def predict(self, board: np.ndarray):
start = time()
board = torch.FloatTensor(board.astype(np.float64)) # TODO: Make sure this doesn't modify game.board in place
if args["cuda"]:
board = board.contiguous().cuda()
board = board.view(1, self.board_x, self.board_y) # Reshape? TODO: Figure out why this is 1 deep instead of two splitting up the players like in the paper
self.net.eval()
with torch.no_grad(): # No need to compute gradients
policy, value = self.net(board)
inference_time = time() - start
return torch.exp(policy).data.cpu().numpy()[0], value.data.cpu().numpy()[0]
def encode_rnn(self, feats: numpy.ndarray) -> torch.Tensor:
"""Encode acoustic features.
Args:
feats: Feature sequence. (F, D_feats)
Returns:
enc_out: Encoded feature sequence. (T, D_enc)
"""
p = next(self.parameters())
feats_len = [feats.shape[0]]
feats = feats[::self.subsample[0], :]
feats = torch.as_tensor(feats, device=p.device, dtype=p.dtype)
feats = feats.contiguous().unsqueeze(0)
enc_out, _, _ = self.enc(feats, feats_len)
return enc_out.squeeze(0)
def predict(self, board: np.ndarray):
"""
board: np array with board
"""
# timing
#start = time.time()
# preparing input
board = torch.as_tensor(board, dtype=torch.float)
if args.cuda:
board = board.contiguous().cuda()
self.nnet.cuda()
board = board.view(1, 2 * self.history + 1, self.board_x, self.board_y)
self.nnet.eval()
with torch.no_grad():
p, v = self.nnet(board)
#print(p.shape, v.shape)
#print(v.data.cpu().numpy()[0])
#print('PREDICTION TIME TAKEN : {0:03f}'.format(time.time()-start))
return p.data.cpu().numpy()[0], v.data.cpu().numpy()[0]
