os.chdir('C:\Git\ZSS')
df = pd.read_csv('ZSSdata.csv', header=None)
dataset = df.to_numpy() # zss is broken, use angle_steers!
keys = [
'angle_steers', 'shitty_angle', 'zss', 'output_steer', 'wheel_speeds.fl',
'wheel_speeds.fr', 'wheel_speeds.rl', 'wheel_speeds.rr'
]
data = [np.take(dataset, indices=i, axis=1) for i in range(dataset.shape[1])]
data_zip = list(zip(keys, data))
scales = dict(zip(keys, [[np.min(i[1]), np.max(i[1])] for i in data_zip]))
# normalize data
data_normalized = norm(data_zip, scales)
inputs = [
'shitty_angle', 'output_steer', 'wheel_speeds.fl', 'wheel_speeds.fr',
'wheel_speeds.rl', 'wheel_speeds.rr'
]
output = 'angle_steers' # this is the accurate tssp2 sensor
# sort data into above format with output at the end for tokenization
seq_len = 20
# if we want to predict the future steer angle, we offset the output list by a few samples and remove that amount from beginning of input lists
data_sorted = np.array([
data_normalized[i] for i in inputs + [output]
]) # formats dict of lists into list of lists sorted by inputs list above
data_sorted = np.stack(
data_sorted, axis=1) # join the array so that each item is a full sample
samples_to_use = 7000000
if samples_to_use != 'all':
y_train = np.load("data/{}/y_train.npy".format(data_dir))[:samples_to_use]
else:
y_train = np.load("data/{}/y_train.npy".format(data_dir))
print(len(y_train))
is_array = False
if not os.path.exists(norm_dir.format(data_dir)):
if samples_to_use != 'all':
x_train = np.load("data/{}/x_train.npy".format(data_dir))[:samples_to_use]
else:
x_train = np.load("data/{}/x_train.npy".format(data_dir))
print("Normalizing...", flush=True)
normalized = norm(x_train)
x_train = normalized['normalized']
scales = normalized['scales']
print("Dumping normalization...", flush=True)
np.save(norm_dir.format(data_dir), x_train)
with open('data/LSTM/scales', "wb") as f:
pickle.dump(normalized['scales'], f)
#with open(norm_dir.format(data_dir), "wb") as f:
#pickle.dump(normalized, f)
else:
is_array = True
print("Loading normalized data...", flush=True)
x_train = np.load(norm_dir.format(data_dir))
with open('data/LSTM/scales', "rb") as f:
scales = pickle.load(f)
print('Loaded!', flush=True)
y_train = json.load(f)
NORM = True
if NORM:
#x_train_copy = list(x_train)
v_ego_scale = get_3d_min_max(x_train, 0)
a_ego_scale = get_3d_min_max(x_train, 1)
v_lead_scale = get_3d_min_max(x_train, 2)
x_lead_scale = get_3d_min_max(x_train, 3)
a_lead_scale = get_3d_min_max(x_train, 4)
x_train_reformat = [] # reformat to sequence data for nn to train, to test
for idx, i in enumerate(x_train):
x_train_reformat.append([[], [], [], [], []])
for x in i:
x_train_reformat[idx][0].append(norm(x[0], v_ego_scale))
x_train_reformat[idx][1].append(norm(x[1], a_ego_scale))
x_train_reformat[idx][2].append(norm(x[2], v_lead_scale))
x_train_reformat[idx][3].append(norm(x[3], x_lead_scale))
x_train_reformat[idx][4].append(norm(x[4], a_lead_scale))
x_train = x_train_reformat
#x_train = x_train * 2
#y_train = y_train * 2
x_train = np.asarray(x_train)
y_train = np.asarray([np.interp(i, [-1, 1], [0, 1]) for i in y_train])
else:
x_train = np.asarray(x_train)
y_train = np.asarray(y_train)
if stop.lower()=="y":
model.stop_training = True'''
#v_ego, v_lead, d_lead
os.chdir("C:/Git/dynamic-follow-tf")
with open("data/x", "r") as f:
x_train = json.load(f)
with open("data/y", "r") as f:
y_train = json.load(f)
NORM = True
if NORM:
v_ego, v_ego_scale = (norm([i[0] for i in x_train]))
a_ego, a_ego_scale = (norm([i[1] for i in x_train]))
v_lead, v_lead_scale = (norm([i[2] for i in x_train]))
x_lead, x_lead_scale = (norm([i[3] for i in x_train]))
a_lead, a_lead_scale = (norm([i[4] for i in x_train]))
x_train = [[v_ego[idx], a_ego[idx], v_lead[idx], x_lead[idx], a_lead[idx]]
for (idx, i) in enumerate(v_ego)]
#x_train.append([v_ego[idx], v_lead[idx], x_lead[idx], a_lead[idx]])
x_train = np.asarray(x_train)
#y_train = np.asarray([np.interp(i, [-1, 1], [0, 1]) for i in y_train])
#y_train = np.asarray(y_train)
#scaler = preprocessing.StandardScaler().fit(x_train)
#x_train = scaler.transform(x_train)
import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt
import os
from normalizer import norm
import itertools
os.chdir("C:/Git/dynamic-follow-tf/models/h5_models")
model = tf.keras.models.load_model("3model-epoch-8.h5")
v_scale, a_scale, x_scale = [0.0, 48.288787841797], [-8.39838886261, 9.78254699707], [0.125, 138.625]
data = [norm(0, v_scale), norm(0, v_scale), norm(0, x_scale), norm(0, a_scale)]
prediction=model.predict([[data]])[0][0]
#print(prediction)
print((prediction - 0.5)*2.0)
with open("data/LSTM/y_train", "r") as f:
y_train = json.load(f)
NORM = False
if NORM:
x_train_copy = list(x_train)
v_ego_scale = get_min_max(x_train, 0)
a_ego_scale = get_min_max(x_train, 1)
v_lead_scale = get_min_max(x_train, 2)
x_lead_scale = get_min_max(x_train, 3)
a_lead_scale = get_min_max(x_train, 4)
x_train = []
for idx, i in enumerate(x_train_copy):
x_train.append([])
for x in i:
x_train[idx].append([norm(x[0], v_ego_scale), norm(x[1], a_ego_scale), norm(x[2], v_lead_scale), norm(x[3], x_lead_scale), norm(x[4], a_lead_scale)])
#x_train = x_train * 2
#y_train = y_train * 2
x_train = np.asarray(x_train)
y_train = np.asarray([np.interp(i, [-1, 1], [0, 1]) for i in y_train])
else:
x_train = np.asarray(x_train)
y_train = np.asarray(y_train)
'''for idx,i in enumerate(y_train):
if i < -.5 and x_train[idx][0] > 8.9:
print(i)
print(idx)
print(x_train[idx])
break'''
import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt
import os
from normalizer import norm
os.chdir("C:/Git/dynamic-follow-tf/models/h5_models")
model = tf.keras.models.load_model("gm-only.h5")
v_ego_scale = [0.0, 36.1995735168457]
a_ego_scale = [-3.0412862300872803, 2.78971791267395]
v_lead_scale = [0.0, 91.02222442626953]
x_lead_scale = [0.9600000381469727, 138.67999267578125]
a_lead_scale = [-3.909122943878174, 25.991727828979492]
data = [
norm(0, v_ego_scale),
norm(0, a_ego_scale),
norm(1, v_lead_scale),
norm(3, x_lead_scale),
norm(.3, a_lead_scale)
]
prediction = model.predict(np.asarray([data]))[0][0]
print((prediction - 0.5) * 2.0)