from hessian_analysis_tools import plot_spectra, compute_hess_corr, plot_consistency_example, plot_consistentcy_mat, \
plot_consistency_hist, average_H, scan_hess_npz
eva_col, evc_col, feat_col, meta = scan_hess_npz(savedir,
"Hessian_cmp_(\d*).npz",
featkey="feat")
feat_col = np.array(feat_col).squeeze()
H_avg, eva_avg, evc_avg = average_H(eva_col, evc_col)
np.savez(join(figdir, "H_avg_%s.npz" % "DCGAN"),
H_avg=H_avg,
eva_avg=eva_avg,
evc_avg=evc_avg,
feats=feat_col)
#%%
fig = plot_spectra(
eva_col,
figdir=figdir,
titstr="DCGAN",
)
#%%
corr_mat_log, corr_mat_lin = compute_hess_corr(eva_col,
evc_col,
figdir=figdir,
use_cuda=False)
# without cuda 2:12 mins, with cuda 6:55
#%%
fig1, fig2 = plot_consistentcy_mat(corr_mat_log,
corr_mat_lin,
figdir=figdir,
titstr="DCGAN")
fig11, fig22 = plot_consistency_hist(corr_mat_log,
corr_mat_lin,
label="all space")
plt.ylabel("eigenvalue(log)")
plt.xlabel("eig id")
plt.legend()
st = plt.suptitle(
"Hessian Spectrum of StyleGAN2\n (error bar for [5,95] percentile among all samples)"
)
plt.savefig(join(figdir, savename),
bbox_extra_artists=[st]) # this is working.
plt.show()
return fig
#%%
plot_spectra(
eigval_col,
savename="Cat-f_spectra.jpg",
)
#%% Go through all the models see how their thing is doing
subpath = [f.path for f in os.scandir(SGdir) if f.is_dir()]
subfdnm = [f.name for f in os.scandir(SGdir) if f.is_dir()]
failnms = []
for fdnm in subfdnm:
try:
eigval_col, _, meta = scan_hess_npz(join(SGdir, fdnm))
eigval_col = eigval_col[::-1]
plot_spectra(eigval_col, savename="%s-spectra.jpg" % fdnm)
np.savez(
join(figdir, "spectra_col_%s.npz" % fdnm),
eigval_col=eigval_col,
)
print(fdnm, "folder finished")
# Load the Hessian NPZ
eva_ctrl, evc_ctrl, feat_ctrl, meta = scan_hess_npz(savedir,
"eig_full_trial(\d*).npz",
evakey='eva',
evckey='evc',
featkey="feat")
# compute the Mean Hessian and save
H_avg, eva_avg, evc_avg = average_H(eva_ctrl, evc_ctrl)
np.savez(join(figdir, "H_avg_%s.npz" % modelnm),
H_avg=H_avg,
eva_avg=eva_avg,
evc_avg=evc_avg,
feats=feat_ctrl)
# compute and plot spectra
fig0 = plot_spectra(eigval_col=eva_ctrl,
savename="%s_spectrum" % modelnm,
figdir=figdir)
np.savez(
join(figdir, "spectra_col_%s.npz" % modelnm),
eigval_col=eva_ctrl,
)
# compute and plot the correlation between hessian at different points
corr_mat_log_ctrl, corr_mat_lin_ctrl = compute_hess_corr(eva_ctrl,
evc_ctrl,
figdir=figdir,
use_cuda=True,
savelabel=modelnm)
fig1, fig2 = plot_consistentcy_mat(corr_mat_log_ctrl,
corr_mat_lin_ctrl,
figdir=figdir,
titstr="%s" % modelnm,
# Load the Hessian NPZ
eva_ctrl, evc_ctrl, feat_ctrl, meta = scan_hess_npz(savedir,
"Hess_BP_(\d*).npz",
evakey='eva_BP',
evckey='evc_BP',
featkey="feat")
# compute the Mean Hessian and save
H_avg, eva_avg, evc_avg = average_H(eva_ctrl, evc_ctrl)
np.savez(join(figdir, "H_avg_%s.npz" % label),
H_avg=H_avg,
eva_avg=eva_avg,
evc_avg=evc_avg,
feats=feat_ctrl)
# compute and plot spectra
fig0 = plot_spectra(eigval_col=eva_ctrl,
savename="%s_spectrum" % label,
figdir=figdir)
np.savez(
join(figdir, "spectra_col_%s.npz" % label),
eigval_col=eva_ctrl,
)
# compute and plot the correlation between hessian at different points
corr_mat_log_ctrl, corr_mat_lin_ctrl = compute_hess_corr(eva_ctrl,
evc_ctrl,
figdir=figdir,
use_cuda=True,
savelabel=label)
fig1, fig2 = plot_consistentcy_mat(corr_mat_log_ctrl,
corr_mat_lin_ctrl,
figdir=figdir,
titstr="%s" % label,
preprocess=lambda img:img)
#preprocess=lambda img: F.interpolate(img, (256, 256), mode='bilinear', align_corners=True))
print("%.2f sec" % (time() - T0)) # 109 sec
np.savez(join(savedir, "Hess_BP_%03d.npz"%triali), eva_BP=eva_BP, evc_BP=evc_BP, H_BP=H_BP,
feat=feat.detach().cpu().numpy(), feat_z=feat_z.detach().cpu().numpy())
#%%
#%
figdir = r"E:\OneDrive - Washington University in St. Louis\Hessian_summary\StyleGAN2_wspace"
os.makedirs(figdir, exist_ok=True)
# Load the Hessian NPZ
eva_col, evc_col, feat_col, meta = scan_hess_npz(savedir, "Hess_BP_(\d*).npz", featkey="feat")
# compute the Mean Hessian and save
H_avg, eva_avg, evc_avg = average_H(eva_col, evc_col, )
np.savez(join(figdir, "H_avg_%s.npz"%modelnm), H_avg=H_avg, eva_avg=eva_avg, evc_avg=evc_avg, feats=feat_col)
# compute and plot spectra
fig0 = plot_spectra(eigval_col=eva_col, savename="%s_spectrum"%modelnm, figdir=figdir)
np.savez(join(figdir, "spectra_col_%s.npz"%modelnm), eigval_col=eva_col, )
# compute and plot the correlation between hessian at different points
corr_mat_log, corr_mat_lin = compute_hess_corr(eva_col, evc_col, figdir=figdir, use_cuda=False, savelabel=modelnm)
corr_mat_vec = compute_vector_hess_corr(eva_col, evc_col, figdir=figdir, use_cuda=False, savelabel=modelnm)
fig1, fig2 = plot_consistentcy_mat(corr_mat_log, corr_mat_lin, figdir=figdir, titstr="%s"%modelnm, savelabel=modelnm)
fig11, fig22 = plot_consistency_hist(corr_mat_log, corr_mat_lin, figdir=figdir, titstr="%s"%modelnm,
savelabel=modelnm)
fig3 = plot_consistency_example(eva_col, evc_col, figdir=figdir, nsamp=5, titstr="%s"%modelnm, savelabel=modelnm)
fig3 = plot_consistency_example(eva_col, evc_col, figdir=figdir, nsamp=3, titstr="%s"%modelnm, savelabel=modelnm)
#%% "stylegan2-cat-config-f"
modelname = "stylegan2-cat-config-f"
modelnm = "stylegan2-cat-config-f"+"_wspace"
modelsnm = "Cat256"
SGAN = loadStyleGAN2(modelname+".pt", size=256, channel_multiplier=2)
plt.ylabel("eigenvalue")
plt.xlabel("eig id")
plt.legend()
plt.sca(axs[1])
plt.plot(range(eigN), np.log10(eigmean), alpha=0.6)
plt.fill_between(range(eigN), np.log10(eiglim[0, :]), np.log10(eiglim[1, :]), alpha=0.3, label=label)
plt.ylabel("eigenvalue(log)")
plt.xlabel("eig id")
plt.legend()
st = plt.suptitle("Hessian Spectrum of StyleGAN\n (error bar for [5,95] percentile among all samples)")
plt.savefig(join(figdir, savename), bbox_extra_artists=[st]) # this is working.
# fig.show()
return fig
fig = plot_spectra(data_BP["eva_BP"][np.newaxis, :], label="BP", savename="spectrum_onetrial.jpg")
fig = plot_spectra(data_FI["H_col"][4, 0][np.newaxis, :], savename="spectrum_method_cmp.jpg", label="ForwardIter 1E-3", fig=fig)
fig = plot_spectra(data_FI["H_col"][5, 0][np.newaxis, :], savename="spectrum_method_cmp.jpg", label="ForwardIter 3E-3", fig=fig)
fig = plot_spectra(data_FI["H_col"][6, 0][np.newaxis, :], savename="spectrum_method_cmp.jpg", label="ForwardIter 1E-2", fig=fig)
plt.show()
#%%
"""
This is the smaller explicit version of StyleGAN. Very easy to work with
"""
#%%
sys.path.append("E:\Github_Projects\style-based-gan-pytorch")
sys.path.append("D:\Github\style-based-gan-pytorch")
from model import StyledGenerator
from generate import get_mean_style
import math
#%%
# feat=feat.cpu().detach().numpy())
#%%
from hessian_analysis_tools import scan_hess_npz, average_H, compute_vector_hess_corr, plot_consistentcy_mat, \
plot_consistency_hist, plot_consistency_example, plot_spectra
savedir = r"E:\OneDrive - Washington University in St. Louis\HessNetArchit\BigBiGAN\rand_Hessians"
figdir = r"E:\OneDrive - Washington University in St. Louis\HessNetArchit\BigBiGAN"
realfigdir = r"E:\OneDrive - Washington University in St. Louis\Hessian_summary\BigBiGAN"
modelnm = "BigBiGAN_randinit"
# Load the Hessian NPZ
eva_ctrl, evc_ctrl, feat_ctrl, meta = scan_hess_npz(savedir, "eig_full_trial(\d*).npz", evakey='eva', evckey='evc', featkey="feat")
# compute the Mean Hessian and save
H_avg, eva_avg, evc_avg = average_H(eva_ctrl, evc_ctrl)
np.savez(join(figdir, "H_avg_%s.npz"%modelnm), H_avg=H_avg, eva_avg=eva_avg, evc_avg=evc_avg, feats=feat_ctrl)
# compute and plot spectra
fig0 = plot_spectra(eigval_col=eva_ctrl, savename="%s_spectrum"%modelnm, figdir=figdir)
fig0 = plot_spectra(eigval_col=eva_ctrl, savename="%s_spectrum_med"%modelnm, figdir=figdir, median=True)
np.savez(join(figdir, "spectra_col_%s.npz"%modelnm), eigval_col=eva_ctrl, )
# compute and plot the correlation between hessian at different points
"""Note the spectra of randinitd BigBiGAN doesn't make sense we use randomly initialized one instead"""
#%%
corr_mat_log_ctrl, corr_mat_lin_ctrl = compute_hess_corr(eva_ctrl+1E-7, evc_ctrl, figdir=figdir, use_cuda=True,
savelabel=modelnm+"shift1E-7")
corr_mat_vec_ctrl = compute_vector_hess_corr(eva_ctrl+1E-7, evc_ctrl, figdir=figdir, use_cuda=True,
savelabel=modelnm+"shift1E-7")
fig1, fig2 = plot_consistentcy_mat(corr_mat_log_ctrl, corr_mat_lin_ctrl, figdir=figdir, titstr="%s"%modelnm+"shift1E-7",
savelabel=modelnm)
fig11, fig22 = plot_consistency_hist(corr_mat_log_ctrl, corr_mat_lin_ctrl, figdir=figdir, titstr="%s"%modelnm+"shift1E-7",
savelabel=modelnm)
fig3 = plot_consistency_example(eva_ctrl+1E-7, evc_ctrl, figdir=figdir, nsamp=5, titstr="%s"%modelnm+"shift1E-7", savelabel=modelnm)
#%%
Hessdir = join(rootdir, 'StyleGAN2')
modelnm = "stylegan2-ffhq-config-f"
modelsnm = "Face1024"
# SGAN = loadStyleGAN2(modelnm+".pt", size=1024,)
# SG = StyleGAN2_wrapper(SGAN, )
eigval_col, eigvec_col, feat_col, meta = scan_hess_npz(join(dataroot, modelnm),
"Hess_BP_(\d*).npz",
featkey="feat")
feat_col = np.array(feat_col).squeeze()
H_avg, eva_avg, evc_avg = average_H(eigval_col, eigvec_col)
np.savez(join(Hessdir, "H_avg_%s.npz" % modelnm),
H_avg=H_avg,
eva_avg=eva_avg,
evc_avg=evc_avg,
feats=feat_col)
fig0 = plot_spectra(eigval_col=eigval_col, savename="%s_spectrum" % modelnm)
corr_mat_log, corr_mat_lin = compute_hess_corr(
eigval_col,
eigvec_col,
figdir=figdir,
use_cuda=False,
savelabel=modelnm,
)
fig1, fig2 = plot_consistentcy_mat(corr_mat_log,
corr_mat_lin,
figdir=figdir,
titstr="StyleGAN2 %s" % modelnm,
savelabel=modelnm)
fig3 = plot_consistency_example(eigval_col,
eigvec_col,
figdir=figdir,
label = modelnm + "_fix_ctrl"
eva_col_ctrl = np.load(
join(saveroot, "summary", label,
"spectra_col_%s.npz" % label))["eigval_col"]
label = modelnm + "_W_fix"
eva_col_W = np.load(
join(saveroot, "summary", label,
"spectra_col_%s.npz" % label))["eigval_col"]
label = modelnm + "_W_fix_ctrl"
eva_col_W_ctrl = np.load(
join(saveroot, "summary", label,
"spectra_col_%s.npz" % label))["eigval_col"]
fig0 = plot_spectra(eva_col,
label="Fix",
titstr="StyleGAN2 %s" % modelsnm,
savename="SG2_%s_shfl_spectrum_cmp" % modelsnm,
figdir=summaryroot)
fig0 = plot_spectra(eva_col_ctrl,
label="Fix_Shuffled",
titstr="StyleGAN2 %s" % modelsnm,
fig=fig0,
savename="SG2_%s_shfl_spectrum_cmp_Zspace" % modelsnm,
figdir=summaryroot)
fig0 = plot_spectra(eva_col_W,
label="Fix Wspace",
titstr="StyleGAN2 %s" % modelsnm,
fig=fig0,
savename="SG2_%s_shfl_spectrum_cmp" % modelsnm,
figdir=summaryroot)
plt.plot(np.log10(eigval_arr).T, color="pink", lw=1.5, alpha=0.5)
if control:
plt.plot(range(len(eigmean)), np.log10(eigmean), alpha=0.7, color="green") # , eigval_arr.std(axis=0)
plt.fill_between(range(len(eigmean)), np.log10(eiglim[0, :]), np.log10(eiglim[1, :]), alpha=0.5,
color="purple", label="control")
plt.ylabel("eigenvalue(log)")
plt.xlabel("eig id")
plt.xlim([-10, 130])
plt.legend()
st = plt.suptitle("Hessian Spectrum of BigBiGAN in Different Spaces\n (error bar for [5,95] percentile)")
plt.savefig(join(figdir, savename), bbox_extra_artists=[st]) # this is working.
# plt.show()
figdir = r"E:\OneDrive - Washington University in St. Louis\Hessian_summary\BigBiGAN"
plot_spectra(control=False, savename="spectrum_stat_120_org.jpg", )
plot_spectra(indiv_trace=True, savename="spectrum_stat_120_traces.jpg", )
# plot_spectra(control=True, savename="spectrum_stat_cmp.jpg", )
#%%
def corr_torch(V1, V2):
C1 = (V1 - V1.mean())
C2 = (V2 - V2.mean())
return torch.dot(C1, C2) / C1.norm() / C2.norm()
def corr_nan_torch(V1, V2):
Msk = torch.isnan(V1) | torch.isnan(V2)
return corr_torch(V1[~Msk], V2[~Msk])
#%%
evc_col = []
eva_col = []
code_all = []