In [1]:

    

%matplotlib inline
import matplotlib.pyplot as plt
plt.rcParams['figure.figsize'] = [20, 20]
import numpy as np
import dill as pickle

from src.data.dataset import *
from src.data import loop_ast
import pandas as pd

import seaborn as sns


    

In [2]:

    

ls data


    

    




old/  speedup_dataset2.pkl  speedup_dataset3.pkl

In [ ]:

    




    

In [4]:

    

DatasetFromPkl.pickle_data('../blur_h/', 'data/blur_h.pkl')


    

    




Reading data






    




100%|██████████| 2/2 [00:00<00:00, 6781.41it/s]





    




data loaded
Serializing
done






    

In [1]:

    




    

    




old  speedup_dataset2.pkl

In [5]:

    

D = DatasetFromPkl('data/speedup_dataset3.pkl', maxsize=None)


    

In [12]:

    




    

    
Out[12]:





array([[0.66116375],
       [2.9512894 ],
       [1.        ],
       ...,
       [0.08929031],
       [0.2774907 ],
       [0.9255319 ]], dtype=float32)

In [5]:

    

df = pd.DataFrame()

df['program'] = D.program_indexes
df['speedup'] = D.Y
df['schedule'] = [D.schedules[index].binary_repr for index in range(len(D))]


    

In [7]:

    

schedules = set(df['schedule'])


    

In [21]:

    

df[(df.schedule==(1, 0, 0)) & (df.program ==110)].speedup


    

    
Out[21]:





34    0.234309
Name: speedup, dtype: float32

In [23]:

    

scores = []

for sch in schedules:
    programs = set(df[df.schedule == sch].program)
    
    #score = np.mean([float(df[(df.schedule==sch) & (df.program == p)].speedup)/max(df[df.program==p].speedup) for p in programs])
    score = [(df[(df.schedule==sch) & (df.program == p)].speedup) for p in programs]
    scores.append(score)


    

In [36]:

    

programs = set(df.program)
schedules = set(df.schedule)

scores = {}
for schedule in schedules:
    scores[schedule] = []
    
for p in programs:
    tmp_df = df[df.program == p]
    
    best = max(tmp_df.speedup)
    
    for _, element in tmp_df.iterrows():
        scores[element.schedule].append(element.speedup/best)
        
for schedule in scores:
    scores[schedule] = np.mean(scores[schedule])


    

In [60]:

    

scores


    

    
Out[60]:





[((0, 0, 1), 0.827033572001672),
 ((0, 1, 0), 0.5479196742465374),
 ((0, 1, 1), 0.5404869417333152),
 ((0, 0, 0), 0.4665233634258242),
 ((1, 0, 1), 0.3725360176938043),
 ((1, 0, 0), 0.3488203177861382),
 ((1, 1, 1), 0.2653837692052882),
 ((1, 1, 0), 0.2541793923647423)]

In [ ]:

    




    

In [64]:

    

df.head()


    

    
Out[64]:







  

    

      

      
program
      
speedup
      
schedule
    
  
  

    

      
0
      
110
      
1.000000
      
(0, 0, 0)
    
    

      
1
      
110
      
1.927645
      
(0, 1, 0)
    
    

      
2
      
110
      
2.009328
      
(0, 1, 1)
    
    

      
3
      
110
      
1.992627
      
(0, 1, 1)
    
    

      
4
      
110
      
1.972955
      
(0, 1, 0)
    
  

In [57]:

    

scores


    

In [ ]:

    

df = pd.DataFramerame()

df['0.1'] = losses[0]
df['0.01'] = losses[1]
df['0.001'] = losses[2]
df['0.0001'] = losses[3]

df['x'] = df.index

df1 = df

plt.plot('x', '0.1', data=df1, color='blue')
plt.plot('x', '0.01', data=df1, color='orange')
plt.plot('x', '0.001', data=df1, color='green')
plt.plot('x', '0.0001', data=df1, color='red')
plt.xlabel('epochs')
plt.ylabel('loss')
plt.legend()


    

In [123]:

    

programs = set(df.program)

df1 = pd.DataFrame()

schedules = [(0, 1, 0), (1, 0, 0), (1, 1, 0)]
unrolled = []
not_unrolled = []

for p in programs:
    tmp_df = df[df.program == p]
    program_schedules = set(tmp_df.schedule)
    
    for sch in schedules:
        unrolled_sch = sch[0], sch[1], 1
        
        if sch in program_schedules and unrolled_sch in program_schedules:
            t = tmp_df[tmp_df.schedule == sch].speedup.mean()
            not_unrolled.append(t)
           
            t = tmp_df[tmp_df.schedule == unrolled_sch].speedup.mean()

            unrolled.append(t)
            
unrolled, not_unrolled = zip(*sorted(list(zip(unrolled, not_unrolled)), key=lambda x:x[1]))
            
df1['unrolled'] = unrolled
df1['not_unrolled'] = not_unrolled
df1['x'] = df1.index


    

In [124]:

    

df1['x'] = df1.index


    

In [2]:

    

plt.plot('x', 'unrolled', data=df1, color='blue')
plt.plot('x', 'not_unrolled', data=df1, label='not unrolled', color='orange')
plt.xlabel('scheduled program')
plt.ylabel('speedup')
plt.legend()


    

    




---------------------------------------------------------------------------
NameError                                 Traceback (most recent call last)
<ipython-input-2-9531d0dc7ee8> in <module>
----> 1 plt.plot('x', 'unrolled', data=df1, color='blue')
      2 plt.plot('x', 'not_unrolled', data=df1, label='not unrolled', color='orange')
      3 plt.xlabel('scheduled program')
      4 plt.xticks(fontsize=14)
      5 plt.ylabel('speedup')

NameError: name 'plt' is not defined

In [111]:

    

sns.lineplot(data=df1) (y + H*x)


    

    
Out[111]:





<matplotlib.axes._subplots.AxesSubplot at 0x7f6d9c63b4e0>






    

In [ ]: