def generate_plots(self, input_data, input_labels=None):
"""
Plot weights, reconstruction, and gradients
Inputs:
input_data: data object containing the current image batch
input_labels: data object containing the current label batch
"""
super(ICA, self).generate_plots(input_data, input_labels)
feed_dict = self.get_feed_dict(input_data, input_labels)
weights = tf.get_default_session().run(self.a, feed_dict)
current_step = str(self.global_step.eval())
#pf.plot_data_tiled(input_data.reshape((self.batch_size,
# np.int(np.sqrt(self.num_pixels)),
# np.int(np.sqrt(self.num_pixels)))),
# normalize=False, title="Images at step "+current_step,
# vmin=np.min(input_data), vmax=np.max(input_data),
# save_filename=(self.disp_dir+"images_"+current_step.zfill(5)+".pdf"))
pf.plot_data_tiled(
weights.reshape(self.num_neurons, int(np.sqrt(self.num_pixels)),
int(np.sqrt(self.num_pixels))),
normalize=True,
title="Dictionary at step " + current_step,
vmin=-1.0,
vmax=1.0,
save_filename=(self.disp_dir + "a_v" + self.version + "-" +
current_step.zfill(5) + ".pdf"))
pf.plot_activity_hist(
self.z.eval(feed_dict),
num_bins=1000,
title="z Activity Histogram at step " + current_step,
save_filename=(self.disp_dir + "z_hist_v" + self.version + "-" +
current_step.zfill(5) + ".pdf"))
pf.plot_activity_hist(
self.u.eval(feed_dict),
num_bins=1000,
title="u Activity Histogram at step " + current_step,
save_filename=(self.disp_dir + "u_hist_v" + self.version + "-" +
current_step.zfill(5) + ".pdf"))
pf.plot_bar(np.linalg.norm(weights, axis=1, keepdims=False),
num_xticks=5,
title="a l2 norm",
xlabel="Basis Index",
ylabel="L2 Norm",
save_filename=(self.disp_dir + "a_norm_v" + self.version +
"-" + current_step.zfill(5) + ".pdf"))
for weight_grad_var in self.grads_and_vars[self.sched_idx]:
grad = weight_grad_var[0][0].eval(feed_dict)
shape = grad.shape
name = weight_grad_var[0][1].name.split('/')[1].split(':')[
0] #np.split
pf.plot_data_tiled(
grad.reshape(self.num_neurons, int(np.sqrt(self.num_pixels)),
int(np.sqrt(self.num_pixels))),
normalize=False,
title="Gradient for " + name + " at step " + current_step,
vmin=None,
vmax=None,
save_filename=(self.disp_dir + "d" + name + "_v" +
self.version + "_" + current_step.zfill(5) +
".pdf"))
def save_data(self, data, datatype):
fig_title = "Activity triggered averages on " + datatype + " data"
for layer_idx, layer in enumerate(data):
ata_filename = (self.analysis_out_dir + "act_trig_avg_layer_" +
str(layer_idx) + "_" + datatype + "_v" +
self.version + self.file_ext)
ata = layer.reshape(layer.shape[0], int(np.sqrt(layer.shape[1])),
int(np.sqrt(layer.shape[1])))
pf.plot_data_tiled(ata,
normalize=True,
title=fig_title,
vmin=-1.0,
vmax=1.0,
save_filename=ata_filename)
def generate_plots(self, input_data, input_labels=None):
"""
Plot weights, reconstruction, and gradients
Inputs: input_data and input_labels used for the session
"""
super(MLP, self).generate_plots(input_data, input_labels)
feed_dict = self.get_feed_dict(input_data, input_labels)
current_step = str(self.global_step.eval())
pf.plot_data_tiled(
self.w2.eval().T.reshape(self.num_classes,
int(np.sqrt(self.num_hidden)), int(np.sqrt(self.num_hidden))),
normalize=True, title="Classification matrix at step number "+current_step,
vmin=None, vmax=None, save_filename=(self.disp_dir+"w2_v"+self.version+"-"
+current_step.zfill(5)+".pdf"))
pf.plot_data_tiled(
self.w1.eval().T.reshape(self.num_hidden,
int(np.sqrt(self.num_pixels)), int(np.sqrt(self.num_pixels))),
normalize=True, title="Dictionary at step "+current_step, vmin=None, vmax=None,
save_filename=(self.disp_dir+"w1_v"+self.version+"-"+current_step.zfill(5)+".pdf"))
for weight_grad_var in self.grads_and_vars[self.sched_idx]:
grad = weight_grad_var[0][0].eval(feed_dict)
shape = grad.shape
name = weight_grad_var[0][1].name.split('/')[1].split(':')[0]#np.split
if name == "w1":
pf.plot_data_tiled(grad.T.reshape(self.num_hidden,
int(np.sqrt(self.num_pixels)), int(np.sqrt(self.num_pixels))),
normalize=True, title="Gradient for w1 at step "+current_step, vmin=None, vmax=None,
save_filename=(self.disp_dir+"dw1_v"+self.version+"_"+current_step.zfill(5)+".pdf"))
elif name == "w2":
pf.plot_data_tiled(grad.T.reshape(self.num_classes,
int(np.sqrt(self.num_hidden)), int(np.sqrt(self.num_hidden))),
normalize=True, title="Gradient for w2 at step "+current_step, vmin=None, vmax=None,
save_filename=(self.disp_dir+"dw2_v"+self.version+"_"+current_step.zfill(5)+".pdf"))
def generate_plots(self, input_data, input_labels=None):
"""
Plot weights, reconstruction, and gradients
Inputs:
input_data: data object containing the current image batch
input_labels: data object containing the current label batch
"""
feed_dict = self.get_feed_dict(input_data, input_labels)
current_step = str(self.global_step.eval())
recon = tf.get_default_session().run(self.x_, feed_dict)
weights = tf.get_default_session().run(self.phi, feed_dict)
pf.plot_data_tiled(
weights.T.reshape(self.num_neurons, int(np.sqrt(self.num_pixels)),
int(np.sqrt(self.num_pixels))),
normalize=False,
title="Dictionary at step " + current_step,
vmin=None,
vmax=None,
save_filename=(self.disp_dir + "phi_v" + self.version + "_" +
current_step.zfill(5) + ".png"))
pf.plot_data_tiled(
input_data.reshape(
(self.batch_size, np.int(np.sqrt(self.num_pixels)),
np.int(np.sqrt(self.num_pixels)))),
normalize=False,
title="Images at step " + current_step,
vmin=None,
vmax=None,
save_filename=(self.disp_dir + "images_" + self.version + "-" +
current_step.zfill(5) + ".png"))
pf.plot_data_tiled(
recon.reshape((self.batch_size, np.int(np.sqrt(self.num_pixels)),
np.int(np.sqrt(self.num_pixels)))),
normalize=False,
title="Recons at step " + current_step,
vmin=None,
vmax=None,
save_filename=(self.disp_dir + "recons_v" + self.version + "-" +
current_step.zfill(5) + ".png"))
for weight_grad_var in self.grads_and_vars[self.sched_idx]:
grad = weight_grad_var[0][0].eval(feed_dict)
shape = grad.shape
name = weight_grad_var[0][1].name.split('/')[1].split(':')[
0] #np.split
pf.plot_data_tiled(
grad.T.reshape(self.num_neurons, int(np.sqrt(self.num_pixels)),
int(np.sqrt(self.num_pixels))),
normalize=True,
title="Gradient for phi at step " + current_step,
vmin=None,
vmax=None,
save_filename=(self.disp_dir + "dphi_v" + self.version + "_" +
current_step.zfill(5) + ".pdf"))
def generate_plots(self, input_data, input_labels=None):
"""
Plot weights, reconstruction, and gradients
Inputs:
input_data: data object containing the current image batch
input_labels: data object containing the current label batch
"""
super(CONV_LCA, self).generate_plots(input_data, input_labels)
feed_dict = self.get_feed_dict(input_data, input_labels)
current_step = str(self.global_step.eval())
pf.plot_data_tiled(
self.phi.eval(), normalize=True,
title="Dictionary at step "+current_step, vmin=None, vmax=None,
save_filename=(self.disp_dir+"phi_v"+self.version+"_"+current_step.zfill(5)+".pdf"))
for weight_grad_var in self.grads_and_vars[self.sched_idx]:
grad = weight_grad_var[0][0].eval(feed_dict)
shape = grad.shape
name = weight_grad_var[0][1].name.split('/')[1].split(':')[0]#np.split
pf.plot_data_tiled(grad, normalize=True,
title="Gradient for phi at step "+current_step, vmin=None, vmax=None,
save_filename=(self.disp_dir+"dphi_v"+self.version+"_"+current_step.zfill(5)+".pdf"))
def generate_plots(self, input_data, input_labels=None):
"""
Plot weights, reconstruction, and gradients
Inputs:
input_data: data object containing the current image batch
input_labels: data object containing the current label batch
"""
super(density_learner, self).generate_plots(input_data, input_labels)
feed_dict = self.get_feed_dict(input_data, input_labels)
current_step = str(self.global_step.eval())
#recon = tf.get_default_session().run(self.x_, feed_dict)
#weights = tf.get_default_session().run(self.phi, feed_dict)
b_weights = tf.get_default_session().run(self.b, feed_dict)
v_vals = tf.get_default_session().run(self.v, feed_dict)
#pf.plot_data_tiled(input_data.reshape((self.batch_size,
# np.int(np.sqrt(self.num_pixels)), np.int(np.sqrt(self.num_pixels)))),
# normalize=False, title="Images at step "+current_step, vmin=None, vmax=None,
# save_filename=(self.disp_dir+"images_"+self.version+"-"+current_step.zfill(5)+".pdf"))
#pf.plot_data_tiled(weights.T.reshape(self.num_neurons,
# int(np.sqrt(self.num_pixels)), int(np.sqrt(self.num_pixels))),
# normalize=False, title="Dictionary at step "+current_step, vmin=None, vmax=None,
# save_filename=(self.disp_dir+"phi_v"+self.version+"_"+current_step.zfill(5)+".pdf"))
pf.plot_data_tiled(
b_weights.T.reshape(self.num_v, int(np.sqrt(self.num_neurons)),
int(np.sqrt(self.num_neurons))),
normalize=True,
title="Density weights at step " + current_step,
vmin=None,
vmax=None,
save_filename=(self.disp_dir + "b_v" + self.version + "_" +
current_step.zfill(5) + ".pdf"))
#pf.plot_bar(np.linalg.norm(weights, axis=1, keepdims=False), num_xticks=5,
# title="phi l2 norm", xlabel="Basis Index", ylabel="L2 Norm",
# save_filename=(self.disp_dir+"phi_norm_v"+self.version+"-"+current_step.zfill(5)+".pdf"))
pf.plot_bar(np.linalg.norm(b_weights, axis=1, ord=1, keepdims=False),
num_xticks=5,
title="b l1 norm",
xlabel="Basis Index",
ylabel="L1 Norm",
save_filename=(self.disp_dir + "b_norm_v" + self.version +
"-" + current_step.zfill(5) + ".pdf"))
pf.plot_activity_hist(
v_vals,
num_bins=1000,
title="v Activity Histogram at step " + current_step,
save_filename=(self.disp_dir + "v_hist_v" + self.version + "-" +
current_step.zfill(5) + ".pdf"))
#pf.plot_data_tiled(recon.reshape((self.batch_size,
# np.int(np.sqrt(self.num_pixels)), np.int(np.sqrt(self.num_pixels)))),
# normalize=False, title="Recons at step "+current_step, vmin=None, vmax=None,
# save_filename=(self.disp_dir+"recons_v"+self.version+"-"+current_step.zfill(5)+".pdf"))
for weight_grad_var in self.grads_and_vars[self.sched_idx]:
grad = weight_grad_var[0][0].eval(feed_dict)
shape = grad.shape
name = weight_grad_var[0][1].name.split('/')[1].split(':')[
0] #np.split
if name == "phi":
pf.plot_data_tiled(
grad.T.reshape(self.num_neurons,
int(np.sqrt(self.num_pixels)),
int(np.sqrt(self.num_pixels))),
normalize=True,
title="Gradient for phi at step " + current_step,
vmin=None,
vmax=None,
save_filename=(self.disp_dir + "dphi_v" + self.version +
"_" + current_step.zfill(5) + ".pdf"))
elif name == "b":
pf.plot_data_tiled(
grad.T.reshape(self.num_v, int(np.sqrt(self.num_neurons)),
int(np.sqrt(self.num_neurons))),
normalize=True,
title="Gradient for b at step " + current_step,
vmin=None,
vmax=None,
save_filename=(self.disp_dir + "db_v" + self.version +
"_" + current_step.zfill(5) + ".pdf"))