input_sequences = sequences[:, :-1]
target_sequences = sequences[:, 1:]
sample_level_outputs, new_h0 = sample_level_rnn(input_sequences, h0, reset)
cost = T.nnet.categorical_crossentropy(
T.nnet.softmax(sample_level_outputs.reshape((-1, Q_LEVELS))),
target_sequences.flatten()
).mean()
# By default we report cross-entropy cost in bits.
# Switch to nats by commenting out this line:
cost = cost * lib.floatX(1.44269504089)
params = lib.search(cost, lambda x: hasattr(x, 'param'))
lib._train.print_params_info(cost, params)
grads = T.grad(cost, wrt=params, disconnected_inputs='warn')
grads = [T.clip(g, lib.floatX(-GRAD_CLIP), lib.floatX(GRAD_CLIP)) for g in grads]
updates = lasagne.updates.adam(grads, params)
train_fn = theano.function(
[sequences, h0, reset],
[cost, new_h0],
updates=updates,
on_unused_input='warn'
)
generate_outputs, generate_new_h0 = sample_level_rnn(sequences, h0, reset)
from lib.pyramid import youtubedl
except Exception:
xbmc.executebuiltin("XBMC.Notification(ThePyramid,Please [COLOR yellow]install the Youtube Addon[/COLOR] module ,10000,"")")
stream_url=youtubedl.single_YD(url)
from lib.pyramid import pyramid;pyramid.playsetresolved(stream_url,name,iconimage)
elif mode==1119:from lib.pyramid import pyramid;pyramid.playsetresolved (pyramid.urlsolver(url),name,iconimage,True)
elif mode==1121:from lib.pyramid import pyramid;pyramid.ytdl_download('',name,'video')
elif mode==1123:from lib.pyramid import pyramid;pyramid.ytdl_download(url,name,'video')
elif mode==1124:from lib.pyramid import pyramid;pyramid.ytdl_download(url,name,'audio')
elif mode==1125:from lib.pyramid import pyramid;pyramid.search(url);xbmcplugin.endOfDirectory(int(sys.argv[1]))
elif mode==1126:
name = name.split(':')
from lib.pyramid import pyramid;pyramid.search(url,search_term=name[1])
xbmcplugin.endOfDirectory(int(sys.argv[1]))
elif mode == 1127:
from lib.pyramid import pyramid;pyramid.pulsarIMDB=search(url)
xbmc.Player().play(pulsarIMDB)
elif mode == 1130:from lib.pyramid import pyramid;pyramid.GetSublinks(name,url,iconimage,fanart)
elif mode == 1140:from lib.pyramid import pyramid;pyramid.SearchChannels();pyramid.SetViewThumbnail();xbmcplugin.endOfDirectory(int(sys.argv[1]))
elif mode == 1141 : from lib.pyramid import pyramid;pyramid.Search_input(url);xbmcplugin.endOfDirectory(int(sys.argv[1]))
elif mode == 1142: from lib.pyramid import pyramid;pyramid.RESOLVE(url)
elif mode == 1153:from lib.pyramid import pyramid;pyramid.pluginquerybyJSON(url);xbmcplugin.endOfDirectory(int(sys.argv[1]))
elif mode == 1200: from lib.freeview import freeview;freeview.CATEGORIES()
elif mode == 1201: from lib.freeview import freeview;freeview.play(url)
elif mode == 1202: from lib.freeview import freeview;freeview.tvplayer(url)
elif mode == 1300: from lib import apprentice;apprentice.apprentice_Main()
elif mode == 1301 : from lib import apprentice;apprentice.Mov_Menu()
elif mode == 1302 : from lib import apprentice;apprentice.Tv_Menu()
elif mode == 1303 : from lib import apprentice;apprentice.Second_Menu(url)
elif mode == 1304 : from lib import apprentice;apprentice.Index_List_Mov()
elif mode == 1305 : from lib import apprentice;apprentice.Main_Loop(url)
test_mask
)
valid_costs.append(valid_loss)
print "Validation Completed! cost: {} time: {}".format(np.mean(np.asarray(valid_costs),axis=0),time.time()-start)
return np.mean(np.asarray(valid_costs),axis=0)[1]
if __name__=='__main__':
batch_size = T.shape(X)[0]
# readout = DurationPredictor(X,drop_prob)
# readout = RecurrentPredictor(X,drop_prob,mask=mask)
readout = DeepVoice(X,drop_prob)
batch_size = T.shape(readout)[0]
cost = T.sum(T.sqr(readout-Y)*mask)/T.sum(mask)
abs_cost = T.sum(T.abs_(readout-Y)*mask)/T.sum(mask)
params = lib.search(cost, lambda x: hasattr(x, "param"))
lib.print_params_info(params)
grads = T.grad(cost, wrt=params, disconnected_inputs='warn')
grads = [T.clip(g, lib.floatX(-GRAD_CLIP), lib.floatX(GRAD_CLIP)) for g in grads]
print "Gradients Computed"
updates = lasagne.updates.adam(grads, params, learning_rate=learn_rate)
train_fn = theano.function(
[drop_prob,X,Y,mask,learn_rate],
[cost,abs_cost],
updates=updates,
on_unused_input='warn'
)
test_fn = theano.function(
mode='open-loop-rnn'
)
readout = lib.ops.Linear(
'Generator.GRU.Output.MLP.1',
T.concatenate([state[:,:,-1],tiled_speaker],-1),
DEC_DIM+SPEAKER_DIM,
OUTPUT_DIM
)
mask_mult = T.shape_padright(mask)
cost = T.sum(T.sqr(X-readout)*mask_mult)/(T.sum(mask)*63.)
test_cost = T.sum(T.sqr(X-predict_readout)*T.shape_padright(mask))/(T.sum(mask)*63.)
params = lib.search(cost, lambda x: hasattr(x, "param") and x.param==True)
lib.print_params_info(params)
grads = T.grad(cost, wrt=params, disconnected_inputs='warn')
grads = [T.clip(g, lib.floatX(-GRAD_CLIP), lib.floatX(GRAD_CLIP)) for g in grads]
print "Gradients Computed"
updates = lasagne.updates.adam(grads, params, learning_rate=learn_rate)
train_fn = theano.function(
[noise_vocoder,X,spkr_ids,ctx,mask,learn_rate],
cost,
updates=updates,
on_unused_input='warn'
)
pyramid.ytdl_download(url, name, 'video')
elif mode == 1124:
from lib.pyramid import pyramid
pyramid.ytdl_download(url, name, 'audio')
elif mode == 1125:
from lib.pyramid import pyramid
pyramid.search(url)
xbmcplugin.endOfDirectory(int(sys.argv[1]))
elif mode == 1126:
name = name.split(':')
from lib.pyramid import pyramid
pyramid.search(url, search_term=name[1])
xbmcplugin.endOfDirectory(int(sys.argv[1]))
elif mode == 1127:
from lib.pyramid import pyramid
pyramid.pulsarIMDB = search(url)
xbmc.Player().play(pulsarIMDB)
elif mode == 1130:
from lib.pyramid import pyramid
pyramid.GetSublinks(name, url, iconimage, fanart)
elif mode == 1140:
from lib.pyramid import pyramid
pyramid.SearchChannels()
pyramid.SetViewThumbnail()
xbmcplugin.endOfDirectory(int(sys.argv[1]))
elif mode == 1141:
from lib.pyramid import pyramid
pyramid.Search_input(url)
xbmcplugin.endOfDirectory(int(sys.argv[1]))
elif mode == 1142:
from lib.pyramid import pyramid
# f = search(gp*100+user, doc, rank, th, 100*(i+1), beta=0.2)
# hit = 0
# for u in f:
# if u[2] == gp: hit += 1
# hr[i].append(hit)
# th += 0.05
# w.writerows(hr)
# print "insert result successfully..."
# test the metric.
w = csv.writer(open('metric_sc.csv', 'wb'), delimiter=',')
for gp in range(10):
for user in range(100): # every user.
hr = [0]*10
for p in range(10):
f = search(gp*100+user, doc, rank, 0, 100*(p+1), beta=0.2)
hit = 0
for u in f:
if u[2] == gp:
hit += 1
hr[p] = hit
w.writerow(hr)
#hr = [0]*10
#for cl in range(10):
# for i in range(10):
# f = search(cl*100+23, doc, rank, 0, 100*(i+1))
# hit = 0
# for u in f:
# if u[2] == cl:
# hit += 1
def search_handler(response):
query = response.get_field('query')
data = {'isloggedin': isloggedin(response), 'results': search(query), 'query':query, 'username': currentuser(response).username}
response.write(template.render_file("templates/search_result.html", data))
if __name__ == '__main__':
rec = Recognizer(X, 0.)
h = rec['CNN.1']
# readout = RecurrentMapper(ctx)
readout = ConvolutionalMapper(ctx)
# predict_readout = create_graph(X_concat,mode='test')
predict_readout = readout
mask_mult = T.shape_padright(mask)
cost = T.sum(T.sqr(readout - h) * mask_mult) / (T.sum(mask_mult))
test_cost = T.sum(
T.sqr(predict_readout - h) * mask_mult) / (T.sum(mask_mult))
params = lib.search(
cost, lambda x: hasattr(x, "param") and x.param == True and
'Mapper.Generator' in x.name)
lib.print_params_info(params)
grads = T.grad(cost, wrt=params, disconnected_inputs='warn')
print "Gradients Computed"
updates = lasagne.updates.adam(grads, params, learning_rate=learn_rate)
# for x,y in lib._updates.iteritems():
# if 'Generator' in x.name:
# updates[x] = y
# print "Adding update: ",x.name
train_fn = theano.function([X, ctx, mask, learn_rate],
cost,
updates=updates,
def train_loop(inputs,
cost,
train_data,
times,
prints=None,
inject_total_iters=False,
test_data=None,
callback=None,
optimizer=lasagne.updates.adam,
save_params=False,
nan_guard=False):
params = lib.search(cost, lambda x: hasattr(x, 'param'))
lib.print_params_info(params)
grads = T.grad(cost, wrt=params, disconnected_inputs='warn')
grads = [T.clip(g, lib.floatX(-1), lib.floatX(1)) for g in grads]
updates = optimizer(grads, params)
if prints is None:
prints = [('cost', cost)]
else:
prints = [('cost', cost)] + prints
print "Compiling train function..."
if nan_guard:
from theano.compile.nanguardmode import NanGuardMode
mode = NanGuardMode(nan_is_error=True,
inf_is_error=True,
big_is_error=True)
else:
mode = None
train_fn = theano.function(inputs, [p[1] for p in prints],
updates=updates,
on_unused_input='warn',
mode=mode)
print "Compiling eval function..."
eval_fn = theano.function(inputs, [p[1] for p in prints],
on_unused_input='warn')
print "Training!"
total_iters = 0
total_seconds = 0.
last_print = 0
last_gen = 0
if len(times) >= 4:
gen_every = times[3]
else:
gen_every = times[1]
if len(times) >= 5:
early_stop = times[4]
if len(times) >= 6:
early_stop_min = times[5]
else:
early_stop_min = 0
else:
early_stop = None
early_stop_min = None
best_test_cost = np.inf
best_test_cost_iter = 0.
all_outputs = []
all_stats = []
for epoch in itertools.count():
generator = train_data()
while True:
try:
inputs = generator.__next__()
except StopIteration:
break
if inject_total_iters:
inputs = [np.int32(total_iters)] + list(inputs)
start_time = time.time()
outputs = train_fn(*inputs)
total_seconds += time.time() - start_time
total_iters += 1
all_outputs.append(outputs)
if total_iters == 1:
try: # This only matters on Ishaan's computer
import experiment_tools
experiment_tools.register_crash_notifier()
except ImportError:
pass
if (times[0]=='iters' and total_iters-last_print == times[1]) or \
(times[0]=='seconds' and total_seconds-last_print >= times[1]):
mean_outputs = np.array(all_outputs).mean(axis=0)
if test_data is not None:
if inject_total_iters:
test_outputs = [
eval_fn(np.int32(total_iters), *inputs)
for inputs in test_data()
]
else:
test_outputs = [
eval_fn(*inputs) for inputs in test_data()
]
test_mean_outputs = np.array(test_outputs).mean(axis=0)
stats = collections.OrderedDict()
stats['epoch'] = epoch
stats['iters'] = total_iters
for i, p in enumerate(prints):
stats['train ' + p[0]] = mean_outputs[i]
if test_data is not None:
for i, p in enumerate(prints):
stats['test ' + p[0]] = test_mean_outputs[i]
stats['secs'] = total_seconds
stats['secs/iter'] = total_seconds / total_iters
if test_data != None and (stats['test cost'] < best_test_cost
or
(early_stop_min != None
and total_iters <= early_stop_min)):
best_test_cost = stats['test cost']
best_test_cost_iter = total_iters
print_str = ""
for k, v in stats.items():
if isinstance(v, int):
print_str += "{}:{}\t".format(k, v)
else:
print_str += "{}:{:.4f}\t".format(k, v)
print print_str[:-1] # omit the last \t
all_stats.append(stats)
all_outputs = []
last_print += times[1]
if (times[0]=='iters' and total_iters-last_gen==gen_every) or \
(times[0]=='seconds' and total_seconds-last_gen >= gen_every):
tag = "iters{}_time{}".format(total_iters, total_seconds)
if callback is not None:
callback(tag)
if save_params:
lib.save_params('params_{}.pkl'.format(tag))
last_gen += gen_every
if (times[0]=='iters' and total_iters == times[2]) or \
(times[0]=='seconds' and total_seconds >= times[2]) or \
(test_data != None and early_stop != None and total_iters > (3*early_stop) and (total_iters-best_test_cost_iter) > early_stop):
if (test_data != None and early_stop != None and total_iters >
(3 * early_stop)
and (total_iters - best_test_cost_iter) > early_stop):
print "Early stop! Best test cost was {} at iter {}".format(
best_test_cost, best_test_cost_iter)
print "Done!"
try: # This only matters on Ishaan's computer
import experiment_tools
experiment_tools.send_sms("done!")
except ImportError:
pass
return all_stats
sample_level_outputs = sample_level_predictor(
frame_level_outputs.reshape((BATCH_SIZE * SEQ_LEN, DIM)), prev_samples)
cost = T.nnet.categorical_crossentropy(T.nnet.softmax(sample_level_outputs),
target_sequences.flatten()).mean()
# By default we report cross-entropy cost in bits.
# Switch to nats by commenting out this line:
cost = cost * lib.floatX(1.44269504089)
ip_cost = lib.floatX(1.44269504089) * T.nnet.categorical_crossentropy(
T.nnet.softmax(big_frame_independent_preds.reshape(
(-1, Q_LEVELS))), target_sequences.flatten()).mean()
all_params = lib.search(cost, lambda x: hasattr(x, 'param'))
ip_params = lib.search(
ip_cost, lambda x: hasattr(x, 'param') and 'BigFrameLevel' in x.name)
other_params = [p for p in all_params if p not in ip_params]
all_params = ip_params + other_params
lib._train.print_params_info(ip_cost, ip_params)
lib._train.print_params_info(cost, other_params)
lib._train.print_params_info(cost, all_params)
ip_grads = T.grad(ip_cost, wrt=ip_params, disconnected_inputs='warn')
ip_grads = [
T.clip(g, lib.floatX(-GRAD_CLIP), lib.floatX(GRAD_CLIP)) for g in ip_grads
]
other_grads = T.grad(cost, wrt=other_params, disconnected_inputs='warn')
other_grads = [
def create_encoder_decoder():
input_var = T.tensor3('input')
input_var_normalised = (input_var - floatX(0.5))
mu, log_square_sigma = Encoder(input_var_normalised)
mu = lib.floatX(2.) * T.tanh(mu / lib.floatX(2.))
sampled_z = gaussian_sampler(mu, log_square_sigma)
reconstructed = Decoder(sampled_z)
reconstruction_cost = T.nnet.binary_crossentropy(
reconstructed.reshape((reconstructed.shape[0], -1)),
input_var.reshape((input_var.shape[0], -1))).sum(axis=1)
kl_cost = KL_with_standard_gaussian(mu, log_square_sigma)
loss = T.mean(kl_cost + reconstruction_cost)
params = lib.search(loss,
lambda x: hasattr(x, 'param') and x.param == True)
lib.print_params_info(params)
grads = T.grad(loss, wrt=params, disconnected_inputs='warn')
grads = [T.clip(g, lib.floatX(-1.), lib.floatX(1.)) for g in grads]
lr = T.scalar('lr')
updates = lasagne.updates.adam(grads,
params,
learning_rate=lr,
epsilon=EPS)
generated_z = T.matrix('generated_z')
generated_samples = Decoder(generated_z)
print "Compiling functions ..."
train_fn = theano.function(
[input_var, lr],
[
loss,
kl_cost.mean(),
mu.min(),
mu.max(), mu,
sampled_z.min(),
sampled_z.max()
],
updates=updates,
# mode=NanGuardMode(nan_is_error=True, inf_is_error=True, big_is_error=True)
)
reconstruct_fn = theano.function([input_var], reconstructed)
val_fn = theano.function(
[input_var],
[
loss,
kl_cost.mean(),
mu.min(),
mu.max(), mu,
sampled_z.min(),
sampled_z.max()
],
# mode=NanGuardMode(nan_is_error=True, inf_is_error=True, big_is_error=True)
)
generate_fn = theano.function([generated_z], generated_samples)
encode_fn = theano.function([input_var], mu)
return train_fn, val_fn, generate_fn, reconstruct_fn, encode_fn
def train_loop(
inputs,
cost,
train_data,
times,
prints=None,
inject_total_iters=False,
test_data=None,
callback=None,
optimizer=lasagne.updates.adam,
save_params=False,
nan_guard=False
):
params = lib.search(cost, lambda x: hasattr(x, 'param'))
lib.print_params_info(params)
grads = T.grad(cost, wrt=params, disconnected_inputs='warn')
grads = [T.clip(g, lib.floatX(-1), lib.floatX(1)) for g in grads]
updates = optimizer(grads, params)
if prints is None:
prints = [('cost', cost)]
else:
prints = [('cost', cost)] + prints
print "Compiling train function..."
if nan_guard:
from theano.compile.nanguardmode import NanGuardMode
mode = NanGuardMode(
nan_is_error=True,
inf_is_error=True,
big_is_error=True
)
else:
mode = None
train_fn = theano.function(
inputs,
[p[1] for p in prints],
updates=updates,
on_unused_input='warn',
mode=mode
)
print "Compiling eval function..."
eval_fn = theano.function(
inputs,
[p[1] for p in prints],
on_unused_input='warn'
)
print "Training!"
total_iters = 0
total_seconds = 0.
last_print = 0
all_outputs = []
all_stats = []
for epoch in itertools.count():
for inputs in train_data():
if inject_total_iters:
inputs = [np.int32(total_iters)] + list(inputs)
start_time = time.time()
outputs = train_fn(*inputs)
total_seconds += time.time() - start_time
total_iters += 1
all_outputs.append(outputs)
if total_iters == 1:
try: # This only matters on Ishaan's computer
import experiment_tools
experiment_tools.register_crash_notifier()
except ImportError:
pass
if (times[0]=='iters' and total_iters-last_print == times[1]) or \
(times[0]=='seconds' and total_seconds-last_print >= times[1]):
mean_outputs = np.array(all_outputs).mean(axis=0)
if test_data is not None:
if inject_total_iters:
test_outputs = [
eval_fn(np.int32(total_iters), *inputs)
for inputs in test_data()
]
else:
test_outputs = [
eval_fn(*inputs)
for inputs in test_data()
]
test_mean_outputs = np.array(test_outputs).mean(axis=0)
stats = collections.OrderedDict()
stats['epoch'] = epoch
stats['iters'] = total_iters
for i,p in enumerate(prints):
stats['train '+p[0]] = mean_outputs[i]
if test_data is not None:
for i,p in enumerate(prints):
stats['test '+p[0]] = test_mean_outputs[i]
stats['secs'] = total_seconds
stats['secs/iter'] = total_seconds / total_iters
print_str = ""
for k,v in stats.items():
if isinstance(v, int):
print_str += "{}:{}\t".format(k,v)
else:
print_str += "{}:{:.4f}\t".format(k,v)
print print_str[:-1] # omit the last \t
all_stats.append(stats)
tag = "iters{}_time{}".format(total_iters, total_seconds)
if callback is not None:
callback(tag)
if save_params:
lib.save_params('params_{}.pkl'.format(tag))
all_outputs = []
last_print += times[1]
if (times[0]=='iters' and total_iters == times[2]) or \
(times[0]=='seconds' and total_seconds >= times[2]):
print "Done!"
try: # This only matters on Ishaan's computer
import experiment_tools
experiment_tools.send_sms("done!")
except ImportError:
pass
return all_stats
cost = T.nnet.categorical_crossentropy(
T.nnet.softmax(sample_level_outputs),
target_sequences.flatten()
).mean()
# By default we report cross-entropy cost in bits.
# Switch to nats by commenting out this line:
cost = cost * lib.floatX(1.44269504089)
ip_cost = lib.floatX(1.44269504089) * T.nnet.categorical_crossentropy(
T.nnet.softmax(big_frame_independent_preds.reshape((-1, Q_LEVELS))),
target_sequences.flatten()
).mean()
all_params = lib.search(cost, lambda x: hasattr(x, 'param'))
ip_params = lib.search(ip_cost, lambda x: hasattr(x, 'param') and 'BigFrameLevel' in x.name)
other_params = [p for p in all_params if p not in ip_params]
all_params = ip_params + other_params
lib._train.print_params_info(ip_cost, ip_params)
lib._train.print_params_info(cost, other_params)
lib._train.print_params_info(cost, all_params)
ip_grads = T.grad(ip_cost, wrt=ip_params, disconnected_inputs='warn')
ip_grads = [T.clip(g, lib.floatX(-GRAD_CLIP), lib.floatX(GRAD_CLIP)) for g in ip_grads]
other_grads = T.grad(cost, wrt=other_params, disconnected_inputs='warn')
other_grads = [T.clip(g, lib.floatX(-GRAD_CLIP), lib.floatX(GRAD_CLIP)) for g in other_grads]
grads = T.grad(cost, wrt=all_params, disconnected_inputs='warn')
grads = [T.clip(g, lib.floatX(-GRAD_CLIP), lib.floatX(GRAD_CLIP)) for g in grads]
disc_out = Discriminator(T.concatenate([real_data, fake_data], axis=0))
disc_real = disc_out[:BATCH_SIZE]
disc_fake = disc_out[BATCH_SIZE:]
gen_cost = -T.mean(Discriminator(Generator(2 * BATCH_SIZE)))
disc_cost = T.mean(disc_fake) - T.mean(disc_real)
alpha = srng.uniform(size=(BATCH_SIZE, 1), low=0., high=1.)
differences = fake_data - real_data
interpolates = real_data + (alpha * differences)
gradients = T.grad(T.sum(Discriminator(interpolates)), interpolates)
slopes = T.sqrt(T.sum(T.sqr(gradients), axis=1))
lipschitz_penalty = T.mean((slopes - 1.)**2)
disc_cost += 10 * lipschitz_penalty
gen_params = lib.search(
gen_cost, lambda x: hasattr(x, 'param') and 'Generator' in x.name)
discrim_params = lib.search(
disc_cost, lambda x: hasattr(x, 'param') and 'Discriminator' in x.name)
gen_grads = T.grad(gen_cost, gen_params)
discrim_grads = T.grad(disc_cost, discrim_params)
# Just so logging code doesn't break
gen_grad_norm, discrim_grad_norm = T.as_tensor_variable(
np.float32(0)), T.as_tensor_variable(np.float32(0))
gen_grads, gen_grad_norm = lasagne.updates.total_norm_constraint(
gen_grads, 50.0, return_norm=True)
discrim_grads, discrim_grad_norm = lasagne.updates.total_norm_constraint(
discrim_grads, 50.0, return_norm=True)
gen_cost = -T.mean(Discriminator(fake_data_4x))
disc_cost = T.mean(disc_fake) - T.mean(disc_real)
alpha = srng.uniform(
size=(BATCH_SIZE,1),
low=0.,
high=1.
)
differences = fake_data - real_data
interpolates = real_data + (alpha*differences)
gradients = T.grad(T.sum(Discriminator(interpolates)), interpolates)
slopes = T.sqrt(T.sum(T.sqr(gradients), axis=1))
lipschitz_penalty = T.mean((slopes-1.)**2)
disc_cost += 10*lipschitz_penalty
gen_params = lib.search(gen_cost, lambda x: hasattr(x, 'param') and 'Generator' in x.name)
discrim_params = lib.search(disc_cost, lambda x: hasattr(x, 'param') and 'Discriminator' in x.name)
gen_grads = T.grad(gen_cost, gen_params)
discrim_grads = T.grad(disc_cost, discrim_params)
gen_grads = [
T.clip(g, lib.floatX(-1.0), lib.floatX(1.0))
for g in gen_grads
]
discrim_grads = [
T.clip(g, lib.floatX(-1.0), lib.floatX(1.0))
for g in discrim_grads
]
gen_updates = lasagne.updates.adam(gen_grads, gen_params, learning_rate=1e-4, beta1=0.5, beta2=0.9)
discrim_updates = lasagne.updates.adam(discrim_grads, discrim_params, learning_rate=1e-4, beta1=0.5, beta2=0.9)
#prev_samples = sequences[:, :-1]
#prev_samples = prev_samples.reshape((1, BATCH_SIZE, 1, -1))
#prev_samples = T.nnet.neighbours.images2neibs(prev_samples, (1, FRAME_SIZE), neib_step=(1, 1), mode='valid')
#prev_samples = prev_samples.reshape((BATCH_SIZE,SEQ_LEN, FRAME_SIZE))
encoder_outputs, new_h0 = encoder(input_sequences, h0, reset)
#decoder_outputs = decoder(encoder_outputs,prev_samples)
cost = T.nnet.categorical_crossentropy(T.nnet.softmax(encoder_outputs),
target_sequences.flatten()).mean()
cost = cost * lib.floatX(1.44269504089)
params = lib.search(cost, lambda x: hasattr(x, 'param'))
lib.print_params_info(cost, params)
grads = T.grad(cost, wrt=params, disconnected_inputs='warn')
grads = [
T.clip(g, lib.floatX(-GRAD_CLIP), lib.floatX(GRAD_CLIP)) for g in grads
]
print "Gradients Computed"
updates = lasagne.updates.adam(grads, params, learning_rate=lr)
train_fn = theano.function([sequences, h0, reset, lr], [cost, new_h0],
updates=updates,
pyramid.ytdl_download(url, name, "audio")
elif mode == 1125:
from lib.pyramid import pyramid
pyramid.search(url)
xbmcplugin.endOfDirectory(int(sys.argv[1]))
elif mode == 1126:
name = name.split(":")
from lib.pyramid import pyramid
pyramid.search(url, search_term=name[1])
xbmcplugin.endOfDirectory(int(sys.argv[1]))
elif mode == 1127:
from lib.pyramid import pyramid
pyramid.pulsarIMDB = search(url)
xbmc.Player().play(pulsarIMDB)
elif mode == 1130:
from lib.pyramid import pyramid
pyramid.GetSublinks(name, url, iconimage, fanart)
elif mode == 1140:
from lib.pyramid import pyramid
pyramid.SearchChannels()
pyramid.SetViewThumbnail()
xbmcplugin.endOfDirectory(int(sys.argv[1]))
elif mode == 1141:
from lib.pyramid import pyramid
pyramid.Search_input(url)
