def main():
# parser
parser = argparse.ArgumentParser()
parser.add_argument("-m", "--model_config", required=False,
default="deep_dnn.json",
help="Model Architecture .json")
parser.add_argument('-s', "--setup", default="agressive_setup.json",
help="Experimental Setup .json")
args = parser.parse_args()
# config paths
model_config = 'configs/model_architecture/' + args.model_config
setup_path = "configs/experimental_setup/" + args.setup
print(model_config)
# Hyper Parameters
setup = load_config(setup_path)
train_setup = setup["Train"]
prune_setup = setup["Prune"]
batch_size = train_setup["batch_size"]
epochs = train_setup["training_epochs"]
lr = train_setup["learning_rate"]
datatype = train_setup["datatype"]
feat_size = train_setup["feature_size"]
n_samples = train_setup["n_samples"]
n_classes = train_setup["n_classes"]
val_ratio = train_setup["val_ratio"]
test_ratio = train_setup["test_ratio"]
labels = 1
if datatype == "multilabel":
labels = train_setup["labels_per_sample"]
# CUDA for PyTorch
use_cuda = torch.cuda.is_available()
device = torch.device("cuda" if use_cuda else "cpu")
# Dataloaders
train_loader, val_loader, test_loader = \
CreateRandomDataset(datatype,
feat_size,
n_samples,
n_classes,
val_ratio,
test_ratio,
batch_size,
labels).get_dataloaders()
data_loaders = {"train": train_loader,
"val": val_loader,
"test": test_loader}
# Init model
model = DNN(config=model_config,
in_features=feat_size,
n_classes=n_classes)
if torch.cuda.is_available():
print('CUDA enabled.')
model.cuda()
print("--- DNN network initialized ---")
print(model)
# Criterion/Optimizer/Pruner
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=lr)
weight_pruner = WeightPruner(model)
# Train the model from scratch
print("--- Training DNN ---")
train_losses, val_losses = \
train_eval(model, criterion, optimizer, epochs, train_loader, val_loader)
test_acc, test_loss = test(model, test_loader, criterion)
learn_curves(train_losses, val_losses, "loss_fig.png")
iterative_pruning(model, weight_pruner, criterion, data_loaders, prune_setup)
def main():
parser = argparse.ArgumentParser(
description='Emotion recognition on CMU-MOSEI')
parser.add_argument('--seed', default=5566, type=int, help="Random seed")
parser.add_argument('--save_dir', default='./results', type=str, help="")
parser.add_argument('--num_epochs',
default=20,
type=int,
help="Number of training epochs")
parser.add_argument('--dropout', default=0.5, type=float, help="")
parser.add_argument('--min_ir',
default=2,
type=float,
help="Minimum imbalance ratio")
parser.add_argument('--lr', default=0.5, type=float, help="")
parser.add_argument('--activation', default='relu', type=str, help="")
parser.add_argument('--batch_size', default=32, type=int, help="")
parser.add_argument('--layers',
default='512.512.256.256.128.128',
type=str,
help="Comma-separted list of hidden dimensions")
parser.add_argument('--gamma',
default=1,
type=float,
help="Weight for negative class")
parser.add_argument('--dataset', default=None, type=str, help="Dataset")
parser.add_argument('--verbose',
default=False,
action='store_true',
help="Verbose")
args = parser.parse_args()
options['num_epochs'] = args.num_epochs
options['dropout'] = args.dropout
options['activation'] = args.activation
options['batch_size'] = args.batch_size
options['layers'] = [int(x) for x in args.layers.split('.')]
options['gamma'] = args.gamma
options['lr'] = args.lr
options['min_ir'] = args.min_ir
grad_clip_value = 10.0
CUDA = True
import torch
torch.manual_seed(args.seed)
import torch.nn as nn
import torch.nn.functional as F
from torch.optim import Adam, Adagrad
from torch.utils.data import Dataset, DataLoader
from losses import WeightedBCELoss
from models import DNN, mosei_dataset
n_epochs = options['num_epochs']
batch_size = options['batch_size']
verbose = args.verbose
save_dir = args.save_dir
ckpt_path = os.path.join(save_dir, "checkpoint.pt")
if not os.path.exists(save_dir):
os.makedirs(save_dir)
with open(os.path.join(save_dir, 'options.json'), 'w') as fout:
fout.write(json.dumps(options))
class_weight = torch.Tensor(options['class_weights'])
class_weight = class_weight / torch.sum(class_weight)
gamma = torch.Tensor([options['gamma']])
if CUDA:
class_weight = class_weight.cuda()
gamma = gamma.cuda()
model = DNN(options)
if CUDA:
model.cuda()
if verbose: print(model)
model.train()
dataset = pickle.load(open(args.dataset, "rb"))
train_dataset = mosei_dataset(dataset,
splits=train_split,
oversample=True,
min_ir=options['min_ir'])
train_loader = DataLoader(train_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=4)
if verbose: print(f"Train set: {len(train_dataset)} samples")
val_dataset = mosei_dataset(dataset, splits=dev_split)
val_loader = DataLoader(val_dataset,
batch_size=256,
shuffle=False,
num_workers=4)
if verbose: print(f"Val set: {len(val_dataset)} samples")
test_dataset = mosei_dataset(dataset, splits=test_split)
test_loader = DataLoader(test_dataset,
batch_size=256,
shuffle=False,
num_workers=4)
if verbose: print(f"Test set: {len(test_dataset)} samples")
optimizer = Adam(
[param for param in model.parameters() if param.requires_grad],
lr=options['lr'])
criterion = WeightedBCELoss(class_weight=class_weight,
PosWeightIsDynamic=True,
gamma=gamma)
best_val = np.Inf
best_metric = 0
train_labels = []
for epoch_no in range(n_epochs):
total_pos = 0
model.train()
for batch_no, batch in enumerate(train_loader, start=1):
embeddings, labels = batch
if CUDA:
embeddings, labels = embeddings.cuda(), labels.cuda()
y_hat = model(embeddings)
loss = criterion(y_hat, labels)
loss.backward()
# torch.nn.utils.clip_grad_norm_([param for param in model.parameters() if param.requires_grad], grad_clip_value)
optimizer.step()
optimizer.zero_grad()
if batch_no % 200 == 0:
if verbose: print(f"Training loss: {loss.item():.5f}")
y_true, y_pred, val_loss = [], [], []
model.eval()
for batch in val_loader:
embeddings, labels = batch
if CUDA:
embeddings, labels = embeddings.cuda(), labels.cuda()
y_hat = model(embeddings)
loss = criterion(y_hat, labels)
val_loss.append(loss.item())
y_true.append(labels.detach().cpu().numpy())
y_pred.append(y_hat.detach().cpu().numpy())
assert not np.any(np.isnan(val_loss))
y_true = np.concatenate(y_true, axis=0).squeeze()
y_pred = np.concatenate(y_pred, axis=0).squeeze()
y_true_bin = y_true > 0
y_pred_bin = y_pred > 0
val_loss = np.average(val_loss)
f1score = [
f1_score(t, p, average="weighted")
for t, p in zip(y_true_bin.T, y_pred_bin.T)
]
wa = np.average(weighted_accuracy(y_true_bin, y_pred_bin))
val_metric = np.average(f1score) + np.average(wa)
f1score = [f'{x*100:.2f}' for x in f1score]
if verbose: print(f"Validation loss: {val_loss:.3f}")
if best_metric < val_metric:
if verbose: print("Validation metric improved")
best_metric = val_metric
checkpoint = {
'options': options,
'model': model,
'epoch': epoch_no
}
torch.save(checkpoint, ckpt_path)
model.train()
# ====================================================================================================
# Final Validation
# ====================================================================================================
checkpoint = torch.load(ckpt_path)
model = checkpoint['model']
if verbose:
print("Loaded best model from epoch {}".format(checkpoint['epoch']))
model.eval()
val_true = []
val_pred = []
for batch in val_loader:
embeddings, labels = batch
if CUDA:
embeddings, labels = embeddings.cuda(), labels.cuda()
val_hat = model(embeddings)
val_true.append(labels.detach().cpu().numpy())
val_pred.append(val_hat.detach().cpu().numpy())
val_true = np.concatenate(val_true, axis=0).squeeze()
val_pred = np.concatenate(val_pred, axis=0).squeeze()
val_true_bin = val_true > 0
val_pred_bin = val_pred > 0
wa = [
weighted_accuracy(t, p) * 100
for t, p in zip(val_true_bin.T, val_pred_bin.T)
]
f1score = [
f1_score(t, p, average="weighted") * 100
for t, p in zip(val_true_bin.T, val_pred_bin.T)
]
if verbose:
print(f"Val WA, {reformat_array(wa)} Avg: {np.average(wa):.2f}")
if verbose:
print(
f"Val F1, {reformat_array(f1score)} Avg: {np.average(f1score):.2f} "
)
# ====================================================================================================
# Final Test
# ====================================================================================================
test_true = []
test_pred = []
for batch in test_loader:
embeddings, labels = batch
if CUDA:
embeddings, labels = embeddings.cuda(), labels.cuda()
pred = model(embeddings)
test_true.append(labels.detach().cpu().numpy())
test_pred.append(pred.detach().cpu().numpy())
test_true = np.concatenate(test_true, axis=0).squeeze()
test_pred = np.concatenate(test_pred, axis=0).squeeze()
test_true_bin = test_true > 0 # Binarized
test_pred_bin = test_pred > 0 # Logit outputs
test_wa = [
weighted_accuracy(t, p)
for t, p in zip(test_true_bin.T, test_pred_bin.T)
]
test_wa = reorder_labels(test_wa)
test_f1score = [
f1_score(t, p, average="weighted")
for t, p in zip(test_true_bin.T, test_pred_bin.T)
]
test_f1score = reorder_labels(test_f1score)
test_wa_str = [f'{x*100:.2f}' for x in test_wa]
test_f1score_str = [f'{x*100:.2f}' for x in test_f1score]
print(f"Test WA: {test_wa_str} Avg: {np.average(test_wa)*100:2.1f}")
print(
f"Test F1: {test_f1score_str} Avg: {np.average(test_f1score)*100:2.1f}"
)
combined = [f" {x} & {y} " for x, y in zip(test_wa_str, test_f1score_str)]
combined.append(
f" {np.average(test_wa)*100:2.1f} & {np.average(test_f1score)*100:2.1f}"
)
if verbose: print("&".join(combined))
