def tier_level_elm(elm,tier_levels):
"""tier_levels=[bool,bool]"""
if elm.og_index in break_points: st()
cd = elm.call_data
t1,t2,t3 = fmt_syms.get_Tier1() if tier_levels[0] else fmt_syms.get_Tier2()
tier = 1 if tier_levels[0] else 2
if isinstance(cd,SimpleFunkWithArgs):
return (
simp_funk_w_args(elm,sym1=t1,sym2=t2,c=True),
simp_funk_w_args(elm,sym1=t1,sym2=t2,c=False)
)
elif isinstance(cd,LineEvent):
return (
prsd_line(elm,le2=t2,c=True),
prsd_line(elm,le2=t2,c=False)
)
elif isinstance(cd,ParsedHTML):
t1,t2,t3 = fmt_syms.get_Tier(tier)
return (
prsd_html(elm,sym1=t1,sym2=t2,c=True),
prsd_html(elm,sym1=t1,sym2=t2,c=False)
)
elif isinstance(cd,ParsedJSON):
t1,t2,t3 = fmt_syms.get_Tier(tier)
return (prsd_json(elm,sym1=t1,sym2=t2,c=True),
prsd_json(elm,sym1=t1,sym2=t2,c=False)
)
elif isinstance(cd,VerboseList):
return prsd_lst(elm,c=True,tier=tier),prsd_lst(elm,c=False,tier=tier)
elif isinstance(cd[0],ParsedTuple):
return prsd_tpl(elm,c=True,tier=tier),prsd_tpl(elm,c=False,tier=tier)
else:
raise
def plot_polygons(img,
poly_dict,
alpha=0.5,
ax=None,
figsize=None,
show_axs=False):
if not ax:
fig, ax = plt.subplots(figsize=figsize)
show_img(img, ax, figsize, show_axs)
patches = []
colors = []
labels = list(poly_dict.keys())
st()
for l in labels:
color = 100 * np.random.rand()
for polygons in poly_dict[l]:
for poly in polygons:
if poly:
# st()
p = np.array(convert_poly(poly, img.shape))
patches.append(Polygon(p, True))
colors.append(color)
patch_collection = PatchCollection(patches, alpha=alpha)
patch_collection.set_array(np.array(colors))
ax.add_collection(patch_collection)
def entry():
query, actions, outputs, filenames, write_func, pkldf = qcfg = QueryConfig(
).fullcall()
dfpath = basepath.joinpath('src/youtube-dl/bin/agg.fullcall/fcdf.pkl')
dfs = {}
if not dfpath.exists():
fcdf = fur.get_df_from_tracefile(filenames[0])
original_index = list(fcdf.index)
fcdf['og_index'] = original_index
fcdf.to_pickle(dfpath)
else:
fcdf = pd.read_pickle(dfpath)
st()
fcdf = fcdf[:30] #!! RIGHT HERE
# if TRUNC: fcdf = fcdf[trunc_start:trunc_end]
columns = [
"filepath", "line_number", "event_kind", "call_data", "og_index"
]
fcdf_agg = fud.aggregate_aggdfs([fcdf], columns=columns)
cds = call_data_series = fcdf_agg.apply(process_row, axis=1)
cdl = call_data_lst = [
SimpleFunkWithArgs(cd) if
not (isinstance(cd, LineEvent) or isinstance(cd, VerboseList)
or isinstance(cd, ParsedHTML) or isinstance(cd, ParsedJSON)
or isinstance(cd[0], ParsedTuple)) else cd
for cd in call_data_series
]
fcdf_agg.call_data = call_data_lst
lwr = lst_w_regions = add_regions(fcdf_agg)
lwr_df = pd.DataFrame(lwr)
wpath = util.homepath.joinpath('bin/lwr')
lwr2 = write_file(df=fcdf_agg, lst_w_regions=lwr, writepath=wpath)
return lwr2
def get_syn(sentence, word):
# return WordNet.get_syn(word, types=[SynonymSubstituition.get_type(sentence, word)])[0]
# st()
word = word.split(' ')[0]
syn = WordNet.get_syn(
word, types=[SynonymSubstituition.get_type(sentence, word)])
if len(syn) == 0:
syn = WordNet.get_syn(
word
) # get all syns, ranked by general commonality (not wd specific)
if len(syn) > 0:
syn = pd.Series(syn)
try:
syn = syn[syn != word].unique()
except:
st()
syn2 = []
for i in range(len(syn)):
if len(syn[i].split(', ')) > 0:
syn[i] = syn[i].split(', ')
key = lambda x: CheckRareWord(
SynonymSubstituition.clean(x).split(' ')[0])
for wd in sorted(syn[i], key=key):
syn2.append(wd.strip())
else:
syn2.append(syn[i].strip())
return ', '.join(syn2[0:SynonymSubstituition.num_syns_display]) if len(syn) > 0 \
else '?'
def get_list_of_scenarios(config_file_name):
"""
Parse *all* config files and provide a list with all scenarios @return
"""
list_of_config_files = glob.glob("{}/srunner/examples/*.xml".format(
os.getenv('ROOT_SCENARIO_RUNNER', "./")))
list_of_config_files += glob.glob("{}/srunner/examples/*.xosc".format(
os.getenv('ROOT_SCENARIO_RUNNER', "./")))
from pdb import set_trace as st
st()
if config_file_name != '':
list_of_config_files.append(config_file_name)
scenarios = []
for file_name in list_of_config_files:
if ".xosc" in file_name:
tree = ET.parse(file_name)
scenarios.append("{} (OpenSCENARIO)".format(
tree.find("FileHeader").attrib.get('description', None)))
else:
tree = ET.parse(file_name)
for scenario in tree.iter("scenario"):
scenarios.append(scenario.attrib.get('name', None))
return scenarios
def train_cifar(args, nb_epochs, trainloader,testloader, net,optimizer,criterion,logging_freq=2000):
# TODO: test loss
for epoch in range(nb_epochs): # loop over the dataset multiple times
running_train_loss = 0.0
#running_test_loss = 0.0
for i, data in enumerate(trainloader, 0):
# get the inputs
start_time = time.time()
inputs, labels = data
if args.enable_cuda:
inputs, labels = Variable(inputs.cuda()), Variable(labels.cuda())
else:
inputs, labels = Variable(inputs), Variable(labels)
# zero the parameter gradients
optimizer.zero_grad()
# forward + backward + optimize
outputs = net(inputs)
loss = criterion(outputs, labels)
loss.backward()
optimizer.step()
# print statistics
running_train_loss += loss.data[0]
seconds,minutes,hours = utils.report_times(start_time)
st()
if i % logging_freq == logging_freq-1: # print every logging_freq mini-batches
# note you dividing by logging_freq because you summed logging_freq mini-batches, so the average is dividing by logging_freq.
print(f'monitoring during training: eptoch={epoch+1}, batch_index={i+1}, loss={running_train_loss/logging_freq}')
running_train_loss = 0.0
def get_weight_stat(self, batch, W_type):
## get lengths of sequences for each sample in the batch
weight_lengths = self.Tensor([len(sample[W_type])
for sample in batch]).long()
## padd
#st()
new_batch = []
for i, sample in enumerate(batch):
try:
print(f'i = {i}')
print(f'sample = {sample}')
tensor_sample = self.Tensor(sample[W_type])
print(f'tensor_sample = {tensor_sample}')
new_batch.append(tensor_sample)
except:
print(f'\n ---- ERROR: i = {i}')
print(f'sample = {sample}')
st()
## padd batch sequences
batch_weight_rep = torch.nn.utils.rnn.pad_sequence(
new_batch,
batch_first=self.batch_first,
padding_value=self.padding_value)
## compute mask
weight_mask = (batch_weight_rep != self.padding_value)
##
return batch_weight_rep.to(self.device), weight_lengths.to(
self.device), weight_mask.to(self.device)
def fix_init(self, elm):
new_filepath = Path(elm.filepath).stem
assert new_filepath is not None, elm
if "__init__" in new_filepath:
lno = int(elm.line_number)
def trunc(f):
return f[lno - 3:lno + 3]
def func(e):
return e.filepath
e, d, p, y = trnctd_fls = [trunc(f) for f in self.initfiles]
try:
if lno > 61:
new_filepath = "__init__"
elif e and any([func(elm) in itm for itm in e]):
new_filepath = "e.__init__"
elif d and any([func(elm) in itm for itm in d]):
new_filepath = "d.__init__"
elif p and any([func(elm) in itm for itm in p]):
new_filepath = "p.__init__"
elif y and any([func(elm) in itm for itm in y]):
new_filepath = "__init__"
except Exception as exc:
tb = stackprinter.format(exc)
print('The front fell off.' + tb)
st()
if new_filepath is None: st()
return new_filepath
def prsd_lst(elm, c=False, tier=0):
"""VerboseList"""
if elm.og_index == 39: st()
cd = elm.call_data
m = re.match(r"\s*", cd.get_funcname())
idt = m.group(0)
if c:
sym_func = f"{cd.symbol} {cd.funk}"
filename_fileno = (f"{util.fmt_filename(elm.filepath)}"
f":{util.fmt_lineno(elm.line_number)}\n")
filename_fileno = util.filename_fileno2(elm.filepath,
elm.line_number) + "\n"
argname_argvals = (f"{cd.get_args(c=True)}")
else:
sym_func = f"{cd.symbol} {cd.funk}"
filename_fileno = (
f"{util.remove_color(util.fmt_filename(elm.filepath))}"
f":{util.remove_color(util.fmt_lineno(elm.line_number))}\n")
argname_argvals = (f"{cd.get_args(c=False)}")
split_args = argname_argvals.split("\n")
num_args = len(split_args)
vl1, vl2 = fmt_syms.get_VL(nargs=num_args)
argname_argvals = "\n".join([
f"{next(vl2,None)}{idt}{line}"
for line in argname_argvals.split("\n")
])
fillchars = util.get_fillchars(sym_func)
elm_as_str = f'{vl1}' + sym_func + fillchars + filename_fileno + argname_argvals
return elm_as_str
def iterate_graph(node_id: int):
node_name = self.node(node_id).name
if node_name not in layer_list:
node = self.node(node_id)
if isinstance(self.node(node_id), Operation):
print(type(self.node(node_id)))
nonlocal weights
if isinstance(node, DenseOp):
input_shape = self.node(node._inputs[0]).shape
output_shape = self.node(node._outputs[0]).shape
syn = input_shape[1] * output_shape[1]
elif (isinstance(node, AvgPoolOp) or
isinstance(node, ReluOp) or
isinstance(node, BatchNormOp)
):
input_shape = self.node(node._inputs[0]).shape
output_shape = self.node(node._outputs[0]).shape
syn = np.prod(input_shape)
elif (isinstance(node, ReshapeOp) or
isinstance(node, PadOp) or
isinstance(node, TransposeOp)
):
input_shape, output_shape, syn = 0, 0, 0
elif isinstance(node, AddOp):
input_shape = node._inputs
output_shape = self.node(node._outputs[0]).shape
syn = np.prod(output_shape) * len(node._inputs)
elif isinstance(node, Conv2dOp):
input_shape = self.node(node._inputs[0]).shape
output_shape = self.node(node._outputs[0]).shape
# print(input_shape[1], output_shape, 3, 3)
syn = input_shape[1] * output_shape[1] *3*3
else:
print(node)
st()
# print(node, input_shape, output_shape, syn)
weights += syn
layer_list.append(node_name)
if node_id not in self.inputs:
node_inputs = self.node(node_id).inputs
if len(node_inputs) == 1:
iterate_graph(node_inputs[0])
else:
node_inputs = list(
filter(
lambda input_id: not (
self.contains_tensor(input_id)
and self.tensor(input_id).name.startswith("add")
),
node_inputs,
)
)
# if len(node_inputs) != 1:
# print(f"guess: choose first input in {node_name}")
iterate_graph(node_inputs[0])
else:
print(f"node {node_name} is already in list")
def __init__(self, hyperparams):
# Get configs, hyperparameters & initialize agent
self.ptconf, self.dconf, self.camconf = hyperparams[
'ptconf'], hyperparams['dconf'], hyperparams['camconf']
self.env_name = hyperparams['env']
self.gui_on = hyperparams['bullet_gui_on']
self.hyperparams = hyperparams
config = copy(AGENT_BULLET)
config.update(hyperparams)
Agent.__init__(self, config)
# Setup bullet environment
self._setup_conditions()
self._setup_world(hyperparams['filename'])
self.setup_bullet()
self.setup_inference_camera()
# Get demo data
self.demo_vid = imageio.get_reader(
join(self.dconf.DEMO_DIR, self.dconf.DEMO_NAME,
'rgb/{}.mp4'.format(self.dconf.SEQNAME)))
self.demo_frames = []
self.reset_condition = hyperparams['reset_condition']
for im in self.demo_vid:
self.demo_frames.append(im)
# self.cost_tgt_mean = hyperparams['cost_tgt_mean']
# self.cost_tgt_std = hyperparams['cost_tgt_std']
self.objects_centroids_mrcnn = hyperparams['objects_centroids_mrcnn']
# Setup feature embedding network if enabled
self.tcn = None
# Setup Mask RCNN
inference_config = InferenceConfig()
with tf.device('/device:GPU:1'):
self.mrcnn = modellib.MaskRCNN(
mode='inference',
model_dir=join(self.ptconf.EXP_DIR, self.ptconf.EXP_NAME,
'mrcnn_logs'),
config=inference_config)
self.mrcnn.load_weights(self.ptconf.WEIGHTS_FILE_PATH_MRCNN,
by_name=True)
self.class_names = gconf.CLASS_NAMES_W_BG
# self.target_ids = gconf.CLASS_IDS
self.target_ids = [1, 2]
self.colors = visualize.random_colors(7)
self.plotting_on = hyperparams['plotting_on']
if self.plotting_on:
self.fig, self.ax = visualize.get_ax()
self.mrcnn_centroids_last_known = {
key: None
for key in self.target_ids
}
st()
# Run MRCNN on one image because first detection always takes long to initialize
self.vid_seqname = 0
rgb_crop, depth_crop = self.get_images(0)
results = self.mrcnn.detect([rgb_crop], verbose=0)
self.blob_detector = BlobDetector()
def get_results_from_normalized_net(self, epoch, seed_id,
path_to_folder_expts):
''' '''
''' get net '''
nets_folders = [
filename for filename in os.listdir(path_to_folder_expts)
if 'nets_folder' in filename
]
net_folder = [
filename for filename in nets_folders
if f'seed_{seed_id}' in filename
][0] # note seed are unique very h.p.
net_path = os.path.join(path_to_folder_expts, net_folder)
if len([
net_name for net_name in os.listdir(net_path)
if f'epoch_{epoch}' in net_name
]) == 0:
st()
net_name = [
net_name for net_name in os.listdir(net_path)
if f'epoch_{epoch}' in net_name
][0]
net_path = os.path.join(net_path, net_name)
net = torch.load(net_path)
''' get unormalized test error '''
train_loss_un, train_error_un = evalaute_mdl_on_full_data_set(
self.loss, self.error, net, self.trainloader, self.device)
test_loss_un, test_error_un = evalaute_mdl_on_full_data_set(
self.loss, self.error, net, self.testloader, self.device)
## random labels
train_loss_un_rand, train_error_un_rand = evalaute_mdl_on_full_data_set(
self.loss, self.error, net, self.trainloader_rand, self.device)
test_loss_un_rand, test_error_un_rand = evalaute_mdl_on_full_data_set(
self.loss, self.error, net, self.testloader_rand, self.device)
''' normalize net '''
net = self.normalize(net)
''' get normalized train errors '''
## natural labels
train_loss_norm, train_error_norm = evalaute_mdl_on_full_data_set(
self.loss, self.error, net, self.trainloader, self.device)
test_loss_norm, test_error_norm = evalaute_mdl_on_full_data_set(
self.loss, self.error, net, self.testloader, self.device)
## random labels
train_loss_norm_rand, train_error_norm_rand = evalaute_mdl_on_full_data_set(
self.loss, self.error, net, self.trainloader_rand, self.device)
test_loss_norm_rand, test_error_norm_rand = evalaute_mdl_on_full_data_set(
self.loss, self.error, net, self.testloader_rand, self.device)
''' pack results '''
normalized_results = (train_loss_norm, train_error_norm,
test_loss_norm, test_error_norm)
unnormalized_results = (train_loss_un, train_error_un, test_loss_un,
test_error_un)
##
normalized_results_rand = (train_loss_norm_rand, train_error_norm_rand,
test_loss_norm_rand, test_error_norm_rand)
unnormalized_results_rand = (train_loss_un_rand, train_error_un_rand,
test_loss_un_rand, test_error_un_rand)
''' return '''
return normalized_results, unnormalized_results, normalized_results_rand, unnormalized_results_rand
def fix_init(self,elm):
new_filepath = Path(elm.filename).stem
assert new_filepath is not None, elm
if "__init__" in new_filepath:
parent_initial = Path(elm.filename).parent.stem[0]
new_filepath = f"{parent_initial}{new_filepath}"
if new_filepath is None: st()
return new_filepath
def check_order_data(trainloader):
for i,data_train in enumerate(trainloader):
if i==0:
print(i)
st()
#print(data_train)
for i,data_train in enumerate(trainloader):
if i==3:
print(i)
def test_mutate(model, dataset, args):
model.eval()
for idx in range(args.test_num):
raw_img, label = dataset[idx]
image = np.array(raw_img)
imgs = [image]
# these parameters need to be carefullly considered in the experiment
# to consider the feedbacks
params = []
# params.append(list(range(-10, 10))) # image_translation
# params.append(list(map(lambda x: x * 0.1, list(range(9, 11))))) # image_scale
# params.append(list(map(lambda x: x * 0.1, list(range(-1, 1))))) # image_shear
# params.append(list(range(-10, 10))) # image_rotation
# params.append(list(map(lambda x: x * 0.1, list(range(7, 13))))) # image_contrast
# params.append(list(range(-15, 15))) # image_brightness
# params.append(list(range(1, 10))) # image_blur
# params.append(list(range(1, 10))) # image_pixel_change
# params.append(list(range(1, 3))) # image_noise [1,2]
params.append([-10, 10]) # image_translation
params.append(list(map(lambda x: x * 0.1, [9, 11]))) # image_scale
params.append(list(map(lambda x: x * 0.1, [-1, 1]))) # image_shear
params.append([-10, 10]) # image_rotation
params.append(list(map(lambda x: x * 0.1, [7, 12]))) # image_contrast
params.append([-15, 15]) # image_brightness
params.append([9]) # image_blur
params.append([9]) # image_pixel_change
params.append([2]) # image_noise
classA = [7, 8] # pixel value transformation
classB = [0, 1, 2, 3, 4, 5, 6] # Affine transformation
x, y, z = image.shape
# random.seed(time.time())
for tid in range(9):
# print(tid)
# tid = random.sample(classA + classB, 1)[0]
# tid = 7
# Randomly select one transformation Line-7 in Algorithm2
transformation = transformations[tid]
param_range = params[tid]
for param_id, param in enumerate(param_range):
img_new = transformation(copy.deepcopy(image), param)
imgs.append(img_new)
row_img = np.zeros((x, y * len(imgs), 3), dtype=np.uint8)
for i in range(len(imgs)):
row_img[:, i * y:(i + 1) * y, :] = imgs[i]
path = osp.join("results/test", f"{idx}.png")
Image.fromarray(row_img).save(path)
st()
def load_nc_info(self, ):
path = osp.join(self.args.nc_info_dir, "accumulate_coverage.npy")
with open(path, "rb") as f:
accumulate_coverage = np.load(f, allow_pickle=True)
path = osp.join(self.args.nc_info_dir, "log_module_names.npy")
with open(path, "rb") as f:
log_names = np.load(f, allow_pickle=True)
return accumulate_coverage, log_names
st()
def merge_trace(acc_trace, trace2):
node_names = acc_trace.ops.keys()
for node_name in node_names:
if TraceKey.IO_AVG in acc_trace.ops[node_name]:
acc_point_logger = acc_trace.ops[node_name][TraceKey.IO_AVG]
if TraceKey.IO_AVG not in trace2.ops[node_name]:
st()
new_point_logger = trace2.ops[node_name][TraceKey.IO_AVG]
acc_point_logger += new_point_logger
acc_trace.ops[node_name][TraceKey.IO_AVG] = acc_point_logger
def children_and_parameters(m: nn.Module):
"""Return the children of `m` and its direct parameters not registered in modules."""
children = list(m.children())
children_p = sum([[id(p) for p in c.parameters()] for c in m.children()],
[])
for p in m.parameters():
if id(p) not in children_p:
st()
children.append(ParameterModule(p))
return children
def compute_features(trip, d_id=None, t_id=None):
# print d_id,t_id
trip_duration = float(len(trip))
speed, acceleration = speeds_and_accelerations(trip)
speed = abs(speed)
trip_length = float(np.sum(speed))
speedup = acceleration[acceleration > 0]
slowdown = -acceleration[acceleration < 0]
try:
curvature = get_curvature(trip)
except:
logging.exception('Cannot compute curvature for trip %d of driver %d' % (t_id, d_id))
curvature = np.array([])
curvature_percentiles = [percentile(curvature, pc) for pc in [90, 75, 50, 25, 10]]
curvature_maxima_x = sp.signal.argrelmax(curvature, order=2)[0]
curvature_maxima_y = curvature[curvature_maxima_x]
curvature_maxima_speed = speed[curvature_maxima_x]
curvature_maxima_percentiles = [percentile(curvature_maxima_y, pc) for pc in [75, 50, 25]]
centrifugal_acceleration = np.power(curvature_maxima_speed, 2)*curvature_maxima_y
# curvature_inter_maxima = [np.diff(maxima) for maxima in curvature_maxima]
# median_curvature_inter_maxima = [np.percentile(cim, 50) if cim.size else trip_length for cim in curvature_inter_maxima]
mean_slowdown, mean_speedup = speeds_near_stops(speed)
features = [
trip_length,
trip_duration,
np.percentile(speed, 75),
np.percentile(speed, 50),
np.percentile(speed, 25),
np.percentile(speedup, 75),
np.percentile(speedup, 50),
np.percentile(speedup, 25),
np.percentile(slowdown, 75),
np.percentile(slowdown, 50),
np.percentile(slowdown, 25),
curvature_percentiles[0],
curvature_percentiles[1],
curvature_percentiles[2],
curvature_percentiles[3],
curvature_percentiles[4],
len(curvature_maxima_x)/trip_length,
curvature_maxima_percentiles[0],
curvature_maxima_percentiles[1],
curvature_maxima_percentiles[2],
mean(centrifugal_acceleration),
mean(curvature_maxima_speed),
gps_signal(speed),
mean_slowdown,
mean_speedup,
]
if not np.all(np.isfinite(features)):
st()
return features
def eval(v):
st()
print(vec.shape)
regu_coef = self.FIM_invert_args['damping'] if regu_coef is None else regu_coef
self.npg_grad = self.cg_solve(self.hvp, self.vpg_grad, x_0=self.vpg_grad.copy(),
cg_iters=self.FIM_invert_args['iters'])
Hvp, npg_grad = self.sess.run([self.hvp, self.npg_grad], {self.obs_ph: observations,
self.action_ph: actions,
self.vec: vec,
self.regu_coef: regu_coef})
return Hvp, npg_grad
def forward(self, x):
# Input is of shape [Batch, 3, channel, height, width]
# Cylindrical representation of data: [log(|z|), x/|z|, y/|z|]
wFMs = self.wFM(x)
# Input is of shape [Batch, 3, channel, height, width]
# Cylindrical representation of data: [log(|z|), x/|z|, y/|z|]
# Separate each component and do convolution along each component
# The input of each component is [Batch, in_channel, height, width]
# The output of each component is [Batch, in_channel, height, width]
# The final output of this layer would be [Batch, in_channel, height, width]
mag_output = wFMs[:, 0, ...]
cos_output = wFMs[:, 1, ...]
sin_output = wFMs[:, 2, ...]
cos_sin_output = wFMs[:, 1:, ...]
# For center-cropping original input
output_xdim = cos_output.shape[2]
output_ydim = cos_output.shape[3]
input_xdim = x.shape[3]
input_ydim = x.shape[4]
start_x = int((input_xdim - output_xdim) / 2)
start_y = int((input_ydim - output_ydim) / 2)
m = nn.Upsample(size=(input_xdim, input_ydim), mode='bilinear')
cropped_input = x
log_output = m(log_output)
cos_sin_output = torch.cat((m(cos_sin_output[:, 0, ...]).unsqueeze(1),
m(cos_sin_output[:, 1, ...]).unsqueeze(1)),
dim=1)
# Compute distance according to sqrt(log^2[(|z2|+b)/(|z1|+b)] + acos^2(x^T * y))
# Need to add noise or else normalization may have zero entries which cause NaN
# Compute distance according to sqrt(log^2[(|z2|+b)/(|z1|+b)] + acos^2(x^T * y))
# Need to add noise or else normalization may have zero entries which cause NaN
direction_noise = 1e-5
directional_difference = cropped_input[:, 1:, ...] - cos_sin_output
directional_difference = (
directional_difference + direction_noise) / torch.sqrt(
torch.sum(directional_difference**2 + direction_noise,
dim=1,
keepdim=True))
if m(directional_difference):
st()
magnitude_difference = cropped_input[:, 0, ...] - log_output
return torch.cat(
(magnitude_difference.unsqueeze(1), directional_difference), dim=1)
def main(unused_argv):
print('Saving results to %s' % FLAGS.output_directory)
if not exists(FLAGS.output_directory):
makedirs(FLAGS.output_directory)
from pdb import set_trace as st
st()
process_mpii(anno,
img_dir,
FLAGS.output_directory,
FLAGS.train_shards,
is_train=True)
def find_path(graphd, src, dest, path=[]):
path = path + [src]
st()
if src == dest:
return path
if not src in graphd:
return None
for node in graphd[src]:
if node not in path:
new_path = find_path(graphd, node, dest, path)
if new_path : return new_path
return None
def restore(self, load_path):
sum_tf, sumsq_tf, count_tf, mean, std = pickle.load(
open(load_path, 'rb'))
restored_value = tf.constant(sum_tf), tf.constant(
sumsq_tf), tf.constant(count_tf), tf.constant(mean), tf.constant(
std)
var_list = [
self.sum_tf, self.sumsq_tf, self.count_tf, self.mean, self.std
]
st()
self.sess.run([
tf.assign(var, restored)
for (var, restored) in zip(var_list, restored_value)
])
def __init__(self,
img_name_list_path,
voc12_root,
resize_long=None,
rescale=None,
img_normal=TorchvisionNormalize(),
hor_flip=False,
crop_size=None,
crop_method=None):
super().__init__(img_name_list_path, voc12_root, resize_long, rescale,
img_normal, hor_flip, crop_size, crop_method)
self.label_list = load_image_label_list_from_npy(self.img_name_list)
st()
def get_random_faltness_radius_repeated_doubling(self, precision=0.0001):
'''
The goal of this is for the easy of computation of the epsilon radius of
a network which is defined as follows:
r(dx,eps,W) = sup{r \in R : |I(W) - I(W+r*dx)|<=eps}
W_all = r*dx
dx = isotropic unit vector from the net
'''
''' record reference errors/losses '''
#stats_collector.record_errors_loss_reference_net(criterion,error_criterion,net,trainloader,testloader,device)
''' get isotropic direction '''
nb_params = nn_mdls.count_nb_params(self.net)
v = torch.normal(torch.zeros(nb_params),
torch.ones(nb_params)).to(self.device)
dx = v / v.norm(2)
''' fill up I list '''
r = self.r_initial
Loss_minima, Error_minima = evalaute_running_mdl_data_set(
self.criterion, self.error_criterion, self.net, self.trainloader,
self.device, self.iterations)
while True:
net_rdx = produce_new_translated_net(self.net, r, dx)
Loss_rdx, Error_rdx = evalaute_running_mdl_data_set(
self.criterion, self.error_criterion, net_rdx,
self.trainloader, self.device, self.iterations)
diff = Error_rdx - Error_minima
print(f'\n--\nr = {r}')
print(f'Error_minima={Error_minima}')
print(f'Error_rdx={Error_rdx}')
print(f'diff = {diff}')
print(f'epsilon={self.epsilon}')
print(f'abs(abs(diff)-eps)={abs(abs(diff)-self.epsilon)}')
print(f'precision={precision}')
print(
f'abs(abs(diff)-eps) < precision={abs(abs(diff)-self.epsilon) < precision}'
)
print(
f'approx_equals(diff,self.epsilon,precision=precision)={approx_equals(diff,self.epsilon,precision=precision)}'
)
''' check if we reached epsilon jump '''
if approx_equals(diff, self.epsilon, precision=precision
): ## 10^-4.5 for half machine precision
''' compute I(W+r*dx) = I(W+W_all)'''
st()
return r
elif diff > self.epsilon: # I(w+rdx) - I(W) > eps, r is too large
r /= 2
else: # I(w+rdx) - I(W) < eps, r is too small
r *= 1.5
def forward(self, x):
y = list()
if self.arch == 'C':
x = self.conv1(x)
x = self.bn1(x)
x = self.relu(x)
elif self.arch == 'D':
x = self.layer0(x)
x = self.layer1(x)
y.append(x)
x = self.layer2(x)
y.append(x)
x = self.layer3(x)
y.append(x)
x = self.layer4(x)
y.append(x)
x = self.layer5(x)
y.append(x)
if self.layer6 is not None:
x = self.layer6(x)
y.append(x)
if self.layer7 is not None:
x = self.layer7(x)
y.append(x)
if self.layer8 is not None:
x = self.layer8(x)
y.append(x)
if self.out_map:
x = self.fc(x)
else:
x = self.avgpool(x)
x = self.fc(x)
x = x.view(x.size(0), -1)
st()
if self.out_middle:
return x, y
else:
return x
def read_frame(vid_name, frame_num, box, x):
img_name = os.path.join(vid_name, str(frame_num + 1) + '.jpg')
if not os.path.isfile(img_name):
img_name = os.path.join(vid_name, str(frame_num + 1) + '.png')
img = cv2.imread(img_name)
joints = x[:, :, frame_num] - 1.0
box_frame = box[frame_num, :]
st()
scale = get_person_scale(joints)
pos = np.zeros(2)
pos[0] = (box_frame[0] + box_frame[2] / 2.0)
pos[1] = (box_frame[1] + box_frame[3] / 2.0)
return img, joints, scale, pos
def tf_get_actions(self, observations, update_filter=True, params=None):
assert len(observations.shape) == 2
st()
obs = self.tf_obs_filters[0](observations, update=update_filter)
if params == None:
actions = tf.transpose(
tf.matmul(self.policy_params["W"],
tf.transpose(obs))) + self.policy_params["b"]
else:
actions = tf.transpose(
tf.matmul(tf.cast(params["W"], obs.dtype),
tf.transpose(obs))) + tf.cast(
params["b"], obs.dtype)
actions = tf.clip_by_value(actions, LOG_SIG_MIN, LOG_SIG_MAX)
return actions, {}
def reinit_model(
begin_layer,
model,
num_classes,
train_all,
prune_ratio,
lr,
random_prune=False,
):
if isinstance(model, torchvision.models.resnet.ResNet):
if len(model.layer1) == 2:
model, model_params, fix_bn_dict = reinit_in_resnet.init_resnet18(
model,
num_classes,
begin_layer,
lr,
train_all,
random_prune=random_prune,
)
model = weight_prune.weight_prune(
model,
begin_layer,
prune_ratio,
reinit_in_resnet.resnet18_layer_depth,
)
return model, model_params, fix_bn_dict
elif len(model.layer1) == 3:
raise RuntimeError("lr")
return reinit_in_resnet.init_resnet50(
model,
num_classes,
begin_layer,
train_all,
)
else:
raise NotImplementedError
elif isinstance(model, torchvision.models.vgg.VGG):
raise RuntimeError("lr")
return reinit_in_vgg.init_vgg16(
model,
num_classes,
begin_layer,
train_all,
)
else:
st()
raise NotImplementedError
def run_machine(machine, env):
daemons = machine.get('daemons', [])
daemons = expand_commands(daemons, env)
hp_node = int(machine.get('hp_node', -1))
if verbose:
print(" hp_node %d" % hp_node)
if hp_node == -1:
print("machine has invalid hp_node")
st()
sys.exit(1)
for d in daemons:
if 'port' not in d or 'command' not in d:
print("machine does not have port or command")
st()
sys.exit(1)
port = int(d['port'])
notify_port_map(hp_node, port)
run_daemons(hp_node, daemons)
def company_create():
NAME_KEY = 'name'
if request.method == 'POST':
try:
db.create_company(ImmutableDict(
name = request.form[NAME_KEY]))
flash("company created")
except DbException as e:
flash("Error creating company: '" + e.msg + "'")
except Exception as e:
print "*************************************"
print "unexpected error creating new company"
print e
print repr(e)
st()
print "*************************************"
flash("unexpected error creating new company")
return render_template('new_company.html', url_self=_url_company_create)
def main(model='mlp', num_epochs=500):
# Load the dataset
print("Loading data...")
X_train, y_train, X_val, y_val, X_test, y_test = load_dataset()
# make effective data set smaller
ind1 = np.arange(y_train.size); ind2 = np.arange(y_test.size)
np.random.shuffle(ind1); np.random.shuffle(ind2)
X_train, y_train = X_train[ind1[:2000]], y_train[ind1[:2000]]
X_test, y_test = X_test[ind2[:1000]], y_test[ind2[:1000]]
from pdb import set_trace as st
# Prepare Theano variables for inputs and targets
input_var = T.tensor4('inputs')
target_var = T.ivector('targets')
# Create neural network model (depending on first command line parameter)
print("Building model and compiling functions...")
if model == 'mlp':
network = build_mlp(input_var)
elif model.startswith('custom_mlp:'):
depth, width, drop_in, drop_hid = model.split(':', 1)[1].split(',')
network = build_custom_mlp(input_var, int(depth), int(width),
float(drop_in), float(drop_hid))
elif model == 'cnn':
network = build_cnn(input_var)
else:
print("Unrecognized model type %r." % model)
return
# Create a loss expression for training, i.e., a scalar objective we want
# to minimize (for our multi-class problem, it is the cross-entropy loss):
prediction = lasagne.layers.get_output(network)
loss = lasagne.objectives.categorical_crossentropy(prediction, target_var)
loss = loss.mean()
# We could add some weight decay as well here, see lasagne.regularization.
# Create update expressions for training, i.e., how to modify the
# parameters at each training step. Here, we'll use Stochastic Gradient
# Descent (SGD) with Nesterov momentum, but Lasagne offers plenty more.
params = lasagne.layers.get_all_params(network, trainable=True)
updates = lasagne.updates.nesterov_momentum(
loss, params, learning_rate=0.01, momentum=0.9)
# Create a loss expression for validation/testing. The crucial difference
# here is that we do a deterministic forward pass through the network,
# disabling dropout layers.
test_prediction = lasagne.layers.get_output(network, deterministic=True)
test_loss = lasagne.objectives.categorical_crossentropy(test_prediction,
target_var)
test_loss = test_loss.mean()
# As a bonus, also create an expression for the classification accuracy:
test_acc = T.mean(T.eq(T.argmax(test_prediction, axis=1), target_var),
dtype=theano.config.floatX)
# Compile a function performing a training step on a mini-batch (by giving
# the updates dictionary) and returning the corresponding training loss:
train_fn = theano.function([input_var, target_var], loss, updates=updates)
# Compile a second function computing the validation loss and accuracy:
val_fn = theano.function([input_var, target_var], [test_loss, test_acc])
# Finally, launch the training loop.
print("Starting training...")
acc_list = np.zeros((num_epochs,))
# We iterate over epochs:
for epoch in range(num_epochs):
# In each epoch, we do a full pass over the training data:
train_err = 0
train_batches = 0
start_time = time.time()
for batch in iterate_minibatches(X_train, y_train, 500, shuffle=True):
inputs, targets = batch
train_err += train_fn(inputs, targets)
train_batches += 1
# And a full pass over the validation data:
val_err = 0
val_acc = 0
val_batches = 0
for batch in iterate_minibatches(X_val, y_val, 200, shuffle=False):
inputs, targets = batch
err, acc = val_fn(inputs, targets)
val_err += err
val_acc += acc
val_batches += 1
# Then we print the results for this epoch:
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(" validation loss:\t\t{:.6f}".format(val_err / val_batches))
print(" validation accuracy:\t\t{:.2f} %".format(
val_acc / val_batches * 100))
acc_list[epoch] = val_acc / val_batches
# After training, we compute and print the test error:
test_err = 0
test_acc = 0
test_batches = 0
for batch in iterate_minibatches(X_test, y_test, 200, shuffle=False):
inputs, targets = batch
err, acc = val_fn(inputs, targets)
test_err += err
test_acc += acc
test_batches += 1
st()
print("Final results:")
print(" test loss:\t\t\t{:.6f}".format(test_err / test_batches))
print(" test accuracy:\t\t{:.2f} %".format(
test_acc / test_batches * 100))
# -*- coding: utf-8 -*-
"""
Created on Tue Jul 26 17:27:11 2016
@author: pi
"""
import pdb as db
from pdb import set_trace as st
import easygui
easygui.msgbox("Hello there!")
st()
user_reponse = easygui("hello there!")
easygui.msgbox("Leo is haunting you for life!")