data_sh = theano.shared(np.asarray(data, dtype=theano.config.floatX),
borrow=True)
data_t_sh = theano.shared(np.asarray(data_t, dtype=theano.config.floatX),
borrow=True)
return data_sh, T.cast(data_t_sh, 'int32')
train_data_sh, train_data_t_sh = make_theano_dataset(
(train_data, train_data_t))
valid_data_sh, valid_data_t_sh = make_theano_dataset(
(valid_data, valid_data_t))
test_data_sh, test_data_t_sh = make_theano_dataset((test_data, test_data_t))
datasets = [(train_data_sh, train_data_t_sh), (valid_data_sh, valid_data_t_sh),
(test_data_sh, test_data_t_sh)]
rbm_stack = RBMStack(num_dims, [500])
dbn = DBN(rbm_stack, 2)
batch_size = 100
max_epoch = 10
train_params = {
'batch_size': batch_size,
'learning_rate': 0.01,
'cd_steps': 2,
'max_epoch': max_epoch,
'persistent': True,
'finetune_learning_rate': 0.1
}
pre_fn = dbn.pretrain_fun(train_data_sh, train_params)
data_target_sh = theano.shared(np.asarray(data_target,
dtype=theano.config.floatX),
borrow=True)
batch_size = 70
num_batches = num_cases / batch_size
max_epoch = 50
train_params = {
'batch_size': batch_size,
'learning_rate': 0.01,
'cd_steps': 5,
'max_epoch': max_epoch,
'persistent': True,
'learning_rate_line': 0.0001,
'finetune_learning_rate': 0.1
}
stack_rbm = RBMStack(num_vis=num_vis, hid_layers_size=[500])
if not load_from_file(stack_rbm, train_params):
stack_rbm.pretrain(data_sh, train_params)
save_to_file(stack_rbm, train_params)
print_top_to_file(sent_ae, "pre_train", data_sh, data_target, range(0, 999))
fine_tune = sent_ae.finetune_fun(data_sh, train_params)
for ep in xrange(0, max_epoch):
for b in xrange(0, num_batches):
cost = fine_tune(b)
print "Epoch # %d:%d cost: %f" % (ep, b, cost)
print_top_to_file(sent_ae, "fine_tune", data_sh, data_target, range(0, 999))
start = time.time()
for k in watches:
name = "%i:%s" % (iter, k)
load_bin(name, watches[k])
db_redis.r0.set("last_it", iter)
iter+=1
end = time.time()
print "Inserted in %s" % (end - start,)
free_en_acc = []
free_en_valid_acc = []
cost_acc = []
cost_valid_acc = []
#rbm.need_train = True
rbms = RBMStack(rbms=[rbm])
def train(rbms, data_sh, data_valid_sh, train_params):
for watches in rbms.pretrain(data_sh, data_valid_sh, train_params):
free_en_acc.append(watches['free_en'])
free_en_valid_acc.append(watches['free_en_valid'])
cost_acc.append(watches['cost'])
cost_valid_acc.append(watches['cost_valid'])
watches['free_en_acc'] = np.asarray(free_en_acc, dtype=theano.config.floatX)
watches['free_en_valid_acc'] = np.asarray(free_en_valid_acc, dtype=theano.config.floatX)
watches['cost_acc'] = np.asarray(cost_acc, dtype=theano.config.floatX)
watches['cost_valid_acc'] = np.asarray(cost_valid_acc, dtype=theano.config.floatX)
del watches['free_en']
del watches['free_en_valid']
return data_sh, T.cast(data_t_sh, 'int32')
train_data_sh, train_data_t_sh = make_theano_dataset((train_data, train_data_t))
valid_data_sh, valid_data_t_sh = make_theano_dataset((valid_data, valid_data_t))
test_data_sh, test_data_t_sh = (None, None) #make_theano_dataset((test_data, test_data_t))
train_data_wp_sh, train_data_wp_t_sh = make_theano_dataset((data_wo_p, data_target_wo_p))
datasets = [(train_data_sh, train_data_t_sh),(valid_data_sh, valid_data_t_sh),(test_data_sh, test_data_t_sh)]
batch_size = 100
max_epoch = 30
train_params = {'batch_size' : batch_size, 'learning_rate' : 0.1, 'cd_steps' : 10, 'max_epoch' : max_epoch, 'persistent' : True, 'cd_steps_line' : 1, 'learning_rate_line' : 0.001, 'finetune_learning_rate' : 0.1, 'validation_frequency' : batch_size }
stack_rbm = RBMStack(num_vis = num_vis, hid_layers_size = hid_layers_size)
if not load_from_file(stack_rbm, train_params):
stack_rbm.pretrain(data_sh, train_params)
save_to_file(stack_rbm, train_params)
dbn = DBN(stack_rbm, 2)
#dbn.finetune(datasets, train_params)
#ae = AutoEncoder(stack_rbm, 100)
#ae.pretrain(train_data_sh, train_params)
#ae.finetune(train_data_sh, train_params)
get_output = theano.function([], dbn.stack[-1].output, givens=[(dbn.input,train_data_sh)])
out = get_output()
np.savetxt("/home/alexeyche/prog/sentiment/out.tsv", train_data, delimiter="\t")
continue
case_arr = np.concatenate(hist, axis=1)
l = case_arr.shape[0]
data_train[i,0:l] = case_arr
i+=1
del hist[:]
if i == num_cases:
break
else:
hist.append(case[:-1])
return data_train
data_train = preprocess()
data_train_sh = theano.shared(np.asarray(data_train, dtype=theano.config.floatX), borrow=True)
rbms = RBMStack(hid_layers_size = [15], bottomRBMtype = RBMMultSoftmax, add_opts = { 'num_units' : num_units, 'unit_size' : unit_size })
train_params = { 'max_epoch' : 50, 'batch_size' : 50, 'learning_rate' : 0.001, 'cd_steps' : 15, 'persistent' : True}
rbms.pretrain(data_train_sh, train_params)
rbm = rbms.stack[0]
rbm.save_model()
def evaluate(data):
def softmax(w):
e = np.exp(w)
dist = e / np.sum(e)
return dist
w = rbm.W.get_value(borrow=True)
vbias = rbm.vbias.get_value(borrow=True)
del hist[:]
if i == num_cases:
break
else:
hist.append(case[:-1])
return data_train
data_train = preprocess()
data_train_sh = theano.shared(np.asarray(data_train,
dtype=theano.config.floatX),
borrow=True)
rbms = RBMStack(hid_layers_size=[15],
bottomRBMtype=RBMMultSoftmax,
add_opts={
'num_units': num_units,
'unit_size': unit_size
})
train_params = {
'max_epoch': 50,
'batch_size': 50,
'learning_rate': 0.001,
'cd_steps': 15,
'persistent': True
}
rbms.pretrain(data_train_sh, train_params)
rbm = rbms.stack[0]
rbm.save_model()
