def process_dir(root_dir, subdir, template):
curr_dir = os.path.join(root_dir, subdir)
# Look for template in current dir
template_file = os.path.join(curr_dir, config.TEMPLATE_FILE)
if os.path.exists(template_file):
template = Template(template_file)
# look for images in current dir to process
paths = config.Paths(os.path.join(args['output_dir'], subdir))
exts = ('*.png', '*.jpg')
omr_files = sorted(
[f for ext in exts for f in glob(os.path.join(curr_dir, ext))])
# Exclude marker image if exists
if (template and template.marker_path):
omr_files = [f for f in omr_files if f != template.marker_path]
print("\n\n\nMarker Found\n\n\n")
subfolders = sorted([
file for file in os.listdir(curr_dir)
if os.path.isdir(os.path.join(curr_dir, file))
])
if omr_files:
args_local = args.copy()
if ("OverrideFlags" in template.options):
args_local.update(template.options["OverrideFlags"])
print(
'\n------------------------------------------------------------------'
)
print(f'Processing directory {curr_dir} with settings- ')
print("\tTotal images : %d" % (len(omr_files)))
print("\tCropping Enabled : " + str(not args_local["noCropping"]))
print("\tAuto Alignment : " + str(args_local["autoAlign"]))
print("\tUsing Template : " +
str(template.path) if (template) else "N/A")
print("\tUsing Marker : " + str(template.marker_path) if (
template.marker is not None) else "N/A")
print('')
if not template:
print(f'Error: No template file when processing {curr_dir}.')
print(
f' Place {config.TEMPLATE_FILE} in the directory or specify a template using -t.'
)
return
utils.setup_dirs(paths)
output_set = setup_output(paths, template)
return process_files(omr_files, template, args_local, output_set)
elif (len(subfolders) == 0):
# the directory should have images or be non-leaf
print(f'Note: No valid images or subfolders found in {curr_dir}')
# recursively process subfolders
results_lists = []
for folder in subfolders:
results_lists.append(
process_dir(root_dir, os.path.join(subdir, folder), template))
return results_lists
def main():
paths = config.Paths()
start_time = time.time()
np.random.seed(0)
evaluate(paths)
# test(paths)
print('Time elapsed: {}s'.format(time.time() - start_time))
def main():
paths = config.Paths()
scene = 'city'
scale = 0.1
# create_labels(paths)
process_ply(paths, scene, scale) # each unit is 10 centimeters
def test_valid():
paths = config.Paths()
grammar_file = os.path.join(paths.tmp_root, 'grammar', 'cad',
'stacking_objects.pcfg')
# sentence = 'null reaching moving placing'
# grammar = grammarutils.read_grammar(grammar_file, index=False)
# test_earley(grammar, sentence.split())
sentence = 'null reaching'
tokens = sentence.split()
grammar = grammarutils.read_grammar(
grammar_file,
index=True,
mapping=datasets.cad_metadata.subactivity_index)
seg_length = 15
correct_prob = 0.8
classifier_output = np.ones((seg_length * 2, 10)) * 1e-10
classifier_output[:seg_length, datasets.cad_metadata.
subactivity_index[tokens[0]]] = correct_prob
classifier_output[
seg_length:,
datasets.cad_metadata.subactivity_index[tokens[1]]] = correct_prob
classifier_output[:seg_length,
datasets.cad_metadata.subactivity_index[tokens[0]] +
1] = 1 - correct_prob
classifier_output[seg_length:,
datasets.cad_metadata.subactivity_index[tokens[1]] +
1] = 1 - correct_prob
test_generalized_earley(grammar, classifier_output)
def main():
paths = config.Paths()
start_time = time.time()
# induce_activity_grammar(paths)
# read_induced_grammar(paths)
test(paths)
print('Time elapsed: {}'.format(time.time() - start_time))
def save_model(cls, iteration, performance):
if not cls.test:
paths = config.Paths()
cls.saver.save(
cls.session,
os.path.join(
paths.tmp_root, 'rnn', cls.timestamp,
'model_{:07d}_{:03d}_{:03d}.ckpt'.format(
iteration, performance[0], performance[1])))
def main():
paths = config.Paths()
scene = 'city'
height_threshold = 1
start_time = time.time()
# compute_all_cost_maps(paths, scene, height_threshold)
# print('Cost maps computed. Time elapsed: {}s'.format(time.time() - start_time))
path_planning_demo(paths, scene, height_threshold)
print('Time elapsed: {}s'.format(time.time() - start_time))
def save_model(cls, iteration, performance):
paths = config.Paths()
np.save(
os.path.join(paths.tmp_root, 'theta', cls.timestamp,
'theta_{:07d}'.format(iteration)), cls.theta)
with open(
os.path.join(paths.tmp_root, 'theta', cls.timestamp,
'theta_{:07d}.csv'.format(iteration)), 'w') as f:
for agent_theta in cls.theta:
np.savetxt(f, agent_theta, fmt='%.5f', footer='====')
np.savetxt(f, performance, fmt='%d')
np.savetxt(f, [cls.r_bar], fmt='%.5f')
def main():
paths = config.Paths()
paths.path_huang()
# model_path = paths.metadata_root + 'models/flipped/layer_3_without_dropout_maxlen_200_epoch_200.h5'
# model_path_per_frame = paths.metadata_root + 'models/flipped/layer_3_per_frame_without_dropout_epoch_1000.h5'
# without_segmentation_sequence_test(paths.metadata_root + 'flipped/', model_path, 200)
# without_segmentation_sequence_test_per_frame(paths.metadata_root + 'flipped/', model_path_per_frame)
# without_segmentation_sequence_test_per_frame_vgg16(paths.metadata_root)
# get_bottleneck_feature(paths.metadata_root)
# without_segmentation_sequence_test_per_frame_sequential(paths.metadata_root)
without_segmentation_sequence_test_per_frame_lstm(paths.data_root,
paths.metadata_root)
def visualize_grammar():
paths = config.Paths()
dataset_name = 'wnp'
for pcfg in os.listdir(os.path.join(paths.tmp_root, 'grammar', dataset_name)):
if not pcfg.endswith('.pcfg'):
continue
grammar_file = os.path.join(paths.tmp_root, 'grammar', dataset_name, pcfg)
grammar = grammarutils.read_grammar(grammar_file, insert=False)
dot_filename = os.path.join(paths.tmp_root, 'visualize', 'grammar', dataset_name, pcfg.replace('.pcfg', '.dot'))
pdf_filename = os.path.join(paths.tmp_root, 'visualize', 'grammar', dataset_name, pcfg.replace('.pcfg', '.pdf'))
grammarutils.grammar_to_dot(grammar, dot_filename)
os.system('dot -Tpdf {} -o {}'.format(dot_filename, pdf_filename))
def test_time():
paths = config.Paths()
start_time = time.time()
np.random.seed(int(start_time))
classifier_output = np.random.rand(100000, 10)
classifier_output = classifier_output / np.sum(classifier_output, axis=1)[:, None] # Normalize to probability
for pcfg in os.listdir(os.path.join(paths.tmp_root, 'grammar', 'cad')):
if not pcfg.endswith('.pcfg'):
continue
grammar_file = os.path.join(paths.tmp_root, 'grammar', 'cad', pcfg)
grammar = grammarutils.read_grammar(grammar_file, index=True, mapping=datasets.cad_metadata.subactivity_index)
test_generalized_earley(grammar, classifier_output)
print('Time elapsed: {}s'.format(time.time() - start_time))
def init_lstm(cls):
cls.max_goal_size = cls.static_goal_num + cls.max_people_tracked
cls.state_size = 2 * (cls.total_drone_num + cls.max_people_tracked - 1)
cls.input_size = cls.state_size + cls.max_goal_size
cls.inputs = tf.placeholder(tf.float32, [None, None, cls.input_size],
name='input')
lstm = tf.contrib.rnn.LSTMCell(num_units=RNNConfig.h_size,
state_is_tuple=True)
batch_size = tf.shape(cls.inputs)[1]
initial_state = lstm.zero_state(batch_size, tf.float32)
rnn_outputs, rnn_states = tf.nn.dynamic_rnn(
lstm, cls.inputs, initial_state=initial_state, time_major=True)
cls.q_values_tf = tf.contrib.layers.fully_connected(
rnn_outputs, num_outputs=128, activation_fn=tf.nn.relu)
cls.q_values_tf = tf.contrib.layers.fully_connected(
cls.q_values_tf, num_outputs=32, activation_fn=tf.nn.relu)
# cls.q_values_tf = tf.contrib.layers.fully_connected(cls.q_values_tf, num_outputs=1, activation_fn=tf.nn.tanh)
cls.q_values_tf = tf.contrib.layers.fully_connected(cls.q_values_tf,
num_outputs=1,
activation_fn=None)
cls.delta_tf = tf.placeholder(tf.float32, name='delta')
optimizer = tf.train.GradientDescentOptimizer(
learning_rate=RNNConfig.alpha)
grads_and_vars = optimizer.compute_gradients(cls.q_values_tf)
grads_and_vars = [(-gv[0] * cls.delta_tf, gv[1])
for gv in grads_and_vars]
cls.train = optimizer.apply_gradients(grads_and_vars)
# # Direct approximation of reward
# cls.r_tf = tf.placeholder(tf.float32, name='r')
# loss = tf.square(cls.q_values_tf - cls.r_tf)
# optimizer = tf.train.GradientDescentOptimizer(learning_rate=RNNConfig.alpha)
# grads_and_vars = optimizer.compute_gradients(loss)
# cls.train = optimizer.apply_gradients(grads_and_vars)
cls.session = tf.Session()
cls.session.run(tf.global_variables_initializer())
# file_writer = tf.summary.FileWriter('/home/siyuan/Downloads/test/log', cls.session.graph)
cls.saver = tf.train.Saver()
if cls.test:
cls.saver.restore(
cls.session,
os.path.join(config.Paths().tmp_root, 'rnn',
'2017-09-08 18:24:57',
'model_0190000_249_1016.ckpt'),
)
def test_grammar():
paths = config.Paths()
for pcfg in os.listdir(os.path.join(paths.tmp_root, 'grammar', 'cad')):
if not pcfg.endswith('.pcfg'):
continue
grammar_file = os.path.join(paths.tmp_root, 'grammar', 'cad', pcfg)
grammar = grammarutils.read_grammar(grammar_file, index=True, mapping=datasets.cad_metadata.subactivity_index)
corpus_file = os.path.join(paths.tmp_root, 'corpus', 'cad', pcfg.replace('pcfg', 'txt'))
with open(corpus_file, 'r') as f:
for line in f:
tokens = [str(datasets.cad_metadata.subactivity_index[token]) for token in line.strip(' *#\n').split(' ')]
earley_parser = nltk.EarleyChartParser(grammar, trace=0)
e_chart = earley_parser.chart_parse(tokens)
print(e_chart.edges()[-1])
def parse_arguments():
# Parser check
def restricted_float(x, inter):
x = float(x)
if x < inter[0] or x > inter[1]:
raise argparse.ArgumentTypeError("%r not in range [1e-5, 1e-4]"%(x,))
return x
paths = config.Paths()
feature_type = 'resnet'
# Path settings
parser = argparse.ArgumentParser(description='VCOCO dataset')
parser.add_argument('--project-root', default=paths.project_root, help='intermediate result path')
parser.add_argument('--tmp-root', default=paths.tmp_root, help='intermediate result path')
parser.add_argument('--data-root', default=paths.vcoco_data_root, help='data path')
parser.add_argument('--log-root', default=os.path.join(paths.log_root, 'vcoco/parsing_{}'.format(feature_type)), help='log files path')
parser.add_argument('--resume', default=os.path.join(paths.tmp_root, 'checkpoints/vcoco/parsing_{}'.format(feature_type)), help='path to latest checkpoint')
parser.add_argument('--eval-root', default=os.path.join(paths.tmp_root, 'evaluation/vcoco/{}'.format(feature_type)), help='path to save evaluation file')
parser.add_argument('--feature-type', default=feature_type, help='feature_type')
parser.add_argument('--visualize', action='store_true', default=False, help='Visualize final results')
parser.add_argument('--vis-top-k', type=int, default=1, metavar='N', help='Top k results to visualize')
# Optimization Options
parser.add_argument('--batch-size', type=int, default=1, metavar='N',
help='Input batch size for training (default: 10)')
parser.add_argument('--no-cuda', action='store_true', default=False,
help='Enables CUDA training')
parser.add_argument('--epochs', type=int, default=30, metavar='N',
help='Number of epochs to train (default: 10)')
parser.add_argument('--start-epoch', type=int, default=0, metavar='N',
help='Index of epoch to start (default: 0)')
parser.add_argument('--link-weight', type=float, default=2, metavar='N',
help='Loss weight of existing edges')
parser.add_argument('--lr', type=lambda x: restricted_float(x, [1e-5, 1e-2]), default=1e-4, metavar='LR',
help='Initial learning rate [1e-5, 1e-2] (default: 1e-3)')
parser.add_argument('--lr-decay', type=lambda x: restricted_float(x, [.01, 1]), default=0.8, metavar='LR-DECAY',
help='Learning rate decay factor [.01, 1] (default: 0.8)')
parser.add_argument('--momentum', type=float, default=0.9, metavar='M',
help='SGD momentum (default: 0.9)')
# i/o
parser.add_argument('--log-interval', type=int, default=100, metavar='N',
help='How many batches to wait before logging training status')
# Accelerating
parser.add_argument('--prefetch', type=int, default=0, help='Pre-fetching threads.')
return parser.parse_args()
def process_dir(root_dir, subdir, template, kesme_islemi, onizleme):
curr_dir = os.path.join(root_dir, subdir)
args['noCropping'] = bool(kesme_islemi)
# Look for template in current dir
template_file = os.path.join(curr_dir, config.TEMPLATE_FILE)
if os.path.exists(template_file):
template = Template(template_file)
# look for images in current dir to process
paths = config.Paths(os.path.join(args['output_dir'], subdir))
exts = ('*.png', '*.jpg')
omr_files = sorted(
[f for ext in exts for f in glob(os.path.join(curr_dir, ext))])
# Exclude marker image if exists
if(template and template.marker_path):
omr_files = [f for f in omr_files if f != template.marker_path]
subfolders = sorted([file for file in os.listdir(
curr_dir) if os.path.isdir(os.path.join(curr_dir, file))])
if omr_files:
args_local = args.copy()
if("OverrideFlags" in template.options):
args_local.update(template.options["OverrideFlags"])
print('\n------------------------------------------------------------------')
print(f'"{curr_dir}" dizini ayarları ile birlikte işleniyor- ')
print("\tToplan resim : %d" % (len(omr_files)))
print("\tKırpma Aktif : " + str(not args_local["noCropping"]))
print("\tOtomatik Hizalama : " + str(args_local["autoAlign"]))
print("\tKullanılan Şablon : " + str(template.path) if(template) else "N/A")
print("\tKullanılan İşaretçi : " + str(template.marker_path)
if(template.marker is not None) else "N/A")
print('')
if not template:
print(f'Hata: işlenirken şablon bulunamadı {curr_dir}.')
print(f' Çalışma dizininde {config.TEMPLATE_FILE} dosyasını konumlandırın veya -t parametresi ile birlikte belirtin.')
return
utils.setup_dirs(paths)
output_set = setup_output(paths, template)
process_files(omr_files, template, args_local, output_set, onizleme)
elif(len(subfolders) == 0):
# the directory should have images or be non-leaf
print(f'Bilgi: {curr_dir} klasöründe geçerli bir resim veya alt klasör bulunamadı.')
# recursively process subfolders
for folder in subfolders:
process_dir(root_dir, os.path.join(subdir, folder), template, kesme_islemi, onizleme)
def main():
paths = config.Paths()
start_time = time.time()
seed = int(time.time())
print 'seed:', seed
np.random.seed(0)
evaluate(paths)
# supplementary.rgb_videos(paths)
# plyutils.create_labels(paths)
# cal_all_affordance(paths)
# planning(paths, seed)
print('Time elapsed: {}'.format(time.time() - start_time))
def main():
paths = config.Paths()
paths.path_huang()
model_path = paths.metadata_root + 'models/affordance/'
cache_path = paths.metadata_root + 'caches/affordance/'
if not os.path.exists(model_path):
os.mkdir(model_path)
if not os.path.exists(cache_path):
os.mkdir(cache_path)
np.random.seed(1337)
print('Loading data')
object_feature, affordance_feature, distance_feature, orientation_feature = load_affordance_data(paths.metadata_root)
x_1_train, x_2_train, y_train, x_1_test, x_2_test, y_test = data_splitting(object_feature, affordance_feature, distance_feature, orientation_feature)
print('Building model')
model_dir = train(x_1_train, x_2_train, y_train, x_1_test, x_2_test, y_test, model_path)
result_output(x_1_train, x_2_train, y_train, x_1_test, x_2_test, y_test, model_dir, cache_path)
def main(opt):
train_datalist = '/home/mcislab/wangruiqi/IJCV2019/data/ucf101Vid_all_lin.txt'
#train_datalist = '/home/mcislab/wangruiqi/IJCV2019/data/test.txt'
save_path = '/media/mcislab/new_ssd/wrq/data/UCF101/res18_rgbflow_same_similiarity/'
paths = config.Paths()
train_dataset = UCF101dataset_creat.dataset(
train_datalist, paths.detect_root_ucf_mmdet, paths.img_root_ucf,
paths.rgb_res18_ucf, paths.rgb_res18_ucf, opt) #rgb_bninc_ucf
train_dataloader = DataLoader(train_dataset,
batch_size=opt.batch_size,
shuffle=False,
num_workers=opt.workers,
drop_last=False)
print('start to read feature...')
count = 0
for i, (edge_features, node_features, adj_mat, graph_label, tem_features,
video_info) in enumerate(train_dataloader, 0):
#print (video_info)
videoinfo = video_info[0]
category, video_name = videoinfo.split('/')
edge_features = edge_features.squeeze(0)
node_features = node_features.squeeze(0)
adj_mat = adj_mat.squeeze(0)
graph_label = graph_label.squeeze(0)
tem_features = tem_features.squeeze(0)
#print (category)
#print (video_name)
#print (edge_features.shape)
feat_path = os.path.join(save_path, category)
if not os.path.exists(feat_path):
os.makedirs(feat_path)
feat_dict = {
'edge_features': edge_features,
'node_features': node_features,
'adj_mat': adj_mat,
'graph_label': graph_label,
'tem_features': tem_features,
'video_info': video_info
}
#print(edge_features.shape)
print(feat_path + '/' + video_name[:-4] + '.pkl')
file = open(feat_path + '/' + video_name[:-4] + '.pkl', 'wb')
pickle.dump(feat_dict, file)
count += 1
print('Has tranformed %d input features. Finished!' % count)
def __init__(self, root, input_imsize, transform, imageset):
self.imageset = 'train' if imageset == 'train' else 'test'
# result for deformable convnet? img feature
anno_file = os.path.join(root, '{}_annotations.mat'.format(imageset))
ld = sio.loadmat(anno_file)
gt_anno = ld['gt_all']
data = dict()
data['img_ids'] = list()
data['bbxs'] = list()
data['actions'] = list()
for hoi_idx, hoi in enumerate(gt_anno):
for img_idx, bboxes in enumerate(hoi):
if bboxes.size != 0:
for row in bboxes:
data['img_ids'].append(img_idx)
data['bbxs'].append(row)
data['actions'].append(metadata.hoi_to_action[hoi_idx])
print('finished for ' + str(hoi_idx))
np.save("anno_tune.npy", data)
self.hico_path = root
self.imsize = input_imsize
self.transform = transform
image_list = list()
with open(
os.path.join(config.Paths().project_root,
'{}_all.txt'.format(imageset))) as f:
for line in f.readlines():
image_list.append(line.strip())
self.img_files = [image_list[x] for x in data['img_ids']]
self.bbxs = data['bbxs']
self.actions = data['actions']
def main():
paths = config.Paths()
# prepare_evaluation(paths)
# evaluate(paths)
qualitative_result(paths)
def main():
paths = config.Paths()
# plot_affordance(paths)
plot_sample(paths)
def main():
paths = config.Paths()
paths.path_huang()
load_affordance_data(paths.metadata_root)
def parse_arguments():
# Parser check
def restricted_float(x, inter):
x = float(x)
if x < inter[0] or x > inter[1]:
raise argparse.ArgumentTypeError("%r not in range [1e-5, 1e-4]" %
(x, ))
return x
paths = config.Paths()
feature_type = 'resnet'
# Path settings
parser = argparse.ArgumentParser(description='VCOCO dataset')
parser.add_argument('--project-root',
default=paths.project_root,
help='intermediate result path')
parser.add_argument('--tmp-root',
default=paths.tmp_root,
help='intermediate result path')
parser.add_argument('--data-root',
default=paths.vcoco_data_root,
help='data path')
parser.add_argument('--log-root',
default=os.path.join(
paths.log_root,
'vcoco/parsing_{}'.format(feature_type)),
help='log files path')
parser.add_argument(
'--resume',
default=os.path.join(
paths.tmp_root,
'checkpoints/vcoco/parsing_{}'.format(feature_type)),
help='path to latest checkpoint')
parser.add_argument('--eval-root',
default=os.path.join(
paths.tmp_root,
'evaluation/vcoco/{}'.format(feature_type)),
help='path to save evaluation file')
parser.add_argument('--feature-type',
default=feature_type,
help='feature_type')
parser.add_argument('--visualize',
action='store_true',
default=False,
help='Visualize final results')
parser.add_argument('--vis-top-k',
type=int,
default=1,
metavar='N',
help='Top k results to visualize')
# Model parameters
parser.add_argument('--prop-layer',
type=int,
default=3,
metavar='N',
help='Number of propogation layers (default: 3)')
parser.add_argument('--update-layer',
type=int,
default=1,
metavar='N',
help='Number of update hidden layers (default: 1)')
parser.add_argument('--link-layer',
type=int,
default=3,
metavar='N',
help='Number of link hidden layers (default: 3)')
parser.add_argument('--model-type',
type=str,
default='V1',
help='GPNN model type')
parser.add_argument('--po-type',
type=str,
default='mult',
help='Part-Object prior type')
parser.add_argument('--suppress-hh',
action='store_true',
default=False,
help='Suppresses human-human edges')
# Optimization Options
parser.add_argument('--batch-size',
type=int,
default=1,
metavar='N',
help='Input batch size for training (default: 10)')
parser.add_argument('--no-cuda',
action='store_true',
default=False,
help='Enables CUDA training')
parser.add_argument('--epochs',
type=int,
default=30,
metavar='N',
help='Number of epochs to train (default: 10)')
parser.add_argument('--start-epoch',
type=int,
default=0,
metavar='N',
help='Index of epoch to start (default: 0)')
parser.add_argument('--link-weight',
type=float,
default=2,
metavar='N',
help='Loss weight of existing edges')
parser.add_argument(
'--lr',
type=lambda x: restricted_float(x, [1e-5, 1e-2]),
default=1e-4,
metavar='LR',
help='Initial learning rate [1e-5, 1e-2] (default: 1e-3)')
parser.add_argument(
'--lr-decay',
type=lambda x: restricted_float(x, [.01, 1]),
default=0.8,
metavar='LR-DECAY',
help='Learning rate decay factor [.01, 1] (default: 0.8)')
parser.add_argument('--momentum',
type=float,
default=0.9,
metavar='M',
help='SGD momentum (default: 0.9)')
# i/o
parser.add_argument(
'--log-interval',
type=int,
default=10,
metavar='N',
help='How many batches to wait before logging training status')
# Accelerating
parser.add_argument('--prefetch',
type=int,
default=0,
help='Pre-fetching threads.')
# Debug
parser.add_argument('--debug',
action='store_true',
default=False,
help='Runs only 10 batch and 1 epoch')
# Feature
parser.add_argument('--extra-feature',
action='store_true',
default=False,
help='Include 101-d extra node feature')
# Negative Rejection
parser.add_argument('--NRT',
action='store_true',
default=False,
help='Reject negative object-action pairs in training')
parser.add_argument(
'--NRE',
action='store_true',
default=False,
help='Reject negative object-action pairs in evaluation')
# AdjMat Loss
parser.add_argument('--ignore-adj',
action='store_true',
default=False,
help='ignores adjacency matrix in loss function')
return parser.parse_args()
Sample Human from the Scene
"""
import config
import os
import pickle
from scipy.io import loadmat
import numpy as np
import matplotlib.pyplot as plt
import visualize
import copy
from camera import rotation_matrix_3d_z, center_to_corners
import random
import sys
from PIL import Image
paths = config.Paths()
metadata_root = paths.metadata_root
proposal_root = paths.proposal_root
evaluation_root = os.path.join(metadata_root, 'evaluation')
stats_root = os.path.join(metadata_root, 'sunrgbd_stats')
office_label = ['reading', 'put-down-item', 'take-item', 'play-computer']
class HumanSample(object):
def __init__(self, pg, skeleton_stats, save_root='.', if_vis=True):
self.pg = copy.deepcopy(pg)
self.skeleton_stats = skeleton_stats
self.learning_rate = 1
self.pg_output = copy.deepcopy(pg)
self.skeleton = list()
self.adjust_index = list()
class RNNAgent(object):
total_drone_num = 0
max_people_tracked = 4
static_goal_num = 0
r_bar = 0
lstm = None
session = None
train = None
saver = None
state_size = 0
input_size = 0
max_goal_size = 0
inputs = None
q_values_tf = None
r_tf = None
delta_tf = None
test = True
if not test:
timestamp = datetime.datetime.fromtimestamp(
time.time()).strftime('%Y-%m-%d %H:%M:%S')
os.mkdir(os.path.join(config.Paths().tmp_root, 'rnn', timestamp))
def __init__(self, index, handle):
self.__class__.total_drone_num += 1
self._index = index
self._handle = handle
self.epsilon = 0.0 if self.__class__.test else 0.1
self.last_state = None
self.last_q = 0
self.last_action = -1
@classmethod
def set_static_goals(cls, goals):
cls.static_goal_num = len(goals[0]) + len(goals[1]) + len(goals[2])
@classmethod
def init_lstm(cls):
cls.max_goal_size = cls.static_goal_num + cls.max_people_tracked
cls.state_size = 2 * (cls.total_drone_num + cls.max_people_tracked - 1)
cls.input_size = cls.state_size + cls.max_goal_size
cls.inputs = tf.placeholder(tf.float32, [None, None, cls.input_size],
name='input')
lstm = tf.contrib.rnn.LSTMCell(num_units=RNNConfig.h_size,
state_is_tuple=True)
batch_size = tf.shape(cls.inputs)[1]
initial_state = lstm.zero_state(batch_size, tf.float32)
rnn_outputs, rnn_states = tf.nn.dynamic_rnn(
lstm, cls.inputs, initial_state=initial_state, time_major=True)
cls.q_values_tf = tf.contrib.layers.fully_connected(
rnn_outputs, num_outputs=128, activation_fn=tf.nn.relu)
cls.q_values_tf = tf.contrib.layers.fully_connected(
cls.q_values_tf, num_outputs=32, activation_fn=tf.nn.relu)
# cls.q_values_tf = tf.contrib.layers.fully_connected(cls.q_values_tf, num_outputs=1, activation_fn=tf.nn.tanh)
cls.q_values_tf = tf.contrib.layers.fully_connected(cls.q_values_tf,
num_outputs=1,
activation_fn=None)
cls.delta_tf = tf.placeholder(tf.float32, name='delta')
optimizer = tf.train.GradientDescentOptimizer(
learning_rate=RNNConfig.alpha)
grads_and_vars = optimizer.compute_gradients(cls.q_values_tf)
grads_and_vars = [(-gv[0] * cls.delta_tf, gv[1])
for gv in grads_and_vars]
cls.train = optimizer.apply_gradients(grads_and_vars)
# # Direct approximation of reward
# cls.r_tf = tf.placeholder(tf.float32, name='r')
# loss = tf.square(cls.q_values_tf - cls.r_tf)
# optimizer = tf.train.GradientDescentOptimizer(learning_rate=RNNConfig.alpha)
# grads_and_vars = optimizer.compute_gradients(loss)
# cls.train = optimizer.apply_gradients(grads_and_vars)
cls.session = tf.Session()
cls.session.run(tf.global_variables_initializer())
# file_writer = tf.summary.FileWriter('/home/siyuan/Downloads/test/log', cls.session.graph)
cls.saver = tf.train.Saver()
if cls.test:
cls.saver.restore(
cls.session,
os.path.join(config.Paths().tmp_root, 'rnn',
'2017-09-08 18:24:57',
'model_0190000_249_1016.ckpt'),
)
def act(self, goals, state, reward, episode_done=False):
if reward == 0 and np.random.randint(
500) != 1 and self.last_action != -1:
return self.last_action
goals_list = [
k for i in range(len(goals)) for k, v in goals[i].items()
]
agent_state = state[1][:]
# Swap the current agent's position to the first
agent_state[0], agent_state[self._index *
2] = agent_state[self._index *
2], agent_state[0]
agent_state[1], agent_state[self._index * 2 +
1] = agent_state[self._index * 2 +
1], agent_state[1]
lstm_input = np.zeros((1, 1, self.__class__.input_size))
if len(agent_state) < self.__class__.state_size:
lstm_input[0, 0, :len(agent_state)] = np.array(agent_state)
else:
lstm_input[0, 0, :self.__class__.state_size] = np.array(
agent_state)[:self.__class__.state_size]
lstm_input /= 210.
# RNN outputs Q(s, a)
q_values = np.zeros(len(goals_list)) if len(
goals_list) < self.__class__.max_goal_size else np.zeros(
self.__class__.max_goal_size)
for i in range(len(q_values)):
lstm_input[0, 0, self.__class__.state_size:] = 0
lstm_input[0, 0, self.__class__.state_size + i] = 1
q_values[i] = self.__class__.session.run(
self.__class__.q_values_tf,
{self.__class__.inputs: lstm_input})[0][0]
# Epsilon Greedy
if np.random.random() < self.epsilon:
a = np.random.choice(len(q_values))
else:
a = random.choice(
np.argwhere(q_values == np.max(q_values)).flatten())
# lstm_input[0, 0, -1] = goals_list[a]
lstm_input[0, 0, self.__class__.state_size:] = 0
lstm_input[0, 0, self.__class__.state_size + a] = 1
# # Boltzmann distribution
# softmax_q_values = np.exp(q_values)
# if np.isinf(sum(softmax_q_values)):
# softmax_q_values = [1./len(q_values) for _ in range(len(q_values))]
# else:
# softmax_q_values /= sum(softmax_q_values)
# a = np.random.choice(len(softmax_q_values), p=softmax_q_values)
# lstm_input[0, 0, -1] = goals_list[a]
# Update network
if reward != 0 and self.epsilon != 0:
# delta = reward - self.__class__.r_bar + q_values[a] - self.last_q # Continuing
if episode_done:
print 'Episode done!'
delta = reward - q_values[a]
else:
delta = reward - RNNConfig.gamma * q_values[
a] - self.last_q # Episodic
self.__class__.r_bar += RNNConfig.beta * delta
self.__class__.session.run(
self.__class__.train, {
self.__class__.inputs: self.last_state,
self.__class__.delta_tf: delta
})
# self.__class__.session.run(self.__class__.train, {self.__class__.inputs: self.last_state, self.__class__.r_tf: reward})
print 'reward', reward, 'r_bar', self.__class__.r_bar, 'a', a, 'q', q_values[
a], 'last_q', self.last_q, 'delta', delta, self.last_state
print q_values
self.last_state = lstm_input
self.last_q = q_values[a]
self.last_action = goals_list[a]
print self._index, 'choose action', self.last_action
return goals_list[a]
@classmethod
def save_model(cls, iteration, performance):
if not cls.test:
paths = config.Paths()
cls.saver.save(
cls.session,
os.path.join(
paths.tmp_root, 'rnn', cls.timestamp,
'model_{:07d}_{:03d}_{:03d}.ckpt'.format(
iteration, performance[0], performance[1])))
def main():
paths = config.Paths()
start_time = time.time()
plot_demo(paths)
# rgb_videos(paths)
print('Time elapsed: {}'.format(time.time() - start_time))
def main():
paths = config.Paths()
start_time = time.time()
collect_data(paths)
print('Time elapsed: {}s'.format(time.time() - start_time))
def main():
paths = config.Paths()
# learn_furniture_wall_relationships(paths)
learn_affordance_distributions(paths)
def parse_arguments():
# Parser check
def restricted_float(x, inter):
x = float(x)
if x < inter[0] or x > inter[1]:
raise argparse.ArgumentTypeError("%r not in range [1e-5, 1e-4]" %
(x, ))
return x
paths = config.Paths()
# Path settings
parser = argparse.ArgumentParser(description='CAD 120 dataset')
parser.add_argument('--project-root',
default=paths.project_root,
help='intermediate result path')
parser.add_argument('--tmp-root',
default=paths.tmp_root,
help='intermediate result path')
parser.add_argument('--log-root',
default=os.path.join(paths.log_root,
'cad120/prediction'),
help='log files path')
parser.add_argument('--resume',
default=os.path.join(paths.tmp_root,
'checkpoints/cad120/prediction'),
help='path to latest checkpoint')
parser.add_argument('--visualize',
action='store_true',
default=True,
help='Visualize final results')
# Optimization Options
parser.add_argument('--batch-size',
type=int,
default=1,
metavar='N',
help='Input batch size for training (default: 10)')
parser.add_argument('--no-cuda',
action='store_true',
default=False,
help='Enables CUDA training')
parser.add_argument('--epochs',
type=int,
default=0,
metavar='N',
help='Number of epochs to train (default: 10)')
parser.add_argument('--start-epoch',
type=int,
default=0,
metavar='N',
help='Index of epoch to start (default: 0)')
parser.add_argument(
'--lr',
type=lambda x: restricted_float(x, [1e-5, 1e-2]),
default=5e-5,
metavar='LR',
help='Initial learning rate [1e-5, 1e-2] (default: 1e-3)')
parser.add_argument(
'--lr-decay',
type=lambda x: restricted_float(x, [.01, 1]),
default=0.8,
metavar='LR-DECAY',
help='Learning rate decay factor [.01, 1] (default: 0.8)')
parser.add_argument('--momentum',
type=float,
default=0.9,
metavar='M',
help='SGD momentum (default: 0.9)')
# i/o
parser.add_argument(
'--log-interval',
type=int,
default=10,
metavar='N',
help='How many batches to wait before logging training status')
# Accelerating
parser.add_argument('--prefetch',
type=int,
default=0,
help='Pre-fetching threads.')
return parser.parse_args()
def main():
paths = config.Paths()
paths.path_huang()
without_segmentation_sequence_test_per_frame_sequential(
paths.data_root, paths.metadata_root)