def restart(event):
global games_played
global done
global currently_passing
# Increment game counter
games_played += 1
done = False
# Generate new initial coordinates and reset agent
env.reset()
# Clear figure
plt.clf()
# Plot initial point
draw.pitch()
plt.scatter(env.x * 100, env.y * 100, color='C1')
# Additional information and stats to the player
plt.text(0, 105, rules_text, color='black')
reward_text = 'Games Played: ' + str(
games_played) + '\nTotal Rewards: ' + str(round(
rewards_cumulative, 3)) + '\nRewards pg: ' + str(
round(rewards_cumulative / games_played, 3))
plt.text(100, 105, reward_text, color='black', horizontalalignment='right')
# Save frame for update
plt.savefig(img_dir)
# Update frame
update_screen()
def update_screen():
global team_1_players
global team_2_players
plt.figure()
draw.pitch()
for coordinates in team_1_players:
plt.scatter(coordinates[0],
coordinates[1],
s=100,
color='navy',
edgecolors='navy',
alpha=0.7,
linewidths=1,
zorder=100)
for coordinates in team_2_players:
plt.scatter(coordinates[0],
coordinates[1],
s=100,
color='firebrick',
edgecolors='firebrick',
alpha=0.7,
linewidth=1,
zorder=100)
plt.savefig('_tmp/b.png')
plt.clf()
img = ImageTk.PhotoImage(Image.open('_tmp/b.png'))
panel.configure(image=img)
panel.image = img
def draw_play(agent, dpi=144, directory='__default', plot_now=False):
draw.pitch(dpi=dpi)
for action in range(len(agent.actions)):
draw_actions(agent, action)
if plot_now:
return
plt.savefig('img/' + directory + '.png')
plt.clf()
def draw_play_from_agent_set(agent_set):
draw.pitch()
max_len = 0
for agent in agent_set:
if len(agent.actions) > max_len:
max_len = len(agent.actions)
for action in range(max_len):
draw_actions_for_set(agent_set, action)
plt.show()
def draw_play_from_agent_set(agent_set,
dpi=144,
plot_now=True,
directory='a.png'):
draw.pitch(dpi=dpi)
max_len = 0
for agent in agent_set:
if len(agent.actions) > max_len:
max_len = len(agent.actions)
for action in range(max_len):
draw_actions_for_set(agent_set, action)
if plot_now:
return
plt.savefig(directory, format='png')
plt.clf()
def draw_play(agent,
dpi=144,
directory='__default',
plot_now=False,
current_action_color='C0'):
draw.pitch(dpi=dpi)
for action in range(len(agent.actions)):
draw_actions(agent, action)
if not agent.end:
plt.scatter(agent.x * 100,
agent.y * 100,
color=current_action_color,
zorder=100)
if plot_now:
return
plt.savefig('img/' + directory + '.png')
plt.clf()
def plot_team(self, dpi=120):
draw.pitch(dpi=dpi)
attacking_team = [(player.x, player.y, 0) for player in self.players]
for i, player in enumerate(attacking_team):
plt.scatter(player[0],
player[1],
s=75,
color='blue',
edgecolors='navy',
linewidths=1,
zorder=10)
plt.text(player[0],
player[1],
i,
zorder=11,
fontsize='xx-small',
color='white',
ha='center',
va='center')
plt.scatter(attacking_team[0][0] + 1,
attacking_team[0][1],
s=48,
color='black',
zorder=11)
plt.scatter(attacking_team[0][0] + 1,
attacking_team[0][1],
s=48,
marker='*',
color='white',
zorder=11)
plt.scatter(attacking_team[0][0] + 1,
attacking_team[0][1],
s=32,
marker='+',
color='black',
zorder=11)
def plot_result(self, proposed_team, save_fig_dir='', dpi=120):
attacking_team = [(player.x, player.y, 0) for player in self.players]
defensive_team = [(player.x, player.y, 1)
for player in proposed_team.players]
draw.pitch(dpi=dpi)
# Draw initial state
for i, player in enumerate(attacking_team):
plt.scatter(player[0],
player[1],
s=75,
color='blue',
edgecolors='navy',
linewidths=1,
zorder=10)
plt.text(player[0],
player[1],
i,
zorder=11,
fontsize='xx-small',
color='white',
ha='center',
va='center')
for i, player in enumerate(defensive_team):
plt.scatter(player[0],
player[1],
s=75,
color='orange',
edgecolors='C1',
linewidths=1,
zorder=10)
plt.text(player[0],
player[1],
i,
zorder=11,
fontsize='xx-small',
color='white',
ha='center',
va='center')
# Draw all passing lines
for passline in self.passlines:
passline.plot(color='tomato')
# Draw all goal-defense lines
for goalline in self.goallines:
goalline.plot(color='green')
# Distance between players
#plt.plot([defensive_team[3][0], attacking_team[8][0]], [defensive_team[3][1], attacking_team[8][1]], color = 'C0')
#plt.plot([defensive_team[4][0], attacking_team[3][0]], [defensive_team[4][1], attacking_team[3][1]], color = 'C0')
#plt.plot([defensive_team[9][0], attacking_team[2][0]], [defensive_team[9][1], attacking_team[2][1]], color = 'C0')
#plt.plot([defensive_team[5][0], attacking_team[7][0]], [defensive_team[5][1], attacking_team[7][1]], color = 'C0')
#plt.plot([defensive_team[1][0], attacking_team[9][0]], [defensive_team[1][1], attacking_team[9][1]], color = 'C0')
#plt.plot([defensive_team[2][0], attacking_team[10][0]], [defensive_team[2][1], attacking_team[10][1]], color = 'C0')
#plt.plot([defensive_team[6][0], attacking_team[6][0]], [defensive_team[6][1], attacking_team[6][1]], color = 'C0')
#plt.plot([defensive_team[7][0], attacking_team[4][0]], [defensive_team[7][1], attacking_team[4][1]], color = 'C0')
#plt.plot([defensive_team[8][0], attacking_team[5][0]], [defensive_team[8][1], attacking_team[5][1]], color = 'C0')
#plt.plot([defensive_team[8][0], attacking_team[1][0]], [defensive_team[8][1], attacking_team[1][1]], color = 'C0')
# Pitch control
print_df = pd.DataFrame([[i, j / (120 / 75)] for i in range(101)
for j in range(101)],
columns=['x', 'y'])
df = pd.DataFrame(attacking_team + defensive_team,
columns=['x', 'y', 'team'])
df.y /= 120 / 75
clf = KNeighborsClassifier(n_neighbors=1)
clf.fit(df[['x', 'y']], df.team)
print_df['team'] = clf.predict(print_df[['x', 'y']])
print_df.y *= 120 / 75
print_df_t1 = print_df[print_df.team == 0]
plt.scatter(print_df_t1['x'],
print_df_t1['y'],
3,
c='C0',
marker='s',
alpha=0.5)
print_df_t2 = print_df[print_df.team == 1]
plt.scatter(print_df_t2['x'],
print_df_t2['y'],
3,
c='C1',
marker='s',
alpha=0.5)
if save_fig_dir != '':
plt.savefig(save_fig_dir, bbox_inches='tight', dpi=dpi)
plt.clf()
else:
plt.show()
return
# Register process every TEST_ITERS iterations
if iteration % TEST_ITERS == 0:
rewards, steps = test_net(act_net, device=device)
print("%d, reward %.5f, steps %d" % (iteration, rewards, steps))
iteration_stats.append([iteration, rewards, steps])
pd.DataFrame(iteration_stats).to_csv('tmp/log' + ATTEMPT + '.csv')
if (best_reward is None or best_reward < rewards) and ENABLE_VIZ:
# Progress visualizer
viz_df = pd.DataFrame([[i, j] for i in np.arange(0, 1.01, 0.01) for j in np.arange(0, 1, 0.015)], columns = ['x','y'])
viz_df[['Shot', 'Pass', 'r', 'a']] = act_net(torch.FloatTensor(viz_df[['x','y']].to_numpy())).cpu().detach().numpy()
viz_df['x'] *= 100
viz_df['y'] *= 100
## Discrete action decisons
draw.pitch()
shot_action = viz_df[viz_df.Shot > viz_df.Pass]
plt.scatter(shot_action.x, shot_action.y, s = 15, c = 'C0', alpha = 0.5, marker = 's', linewidth=0, zorder = 10)
pass_action = viz_df[viz_df.Shot < viz_df.Pass]
plt.scatter(pass_action.x, pass_action.y, s = 15, c = 'C1', alpha = 0.5, marker = 's', linewidth=0, zorder = 10)
plt.savefig('img/decision_maps_' + ATTEMPT + '/passshotmap_' + str(iteration) + '.png', dpi = 900)
plt.clf()
## Continuous action decisions
draw.pitch()
for i, row in viz_df.iterrows():
if i % 3 == 0:
plt.arrow(row['x'], row['y'], row.r/abs(row.r) * np.cos((row['a'] - 0.5) * 2 * np.pi), row.r/abs(row.r) * np.sin((row['a'] - 0.5) * 2 * np.pi), length_includes_head = True, head_width = .5, head_length = .5, color = 'white')
viz_df['i'] = [i for i in range(len(viz_df))]
plt.scatter(viz_df[(viz_df.i % 3) == 0].x, viz_df[(viz_df.i % 3) == 0].y, s=(viz_df[(viz_df.i % 3) == 0].r + 1) ** 2, zorder = 99)
plt.savefig('img/decision_maps_' + ATTEMPT + '/orientationmap_' + str(iteration) + '.png', dpi = 900)