print('Create Model!')
som = som.to(device)
if train == True:
losses = list()
for epoch in range(total_epoch):
running_loss = 0
start_time = time.time()
for idx, (X, Y) in enumerate(train_loader):
X = X.view(-1, 28 * 28 * 1).to(device) # flatten
loss = som.self_organizing(X, epoch, total_epoch) # train som
running_loss += loss
losses.append(running_loss)
print('epoch = %d, loss = %.2f, time = %.2fs' %
(epoch + 1, running_loss, time.time() - start_time))
if epoch % 5 == 0:
# save
som.save_result('%s/som_epoch_%d.png' % (RES_DIR, epoch),
(1, 28, 28))
torch.save(som.state_dict(), '%s/som.pth' % MODEL_DIR)
torch.save(som.state_dict(), '%s/som.pth' % MODEL_DIR)
plt.title('SOM loss')
plt.plot(losses)
plt.show()
som.save_result('%s/som_result.png' % (RES_DIR), (1, 28, 28))
torch.save(som.state_dict(), '%s/som.pth' % MODEL_DIR)
if train == True:
losses = list()
for epoch in tqdm(range(total_epoch)):
running_loss = 0
start_time = time.time()
for idx, b in enumerate(train_dataloader):
X=torch.FloatTensor(b['mol_fp'])
#print(f'X.size{X.size()}')
X = X.view(-1, input_dim).to(device) # flatten
loss = som.self_organizing(X, epoch, total_epoch) # train som
running_loss += loss
losses.append(running_loss)
print('epoch = %d, loss = %.2f, time = %.2fs' % (epoch + 1, running_loss, time.time() - start_time))
if epoch % 100 == 0:
# save TODO som.save_result for molecules save
torch.save(som.state_dict(), f'{args.save_dir}/som_F{f_i}Epoch{epoch}.pth')
torch.save(som.state_dict(), f'{args.save_dir}/som_F{f_i}.pth')
#plot loss trend for each fold training
plt.title('SOM loss')
plt.plot(losses)
#plt.show()
plt.savefig(f'{args.save_dir}/train_loss{f_i}.png')
plt.clf()#https://stackoverflow.com/questions/21884271/warning-about-too-many-open-figures
plt.close()
save_f_val = args.save_dir + '/' + f'val_heatmapF{f_i}.png'
save_f_test = args.save_dir + '/' + f'test_heatmapF{f_i}.png'
# testing and pliot heatmap each fold
x_val=torch.FloatTensor(val)
x_val = x_val.view(-1, input_dim).to(device)
val_clusters = predict(som,x_val)
