def evaluate(self):
game = self.trainer.game
game.eval()
old_loss = game.loss
for loader_name, loader, metric in self.loaders_metrics:
acc_or, acc = 0.0, 0.0
n_batches = 0
game.loss = metric
for batch in loader:
n_batches += 1
batch = core.move_to(batch, self.device)
with torch.no_grad():
_, rest = game(*batch)
acc += rest['acc']
acc_or += rest['acc_or']
self.results[loader_name] = {
'acc': acc / n_batches,
'acc_or': acc_or / n_batches
}
self.results['epoch'] = self.epoch
output_json = json.dumps(self.results)
print(output_json, flush=True)
game.loss = old_loss
game.train()
def on_epoch_end(self, loss, logs, epoch):
dump_dir = pathlib.Path.cwd() / "dump" / str(epoch)
dump_dir.mkdir(exist_ok=True, parents=True)
self.trainer.game.eval()
len_dataset = len(self.eval_dataset)
for i in range(5):
example_id = np.random.randint(0, len_dataset)
example = self.eval_dataset[example_id]
example = (
example[0].unsqueeze(0),
example[1].unsqueeze(0),
example[2].unsqueeze(0),
)
device = torch.device(
"cuda" if torch.cuda.is_available() else "cpu")
example = core.move_to(example, device)
_, interaction = self.trainer.game(*example)
image = example[0][0]
output = interaction.receiver_output.view(*self.image_shape)
image = image.view(*self.image_shape)
utils.save_image(torch.cat([image, output], dim=1),
dump_dir / (str(i) + ".png"))
self.trainer.game.train()
def on_epoch_end(self, loss, logs, epoch):
dump_dir = pathlib.Path.cwd() / "dump" / str(epoch)
dump_dir.mkdir(exist_ok=True, parents=True)
self.trainer.game.eval()
for i in range(5):
example = self.eval_dataset[i]
example = core.move_to(example, self.device)
_, interaction = self.trainer.game(*example)
image = example[0][0]
output = interaction.receiver_output.view(28, 28)
image = image.view(28, 28)
utils.save_image(
torch.cat([image, output], dim=1), dump_dir / (str(i) + ".png")
)
self.trainer.game.train()