I have been successful in performing a steady-state parameter estimation employing the same techniques presented in the Gekko tutorials (linear and non-linear regression). Below is the code:
# -*- coding: utf-8 -*-
"""
Spyder Editor
This is a temporary script file.
"""
from io import StringIO
from gekko import GEKKO
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
# Tb, Dy, Ho, Y, H, (HA)2
# measurements of the initial and final aqueous concentrations
aqueous_species = ["Tb", "Dy", "Ho", "Y"]
# import dataframe
x_i_a_Lns_v = (
pd.read_csv(
StringIO(
"""Tb,Dy,Ho,Y
0.19538,1.22628,0.2242,3.39823
0.28462,1.83411,0.32435,4.90551
0.34769,2.1979,0.39609,5.89209
0.37692,2.39495,0.43794,6.57722
0.41231,2.62232,0.47382,7.09791
0.43538,2.78906,0.5052,7.45418
0.44462,2.88,0.51866,7.89266
0.46,2.92548,0.52912,7.83785
0.46615,2.94064,0.5351,7.97488
0.45846,2.91032,0.52613,7.94747
0.46,2.86485,0.52613,7.89266
0.46769,2.92548,0.53659,8.00228
0.45692,2.92548,0.5351,8.02969
0.47385,2.98611,0.55005,8.13931
0.47385,2.97095,0.54108,8.1119"""
)
)
/ 1000
)
x_i_a_H_v = pd.read_csv(
StringIO(
"""H
10.01809
7.28346
5.16795
3.62036
2.50218
1.7173
1.17411
0.80491
0.54616
0.37078
0.25406
0.16932
0.11262
0.07574
0.04455"""
)
)
x_i_o_HA2_v = pd.read_csv(
StringIO(
"""(HA)2
0.4746
0.4746
0.4746
0.4746
0.4746
0.4746
0.4746
0.4746
0.4746
0.4746
0.4746
0.4746
0.4746
0.4746
0.4746"""
)
)
x_f_a_Lns_v = (
pd.read_csv(
StringIO(
"""Tb,Dy,Ho,Y
0.19231,1.23234,0.21822,3.34342
0.26923,1.74316,0.31239,4.60405
0.33692,2.13727,0.38862,5.72766
0.37385,2.33432,0.41253,6.05652
0.40923,2.50106,0.43196,5.97431
0.38769,2.1979,0.3363,3.91892
0.33692,1.53095,0.19431,1.91287
0.24,0.82307,0.0852,0.77282
0.12,0.33347,0.03139,0.27679
0.05846,0.14552,0.01495,0.1151
0.02615,0.06669,0.00598,0.05207
0.01231,0.03032,,0.02466
,,,0.00548
0.00615,0.01364,,0.01096
,,,0.00548"""
)
)
/ 1000
)
# Issue with NaN (missing value)
idx = x_f_a_Lns_v[x_f_a_Lns_v["Ho"] >= 0].index
# idx = x_f_a_Lns_v.index
#### Solution
# create model to be solved locally
m = GEKKO(remote=False)
x_i_a_Lns = [m.Param(value=x_i_a_Lns_v.loc[idx, v].values) for v in aqueous_species]
x_i_a_H = m.Param(value=x_i_a_H_v.loc[idx, "H"].values)
x_i_o_HA2 = m.Param(value=x_i_o_HA2_v.loc[idx, "(HA)2"].values)
x_f_a_Lns = [m.CV(value=x_f_a_Lns_v.loc[idx, v].values) for v in aqueous_species]
for i in range(len(aqueous_species)):
x_f_a_Lns[i].FSTATUS = 1
# equilibrium constants
K_eqs = [m.FV(value=1) for i in range(len(aqueous_species))]
for i in range(len(x_f_a_Lns_v.columns)):
K_eqs[i].STATUS = 1
# equilibrium model
x_f_a_H = m.Intermediate(
x_i_a_H
+ 3 * m.sum([(x_i_a_Lns[i] - x_f_a_Lns[i]) for i in range(len(aqueous_species))])
)
x_f_o_HA2 = m.Intermediate(
x_i_o_HA2
- 3 * m.sum([(x_i_a_Lns[i] - x_f_a_Lns[i]) for i in range(len(aqueous_species))])
)
m.Equations(
[
K_eqs[i]
== ((x_i_a_Lns[i] - x_f_a_Lns[i]) * x_f_a_H ** 3)
/ (x_f_a_Lns[i] * x_f_o_HA2 ** 3)
for i in range(len(aqueous_species))
]
)
# regression
m.options.IMODE = 2
m.options.EV_TYPE = 2
m.solve(disp=True)
for i, v in enumerate(aqueous_species):
print(f"K_eq_{v}: {K_eqs[i][0]}")
Output:
K_eq_Tb: 5.7327051238
K_eq_Dy: 15.581534791
K_eq_Ho: 27.414244408
K_eq_Y: 46.925190325
This is as expected.
However, when I try and implement a specific objective function I encounter an error (-13). I have tried setting a lower bound for the Var(s) however I then receive a different error code of (-1). I hope that that somebody can identify where I have made an error. I am of the view that the objective function is such that a similar result should be found.
# -*- coding: utf-8 -*-
"""
Spyder Editor
This is a temporary script file.
"""
from io import StringIO
from gekko import GEKKO
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
# Tb, Dy, Ho, Y, H, (HA)2
# measurements of the initial and final aqueous concentrations
aqueous_species = ["Tb", "Dy", "Ho", "Y"]
# import dataframes
x_i_a_Lns_v = (
pd.read_csv(
StringIO(
"""Tb,Dy,Ho,Y
0.19538,1.22628,0.2242,3.39823
0.28462,1.83411,0.32435,4.90551
0.34769,2.1979,0.39609,5.89209
0.37692,2.39495,0.43794,6.57722
0.41231,2.62232,0.47382,7.09791
0.43538,2.78906,0.5052,7.45418
0.44462,2.88,0.51866,7.89266
0.46,2.92548,0.52912,7.83785
0.46615,2.94064,0.5351,7.97488
0.45846,2.91032,0.52613,7.94747
0.46,2.86485,0.52613,7.89266
0.46769,2.92548,0.53659,8.00228
0.45692,2.92548,0.5351,8.02969
0.47385,2.98611,0.55005,8.13931
0.47385,2.97095,0.54108,8.1119"""
)
)
/ 1000
)
x_i_a_H_v = pd.read_csv(
StringIO(
"""H
10.01809
7.28346
5.16795
3.62036
2.50218
1.7173
1.17411
0.80491
0.54616
0.37078
0.25406
0.16932
0.11262
0.07574
0.04455"""
)
)
x_i_o_HA2_v = pd.read_csv(
StringIO(
"""(HA)2
0.4746
0.4746
0.4746
0.4746
0.4746
0.4746
0.4746
0.4746
0.4746
0.4746
0.4746
0.4746
0.4746
0.4746
0.4746"""
)
)
x_f_a_Lns_v = (
pd.read_csv(
StringIO(
"""Tb,Dy,Ho,Y
0.19231,1.23234,0.21822,3.34342
0.26923,1.74316,0.31239,4.60405
0.33692,2.13727,0.38862,5.72766
0.37385,2.33432,0.41253,6.05652
0.40923,2.50106,0.43196,5.97431
0.38769,2.1979,0.3363,3.91892
0.33692,1.53095,0.19431,1.91287
0.24,0.82307,0.0852,0.77282
0.12,0.33347,0.03139,0.27679
0.05846,0.14552,0.01495,0.1151
0.02615,0.06669,0.00598,0.05207
0.01231,0.03032,,0.02466
,,,0.00548
0.00615,0.01364,,0.01096
,,,0.00548"""
)
)
/ 1000
)
# Issue with NaN (missing value)
idx = x_f_a_Lns_v[x_f_a_Lns_v["Ho"] >= 0].index
# idx = x_f_a_Lns_v.index
#### Solution
# create model to be solved locally
m = GEKKO(remote=False)
x_i_a_Lns = [m.Param(value=x_i_a_Lns_v.loc[idx, v].values) for v in aqueous_species]
x_i_a_H = m.Param(value=x_i_a_H_v.loc[idx, "H"].values)
x_i_o_HA2 = m.Param(value=x_i_o_HA2_v.loc[idx, "(HA)2"].values)
x_f_a_Lns_m = [m.Param(value=x_f_a_Lns_v.loc[idx, v].values) for v in aqueous_species]
x_f_a_Lns = [m.Var() for v in aqueous_species]
# x_f_a_Lns = [m.Var(lb=1e-12) for v in aqueous_species]
# equilibrium constants
K_eqs = [m.FV(value=1) for i in range(len(aqueous_species))]
for i in range(len(aqueous_species)):
K_eqs[i].STATUS = 1
# equilibrium model
x_f_a_H = m.Intermediate(
x_i_a_H
+ 3 * m.sum([(x_i_a_Lns[i] - x_f_a_Lns[i]) for i in range(len(aqueous_species))])
)
x_f_o_HA2 = m.Intermediate(
x_i_o_HA2
- 3 * m.sum([(x_i_a_Lns[i] - x_f_a_Lns[i]) for i in range(len(aqueous_species))])
)
m.Equations(
[
K_eqs[i]
== ((x_i_a_Lns[i] - x_f_a_Lns[i]) * x_f_a_H ** 3)
/ (x_f_a_Lns[i] * x_f_o_HA2 ** 3)
for i in range(len(aqueous_species))
]
)
m.Obj(
m.sum(
[
((x_f_a_Lns_m[i] - x_f_a_Lns[i]) / x_f_a_Lns_m[i]) ** 2
for i in range(len(aqueous_species))
]
)
)
# regression
m.options.IMODE = 2
# m.options.EV_TYPE = 2
m.solve(disp=True)
for i, v in enumerate(aqueous_species):
print(f"K_eq_{v}: {K_eqs[i][0]}")
Error code -13 description for solver IPOPT is reported in the output:
EXIT: Invalid number in NLP function or derivative detected.
An error occured.
The error code is -13
This means that there is likely divide by zero somewhere in the model. The other error -1 is when an infeasible solution is detected. The additional constraints prevent the solver from finding a solution because the equations and variable constraints can't be satisfied.
Two things are needed to find a successful solution:
Rearrange equation to avoid divide by zero.
m.Equations(
[
K_eqs[i]
== ((x_i_a_Lns[i] - x_f_a_Lns[i]) * x_f_a_H ** 3)
/ (x_f_a_Lns[i] * x_f_o_HA2 ** 3)
for i in range(len(aqueous_species))
]
)
Rearrange to:
m.Equations(
[
K_eqs[i] * (x_f_a_Lns[i] * x_f_o_HA2 ** 3)
== ((x_i_a_Lns[i] - x_f_a_Lns[i]) * x_f_a_H ** 3)
for i in range(len(aqueous_species))
]
)
Switch to APOPT solver:
m.options.SOLVER=1
There is now a successful solution.
# -*- coding: utf-8 -*-
"""
Spyder Editor
This is a temporary script file.
"""
from io import StringIO
from gekko import GEKKO
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
# Tb, Dy, Ho, Y, H, (HA)2
# measurements of the initial and final aqueous concentrations
aqueous_species = ["Tb", "Dy", "Ho", "Y"]
# import dataframes
x_i_a_Lns_v = (
pd.read_csv(
StringIO(
"""Tb,Dy,Ho,Y
0.19538,1.22628,0.2242,3.39823
0.28462,1.83411,0.32435,4.90551
0.34769,2.1979,0.39609,5.89209
0.37692,2.39495,0.43794,6.57722
0.41231,2.62232,0.47382,7.09791
0.43538,2.78906,0.5052,7.45418
0.44462,2.88,0.51866,7.89266
0.46,2.92548,0.52912,7.83785
0.46615,2.94064,0.5351,7.97488
0.45846,2.91032,0.52613,7.94747
0.46,2.86485,0.52613,7.89266
0.46769,2.92548,0.53659,8.00228
0.45692,2.92548,0.5351,8.02969
0.47385,2.98611,0.55005,8.13931
0.47385,2.97095,0.54108,8.1119"""
)
)
/ 1000
)
x_i_a_H_v = pd.read_csv(
StringIO(
"""H
10.01809
7.28346
5.16795
3.62036
2.50218
1.7173
1.17411
0.80491
0.54616
0.37078
0.25406
0.16932
0.11262
0.07574
0.04455"""
)
)
x_i_o_HA2_v = pd.read_csv(
StringIO(
"""(HA)2
0.4746
0.4746
0.4746
0.4746
0.4746
0.4746
0.4746
0.4746
0.4746
0.4746
0.4746
0.4746
0.4746
0.4746
0.4746"""
)
)
x_f_a_Lns_v = (
pd.read_csv(
StringIO(
"""Tb,Dy,Ho,Y
0.19231,1.23234,0.21822,3.34342
0.26923,1.74316,0.31239,4.60405
0.33692,2.13727,0.38862,5.72766
0.37385,2.33432,0.41253,6.05652
0.40923,2.50106,0.43196,5.97431
0.38769,2.1979,0.3363,3.91892
0.33692,1.53095,0.19431,1.91287
0.24,0.82307,0.0852,0.77282
0.12,0.33347,0.03139,0.27679
0.05846,0.14552,0.01495,0.1151
0.02615,0.06669,0.00598,0.05207
0.01231,0.03032,,0.02466
,,,0.00548
0.00615,0.01364,,0.01096
,,,0.00548"""
)
)
/ 1000
)
# Issue with NaN (missing value)
idx = x_f_a_Lns_v[x_f_a_Lns_v["Ho"] >= 0].index
# idx = x_f_a_Lns_v.index
#### Solution
# create model to be solved locally
m = GEKKO(remote=False)
x_i_a_Lns = [m.Param(value=x_i_a_Lns_v.loc[idx, v].values) for v in aqueous_species]
x_i_a_H = m.Param(value=x_i_a_H_v.loc[idx, "H"].values)
x_i_o_HA2 = m.Param(value=x_i_o_HA2_v.loc[idx, "(HA)2"].values)
x_f_a_Lns_m = [m.Param(value=x_f_a_Lns_v.loc[idx, v].values) for v in aqueous_species]
x_f_a_Lns = [m.Var() for v in aqueous_species]
# x_f_a_Lns = [m.Var(lb=1e-12) for v in aqueous_species]
# equilibrium constants
K_eqs = [m.FV(value=1) for i in range(len(aqueous_species))]
for i in range(len(aqueous_species)):
K_eqs[i].STATUS = 1
# equilibrium model
x_f_a_H = m.Intermediate(
x_i_a_H
+ 3 * m.sum([(x_i_a_Lns[i] - x_f_a_Lns[i]) for i in range(len(aqueous_species))])
)
x_f_o_HA2 = m.Intermediate(
x_i_o_HA2
- 3 * m.sum([(x_i_a_Lns[i] - x_f_a_Lns[i]) for i in range(len(aqueous_species))])
)
m.Equations(
[
K_eqs[i] * (x_f_a_Lns[i] * x_f_o_HA2 ** 3)
== ((x_i_a_Lns[i] - x_f_a_Lns[i]) * x_f_a_H ** 3)
for i in range(len(aqueous_species))
]
)
m.Obj(
m.sum(
[
((x_f_a_Lns_m[i] - x_f_a_Lns[i]) / x_f_a_Lns_m[i]) ** 2
for i in range(len(aqueous_species))
]
)
)
# regression
m.options.IMODE = 2
m.options.SOLVER=1
# m.options.EV_TYPE = 2
m.solve(disp=True)
for i, v in enumerate(aqueous_species):
print(f"K_eq_{v}: {K_eqs[i][0]}")
Successful solution
---------------------------------------------------
Solver : APOPT (v1.0)
Solution time : 0.06899999999999999 sec
Objective : 0.32417750846391324
Successful solution
---------------------------------------------------
K_eq_Tb: 5.489166594
K_eq_Dy: 15.245134386
K_eq_Ho: 30.529918885
K_eq_Y: 54.83667882
Related
I'm new in AI, and i want to get in the field, i have spent some time finishing a program to train an agent for a simple customized environment, but when i perform the training in colab for 10000 episodes, it still can not get well performance. I guess whether there is something wrong with the customized env or there is something wrong with the training process.
Env: a helicopter tries to get throw the continous flow of birds (max num: 10), the birds moves from the right to the left, and there is fuel randomly. If the helicopter is still alive, i.e., it has not collided with a bird and still has fuel (initialized by 1000, when it collides with the fuel icon (max num: 2), fuel_left will be reset to 1000), its rewards plus 1.
the environment is shown in the figure:
after 10000 episode in ddpg/dqn, the agent still can not play more than 15 seconds, could you point out where the problem is?
Action space(1 dim): 0, 1, 2, 3, 4 -> helicopter moves up, down, left, right and keep static.
State space(28 dim): (x,y) for 10 birds, 2 fuel, and 1 helicopter. Besides, there is fuel left and rewards obtained.
Rewards: If the helicopter is alive, rewards plus 1.
the env settings code is as follwos (custom.py):
import numpy as np
import cv2
import matplotlib.pyplot as plt
import random
import math
import time
from gym import Env, spaces
import time
font = cv2.FONT_HERSHEY_COMPLEX_SMALL
class ChopperScape(Env):
def __init__(self):
super(ChopperScape,self).__init__()
self.maxbirdnum = 10
self.maxfuelnum = 2
self.observation_shape = (28,)
self.canvas_shape = (600,800,3)
self.action_space = spaces.Discrete(5,)
self.last_action = 0
self.obs = np.zeros(self.observation_shape)
self.canvas = np.ones(self.canvas_shape) * 1
self.elements = []
self.maxfuel = 1000
self.y_min = int (self.canvas_shape[0] * 0.1)
self.x_min = 0
self.y_max = int (self.canvas_shape[0] * 0.9)
self.x_max = self.canvas_shape[1]
def draw_elements_on_canvas(self):
self.canvas = np.ones(self.canvas_shape) * 1
for elem in self.elements:
elem_shape = elem.icon.shape
x,y = elem.x, elem.y
self.canvas[y : y + elem_shape[1], x:x + elem_shape[0]] = elem.icon
text = 'Fuel Left: {} | Rewards: {}'.format(self.fuel_left, self.ep_return)
self.canvas = cv2.putText(self.canvas, text, (10,20), font, 0.8, (0,0,0), 1, cv2.LINE_AA)
def reset(self):
self.fuel_left = self.maxfuel
self.ep_return = 0
self.obs = np.zeros(self.observation_shape)
self.obs[26] = self.maxfuel
self.bird_count = 0
self.fuel_count = 0
x = random.randrange(int(self.canvas_shape[0] * 0.05), int(self.canvas_shape[0] * 0.90))
y = random.randrange(int(self.canvas_shape[1] * 0.05), int(self.canvas_shape[1] * 0.90))
self.chopper = Chopper("chopper", self.x_max, self.x_min, self.y_max, self.y_min)
self.chopper.set_position(x,y)
self.obs[24] = x
self.obs[25] = y
self.elements = [self.chopper]
self.canvas = np.ones(self.canvas_shape) * 1
self.draw_elements_on_canvas()
return self.obs
def get_action_meanings(self):
return {0: "Right", 1: "Left", 2: "Down", 3: "Up", 4: "Do Nothing"}
def has_collided(self, elem1, elem2):
x_col = False
y_col = False
elem1_x, elem1_y = elem1.get_position()
elem2_x, elem2_y = elem2.get_position()
if 2 * abs(elem1_x - elem2_x) <= (elem1.icon_w + elem2.icon_w):
x_col = True
if 2 * abs(elem1_y - elem2_y) <= (elem1.icon_h + elem2.icon_h):
y_col = True
if x_col and y_col:
return True
return False
def step(self, action):
done = False
reward = 1
assert self.action_space.contains(action), "invalid action"
if action == 4:
self.chopper.move(0,5)
elif action == 1:
self.chopper.move(0,-5)
elif action == 2:
self.chopper.move(5,0)
elif action == 0:
self.chopper.move(-5,0)
elif action == 3:
self.chopper.move(0,0)
if random.random() < 0.1 and self.bird_count<self.maxbirdnum:
spawned_bird = Bird("bird_{}".format(self.bird_count), self.x_max, self.x_min, self.y_max, self.y_min)
self.bird_count += 1
bird_y = random.randrange(self.y_min, self.y_max)
spawned_bird.set_position(self.x_max, bird_y)
self.elements.append(spawned_bird)
if random.random() < 0.05 and self.fuel_count<self.maxfuelnum:
spawned_fuel = Fuel("fuel_{}".format(self.bird_count), self.x_max, self.x_min, self.y_max, self.y_min)
self.fuel_count += 1
fuel_x = random.randrange(self.x_min, self.x_max)
fuel_y = self.y_max
spawned_fuel.set_position(fuel_x, fuel_y)
self.elements.append(spawned_fuel)
for elem in self.elements:
if isinstance(elem, Bird):
if elem.get_position()[0] <= self.x_min:
self.elements.remove(elem)
self.bird_count -= 1
else:
elem.move(-5,0)
if self.has_collided(self.chopper, elem):
done = True
reward = -100000.0*(1.0/self.ep_return+1)
if isinstance(elem, Fuel):
flag1 = False
flag2 = False
if self.has_collided(self.chopper, elem):
self.fuel_left = self.maxfuel
flag1 = True
reward += 2
# time.sleep(0.5)
if elem.get_position()[1] <= self.y_min:
flag2 = True
self.fuel_count -= 1
else:
elem.move(0, -5)
if flag1 == True or flag2 == True:
self.elements.remove(elem)
self.fuel_left -= 1
if self.fuel_left == 0:
done = True
self.draw_elements_on_canvas()
self.ep_return += 1
birdnum = 0
fuelnum = 0
x_, y_ = self.chopper.get_position()
dis = 0.0
for elem in self.elements:
x,y = elem.get_position()
if isinstance(elem,Bird):
self.obs[2*birdnum] = x
self.obs[2*birdnum+1] = y
birdnum += 1
dis += math.hypot(x_-x,y_-y)
if isinstance(elem,Fuel):
base = self.maxbirdnum*2
self.obs[base+2*fuelnum] = x
self.obs[base+2*fuelnum+1] = y
fuelnum += 1
self.obs[24] = x_
self.obs[25] = y_
self.obs[26] = self.fuel_left
self.obs[27] = self.ep_return
if x_ == self.x_min or x_ == self.x_max or y_ == self.y_max or y_ == self.y_min:
reward -= random.random()
for i in range(26):
if i%2 == 0:
self.obs[i]/=800.0
else:
self.obs[i]/=600.0
self.obs[26]/=1000.0
self.obs[27]/=100.0
# print('reward:',reward)
# if done == True:
# time.sleep(1)
return self.obs, reward, done, {}
def render(self, mode = "human"):
assert mode in ["human", "rgb_array"], "Invalid mode, must be either \"human\" or \"rgb_array\""
if mode == "human":
cv2.imshow("Game", self.canvas)
cv2.waitKey(10)
elif mode == "rgb_array":
return self.canvas
def close(self):
cv2.destroyAllWindows()
class Point(object):
def __init__(self, name, x_max, x_min, y_max, y_min):
self.x = 0
self.y = 0
self.x_min = x_min
self.x_max = x_max
self.y_min = y_min
self.y_max = y_max
self.name = name
def set_position(self, x, y):
self.x = self.clamp(x, self.x_min, self.x_max - self.icon_w)
self.y = self.clamp(y, self.y_min, self.y_max - self.icon_h)
def get_position(self):
return (self.x, self.y)
def move(self, del_x, del_y):
self.x += del_x
self.y += del_y
self.x = self.clamp(self.x, self.x_min, self.x_max - self.icon_w)
self.y = self.clamp(self.y, self.y_min, self.y_max - self.icon_h)
def clamp(self, n, minn, maxn):
return max(min(maxn, n), minn)
class Chopper(Point):
def __init__(self, name, x_max, x_min, y_max, y_min):
super(Chopper, self).__init__(name, x_max, x_min, y_max, y_min)
self.icon = cv2.imread("chopper1.jpg") / 255.0
self.icon_w = 64
self.icon_h = 64
self.icon = cv2.resize(self.icon, (self.icon_h, self.icon_w))
class Bird(Point):
def __init__(self, name, x_max, x_min, y_max, y_min):
super(Bird, self).__init__(name, x_max, x_min, y_max, y_min)
self.icon = cv2.imread("bird1.jpg") / 255.0
self.icon_w = 32
self.icon_h = 32
self.icon = cv2.resize(self.icon, (self.icon_h, self.icon_w))
class Fuel(Point):
def __init__(self, name, x_max, x_min, y_max, y_min):
super(Fuel, self).__init__(name, x_max, x_min, y_max, y_min)
self.icon = cv2.imread("fuel1.jpg") / 255.0
self.icon_w = 32
self.icon_h = 32
self.icon = cv2.resize(self.icon, (self.icon_h, self.icon_w))
if __name__ == '__main__':
from IPython import display
env = ChopperScape()
obs = env.reset()
while True:
# random agent
action = random.randrange(-1,1)
obs, reward, done, info = env.step(action)
# Render the game
env.render()
if done == True:
break
env.close()
the ddpg algorithm to train the agent is as follows (ddpg.py):
from custom import ChopperScape
import random
import collections
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
#超参数
lr_mu = 0.005
lr_q = 0.01
gamma = 0.99
batch_size = 32
buffer_limit = 50000
tau = 0.005 # for target network soft update
class ReplayBuffer():
def __init__(self):
self.buffer = collections.deque(maxlen=buffer_limit)
def put(self, transition):
self.buffer.append(transition)
def sample(self, n):
mini_batch = random.sample(self.buffer, n)
s_lst, a_lst, r_lst, s_prime_lst, done_mask_lst = [], [], [], [], []
for transition in mini_batch:
s, a, r, s_prime, done = transition
s_lst.append(s)
a_lst.append([a])
r_lst.append(r)
s_prime_lst.append(s_prime)
done_mask = 0.0 if done else 1.0
done_mask_lst.append(done_mask)
return torch.tensor(s_lst, dtype=torch.float), torch.tensor(a_lst, dtype=torch.float), \
torch.tensor(r_lst, dtype=torch.float), torch.tensor(s_prime_lst, dtype=torch.float), \
torch.tensor(done_mask_lst, dtype=torch.float)
def size(self):
return len(self.buffer)
class MuNet(nn.Module):
def __init__(self):
super(MuNet, self).__init__()
self.fc1 = nn.Linear(28, 128)
self.fc2 = nn.Linear(128, 64)
self.fc_mu = nn.Linear(64, 1)
def forward(self, x):
x = F.relu(self.fc1(x))
x = F.relu(self.fc2(x))
mu = torch.tanh(self.fc_mu(x))
return mu
class QNet(nn.Module):
def __init__(self):
super(QNet, self).__init__()
self.fc_s = nn.Linear(28, 64)
self.fc_a = nn.Linear(1,64)
self.fc_q = nn.Linear(128, 32)
self.fc_out = nn.Linear(32,1)
def forward(self, x, a):
h1 = F.relu(self.fc_s(x))
h2 = F.relu(self.fc_a(a))
cat = torch.cat([h1,h2], dim=1)
q = F.relu(self.fc_q(cat))
q = self.fc_out(q)
return q
class OrnsteinUhlenbeckNoise:
def __init__(self, mu):
self.theta, self.dt, self.sigma = 0.1, 0.01, 0.1
self.mu = mu
self.x_prev = np.zeros_like(self.mu)
def __call__(self):
x = self.x_prev + self.theta * (self.mu - self.x_prev) * self.dt + \
self.sigma * np.sqrt(self.dt) * np.random.normal(size=self.mu.shape)
self.x_prev = x
return x
def train(mu, mu_target, q, q_target, memory, q_optimizer, mu_optimizer):
s,a,r,s_prime,done_mask = memory.sample(batch_size)
core = q_target(s_prime, mu_target(s_prime)) * done_mask
target = r + gamma * core
q_loss = F.smooth_l1_loss(q(s,a), target.detach())
q_optimizer.zero_grad()
q_loss.backward()
q_optimizer.step()
mu_loss = -q(s,mu(s)).mean() # That's all for the policy loss.
mu_optimizer.zero_grad()
mu_loss.backward()
mu_optimizer.step()
def soft_update(net, net_target):
for param_target, param in zip(net_target.parameters(), net.parameters()):
param_target.data.copy_(param_target.data * (1.0 - tau) + param.data * tau)
def main():
env = ChopperScape()
memory = ReplayBuffer()
q, q_target = QNet(), QNet()
q_target.load_state_dict(q.state_dict())
mu, mu_target = MuNet(), MuNet()
mu_target.load_state_dict(mu.state_dict())
score = 0.0
print_interval = 20
mu_optimizer = optim.Adam(mu.parameters(), lr=lr_mu)
q_optimizer = optim.Adam(q.parameters(), lr=lr_q)
ou_noise = OrnsteinUhlenbeckNoise(mu=np.zeros(1))
for n_epi in range(10000):
s = env.reset()
done = False
while not done:
a = mu(torch.from_numpy(s).float())
a = a.item() + ou_noise()[0]
print('action:',a)
s_prime, r, done, info = env.step(a)
env.render()
memory.put((s,a,r/100.0,s_prime,done))
score += r
s = s_prime
if memory.size()>20000:
for _ in range(10):
train(mu, mu_target, q, q_target, memory, q_optimizer, mu_optimizer)
soft_update(mu, mu_target)
soft_update(q, q_target)
if n_epi%print_interval==0 and n_epi!=0:
print("# of episode :{}, avg score : {:.1f}".format(n_epi, score/print_interval))
score = 0.0
env.close()
if __name__ == '__main__':
main()
and the dqn algorithm is as follows(dqn.py):
import gym
import collections
import random
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from custom import ChopperScape
#Hyperparameters
learning_rate = 0.0005
gamma = 0.98
buffer_limit = 50000
batch_size = 32
class ReplayBuffer():
def __init__(self):
self.buffer = collections.deque(maxlen=buffer_limit)
def put(self, transition):
self.buffer.append(transition)
def sample(self, n):
mini_batch = random.sample(self.buffer, n)
s_lst, a_lst, r_lst, s_prime_lst, done_mask_lst = [], [], [], [], []
for transition in mini_batch:
s, a, r, s_prime, done_mask = transition
s_lst.append(s)
a_lst.append([a])
r_lst.append([r])
s_prime_lst.append(s_prime)
done_mask_lst.append([done_mask])
return torch.tensor(s_lst, dtype=torch.float), torch.tensor(a_lst), \
torch.tensor(r_lst), torch.tensor(s_prime_lst, dtype=torch.float), \
torch.tensor(done_mask_lst)
def size(self):
return len(self.buffer)
class Qnet(nn.Module):
def __init__(self):
super(Qnet, self).__init__()
self.fc1 = nn.Linear(28, 128)
self.fc2 = nn.Linear(128, 128)
self.fc3 = nn.Linear(128, 5)
def forward(self, x):
x = F.relu(self.fc1(x))
x = F.relu(self.fc2(x))
x = self.fc3(x)
return x
def sample_action(self, obs, epsilon):
out = self.forward(obs)
# coin = random.random()
# if coin < epsilon:
# return random.randint(0,1)
# else :
# return out.argmax().item()
return out.argmax().item()
def train(q, q_target, memory, optimizer):
for _ in range(10):
s,a,r,s_prime,done_mask = memory.sample(batch_size)
q_out = q(s)
q_a = q_out.gather(1,a)
max_q_prime = q_target(s_prime).max(1)[0].unsqueeze(1)
target = r + gamma * max_q_prime * done_mask
loss = F.smooth_l1_loss(q_a, target)
optimizer.zero_grad()
loss.backward()
optimizer.step()
def main():
env = ChopperScape()
q = torch.load('10000_dqn_3.pt')
q_target = torch.load('10000_dqn_3_qtarget.pt')
# q_target.load_state_dict(q.state_dict())
memory = ReplayBuffer()
print_interval = 20
score = 0.0
optimizer = optim.Adam(q.parameters(), lr=learning_rate)
for n_epi in range(10000):
epsilon = max(0.01, 0.08 - 0.01*(n_epi/200)) #Linear annealing from 8% to 1%
s = env.reset()
done = False
while not done:
a = q.sample_action(torch.from_numpy(s).float(), epsilon)
s_prime, r, done, info = env.step(a)
env.render()
done_mask = 0.0 if done else 1.0
memory.put((s,a,r,s_prime, done_mask))
s = s_prime
if done:
break
score += r
if memory.size()>20000:
train(q, q_target, memory, optimizer)
if n_epi%print_interval==0 and n_epi!=0:
q_target.load_state_dict(q.state_dict())
print("n_episode :{}, score : {:.1f}, n_buffer : {}, eps : {:.1f}%".format(n_epi, score/print_interval, memory.size(), epsilon*100))
score = 0.0
env.close()
def test():
env = ChopperScape()
q = torch.load('10000_dqn_q.pt')
done = False
s = env.reset()
while not done:
a = q.sample_action(torch.from_numpy(s).float(), 1)
s_prime, r, done, info = env.step(a)
env.render()
s = s_prime
if done:
break
if __name__ == '__main__':
main()
when perform dqn, please annotate the action convert part in custom.py/class ChoperScape/step
after 10000 episode in ddpg/dqn, the agent still can not play more than 15 seconds, could you point out where the problem is?
I tried to create a CNN but I got an error. Could you figure out what I did wrong?
C1, C2 = tf.constant(70, dtype='float32'), tf.constant(1000, dtype="float32")
def score(y_true, y_pred):
tf.dtypes.cast(y_true, tf.float32)
tf.dtypes.cast(y_pred, tf.float32)
sigma = y_pred[:, 2] - y_pred[:, 0]
fvc_pred = y_pred[:, 1]
# sigma_clip = sigma + C1
sigma_clip = tf.maximum(sigma, C1)
delta = tf.abs(y_true[:, 0] - fvc_pred)
delta = tf.minimum(delta, C2)
sq2 = tf.sqrt(tf.dtypes.cast(2, dtype=tf.float32))
metric = (delta / sigma_clip) * sq2 + tf.math.log(sigma_clip * sq2)
return K.mean(metric)
def mloss(_lambda):
def loss(y_true, y_pred):
return _lambda * qloss(y_true, y_pred) + (1 - _lambda) * score(y_true, y_pred)
return loss
def make_model():
z = L.Input((9,), name="Patient")
x = L.Dense(100, activation="relu", name="d1")(z)
x = L.Dense(100, activation="relu", name="d2")(x)
p1 = L.Dense(3, activation="linear", name="p1")(x)
p2 = L.Dense(3, activation="relu", name="p2")(x)
preds = L.Lambda(lambda x: x[0] + tf.cumsum(x[1], axis=1), name="preds")([p1, p2])
model = M.Model(z, preds, name="CNN")
model.compile(loss=mloss(0.8), optimizer="adam", metrics=[score])
return model
net = make_model()
net.fit(z[tr_idx], y[tr_idx], batch_size=200, epochs=1000,
validation_data=(z[val_idx], y[val_idx]), verbose=0)
Error here: ValueError: Tensor conversion requested dtype int64 for
Tensor with dtype float32: <tf.Tensor 'CNN/preds/add:0' shape=(None,
3) dtype=float32>
type cast but didn't solve the problem
I made my custom model, AlexNetQIL (Alexnet with QIL layer)
'QIL' means quantization intervals learning
I trained my model and loss value didn't decrease at all and I found out parameters in my model were not updated at all because of QIL layer I added
I attached my codes AlexNetQil and qil
please someone let me know what's the problem in my codes
AlexNetQIL
import torch
import torch.nn as nn
from qil import *
class AlexNetQIL(nn.Module):
#def __init__(self, num_classes=1000): for imagenet
def __init__(self, num_classes=10): # for cifar-10
super(AlexNetQIL, self).__init__()
self.conv1 = nn.Conv2d(3, 64, kernel_size=3, stride=2, padding=1)
self.bn1 = nn.BatchNorm2d(64)
self.relu2 = nn.ReLU(inplace=True)
self.maxpool1 = nn.MaxPool2d(kernel_size=2)
self.qil2 = Qil()
self.conv2 = nn.Conv2d(64, 192, kernel_size=3, padding=1)
self.bn2 = nn.BatchNorm2d(192)
self.relu2 = nn.ReLU(inplace=True)
self.maxpool2 = nn.MaxPool2d(kernel_size=2)
self.qil3 = Qil()
self.conv3 = nn.Conv2d(192, 384, kernel_size=3, padding=1)
self.bn3 = nn.BatchNorm2d(384)
self.relu3 = nn.ReLU(inplace=True)
self.qil4 = Qil()
self.conv4 = nn.Conv2d(384, 256, kernel_size=3, padding=1)
self.bn4 = nn.BatchNorm2d(256)
self.relu4 = nn.ReLU(inplace=True)
self.qil5 = Qil()
self.conv5 = nn.Conv2d(256, 256, kernel_size=3, padding=1)
self.bn5 = nn.BatchNorm2d(256)
self.relu5 = nn.ReLU(inplace=True)
self.maxpool5 = nn.MaxPool2d(kernel_size=2)
self.classifier = nn.Sequential(
nn.Linear(256 * 2 * 2, 4096),
nn.ReLU(inplace=True),
nn.Linear(4096, 4096),
nn.ReLU(inplace=True),
nn.Linear(4096, num_classes),
)
def forward(self,x,inference = False):
x = self.conv1(x)
x = self.bn1(x)
x = self.relu2(x)
x = self.maxpool1(x)
x,self.conv2.weight = self.qil2(x,self.conv2.weight,inference ) # if I remove this line, No problem
x = self.conv2(x)
x = self.bn2(x)
x = self.relu2(x)
x = self.maxpool2(x)
x,self.conv3.weight = self.qil3(x,self.conv3.weight,inference ) # if I remove this line, No problem
x = self.conv3(x)
x = self.bn3(x)
x = self.relu3(x)
x,self.conv4.weight = self.qil4(x,self.conv4.weight,inference ) # if I remove this line, No problem
x = self.conv4(x)
x = self.bn4(x)
x = self.relu4(x)
x,self.conv5.weight = self.qil5(x,self.conv5.weight,inference ) # if I remove this line, No problem
x = self.conv5(x)
x = self.bn5(x)
x = self.relu5(x)
x = self.maxpool5(x)
x = x.view(x.size(0),256 * 2 * 2)
x = self.classifier(x)
return x
QIL
forward
quantize weights and input activation with 2 steps
transformer(params) -> discretizer(params)
import torch
import torch.nn as nn
import numpy as np
import copy
#Qil (Quantize intervals learning)
class Qil(nn.Module):
discretization_level = 32
def __init__(self):
super(Qil,self).__init__()
self.cw = nn.Parameter(torch.rand(1)) # I have to train this interval parameter
self.dw = nn.Parameter(torch.rand(1)) # I have to train this interval parameter
self.cx = nn.Parameter(torch.rand(1)) # I have to train this interval parameter
self.dx = nn.Parameter(torch.rand(1)) # I have to train this interval parameter
self.gamma = nn.Parameter(torch.tensor(1.0)) # I have to train this transformer parameter
self.a = Qil.discretization_level
def forward(self,x,weights,Inference = False):
if not Inference:
weights = self.transfomer_weights(weights)
weights = self.discretizer(weights)
x = self.transfomer_activation(x)
x = self.discretizer(x)
return torch.nn.Parameter(x), torch.nn.Parameter(weights)
def transfomer_weights(self,weights):
device = weights.device
aw,bw = (0.5 / self.dw) , (-0.5*self.cw / self.dw + 0.5)
weights = torch.where( abs(weights) < self.cw - self.dw,
torch.tensor(0.).to(device),weights)
weights = torch.where( abs(weights) > self.cw + self.dw,
weights.sign(), weights)
weights = torch.where( (abs(weights) >= self.cw - self.dw) & (abs(weights) <= self.cw + self.dw),
(aw*abs(weights) + bw)**self.gamma * weights.sign() , weights)
return weights
def transfomer_activation(self,x):
device = x.device
ax,bx = (0.5 / self.dx) , (-0.5*self.cx / self.dx + 0.5)
x = torch.where(x < self.cx - self.dx,
torch.tensor(0.).to(device),x)
x = torch.where(x > self.cx + self.dx,
torch.tensor(1.0).to(device),x)
x = torch.where( (abs(x) >= self.cx - self.dx) & (abs(x) <= self.cx + self.dx),
ax*abs(x) + bx, x)
return x
def discretizer(self,tensor):
q_D = pow(2, Qil.discretization_level)
tensor = torch.round(tensor * q_D) / q_D
return tensor
I have a server with two GPUs, if I use one GPU I need more than 10 days to finish 1000 epoch. But when I tried to use Dataparallel, the programm didn’t work. It seems because there is a ‘for’ cycle in my network. So how can I use Dataparallel in this case. Or is there any other solution to speed up training?
class WaveNet( nn.Module ):
def __init__(self, mu, n_residue, n_skip, dilation_depth, n_repeat):
# mu: audio quantization size
# n_residue: residue channels
# n_skip: skip channels
# dilation_depth & n_repeat: dilation layer setup
self.mu = mu
super( WaveNet, self ).__init__()
self.dilation_depth = dilation_depth
dilations = self.dilations = [2 ** i for i in range( dilation_depth )] * n_repeat
self.one_hot = One_Hot( mu )
self.from_input = nn.Conv1d( in_channels=mu, out_channels=n_residue, kernel_size=1 )
self.from_input = nn.DataParallel(self.from_input)
self.conv_sigmoid = nn.ModuleList(
[nn.Conv1d( in_channels=n_residue, out_channels=n_residue, kernel_size=2, dilation=d )
for d in dilations] )
self.conv_sigmoid = nn.DataParallel(self.conv_sigmoid)
self.conv_tanh = nn.ModuleList(
[nn.Conv1d( in_channels=n_residue, out_channels=n_residue, kernel_size=2, dilation=d )
for d in dilations] )
self.conv_tanh = nn.DataParallel(self.conv_tanh)
self.skip_scale = nn.ModuleList( [nn.Conv1d( in_channels=n_residue, out_channels=n_skip, kernel_size=1 )
for d in dilations] )
self.skip_scale = nn.DataParallel(self.skip_scale)
self.residue_scale = nn.ModuleList( [nn.Conv1d( in_channels=n_residue, out_channels=n_residue, kernel_size=1 )
for d in dilations] )
self.residue_scale = nn.DataParallel(self.residue_scale)
self.conv_post_1 = nn.Conv1d( in_channels=n_skip, out_channels=n_skip, kernel_size=1 )
self.conv_post_1 = nn.DataParallel(self.conv_post_1)
self.conv_post_2 = nn.Conv1d( in_channels=n_skip, out_channels=mu, kernel_size=1 )
self.conv_post_2 = nn.DataParallel(self.conv_post_2)
def forward(self, input, train=True):
output = self.preprocess( input, train )
skip_connections = [] # save for generation purposes
for s, t, skip_scale, residue_scale in zip( self.conv_sigmoid, self.conv_tanh, self.skip_scale,
self.residue_scale ):
output, skip = self.residue_forward( output, s, t, skip_scale, residue_scale )
skip_connections.append( skip )
# sum up skip connections
output = sum( [s[:, :, -output.size( 2 ):] for s in skip_connections] )
output = self.postprocess( output, train )
return output
TypeError: zip argument #1 must support iteration
I'm trying to implement a CNN using Theano and tried to test my code with a small sample-set of my bigger dataset. I'm trying to categorize a set of 8280 pictures(of 250*250 sizes) into 115 classes and my sample set is a set of 32 pictures of the first two classes(16 pictures from each). The problem I'm experiencing is that from the first epoch, the training loss in NaN and It will not change in the further epochs.
from __future__ import print_function
import sys
import os
import time
import numpy as np
import theano
import theano.tensor as T
import lasagne
import re
import cv2
from lasagne.layers import Conv2DLayer, MaxPool2DLayer , DropoutLayer
from lasagne.layers import InputLayer, DenseLayer, batch_norm
def split_list(a_list):
half = len(a_list)/2
return a_list[:half], a_list[half:]
def load_dataset(path=''):
cat_list = []
filelist = sorted(os.listdir(path))
trainlist = []
testlist = []
tmptrain = []
tmptest = []
max_id = 0
for f in filelist:
match = re.match(r'C(\d+)([F|G])(\d+)\.PNG', f)
id = int(match.group(1)) - 1
max_id = max(max_id,id)
fg_class = match.group(2)
fg_id = int(match.group(3))
if id not in [p[0] for p in cat_list]:
cat_list.append([id, [], []])
if fg_class == 'G':
cat_list[-1][1].append(f)
else:
cat_list[-1][2].append(f)
for f in cat_list:
id = f[0]
trainG, testG = split_list(f[1])
trainF, testF = split_list(f[2])
tmptrain = tmptrain + [(id, 1, F) for F in trainF] + [(id, 0, G) for G in trainG] # (Class_id,Forgery,Img)
tmptest = tmptest + [(id, 1, F) for F in testF] + [(id, 0, F) for F in testG]
X_train = np.array([cv2.imread(path+f[2],0) for f in tmptrain]).astype(np.int32)
y_train = np.array([f[0] for f in tmptrain]).astype(np.int32)
X_test = np.array([cv2.imread(path+f[2],0) for f in tmptest]).astype(np.int32)
y_test = np.array([f[0] for f in tmptest]).astype(np.int32)
fg_train = np.array([f[1] for f in tmptrain]).astype(np.int32)
fg_test = np.array([f[1] for f in tmptest]).astype(np.int32)
X_train = np.expand_dims(X_train,axis=1).astype(np.int32)
X_test = np.expand_dims(X_test, axis=1).astype(np.int32)
return X_train, y_train, X_test, y_test, fg_train , fg_test
def ExplicitNegativeCorrelation(net,layer='fc2',lr=0.00001):
for param in lasagne.layers.get_all_params(net[layer]):
if param.name.startswith('W'):
W = param
mean = T.mean(W,0) * lr
W = W - mean#T.mean(T.mean(W,0))
def ImplicitNegativeCorrelation(MSE,Cross,Hinge):
mean = T.mean((MSE+Cross+Hinge),axis=0)
return ((MSE-mean)**2+(Cross-mean)**2+(Hinge-mean)**2)/3
def build_cnn(inputvar,input_shape, trained_weights=None):
net = {}
net['input'] = InputLayer(input_shape,input_var=inputvar)
net['drop_input'] = DropoutLayer(net['input'],p=0.2)
net['conv1'] = batch_norm(Conv2DLayer(net['input'], num_filters=96, filter_size=11, stride=4, flip_filters=False))#,W=lasagne.init.HeNormal()))
net['pool1'] = MaxPool2DLayer(net['conv1'], pool_size=3, stride=2)
net['conv2'] = batch_norm(Conv2DLayer(net['pool1'], num_filters=256, filter_size=5, pad=2, flip_filters=False))#, W=lasagne.init.HeNormal()))
net['pool2'] = MaxPool2DLayer(net['conv2'], pool_size=3, stride=2)
net['conv3'] = batch_norm(Conv2DLayer(net['pool2'], num_filters=384, filter_size=3, pad=1, flip_filters=False))#, W=lasagne.init.HeNormal()))
net['conv4'] = batch_norm(Conv2DLayer(net['conv3'], num_filters=384, filter_size=3, pad=1, flip_filters=False))#, W=lasagne.init.HeNormal()))
net['conv5'] = batch_norm(Conv2DLayer(net['conv4'], num_filters=256, filter_size=3, pad=1, flip_filters=False))#, W=lasagne.init.HeNormal()))
net['pool5'] = MaxPool2DLayer(net['conv5'], pool_size=3, stride=2)
net['fc1'] = batch_norm(DenseLayer(net['pool5'], num_units=2048))
net['drop_fc1'] = DropoutLayer(net['fc1'])
net['fc2'] = batch_norm(DenseLayer(net['drop_fc1'], num_units=2048))
net['fc_class'] = batch_norm(DenseLayer(net['fc2'],num_units=115))
return net
def iterate_minibatches(inputs, targets_class,targets_verif, batchsize, shuffle=False):
assert len(inputs) == len(targets_class)
assert len(inputs) == len(targets_verif)
if shuffle:
indices = np.arange(len(inputs))
np.random.shuffle(indices)
for start_idx in range(0, len(inputs) - batchsize + 1, batchsize):
if shuffle:
excerpt = indices[start_idx:start_idx + batchsize]
else:
excerpt = slice(start_idx, start_idx + batchsize)
yield inputs[excerpt], targets_class[excerpt], targets_verif[excerpt]
def main(num_epochs=500):
print("Loading data...")
X_train, y_train, X_test, y_test, fg_train, fg_test = load_dataset('./signatures/tmp4/')
X_val, y_val, fg_val = X_train, y_train, fg_train
print(y_train.shape)
input_var = T.tensor4('inputs')
target_var_class = T.ivector('targets')
network = build_cnn(input_var, (None, 1, 250, 250))
class_prediction = lasagne.layers.get_output(network['fc_class']) # ,inputs={network['input']:input_var})
loss_class = lasagne.objectives.categorical_crossentropy(class_prediction, target_var_class)
loss = loss_class.mean()
params = lasagne.layers.get_all_params([network['fc_class']], trainable=True)
lr = 0.01
updates = lasagne.updates.nesterov_momentum(
loss, params, learning_rate=lr, momentum=0.9)
test_prediction_class = lasagne.layers.get_output(network['fc_class'], deterministic=True)
test_loss_class = lasagne.objectives.categorical_crossentropy(test_prediction_class,
target_var_class)
test_loss_class = test_loss_class.mean()
test_acc_class = T.mean(T.eq(T.argmax(test_prediction_class, axis=1), target_var_class),
dtype=theano.config.floatX)
predict_class = theano.function([input_var], T.argmax(test_prediction_class,axis=1))
train_fn = theano.function([input_var, target_var_class], loss, updates=updates)
val_fn_class = theano.function([input_var, target_var_class], [test_loss_class, test_acc_class])
print("Starting training...")
BatchSize = 2
for epoch in range(num_epochs):
train_err = 0
train_batches = 0
start_time = time.time()
for batch in iterate_minibatches(X_train, y_train,fg_train, BatchSize, shuffle=True):
inputs, targets_class, targets_verif = batch
train_err += train_fn(inputs, targets_class)
#ExplicitNegativeCorrelation(network, layer='fc2',lr=lr/10)
print(targets_class,predict_class(inputs))
train_batches += 1
val_err_class = 0
val_acc_class = 0
val_batches = 0
for batch in iterate_minibatches(X_val, y_val, fg_val, BatchSize, shuffle=False):
inputs, targets_class, targets_verif = batch
err_class, acc_class = val_fn_class(inputs, targets_class)
val_err_class += err_class
val_acc_class += acc_class
val_batches += 1
print("Epoch {} of {} took {:.3f}s".format(
epoch + 1, num_epochs, time.time() - start_time))
print(" training loss:\t\t{:.6f}".format(train_err / train_batches))
print(" Classification loss:\t\t{:.6f}".format(val_err_class / val_batches))
print(" Classification accuracy:\t\t{:.2f} %".format(
val_acc_class / val_batches * 100))
test_err_class = 0
test_acc_class = 0
test_err_verif = 0
test_acc_verif = 0
test_batches = 0
for batch in iterate_minibatches(X_test, y_test, fg_test, BatchSize, shuffle=False):
inputs, targets_class, targets_verif = batch
err_class, acc_class = val_fn_class(inputs, targets_class)
test_err_class += err_class
test_acc_class += acc_class
test_batches += 1
print("Final results:")
print(" test loss (Classification):\t\t\t{:.6f}".format(test_err_class / test_batches))
print(" test accuracy (Classification):\t\t{:.2f} %".format(
test_acc_class / test_batches * 100))
if __name__ == '__main__':
main()
I've tried to put lasagne.nonlinearities.softmax in the DenseLayers but it does fix the NaN issue but the accuracy of the Training model will not be any good, it will be fluctuating between 0 to 25%.(after 50 epochs!).
I have implemented a load_dataset function which I think that works correctly (I've tested the function multiple times), and I'm giving the class id of each picture as the target in the loss function. So my inputs and Targets would be like this:
Input Shape: (BatchSize, 1, 250, 250)
Target Shape: (BatchSize, 1) : vector of class ids
I've uploaded my sample-set here in this link.
It looks like we have 4 classes, according to the data, so I changed loading code to reflect it:
y_train = np.array([f[0] * 2 + f[1] for f in tmptrain]).astype(np.int32)
y_test = np.array([f[0] * 2 + f[1] for f in tmptest]).astype(np.int32)
Number of units in output layer should be equal to the number of classes, so I added an output layer with SoftMax:
net['fo_class'] = DenseLayer(net['fc_class'],num_units=4,
nonlinearity=lasagne.nonlinearities.softmax)
I suggest removing dropout layer just after inputs – you can compare outcomes with it and without it to make sure of that
Batch size = 2 is too small and learning rate is too high
Here is an example of code with those changes:
from __future__ import print_function
import sys
import os
import time
import numpy as np
import theano
import theano.tensor as T
import lasagne
import re
import cv2
from lasagne.layers import Conv2DLayer, MaxPool2DLayer , DropoutLayer
from lasagne.layers import InputLayer, DenseLayer
def split_list(a_list):
half = len(a_list)/2
return a_list[:half], a_list[half:]
def load_dataset(path=''):
cat_list = []
filelist = sorted(os.listdir(path))
tmptrain = []
tmptest = []
max_id = 0
for f in filelist:
match = re.match(r'C(\d+)([F|G])(\d+)\.PNG', f)
id = int(match.group(1)) - 1
max_id = max(max_id,id)
fg_class = match.group(2)
if id not in [p[0] for p in cat_list]:
cat_list.append([id, [], []])
if fg_class == 'G':
cat_list[-1][1].append(f)
else:
cat_list[-1][2].append(f)
for f in cat_list:
id = f[0]
trainG, testG = split_list(f[1])
trainF, testF = split_list(f[2])
tmptrain = tmptrain + [(id, 1, F) for F in trainF] + [(id, 0, G) for G in trainG]
tmptest = tmptest + [(id, 1, F) for F in testF] + [(id, 0, F) for F in testG]
X_train = np.array([cv2.imread(path+f[2],0) for f in tmptrain]).astype(np.float32)
y_train = np.array([f[0] * 2 + f[1] for f in tmptrain]).astype(np.int32)
X_test = np.array([cv2.imread(path+f[2],0) for f in tmptest]).astype(np.float32)
y_test = np.array([f[0] * 2 + f[1] for f in tmptest]).astype(np.int32)
fg_train = np.array([f[1] for f in tmptrain]).astype(np.float32)
fg_test = np.array([f[1] for f in tmptest]).astype(np.float32)
X_train = np.expand_dims(X_train,axis=1).astype(np.float32)
X_test = np.expand_dims(X_test, axis=1).astype(np.float32)
return X_train, y_train, X_test, y_test, fg_train , fg_test
def ExplicitNegativeCorrelation(net,layer='fc2',lr=0.00001):
for param in lasagne.layers.get_all_params(net[layer]):
if param.name.startswith('W'):
W = param
mean = T.mean(W,0) * lr
W = W - mean
def ImplicitNegativeCorrelation(MSE,Cross,Hinge):
mean = T.mean((MSE+Cross+Hinge),axis=0)
return ((MSE-mean)**2+(Cross-mean)**2+(Hinge-mean)**2)/3
def build_cnn(inputvar,input_shape, trained_weights=None):
net = {}
net['input'] = InputLayer(input_shape,input_var=inputvar)
net['conv1'] = Conv2DLayer(net['input'], num_filters=96, filter_size=11, stride=4)
net['pool1'] = MaxPool2DLayer(net['conv1'], pool_size=3, stride=2)
net['conv2'] = Conv2DLayer(net['pool1'], num_filters=256, filter_size=5, pad=2)
net['pool2'] = MaxPool2DLayer(net['conv2'], pool_size=3, stride=2)
net['conv3'] = Conv2DLayer(net['pool2'], num_filters=384, filter_size=3, pad=1)
net['conv4'] = Conv2DLayer(net['conv3'], num_filters=384, filter_size=3, pad=1)
net['conv5'] = Conv2DLayer(net['conv4'], num_filters=256, filter_size=3, pad=1)
net['pool5'] = MaxPool2DLayer(net['conv5'], pool_size=3, stride=2)
net['fc1'] = DenseLayer(net['pool5'], num_units=2048)
net['drop_fc1'] = DropoutLayer(net['fc1'])
net['fc2'] = DenseLayer(net['drop_fc1'], num_units=2048)
net['fc_class'] = DenseLayer(net['fc2'],num_units=115)
net['fo_class'] = DenseLayer(net['fc_class'],num_units=4,
nonlinearity=lasagne.nonlinearities.softmax)
return net
def iterate_minibatches(inputs, targets_class,targets_verif, batchsize, shuffle=False):
assert len(inputs) == len(targets_class)
assert len(inputs) == len(targets_verif)
if shuffle:
indices = np.arange(len(inputs))
np.random.shuffle(indices)
for start_idx in range(0, len(inputs) - batchsize + 1, batchsize):
if shuffle:
excerpt = indices[start_idx:start_idx + batchsize]
else:
excerpt = slice(start_idx, start_idx + batchsize)
yield inputs[excerpt], targets_class[excerpt], targets_verif[excerpt]
def main(num_epochs=500):
print("Loading data...")
X_train, y_train, X_test, y_test, fg_train, fg_test = load_dataset('./signatures/tmp4/')
X_train /= 255
X_val, y_val, fg_val = X_train, y_train, fg_train
print(y_train.shape)
check = X_train[0][0]
print(check)
input_var = T.tensor4('inputs')
target_var_class = T.ivector('targets')
network = build_cnn(input_var, (None, 1, 250, 250))
class_prediction = lasagne.layers.get_output(network['fo_class'])
loss_class = lasagne.objectives.categorical_crossentropy(class_prediction, target_var_class)
loss = loss_class.mean()
params = lasagne.layers.get_all_params([network['fo_class']], trainable=True)
lr = 0.0007
updates = lasagne.updates.nesterov_momentum(
loss, params, learning_rate=lr, momentum=0.9)
test_prediction_class = lasagne.layers.get_output(network['fo_class'], deterministic=True)
test_loss_class = lasagne.objectives.categorical_crossentropy(test_prediction_class,
target_var_class)
test_loss_class = test_loss_class.mean()
test_acc_class = T.mean(T.eq(T.argmax(test_prediction_class, axis=1), target_var_class),
dtype=theano.config.floatX)
predict_class = theano.function([input_var], T.argmax(test_prediction_class,axis=1))
train_fn = theano.function([input_var, target_var_class], loss, updates=updates)
val_fn_class = theano.function([input_var, target_var_class], [test_loss_class, test_acc_class])
print("Starting training...")
BatchSize = 16
for epoch in range(num_epochs):
train_err = 0
train_batches = 0
start_time = time.time()
for batch in iterate_minibatches(X_train, y_train,fg_train, BatchSize, shuffle=True):
inputs, targets_class, targets_verif = batch
train_err += train_fn(inputs, targets_class)
print(targets_class,predict_class(inputs))
train_batches += 1
val_err_class = 0
val_acc_class = 0
val_batches = 0
for batch in iterate_minibatches(X_val, y_val, fg_val, BatchSize, shuffle=False):
inputs, targets_class, targets_verif = batch
err_class, acc_class = val_fn_class(inputs, targets_class)
val_err_class += err_class
val_acc_class += acc_class
val_batches += 1
print("Epoch {} of {} took {:.3f}s".format(
epoch + 1, num_epochs, time.time() - start_time))
print(" training loss:\t\t{:.6f}".format(train_err / train_batches))
print(" Classification loss:\t\t{:.6f}".format(val_err_class / val_batches))
print(" Classification accuracy:\t\t{:.2f} %".format(
val_acc_class / val_batches * 100))
test_err_class = 0
test_acc_class = 0
test_batches = 0
for batch in iterate_minibatches(X_test, y_test, fg_test, BatchSize, shuffle=False):
inputs, targets_class, targets_verif = batch
err_class, acc_class = val_fn_class(inputs, targets_class)
test_err_class += err_class
test_acc_class += acc_class
test_batches += 1
print("Final results:")
print(" test loss (Classification):\t\t\t{:.6f}".format(test_err_class / test_batches))
print(" test accuracy (Classification):\t\t{:.2f} %".format(
test_acc_class / test_batches * 100))
if __name__ == '__main__':
main()