with torch.no_grad():
scores = model(imgs)
emd_loss = criterion(scores, labels)
emd_losses.update(emd_loss.item(), imgs.shape[0])
scores_hist.append(scores.cpu().data.numpy())
labels_hist.append(labels.cpu().data.numpy())
print()
logging.info("Test emd_loss@r=1: {:.4f}".format(emd_losses.avg))
scores_hist = np.concatenate(scores_hist)
labels_hist = np.concatenate(labels_hist)
scores_mean = np.dot(scores_hist, np.arange(1, 11))
labels_mean = np.dot(labels_hist, np.arange(1, 11))
SRCC_mean, _ = spearmanr(scores_mean, labels_mean)
logging.info("Test SRCC_mean: {:.4f}".format(SRCC_mean))
return emd_losses.avg, SRCC_mean
if __name__ == "__main__":
for epoch in range(1, epochs + 1):
# Train Phase
train(model, train_loader, scheduler, optimizer)
# Valid Phase
emd_loss, SRCC_mean = eval(model, test_loader)
# Save model
fname = "model_{}{}.pth".format(__file__.split(".")[0], comment)
saver.save(SRCC_mean, model.state_dict(), fname)
# print(winner)
if winner == 'BD':
points_ac.append(points)
else:
points_ac.append(-points)
points = -points
for p in game.players:
if isinstance(p, MyPlayer):
loss += learn(p.probabilities, p.actions, points)
rollout_points.append(sum(points_ac))
loss /= rollouts
optimizer.zero_grad()
loss.backward()
optimizer.step()
if e % print_epochs == 0:
print(
f'Epoch {e}: Mean AC total {rollouts} rollouts points in last {print_epochs} episodes: {np.mean(rollout_points[-print_epochs:])}'
)
# plt.plot(points_ac)
plt.plot(rollout_points)
plt.plot(running_average(rollout_points))
# plt.plot(points_bd)
# plt.plot(running_average(points_bd))
plt.show()
saver.save()
