def animate_training_comparison(self, gridworld, learner_1, learner_2,
episode):
# make local copies of input
grid = gridworld
training_episodes_history_v1 = learner_1.training_episodes_history
training_episodes_history_v2 = learner_2.training_episodes_history
# initialize figure
fig = plt.figure(figsize=(10, 3))
axs = []
for i in range(2):
ax = fig.add_subplot(1, 2, i + 1, aspect='equal')
axs.append(ax)
# compute maximum length of episodes animated
max_len = 0
key = episode
L1 = len(training_episodes_history_v1[str(key)])
L2 = len(training_episodes_history_v2[str(key)])
max_len = max(L1, L2)
# loop over the episode histories and plot the results
rewards = np.zeros((2, 1))
def show_episode(step):
# loop over subplots and plot current step of each episode history
artist = fig
for k in range(len(axs)):
ax = axs[k]
# take correct episode
current_episode = 0
if k == 0:
current_episode = training_episodes_history_v1[str(key)]
else:
current_episode = training_episodes_history_v2[str(key)]
# define new location of agent
grid.agent = current_episode[min(step,
len(current_episode) - 1)]
# color gridworld for this episode and step
grid.color_gridworld(ax=ax)
return artist,
# create animation object
anim = animation.FuncAnimation(fig,
show_episode,
frames=min(100, max_len),
interval=min(100, max_len),
blit=True)
# set frames per second in animation
IPython_display.anim_to_html(anim, fps=min(100, max_len) / float(5))
return (anim)
def animate_training_runs(self, gridworld, learner, episodes):
# make local copies of input
grid = gridworld
training_episodes_history = learner.training_episodes_history
# initialize figure
fig = plt.figure(figsize=(10, 3))
axs = []
for i in range(len(episodes)):
ax = fig.add_subplot(1, len(episodes), i + 1, aspect='equal')
axs.append(ax)
if len(episodes) == 1:
axs = np.array(axs)
# compute maximum length of episodes animated
max_len = 0
for key in episodes:
l = len(training_episodes_history[str(key)])
if l > max_len:
max_len = l
# loop over the episode histories and plot the results
def show_episode(step):
# loop over subplots and plot current step of each episode history
artist = fig
for k in range(len(axs)):
ax = axs[k]
# take correct episode
episode_num = episodes[k]
current_episode = training_episodes_history[str(episode_num)]
# define new location of agent
grid.agent = current_episode[min(step,
len(current_episode) - 1)]
# color gridworld for this episode and step
grid.color_gridworld(ax=ax)
ax.set_title('episode = ' + str(episode_num + 1))
return artist,
# create animation object
anim = animation.FuncAnimation(fig,
show_episode,
frames=min(100, max_len),
interval=min(100, max_len),
blit=True)
# set frames per second in animation
IPython_display.anim_to_html(anim, fps=min(100, max_len) / float(5))
return (anim)
def plot_movie_js2(enc_array, image_array, save=None):
#Shows encoding and frames
fig = plt.figure(figsize=(10, 3), dpi=72)
ax1 = fig.add_subplot(2, 2, 1)
plt.title('Visual Encoding', fontsize=15)
plt.axis('off')
im = plt.imshow(enc_array[0][:8, :], vmin=-1, vmax=25)
ax2 = fig.add_subplot(2, 2, 3)
plt.title('Vector Encoding', fontsize=15)
plt.axis('off')
im2 = plt.imshow(enc_array[0][8:, :], vmin=-1, vmax=25)
ax3 = fig.add_subplot(1, 2, 2)
im3 = plt.imshow(image_array[0])
plt.axis('off')
Writer = animation.writers['ffmpeg']
writer = Writer(fps=10, metadata=dict(artist='Me'), bitrate=1800)
def animate(i):
im.set_array(enc_array[i][:8, :])
im2.set_array(enc_array[i][8:, :])
im3.set_array(image_array[i])
return (im, )
anim = animation.FuncAnimation(fig, animate, frames=len(image_array))
display(IPython_display.display_animation(anim))
if save != None:
anim.save(save, writer=writer)
def plot_movie_js3(enc_array, image_array, cluster, save=None):
#Plot Encodings and frames + information about which cluster the frame is in
fig = plt.figure(figsize=(10, 3), dpi=72)
ax1 = fig.add_subplot(2, 2, 1)
plt.axis('off')
im = plt.imshow(enc_array[0][:8, :], vmin=-1, vmax=25)
plt.title('Visual Encoding', fontsize=15)
ax2 = fig.add_subplot(2, 2, 3)
plt.axis('off')
im2 = plt.imshow(enc_array[0][8:, :], vmin=-1, vmax=25)
plt.title('Vector Encoding', fontsize=15)
ax4 = fig.add_subplot(1, 2, 2)
plt.axis('off')
im4 = plt.imshow(image_array[0])
ax3 = fig.add_subplot(6, 2, 2)
im3 = plt.text(0.3,
0.1,
'Cluster ' + str(cluster[0]),
fontsize=20,
color='white',
bbox=dict(facecolor='blue', alpha=0.5))
plt.axis('off')
Writer = animation.writers['ffmpeg']
writer = Writer(fps=10, metadata=dict(artist='Me'), bitrate=1800)
def animate(i):
im.set_array(enc_array[i][:8, :])
im2.set_array(enc_array[i][8:, :])
im3.set_text('Cluster ' + str(cluster[i]))
im4.set_array(image_array[i])
return (im, )
anim = animation.FuncAnimation(fig, animate, frames=len(image_array))
display(IPython_display.display_animation(anim))
if save != None:
anim.save(save, writer=writer)
def classification_slider(self, **args):
# run logistic regression
x = self.x_orig
y = self.data[:, 1]
##### precomputations #####
# precompute fits input
x_fit = np.linspace(np.min(x) - 1, np.max(x) + 1, 100)
# precompute surface
costs1 = [v[0] for v in self.cost_history]
costs2 = [v[1] for v in self.cost_history]
a = np.linspace(min(costs1), max(costs1))
b = np.linspace(min(costs2), max(costs2))
s, t = np.meshgrid(a, b)
shape = np.shape(s)
s = np.reshape(s, (np.size(s), 1))
t = np.reshape(t, (np.size(t), 1))
# generate surface based on given data - done very lazily - recomputed each time
g = 0
P = len(y)
for p in range(0, P):
g += np.log(1 + np.exp(-y[p] * (s + t * x[p])))
# reshape and plot the surface, as well as where the zero-plane is
s.shape = (shape)
t.shape = (shape)
g.shape = (shape)
##### start plotting #####
fig = plt.figure(num=None,
figsize=(12, 4),
dpi=80,
facecolor='w',
edgecolor='k')
ax1 = plt.subplot(131)
ax2 = plt.subplot(132)
ax3 = plt.subplot(133, projection='3d')
def show_fit(step):
### initialize plot data points and fit
# initialize fit
vals = self.cost_history[step]
b = vals[0]
w = vals[1]
#### print left panel ####
# transform input if needed for plotting
y_fit = np.tanh(b + x_fit * w)
# plot fit to data
ax1.cla()
ax1.plot(x_fit, y_fit, '-k', linewidth=3, zorder=0)
# initialize points
pos_inds = np.argwhere(y > 0)
pos_inds = [v[0] for v in pos_inds]
ax1.scatter(x[pos_inds],
y[pos_inds],
color='salmon',
linewidth=1,
marker='o',
edgecolor='k',
s=80)
neg_inds = np.argwhere(y < 0)
neg_inds = [v[0] for v in neg_inds]
ax1.scatter(x[neg_inds],
y[neg_inds],
color='cornflowerblue',
linewidth=1,
marker='o',
edgecolor='k',
s=80)
# clean up panel
xgap = float(max(x) - min(x)) / float(10)
ax1.set_xlim([min(x) - xgap, max(x) + xgap])
ygap = float(max(y) - min(y)) / float(10)
ax1.set_ylim([min(y) - ygap, max(y) + ygap])
ax1.set_xticks([])
ax1.set_yticks([])
ax1.set_title('logistic fit')
#### print middle panel ####
# plot fit to data
ax2.cla()
y_fit = np.sign(b + x_fit * w)
ax2.plot(x_fit, y_fit, '-k', linewidth=3, zorder=0)
# initialize points
pos_inds = np.argwhere(y > 0)
pos_inds = [v[0] for v in pos_inds]
ax2.scatter(x[pos_inds],
y[pos_inds],
color='salmon',
linewidth=1,
marker='o',
edgecolor='k',
s=80)
neg_inds = np.argwhere(y < 0)
neg_inds = [v[0] for v in neg_inds]
ax2.scatter(x[neg_inds],
y[neg_inds],
color='cornflowerblue',
linewidth=1,
marker='o',
edgecolor='k',
s=80)
# clean up panel
xgap = float(max(x) - min(x)) / float(10)
ax2.set_xlim([min(x) - xgap, max(x) + xgap])
ygap = float(max(y) - min(y)) / float(10)
ax2.set_ylim([min(y) - ygap, max(y) + ygap])
ax2.set_xticks([])
ax2.set_yticks([])
ax2.set_title('final classifier')
#### print right panel ####
ax3.cla()
ax3.plot_surface(s, t, g, alpha=0.15)
ax3.plot_surface(s, t, g * 0, alpha=0.1)
# plot all gradient descent steps faintly for visualization purposes
bs = []
ws = []
costs = []
for i in range(len(self.cost_history)):
bwg = self.cost_history[i]
b = bwg[0]
w = bwg[1]
cost = bwg[2]
bs.append(b)
ws.append(w)
costs.append(cost)
ax3.scatter(bs,
ws,
costs,
color='m',
marker='x',
linewidth=3,
alpha=0.1)
# plot current gradient descent step in bright red
b = vals[0]
w = vals[1]
cost = vals[2]
artist = ax3.scatter(b,
w,
cost,
marker='o',
color='r',
s=60,
edgecolor='k',
linewidth=4)
# clean up panel
ax3.view_init(args['view'][0], args['view'][1])
ax3.set_xticks([])
ax3.set_yticks([])
ax3.set_zticks([])
ax3.set_title('cost function')
ax3.set_xlabel(args['xlabel'], fontsize=14, labelpad=-5)
ax3.set_ylabel(args['ylabel'], fontsize=14, labelpad=-5)
ax3.zaxis.set_rotate_label(False) # disable automatic rotation
ax3.set_zlabel('cost ', fontsize=14, rotation=0, labelpad=1)
return artist,
anim = animation.FuncAnimation(fig,
show_fit,
frames=len(self.cost_history),
interval=len(self.cost_history),
blit=True)
# set frames per second in animation
IPython_display.anim_to_html(anim,
fps=int(
len(self.cost_history) / float(3)))
return (anim)
def fitting_slider(self, **args):
# pull out coordinates
x_orig = self.x_orig
x_tran = self.data[:, 0]
y = self.data[:, 1]
##### precomputations #####
# precompute fits input
x_fit = np.linspace(np.min(x_orig) - 1, np.max(x_orig) + 1, 100)
# precompute surface
xs = max([abs(v[0]) for v in self.cost_history])
ys = max([abs(v[1]) for v in self.cost_history])
minval = min(-xs, -ys)
maxval = max(xs, ys)
gap = (maxval - minval) * 0.2
r = np.linspace(minval - gap, maxval + gap)
s, t = np.meshgrid(r, r)
s = np.reshape(s, (np.size(s), 1))
t = np.reshape(t, (np.size(t), 1))
# generate surface based on given data - done very lazily - recomputed each time
g = 0
P = len(y)
if args['fit_type'] == 'line fit' or args['fit_type'] == 'sine fit':
for p in range(0, P):
g += (s + t * x_tran[p] - y[p])**2
if args['fit_type'] == 'logistic fit':
for p in range(0, P):
g += np.log(1 + np.exp(-y[p] * (s + t * x_tran[p])))
# reshape and plot the surface, as well as where the zero-plane is
s.shape = (np.size(r), np.size(r))
t.shape = (np.size(r), np.size(r))
g.shape = (np.size(r), np.size(r))
# setup figure to plot
fig = plt.figure(num=None,
figsize=(12, 4),
dpi=80,
facecolor='w',
edgecolor='k')
ax1 = plt.subplot(121)
ax2 = plt.subplot(122, projection='3d')
# slider mechanism
def show_fit(step):
### initialize plot data points and fit
# initialize fit
vals = self.cost_history[step]
b = vals[0]
w = vals[1]
yfit = 0
# transform input if needed for plotting
if args['fit_type'] == 'line fit':
y_fit = b + x_fit * w
if args['fit_type'] == 'sine fit':
y_fit = b + np.sin(2 * np.pi * x_fit) * w
if args['fit_type'] == 'logistic fit':
y_fit = np.tanh(b + x_fit * w)
# plot fit to data
ax1.cla()
ax1.plot(x_fit, y_fit, '-r', linewidth=3)
# initialize points
ax1.scatter(x_orig, y)
# clean up panel
xgap = float(max(x_orig) - min(x_orig)) / float(10)
ax1.set_xlim([min(x_orig) - xgap, max(x_orig) + xgap])
ygap = float(max(y) - min(y)) / float(10)
ax1.set_ylim([min(y) - ygap, max(y) + ygap])
ax1.set_xticks([])
ax1.set_yticks([])
### plot surface
ax2.cla()
artist = ax2.plot_surface(s, t, g, alpha=0.15)
ax2.plot_surface(s, t, g * 0, alpha=0.1)
# plot all gradient descent steps faintly for visualization purposes
bs = []
ws = []
costs = []
for i in range(len(self.cost_history)):
bwg = self.cost_history[i]
b = bwg[0]
w = bwg[1]
cost = bwg[2]
bs.append(b)
ws.append(w)
costs.append(cost)
ax2.scatter(bs,
ws,
costs,
color='m',
marker='x',
linewidth=3,
alpha=0.1)
# plot current gradient descent step in bright red
b = vals[0]
w = vals[1]
cost = vals[2]
ax2.scatter(b,
w,
cost,
marker='o',
color='r',
s=50,
edgecolor='k',
linewidth=1)
# clean up panel
ax2.view_init(args['view'][0], args['view'][1])
ax2.set_xticks([])
ax2.set_yticks([])
ax2.set_zticks([])
ax2.set_xlabel(args['xlabel'], fontsize=14, labelpad=-5)
ax2.set_ylabel(args['ylabel'], fontsize=14, labelpad=-5)
ax2.zaxis.set_rotate_label(False) # disable automatic rotation
ax2.set_zlabel('cost ', fontsize=14, rotation=0, labelpad=1)
return artist,
anim = animation.FuncAnimation(fig,
show_fit,
frames=len(self.cost_history),
interval=len(self.cost_history),
blit=True)
# set frames per second in animation
IPython_display.anim_to_html(anim,
fps=int(
len(self.cost_history) / float(3)))
return (anim)
def animate_training_runs(self, **args):
grid = args['gridworld']
episodes = args['episodes']
learner = args['learner']
# make local copies of input
training_episodes_history = learner.training_episodes_history
# initialize figure
fsize = 3
if grid.width:
fsize = 5
fig = plt.figure(figsize=(12, fsize))
axs = []
for i in range(len(episodes)):
ax = fig.add_subplot(1, len(episodes), i + 1, aspect='equal')
axs.append(ax)
if len(episodes) == 1:
axs = np.array(axs)
# compute maximum length of episodes animated
max_len = 0
for key in episodes:
l = len(training_episodes_history[str(key)])
if l > max_len:
max_len = l
# loop over the episode histories and plot the results
print 'animating run...'
def show_episode(step):
# loop over subplots and plot current step of each episode history
artist = fig
for k in range(len(axs)):
ax = axs[k]
# take correct episode
episode_num = episodes[k]
current_episode = training_episodes_history[str(episode_num)]
# define new location of agent
grid.agent = current_episode[min(step,
len(current_episode) - 1)]
# color gridworld for this episode and step
if 'lights' not in args:
grid.color_gridworld(ax=ax)
else:
grid.color_gridworld(ax=ax, lights=args['lights'])
ax.set_title('episode = ' + str(episode_num + 1))
return artist,
# create animation object
anim = animation.FuncAnimation(fig,
show_episode,
frames=min(100, max_len),
interval=min(100, max_len),
blit=True)
# set frames per second in animation
IPython_display.anim_to_html(anim, fps=min(100, max_len) / float(10))
print '...done!'
time.sleep(1)
clear_output()
return (anim)
def animate_validation_runs(self, **args):
gridworld = args['gridworld']
learner = args['learner']
starting_locations = args['starting_locations']
# make local copies of input
grid = gridworld
Q = learner.Q
starting_locs = starting_locations
# initialize figure
fsize = 3
if grid.width:
fsize = 5
fig = plt.figure(figsize=(12, fsize))
axs = []
for i in range(len(starting_locs)):
ax = fig.add_subplot(1, len(starting_locs), i + 1, aspect='equal')
axs.append(ax)
# only one added subplot, axs must be in array format
if len(starting_locs) == 1:
axs = np.array(axs)
### produce validation runs ###
print 'animating run...'
validation_run_history = []
for i in range(len(starting_locs)):
# take random starting point - for short just from validation schedule
grid.agent = starting_locs[i]
# loop over max number of steps and try reach goal
episode_path = []
for j in range(grid.max_steps):
# store current location
episode_path.append(grid.agent)
### if you reach the goal end current episode immediately
if grid.agent == grid.goal:
break
# translate current agent location tuple into index
s_k_1 = grid.state_tuple_to_index(grid.agent)
# get action
a_k = grid.get_action(method='optimal', Q=Q)
# move based on this action - if move takes you out of gridworld don't move and instead move randomly
s_k = grid.get_movin(action=a_k,
illegal_move_response='random')
# record next step in path
grid.agent = grid.state_index_to_tuple(state_index=s_k)
# record this episode's path
validation_run_history.append(episode_path)
### compute maximum length of episodes animated ###
max_len = 0
for i in range(len(starting_locs)):
l = len(validation_run_history[i])
if l > max_len:
max_len = l
### loop over the episode histories and plot the results ###
def show_episode(step):
# loop over subplots and plot current step of each episode history
artist = fig
for k in range(len(axs)):
ax = axs[k]
# take correct episode
current_episode = validation_run_history[k]
# define new location of agent
loc = current_episode[min(step, len(current_episode) - 1)]
grid.agent = loc
# color gridworld for this episode and step
if 'lights' not in args:
grid.color_gridworld(ax=ax)
else:
grid.color_gridworld(ax=ax, lights=args['lights'])
ax.set_title('fully trained run ' + str(k + 1))
# fig.subplots_adjust(left=0,right=1,bottom=0,top=1) ## gets rid of the white space around image
return artist,
# create animation object
anim = animation.FuncAnimation(fig,
show_episode,
frames=min(100, max_len),
interval=min(100, max_len),
blit=True)
# set frames per second in animation
IPython_display.anim_to_html(anim, fps=min(100, max_len) / float(10))
print '...done!'
time.sleep(1)
clear_output()
return (anim)
def animate_training_comparison(self, **args):
grid = args['gridworld']
learner_1 = args['learner_1']
learner_2 = args['learner_2']
episode = args['episode']
# make local copies of input
training_episodes_history_v1 = learner_1.training_episodes_history
training_episodes_history_v2 = learner_2.training_episodes_history
# initialize figure
fsize = 3
if grid.width > 10:
fsize = 5
fig = plt.figure(figsize=(12, fsize))
axs = []
for i in range(2):
ax = fig.add_subplot(1, 2, i + 1, aspect='equal')
axs.append(ax)
# compute maximum length of episodes animated
max_len = 0
key = episode
L1 = len(training_episodes_history_v1[str(key)])
L2 = len(training_episodes_history_v2[str(key)])
max_len = max(L1, L2)
# loop over the episode histories and plot the results
print 'animating run...'
rewards = np.zeros((2, 1))
def show_episode(step):
# loop over subplots and plot current step of each episode history
artist = fig
for k in range(len(axs)):
ax = axs[k]
# take correct episode
current_episode = 0
if k == 0:
current_episode = training_episodes_history_v1[str(key)]
else:
current_episode = training_episodes_history_v2[str(key)]
# define new location of agent
grid.agent = current_episode[min(step,
len(current_episode) - 1)]
# color gridworld for this episode and step
if 'lights' not in args:
grid.color_gridworld(ax=ax)
else:
grid.color_gridworld(ax=ax, lights=args['lights'])
# set title
if k == 0:
ax.set_title('random')
else:
ax.set_title('exploration/exploitation')
return artist,
# create animation object
anim = animation.FuncAnimation(fig,
show_episode,
frames=min(100, max_len),
interval=min(100, max_len),
blit=True)
# set frames per second in animation
IPython_display.anim_to_html(anim, fps=min(100, max_len) / float(10))
print '...done!'
time.sleep(1)
clear_output()
return (anim)
def plot_movie_jsInfo(enc_array, image_array, acts, vals, rews, save=None):
def getImage(act, num):
if act == 0:
return stand
if num == 0:
if act == 1:
return up
if act == 2:
return down
if num == 1:
if act == 1:
return turn_l
if act == 2:
return turn_r
if num == 2 and act == 1:
return jump
if num == 3:
if act == 1:
return right
if act == 2:
return left
jump = imageio.imread('./symbols/jump.png')
left = imageio.imread('./symbols/arrow-left.png')
right = imageio.imread('./symbols/arrow_right.png')
down = imageio.imread('./symbols/down-arrow.png')
up = imageio.imread('./symbols/up-arrow.png')
turn_l = imageio.imread('./symbols/turn-left.png')
turn_r = imageio.imread('./symbols/turn-right.png')
stand = imageio.imread('./symbols/Stand.png')
fig = plt.figure(figsize=(10, 3), dpi=72)
ax1 = fig.add_subplot(2, 2, 1)
plt.axis('off')
if not isinstance(enc_array, list):
im = plt.imshow(enc_array[0][:8, :], vmin=-1, vmax=25)
plt.title('Visual Encoding', fontsize=15)
else:
icaLen = enc_array[0][0].shape[0]
im = plt.imshow(enc_array[0][0].reshape(1, icaLen), vmin=-1, vmax=25)
plt.title('Visual Encoding - ICs', fontsize=15)
ax2 = fig.add_subplot(2, 2, 3)
plt.axis('off')
if not isinstance(enc_array, list):
im2 = plt.imshow(enc_array[0][8:, :], vmin=-1, vmax=25)
plt.title('Vector Encoding', fontsize=15)
else:
im2 = plt.imshow(enc_array[1][0].reshape(1, icaLen), vmin=-1, vmax=25)
plt.title('Vector Encoding - ICs', fontsize=15)
ax4 = fig.add_subplot(1, 2, 2)
im4 = plt.imshow(image_array[0])
plt.axis('off')
ax3 = fig.add_subplot(6, 2, 2)
im3 = plt.text(0.2,
0.1,
"R: " + str(rews[0]) + ' V: ' + str(vals[0]),
fontsize=15,
color='white',
bbox=dict(facecolor='blue', alpha=0.5))
plt.axis('off')
ax5 = fig.add_subplot(4, 10, 10)
im5 = plt.imshow(getImage(acts[0][0][0], 0))
plt.axis('off')
ax6 = fig.add_subplot(4, 10, 20)
im6 = plt.imshow(getImage(acts[0][0][1], 1))
plt.axis('off')
ax7 = fig.add_subplot(4, 10, 30)
im7 = plt.imshow(getImage(acts[0][0][2], 2))
plt.axis('off')
ax8 = fig.add_subplot(4, 10, 40)
im8 = plt.imshow(getImage(acts[0][0][3], 3))
plt.axis('off')
Writer = animation.writers['ffmpeg']
writer = Writer(fps=10, metadata=dict(artist='Me'), bitrate=1800)
def animate(i):
if not isinstance(enc_array, list):
im.set_array(enc_array[i][:8, :])
im2.set_array(enc_array[i][8:, :])
else:
im.set_array(enc_array[0][i].reshape(1, icaLen))
im2.set_array(enc_array[1][i].reshape(1, icaLen))
im3.set_text("R: " + str(rews[i])[:3] + ' V: ' +
str(vals[i][0][0])[:4])
im4.set_array(image_array[i])
im5.set_array(getImage(acts[i][0][0], 0))
im6.set_array(getImage(acts[i][0][1], 1))
im7.set_array(getImage(acts[i][0][2], 2))
im8.set_array(getImage(acts[i][0][3], 3))
return (im, )
anim = animation.FuncAnimation(fig, animate, frames=len(image_array))
display(IPython_display.display_animation(anim))
if save != None:
anim.save(save, writer=writer)
def plot_movie_semantic2(semantic, image_array, acts, vals, rews, save=None):
label_dict = {
"Unknown": 0,
"Agent": 1,
"Level Door": 2,
"Regular Door": 3,
"Key Door": 4,
"Entry Door": 5,
"Puzzle Door": 6,
"Key": 7,
"Time Orb": 8,
"Wall": 9,
"Floor": 10
}
inv_map = {v: k for k, v in label_dict.items()}
def getImage(act, num):
if act == 0:
return stand
if num == 0:
if act == 1:
return up
if act == 2:
return down
if num == 1:
if act == 1:
return turn_l
if act == 2:
return turn_r
if num == 2 and act == 1:
return jump
if num == 3:
if act == 1:
return right
if act == 2:
return left
jump = imageio.imread('./symbols/jump.png')
left = imageio.imread('./symbols/arrow-left.png')
right = imageio.imread('./symbols/arrow_right.png')
down = imageio.imread('./symbols/down-arrow.png')
up = imageio.imread('./symbols/up-arrow.png')
turn_l = imageio.imread('./symbols/turn-left.png')
turn_r = imageio.imread('./symbols/turn-right.png')
stand = imageio.imread('./symbols/Stand.png')
fig = plt.figure(figsize=(10, 3), dpi=72)
ax1 = fig.add_subplot(1, 2, 1)
plt.axis('off')
im1 = plt.imshow(rgb2L(semantic[0]), vmin=0, vmax=11, cmap='tab20')
values = np.linspace(0, 10, 11)
colors = [im1.cmap(im1.norm(value)) for value in values]
patches = [
mpatches.Patch(color=colors[i], label=inv_map[values[i]])
for i in range(len(values))
]
plt.legend(handles=patches,
bbox_to_anchor=(1.05, 1),
loc=2,
borderaxespad=0.)
ax4 = fig.add_subplot(1, 2, 2)
im4 = plt.imshow(image_array[0])
plt.axis('off')
ax3 = fig.add_subplot(6, 2, 2)
im3 = plt.text(0.2,
0.1,
"0 R: " + str(rews[0]) + ' V: ' + str(vals[0]),
fontsize=15,
color='white',
bbox=dict(facecolor='blue', alpha=0.5))
plt.axis('off')
ax5 = fig.add_subplot(4, 10, 10)
im5 = plt.imshow(getImage(acts[0][0][0], 0))
plt.axis('off')
ax6 = fig.add_subplot(4, 10, 20)
im6 = plt.imshow(getImage(acts[0][0][1], 1))
plt.axis('off')
ax7 = fig.add_subplot(4, 10, 30)
im7 = plt.imshow(getImage(acts[0][0][2], 2))
plt.axis('off')
ax8 = fig.add_subplot(4, 10, 40)
im8 = plt.imshow(getImage(acts[0][0][3], 3))
plt.axis('off')
Writer = animation.writers['ffmpeg']
writer = Writer(fps=10, metadata=dict(artist='Me'), bitrate=1800)
def animate(i):
try:
im1.set_array(rgb2L(semantic[i]))
except:
print(str(i) + " - " + str(broken.shape))
im3.set_text(
str(i) + " R: " + str(rews[i])[:3] + ' V: ' +
str(vals[i][0][0])[:4])
im4.set_array(image_array[i])
im5.set_array(getImage(acts[i][0][0], 0))
im6.set_array(getImage(acts[i][0][1], 1))
im7.set_array(getImage(acts[i][0][2], 2))
im8.set_array(getImage(acts[i][0][3], 3))
return (im1, )
anim = animation.FuncAnimation(fig, animate, frames=len(image_array))
display(IPython_display.display_animation(anim))
if save != None:
anim.save(save, writer=writer)
def plot_movie_semantic(semantic, image_array, acts, vals, rews, save=None):
def getImage(act, num):
if act == 0:
return stand
if num == 0:
if act == 1:
return up
if act == 2:
return down
if num == 1:
if act == 1:
return turn_l
if act == 2:
return turn_r
if num == 2 and act == 1:
return jump
if num == 3:
if act == 1:
return right
if act == 2:
return left
jump = imageio.imread('./symbols/jump.png')
left = imageio.imread('./symbols/arrow-left.png')
right = imageio.imread('./symbols/arrow_right.png')
down = imageio.imread('./symbols/down-arrow.png')
up = imageio.imread('./symbols/up-arrow.png')
turn_l = imageio.imread('./symbols/turn-left.png')
turn_r = imageio.imread('./symbols/turn-right.png')
stand = imageio.imread('./symbols/Stand.png')
fig = plt.figure(figsize=(10, 3), dpi=72)
ax1 = fig.add_subplot(1, 2, 1)
plt.axis('off')
lbls = np.rot90(
np.array([int(n) for n in semantic[0][1:-4].split(",")]).reshape(
(128, 128)))
im1 = plt.imshow(lbls, vmin=-1, vmax=10, cmap='tab20')
values = np.linspace(-1, 10, 12)
colors = [im1.cmap(im1.norm(value)) for value in values]
patches = [
mpatches.Patch(color=colors[i], label=inv_map[values[i]])
for i in range(len(values))
]
plt.legend(handles=patches,
bbox_to_anchor=(1.05, 1),
loc=2,
borderaxespad=0.)
ax4 = fig.add_subplot(1, 2, 2)
im4 = plt.imshow(image_array[0])
plt.axis('off')
ax3 = fig.add_subplot(6, 2, 2)
im3 = plt.text(0.2,
0.1,
"R: " + str(rews[0]) + ' V: ' + str(vals[0]),
fontsize=15,
color='white',
bbox=dict(facecolor='blue', alpha=0.5))
plt.axis('off')
ax5 = fig.add_subplot(4, 10, 10)
im5 = plt.imshow(getImage(acts[0][0][0], 0))
plt.axis('off')
ax6 = fig.add_subplot(4, 10, 20)
im6 = plt.imshow(getImage(acts[0][0][1], 1))
plt.axis('off')
ax7 = fig.add_subplot(4, 10, 30)
im7 = plt.imshow(getImage(acts[0][0][2], 2))
plt.axis('off')
ax8 = fig.add_subplot(4, 10, 40)
im8 = plt.imshow(getImage(acts[0][0][3], 3))
plt.axis('off')
Writer = animation.writers['ffmpeg']
writer = Writer(fps=10, metadata=dict(artist='Me'), bitrate=1800)
def animate(i):
try:
lbls = np.rot90(
np.array([int(n)
for n in semantic[i][1:-4].split(",")]).reshape(
(128, 128)))
im1.set_array(lbls)
except:
broken = np.array([int(n) for n in data[i][1:-4].split(",")])
lbls = np.rot90(
np.append(broken,
np.zeros((128 * 128) - broken.shape[0])).reshape(
(128, 128)))
im1.set_array(lbls)
print(str(i) + " - " + str(broken.shape))
im3.set_text("R: " + str(rews[i])[:3] + ' V: ' +
str(vals[i][0][0])[:4])
im4.set_array(image_array[i])
im5.set_array(getImage(acts[i][0][0], 0))
im6.set_array(getImage(acts[i][0][1], 1))
im7.set_array(getImage(acts[i][0][2], 2))
im8.set_array(getImage(acts[i][0][3], 3))
return (im1, )
anim = animation.FuncAnimation(fig, animate, frames=len(image_array))
display(IPython_display.display_animation(anim))
if save != None:
anim.save(save, writer=writer)
def plot_movie_curInfo(enc_array,
image_array,
pred_s,
enc_cur,
acts,
vals,
rews,
range_vis,
range_vec,
save=None):
fig = plt.figure(figsize=(10, 3), dpi=72)
ax1 = fig.add_subplot(3, 3, 1)
plt.axis('off')
if not isinstance(enc_array, list):
im = plt.imshow(enc_array[0][:4, :],
vmin=range_vis[0],
vmax=range_vis[1])
plt.title('Visual Encoding', fontsize=15)
else:
icaLen = enc_array[0][0].shape[0]
im = plt.imshow(enc_array[0][0].reshape(1, icaLen))
plt.title('Visual Encoding - ICs', fontsize=15)
ax2 = fig.add_subplot(3, 3, 2)
plt.axis('off')
if not isinstance(enc_array, list):
im2 = plt.imshow(enc_array[0][4:, :],
vmin=range_vec[0],
vmax=range_vec[1])
plt.title('Vector Encoding', fontsize=15)
else:
im2 = plt.imshow(enc_array[1][0].reshape(1, icaLen))
plt.title('Vector Encoding - ICs', fontsize=15)
ax5 = fig.add_subplot(3, 3, 4)
plt.axis('off')
im5 = plt.imshow(pred_s[0][:4, :], vmin=range_vis[0], vmax=range_vis[1])
plt.title('Predicted')
ax6 = fig.add_subplot(3, 3, 5)
plt.axis('off')
im6 = plt.imshow(pred_s[0][4:, :], vmin=range_vec[0], vmax=range_vec[1])
plt.title('Predicted')
ax7 = fig.add_subplot(3, 3, 7)
plt.axis('off')
im7 = plt.imshow(enc_cur[0][:4, :], vmin=range_vis[0], vmax=range_vis[1])
plt.title('Actual')
ax8 = fig.add_subplot(3, 3, 8)
plt.axis('off')
im8 = plt.imshow(enc_cur[0][4:, :], vmin=range_vec[0], vmax=range_vec[1])
plt.title('Actual')
ax4 = fig.add_subplot(1, 3, 3)
im4 = plt.imshow(image_array[0])
plt.axis('off')
ax3 = fig.add_subplot(6, 3, 3)
im3 = plt.text(0.2,
0.1,
"R: " + str(rews[0]) + ' V: ' + str(vals[0]),
fontsize=15,
color='white',
bbox=dict(facecolor='blue', alpha=0.5))
plt.axis('off')
Writer = animation.writers['ffmpeg']
writer = Writer(fps=10, metadata=dict(artist='Me'), bitrate=1800)
def animate(i):
if not isinstance(enc_array, list):
im.set_array(enc_array[i][:4, :])
im2.set_array(enc_array[i][4:, :])
else:
im.set_array(enc_array[0][i].reshape(1, icaLen))
im2.set_array(enc_array[1][i].reshape(1, icaLen))
im3.set_text("R: " + str(rews[i])[:3] + ' V: ' +
str(vals[i][0][0])[:4])
im4.set_array(image_array[i])
im5.set_array(pred_s[i][:4, :])
im6.set_array(pred_s[i][4:, :])
im7.set_array(enc_cur[i][:4, :])
im8.set_array(enc_cur[i][4:, :])
return (im, )
anim = animation.FuncAnimation(fig, animate, frames=len(image_array))
display(IPython_display.display_animation(anim))
if save != None:
anim.save(save, writer=writer)
def transformation_slider(self):
print "rendering JS animation, this can take several minutes..."
#### make points start and end
X1 = self.orig_data
X2 = self.transformed_data
### make separator start and end
r = np.linspace(-10.1,10.1,2000)
a,b = np.meshgrid(r,r)
a = np.reshape(a,(np.size(a),1))
b = np.reshape(b,(np.size(b),1))
h = np.concatenate((a,b),axis = 1)
### use rule to partition the input space
f1,f2 = self.make_rule(h)
z = f1 + f2
z = (np.sign(z))
a.shape = (np.size(r),np.size(r))
b.shape = (np.size(r),np.size(r))
z.shape = (np.size(r),np.size(r))
f1.shape = (np.size(r),np.size(r))
f2.shape = (np.size(r),np.size(r))
#### make grid start and end
grid1 = self.grid
grida,gridb = self.make_rule(self.grid)
grid2 = np.concatenate((grida,gridb),axis = 1)
# make figure
fig = plt.figure(figsize = (5,5))
ax1 = fig.add_subplot(111) # panel for original space
def show_fit(p):
ax1.cla()
## make alpha
alpha = p/float(10)
#### setup current points and print
T = (1-alpha)*X1 + alpha*X2
# custom colors
red = 'salmon' # custom color for plotting purposes
blue = 'cornflowerblue' # custom color for plotting purposes
# plot points on desired panel
ax1.scatter(T[self.ind0,0],T[self.ind0,1],s = 60, color = blue, edgecolor = 'k')
artist = ax1.scatter(T[self.ind1,0],T[self.ind1,1],s = 60, color = red, edgecolor = 'k')
#### setup separator and print
sep1 = (1-alpha)*a + alpha*f1
sep2 = (1-alpha)*b + alpha*f2
# the cntr command grabs a contour without plotting it
c = cntr.Cntr(sep1, sep2, z)
res = c.trace(0) # here trace grabs a contour at slice z = value
# plot points
for k in range(len(res) - 1):
data = res[k] # extract the right array from the trace object
ax1.plot(data[:,0],data[:,1],'k', linewidth = 3)
### setup current grid and print
grid = (1-alpha)*grid1 + alpha*grid2
# plot points
for i in range(80):
ax1.plot(grid[200*i:(i+1)*200,0],grid[200*i:(i+1)*200,1],color = [0.75,0.75,0.75],linewidth = 1,zorder = 0)
### set axis limits for current dataset
hgap = (max(T[:,0]) - min(T[:,0]))*0.05
vgap = (max(T[:,1]) - min(T[:,1]))*0.05
ax1.set_xlim([min(T[:,0])-hgap,max(T[:,0])+hgap])
ax1.set_ylim([min(T[:,1])-vgap,max(T[:,1])+vgap])
ax1.axis('off')
fig.subplots_adjust(left=0,right=1,bottom=0,top=1) ## gets rid of the white space around image
return artist,
anim = animation.FuncAnimation(fig, show_fit,frames=10, interval=5, blit=True)
# set frames per second in animation
IPython_display.anim_to_html(anim,fps = 2)
return(anim)
def animate_fit(self, **args):
# run logistic regression
classifier = demo_logistic_regression()
self.w, self.cost_history = classifier.fit(self.X, self.y, **args)
# setup range for plots
costs = [v[-1] for v in self.cost_history]
X = self.X
y = self.y
xgap = float(max(X[1, :]) - min(X[1, :])) / float(10)
ygap = float(max(X[2, :]) - min(X[2, :])) / float(10)
cgap = float(max(costs) - min(costs)) / float(10)
# take positive and negative label indicies
pos_inds = np.argwhere(y > 0)
pos_inds = [v[0] for v in pos_inds]
neg_inds = np.argwhere(y < 0)
neg_inds = [v[0] for v in neg_inds]
# setup surface data for left and middle plot
colors = ['salmon', 'cornflowerblue']
a = np.linspace(np.min(X[:, 0]) - xgap, np.max(X[:, 0]) + xgap, 200)
b = np.linspace(np.min(X[:, 1]) - ygap, np.max(X[:, 1]) + ygap, 200)
s, t = np.meshgrid(a, b)
shape = np.shape(s)
s = np.reshape(s, (np.size(s), 1))
t = np.reshape(t, (np.size(t), 1))
h = np.concatenate((s, t), 1)
s.shape = (shape)
t.shape = (shape)
# get kernelized data
if "kernel" in args:
h = classifier.kernelize_test(h)
else:
o = np.ones((np.shape(h)[0], 1))
h = np.concatenate((o, h), 1)
# plotting
fig = plt.figure(num=None,
figsize=(15, 5),
dpi=80,
facecolor='w',
edgecolor='k')
ax1 = plt.subplot(131)
ax2 = plt.subplot(132, projection='3d')
ax3 = plt.subplot(133)
# slider mechanism
def show_fit(step):
#### print left panel ####
ax1.cla()
ax1.scatter(X[pos_inds, 0],
X[pos_inds, 1],
color=colors[0],
linewidth=1,
marker='o',
edgecolor='k',
s=80)
ax1.scatter(X[neg_inds, 0],
X[neg_inds, 1],
color=colors[1],
linewidth=1,
marker='o',
edgecolor='k',
s=80)
# print approximation
w_now = self.cost_history[step][:-1]
w_now = np.asarray([u[0] for u in w_now])
w_now.shape = (len(w_now), 1)
z = np.tanh(np.dot(h, w_now))
z2 = np.sign(np.copy(z))
z2.shape = (shape)
unique_labels = np.unique(y)
levels = unique_labels
ax1.contourf(s, t, z2, colors=colors, alpha=0.2)
# show the classification boundary if it exists
if len(np.unique(z2)) > 1:
ax1.contour(s,
t,
z2,
colors='k',
linewidths=2.5,
levels=unique_labels)
# clean up panel
ax1.set_xlim([min(X[:, 0]) - xgap, max(X[:, 0]) + xgap])
ax1.set_ylim([min(X[:, 1]) - ygap, max(X[:, 1]) + ygap])
ax1.set_xticks([])
ax1.set_yticks([])
#### print middle panel ####
ax2.cla()
# compute surface - due to bug in matplotlib surface will sometimes appear in front of points, sometimes behind, depending on angle and regardless of zorder setting
z.shape = (shape)
ax2.plot_surface(s, t, z, alpha=0.1, color='k', zorder=0)
# plot points
ax2.scatter(X[pos_inds, 0],
X[pos_inds, 1],
y[pos_inds],
color=colors[0],
linewidth=1,
marker='o',
edgecolor='k',
s=80,
zorder=1)
ax2.scatter(X[neg_inds, 0],
X[neg_inds, 1],
y[neg_inds],
color=colors[1],
linewidth=1,
marker='o',
edgecolor='k',
s=80,
zorder=1)
## clean up panel
# set viewing angle
ax2.view_init(args['view'][0], args['view'][1])
# set viewing limits
ax2.set_xlim([min(X[:, 0]) - xgap, max(X[:, 0]) + xgap])
ax2.set_ylim([min(X[:, 1]) - ygap, max(X[:, 1]) + ygap])
# turn off tick labels
ax2.set_xticklabels([])
ax2.set_yticklabels([])
ax2.set_zticklabels([])
# Get rid of the spines on the 3d plot
ax2.w_xaxis.line.set_color((1.0, 1.0, 1.0, 0.0))
ax2.w_yaxis.line.set_color((1.0, 1.0, 1.0, 0.0))
ax2.w_zaxis.line.set_color((1.0, 1.0, 1.0, 0.0))
# turn off tick marks
ax2.xaxis.set_tick_params(size=0, color='w')
ax2.yaxis.set_tick_params(size=0, color='w')
ax2.zaxis.set_tick_params(size=0, color='w')
#### print right panel ####
# print all of the step cost function values lightly
ax3.cla()
ax3.plot(costs, color='m', marker='o', linewidth=3, alpha=0.05)
# print current value
artist = ax3.scatter(step,
costs[step],
marker='o',
color='r',
s=60,
edgecolor='k',
linewidth=2)
# dress up plot
ax3.set_xlabel('step', fontsize=15, labelpad=5)
ax3.set_ylabel('cost function value',
fontsize=12,
rotation=90,
labelpad=5)
ax3.set_ylim(min(costs) - cgap, max(costs) + cgap)
# ax3.set_xticks([])
# ax3.set_yticks([])
return artist,
anim = animation.FuncAnimation(fig,
show_fit,
frames=len(self.cost_history),
interval=len(self.cost_history),
blit=True)
# set frames per second in animation
IPython_display.anim_to_html(anim,
fps=int(
len(self.cost_history) / float(3)))
return (anim)