def solve_contractive_sign(known_T, weight, bias, LAYER):
print("Solve the extraction problem for contractive networks")
def get_preimage(hidden):
preimage = hidden
for i, (my_A, my_B) in reversed(
list(enumerate(zip(known_T.A + [weight],
known_T.B + [bias])))):
if i == 0:
res = scipy.optimize.lsq_linear(my_A.T,
preimage - my_B,
bounds=(-np.inf, np.inf))
else:
res = scipy.optimize.lsq_linear(my_A.T,
preimage - my_B,
bounds=(0, np.inf))
preimage = res.x
return preimage[np.newaxis, :]
hidden = np.zeros((sizes[LAYER + 1]))
preimage = get_preimage(hidden)
extended_T = known_T.extend_by(weight, bias)
standard_out = run(preimage)
signs = []
for axis in range(len(hidden)):
h = np.array(hidden)
h[axis] = 10
preimage_plus = get_preimage(h)
h[axis] = -10
preimage_minus = get_preimage(h)
print("Confirm preimage")
if np.any(extended_T.forward(preimage) > 1e-5):
raise AcceptableFailure()
out_plus = run(preimage_plus)
out_minus = run(preimage_minus)
print(standard_out, out_plus, out_minus)
inverted_if_small = np.sum(np.abs(out_plus - standard_out))
not_inverted_if_small = np.sum(np.abs(out_minus - standard_out))
print("One of these should be small", inverted_if_small,
not_inverted_if_small)
if inverted_if_small < not_inverted_if_small:
signs.append(-1)
else:
signs.append(1)
return signs
def run_spiders():
bots = (
(AvitoSpider, {
'thread_number': 2
}),
(IRRSpider, {
'thread_number': 2
}),
)
run(bots, debug=True)
def get_grad(x, direction, eps=1e-6):
"""
Finite differences to estimate the gradient.
Uses just two coordinates---that's sufficient for most of the code.
Can fail if we're right at a critical point and we get the left and right side.
/
X
/
-X--/
"""
x = x[np.newaxis,:]
a = run(x-eps*direction)
b = run(x)
g1 = (b-a)/eps
return g1
def solve_final_layer(known_T, inputs, outputs):
if CHEATING:
for i, (normal, bias) in enumerate(zip(known_T.A, known_T.B)):
logger.log('', level=Logger.INFO)
logger.log("LAYER", i, level=Logger.INFO)
check_quality(i, normal, bias)
outputs = run(inputs)
hidden = known_T.forward(inputs, with_relu=True)
hidden = np.concatenate([hidden, np.ones((hidden.shape[0], 1))], axis=1)
solution = np.linalg.lstsq(hidden, outputs)
vector = solution[0]
At = known_T.A + [vector[:-1]]
Bt = known_T.B + [vector[-1]]
logger.log("SAVING",
"./models/extracted-%s.p" % "-".join(map(str, sizes)),
level=Logger.INFO)
pickle.dump([At, Bt],
open("./models/extracted-%s.p" % "-".join(map(str, sizes)),
"wb"))
from src.global_vars import __cheat_A, __cheat_B
pickle.dump([__cheat_A, __cheat_B],
open("./models/real-%s.p" % "-".join(map(str, sizes)), "wb"))
def loss(x):
return (run(x, inner_A=At, inner_B=Bt) -
run(x, inner_A=__cheat_A, inner_B=__cheat_B))
ls = []
for _ in range(1):
print(_)
inp = np.random.normal(0, 1, (100000, A[0].shape[0]))
ls.extend(loss(inp).flatten())
logger.log("\n\n", level=Logger.INFO)
logger.log("Finally we are done.\n", level=Logger.INFO)
max_loss = np.max(np.abs(ls))
res = open("results.txt", "a")
res.write(str(max_loss))
res.close()
logger.log('Maximum logit loss on the unit sphere',
max_loss,
level=Logger.INFO)
logger.log("\nfin", level=Logger.INFO)
def get_second_grad_unsigned(x, direction, eps, eps2):
"""
Same as the above but batched so it's more efficient.
"""
x = np.array([x + direction * (eps - eps2),
x + direction * (eps),
x - direction * (eps - eps2),
x - direction * (eps)])
out = run(x)
return np.dot(out.flatten(), MASK)/eps
def loss(x):
return (run(x, inner_A=At, inner_B=Bt) -
run(x, inner_A=__cheat_A, inner_B=__cheat_B))
make_data_and_env(args)
agent, evaluator = make_follower_models(args, len(vocab), all_val_data, env)
if args.reward and not args.no_speaker:
speaker = make_speaker_models(args, len(vocab), env, tok)
speaker.load(args.speaker_prefix, **{})
print("Load speaker model %s" % args.speaker_prefix)
else:
speaker = None
if args.follower_prefix is not None:
agent.load(args.follower_prefix, args.load_opt, **{})
print("Load follower model %s" % args.follower_prefix)
return agent, train_data, val_data, evaluator, speaker, train_tag
def train_val(args):
''' Train on the training set, and validate on validation (seen/unseen) set. '''
agent, train_data, val_data, evaluator, speaker, train_tag = \
train_setup(args)
train(args, agent, train_data, val_data, evaluator, speaker, train_tag)
def make_arg_parser():
parser = argparse.ArgumentParser()
ImageFeatures.add_args(parser)
add_general_args(parser)
return parser
if __name__ == "__main__":
run(make_arg_parser(), train_val)
def memo_forward_pass(x):
if x not in c:
c[x] = run((offset + direction * x)[np.newaxis, :])
return c[x]
