shape = [1, 100, 100]
input_size = np.prod([1, resized_shape, resized_shape])
dir = "data/kaggle_dessins/"
train_labels = genfromtxt(dir + train_labels_fname,
delimiter=",",
dtype=str,
skip_header=True)[:, 1]
train_labels_set = set(train_labels)
data_transform = transforms.Compose([
transforms.RandomSizedCrop((resized_shape)),
transforms.RandomRotation(45),
transforms.RandomHorizontalFlip(),
transforms.RandomVerticalFlip(),
transforms.Grayscale(),
transforms.ToTensor(),
])
dessins_train_data = datasets.ImageFolder(
root='data/kaggle_dessins/train/', transform=data_transform)
dessins_valid_data = datasets.ImageFolder(
root='data/kaggle_dessins/valid/', transform=data_transform)
train_ds = torch.utils.data.DataLoader(dessins_train_data,
batch_size=batch_size,
shuffle=True,
num_workers=0)
valid_ds = torch.utils.data.DataLoader(dessins_valid_data,
batch_size=batch_size,
shuffle=True,
num_workers=0)
mlp = MLP(input_size=input_size,
input_shape=input_shape,
indices_names=list(range(len(train_labels))),
num_classes=len(train_labels_set),
h_dims=h_dims,
extra_class=extra_class,
l1=l1,
l2=l2,
batch_norm=True)
mlp.x_train = train_ds
mlp.x_valid = valid_ds
mlp.x_test = None
mlp.batch_size = batch_size
mlp.labels_set = os.listdir('data/kaggle_dessins/train/')
mlp.num_classes = len(mlp.labels_set)
mlp.labels_train = train_labels
mlp.labels = train_labels
mlp.labels_set = train_labels_set
mlp.set_configs(home_path=home_path,
results_folder=results_folder,
data_folder=data_folder,
destination_folder=destination_folder,
dataset_name=dataset_name,
lr=lr,
meta_destination_folder="meta_pandas_dataframes",
csv_filename="csv_loggers",
is_unlabelled=False)
mlp.make_loaders(train_ds=train_ds,
valid_ds=valid_ds,
test_ds=None,
labels_per_class=-1,
unlabelled_train_ds=None,
unlabelled_samples=True)
train_total_loss_histories = [[] for x in range(10)]
train_accuracy_histories = [[] for x in range(10)]
valid_total_loss_histories = [[] for x in range(10)]
valid_accuracy_histories = [[] for x in range(10)]
mlp.train_loader = mlp.train_loader.dataset
mlp.valid_loader = mlp.valid_loader.dataset
for i in range(10):
print("Random train/valid split", i)
mlp.set_data(labels_per_class=-1,
is_example=False,
is_split=False,
extra_class=extra_class,
is_custom_data=True)
mlp.glorot_init()
mlp.run(n_epochs,
verbose=3,
show_progress=10,
hist_epoch=20,
is_balanced_relu=False,
all0=False)