def plt_heat_map_results(plt_file_name = None):
opt = Options()
opt.disp_on = False
pob_siz_len = np.arange(start=3, stop=11,step=2) #not even steps needed
hist_len = np.arange(start=1, stop=5)
# pob_siz_len =[5,3,5,9,1]
results_success_rate_map_zero = []
results_astar_diff = []
cnt_Test=0
with K.get_session():
for p in pob_siz_len:
opt.pob_siz = p
opt.state_siz = (p * opt.cub_siz) ** 2
print("start with opt.pob_siz: {}".format(opt.pob_siz))
print("start with opt.state_siz: {}".format(opt.state_siz))
# get_data(opt)#generaten new data
for l in hist_len:
opt.hist_len = l
print("start with opt.hist_len: {}".format(opt.hist_len))
# train_model(opt, mdl_name,epochs=EPOCHS)
# [success_rate, astar_diff] = test_model(opt,mdl_name)
# results_success_rate_map_zero.append(success_rate)
results_success_rate_map_zero.append(cnt_Test)
cnt_Test1+=1
# results_astar_diff.append(astar_diff)
results_success_rate_map_zero=np.array(results_success_rate_map_zero)
plt.imshow(results_success_rate_map_zero.reshape(len(pob_siz_len),len(hist_len)), cmap='hot', interpolation='nearest')
plt.colorbar()
# plt.show()
helper_save(plt_file_name)
default=1000)
parser.add_argument("-ss",
"--silent",
help="Runs test without displaying the simulation",
action="store_false")
args = parser.parse_args()
model_path = args.model
# --------------------------------------------------------
# 0. initialization
opt = Options()
sim = Simulator(opt.map_ind, opt.cub_siz, opt.pob_siz, opt.act_num)
opt.disp_on = args.silent
if args.steps:
opt.steps = args.steps
# --------------------------------------------------------
# Input layer
image_dimension = opt.cub_siz * opt.pob_siz
img_rows = img_cols = image_dimension
input_shape = [img_rows, img_cols, opt.hist_len]
# --------------------------------------------------------
# Model
agent = model.model(input_shape)
agent.load_weights(model_path)
if not agent:
import numpy as np
np.random.seed(0)
from random import randrange
# custom modules
from utils import Options, rgb2gray
from simulator import Simulator
# 0. initialization
opt = Options()
sim = Simulator(opt.map_ind, opt.cub_siz, opt.pob_siz, opt.act_num)
states = np.zeros([opt.data_steps, opt.state_siz], float)
labels = np.zeros([opt.data_steps], int)
# Note I am forcing the display to be off here to make data collection fast
# you can turn it on again for debugging purposes
opt.disp_on = False
# 1. control loop
if opt.disp_on:
win_all = None
win_pob = None
epi_step = 0 # #steps in current episode
nepisodes = 1 # total #episodes executed
state = sim.newGame(opt.tgt_y, opt.tgt_x)
for step in range(opt.data_steps):
if state.terminal or epi_step >= opt.early_stop:
epi_step = 0
nepisodes += 1
state = sim.newGame(opt.tgt_y, opt.tgt_x)
else:
args = parser.parse_args() args.func(args) '''if not args.cfg_file: opt.prefix = args.prefix opt.set_algo(args.algo) opt.steps = args.steps opt.mtype = args.mtype opt.rtype = args.rtype opt.hist_len = args.hlength from_file = False else: opt.load_config_file(args.file) from_file = True''' opt.disp_on = args.visualize plot = args.plot_output model_name = opt.generate_model_name() run_name = opt.generate_run_name() # Environment setup opt.disp_on = args.visualize sim = Simulator(opt.map_ind, opt.cub_siz, opt.pob_siz, opt.act_num) if opt.rtype == opt.supported_rtypes[0]: opt.training = True else: opt.testing = True # Default model architecture layer_params = [
hours = int(sec / 3600)
rem = int(sec - (hours * 3600))
mins = rem / 60
rem = rem - (mins * 60)
secs = rem
print 'Training time:', hours, ':', mins, ':', secs
#Save the weights
model.save_weights(opt.weights_fil, overwrite=True)
print('Saved weights')
with open(opt.network_fil, "w") as outfile:
json.dump(model.to_json(), outfile)
#Testing
opt.disp_on = True
win_all = None
win_pob = None
action = np.argmax(
model.predict((state_with_history).reshape(1, img_rows, img_cols,
opt.hist_len)))
state_with_history = np.zeros((opt.hist_len, opt.state_siz))
append_to_hist(state_with_history, rgb2gray(state.pob).reshape(opt.state_siz))
next_state_with_history = np.copy(state_with_history)
epi_step = 0
nepisodes = 0
n_reached = 0.0
reward_acc_test = 0
reward_acc_list_test = []
print('Test Phase')
parser.add_argument(
"-s",
"--steps",
help="(Optional) Number of steps to train the model for. Default is 1000",
type=int,
default=10000)
parser.add_argument("-ss",
"--silent",
help="Runs test without displaying the simulation",
action="store_false")
args = parser.parse_args()
# --------------------------------------------------------
# 0. initialization
opt = Options()
opt.disp_on = True
sim = Simulator(opt.map_ind, opt.cub_siz, opt.pob_siz, opt.act_num)
# Display
if opt.disp_on:
win_all = None
win_pob = None
opt.disp_on = args.silent
opt.steps = args.steps
run_name = args.model
sess = tf.Session()
agent = DqnAgent(sess, is_duelling_dqn=False)
sess.run(tf.global_variables_initializer())
