In [4]:

    

scratch = 'scratch'

folders = ! ls -t -d $scratch/logs_*/ $scratch/zee_*/
folders = map(lambda x: x.replace(scratch+'/',''), folders)

print("\n".join(folders))


    

    




zee_x7c5_G1024to64x8_r1m61m5_D1024to32x6_lr1m5_dec1m6_cleaning/
zee_x7c5_G1024to64x8_r1m61m5_D1024to32x6_lr5m6_dec1m6_cleaning/
zee_x7c5_G1024to128x8_r1m61m5_D1024to32x6_lr1m5_dec1m6_cleaning/
zee_x7c5_bdtrw5_G512to64x8_r1m61m6_D1024to32x6_lr1m5_dec5m4_cleaning_m/
zee_x7c5_bdtrw5_G512to64x8_r1m51m5_D1024to32x6_lr5m6_dec1m6_cleaning_m/
zee_x7c5_bdtrw5_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec5m4_cleaning_m/
zee_x7c5_bdtrw5_G512to64x8_r0_D1024to32x6_lr1m5_dec1m6_cleaning_m_pretrain./
zee_x7c5_bdtrw5_G512to64x8_r0_D1024to32x6_lr1m5_dec1m6_cleaning_m_gaus_match_pretrain/
zee_x7c5_bdtrw5_G512to64x8_r0_D1024to32x6_lr1m5_dec1m6_cleaning_m_gaus_match/
zee_x5c2_G512x8_r1m51m5_D1024to32x6_lr1m6_dec1m6/
zee_x5c2_G512x8_r1m41m4_D1024to32x6_lr1m6_dec1m6/
zee_x5c2_G512to64x8_tnh_rlu_r1m41m4_D1024to32x6_lr1m6_dec1m6/
zee_x5c2_G512to64x8_r1m51m5_D1024to32x6_lr1m6_dec1m6/
logs_toy_1D_conditional_2M_200_lambda2em4/
logs_toy_1D_conditional_2M_200_lambda5em4/
logs_toy_1D_conditional_2M_200_lambda1em3/
logs_toy_1D_conditional_2M_200_lambda1em4/
logs_reweight_2Dx2D_clip5_500/
logs_reweight_2Dx2D_clip5_500_fix_last_layer/
logs_reweight_fix_clip5_500/
logs_reweight_G128x7_10000/
logs_toy_1D_conditional_2M_200_lambda1em2/
logs_reweight_2Dx2D_clip2_1000_lr2_tanh_only/
logs_reweight_2Dx2D_clip2_1000_lr2_tanh_only_D_bn/
logs_reweight_2Dx2D_clip2_500/
logs_reweight_2Dx2D_clip2_1000_lr2_tanh_G_128to32x7_D_bn/
logs_reweight_2Dx2D_clip2_1000_lr2_tanh_G_128x7/
logs_reweight_2Dx2D_clip2_1000_lr2/
logs_reweight_2Dx2D_clip2_1000_lr2_tanh/
logs_reweight_2Dx2D_clip2_1000_lr2_tanh_G_128to32x7/
logs_reweight_2Dx2D_clip1_500_D_dropoout0p5/
logs_reweight_2Dx2D_clip1_500_fix_last_layer/
logs_reweight_10000/
logs_reweight_2Dx2D_clip1_500/
logs_cms_zee_conditional_wgan_x5_c2_G512x8_D512x5_lr005_dec06_reg_l2_1em4_l1_1em4_1000_frac_data2/
logs_rewei_500/
logs_cms_zee_conditional_wgan_x5_c2_G512x8_D512x5_lr005_dec06_reg1em4_1000_frac_data2/
logs_cms_zee_conditional_wgan_G512x8_D512x8_lr005_dec06_reg1em6_1000_frac_data2/
logs_cms_zee_conditional_wgan_reg5em4_1000/
logs_cms_zee_conditional_wgan_x5_c2_G2048x8_do05_D512x5_lr005_dec06_reg1em4_1000_frac_data2/
logs_cms_zee_conditional_wgan_G512x8_D512x5_lr005_dec06_reg1em6_1000_frac_data2/
logs_cms_zee_conditional_wgan_G512x8_D512x8_lr005_dec05_reg1em6_1000_frac_data2/
logs_cms_zee_conditional_wgan_G512x8_D512x8_lr005_dec06_reg1em4_1000_frac_data2/
logs_cms_zee_conditional_wgan_G512to64x8_lr01_dec05_reg1em4_D_bn_1000_frac_data20/
logs_cms_zee_conditional_wgan_G512to64x8_lr01_dec05_reg1em5_1000_frac_data5/
logs_cms_zee_conditional_wgan_G512to64x8_lr01_dec05_reg1em4_1000_frac_data20/
logs_cms_zee_conditional_wgan_G512to64x8_lr01_dec05_reg1em4_1000_frac_data5/
logs_cms_zee_conditional_wgan_G512to64x8_lr005_dec05_reg1em5_1000_frac_data5/
logs_cms_zee_conditional_wgan_G512to64x8_lr005_dec05_reg1em5_1000_frac_data2/
logs_cms_zee_conditional_wgan_G512to64x7_lr01_reg1em4_1000_frac_data20/
logs_cms_zee_conditional_wgan_G512to64x8_lr001_dec05_reg1em5_1000_frac_data5/
logs_cms_zee_conditional_wgan_G512to64x7_lr01_dec05_reg1em4_1000_frac_data20/
logs_cms_zee_conditional_wgan_G512to64x7_lr01_reg1em4_1000/
logs_cms_zee_conditional_wgan_G512to64_lr01_reg1em4_1000/
logs_cms_zee_conditional_wgan_G512to64_reg5em4_1000/
zee_x7c5_bdtrw5_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning_m/
zee_x7c5_bdtrw5_G512to64x8_r0_D1024to32x6_lr1m5_dec1m6_cleaning_m/
zee_x7c5_bdtrw5_G512to64x8_r0_D1024to32x6_lr1m5_dec1m6_cleaning_m_pretrain/
zee_x7c5_bdtrw5_G512to64x8_r1m51m5_D1024to32x6_lr1m4_dec1m6_cleaning_m/
zee_x7c5_bdtrw5_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning_m_gaus/
zee_x7c5_bdtrw5_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning_m_gaus_match/
zee_x7c5_bdtrw5_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning_m_pt/
zee_x7c5_bdtrw5_G512to64x8_r1m51m6_D1024to32x6_lr1m5_dec1m6_cleaning_m/
zee_x7c5_bdtrw5_G512to64x8_r1m51m6_D1024to32x6_lr1m5_dec1m6_cleaning_m_gaus/
zee_x7c5_bdtrw5_G512to64x8_r1m61m5_D1024to32x6_lr1m5_dec1m6_cleaning_m/
zee_x7c5_bdtrw5_G512to64x8_r1m61m6_D1024to32x6_lr1m5_dec1m6_cleaning_m/
zee_x7c5_G1024to128x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning/
zee_x7c5_G1024to128x8_r1m51m6_D1024to32x6_lr1m5_dec1m6_cleaning/
zee_x7c5_G1024to128x8_r1m51m6_D1024to32x6_lr5m6_dec1m6_cleaning/
zee_x7c5_G1024to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning/
zee_x7c5_G1024to64x8_r1m51m6_D1024to32x6_lr1m5_dec1m6_cleaning/
zee_x7c5_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning/
zee_x7c5_rw5_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning_m_pt/
zee_x7c3t_rw5_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning/
zee_x7c4_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning/
zee_x7c4put_rw5_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning/
zee_x7c4put_rw5_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning_m_pt/
zee_x7c5_absrw5_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning/
zee_x7c5_absrw5_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning_m_pt/
zee_x5c2_G512to64x8_r1m51m4_D1024to32x6_lr1m5_dec1m6_cleaning/
zee_x5c2_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning/
zee_x5c2_G512to64x8_r1m51m5_D1024to32x6_lr1m6_dec1m6_cleaning/
zee_x5c2_G512to64x8_r1m61m4_D1024to32x6_lr1m5_dec1m6_cleaning/
zee_x5c2_G512to64x8_r1m61m6_D1024to32x6_lr1m5_dec1m6_cleaning/
zee_x7c2_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning/
zee_x7c3pu_rw5_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning/
zee_x7c3_rw3_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning/
zee_x7c3_rw5_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning/
zee_x7c3_rw5_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning_m_pt/

In [15]:

    

import os

targets = [
            'zee_x5c2_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning',
            'zee_x7c5_bdtrw5_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning_m',
    
          ]

targets = list(map(lambda x: os.path.join(scratch,x), targets ))

# print(list(targets))
dfs = [ pd.read_csv('%s/metrics.csv' % target) if os.path.exists('%s/metrics.csv' % target)  
       else None for target in targets ]

# df = pd.read_csv('%s/metrics.csv' % target)

for df,target in zip(dfs,targets):
    if df is None:
        continue
    df['emd_est'] = df['generator_loss']

    first = 0
    last = df.shape[0]
    last = min(last,300)
    
    fig = plt.figure(figsize=(15,3))
    ax = fig.add_subplot(111)
    # ax.set_yscale('log')
    
    df.loc[first:last].plot.scatter('epoch','emd_est',ax=ax)

    max_epoch = df['epoch'].max()
    minval = df.loc[first:last]['emd_est'].min()
    lastval = df.loc[last-1]['emd_est']
    change = df.loc[last-5:last]['emd_est'].mean() - lastval
    maxval = df.loc[first:last]['emd_est'].max()

    # plt.ylim(-0.1*maxval,maxval*1.2)

    #print(maxval,minval,lastval/minval,change/lastval)
    plt.ylim(minval*0.5,maxval*1.5)
    plt.title('%s\nepc max min/max last/min run_avg/last\n%d %1.2g %1.2g %1.2g %1.2g' % (os.path.basename(target),
                                                                                         max_epoch,maxval,minval/maxval,lastval/minval,change/lastval))
    plt.show()


    

In [17]:

    

from IPython.core.display import Image, display

from functools import reduce
from glob import glob

def display_images(which,targets=targets):
    images = reduce(lambda x,y: x+y, map(lambda x: glob('%s/%s' % (x,which)), targets) )
    for img in images:
        print(img)
        display(Image(img))
        
for var in range(7):
    display_images('*epoch_50_var%d*.png' % var) 
    display_images('*epoch_250_var%d*.png' % var) 
    #display_images('*epoch_300_var%d*.png' % var) 
    #display_images('*epoch_350_var%d*.png' % var)


    

    




scratch/zee_x5c2_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning/sampleepoch_50_var0.png






    













    




scratch/zee_x7c5_bdtrw5_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning_m/sampleepoch_50_var0.png






    













    




scratch/zee_x5c2_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning/sampleepoch_250_var0.png






    













    




scratch/zee_x7c5_bdtrw5_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning_m/sampleepoch_250_var0.png






    













    




scratch/zee_x5c2_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning/sampleepoch_50_var1.png






    













    




scratch/zee_x7c5_bdtrw5_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning_m/sampleepoch_50_var1.png






    













    




scratch/zee_x5c2_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning/sampleepoch_250_var1.png






    













    




scratch/zee_x7c5_bdtrw5_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning_m/sampleepoch_250_var1.png






    













    




scratch/zee_x5c2_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning/sampleepoch_50_var2.png






    













    




scratch/zee_x7c5_bdtrw5_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning_m/sampleepoch_50_var2.png






    













    




scratch/zee_x5c2_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning/sampleepoch_250_var2.png






    













    




scratch/zee_x7c5_bdtrw5_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning_m/sampleepoch_250_var2.png






    













    




scratch/zee_x5c2_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning/sampleepoch_50_var3.png






    













    




scratch/zee_x7c5_bdtrw5_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning_m/sampleepoch_50_var3.png






    













    




scratch/zee_x5c2_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning/sampleepoch_250_var3.png






    













    




scratch/zee_x7c5_bdtrw5_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning_m/sampleepoch_250_var3.png






    













    




scratch/zee_x5c2_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning/sampleepoch_50_var4.png






    













    




scratch/zee_x7c5_bdtrw5_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning_m/sampleepoch_50_var4.png






    













    




scratch/zee_x5c2_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning/sampleepoch_250_var4.png






    













    




scratch/zee_x7c5_bdtrw5_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning_m/sampleepoch_250_var4.png






    













    




scratch/zee_x7c5_bdtrw5_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning_m/sampleepoch_50_var5.png






    













    




scratch/zee_x7c5_bdtrw5_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning_m/sampleepoch_250_var5.png






    













    




scratch/zee_x7c5_bdtrw5_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning_m/sampleepoch_50_var6.png






    













    




scratch/zee_x7c5_bdtrw5_G512to64x8_r1m51m5_D1024to32x6_lr1m5_dec1m6_cleaning_m/sampleepoch_250_var6.png






    

In [48]:

    

import GAN.ffwd as ffwd

reload(ffwd)


    

    




Using TensorFlow backend.






    
Out[48]:





<module 'GAN.ffwd' from '/users/musella/jupyter/GAN/GAN/ffwd.py'>

In [ ]:

    

# discriminator = ffwd.FFDBuilder(**{'activation': None,
#   'clip_weights': 0.01,
#   'kernel_sizes': [256, 256, 256, 256, 256],
#   'name': 'D_256x5'})((1,1),(1,1))


    

In [ ]:

    

from keras.models import load_model
import GAN.models

from glob import glob 

discriminators = glob('%s/discriminator1-epoch*.hdf5' % target)

# discriminator = load_model('%s/discriminator-epoch400.hdf5' % target)


    

In [ ]:

    

def weight_lims(disc):
    maxes = []
    mins = []
    means = []
    for layer in discriminator.layers:
        print(layer)
        for weights in layer.get_weights():
            means.append(np.mean(weights))
            maxes.append(np.max(weights))
            mins.append(np.min(weights))
    return np.array( [means,mins,maxes] )

lims = []
for fname in discriminators:
    discriminator = load_model(fname)
    lims.append(weight_lims(discriminator))


    

In [ ]:

    

# maxes = []
# mins = []
# for layer in discriminator.layers:
#     print(layer)
#     for weights in layer.get_weights():
#         maxes.append(np.max(weights))
#         mins.append(np.min(weights))


    

In [ ]:

    

lims = np.array(lims)


    

In [ ]:

    

print(lims.shape)


    

In [ ]:

    

absmax = lims[:,2,:].max(axis=1)
absmin = lims[:,1,:].min(axis=1)


    

In [ ]:

    

absmin,absmax


    

In [ ]:

    

plt.plot(lims[:,2,[0,2,4,6,8,10]])
plt.plot(lims[:,1,[0,2,4,6,8,10]])


    

In [ ]:

    

plt.plot(lims[:,2,[1,3,5,7,9,11]])
plt.plot(lims[:,1,[1,3,5,7,9,11]])


    

In [ ]:

    

plt.plot(lims[:,0,[1,3,5,7,9,11]])
plt.plot(lims[:,0,[0,2,4,6,8,10]])


    

In [ ]:

    

generators = glob('%s/generator-epoch*.hdf5' % target)

generators


    

In [ ]:

    

generator = load_model(generators[-1])

print(generator.summary())


    

In [ ]:

    

from keras.layers import PReLU, Dense

prelus = []

for layer in generator.layers:
    if type(layer) == PReLU:
        prelus.append(layer)


    

In [ ]:

    

for layer in prelus:
    print(layer)
    for weights in layer.get_weights():
        print(weights.shape)
        plt.hist(weights.ravel())
        plt.show()


    

In [ ]:

    




    

In [ ]:

    




    

In [ ]: