misc plotting code for our NN discussion

activations.py

fig = plt.figure(figsize=(30,10))

# loss/acc
ax11 = plt.subplot(241)
ax11.plot(range(len(epoch_loss)),np.array(epoch_loss));
ax11.set_title('Epoch loss')
ax21 = plt.subplot(245)
ax21.plot(range(len(epoch_acc)),np.array(epoch_acc));
ax21.set_title('Epoch acc')

# space embeddings
ax12 = plt.subplot(242,projection='3d')
ax13 = plt.subplot(243,projection='3d')
ax14 = plt.subplot(244)

# H1
ax12.plot(
    np.ravel(H1_blue[:,0]), np.ravel(H1_blue[:,1]), np.ravel(H1_blue[:,2]),
    'bo', label='class: blue circle', alpha=0.5)
ax12.plot(
    np.ravel(H1_red[:,0]), np.ravel(H1_red[:,1]), np.ravel(H1_red[:,2]), 
    'r*', label='class: red star', alpha=0.5)
ax12.set_xlabel('$h_1$', fontsize=20);ax12.set_ylabel('$h_2$', fontsize=20);ax12.set_zlabel('$h_3$', fontsize=20);
# ax12.view_init(elev=30, azim=-200)
ax12.set_title('H1 space');

# H2
ax13.plot(
    np.ravel(H2_blue[:,0]), np.ravel(H2_blue[:,1]), np.ravel(H2_blue[:,2]),
    'bo', label='class: blue circle', alpha=0.5)
ax13.plot(
    np.ravel(H2_red[:,0]), np.ravel(H2_red[:,1]), np.ravel(H2_red[:,2]), 
    'r*', label='class: red star', alpha=0.5)
ax13.set_xlabel('$h_1$', fontsize=20);ax13.set_ylabel('$h_2$', fontsize=20);ax13.set_zlabel('$h_3$', fontsize=20);
# ax13.view_init(elev=30, azim=-200)
ax13.set_title('H2 space');

# output layer
ax14.scatter(output_red,np.zeros_like(output_red),color='red');
ax14.scatter(output_blue,np.zeros_like(output_blue),color='blue');
ax14.axes.yaxis.set_visible(False);
ax14.set_title('Output space');

# activations
ax22 = plt.subplot(246)
sharey = None;
# sharey = ax22
ax23 = plt.subplot(247,sharey=sharey)
ax24 = plt.subplot(248,sharey=sharey)

ax22.plot((np.array(act_info[-3]).T)[0,:],label='Node 1');
ax22.plot((np.array(act_info[-3]).T)[1,:],label='Node 2');
ax22.plot((np.array(act_info[-3]).T)[2,:],label='Node 3');
ax22.legend();
ax22.set_title('H1 Activations');

ax23.plot((np.array(act_info[-2]).T)[0,:],label='Node 1');
ax23.plot((np.array(act_info[-2]).T)[1,:],label='Node 2');
ax23.plot((np.array(act_info[-2]).T)[2,:],label='Node 3');
ax23.legend();
ax23.set_title('H2 Activations');

ax24.plot(np.array(np.squeeze(act_info[-1])),color='red');
ax24.set_ylim([0,1]);
ax24.set_title('Output Activations');

animation.py

# red and blue activations lists generated during training
# must modify `.SGD` to generate these (not shown here)
act_red = rb_info[0]
act_blue = rb_info[1]

fig = plt.figure(figsize=(30,7))

ax1 = plt.subplot(141)
ax1.set_xlim([0, len(act_red)])
epoch_loss_min = np.array(epoch_loss).min()
epoch_loss_max = np.array(epoch_loss).max()
ax1.set_ylim([epoch_loss_min*0.999, epoch_loss_max*1.001])
ax1t = ax1.twinx()
ax1t.set_ylim([0,1.1])
ax2 = plt.subplot(142,projection='3d')
ax3 = plt.subplot(143,projection='3d')
ax4 = plt.subplot(144)

# unpack data:
curr_act_red = act_red[-1]
H1_red = np.array(curr_act_red[-3].T)
H2_red = np.array(curr_act_red[-2].T)
output_red = np.array(curr_act_red[-1])

curr_act_blue = act_blue[-1]
H1_blue = np.array(curr_act_blue[-3].T)
H2_blue = np.array(curr_act_blue[-2].T)
output_blue = np.array(curr_act_blue[-1])

# stack data for plotting
H1 = np.vstack((H1_blue,H1_red))
H2 = np.vstack((H2_blue,H2_red))
output = np.squeeze(np.hstack((output_blue,output_red)))
join_labels = np.array([0]*500 + [1]*500)

# loss/acc
e_loss, = ax1.plot(epoch_loss,color='red')
ax1.set_ylabel('loss',color='red')
ax1.tick_params(axis='y', labelcolor='red')
ax1.set_title('Loss vs Acc')

e_acc, = ax1t.plot(epoch_acc,color='green')
ax1t.set_ylabel('acc',color='green')
ax1t.tick_params(axis='y', labelcolor='green')

# H1
ax2.set_xlabel('$h_1$', fontsize=20);
ax2.set_ylabel('$h_2$', fontsize=20);
ax2.set_zlabel('$h_3$', fontsize=20);
ax2.set_title('H1 space');
h1_scat= ax2.scatter(H1[:,0],H1[:,1],H1[:,2], alpha=0.5,cmap='seismic',c=join_labels)

# H2
ax3.set_xlabel('$h_1$', fontsize=20);
ax3.set_ylabel('$h_2$', fontsize=20);
ax3.set_zlabel('$h_3$', fontsize=20);
ax3.set_title('H2 space');
h2_scat = ax3.scatter(H2[:,0],H2[:,1],H2[:,2], alpha=0.5, cmap='seismic',c=join_labels)

# output
ax4.axes.yaxis.set_visible(False);
ax4.set_title('Output space');
output_scat = ax4.scatter(output,np.zeros_like(output),cmap='seismic',c=join_labels)

# animation function
def update_graph(i):
    
    curr_act_red = act_red[i]
    H1_red = np.array(curr_act_red[-3].T)
    H2_red = np.array(curr_act_red[-2].T)
    output_red = np.array(curr_act_red[-1])
    
    curr_act_blue = act_blue[i]
    H1_blue = np.array(curr_act_blue[-3].T)
    H2_blue = np.array(curr_act_blue[-2].T)
    output_blue = np.array(curr_act_blue[-1])
    
    # stack data for plotting
    H1 = np.vstack((H1_blue,H1_red))
    H2 = np.vstack((H2_blue,H2_red))
    output_c = np.squeeze(np.hstack((output_blue,output_red)))
    
    # loss/acc
    e_loss.set_data(range(i),epoch_loss[0:i*50:50])
    e_acc.set_data(range(i),epoch_acc[0:i*50:50])
    
    # h1 space
    h1_scat._offsets3d = (H1[:,0],H1[:,1],H1[:,2])
    h1_scat.set_array(np.array([0]*500 + [1]*500))
    ax2.view_init(elev=30, azim=i*2)
    
    # h2 space
    h2_scat._offsets3d = (H2[:,0],H2[:,1],H2[:,2])
    h2_scat.set_array(np.array([0]*500 + [1]*500))
    ax3.view_init(elev=30, azim=i*2)
    
    # output space
    output_scat.set_offsets(np.vstack((output_c,np.zeros_like(output_c))).T)
    output_scat.set_array(np.array([0]*500 + [1]*500))
    
# create animation
anim = matplotlib.animation.FuncAnimation(fig, update_graph,
                                         frames=range(len(act_red)),
                                         interval=40, blit=False)
# show in notebook                           
from IPython.display import HTML
HTML(anim.to_html5_video())

# save to file
#anim.save('slow_training.gif')