def plot_diagnostics(self, figdir=None):
"""
FUNCTION: plots diagnostics of sequences generated.
Including MSD, autocorrelation, covering percentage vs time.
INPUTS: self = propagator with propagated sequences.
figdir = output figure diretory
"""
assert self.diagnostics_computed, 'EXPLORER: diagnostics not computed'
self.fig, self.axes = plt.subplots(1, 3, figsize=(18, 6), sharey=False)
self.set_target_axis(ax=self.axes[0])
self.plot_coverage()
self.set_target_axis(ax=self.axes[1])
self.plot_exploration_efficiency()
self.set_target_axis(ax=self.axes[2])
self.plot_msd()
plt.tight_layout()
if figdir is not None:
save_figure(fig=self.fig,
figdir=figdir,
fname_base='exploration_analysis',
file_ext='.png')
def plot_learning_errors(self, objs=["T_error"], figdir=None):
"""
FUNCTION: Plots learning objectives as a function of number of sequence samples.
INPUTS:
objs = ['Q_error', 'Q_corr', 'T_error', 'T_corr', 'SR_error', 'SR_corr', 'SPL_error', 'KLtraj', 'R_error', 'policy_error'] objectives to learn
[Superseded by plot_performance_metrics]
"""
assert self.error_curves_computed, "Learning curves not yet computed."
self.fig_axis()
ax = self.ax
for key in objs:
loss = self._retrieve_learn_var(n_samp=None, key=key).values
ax.plot(range(1, self.n_samp + 1), loss, "--o", label=key)
ax.set_ylabel("Euclidean error")
ax.set_xlabel("#samples")
ax.legend(loc="center left")
if figdir is not None:
save_figure(
fig=self.fig,
figdir=figdir,
fname_base="learning_analysis",
file_ext=".png",
)
options.periodic = False
options.norm_cov = False
options.gauss_norm = False
options.box_width = 2.2
options.box_height = 2.2
# run grid pattern formation
Q = diff_gen(options=options, sigma=0.5, surround_scale=1.5)
options.Q = to.as_tensor(Q)
place_cells = PredictivePlaceCells(options=options)
place_cells.compute_covariance(predmap=True)
place_cells.Cmean = gaussian_filter(input=place_cells.Cmean, sigma=4., truncate=6., order=0, mode='wrap', cval=0.0)
place_cells.fft_covariance(predmap=True)
# rescale/smooth in fourier space
place_cells.Ctilde /= place_cells.Ctilde.max()
place_cells.Ctilde *= options.Ng
# run pattern formation
Gmaps = ortho_grid_pattern_formation(place_cells, options, activation='relu')
# plot resulting maps
fig = plot_ratemaps(np.flip(Gmaps), options.Ng)
if save_output:
save_figure(fig=fig, figdir=figdir, fname_base=fname_base, file_ext='.png')
save_figure(fig=fig, figdir=figdir, fname_base=fname_base, file_ext='.pdf')
PROP.plot_prop_kernels(n=1,
first_state=int(samps[ix_t]),
wrap_col=1,
autoprop_off=False)
axes[ix_tau][ix_t].set_title("")
fig.subplots_adjust(left=0.01,
bottom=0.1,
right=0.99,
top=0.9,
wspace=0.5,
hspace=0.6)
fig.set_size_inches(width, height)
if save_output:
save_figure(
fig=fig,
figdir=figdir,
fname_base=fname_base + "_TauxT_alpha%.1f" % alpha,
file_ext=".png",
)
save_figure(
fig=fig,
figdir=figdir,
fname_base=fname_base + "_TauxT_alpha%.1f" % alpha,
file_ext=".pdf",
)
# %% 3x3 figure
if run_3x3:
ENV = RoomWorld(start=0, scale=scale_lowres)
n_step = 30
start = ENV.n_state // 4 - 1
width = page_width
def plot_performance_metrics(
self,
metrics=["Q_error", "SR_error", "SPL_error", "policy_error"],
figdir=None):
"""
FUNCTION: Plots learning objectives on different panels as a function of number of sequences.
INPUTS: metrics = list of what metrics to plot from self.metrics_dict.keys()
"""
self.metrics_dict = {
"Q_error": "generator error",
"Q_corr": "generator correlation",
"T_error": "transition error",
"T_corr": "transition correlation",
"SR_error": "successor representation error",
"SR_corr": "successor representation correlation",
"R_error": "reward function error",
"SPL_error": "shortest path length error",
"policy_error": "expected cumulative reward loss",
"KLtraj": "trajectory KL-divergence",
}
metrics_df = pd.melt(
self.output_learn.reset_index(),
id_vars=["n_samp", "alpha", "tau", "lr"],
value_vars=self.metrics_dict.keys(),
var_name="metric",
value_name="value",
)
metrics_df.rename(columns={"n_samp": "#samples"}, inplace=True)
for metric_key in self.metrics_dict.keys():
if metric_key in metrics:
metrics_df.loc[metrics_df.metric == metric_key,
"metric"] = self.metrics_dict[metric_key]
else:
metrics_df = metrics_df[metrics_df.metric != metric_key]
self.metrics_learn = metrics_df
g = sb.FacetGrid(
self.metrics_learn,
row="lr",
col="metric",
hue="alpha",
margin_titles=True,
sharex=True,
sharey=False,
)
g.map(
sb.regplot,
"#samples",
"value",
ci=None,
fit_reg=False,
logx=False,
order=1,
truncate=False,
)
[plt.setp(ax.texts, text="") for ax in g.axes.flat]
g.set_titles(row_template="{row_name}", col_template="{col_name}")
if figdir is not None:
save_figure(fig=g,
figdir=figdir,
fname_base="learning_performance",
file_ext=".png")
label_panel(axes[2][1], label="J", x=x, y=y)
label_panel(axes[2][2], label="K", x=x, y=y)
label_panel(axes[2][3], label="L", x=x, y=y)
fig.subplots_adjust(left=0.01, bottom=0.1, right=0.99, top=0.9, wspace=0.6, hspace=0.8)
fig.set_size_inches(width, height)
if save_output:
fname_base = "FIGURE_7_SamplingOptim_tau%.1f_%.1f_alpha%.1f_%.1f_lr%.1f" % (
tau_diff,
tau_supdiff,
alpha_diff,
alpha_supdiff,
lr,
)
save_figure(fig=fig, figdir=figdir, fname_base=fname_base, file_ext=".png")
save_figure(fig=fig, figdir=figdir, fname_base=fname_base, file_ext=".pdf")
# %% SUPPLEMENTARY PLOT
width = page_width * 1.3
height = row_height * 1.6 * 2.5
nrows = 5
ncols = 3
fig, axes = plt.subplots(nrows=nrows, ncols=ncols, figsize=(width, height))
state_msize = 5
jitter_state = False
traj_width = 0.5
n_step = 10
rho_start = rho_init_exp
title_fontsize = 24
n_state = env.n_state
states = np.arange(n_state)
# sequence generation
exp_diff_g1.sample_sequences(n_samp=1, n_step=n_step)
# sequence plots
figsize = (12, 12)
exp_diff_g1.plot_trajectory(samp=0,
plot_env=True,
state_func_env=False,
figsize=figsize)
if save_output:
fname_base = 'FIGURE_4_sample_diff_g1'
save_figure(fig=exp_diff_g1.fig,
figdir=figdir,
fname_base=fname_base,
file_ext='.png')
save_figure(fig=exp_diff_g1.fig,
figdir=figdir,
fname_base=fname_base,
file_ext='.pdf')
exp_sdiff_g1.plot_trajectory(samp=0,
plot_env=True,
state_func_env=False,
figsize=figsize)
if save_output:
fname_base = 'FIGURE_4_sample_sdiff_g1'
save_figure(fig=exp_sdiff_g1.fig,
figdir=figdir,
fname_base=fname_base,
file_ext='.png')