In [ ]:
#Visualize Samples from the model
import sys,os,glob
from collections import OrderedDict
sys.path.append('../../')
import numpy as np
%matplotlib inline
import matplotlib.pyplot as plt
import matplotlib as mpl
mpl.rcParams['lines.linewidth']=5
mpl.rcParams['lines.markersize']=10
mpl.rcParams['text.usetex']=True
mpl.rcParams['text.latex.unicode']=True
mpl.rcParams['font.family'] = 'serif'
mpl.rcParams['font.serif'] = 'Times New Roman'
mpl.rcParams['text.latex.preamble']= ['\\usepackage{amsfonts}','\\usepackage{amsmath}']
mpl.rcParams['font.size'] = 40
mpl.rcParams['axes.labelsize']=40
mpl.rcParams['legend.fontsize']=40
#http://stackoverflow.com/questions/22408237/named-colors-in-matplotlib
from utils.misc import getConfigFile, readPickle, loadHDF5, getUniqueIDFromParams
from optvaeutils.viz import getName
subdirectories = ['none','finopt']
#DIR = '../../expt/chkpt-20newsgroups-';MAXEPOCH='200'
#DIR = '../../expt/chkpt-rcv2-';MAXEPOCH='200'
DIR = '../../expt/chkpt-wikicorp-';MAXEPOCH='20'
short_names = {}
if os.path.exists('../../optvaeutils/default-hmap.pkl'):
short_names = readPickle('../../optvaeutils/default-hmap.pkl')[0]
colors = {}
colors[0] = 'r'
colors[1] = 'b'
colors[2] = 'g'
colors[3] = 'k'
colors[4] = 'y'
colors[5] = 'k'
colors[6] = 'm'
colors[7] = 'c'
colors[8] = 'b'
markers = {}
markers[0]= '*'
markers[1]= '<'
markers[2]= '>'
markers[3]= '8'
markers[4]= 'p'
markers[5]= 'v'
markers[6]= '3'
markers[7]= '2'
markers[8]= '4'
#Evalaute POB
from datasets.load import loadDataset
from optvaedatasets.load import loadDataset as loadDataset_OVAE
dataset =DIR.split('chkpt-')[1][:-1]
print 'Dataset:', dataset
dset = loadDataset_OVAE(dataset)
NLL_train_prob,NLL_valid_prob = np.nan,np.nan
In [ ]:
#Visualize Training Curves (train/valid/test) across different
datalist, l_params, namelist, svallist = [],[],[],[]
ctr = 0
for sdir in subdirectories:
print sdir,DIR+sdir+'/*EP'+MAXEPOCH+'*.h5'
#print DIR+sdir+'/*EP200*.h5'
for f in glob.glob(DIR+sdir+'/*EP'+MAXEPOCH+'*.h5'):
print f,
if 'normalize' in f:# and 'none' in sdir:
continue
params = readPickle(getConfigFile(f))[0]
params['replicate_K'] = None
if params['anneal_rate']>100 or params['p_layers']<1:
continue
name = getName(params)
data = loadHDF5(f)
l_params.append(params)
datalist.append(data)
namelist.append(name)
if 'valid_bound_0' in data:
min_0 = np.min(data['valid_bound_0'][:,1])
min_f = np.min(data['valid_bound_f'][:,1])
amin = np.argmin(data['valid_bound_f'][:,1])
epmin = data['valid_bound_f'][amin,0]
elif 'valid_perp_0' in data:
min_0 = np.min(data['valid_perp_0'][:,1])
min_f = np.min(data['valid_perp_f'][:,1])
amin = np.argmin(data['valid_perp_f'][:,1])
epmin = data['valid_perp_f'][amin,0]
else:
print data.keys()
min_0 = np.min(data['valid_perp_bound_0'][:,1])
min_f = np.min(data['valid_perp_bound_f'][:,1])
amin = np.argmin(data['valid_perp_bound_f'][:,1])
epmin = data['valid_perp_bound_f'][amin,0]
print epmin,amin
Wfiles = np.load(f.split('-final')[0].split('-EP')[0]+'-EP'+str(int(epmin))+'-params.npz')
if 'p_0_W' in Wfiles:
svals = np.sort(np.linalg.svd(Wfiles['p_0_W'],compute_uv=False))
else:
svals = np.sort(np.linalg.svd(Wfiles['p_mean_W'],compute_uv=False))
svallist.append(svals)
print ctr,name,min_0,min_f
ctr+=1
p_names = getUniqueIDFromParams(l_params, short_names = short_names)
names = []
for a,b in zip(namelist,p_names):
names.append(a+b)
idxlist = []
print '\n'
#Restriction
for idx,name in enumerate(names):
print idx,name
idxlist.append(idx)
svallist = [svallist[k] for k in idxlist]
datalist = [datalist[k] for k in idxlist]
def update_name(name):
pl = str(int(name.split('q_layers-')[1].split('-')[0])+1)
if 'baseline' in name:
return pl+'-M1'
else:
if 'normalize' in name:
return pl+'-M'+name.split('fin')[1].split('-')[0]+'-norm'
else:
return pl+'-M'+name.split('fin')[1].split('-')[0]
#names = [names[k] for k in idxlist]
names = [update_name(names[k]) for k in idxlist]
results = {}
for idx,name in enumerate(names):
results[name] = datalist[idx]
print '\n Restricted Plots to: ',names
In [ ]:
#Visualize Training Curves (train/valid/test)
NS = 'M100'
if 'wikicorp' in dataset:
NS = 'M100'
colorFinal = OrderedDict()
colorFinal['1-M1'] = 'r'
colorFinal['1-'+NS] = 'b'
colorFinal['2-M1'] = 'g'
colorFinal['2-'+NS] = 'k'
colorFinal['3-M1'] = 'orange'
colorFinal['3-'+NS] = 'cyan'
markerFinal = {}
markerFinal['1-M1'] = 'o'
markerFinal['1-'+NS] = '>'
markerFinal['2-M1'] = '<'
markerFinal['2-'+NS] = 's'
markerFinal['3-M1'] = 'x'
markerFinal['3-'+NS] = 'v'
namemap = OrderedDict()
namemap['1-M1'] = '$q1$-$\\psi(x)$'
namemap['1-'+NS] = '$q1$-$\\psi^*$'
namemap['2-M1'] = '$q2$-$\\psi(x)$'
namemap['2-'+NS] = '$q2$-$\\psi^*$'
namemap['3-M1'] = '$q3$-$\\psi(x)$'
namemap['3-'+NS] = '$q3$-$\\psi^*$'
fig,axlist = plt.subplots(1,1,figsize=(10,8))
ax = axlist
for name in colorFinal:
if name not in results:
print 'Not found: ',name
continue
data = results[name]
print data.keys()
if 'valid_perp_bound_0' in data:
valid_bound_0 = data['valid_perp_bound_0']
valid_bound_f = data['valid_perp_bound_f']
elif 'valid_perp_0' in data:
valid_bound_0 = data['valid_perp_0']
valid_bound_f = data['valid_perp_f']
else:
valid_bound_0 = data['valid_bound_0']
valid_bound_f = data['valid_bound_f']
print name,np.min(valid_bound_f[:,1]),np.argmin(valid_bound_f[:,1])
MARKER = markerFinal[name]
COLOR = colorFinal[name]
X = valid_bound_f[:,0]
#ax.plot(valid_bound_0[:,0],valid_bound_0[:,1],'--',color=COLOR,marker = MARKER)
ax.plot(valid_bound_f[:,0],valid_bound_f[:,1],marker = MARKER,ms=30,color=COLOR,label=namemap[name])
ax.set_ylabel('Held-out [Perplexity]')
ax.set_xlabel('Epochs')
ax.hlines(NLL_valid_prob, 0, ax.get_xlim()[1], linestyles='dashdot',colors='k')
if 'rcv2' in dataset:
ax.set_ylim([300,700])
if 'wikicorp' in dataset:
ax.set_xticks(X)
ax.set_ylim([1150,1600])
if 'wikicorp' in dataset:
pass
ax.legend(loc='upper center', bbox_to_anchor=(0.4, 1.35),ncol=3, frameon=True,columnspacing=0.1, prop={'size': 35})
else:
ax.legend(loc='upper center', bbox_to_anchor=(.52, 1.05),ncol=2, frameon=False,columnspacing=0.1)
fname = 'bounds-'+dataset+'-qVary.pdf'
print fname,'saved'
plt.savefig(fname,bbox_inches='tight')
In [ ]:
#Visualize Training Curves (train/valid/test)
NS = 'M100'
if 'wikicorp' in dataset:
NS = 'M100'
fig,axlist = plt.subplots(1,1,figsize=(10,8))
ax = axlist
for name in colorFinal:
if name not in results:
print 'Not found:' , name
continue
data = results[name]
print data.keys()
train_bound_0 = data['train_perp_0']
train_bound_f = data['train_perp_f']
print name
MARKER = markerFinal[name]
COLOR = colorFinal[name]
if 'rcv2' in dataset:
X = train_bound_0[:,0][::20]
Y = train_bound_0[:,1][::20]
else:
X = train_bound_0[:,0][::5]
Y = train_bound_0[:,1][::5]
#ax.plot(X,Y,'--',color=COLOR,marker = MARKER)
if 'rcv2' in dataset:
X = train_bound_f[:,0]
Y = train_bound_f[:,1]
else:
X = train_bound_f[:,0]#[::5]
Y = train_bound_f[:,1]#[::5]
print X
ax.plot(X,Y,marker = MARKER,ms=30,color=COLOR,label=namemap[name])
ax.set_ylabel('Train [Perplexity]')
ax.set_xlabel('Epochs')
ax.hlines(NLL_valid_prob, 0, ax.get_xlim()[1], linestyles='dashdot',colors='k')
if 'rcv2' in dataset:
ax.set_ylim([300,700])
if 'wikicorp' in dataset:
ax.set_xticks(X)
ax.set_ylim([1100,1600])
if 'wikicorp' in dataset:
pass
ax.legend(loc='upper center', bbox_to_anchor=(0.4, 1.35),ncol=3, frameon=True,columnspacing=0.1, prop={'size': 35})
else:
ax.legend(loc='upper center', bbox_to_anchor=(.52, 1.05),ncol=2, frameon=False,columnspacing=0.1)
fname = 'train-bounds-'+dataset+'-qVary.pdf'
print fname,'saved'
plt.savefig(fname,bbox_inches='tight')
In [ ]:
#Plot Eigenspectrum of first W matrix
for idx,(svals,name) in enumerate(zip(svallist,namemap.keys())):
if 'res' in name:
continue
MARKER = markers[idx]
COLOR = colors[idx]
if name in markerFinal:
MARKER = markerFinal[name]
COLOR = colorFinal[name]
print name
plt.plot(np.arange(len(svals)),np.log(svals[::-1]),marker=MARKER,color=COLOR,ms=7,lw=1,label = name)
if params['dataset']=='wikicorp':
plt.legend(loc='upper center', bbox_to_anchor=(0.6, 0.75),ncol=2, frameon=True, prop={'size': 20},columnspacing=0.1)
else:
plt.legend(loc='upper center', bbox_to_anchor=(0.35, 1.25),ncol=4, frameon=True, prop={'size': 20},columnspacing=0.1)
plt.ylabel('Log\n Singular Values')
plt.xlabel('Dimensions')
plt.xticks([])
fname = 'logsingular-'+dataset+'.pdf'
print fname,'saved'
plt.savefig(fname,bbox_inches='tight')