def main(agent):
#all_results = np.load('data/{}.npy'.format(text))
all_results = np.zeros((1,17))
#np.save('data/{}.npy'.format(text), all_results)
model_path = 'model/{}.pt'.format(agent)
new_model_path = 'model/{}_new.pt'.format(agent)
while True:
for i in range(n_train):
print('Training:', i)
(results, winner) = game(agent, 0, i, new_model_path)
new_result = np.append(results, np.transpose(2 * np.abs(np.transpose(np.reshape(results[:,0], newshape=[-1,1])) + np.transpose(winner * np.ones((np.size(results, 0), 1))) -1) - 1), 1)
#all_results = new_result #if all_results == None else np.append(all_results, new_result, 0)
all_results = np.append(all_results, new_result, 0)
#np.save('data/{}.npy'.format(agent), all_results)
training(all_results, model_path, new_model_path, agent)
all_results = np.zeros((1,17))
# Validation
print("Validation...")
victory = 0
for i in range(n_valid):
(results, winner) = game(agent, 1, i, new_model_path)
victory += winner # number of victories for player 1
print('Validation', i, ':', 100 * victory / (i+1), '%')
if i > n_valid / 2 and victory / (i+1) < verif_prob - 0.1:
break
if victory / n_valid > verif_prob:
# model validated, replaced with new one
save_new_model(model_path, new_model_path)
def main(args):
start_time = time.time()
if args.preprocessing:
preprocessing(args)
if args.training:
training(args)
if args.testing:
testing(args)
end_time = round((time.time() - start_time) / 60, 4)
print(f'Done!; {end_time}min spend')
def run_training():
initial_time = get_time()
logger.info('Training start.')
data = serializer.load('data')
model = training(data)
serializer.dump(model, 'model')
logger.info('Training end: ' + str(calculate_time(initial_time)) + ' seconds.')
def training(self):
self.pushButtonStartPredict.setEnabled(False)
self.pushButtonStartTraining.setEnabled(False)
rapport = training.training()
self.plainTextResultTraining.setPlainText(rapport)
self.pushButtonStartPredict.setEnabled(True)
self.pushButtonStartTraining.setEnabled(True)
def train(params,id):
# make the folders
save_model_path = os.path.join(root_save_dir, str(id))
save_result_path = os.path.join(save_model_path,"test_results")
save_log_path = os.path.join(root_log_dir, "{}.txt".format(id))
if (not os.path.exists(save_model_path)):
os.makedirs(save_model_path)
os.makedirs(save_result_path)
# perform training
train_results = training(params=params, logger=LoggerGenerator.get_logger(
log_file_path=save_log_path), save_model_to=save_model_path,model_def=test_model,train_func=do_forward_pass)
# plot loss vs. iterations
lines = [str(l) for l in train_results["total_loss_records"]]
plot_trend_graph(var_names=["total loss"], var_indexes=[-1], var_types=["float"], var_colors=["r"], lines=lines,
title="total loss",save_to=os.path.join(save_result_path,"train-total_loss.png"),show_fig=False)
# perform testing
results = run_simple_test(params=params, saved_model_path=save_model_path,model_def=test_model)
# save test results
results["records"]["precision-recall-curve.jpg"].save(os.path.join(save_result_path,"precision-recall.png"))
with open(os.path.join(save_result_path,"metrics.txt"),"w") as f:
f.write(str(results["results"]))
print("finish testing for parameter set #{}".format(id))
def test_im(im_num_start, im_num_end):
# after training this function will predict the values on test images in a specific range
global train_images
global train_labels
global test_images
# get trained values
p_list, probs = training.training(train_labels,train_images)
test_result = [] # test_result will store the prediction of test images
for i in xrange(im_num_start, im_num_end + 1):
im = utils.read_image(i,test_images)
prob_v = [0,0,0,0,0,0,0,0,0,0] # prob_v stored the prob. for each value vj
for j in xrange(10):
p_a_v = 0
for k in xrange(28):
for l in xrange(28):
# work with sums of log probabilities rather than products of probabilities to avoid underflow errors
if im[k][l] == 0:
p_a_v = p_a_v + math.log10(probs[j][k][l][0])
if p_a_v == 0:
print 'problem'
else:
p_a_v = p_a_v + math.log10(probs[j][k][l][1])
if p_a_v == 0:
print 'problem'
prob_v[j] = math.log10(p_list[j]) + p_a_v
predict = prob_v.index(max(prob_v))
test_result.append(predict)
return test_result
def main():
rootPath = os.getcwd() + '\\' + 'training\\'
stopWordList = ['a', 'about', 'above', 'after', 'again', 'against', 'all', 'am', 'an', 'and', 'any', 'are', "aren't", 'as', 'at', 'be', 'because', 'been', 'before', 'being', 'below', 'between', 'both', 'but', 'by', "can't", 'cannot', 'could', "couldn't", 'did', "didn't", 'do', 'does', "doesn't", 'doing', "don't", 'down', 'during', 'each', 'few', 'for', 'from', 'further', 'had', "hadn't", 'has', "hasn't", 'have', "haven't", 'having', 'he', "he'd", "he'll", "he's", 'her', 'here', "here's", 'hers', 'herself', 'him', 'himself', 'his', 'how', "how's", 'i', "i'd", "i'll", "i'm", "i've", 'if', 'in', 'into', 'is', "isn't", 'it', "it's", 'its', 'itself', "let's", 'me', 'more', 'most', "mustn't", 'my', 'myself', 'no', 'nor', 'not', 'of', 'off', 'on', 'once', 'only', 'or', 'other', 'ought', 'our', 'ours' ]
classNames = ['Business', 'Entertainment', 'Politics']
trainingQueries = [ 'bing', 'amazon', 'twitter', 'yahoo', 'google', 'beyonce', 'bieber', 'television', 'movies', 'music', 'obama', 'america', 'congress', 'senate', 'lawmakers' ]
#docRetrieve(classNames, trainingQueries, rootPath)
training = training.training()
trainingDict, vocabulary = training.buildTrainingDict(rootPath , classNames, trainingQueries, stopWordList)
f = open(rootPath +'trainingDict.json', 'w')
f.write(json.dumps(trainingDict))
f.close()
f = open(rootPath + 'vocabulary.json', 'w')
f.write(json.dumps(list(vocabulary)))
rootPath = os.getcwd() + '\\' + 'testing\\'
testingQueries = [ 'apple', 'facebook', 'westeros', 'gonzaga', 'banana' ]
#docRetrieve(classNames, testingQueries, rootPath)
testing = testing.testing()
testingDict = testing.buildTestDict( rootPath , classNames, testingQueries, stopWordList)
f = open(rootPath +'testingDict.json', 'w')
f.write(json.dumps(trainingDict))
f.close()
score, class_tag = testing.naiveBayesTesting(classNames, trainingDict, testingDict)
print score
print class_tag
def main(args):
# updating all the global variables based on the input arguments
if (args.freeze_epochs):
config.FREEZE_EPOCHS = args.freeze_epochs
if (args.unfreeze_epochs):
config.UNFREEZE_EPOCHS = args.unfreeze_epochs
# updating batch size
if (args.batch_size):
config.PARAMS["batch_size"] = args.batch_size
# updating command line arguments to the ARGS variable
config.ARGS = args
# calling required functions based on the input arguments
if args.mode == "inference":
inference.inference()
else:
training.training(args)
def _parse_train(args):
print("starting training with:\n\t{} epochs\n\t{} batch size\n\t{} bars".format(args.epochs, args.batch, args.bars))
repr = "standard MIDI-like"
if args.pianoroll:
repr = "pianoroll"
print("\tusing " + repr + " representation")
if args.transpose:
print('\tusing traning data transposed to only one major and one minor key')
else:
print('\tusing training data transposed to every possible key')
print("\tverbose is set to " + str(args.verbose))
if args.resume:
if args.initialize:
raise ValueError("Resuming training from previous checkpoint and using pretrained weights is not possible!")
print("\ttrying to resume training from previous checkpoint")
if args.initialize:
print("\ttrying to load pretrained weights from 2 bar model\n")
training.training(args.epochs, args.batch, args.bars, args.pianoroll, args.transpose, args.verbose, args.save_location, args.resume, args.initialize)
def model_train(args, max_iter, learning_rate):
theta0, theta1, error = training(args, max_iter, learning_rate)
if error != "":
print(error)
return
save_theta(theta0, theta1)
theta = pd.read_csv(".save_model.csv", delimiter=',')
t0 = theta.Theta0[0]
t1 = theta.Theta1[0]
data = pd.read_csv(args.file, delimiter=',')
if args.visualise:
plot_graph_model(t0, t1, data)
def evaluate_optimization(candidates, args):
global BEST_FITNESS
global BEST_GENE
fitness = []
for cs in candidates:
fn = training(list(cs), EMBED)
fitness.append(fn)
if BEST_FITNESS > min(fitness):
BEST_FITNESS = min(fitness)
for i in cs:
BEST_GENE = [i[0], i[1], i[2], i[3], i[4], i[5], i[6], i[7], i[8], i[9], i[10], i[11], i[12]]
return fitness
def grid_search(bool, discount, softmax, sarsa, filepath):
episodes = 2000 # number of training episodes
alpha_list = [0.1, 0.3, 0.5]
epsilon_list = ['Lin', 'Exp']
epsilon_list_values = [
np.linspace(0.8, 0.001, episodes),
training.exp_decay(episodes)
]
if bool:
print('Discount:', discount, 'Softmax:', softmax, 'Sarsa:', sarsa)
# wall
wall = [[2, 1], [2, 2], [2, 3], [3, 3], [3, 4], [3, 9], [3, 8], [9, 6],
[7, 6], [6, 6], [6, 6], [6, 7], [6, 8], [8, 3], [7, 0]]
portal = [[0, 6], [9, 8]]
sand = [[2, 0], [8, 6], [4, 3], [5, 3], [7, 1], [3, 7], [9, 3]]
reward_val = np.zeros((2, 3))
for a, alpha in enumerate(alpha_list):
alpha_values = np.ones(episodes) * alpha
for e, epsilon_values in enumerate(epsilon_list_values):
_, reward, _ = training.training(10, episodes, wall, portal,
sand, alpha_values,
epsilon_values, discount,
softmax, sarsa)
reward_val[e, a] = reward
print('Alpha:', alpha, 'Epsilon:', epsilon_list[e], 'Reward:',
reward)
print('--------------------------------------------------')
print('\n')
np.save(path + filepath, reward_val)
reward_val = np.load(path + filepath + '.npy')
plot_map(reward_val, alpha_list, epsilon_list, filepath)
'max_examples': 10000,
'epochs': 5,
'lr': 0.01,
'train_files': ['/home/Liz/all_gis_islandviewer_iv4aa_data.csv.gz'],
'valid_files': ['/home/Liz/all_gis_islandviewer_iv4ag_data.csv.gz']
}
perform = []
for _dropout_keep in [0.5, 0.8]:
training_params['dropout_keep_prob'] = _dropout_keep
for _l2_coef in [1e-03, 1e-06]:
training_params['l2'] = _l2_coef
for _lr in [1e-1, 1e-3]:
training_params['lr'] = _lr
_best_cost = training(model_name, model_type, model_params,
training_params)
msg = 'dropout_keep:', _dropout_keep, 'l2 coef:', _l2_coef, 'lr:', _lr, 'Best val loss', _best_cost
print msg
logging.info(msg)
perform.append([_dropout_keep, _l2_coef, _lr, _best_cost])
logging.info('Finished with iterations.')
p = pd.DataFrame(perform,
columns=['dropout_keep', 'l2_coeff', 'lr', 'val_loss'
]).pivot_table(values=['val_loss'],
columns=['l2_coeff'],
index=['dropout_keep', 'lr'])
print p
logging.info('Writing to csv')
p.to_csv('~model_hyperparams_final.csv')
import sys
sys.path.append('../src')
from training import training
training('../dict/ntusd-positive.txt','../dict/ntusd-negative.txt','../model/','../dict/ntusd-full.dic')
def train(is_finetune=False):
tf.reset_default_graph()
startstep = 0 if not is_finetune else int(
FLAGS.finetune_dir.split('-')[-1])
with tf.Graph().as_default():
# ++++++++++++++++++++++++ TRAINING INPUT LAODING ++++++++++++++++++++++++
x_train, y_train, id_train = batch_inputs.inputs(
['./record/train.tfrecords'], FLAGS.batch_size, False)
print(x_train.shape)
y_train = tf.one_hot(y_train, FLAGS.num_class)
tf.summary.image('images', x_train)
x_train = tf.image.resize_image_with_crop_or_pad(
x_train, IMAGE_SIZE, IMAGE_SIZE)
y_train = tf.image.resize_image_with_crop_or_pad(
y_train, IMAGE_SIZE, IMAGE_SIZE)
# ++++++++++++++++++++++++ TESTING INPUT LAODING ++++++++++++++++++++++++
x_test, y_test, id_test = batch_inputs.inputs(
['./record/test.tfrecords'], FLAGS.batch_size, True)
y_test = tf.one_hot(y_test, FLAGS.num_class)
tf.summary.image('images', x_test)
x_test = tf.image.resize_image_with_crop_or_pad(
x_test, IMAGE_SIZE, IMAGE_SIZE)
y_test = tf.image.resize_image_with_crop_or_pad(
y_test, IMAGE_SIZE, IMAGE_SIZE)
# ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
is_training = tf.placeholder(tf.bool, name='is_training')
keep_prob = tf.placeholder(tf.float32, name="keep_probabilty")
images = tf.placeholder(
tf.float32,
shape=[None, FLAGS.image_h, FLAGS.image_w, FLAGS.image_c])
labels = tf.placeholder(
tf.int64, [None, FLAGS.image_h, FLAGS.image_w, FLAGS.num_class])
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
print('++++++++ Mode building starts here +++++++++')
if FLAGS.model == "basic":
logits = inference.inference_basic(images, is_training)
elif FLAGS.model == "extended":
logits = inference.inference_extended(images, is_training)
elif FLAGS.model == "basic_dropout":
logits = inference.inference_basic_dropout(images, is_training,
keep_prob)
elif FLAGS.model == "extended_dropout":
logits = inference.inference_extended_dropout(
images, is_training, keep_prob)
else:
raise ValueError("The selected model does not exist")
loss = evaluation.loss_calc(logits=logits, labels=labels)
train_op, global_step = training.training(loss=loss)
accuracy = tf.argmax(logits, axis=3)
summary = tf.summary.merge_all()
saver = tf.train.Saver(max_to_keep=1000)
with tf.Session() as sess:
if (is_finetune):
print(
"\n =====================================================")
print(" Finetuning with model: ", FLAGS.model)
print("\n Batch size is: ", FLAGS.batch_size)
print(" ckpt files are saved to: ", FLAGS.log_dir)
print(" Max iterations to train is: ", config.n_train_steps)
print(" =====================================================")
saver.restore(sess, FLAGS.finetune_dir)
else:
print(
"\n =====================================================")
print(" Training from scratch with model: ", FLAGS.model)
print("\n Batch size is: ", FLAGS.batch_size)
print(" ckpt files are saved to: ", FLAGS.log_dir)
print(" Max iterations to train is: ", config.n_train_steps)
print(" =====================================================")
sess.run(tf.variables_initializer(tf.global_variables()))
sess.run(tf.local_variables_initializer())
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
train_writer = tf.summary.FileWriter(FLAGS.log_dir, sess.graph)
for step in range(startstep + 1,
startstep + config.n_train_steps + 1):
images_batch, labels_batch = sess.run(
fetches=[x_train, y_train])
train_feed_dict = {
images: images_batch,
labels: labels_batch,
is_training: True,
keep_prob: 0.5
}
start_time = time.time()
_, train_loss_value, \
train_accuracy_value, \
train_summary_str = sess.run([train_op, loss, accuracy, summary], feed_dict=train_feed_dict)
# Finding duration for training batch
duration = time.time() - start_time
if step % 10 == 0: # Print info about training
examples_per_sec = FLAGS.batch_size / duration
sec_per_batch = float(duration)
print('\n--- Normal training ---')
format_str = (
'%s: step %d, loss = %.2f (%.1f examples/sec; %.3f '
'sec/batch)')
print(format_str % (datetime.now(), step, train_loss_value,
examples_per_sec, sec_per_batch))
# eval current training batch pre - class accuracy
pred = sess.run(logits, feed_dict=train_feed_dict)
# evaluation.per_class_acc(pred, labels_batch) # printing class accuracy
train_writer.add_summary(train_summary_str, step)
train_writer.flush()
if step % 100 == 0 or (step + 1) == config.n_train_steps:
# test_iter = FLAGS.num_examples_epoch_test // FLAGS.test_batch_size
test_iter = FLAGS.num_examples_epoch_test // FLAGS.batch_size
""" Validate training by running validation dataset """
print(
"\n==========================================================="
)
print("--- Running test on VALIDATION dataset ---")
total_val_loss = 0.0
# hist = np.zeros((FLAGS.num_class, FLAGS.num_class))
for val_step in range(test_iter):
test_img_batch, test_lbl_batch = sess.run(
fetches=[x_test, y_test])
val_feed_dict = {
images: test_img_batch,
labels: test_lbl_batch,
is_training: True,
keep_prob: 1.0
}
_val_loss, _val_pred = sess.run(
fetches=[loss, logits], feed_dict=val_feed_dict)
total_val_loss += _val_loss
# hist += evaluation.get_hist(_val_pred, val_labels_batch)
print(
"Validation Loss: ", total_val_loss / test_iter,
". If this value increases the model is likely overfitting."
)
# evaluation.print_hist_summery(hist)
print(
"==========================================================="
)
# Save the model checkpoint periodically.
if step % 1000 == 0 or step % 200 == 0 \
or (step + 1) == config.n_train_steps:
print("\n--- SAVING SESSION ---")
checkpoint_path = os.path.join(FLAGS.log_dir, 'model.ckpt')
saver.save(sess, checkpoint_path, global_step=step)
print("=========================")
coord.request_stop()
coord.join(threads)
x = self.pool2(x)
x = self.res31(x)
x = self.res32(x)
x = self.res33(x)
x = self.pool3(x)
x = self.res41(x)
x = self.res42(x)
x = self.res43(x)
x = self.pool4(x)
x = self.res51(x)
x = self.res52(x)
x = self.res53(x)
x = self.res54(x)
x = x.view(valid_size, 1, 512)
x = self.fc1(x)
x = self.softmax(x)
return x
if __name__ == "__main__":
classifier = ResNet()
lossF = nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(classifier.parameters(), lr=0.005)
optimizer.zero_grad()
training(classifier, lossF, optimizer, "resnet_best")
import torch
if __name__ == "__main__":
# input parameters
parser = argparse.ArgumentParser(description='Documentation in the following link: https://github.com/RualPerez/AutoML', formatter_class=argparse.RawTextHelpFormatter)
parser.add_argument('--batch', help='Batch size of the policy (int)', nargs='?', const=1, type=int, default=15)
parser.add_argument('--max_layer', help='Maximum nb layers of the childNet (int)', nargs='?', const=1, type=int, default=6)
parser.add_argument('--possible_hidden_units', default=[1,2,4,8,16,32], nargs='*',
type=int, help='Possible hidden units of the childnet (list of int)')
parser.add_argument('--possible_act_functions', default=['Sigmoid', 'Tanh', 'ReLU', 'LeakyReLU'], nargs='*',
type=int, help='Possible activation funcs of the childnet (list of str)')
parser.add_argument('--verbose', help='Verbose while training the controller/policy (bool)', nargs='?', const=1,
type=bool, default=False)
parser.add_argument('--num_episodes', help='Nb of episodes the policy net is trained (int)', nargs='?', const=1,
type=int, default=500)
args = parser.parse_args()
# parameter settings
args.possible_hidden_units += ['EOS']
total_actions = args.possible_hidden_units + args.possible_act_functions
n_outputs = len(args.possible_hidden_units) + len(args.possible_act_functions) #of the PolicyNet
# setup policy network
policy = PolicyNet(args.batch, n_outputs, args.max_layer)
# train
policy = training(policy, args.batch, total_actions, args.verbose, args.num_episodes)
# save model
torch.save(policy.state_dict(), 'policy.pt')
people_num // n_fold_num * (n_fold_num - 1))
val_y = val_y.reshape(n_fold_num, people_num // n_fold_num)
if (n_fold_num == 1):
temp = train_x
train_x = val_x
val_x = temp
temp = train_y
train_y = val_y
val_y = temp
#train the folds
acc = 0.0
for f in range(n_fold_num):
b_sol, w_sol, mean, sigma, feature_enco = training.training(
train_x, train_y, f, feature_num, n_fold_num, people_num, feature,
norm)
if (f == 0):
w_his1 = w_sol
b_his1 = [b_sol]
else:
w_his1 = np.append(w_his1, w_sol)
b_his1 = np.append(b_his1, b_sol)
if (n_fold_num != 1):
acc_temp = training.validation(val_x, val_y, w_sol, b_sol, f,
feature_num, feature, norm, mean, sigma,
ratio)
print(acc_temp)
acc = acc + acc_temp
import sys
sys.path.append('../src')
from training import training
training('../data/positive.txt','../data/negative.txt','../model/','../dict/user_dic.dic')
def get_fitness(self, vector):
"""
maximize -(x^2 + (y+1)^2) + 4
The maximum is 3.75 at (0.5, -1) (remember that x is fixed at 0.5 here).
"""
return training(vector, EMBED)
cropped_face=cropped_faces[i]
label,confidence=face_recognizer.predict(cropped_face) #confidence is an integer returned by predict method which gives associated confidence (e.g. distance) for the predicted label.
cv.rectangle(img,(x, y),(x+ width, y + height),(0, 255, 0),2) #draw a rectangle around detected face
cv.putText(img,str(names[label]),(x,y),cv.FONT_HERSHEY_PLAIN,1,(0,255,0),1) #put the correct label above the detected face
return img
#train the data using all the images in the dataset
faces,labels=training("training-data")
#LFBP Face Recognizer
face_recognizer=cv.face.LBPHFaceRecognizer_create()
#convert labels (list) to numpy array as LBPH Face Recognizer needs a numpy array as its second argument
labels=np.array(labels)
face_recognizer.train(faces,labels)
testimagepath="test-data/1.jpeg"
names={1:'Evans',2:'Hemsworth'} # a dictionary to store name associated with each label
predictedimg=predict(testimagepath,face_recognizer,names)
if predictedimg is not None:
return draw_img
if __name__ == '__main__':
global history, svc, X_scaler, orient, pix_per_cell, cell_per_block, spatial_size, hist_bins
history = deque(maxlen = 8)
pix_per_cell = 8 # HOG pixels per cell
cell_per_block = 2 # HOG cells per block
spatial_size = (32, 32) # Spatial binning dimensions
hist_bins = 32 # Number of histogram bins
orient = 15 # HOG orientations
images = glob.glob('./training-data/*/*/*.png')
print(images)
svc, X_scaler = training(images, orient, pix_per_cell, cell_per_block, spatial_size, hist_bins)
image = mpimg.imread('./test_images/test4.jpg')
img = detect_cars(image)
fig = plt.gcf()
fig.set_size_inches(16.5, 8.5)
plt.imshow(img)
plt.show()
output = 'project_video_result.mp4'
clip = VideoFileClip("project_video.mp4")
video_clip = clip.fl_image(detect_cars)
video_clip.write_videofile(output, audio=False)
def main():
# Safety checkpoint
if (len(sys.argv) < 2 or len(sys.argv) > 3):
print('Number of arguments is incorrect')
Usage()
''' Creating the Model '''
network, criterion, optimizer = create_smaller_UNET() # 3 layers model
if (sys.argv[1] == '--predict'):
print("Loading the model... ")
checkpoint = torch.load(BEST_MODEL_PATH)
network.load_state_dict(checkpoint['state_dict'])
elif (sys.argv[1] == '--train'):
print("Reading", TRAINING_SIZE, "training image(s)")
trainset = DatasetPatched(TRAIN_IMAGE_PATH, TRAIN_GROUNDTRUTH_PATH,
TRAINING_SIZE, OVERLAY_SIZE, ROTATION)
trainloader = DataLoader(trainset, batch_size=BATCH_SIZE, shuffle=True)
print("Reading", VAL_SIZE, "validation image(s)")
valset = DatasetPatched(TRAIN_IMAGE_PATH, TRAIN_GROUNDTRUTH_PATH,
VAL_SIZE, OVERLAY_SIZE, ROTATION)
valloader = DataLoader(valset, batch_size=BATCH_SIZE, shuffle=True)
''' Training phase '''
# TODO: optional, add: ReduceLROnPlateau
# TODO: optional, put early stopping
val_loss_hist, val_loss_hist_std, train_loss_hist, train_loss_hist_std, val_acc_hist, val_acc_hist_std, train_acc_hist, train_acc_hist_std = training(
network, criterion, optimizer, score, trainloader, valloader,
PATCH_SIZE, NUM_EPOCHS)
print("Saving the model... ")
torch.save({'state_dict': network.state_dict()}, MODEL_PATH)
else:
print("Argument not recognized")
Usage()
print("Load testing data... ")
testset = TestDataset(TEST_IMAGE_PATH, NR_TEST_IMAGES)
loader_test = DataLoader(testset, batch_size=1,
shuffle=False) # To do : increase the batch_size
print("Predict labels... ")
roadsPredicted = predict_test_images(network, loader_test)
# Transform pixel-wise prediction to patchwise
patched_images = patch_prediction(roadsPredicted, TEST_IMG_SIZE,
IMG_PATCH_SIZE)
''' Get patches for submission '''
patches = getPatches(patched_images)
''' Generate submission '''
reconstruct_img(NR_TEST_IMAGES, TEST_IMG_SIZE, IMG_PATCH_SIZE, patches,
PREDICTED_PATH)
submission_to_csv(SUBMISSION_PATH, PREDICTED_PATH)
print('See latest submission file ', SUBMISSION_PATH)
print('See predicted images at ', PREDICTED_PATH)
def startTraining(self):
training.training(self.get_next_input,self.get_next_reference_sequence)
return 0
import preprocessing
import training as tr
# Load dataset
dataset = np.load(params.PATH_DATA)
# Splitting the dataset
index_train, index_valid, index_test = preprocessing.split_dataset(dataset)
# Define the model
model = models.get_LSTM_v1(params.T - 1, params.D - 1, params.LR,
params.NHIDDEN, params.NNEURONSH,
params.DROPOUT_RATE)
#model = models.get_CausalCNN_v5(params.T-1, params.D-1, params.LR, params.DROPOUT_RATE)
# Training
tr.training(params.PATH_EXPERIMENT,
model,
dataset,
index_train,
index_valid,
params.D,
params.B,
params.NB_SAMPLES_PER_EPOCH,
params.NB_EPOCHS,
params.PATIENCE,
params.LR,
weighted_samples=params.WEIGHTED_SAMPLES,
pretrained=params.PRETRAINED,
h5py=params.H5PY)
from training import training import numpy as np from loadData import loadData with open( 'twidf_window4_directed_weighted' ,"r") as File: X = np.loadtxt(File , delimiter=',') path = '../data/r8_train_stemmed.txt' trainData = True data = loadData(path,trainData) labels = data['labels'] (dictionnaryOfClasses , labelsInNumbers) = labelDictionnary(labels) lsi = True numberOfComponents = 100 (reducedMatrix , Y) = dimensialityReduction(X , labelsInNumbers , lsi , numberOfComponents) svm = True scores = training(reducedMatrix , Y , svm ) print scores['micro'] print scores['macro']
#!/usr/bin/env python3 # -*- coding: utf-8 -*- """ Created on Thu Mar 16 19:42:07 2017 @author: axionl """ from train_data import get_train_data from training import training from testing import testing import numpy as np import matplotlib.pyplot as plt root_path = "../Training_data/" [train_data, distance] = get_train_data(root_path) alpha = 0.01 # 下降速度 num_iters = 300 # 迭代次数 training_summary = np.zeros((1, 6)) T = 5 # 矩阵行数 theta_summary = training(train_data, distance, T, alpha, num_iters) # testing_result = testing(theta_summary) error = testing(train_data, distance, T, theta_summary) plt.plot(error[:, 1:2], error[:, 0:1], 'ro') plt.show()
o_file.write('#file_name, file_size(octect), number_of_lines_model, number_of_lines_test, mean_time_training, mean_time_prediction, accuracy\n')
dataFiles = os.listdir("./data")
dataFiles = ["./data/"+i for i in dataFiles]
dataFiles.sort()
#MLlib seems to take time to have maximum speed so we "train" it
print('"Warming up" MLLib to get maximum speed')
for i in range(0, 5):
testTree.main(sc, dataFiles[0], NB_PARTITIONS)
print("----------------")
print("Real test begins")
print("----------------")
for it, dataFile in enumerate(dataFiles):
stats = []
partialStat = training.training(sc, dataFile, NB_PARTITIONS)
#Compute statistics on dataFile
size = os.path.getsize(dataFile)
stats += [dataFile, size]
stats += partialStat
with open(OUTPUT_FILE, 'a') as o_file:
wr = csv.writer(o_file)
wr.writerow(stats)
print("----------------")
print("Test finished")
print("----------------")
print("Writing results on " + OUTPUT_FILE)
Plots the total error for L constant, and W taking values chosen in constant.py
"""
from testing import testing
from training import training
from constant import W, LL, N, N_ite, learning_rate, NB_EPOCH, id_function
import numpy as np
import matplotlib.pyplot as plt
errormax = []
errorplot = []
errormin = []
for lite in LL:
errorl = []
for n in range(N_ite):
training(W, lite, N, NB_EPOCH, learning_rate, id_function)
errorl.append(testing(W, lite, N, NB_EPOCH, learning_rate,
id_function))
means = np.mean(errorl)
sigma = np.sqrt(np.var(errorl, ddof=1))
errormax.append(means + sigma)
errorplot.append(means)
errormin.append(means - sigma)
plt.clf()
plt.figure()
fig, ax1 = plt.subplots()
ax1.set_xlabel("L", fontsize=14)
ax1.set_ylabel("Error", fontsize=14)
from training import training from data import reprod_init from vizualization import vizualization from MNISTNet import MNISTNet reprod_init() mnist_net = MNISTNet(100) test_loss, test_accuracy = training(mnist_net) vizualization(test_loss, test_accuracy)
def main():
global scanning
global ser
# Creazione modello
model = training.createModel()
#TOGLIERE IL COMMENTO ALLA RIGA QUI SOTTO PER AGGIUNGERE LA FASE DI TRAINING
training.training(paintsPath, size, model)
model.load("model.xml")
#Legge l'immagine da riconoscere
[W, w] = imageUtils.read_images(picPath, size, 1)
if W is None:
print("Faccia non riconosciuta")
scanning = True
cv2.imshow('display',again)
cv2.waitKey(1)
time.sleep(4)
cv2.imshow('display',home)
cv2.waitKey(1)
ser = serial.Serial(serialPort, 9600)
return None
#Effettua la classificazione
[p_label, p_confidence] = model.predict(np.asarray(W[0]))
#mostra l'immagine
#creo un'immagine nera
bg = np.zeros(scsize, np.uint8)
vis = imageUtils.showImage(W, paintsPath, p_label, size)
#metto al centro dell'immagine nera il confronto tra volto rilevato e quadro
bg[bg.shape[0]/2-vis.shape[0]/2:bg.shape[0]/2+vis.shape[0]/2, bg.shape[1]/2-vis.shape[1]/2:bg.shape[1]/2+vis.shape[1]/2] = vis
cv2.imshow("display", bg)
#mostra le immagini di confronto per 5 secondi, poi
#chiude le finestre e si rimette in ascolto dell'arduino
cv2.waitKey(5000)
siz = cv2.cv.GetSize(cv2.cv.fromarray(vis))
#thumbnail = cv2.cv.CreateImage( ( siz[0] / 10, siz[1] / 10), 8, 1)
newsz = (siz[0]/2, siz[1]/2)
print "newsz", newsz
thumbnail=cv2.resize(vis, newsz)
# mostra in dipinto originale
filename = paintsOriginalPath+ "/" +str(p_label) +"/" + str(p_label)+ ".jpg"
im = cv2.imread(filename, cv2.CV_LOAD_IMAGE_COLOR)
#carico il confronto tra le due facce in alto a sinistra
im[0:thumbnail.shape[0], 0:thumbnail.shape[1],0] = thumbnail
im[0:thumbnail.shape[0], 0:thumbnail.shape[1],1] = thumbnail
im[0:thumbnail.shape[0], 0:thumbnail.shape[1],2] = thumbnail
cv2.imshow('display',im)
cv2.waitKey(15000)
#cv2.destroyAllWindows()
scanning = True
ser = serial.Serial(serialPort, 9600)
cv2.imshow('display',home)
cv2.waitKey(1)
'convolutional_size_params': [["conv1", [1, 10, 32], 'conv'],
["pool1", [5, 5, 5, 5], 'pool'],
["conv2", [1, 10, 32], 'conv'],
["pool2", [5, 5, 5, 5], 'pool']],
'fc_size_params': [['h_fc1', 128], ['out', 2]],
'l2':
0.001,
'batch_size':
10,
'valid_size':
10
}
training_params = {
'dropout_keep_prob':
1.0,
'max_grad':
0.01,
'epochs':
10,
'lr':
0.01,
'train_files': [
'/afs/csail.mit.edu/u/p/priyav/PAI_data/final_data/all_gis_islandviewer_iv4aa_data.csv.gz'
],
'valid_files': [
'/afs/csail.mit.edu/u/p/priyav/PAI_data/final_data/all_gis_islandviewer_iv4ag_data.csv.gz'
]
}
training(model_name, model_type, model_params, training_params)
"""
Plot the total error for L constant, and W taking values chosen in constant.py
"""
from testing import testing
from training import training
from constant import WW, L, N, N_ite, learning_rate, NB_EPOCH, id_function
import numpy as np
import matplotlib.pyplot as plt
errormax = []
errorplot = []
errormin = []
for wite in WW:
errorw = []
for n in range(N_ite):
training(wite, L, N, NB_EPOCH, learning_rate, id_function)
errorw.append(testing(wite, L, N, NB_EPOCH, learning_rate,
id_function))
means = np.mean(errorw)
sigma = np.sqrt(np.var(errorw, ddof=1))
errormax.append(means + sigma)
errorplot.append(means)
errormin.append(means - sigma)
plt.clf()
plt.figure()
fig, ax1 = plt.subplots()
ax1.set_xlabel("W", fontsize=14)
ax1.set_ylabel("Error", fontsize=14)
def planning(self, currLoc, utilObj, params):
u = currLoc
trainObj = training(self.dirName, self.ownNo)
allV = utilObj.readMap()
it = initialD(allV, currLoc)
update = 0
while update == 0:
#it.Q.remove(u)
lenPred = utilObj.findLenPredecessor(it, u)
k = lenPred + 1
self.B_curr[u] = utilObj.findNeighbor(u)
f1 = os.path.join(self.dirName,
self.base1 + str(self.ownNo) + self.suffix)
fid1 = open(f1, 'a')
outtxt1 = 'AGV: ' + str(
self.ownNo) + ' ' + 'U: ' + str(u) + ' ' + 'B[u]' + '\n'
fid1.write(outtxt1)
fid1.close()
fid1 = open(f1, 'a')
np.savetxt(fid1, self.B_curr[u], delimiter=',')
fid1.close()
#print("Neighor of u", self.B_curr[u])
if (len(self.B_curr[u]) > 0):
for e in self.B_curr[u]:
if (k <= 1):
prevTask = currLoc
elif k > 1: # rootDic, k, u, utilObj, prevTask
prevTask = it.pi_v[u]
#print("K: ", k)
#print("PrevTask: ", prevTask)
estimatedCost = trainObj.computeCost(
rootDic=self.dirName,
noAGV=self.ownNo,
k=k,
u=u,
utilObj=utilObj,
prevTask=prevTask
) #rootDic, noAGV, k, u, utilObj, prevTask
estimatedCost = estimatedCost.flatten()
f2 = os.path.join(
self.dirName,
self.base2 + str(self.ownNo) + self.suffix)
outtxt2 = 'AGV: ' + str(self.ownNo) + ' ' + 'U: ' + str(
u) + ' ' + 'to' + ' ' + 'E: ' + str(e) + ' ' + str(
estimatedCost[0]) + ' ' + str(
estimatedCost[1]) + ' ' + str(
estimatedCost[2]) + '\n'
fid2 = open(f2, 'a')
fid2.write(outtxt2)
fid2.close()
self.stateDict[k] = estimatedCost
utilObj.storeObs(k, e, estimatedCost)
sumCost = estimatedCost[0] + estimatedCost[
1] + estimatedCost[2]
utilObj.relax(u, e, it, sumCost) #correct relax
self.B_curr[e] = utilObj.findNeighbor(e)
fid1 = open(f1, 'a')
outtxt1 = 'AGV: ' + str(self.ownNo) + ' ' + 'E: ' + str(
e) + ' ' + 'B[e]' + '\n'
fid1.write(outtxt1)
fid1.close()
fid1 = open(f1, 'a')
np.savetxt(fid1, self.B_curr[e], delimiter=',')
fid1.close()
if (len(self.B_curr[e]) > 0):
for j in self.B_curr[e]:
if (j not in self.B_curr[u]) and (j != currLoc):
print("j: ", j)
lenPred = utilObj.findLenPredecessor(it, e)
k_dash = lenPred + 1
prevTask = it.pi_v[e]
#print("K_dash: ", k_dash)
#print("PrevTask: ", prevTask)
estimatedCost = trainObj.computeCost(
rootDic=self.dirName,
noAGV=self.ownNo,
k=k_dash,
u=e,
utilObj=utilObj,
prevTask=prevTask
) #(self.dirName, k, e, utilObj, prevTask)
estimatedCost = estimatedCost.flatten()
self.stateDict[k_dash] = estimatedCost
utilObj.storeObs(k_dash, j, estimatedCost)
f3 = os.path.join(
self.dirName,
self.base2 + str(self.ownNo) + self.suffix)
outtxt3 = 'AGV: ' + str(
self.ownNo) + ' ' + 'E: ' + str(
e) + ' ' + 'to' + ' ' + 'J: ' + str(
j) + ' ' + str(
estimatedCost[0]) + ' ' + str(
estimatedCost[1]
) + ' ' + str(
estimatedCost[2]) + '\n'
fid3 = open(f3, 'a')
fid3.write(outtxt3)
fid3.close()
sumCost = estimatedCost[0] + estimatedCost[
1] + estimatedCost[2]
utilObj.relax(e, j, it, sumCost)
self.B_curr[j] = utilObj.findNeighbor(j)
fid1 = open(f1, 'a')
outtxt1 = 'AGV: ' + str(
self.ownNo) + ' ' + 'J: ' + str(
j) + ' ' + 'B[j]' + '\n'
fid1.write(outtxt1)
fid1.close()
fid1 = open(f1, 'a')
np.savetxt(fid1, self.B_curr[j], delimiter=',')
fid1.close()
if (len(self.B_curr[j]) > 0):
for h in self.B_curr[j]:
if (h not in self.B_curr[u]) and (
h not in self.B_curr[e]) and (
h != currLoc):
lenPred = utilObj.findLenPredecessor(
it, h)
k_ddash = lenPred + 1
prevTask = it.pi_v[j]
#print("K_ddash: ", k_ddash)
#print("PrevTask: ", prevTask)
estimatedCost = trainObj.computeCost(
rootDic=self.dirName,
noAGV=self.ownNo,
k=k_ddash,
u=j,
utilObj=utilObj,
prevTask=prevTask
) #(self.dirName, k, j, utilObj, prevTask)
estimatedCost = estimatedCost.flatten(
)
self.stateDict[
k_ddash] = estimatedCost
utilObj.storeObs(
k_ddash, h, estimatedCost)
f4 = os.path.join(
self.dirName, self.base2 +
str(self.ownNo) + self.suffix)
outtxt4 = 'AGV: ' + str(
self.ownNo
) + ' ' + 'J: ' + str(
j
) + ' ' + 'to' + ' ' + 'H: ' + str(
h) + ' ' + str(
estimatedCost[0]
) + ' ' + str(
estimatedCost[1]
) + ' ' + str(
estimatedCost[2]) + '\n'
fid4 = open(f4, 'a')
fid4.write(outtxt4)
fid4.close()
sumCost = estimatedCost[
0] + estimatedCost[
1] + estimatedCost[2]
utilObj.relax(j, h, it, sumCost)
self.B_curr[
h] = utilObj.findNeighbor(h)
fid1 = open(f1, 'a')
outtxt1 = 'AGV: ' + str(
self.ownNo
) + ' ' + 'H: ' + str(
h) + ' ' + 'B[h]' + '\n'
fid1.write(outtxt1)
fid1.close()
fid1 = open(f1, 'a')
np.savetxt(fid1,
self.B_curr[h],
delimiter=',')
fid1.close()
if (len(self.B_curr[h]) > 0):
for l in self.B_curr[h]:
if (l not in self.B_curr[u]
) and (l not in self.
B_curr[e]) and (
l
not in self.
B_curr[j]
) and (l !=
currLoc):
lenPred = utilObj.findLenPredecessor(
it, h)
k_dddash = lenPred + 1
prevTask = it.pi_v[h]
#print("K_dddash: ", k_dddash)
#print("PrevTask: ", prevTask)
estimatedCost = trainObj.computeCost(
rootDic=self.
dirName,
noAGV=self.ownNo,
k=k_dddash,
u=h,
utilObj=utilObj,
prevTask=prevTask
) #(self.dirName, k, h, utilObj, prevTask)
estimatedCost = estimatedCost.flatten(
)
self.stateDict[
k_dddash] = estimatedCost
utilObj.storeObs(
k_dddash, l,
estimatedCost)
f5 = os.path.join(
self.dirName,
self.base2 +
str(self.ownNo) +
self.suffix)
outtxt5 = 'AGV: ' + str(
self.ownNo
) + ' ' + 'H: ' + str(
h
) + ' ' + 'to' + ' ' + 'L: ' + str(
l
) + ' ' + str(
estimatedCost[0]
) + ' ' + str(
estimatedCost[1]
) + ' ' + str(
estimatedCost[2]
) + '\n'
fid5 = open(f5, 'a')
fid5.write(outtxt5)
fid5.close()
sumCost = estimatedCost[
0] + estimatedCost[
1] + estimatedCost[
2]
utilObj.relax(
h, l, it, sumCost)
else:
break
else:
break
else:
break
else:
break
self.taskSequence = utilObj.findTaskPath(u, it)
print("Tasksequence", self.taskSequence)
f6 = os.path.join(self.dirName,
self.base3 + str(self.ownNo) + self.suffix)
fid6 = open(f6, 'a')
# self.fid1.write(outtxt1)
np.savetxt(fid6, self.taskSequence, delimiter=',')
fid6.close()
self.lenT = len(self.taskSequence)
self.endTask = self.taskSequence[self.lenT - 1]
update = 1
print("Len of stateDict", len(self.stateDict))
return self.stateDict, self.lenT