def comparaison():
text = 'Nous sommes en guerre. Aussi, comme je vous l’ai dit jeudi, pour nous protéger et contenir la dissémination du virus, mais aussi préserver nos systèmes de soins, nous avons pris ce matin, entre Européens, une décision commune. Dès demain midi, les frontières à l’entrée de l’Union européenne et de l’espace Schengen seront fermées z w . Concrètement, tous les voyages entre les pays non européens et l’Union européenne seront suspendus pendant trente jours. Les Françaises et les Français qui sont actuellement à l’étranger et souhaitent rentrer pourront bien entendu rejoindre leur pays. Nous devons prendre cette décision parce que je vous demande ce soir d’importants efforts et que nous devons, dans la durée, nous protéger. Et je veux dire à tous nos compatriotes qui vivent à l’étranger que là aussi, en bon ordre, ils doivent se rapprocher des ambassades et consulats et que nous organiserons pour celles et ceux qui le souhaitent, et là où c’est nécessaire, le rapatriement. ? ok '.lower(
)
# Désolé pour le texte en ligne c'est aps très PEP-8 friendly
builder = TreeBuilder(text)
binary_tree = builder.tree()
codec = Codec(binary_tree)
texte_french = 'Mes chers compatriotes, alors que je vous parle, les résultats connus nous montrent que vous avez décidé de me confier la plus haute charge de l Etat. J exprime ma profonde gratitude à toutes celles et à tous ceux qui m ont accordé leur confiance et je salue tous les autres avec respect. Mes chers compatriotes, je serai le Président de tous les français'.lower(
)
texte_allemand = 'Die Stadt Paris ist fur viele ein Traumziel, das sie wenigstens einmal im Leben besuchen mochten. Wer mochte nicht einmal unter dem Eiffelturm stehen und in den Pariser Himmel hoch schauen? Und dann zu Fuss und mit dem Aufzug wenigstens auf halbe Hohe gelangen und den Blick uber die ganze franzosische Hauptstadt geniessen?'.lower(
)
compress_fr, compress_de = [], []
for i in range(max(len(texte_french), len(texte_allemand))):
texte_fr = texte_french[:i] # Création d'un texte
texte_de = texte_allemand[:i]
taille_fr = sys.getsizeof(texte_fr) # Espace système de ce texte
taille_de = sys.getsizeof(texte_de)
compress_fr.append(taille_fr /
sys.getsizeof(codec.encode_bin(texte_fr)))
compress_de.append(taille_de /
sys.getsizeof(codec.encode_bin(texte_de)))
plt.close()
plt.title("Comparaison de l'effet de la langue sur la taux de compression")
plt.plot(compress_fr,
label=" facteur de compression du texte en français ")
plt.plot(compress_de, label=" facteur de compression du texte en allemand")
plt.legend()
plt.show()
def cli():
"""\
Codec CLI for instant usage: Read and write.
"""
argv = parse_args()
c = Codec(**vars(argv))
c.run()
def find_codec(self, name):
cmdstr = "adb shell find " + self.path + " -maxdepth 1"
#print cmdstr
result = os.popen(cmdstr)
ret = result.read()
lines = ret.split('\n')
#self.codecs = codecs
for i in range(len(lines)):
if (-1 != lines[i].find(name)):
#print lines[i]
codec = Codec(name, lines[i])
codec.find_widgets()
#print codec
self.codecs.append(codec)
def __init__(self, engine, protocol, sock=None, addr=None, observer=None, auth=None):
self.codec = Codec(protocol, auth)
self.engine = engine
self.connectionState = ConnectionState.CONNECTED
self.session = None
self.addr = addr
self.observer = observer
self.msgBuffer = b''
self.heartbeatPeriod = 30.0
self.msgHandlers = []
self.sock = sock
self.heartbeatTimerRegistration = None
self.expectedHeartbeatRegistration = None
self.socketEvent = FileDescriptorEventRegistration(self.handle_read, sock, EventType.READ)
self.engine.eventManager.registerHandler(self.socketEvent)
def main():
with open(CONFIG_FN) as f:
conf = json.load(f)
global sent_groups
with open(conf["sent_groups"]) as f:
sent_groups = json.load(f)["groups"]
kb = load_kb(conf["kb"], 'name')
sys_vocab, sys_word2idx = load_sys_vocab(conf["sys_vocab"])
sys_codec = Codec(sys_vocab, sys_word2idx)
onto, onto_idx = load_ontology(conf["ontology"])
word2idx, embed = load_embed(**conf)
usr_codec = Codec([], word2idx)
trk_model, slot_len_sum = load_tracker_model(onto, embed, conf, kb)
trk_model.eval()
hidden = trk_model.state_tracker.init_hidden()
kb_vec = Variable(torch.zeros(1, conf["kb_indicator_len"]))
sentence_generator = SentenceGenerator(kb, onto, sent_groups)
for line in iter(sys.stdin.readline, ''):
inp = usr_codec.encode(line.strip())
inp = Variable(torch.LongTensor([
inp,
]))
sentvecs, states_reps, states_preds, hidden, sent_grp_preds = trk_model(
inp, None, hidden)
criteria = to_search_criteria(states_preds, onto)
ret, kb_vec = get_kb_result(kb, criteria, conf["kb_indicator_len"])
# print criteria, kb_vec
sentvecs = sentvecs.view(1, -1)
states_reps = states_reps.view(1, -1)
print_ret(states_preds, sent_grp_preds, onto, sentence_generator)
def __init__(self):
from codec import Codec
from window import Window
from system import System
from datajar import DataJar
from filesystem import FileSystem
try:
manifest = json.load(codecs.open('manifest.json', 'r', 'utf-8'))
except:
manifest = {}
for key in assets.manifest:
if key in manifest:
assets.manifest[key] = manifest[key]
self.app = QApplication(sys.argv)
self.app.setApplicationName(assets.manifest['name'])
self.app.setApplicationVersion(assets.manifest['version'])
assets.sys = System()
assets.codec = Codec()
assets.fs = FileSystem()
assets.dataJar = DataJar()
translator = QTranslator()
if translator.load("zh_CN.qm"):
self.app.installTranslator(translator)
self.window = Window(None, assets.manifest['path'] + 'index.html')
sys.exit(self.app.exec_())
def check_add_codec(ip):
net = get_codec_by_ip(ip)
if net:
return net
net = Codec(ip)
audio_codecs.append(net)
return net
def main():
text = "a dead dad ceded a bad babe a beaded abaca bed"
# on analyse les fréquences d'occurrence dans text
# pour fabriquer un arbre binaire
builder = TreeBuilder(text)
binary_tree = builder.tree()
# on passe l'arbre binaire à un encodeur/décodeur
codec = Codec(binary_tree)
# qui permet d'encoder
encoded = codec.encode_bin(text)
# et de décoder
decoded = codec.decode_bin(encoded)
# si cette assertion est fausse il y a un gros problème avec le code
assert text == decoded
# on affiche le résultat
print(f"{text}\n{encoded}")
if decoded != text:
print("OOPS")
text = 'ab'
compress_encode = []
compress_huffman = []
compress_binary = []
for rang in range(300):
texte = text * rang # Création d'un texte
taille = sys.getsizeof(texte) # Espace système de ce texte
compress_encode.append(taille / sys.getsizeof(codec.encode(texte)))
compress_huffman.append(taille /
sys.getsizeof(codec.encode_bin(texte)))
compress_binary.append(
sys.getsizeof(codec.encode(texte)) /
sys.getsizeof(codec.encode_bin(texte)))
# Affichage
plt.close()
plt.title('Facteurs de compression')
plt.plot(compress_encode,
label="Facteur de compression par \n Codec.encode sortie str")
plt.plot(
compress_huffman,
label=
"Facteur de compression de \n l'algorithme de Huffman (Avec Binaire)")
plt.plot(compress_binary,
label="Facteur de compression \n d'une str encoded en binaire ")
plt.legend()
plt.show()
def recall_trial(num_episodes, save_file):
# Configuration
layer_sizes = {"rinp": 64, "rout": 64, "rtmp": 64}
hidden_size = 64
rho = .99
plastic = ["rout<rtmp"]
# remove_pathways = ["rinp<rout", "rinp<rtmp", "rtmp<rout"]
remove_pathways = []
# Setup GHU
num_symbols = 5
chars = "abcdefghi"
numbs = "123456789"
symbols = chars[:num_symbols] + "0" + numbs[:num_symbols - 1]
pathways, associations = default_initializer( # all to all
layer_sizes.keys(), symbols)
for p in remove_pathways:
pathways.pop(p)
associations = list(
filter(lambda x: x[0] not in remove_pathways, associations))
codec = Codec(layer_sizes, symbols, rho=rho, ortho=True)
controller = Controller(layer_sizes,
pathways,
hidden_size,
plastic,
nonlinearity='relu')
# controller = Controller(layer_sizes, pathways, hidden_size, plastic, nonlinearity='tanh')
ghu = GatedHebbianUnit(layer_sizes,
pathways,
controller,
codec,
batch_size=num_episodes,
plastic=plastic)
ghu.associate(associations)
# Initialize layers
separator = "0"
ghu.fill_layers(separator)
# Dataset of all possible input lists
all_inputs = [
np.array([k1, v1, k2, v2, k, "0"])
for (k1, k2) in it.permutations(chars[:num_symbols], 2)
for (v1, v2) in it.permutations(numbs[:num_symbols - 1], 2)
for k in [k1, k2]
]
input_length = 6
output_window = 2
episode_duration = input_length + output_window
split = int(.80 * len(all_inputs))
# example generation
def example(dataset):
inputs = dataset[np.random.randint(len(dataset))]
targets = inputs[[1 if inputs[0] == inputs[4] else 3]]
return inputs, targets
def training_example():
return example(all_inputs[:split])
def testing_example():
return example(all_inputs[split:])
# reward calculation based on leading LVD at individual steps
def reward(ghu, targets, outputs):
# all or nothing at final time-step
r = np.zeros(len(outputs))
outputs = np.array(
[out for out in outputs[input_length - 1:] if out != separator])
if len(outputs) == len(targets): r[-1] = (targets == outputs).all()
return r
# ################### Sanity check
correct_choices = [
({
"rinp": "rinp<rinp",
"rout": "rout<rout",
"rtmp": "rtmp<rinp"
}, []),
({
"rinp": "rinp<rinp",
"rout": "rout<rinp",
"rtmp": "rtmp<rtmp"
}, []),
({
"rinp": "rinp<rinp",
"rout": "rout<rout",
"rtmp": "rtmp<rinp"
}, ["rout<rtmp"]),
({
"rinp": "rinp<rinp",
"rout": "rout<rinp",
"rtmp": "rtmp<rtmp"
}, []),
({
"rinp": "rinp<rinp",
"rout": "rout<rout",
"rtmp": "rtmp<rinp"
}, ["rout<rtmp"]),
({
"rinp": "rinp<rinp",
"rout": "rout<rtmp",
"rtmp": "rtmp<rtmp"
}, []),
({
"rinp": "rinp<rinp",
"rout": "rout<rinp",
"rtmp": "rtmp<rtmp"
}, []),
({
"rinp": "rinp<rinp",
"rout": "rout<rinp",
"rtmp": "rtmp<rtmp"
}, []),
]
# ################### Sanity check
# Run optimization
avg_rewards, avg_general, grad_norms = reinforce(
ghu,
num_epochs=500,
episode_duration=episode_duration,
training_example=training_example,
testing_example=None,
reward=reward,
task="recall",
learning_rate=.1,
# line_search_iterations = 5,
# distribution_cap = .1,
# likelihood_cap = .7,
distribution_variance_coefficient=.05,
# choices=correct_choices, # perfect rewards with this
verbose=1,
save_file=save_file)
# Assess generalization similarly after run
print("Cloning GHU for generalization...")
ghu_gen = ghu.clone()
print("Sampling problem instances...")
inputs, targets = zip(
*[testing_example() for b in range(ghu_gen.batch_size)])
print("Running on test data...")
outputs, rewards = ghu_gen.run(episode_duration,
inputs,
targets,
reward,
verbose=1)
R_gen = rewards.sum(axis=1)
# Overwrite file dump with R_gen in place of avg_general
with open(save_file, "rb") as f:
result = pk.load(f)
result = list(result)
result[2] = R_gen
result = tuple(result)
with open(save_file, "wb") as f:
pk.dump(result, f)
intranoise_levels = np.arange(0.05, 0.35, 0.05)
num_cs = [4, 8, 12, 16, 20]
for inter in tqdm(internoise_levels, desc="eta_1"):
for intra in tqdm(intranoise_levels, desc="eta_2"):
for num_c in tqdm(num_cs, desc="num_c"):
#print(f"\n### n:{n} num_c:{num_c} ###")
#print("Generation of G ...")
clusters = pd.fixed_clusters(n, num_c)
G, GT, labels = pd.custom_cluster_matrix(n, clusters, inter, 1,
intra, 0)
# epsilon parameter for the approx Alon et al
c = Codec(0.285, 0.285, 1)
c.verbose = False
#print("Generation of compressed/reduced graph ...")
tm = time.time()
k, epsilon, classes, sze_idx, reg_list, nirr = c.compress(
G, "indeg_guided")
red = c.reduced_matrix(G, k, epsilon, classes, reg_list)
t_compression = time.time() - tm
#print(f"s:{t_compression:.2f}")
# Precomputation of the eigenvalues
#print("Eigenvalues of G ... ")
tm = time.time()
G_eig, aux = metrics.eigs(G)
t_G_eig = time.time() - tm
def basic_plastic_trial(num_episodes, save_file):
# Configuration
register_names = ["rinp", "rout", "m"]
layer_sizes = {q: 32 for q in register_names}
hidden_size = 32
rho = .99
plastic = ["rinp<m"]
remove_pathways = ["rinp<rout", "m<rinp", "m<rout", "rout<m"]
# Setup GHU
num_symbols = 3
symbols = [str(a) for a in range(num_symbols)]
pathways, associations = default_initializer(register_names, symbols)
for p in remove_pathways:
pathways.pop(p)
associations = list(
filter(lambda x: x[0] not in remove_pathways, associations))
codec = Codec(layer_sizes, symbols, rho=rho, ortho=True)
controller = Controller(layer_sizes,
pathways,
hidden_size,
plastic,
nonlinearity='relu')
# controller = Controller(layer_sizes, pathways, hidden_size, plastic, nonlinearity='tanh')
ghu = GatedHebbianUnit(layer_sizes,
pathways,
controller,
codec,
batch_size=num_episodes,
plastic=plastic)
ghu.associate(associations)
# Initialize layers
separator = "0"
ghu.fill_layers(separator)
# training example generation
episode_duration = 3
def training_example():
inputs = np.random.choice(symbols[1:], size=2, replace=False)
targets = np.array(["0", "0", inputs[0]])
return inputs, targets
def reward(ghu, targets, outputs):
# All or nothing
r = np.zeros(len(outputs))
if lvd(outputs, targets)[0] == 0: r[-1] = +1.
return r
# ################### Sanity check
inputs = [["2", "1"]]
correct_choices = [
({
"rinp": "rinp<rinp",
"rout": "rout<rout",
"m": "m<m"
}, [1.0]),
({
"rinp": "rinp<m",
"rout": "rout<rout",
"m": "m<m"
}, [0.0]),
({
"rinp": "rinp<rinp",
"rout": "rout<rinp",
"m": "m<m"
}, [0.0]),
]
# ghu.clone().dbg_run(inputs, episode_duration, correct_choices)
# input("???????")
# ################### Sanity check
# Run optimization
avg_rewards, grad_norms = reinforce(
ghu,
num_epochs=500,
episode_duration=episode_duration,
training_example=training_example,
reward=reward,
task="basic_plastic",
learning_rate=.05,
# line_search_iterations = 5,
# distribution_cap = .1,
# likelihood_cap = .7,
# distribution_variance_coefficient = 0.05,
# choices=correct_choices, # perfect rewards with this
verbose=1,
save_file=save_file)
from codec import TreeBuilder, Codec
text = "a dead dad ceded a bad babe a beaded abaca bed"
builder = TreeBuilder(text)
binary_tree = builder.tree()
# on passe l'arbre binaire à un encodeur/décodeur
codec = Codec(binary_tree)
# qui permet d'encoder
encoded = codec.encode(text)
# et de décoder
decoded = codec.decode(encoded)
# si cette assertion est fausse il y a un gros problème avec le code
# on affiche le résultat
print(f"{text}\n{decoded}")
if decoded != text:
print("OOPS")
def echo_trial(episode_duration, save_file):
# Configuration
num_symbols = 10
layer_sizes = {"rinp": 32, "rout": 32, "rtmp": 32}
hidden_size = 32
rho = .99
plastic = []
num_episodes = 250
# Setup GHU
symbols = [str(a) for a in range(num_symbols)]
pathways, associations = default_initializer( # all to all
layer_sizes.keys(), symbols)
codec = Codec(layer_sizes, symbols, rho=rho, ortho=True)
controller = Controller(layer_sizes, pathways, hidden_size, plastic)
ghu = GatedHebbianUnit(layer_sizes,
pathways,
controller,
codec,
batch_size=num_episodes,
plastic=plastic)
ghu.associate(associations)
# Initialize layers
separator = "0"
ghu.fill_layers(separator)
# Generate dataset
input_length = int(episode_duration / 2)
all_inputs = []
for symbol in symbols[1:]:
for t in range(input_length):
inputs = [separator] * input_length
inputs[t] = symbol
all_inputs.append(inputs)
split = int(.80 * len(all_inputs))
# example generation
def example(dataset):
# Randomly choose echo symbol (excluding 0 separator)
inputs = dataset[np.random.randint(len(dataset))]
targets = [inp for inp in inputs if inp != separator]
return inputs, targets
def training_example():
return example(all_inputs[:split])
def testing_example():
return example(all_inputs[split:])
# all or nothing reward
def reward(ghu, targets, outputs):
r = np.zeros(len(outputs))
outputs = np.array(
[out for out in outputs[input_length:] if out != separator])
if len(outputs) == len(targets): r[-1] = (targets == outputs).all()
return r
# # correct choices for debugging
# correct_choices = \
# [({"rinp": "rinp<rout", "rout": "rout<rinp"}, [])]*2 + \
# [({"rinp": "rinp<rinp", "rout": "rout<rout"}, [])]*(episode_duration-2)
# # run it to debug:
# inputs, targets = zip(*[training_example() for b in range(ghu.batch_size)])
# ghu.run(episode_duration, inputs, targets, reward, choices=correct_choices, verbose=3)
# input("????")
# Run optimization
avg_rewards, avg_general, grad_norms = reinforce(
ghu,
num_epochs=200,
episode_duration=episode_duration,
training_example=training_example,
testing_example=testing_example,
reward=reward,
task="echo",
learning_rate=.1,
verbose=1,
save_file=save_file)
return avg_rewards, avg_general, grad_norms
pathways.update({k: ("m", "m") for k in ["inc-m", "dec-m"]})
associations += [(k, str(a), str((a + x) % num_addresses))
for k, x in [("inc-m", 1), ("dec-m", -1)]
for a in range(num_addresses)]
return pathways, associations
if __name__ == "__main__":
layer_sizes = {"r0": 3, "r1": 3}
pathways = {0: ("r0", "r0"), 1: ("r1", "r0"), 2: ("r1", "r1")}
hidden_size = 5
batch_size = 2
codec = Codec(layer_sizes, "01")
controller = Controller(layer_sizes, pathways, hidden_size)
ghu = GatedHebbianUnit(layer_sizes,
pathways,
controller,
codec,
batch_size=batch_size)
ghu.associate([
(0, "0", "0"),
(1, "0", "0"),
(2, "0", "0"),
(0, "1", "1"),
(1, "1", "1"),
(2, "1", "1"),
])
def main():
with open(CONFIG_FN) as f:
conf = json.load(f)
args = parse_args()
kb = load_kb(conf["kb"], 'name')
sys_vocab, sys_word2idx = load_sys_vocab(conf["sys_vocab"])
sys_codec = Codec(sys_vocab, sys_word2idx)
onto, onto_idx = load_ontology(conf["ontology"])
word2idx, embed = load_embed(**conf)
usr_codec = Codec([], word2idx)
trk_model, slot_len_sum = load_tracker_model(onto, embed, conf)
cond_net, generator = load_generator_model(conf, args.cond_net,
args.gen_net, slot_len_sum,
len(sys_vocab))
cond_net.eval()
generator.eval()
trk_model.eval()
hidden = trk_model.state_tracker.init_hidden()
kb_found = Variable(torch.zeros(1, conf["kb_indicator_len"]))
def gen_sent(cond):
'''train one sentence'''
hidden = generator.init_hidden()
inp = Variable(torch.LongTensor([[sys_word2idx['<sos>']]]))
sent_out = []
for i in range(MAX_SENT_LEN):
out, hidden = generator(inp, cond, hidden)
topv, topi = out.data.topk(1)
out_word = int(topi[0][0])
if out_word == sys_word2idx['<eos>']:
break
inp = Variable(torch.LongTensor([[out_word]]))
sent_out.append(out_word)
return sys_codec.decode(sent_out)
for line in iter(sys.stdin.readline, ''):
inp = usr_codec.encode(line.strip())
inp = Variable(torch.LongTensor([
inp,
]))
sentvecs, states_reps, states_preds, hidden = trk_model(
inp, kb_found, hidden)
# print_ret(states_preds)
criteria = to_search_criteria(states_preds, onto)
ret, kb_vec = get_kb_result(kb, criteria, conf["kb_indicator_len"])
sentvecs = sentvecs.view(1, -1)
states_reps = states_reps.view(1, -1)
for slot in states_preds:
states_preds[slot] = states_preds[slot].view(1, -1)
cond = cond_net(sentvecs, states_reps, states_preds, kb_found)
print gen_sent(cond)
def test():
minibatch_size = 100
num_words = 20
tuplesize = 3
num_visible = tuplesize*num_words
num_hidden = 30
codec = Codec(tuplesize, num_words)
tuples = codec.tuples
words = codec.words
encoded = codec.tuples_to_matrix()
(num_data, _) = encoded.shape
print(words)
print('data count: ', num_data)
rbm = RBM(num_visible = num_visible,
num_hidden = num_hidden,
minibatch_size = minibatch_size)
id_indices = numpy.random.randint(low=0, high=num_data, size=minibatch_size)
input_data = T.constant(encoded[id_indices])
#print(input_data)
#print(rbm.propup(input_data).eval())
h1samples = rbm.sample_h_given_v(input_data).eval()
#print(h1samples)
#print(rbm.propdown(h1samples).eval())
v2samples = rbm.sample_v_given_h(h1samples).eval()
#print(v2samples)
(W,H,V) = rbm.contrastive_divergence_1(input_data)
#print(W.eval())
#print(H.eval())
#print(V.eval())
xvis = T.fmatrix('xvis')
h1samples = rbm.sample_h_given_v(xvis)
v2samples = rbm.sample_v_given_h(h1samples)
sample_vhv = theano.function([xvis], v2samples)
example_indices = numpy.random.randint(low=0, high=num_data, size=minibatch_size)
example_input_data = encoded[example_indices]
num_examples = min(10, minibatch_size)
def show_examples():
rec = sample_vhv(example_input_data)
for example in range(num_examples):
print('input words:',
[(t+1, words[idx])
for t in range(tuplesize)
for idx in range(num_words)
if example_input_data[example, t*num_words + idx]])
print('reconstructed words:',
[(t+1, words[idx])
for t in range(tuplesize)
for idx in range(num_words)
if rec[example, t*num_words + idx]])
vis = T.fmatrix('vis')
train = rbm.cd1_fun(vis)
draw = VisualizeWeights('Minibatches', rbm, tuplesize, words, num_hidden)
for epoch in range(1000):
show_examples()
all_vdiffs = numpy.zeros(num_visible)
print('epoch ', epoch)
numpy.random.shuffle(encoded)
for minibatch in range(num_data // minibatch_size):
mb_start = minibatch * minibatch_size;
mb_end = mb_start + minibatch_size;
input_data_indices = numpy.arange(mb_start, mb_end)
encoded_input = encoded[input_data_indices]
input_data = encoded_input
vdiffs = train(input_data)
all_vdiffs = all_vdiffs + numpy.abs(vdiffs)
print('reconstruction error: ', numpy.sum(all_vdiffs) * minibatch_size)
print(numpy.ndarray.astype(rbm.weights.get_value()*100, numpy.int32))
print(numpy.ndarray.astype(rbm.vbias.get_value()*100, numpy.int32))
print(numpy.ndarray.astype(rbm.hbias.get_value()*100, numpy.int32))
draw.epoch_finished(epoch)
def trials(i, avgrew, gradnorm):
print("***************************** Trial ", str(i + 1),
"*******************************")
num_addresses = 4
register_names = ["rinp", "rout"]
num_episodes = 5000
layer_sizes = {q: 128 for q in register_names + ["m"]}
hidden_size = 32
# plastic = ["%s<m"%q for q in register_names]
plastic = ["rinp<m"]
symbols = [str(a) for a in range(num_addresses)]
pathways, associations = turing_initializer(register_names, num_addresses)
# constrain pathways inductive bias
remove_pathways = ["rout<m", "m<rout"]
for p in remove_pathways:
pathways.pop(p)
associations = list(
filter(lambda x: x[0] not in remove_pathways, associations))
codec = Codec(layer_sizes, symbols, rho=.999)
controller = Controller(layer_sizes, pathways, hidden_size, plastic)
# Sanity check
ghu = GatedHebbianUnit(layer_sizes,
pathways,
controller,
codec,
plastic=plastic,
batch_size=num_episodes)
ghu.associate(associations)
# Initialize layers
separator = "0"
ghu.fill_layers(separator)
# training example generation
list_symbols = 4
min_length = 3
max_length = 3
def training_example():
list_length = np.random.randint(min_length, max_length + 1)
inputs = np.array([separator] * (list_length))
inputs[:] = np.random.choice(symbols[1:list_symbols],
size=list_length,
replace=False)
targets = inputs[::-1]
return inputs, targets
# # reward calculation from LVD
# def reward(ghu, targets, outputs):
# # Assess reward: negative LVD after separator filtering
# outputs_ = [out for out in outputs if out != separator]
# l, _ = lvd(outputs_, targets)
# return -l
def reward(ghu, targets, outputs):
outputs_ = outputs[len(targets) - 1:]
#zeros = [o for o in outputs[len(targets)-1:] if o==separator]
#totzeros = len(zeros)
r = np.zeros(len(outputs))
# if len(outputs_)==0:
# r[-1] -= 2*(len(outputs[len(targets)-1:])+1)
# else:
_, d = lvd(outputs_, targets)
for i in range(1, d.shape[0]):
r[-1] += 1. if (i < d.shape[1]
and d[i, i] == d[i - 1, i - 1]) else -2.
#r[-1] -= 0.1*totzeros
return r
filename = "reverse" + str(i + 1) + ".png"
avg_rewards, grad_norms = reinforce(ghu,
num_epochs=1000,
episode_duration=2 * max_length - 1,
training_example=training_example,
reward=reward,
task="reverse",
learning_rate=.03,
verbose=2)
gradnorm[i + 1] = grad_norms.tolist()
avgrew[i + 1] = avg_rewards.tolist()
pt.subplot(2, 1, 1)
pt.plot(avg_rewards)
pt.title("Learning curve of reverse")
pt.ylabel("Avg Reward")
pt.subplot(2, 1, 2)
pt.plot(grad_norms)
pt.xlabel("Epoch")
pt.ylabel("||Grad||")
pt.savefig(filename)
def build(self, filename):
# check the package has .desktop files
pkg = Package(filename)
for b in self.cfg.get_package_blacklist():
if fnmatch.fnmatch(pkg.name, b):
pkg.log.write(LoggerItem.INFO, "package is blacklisted")
return
# set up state
if not os.path.exists('./appstream'):
os.makedirs('./appstream')
if not os.path.exists('./icons'):
os.makedirs('./icons')
if not os.path.exists('./screenshot-cache'):
os.makedirs('./screenshot-cache')
if not os.path.exists('./screenshots'):
os.makedirs('./screenshots')
if not os.path.exists('./screenshots/source'):
os.makedirs('./screenshots/source')
for size in self.cfg.get_screenshot_thumbnail_sizes():
path = './screenshots/' + str(size[0]) + 'x' + str(size[1])
if not os.path.exists(path):
os.makedirs(path)
# remove tmp
if os.path.exists('./tmp'):
shutil.rmtree('./tmp')
os.makedirs('./tmp')
# decompress main file and search for desktop files
package_decompress(pkg)
files = []
for f in self.cfg.get_interesting_installed_files():
files.extend(glob.glob("./tmp" + f))
files.sort()
# we only need to install additional files if we're not running on
# the builders
for c in self.cfg.get_package_data_list():
if fnmatch.fnmatch(pkg.name, c[0]):
extra_files = glob.glob("./packages/%s*.rpm" % c[1])
for f in extra_files:
extra_pkg = Package(f)
pkg.log.write(LoggerItem.INFO, "adding extra package %s" % extra_pkg.name)
package_decompress(extra_pkg)
# check for duplicate apps in the package
self.has_valid_content = False
# check for codecs
if pkg.name.startswith('gstreamer'):
app = Codec(pkg, self.cfg)
if app.parse_files(files):
if self.add_application(app):
self.add_completed(pkg, app)
else:
# process each desktop file in the original package
for f in files:
pkg.log.write(LoggerItem.INFO, "reading %s" % f)
fi = Gio.file_new_for_path(f)
info = fi.query_info('standard::content-type', 0, None)
# create the right object depending on the content type
content_type = info.get_content_type()
if content_type == 'inode/symlink':
continue
if content_type == 'application/x-font-ttf':
app = FontFile(pkg, self.cfg)
elif content_type == 'application/x-font-otf':
app = FontFile(pkg, self.cfg)
elif content_type == 'application/x-desktop':
app = DesktopFile(pkg, self.cfg)
elif content_type == 'application/xml':
app = InputMethodComponent(pkg, self.cfg)
elif content_type == 'application/x-sqlite3':
app = InputMethodTable(pkg, self.cfg)
else:
pkg.log.write(LoggerItem.INFO, "content type %s not supported" % content_type)
continue
# the ID is the filename
app_id = os.path.basename(f).decode('utf-8')
app.set_id(app_id)
# parse file
if not app.parse_file(f):
continue
# write the application
if self.add_application(app):
self.add_completed(pkg, app)
def main(args, arch):
adv_models = None
train_loader, test_loader = create_loaders(args, root='../data')
if args.dataset == 'cifar':
args.nc, args.h, args.w = 3, 32, 32
args.input_size = 32
model, l_test_classif_paths = load_all_classifiers(args, load_archs=[args.source_arch])
model_type = args.source_arch
if args.target_arch is not None:
model_target, l_test_classif_paths = load_all_classifiers(args, load_archs=[args.target_arch])
model_type = args.target_arch
del model_target
torch.cuda.empty_cache()
elif args.dataset == 'mnist':
args.input_size = 28
if args.source_arch == 'natural':
model, l_test_classif_paths = load_all_classifiers(args, load_archs=["natural"])
model_type = 'natural'
elif args.source_arch == 'ens_adv':
adv_model_names = args.adv_models
adv_models = [None] * len(adv_model_names)
for i in range(len(adv_model_names)):
type = get_model_type(adv_model_names[i])
adv_models[i] = load_model(args, adv_model_names[i], type=type).to(args.dev)
path = os.path.join(args.dir_test_models, "pretrained_classifiers",
args.dataset, "ensemble_adv_trained", args.model)
model = load_model(args, args.model, type=args.type)
l_test_classif_paths = [path]
model_type = 'Ensemble Adversarial'
model.to(args.dev)
model.eval()
test_classifier(args, model, args.dev, test_loader, epoch=1)
print("Testing on %d Test Classifiers" %(len(l_test_classif_paths)))
# attack related settings
if args.attack_method == "zoo" or args.attack_method == "autozoom_bilin":
if args.img_resize is None:
args.img_resize = args.input_size
print("Argument img_resize is not set and not using autoencoder, set to image original size:{}".format(
args.img_resize))
codec = None
if args.attack_method == "zoo_ae" or args.attack_method == "autozoom_ae":
codec = Codec(args.input_size, IN_CHANNELS[args.dataset],
args.compress_mode, args.resize, use_tanh=args.use_tanh)
codec.load_codec(args,codec_path)
codec.cuda()
decoder = codec.decoder
args.img_resize = decoder.input_shape[1]
print("Loading autoencoder: {}, set the attack image size to:{}".format(args.codec_path, args.img_resize))
# setup attack
if args.attack_method == "zoo":
blackbox_attack = ZOO(model, args.dataset, args)
elif args.attack_method == "zoo_ae":
blackbox_attack = ZOO_AE(model, args.dataset, args, decoder)
elif args.attack_method == "autozoom_bilin":
blackbox_attack = AutoZOOM_BiLIN(model, args.dataset, args)
elif args.attack_method == "autozoom_ae":
blackbox_attack = AutoZOOM_AE(model, args["dataset"], args, decoder)
target_str = "untargeted" if args.attack_type!="targeted" else "targeted_{}".format(args.target_type)
attack_framework = AutoZoomAttackFramework(args, test_loader)
attack_framework.attack_dataset_images(args, blackbox_attack, arch, model,
codec, l_test_classif_paths=l_test_classif_paths,
adv_models=adv_models)
model.cpu()
import time
n = 100
repetitions = 5
inter_s = [0, 0.05, 0.1, 0.2, 0.3]
intra_s = inter_s
num_cs = [2, 4, 8, 10, 16, 20, 40]
ksize = 23
imbalanced = True
inter_v = 1
intra_v = 0
refinement = 'indeg_guided'
cdc = Codec(0, 0.5, 20)
cdc.fast_search = False
s = Stats("/tmp/test.csv")
for repetition in range(repetitions):
for inter in inter_s:
for intra in intra_s:
for num_c in num_cs:
print(
pu.to_header(
f"r:{repetition+1}/{repetitions} n:{n} num_c:{num_c} inter:{inter} intra:{intra}"
))
tm = time.time()
clusters = pd.random_clusters(n, num_c)
G, GT, labeling = pd.custom_cluster_matrix(
try:
GT = data['GT']
labeling = data['labels']
except:
print("[x] No GT found, No labeling found")
G = data['G']
n = G.shape[0]
repetitions = 1
refinement = 'indeg_guided'
ksize = 23
s = Stats(f"/tmp/{dset_name}.csv")
c = Codec(0.1, 0.4, 40)
for r in range(repetitions):
print(f"### r={r} ###")
tm = time.time()
k, epsilon, classes, sze_idx, reg_list, nirr = c.compress(G, refinement)
tcompression = time.time() - tm
sze = c.decompress(G, 0, classes, k, reg_list)
tdecompression = time.time() - tm
fsze = c.post_decompression(sze, ksize)
tpostdecompression = time.time() - tm
red = c.reduced_matrix(G, k, epsilon, classes, reg_list)
for n in tqdm([500, 1000, 1500, 2000, 2500, 3000, 3500, 4000, 4500, 5000, 5500, 6000, 6500, 7000, 7500, 8000, 8500, 9000, 9500, 10000], desc="Dimension G"):
# First graph
tm = time.time()
clusters1 = proc_d.random_clusters(n, 8)
G, GT, labeling = proc_d.custom_cluster_matrix(n, clusters1, 0.3, 1, 0.2, 0)
tgen = time.time() -tm
#print(f"TIME: tgen:{tgen:.2f}")
# Second graph
clusters2 = proc_d.random_clusters(n, 8)
G2, GT2, labeling2 = proc_d.custom_cluster_matrix(n, clusters2, 0.6, 1, 0.1, 0)
# Compression G1
c = Codec(0.285, 0.285, 1)
c.verbose = False
tm = time.time()
k, epsilon, classes, sze_idx, reg_list, nirr = c.compress(G, 'indeg_guided')
tcompression1 = time.time() - tm
#print(f"TIME: tcompression1:{tcompression1:.2f}")
# Compression G1
c = Codec(0.285, 0.285, 1)
c.verbose = False
tm = time.time()
k2, epsilon, classes2, sze_idx, reg_list2, nirr2 = c.compress(G2, 'indeg_guided')
tcompression2 = time.time() - tm
#print(f"TIME: tcompression2:{tcompression2:.2f}")
# Reduced graph
GT = []
labels = []
try:
GT = data['GT']
labeling = data['labels']
except:
print("[x] No GT found, No labeling found")
refinement = 'indeg_guided'
ksize = 23
s = Stats(f"/tmp/{dset_name}.csv")
c = Codec(0, 0.5, 20)
c.verbose = False
#f = []
#g = []
#thresholds = np.arange(0,0.7, 0.05)
thresholds = np.arange(0, 0.5, 0.05)
for t in thresholds:
print(f"### t={t} ###")
#G = (data['G']*10**4)>(t*np.max(data['G']))
G = data['G'] > 0
n = G.shape[0]
k, epsilon, classes, sze_idx, reg_list, nirr = c.compress(
return s
if __name__ == "__main__":
# GHU settings
num_symbols = 3
layer_sizes = {"rinp": 64, "rout": 64}
hidden_size = 16
plastic = []
symbols = [str(a) for a in range(num_symbols)]
pathways, associations = default_initializer( # all to all
layer_sizes.keys(), symbols)
codec = Codec(layer_sizes, symbols, rho=.9)
controller = Controller(layer_sizes, pathways, hidden_size, plastic)
# Sanity check
ghu = GatedHebbianUnit(layer_sizes,
pathways,
controller,
codec,
plastic=plastic)
ghu.associate(associations)
for p, s, t in associations:
q, r = ghu.pathways[p]
assert (codec.decode(q, tr.mv(ghu.W[p], codec.encode(r, s))) == t)
# Initialize layers
ghu.v[0]["rinp"] = codec.encode("rinp", "0")
##############################################
############## Main script code ##############
##############################################
refinement = 'indeg_guided'
vals = [0.1, 0.2, 0.5, 0.7]
first = True
n = 1000
# Query graph, the bigger one
clusters = pd.random_clusters(n, 8)
inter = 0.1
intra = 0.1
oG, oGT, olabeling = pd.custom_cluster_matrix(n, clusters, inter, 1, intra, 0)
c = Codec(0.15, 0.5, 5)
k, epsilon, classes, sze_idx, reg_list, nirr = c.compress(oG, refinement)
query = c.reduced_matrix(oG, k, epsilon, classes, reg_list)
# Database
c = Codec(0.2, 0.5, 5)
num_c = 8
for r in range(5):
for i in vals:
clusters = pd.random_clusters(n, num_c)
G, GT, labeling = pd.custom_cluster_matrix(n, clusters, i, 1, i, 0)
k, epsilon, classes, sze_idx, reg_list, nirr = c.compress(
G, refinement)
red = c.reduced_matrix(G, k, epsilon, classes, reg_list)
def test():
minibatch_size = 100
num_words = 40
tuplesize = 5
num_visible = tuplesize*num_words
num_hidden = 140
codec = Codec(tuplesize, num_words)
tuples = codec.tuples
words = codec.words
encoded = codec.tuples_to_matrix()
(num_data, _) = encoded.shape
print(words)
print('data count: ', num_data)
rbm = RBM(num_visible = num_visible,
num_hidden = num_hidden,
minibatch_size = minibatch_size,
venabledp=1.0,
henabledp=0.7)
id_indices = numpy.random.randint(low=0, high=num_data, size=minibatch_size)
input_data = T.constant(encoded[id_indices])
#print(input_data)
#print(rbm.propup(input_data).eval())
#h1samples = rbm.sample_h_given_v(input_data).eval()
#print(h1samples)
#print(rbm.propdown(h1samples).eval())
#v2samples = rbm.sample_v_given_h(h1samples).eval()
#print(v2samples)
#(W,H,V) = rbm.contrastive_divergence_1(input_data)
#print(W.eval())
#print(H.eval())
#print(V.eval())
all_h_enabled = numpy.ones(num_hidden)
all_v_enabled = numpy.ones(num_visible)
xvis = T.fmatrix('xvis')
h1samples = rbm.sample_h_given_v(xvis, all_h_enabled)
v2samples = rbm.sample_v_given_h(h1samples, all_v_enabled)
sample_vhv = theano.function([xvis], v2samples)
example_indices = numpy.random.randint(low=0, high=num_data, size=minibatch_size)
example_input_data = encoded[example_indices]
num_examples = min(10, minibatch_size)
def show_examples():
rec = sample_vhv(example_input_data)
for example in range(num_examples):
print('input words:',
[(t+1, words[idx])
for t in range(tuplesize)
for idx in range(num_words)
if example_input_data[example, t*num_words + idx]])
print('reconstructed words:',
[(t+1, words[idx])
for t in range(tuplesize)
for idx in range(num_words)
if rec[example, t*num_words + idx]])
def report_hidden():
weights = rbm.weights.get_value()
for h in range(num_hidden):
print('hidden ', h)
for block in range(tuplesize):
for word in range(num_words):
w = weights[block*num_words+word, h]
if w > 0.5:
print(' %2i %8s %4.1f' % (block, words[word], w))
vis = T.fmatrix('vis')
train = rbm.cd1_fun(vis)
draw = VisualizeWeights('Dropout (vp:%4.2f, hp:%4.2f)' % (rbm.venabledp, rbm.henabledp),
rbm, tuplesize, words, num_hidden,
num_visible)
for epoch in range(1000):
show_examples()
all_vdiffs = numpy.zeros(num_visible)
print('epoch ', epoch)
numpy.random.shuffle(encoded)
for minibatch in range(num_data // minibatch_size):
mb_start = minibatch * minibatch_size;
mb_end = mb_start + minibatch_size;
input_data_indices = numpy.arange(mb_start, mb_end)
encoded_input = encoded[input_data_indices]
input_data = encoded_input
(vdiffs, venabled, henabled) = train(input_data)
all_vdiffs = all_vdiffs + numpy.abs(vdiffs)
#print('venabled', venabled)
#print('henabled', henabled)
print('reconstruction error: ', numpy.sum(all_vdiffs) * minibatch_size)
#print(numpy.ndarray.astype(rbm.weights.get_value()*100, numpy.int32))
#print(numpy.ndarray.astype(rbm.vbias.get_value()*100, numpy.int32))
#print(numpy.ndarray.astype(rbm.hbias.get_value()*100, numpy.int32))
draw.epoch_finished(epoch)
report_hidden()
# plastic = ["%s<m"%q for q in register_names]
plastic = ["rinp<m"]
symbols = [str(a) for a in range(num_addresses)]
pathways, associations = turing_initializer(register_names, num_addresses)
# constrain pathways inductive bias
remove_pathways = ["rout<m", "m<rout"]
for p in remove_pathways:
pathways.pop(p)
associations = list(
filter(lambda x: x[0] not in remove_pathways, associations))
# print(pathways)
# print(associations)
codec = Codec(layer_sizes, symbols, rho=.999)
controller = Controller(layer_sizes, pathways, hidden_size, plastic)
# controller = Controller(layer_sizes, pathways, hidden_size, input_keys=["m","rinp"])
# controller.rnn.weight_hh_l0.data = 1*tr.eye(hidden_size) # favor repeated actions
# Sanity check
ghu = GatedHebbianUnit(layer_sizes,
pathways,
controller,
codec,
plastic=plastic)
ghu.associate(associations)
for p, s, t in associations:
q, r = ghu.pathways[p]
assert (codec.decode(q, tr.mv(ghu.W[p], codec.encode(r, s))) == t)
ghu_init = ghu
class FIXConnectionHandler(object):
def __init__(self, engine, protocol, sock=None, addr=None, observer=None):
self.codec = Codec(protocol)
self.engine = engine
self.connectionState = ConnectionState.CONNECTED
self.session = None
self.addr = addr
self.observer = observer
self.msgBuffer = b''
self.heartbeatPeriod = 30.0
self.msgHandlers = []
self.sock = sock
self.heartbeatTimerRegistration = None
self.expectedHeartbeatRegistration = None
self.socketEvent = FileDescriptorEventRegistration(
self.handle_read, sock, EventType.READ)
self.engine.eventManager.registerHandler(self.socketEvent)
def address(self):
return self.addr
def disconnect(self):
self.handle_close()
def _notifyMessageObservers(self, msg, direction, persistMessage=True):
if persistMessage is True:
self.engine.journaller.persistMsg(msg, self.session, direction)
for handler in filter(
lambda x: (x[1] is None or x[1] == direction) and
(x[2] is None or x[2] == msg.msgType), self.msgHandlers):
handler[0](self, msg)
def addMessageHandler(self, handler, direction=None, msgType=None):
self.msgHandlers.append((handler, direction, msgType))
def removeMessageHandler(self, handler, direction=None, msgType=None):
remove = filter(
lambda x: x[0] == handler and
(x[1] == direction or direction is None) and
(x[2] == msgType or msgType is None), self.msgHandlers)
for h in remove:
self.msgHandlers.remove(h)
def _sendHeartbeat(self):
self.sendMsg(self.codec.protocol.messages.Messages.heartbeat())
def _expectedHeartbeat(self, type, closure):
logging.warning("Expected heartbeat from peer %s" %
(self.expectedHeartbeatRegistration, ))
self.sendMsg(self.codec.protocol.messages.Messages.test_request())
def registerLoggedIn(self):
self.heartbeatTimerRegistration = TimerEventRegistration(
lambda type, closure: self._sendHeartbeat(), self.heartbeatPeriod)
self.engine.eventManager.registerHandler(
self.heartbeatTimerRegistration)
# register timeout for 10% more than we expect
self.expectedHeartbeatRegistration = TimerEventRegistration(
self._expectedHeartbeat, self.heartbeatPeriod * 1.10)
self.engine.eventManager.registerHandler(
self.expectedHeartbeatRegistration)
def registerLoggedOut(self):
if self.heartbeatTimerRegistration is not None:
self.engine.eventManager.unregisterHandler(
self.heartbeatTimerRegistration)
self.heartbeatTimerRegistration = None
if self.expectedHeartbeatRegistration is not None:
self.engine.eventManager.unregisterHandler(
self.expectedHeartbeatRegistration)
self.expectedHeartbeatRegistration = None
def _handleResendRequest(self, msg):
protocol = self.codec.protocol
responses = []
beginSeqNo = msg[protocol.fixtags.BeginSeqNo]
endSeqNo = msg[protocol.fixtags.EndSeqNo]
if int(endSeqNo) == 0:
endSeqNo = sys.maxsize
logging.info("Received resent request from %s to %s", beginSeqNo,
endSeqNo)
replayMsgs = self.engine.journaller.recoverMsgs(
self.session, MessageDirection.OUTBOUND, beginSeqNo, endSeqNo)
gapFillBegin = int(beginSeqNo)
gapFillEnd = int(beginSeqNo)
for replayMsg in replayMsgs:
msgSeqNum = int(replayMsg[protocol.fixtags.MsgSeqNum])
if replayMsg[protocol.fixtags.
MsgType] in protocol.msgtype.sessionMessageTypes:
gapFillEnd = msgSeqNum + 1
else:
if self.engine.shouldResendMessage(self.session, replayMsg):
if gapFillBegin < gapFillEnd:
# we need to send a gap fill message
gapFillMsg = FIXMessage(protocol.msgtype.SEQUENCERESET)
gapFillMsg.setField(protocol.fixtags.GapFillFlag, 'Y')
gapFillMsg.setField(protocol.fixtags.MsgSeqNum,
gapFillBegin)
gapFillMsg.setField(protocol.fixtags.NewSeqNo,
str(gapFillEnd))
responses.append(gapFillMsg)
# and then resent the replayMsg
replayMsg.removeField(protocol.fixtags.BeginString)
replayMsg.removeField(protocol.fixtags.BodyLength)
replayMsg.removeField(protocol.fixtags.SendingTime)
replayMsg.removeField(protocol.fixtags.SenderCompID)
replayMsg.removeField(protocol.fixtags.TargetCompID)
replayMsg.removeField(protocol.fixtags.CheckSum)
replayMsg.setField(protocol.fixtags.PossDupFlag, "Y")
responses.append(replayMsg)
gapFillBegin = msgSeqNum + 1
else:
gapFillEnd = msgSeqNum + 1
responses.append(replayMsg)
if gapFillBegin < gapFillEnd:
# we need to send a gap fill message
gapFillMsg = FIXMessage(protocol.msgtype.SEQUENCERESET)
gapFillMsg.setField(protocol.fixtags.GapFillFlag, 'Y')
gapFillMsg.setField(protocol.fixtags.MsgSeqNum, gapFillBegin)
gapFillMsg.setField(protocol.fixtags.NewSeqNo, str(gapFillEnd))
responses.append(gapFillMsg)
return responses
def handle_read(self, type, closure):
protocol = self.codec.protocol
try:
msg = self.sock.recv(8192)
if msg:
self.msgBuffer = self.msgBuffer + msg
(decodedMsg, parsedLength) = self.codec.decode(self.msgBuffer)
self.msgBuffer = self.msgBuffer[parsedLength:]
while decodedMsg is not None and self.connectionState != ConnectionState.DISCONNECTED:
self.processMessage(decodedMsg)
(decodedMsg,
parsedLength) = self.codec.decode(self.msgBuffer)
self.msgBuffer = self.msgBuffer[parsedLength:]
if self.expectedHeartbeatRegistration is not None:
self.expectedHeartbeatRegistration.reset()
else:
logging.debug("Connection has been closed")
self.disconnect()
except ConnectionError as why:
logging.debug("Connection has been closed %s" % (why, ))
self.disconnect()
def handleSessionMessage(self, msg):
return -1
def processMessage(self, decodedMsg):
protocol = self.codec.protocol
beginString = decodedMsg[protocol.fixtags.BeginString]
if beginString != protocol.beginstring:
logging.warning(
"FIX BeginString is incorrect (expected: %s received: %s)",
(protocol.beginstring, beginString))
self.disconnect()
return
msgType = decodedMsg[protocol.fixtags.MsgType]
try:
responses = []
if msgType in protocol.msgtype.sessionMessageTypes:
(recvSeqNo, responses) = self.handleSessionMessage(decodedMsg)
else:
recvSeqNo = decodedMsg[protocol.fixtags.MsgSeqNum]
# validate the seq number
(seqNoState,
lastKnownSeqNo) = self.session.validateRecvSeqNo(recvSeqNo)
if seqNoState is False:
# We should send a resend request
logging.info("Requesting resend of messages: %s to %s" %
(lastKnownSeqNo, 0))
responses.append(
protocol.messages.Messages.resend_request(
lastKnownSeqNo, 0))
# we still need to notify if we are processing Logon message
if msgType == protocol.msgtype.LOGON:
self._notifyMessageObservers(decodedMsg,
MessageDirection.INBOUND,
False)
else:
self.session.setRecvSeqNo(recvSeqNo)
self._notifyMessageObservers(decodedMsg,
MessageDirection.INBOUND)
for m in responses:
self.sendMsg(m)
except SessionWarning as sw:
logging.warning(sw)
except SessionError as se:
logging.error(se)
self.disconnect()
except DuplicateSeqNoError:
try:
if decodedMsg[protocol.fixtags.PossDupFlag] == "Y":
logging.debug(
"Received duplicate message with PossDupFlag set")
except KeyError:
pass
finally:
logging.error(
"Failed to process message with duplicate seq no (MsgSeqNum: %s) (and no PossDupFlag='Y') - disconnecting"
% (recvSeqNo, ))
self.disconnect()
def handle_close(self):
if self.connectionState != ConnectionState.DISCONNECTED:
logging.info("Client disconnected")
self.registerLoggedOut()
self.sock.close()
self.connectionState = ConnectionState.DISCONNECTED
self.msgHandlers.clear()
if self.observer is not None:
self.observer.notifyDisconnect(self)
self.engine.eventManager.unregisterHandler(self.socketEvent)
def sendMsg(self, msg):
if self.connectionState != ConnectionState.CONNECTED and self.connectionState != ConnectionState.LOGGED_IN:
raise FIXException(FIXException.FIXExceptionReason.NOT_CONNECTED)
encodedMsg = self.codec.encode(msg, self.session).encode('utf-8')
self.sock.send(encodedMsg)
if self.heartbeatTimerRegistration is not None:
self.heartbeatTimerRegistration.reset()
decodedMsg, junk = self.codec.decode(encodedMsg)
try:
self._notifyMessageObservers(decodedMsg, MessageDirection.OUTBOUND)
except DuplicateSeqNoError:
logging.error(
"We have sent a message with a duplicate seq no, failed to persist it (MsgSeqNum: %s)"
% (decodedMsg[self.codec.protocol.fixtags.MsgSeqNum]))
if __name__ == "__main__":
print("*******************************************************")
# Configuration
num_symbols = 4
layer_sizes = {"rinp": 64, "rout": 64, "rtmp": 64}
hidden_size = 32
rho = .99
plastic = []
num_episodes = 1000
# Setup GHU
symbols = [str(a) for a in range(num_symbols)]
pathways, associations = default_initializer( # all to all
layer_sizes.keys(), symbols)
codec = Codec(layer_sizes, symbols, rho=rho)
controller = Controller(layer_sizes, pathways, hidden_size, plastic)
ghu = GatedHebbianUnit(layer_sizes,
pathways,
controller,
codec,
batch_size=num_episodes,
plastic=plastic)
ghu.associate(associations)
# Initialize layers
separator = "0"
ghu.fill_layers(separator)
# training example generation
def training_example():
def reverse_trial(num_episodes, save_file):
# Configuration
register_names = ["rinp", "rout"]
layer_sizes = {q: 64 for q in register_names + ["m"]}
hidden_size = 64
rho = .99
plastic = ["rout<m"]
# remove_pathways = ["rinp<rout", "m<rinp", "m<rout", "rout<m"]
remove_pathways = ["rinp<rout", "rinp<m", "m<rinp", "m<rout"]
# remove_pathways = []
input_keys = None
# Setup GHU
num_addresses = 4
symbols = [str(a) for a in range(num_addresses)]
pathways, associations = turing_initializer(register_names, num_addresses)
for p in remove_pathways:
pathways.pop(p)
associations = list(
filter(lambda x: x[0] not in remove_pathways, associations))
codec = Codec(layer_sizes, symbols, rho=rho, ortho=True)
controller = Controller(layer_sizes,
pathways,
hidden_size,
plastic,
input_keys=input_keys,
nonlinearity='relu')
# controller = Controller(layer_sizes, pathways, hidden_size, plastic, nonlinearity='tanh')
ghu = GatedHebbianUnit(layer_sizes,
pathways,
controller,
codec,
batch_size=num_episodes,
plastic=plastic)
ghu.associate(associations)
# Initialize layers
separator = "0"
ghu.fill_layers(separator)
# Dataset of all possible input lists
min_length = 4
max_length = 4
episode_duration = 2 * max_length
all_inputs = [
np.array(inputs + (separator, ))
for list_length in range(min_length, max_length + 1)
for inputs in it.product(symbols[1:], repeat=list_length)
]
split = int(.80 * len(all_inputs))
# example generation
def example(dataset):
inputs = dataset[np.random.randint(len(dataset))]
targets = inputs[:-1][::-1]
return inputs, targets
def training_example():
return example(all_inputs[:split])
def testing_example():
return example(all_inputs[split:])
# all or nothing reward calculation
def reward(ghu, targets, outputs):
r = np.zeros(len(outputs))
outputs = outputs[len(targets) - 1:]
outputs = np.array([out for out in outputs if out != separator])
if len(outputs) == len(targets): r[-1] = (outputs == targets).all()
return r
# ################### Sanity check
correct_choices = [
({
"rinp": "rinp<rinp",
"rout": "rout<rinp",
"m": "inc-m"
}, []),
({
"rinp": "rinp<rinp",
"rout": "rout<rinp",
"m": "inc-m"
}, ["rout<m"]),
({
"rinp": "rinp<rinp",
"rout": "rout<rinp",
"m": "inc-m"
}, ["rout<m"]),
({
"rinp": "rinp<rinp",
"rout": "rout<rinp",
"m": "inc-m"
}, ["rout<m"]),
({
"rinp": "rinp<rinp",
"rout": "rout<rout",
"m": "dec-m"
}, []),
({
"rinp": "rinp<rinp",
"rout": "rout<m",
"m": "dec-m"
}, []),
({
"rinp": "rinp<rinp",
"rout": "rout<m",
"m": "dec-m"
}, []),
({
"rinp": "rinp<rinp",
"rout": "rout<m",
"m": "dec-m"
}, []),
]
# ################### Sanity check
# Run optimization
avg_rewards, _, grad_norms = reinforce(
ghu,
num_epochs=500,
episode_duration=episode_duration,
training_example=training_example,
# testing_example = testing_example,
testing_example=None,
reward=reward,
task="reverse",
learning_rate=.1,
# line_search_iterations = 5,
# distribution_cap = .1,
# likelihood_cap = .7,
distribution_variance_coefficient=0.05,
# choices = correct_choices, # perfect reward with this
verbose=1,
save_file=save_file)
# Assess generalization similarly after run
print("Cloning GHU for generalization...")
ghu_gen = ghu.clone()
print("Sampling problem instances...")
inputs, targets = zip(
*[testing_example() for b in range(ghu_gen.batch_size)])
print("Running on test data...")
outputs, rewards = ghu_gen.run(episode_duration,
inputs,
targets,
reward,
verbose=1)
R_gen = rewards.sum(axis=1)
# Overwrite file dump with R_gen in place of avg_general
with open(save_file, "rb") as f:
result = pk.load(f)
result = list(result)
result[2] = R_gen
result = tuple(result)
with open(save_file, "wb") as f:
pk.dump(result, f)
def test():
num_words = 20
tuplesize = 3
num_visible = tuplesize*num_words
num_hidden = 30
codec = Codec(tuplesize, num_words)
tuples = codec.tuples
words = codec.words
encoded = codec.tuples_to_matrix()
(num_data, _) = encoded.shape
print('data count: ', num_data)
rbm = RBM1(num_visible = num_visible,
num_hidden = num_hidden)
input_data = T.constant(encoded[2])
#print(pairs[2])
#print(encoded[2])
#print(input_data)
#print(rbm.propup(input_data).eval())
h1samples = rbm.sample_h_given_v(input_data).eval()
#print(h1samples)
#print(rbm.propdown(h1samples).eval())
v2samples = rbm.sample_v_given_h(h1samples).eval()
#print(v2samples)
(W,H,V) = rbm.contrastive_divergence_1(input_data)
#print(W.eval())
xvis = T.fvector('xvis')
h1samples = rbm.sample_h_given_v(xvis)
v2samples = rbm.sample_v_given_h(h1samples)
sample_vhv = theano.function([xvis], v2samples)
num_examples = 20
example_indices = numpy.random.randint(low=0, high=num_data, size=num_examples)
def show_examples():
for example in example_indices:
dat = encoded[example]
v2samples = sample_vhv(dat)
print('input words:',
[(t+1, words[idx])
for t in range(tuplesize)
for idx in range(num_words)
if encoded[example, t*num_words + idx]])
print('reconstructed words:',
[(t+1, words[idx])
for t in range(tuplesize)
for idx in range(num_words)
if v2samples[t*num_words + idx]])
print('')
def report_hidden():
weights = rbm.weights.get_value()
for h in range(num_hidden):
print('hidden ', h)
for block in range(tuplesize):
for word in range(num_words):
w = weights[block*num_words+word, h]
if w > 0.5:
print(' %2i %8s %4.1f' % (block, words[word], w))
vis = T.fvector('vis')
train = rbm.cd1_fun(vis)
input_data = numpy.reshape(encoded[2],
num_visible)
train(input_data)
print(rbm.weights.get_value())
draw = VisualizeWeights('', rbm, tuplesize, words, num_hidden)
for epoch in range(500):
show_examples()
all_vdiffs = numpy.zeros(num_visible)
print('epoch ', epoch)
for i in range(num_data):
input_data = numpy.reshape(encoded[i],
num_visible)
vdiffs = train(input_data)
all_vdiffs = all_vdiffs + numpy.abs(vdiffs)
print('reconstruction error: ', numpy.sum(all_vdiffs))
print(T.cast(rbm.weights.get_value()*100, 'int32').eval())
draw.epoch_finished(epoch)
report_hidden()
file_handler_except = log.FileHandler(sigs.except_path)
file_handler_except.setLevel(log.ERROR)
file_handler_except.setFormatter(formatter_except)
logger_except.addHandler(file_handler_except)
# </editor-fold>
# <editor-fold desc="globals"
BP = g.BREAK_POINT
SET = g.OP_SET
PUZZLE_DEFAULT =\
'.....627........5....54..311.....5..97.8.2.64..6.....764..29....8........526.....'
# </editor-fold>
codec = Codec()
encoded_puzzle = ''
solved_puzzle = ''
def load(cb, puz=PUZZLE_DEFAULT):
try:
global codec
global encoded_puzzle
global solved_puzzle
if sigs.is_grid:
puzzle_list = puz
num_list = [
item if len(item) == 1 else '.' for item in puzzle_list
]
def swap_trial(distribution_variance_coefficient, save_file):
# Configuration
num_symbols = 10
layer_sizes = {"rinp": 32, "rout": 32, "rtmp": 32}
hidden_size = 32
rho = .99
plastic = []
num_episodes = 100
# Setup GHU
symbols = [str(a) for a in range(num_symbols)]
pathways, associations = default_initializer( # all to all
layer_sizes.keys(), symbols)
codec = Codec(layer_sizes, symbols, rho=rho, ortho=True)
controller = Controller(layer_sizes, pathways, hidden_size, plastic)
ghu = GatedHebbianUnit(layer_sizes,
pathways,
controller,
codec,
batch_size=num_episodes,
plastic=plastic)
ghu.associate(associations)
# Initialize layers
separator = "0"
ghu.fill_layers(separator)
# Generate dataset
all_inputs = list(it.permutations(symbols[1:], 2))
split = int(.80 * len(all_inputs))
# example generation
def example(dataset):
# Randomly choose echo symbol (excluding 0 separator)
inputs = dataset[np.random.randint(len(dataset))]
targets = inputs[::-1]
return inputs, targets
def training_example():
return example(all_inputs[:split])
def testing_example():
return example(all_inputs[split:])
# all or nothing reward
def reward(ghu, targets, outputs):
r = np.zeros(len(outputs))
outputs = np.array([out for out in outputs[1:] if out != separator])
if len(outputs) == len(targets): r[-1] = (targets == outputs).all()
return r
# Run optimization
avg_rewards, avg_general, grad_norms = reinforce(
ghu,
num_epochs=100,
episode_duration=3,
training_example=training_example,
testing_example=testing_example,
reward=reward,
task="swap",
learning_rate=.1,
distribution_variance_coefficient=distribution_variance_coefficient,
verbose=1,
save_file=save_file)
def trials(i, avgrew, avggen, gradnorm):
print("***************************** Trial ", str(i + 1),
" *******************************")
# Configuration
num_symbols = 10
layer_sizes = {"rinp": 32, "rout": 32, "rtemp": 32}
hidden_size = 32
rho = .99
plastic = []
num_episodes = 500
# Setup GHU
symbols = [str(a) for a in range(num_symbols)]
pathways, associations = default_initializer( # all to all
layer_sizes.keys(), symbols)
codec = Codec(layer_sizes, symbols, rho=rho, ortho=True)
controller = Controller(layer_sizes, pathways, hidden_size, plastic)
ghu = GatedHebbianUnit(layer_sizes,
pathways,
controller,
codec,
plastic=plastic,
batch_size=num_episodes)
ghu.associate(associations)
# Initialize layers
separator = symbols[0]
ghu.fill_layers(separator)
# Generate dataset
input_length = 5
all_inputs = [
np.array(inputs)
for inputs in it.product(symbols[1:], repeat=input_length)
]
split = int(.80 * len(all_inputs))
# example generation
def example(dataset):
inputs = dataset[np.random.randint(len(dataset))]
targets = np.array([max(inputs)])
return inputs, targets
def training_example():
return example(all_inputs[:split])
def testing_example():
return example(all_inputs[split:])
# all or nothing reward
def reward(ghu, targets, outputs):
r = np.zeros(len(outputs))
r[-1] = (outputs[-1] == targets[0])
return r
# Optimization settings
avg_rewards, avg_general, grad_norms = reinforce(
ghu,
num_epochs=100,
episode_duration=input_length,
training_example=training_example,
testing_example=testing_example,
reward=reward,
task="max",
learning_rate=.1,
verbose=1)
gradnorm[i + 1] = grad_norms.tolist()
avgrew[i + 1] = avg_rewards.tolist()
avggen[i + 1] = avg_general.tolist()