criterion = torch.nn.CrossEntropyLoss()
alpha = 1.0
epochs = 100
for epoch in range(epochs):
total_loss = 0
for x, y in trainloader:
optimizer.zero_grad()
output, embeddings = model(x.to(device))
loss, _ = margin_hard_loss(embeddings, y.to(device), alpha)
loss.backward()
optimizer.step()
total_loss += loss.item()
print('Epoch {} \t Loss: {:.3f}'.format(epoch, total_loss))
train_X, train_Y = get_embeddings(model, trainloader, device)
test_X, test_Y = get_embeddings(model, testloader, device)
reduced_umap = umap.UMAP().fit_transform(test_X.numpy())
reduced_tsne = TSNE(n_components=2).fit_transform(test_X.numpy())
f, (ax1, ax2) = plt.subplots(1, 2, figsize=(20, 10))
ax1.scatter(reduced_umap[:, 0], reduced_umap[:, 1], c=test_Y)
ax2.scatter(reduced_tsne[:, 0], reduced_tsne[:, 1], c=test_Y)
plt.tight_layout()
plt.savefig(results_path+'scatter.pdf', format='pdf')
plt.close()
pred_class = euclidean_knn(test_X, train_X, train_Y)
mAP = average_precision(test_Y.reshape(-1, 1), pred_class)
len(trainset) - int(len(trainset) * 0.9)))
trainloader = DataLoader(trainset,
batch_size=256,
shuffle=True,
num_workers=0)
validationloader = DataLoader(validationset,
batch_size=256,
shuffle=True,
num_workers=0)
model = give_me_resnet(pretrained=True)
model = model.to(device)
optimizer = torch.optim.Adam(model.parameters())
for alpha in alpha_list:
for epoch in range(epochs):
total_loss = 0
for x, y in trainloader:
optimizer.zero_grad()
output, embeddings = model(x.to(device))
loss = margin_loss(embeddings, y.to(device), alpha)
loss.backward()
optimizer.step()
total_loss += loss.item()
print('Epoch {} \t Loss: {}'.format(epoch, total_loss))
train_X, train_Y = get_embeddings(model, trainloader, device)
valid_X, valid_Y = get_embeddings(model, validationloader, device)
pred_class = euclidean_knn(valid_X, train_X, train_Y)
mAP = average_precision(valid_Y.reshape(-1, 1), pred_class)
results[str(alpha)].append(mAP.item())