def main(args):
### Part 1 ###
train_set, train_labels, dev_set, dev_labels = reader.load_dataset(
args.dataset_file, part=1)
train_set = torch.tensor(train_set, dtype=torch.float32)
train_labels = torch.tensor(train_labels, dtype=torch.int64)
dev_set = torch.tensor(dev_set, dtype=torch.float32)
losses, predicted_labels, net = p1.fit(train_set, train_labels, dev_set,
args.max_iter)
accuracy, f1, precision, recall = compute_accuracies(predicted_labels,
dev_set,
dev_labels,
N_class=3)
print(' ##### Part 1 results #####')
print("Accuracy:", accuracy)
print("F1-Score:", f1)
print("Precision:", precision)
print("Recall:", recall)
print("num_parameters:",
sum([np.prod(w.shape) for w in net.get_parameters()]))
torch.save(net, "net_p1.model")
### Part 2 ###
train_set, train_labels, dev_set, dev_labels = reader.load_dataset(
args.dataset_file, part=2)
train_set = torch.tensor(train_set, dtype=torch.float32)
train_labels = torch.tensor(train_labels, dtype=torch.int64)
dev_set = torch.tensor(dev_set, dtype=torch.float32)
_, predicted_labels, net = p2.fit(train_set, train_labels, dev_set,
10 * args.max_iter)
accuracy, f1, precision, recall = compute_accuracies(predicted_labels,
dev_set,
dev_labels,
N_class=5)
print(' ##### Part 2 results #####')
print("Accuracy:", accuracy)
print("F1-Score:", f1)
print("Precision:", precision)
print("Recall:", recall)
print("num_parameters:",
sum([np.prod(w.shape) for w in net.get_parameters()]))
torch.save(net, "net_p2.model")
### Part 3 ###
# input provided will be normalized to 0.0 - 1.0
train_set, _, dev_set, _ = reader.load_dataset(
args.dataset_file, part=2) # use same data as part 2
train_set = torch.tensor(train_set, dtype=torch.float32) / 255.0
dev_set = torch.tensor(dev_set, dtype=torch.float32) / 255.0
_, x_recon, net = p3.fit(train_set, dev_set, args.max_iter * 10)
diff = (x_recon - dev_set.numpy())**2
MSE = diff.mean()
print(' ##### Part 3 results #####')
print("MSE:", MSE)
print("num_parameters:",
sum([np.prod(w.shape) for w in net.get_parameters()]))
torch.save(net, "net_p3.model")
def main(args):
# ### Part 1 ###
# train_set, train_labels, dev_set,dev_labels = reader.load_dataset(args.dataset_file, part=1)
# train_set = torch.tensor(train_set,dtype=torch.float32)
# train_labels = torch.tensor(train_labels,dtype=torch.int64)
# dev_set = torch.tensor(dev_set,dtype=torch.float32)
# losses,predicted_labels,net = p1.fit(train_set,train_labels, dev_set,args.max_iter)
# accuracy,f1,precision,recall = compute_accuracies(predicted_labels,dev_set,dev_labels, N_class=3)
# print(' ##### Part 1 results #####')
# print("Accuracy:",accuracy)
# print("F1-Score:",f1)
# print("Precision:",precision)
# print("Recall:",recall)
# print("num_parameters:", sum([ np.prod(w.shape) for w in net.get_parameters()]))
# torch.save(net, "net_p1.model")
# ### Part 2 ###
# train_set, train_labels, dev_set,dev_labels = reader.load_dataset(args.dataset_file, part=2)
# train_set = torch.tensor(train_set,dtype=torch.float32)
# train_labels = torch.tensor(train_labels,dtype=torch.int64)
# dev_set = torch.tensor(dev_set,dtype=torch.float32)
# _,predicted_labels,net = p2.fit(train_set,train_labels, dev_set,10*args.max_iter)
# accuracy,f1,precision,recall = compute_accuracies(predicted_labels,dev_set,dev_labels, N_class=5)
# print(' ##### Part 2 results #####')
# print("Accuracy:",accuracy)
# print("F1-Score:",f1)
# print("Precision:",precision)
# print("Recall:",recall)
# print("num_parameters:", sum([ np.prod(w.shape) for w in net.get_parameters()]))
# torch.save(net, "net_p2.model")
### Part 3 ###
# input provided will be normalized to 0.0 - 1.0
train_set, _, dev_set, _ = reader.load_dataset(
args.dataset_file, part=2) # use same data as part 2
train_set = torch.tensor(train_set, dtype=torch.float32) / 255.0
dev_set = torch.tensor(dev_set, dtype=torch.float32) / 255.0
_, x_recon, net = p3.fit(train_set, dev_set, args.max_iter * 10)
diff = (x_recon - dev_set.numpy())**2
MSE = diff.mean()
print(' ##### Part 3 results #####')
print("MSE:", MSE)
print("num_parameters:",
sum([np.prod(w.shape) for w in net.get_parameters()]))
torch.save(net, "net_p3.model")
# Show some original images in 1st row & reconstructed images in 2nd row
# For debug only
import matplotlib.pyplot as plt
idx = np.random.choice(x_recon.shape[0], 10)
fig, axes = plt.subplots(2, 10)
for i in range(10):
axes[0, i].imshow(dev_set[idx[i]].reshape(28, 28), cmap='gray')
axes[1, i].imshow(x_recon[idx[i]].reshape(28, 28), cmap='gray')
plt.show()