Beispiel #1
0
    def network_simulation(self, networkpath):
        # load model
        torch.set_default_dtype(torch.float64)
        state_dict = torch.load(networkpath)
        #hidden_layer_size = state_dict["lstm.weight_hh_l0"][1]
        model = LSTM_multi_modal() if output_model == "multi_modal" else LSTM_fixed()
        model.load_state_dict(state_dict)
        model.eval()

        # init hidden
        #hidden_state = [model.init_hidden(1, num_layers) for agent in range(self.num_guppys)]
        states = [[model.init_hidden(1, 1, hidden_layer_size)
                  for i in range(num_layers * 2)]
                  for j in range(self.num_guppys)]
        for i in range(1, len(self.agent_data) - 1):
            for agent in range(self.num_guppys):
                with torch.no_grad():
                    # get input data for this frame
                    sensory = self.craft_vector(i, agent)
                    data = torch.from_numpy(numpy.concatenate((self.loc_vec, sensory)))
                    data = data.view(1, 1, -1)

                    # predict the new ang_turn, lin_speed
                    #out, hidden_state[agent] = model.predict(data, hidden_state[agent])
                    out, states[agent] = model.predict(data, states[agent])
                    ang_turn = out[0].item() if output_model == "multi_modal" else out[0][0][0].item()
                    lin_speed = out[1].item() if output_model == "multi_modal" else out[0][0][1].item()

                    # rotate agent position by angle calculated by network
                    cos_a = cos(ang_turn)
                    sin_a = sin(ang_turn)
                    agent_pos = self.data[agent][i][0], self.data[agent][i][1]
                    agent_ori = self.data[agent][i][2], self.data[agent][i][3]
                    new_ori = [cos_a * agent_ori[0] - sin_a * agent_ori[1], \
                               sin_a * agent_ori[0] + cos_a * agent_ori[1]]
                    # normally the rotation of a normalized vector by a normalized vector should again be a
                    # normalized vector, but it seems there are some numerical errors, so normalize the orientation
                    # again
                    normalize_ori(new_ori)

                    # multiply new orientation by linear speed and add to old position
                    translation_vec = scalar_mul(lin_speed, new_ori)
                    new_pos = vec_add(agent_pos, translation_vec)
                    # network does not learn the tank walls properly sometimes, let fish bump against the wall
                    normalize_pos(new_pos)

                    # update the position for the next timestep
                    self.data[agent][i + 1][0], self.data[agent][i + 1][1] = new_pos
                    self.data[agent][i + 1][2], self.data[agent][i + 1][3] = new_ori

            self.plot_guppy_bins(bins=False)
Beispiel #2
0
        print("###################################")
        print(f'epoch: {i:3} loss: {loss.item():10.10f}')
        print("###################################")
        # torch.nn.utils.clip_grad_norm_(model.parameters(), 0.25)
    #loss.backward()
    # optimizer.step()

    except KeyboardInterrupt:
        if input("Do you want to save the model trained so far? y/n") == "y":
            torch.save(model.state_dict(), network_path + f".epochs{i}")
            print("network saved at " + network_path + f".epochs{i}")
        sys.exit(0)

        ########validation#######

        model.eval()
        for inputs, targets in valloader:

            if output_model == "multi_modal":
                targets = targets.type(torch.LongTensor)

                for s in range(0, inputs.size()[1] - seq_len, seq_len):
                    states = [tuple([each.data for each in s]) for s in states] if arch == "ey" else \
                        tuple([each.data for each in states])
                    angle_pred, speed_pred, _ = model.forward(
                        inputs[:, s:s + seq_len, :], states)

                    # angle_pred, speed_pred, states = model.forward(inputs[:, s:s + seq_len, :], states)
                    angle_pred = angle_pred.view(
                        angle_pred.shape[0] * angle_pred.shape[1], -1)
                    speed_pred = speed_pred.view(