def Predict(max_step, prefix):
edf.components = []
T = max_step
h = edf.Value(np.zeros((1, hidden_dim)))
c = edf.Value(np.zeros((1, hidden_dim)))
prediction = []
for t in range(T):
if t < len(prefix):
pred = edf.Value(prefix[t])
prediction.append(pred)
else:
prediction.append(pred)
wordvec = edf.Embed(pred, C2V)
xt = edf.Reshape(wordvec, [-1, hidden_dim])
h_next, c_next = LSTMCell(xt, h, c)
p = edf.SoftMax(edf.VDot(h_next, V))
pred = edf.ArgMax(p)
h = h_next
c = c_next
edf.Forward()
idx = [pred.value for pred in prediction]
stop_idx = utils.to_index('}')
if stop_idx in idx:
return idx[0:idx.index(stop_idx) + 1]
else:
return idx
def BuildModel():
edf.components = []
B = inp.value.shape[0]
T = inp.value.shape[1]
h = edf.Value(np.zeros((B, hidden_dim)))
c = edf.Value(np.zeros((B, hidden_dim)))
score = []
for t in range(T - 1):
wordvec = edf.Embed(edf.Value(inp.value[:, t]), C2V)
xt = edf.Reshape(wordvec, [-1, hidden_dim])
h_next, c_next = LSTMCell(xt, h, c)
p = edf.SoftMax(edf.VDot(h_next, V))
logloss = edf.Reshape(
edf.LogLoss(edf.Aref(p, edf.Value(inp.value[:, t + 1]))),
(B, 1))
if t == 0:
loss = logloss
else:
loss = edf.ConCat(loss, logloss)
score.append(p)
h = h_next
c = c_next
masks = np.zeros((B, T - 1), dtype=np.int32)
masks[inp.value[:, 1:] != 0] = 1
loss = edf.MeanwithMask(loss, edf.Value(masks))
return loss, score
def Predict(max_step, prefix):
edf.components = []
T = max_step
h = [[None] * layer] * (T + 1)
c = [[None] * layer] * (T + 1)
for i in range(layer):
h[0][i] = edf.Value(np.zeros((1, hidden_dim)))
c[0][i] = edf.Value(np.zeros((1, hidden_dim)))
prediction = []
for t in range(T):
if t < len(prefix):
pred = edf.Value(prefix[t])
prediction.append(pred)
else:
prediction.append(pred)
wordvec = edf.Embed(pred, C2V)
xt = edf.Reshape(wordvec, [-1, hidden_dim])
for i in range(layer):
h[t + 1][i], c[t + 1][i] = LSTMCell(xt, h[t][i], c[t][i], i)
xt = h[t + 1][i]
p = edf.SoftMax(edf.VDot(xt, V))
pred = edf.ArgMax(p)
edf.Forward()
idx = [pred.value for pred in prediction]
stop_idx = utils.to_index('}')
if stop_idx in idx:
return idx[0:idx.index(stop_idx) + 1]
else:
return idx
def BuildModel():
edf.components = []
B = inp.value.shape[0]
T = inp.value.shape[1]
h = [[None] * layer] * T
c = [[None] * layer] * T
for i in range(layer):
h[0][i] = edf.Value(np.zeros((B, hidden_dim)))
c[0][i] = edf.Value(np.zeros((B, hidden_dim)))
score = []
for t in range(T - 1):
wordvec = edf.Embed(edf.Value(inp.value[:, t]), C2V)
xt = edf.Reshape(wordvec, [-1, hidden_dim])
for i in range(layer):
h[t + 1][i], c[t + 1][i] = LSTMCell(xt, h[t][i], c[t][i], i)
xt = h[t + 1][i]
p = edf.SoftMax(edf.VDot(xt, V))
logloss = edf.Reshape(edf.LogLoss(edf.Aref(p, edf.Value(inp.value[:, t + 1]))), (B, 1))
if t == 0:
loss = logloss
else:
loss = edf.ConCat(loss, logloss)
score.append(p)
masks = np.zeros((B, T - 1), dtype=np.int32)
masks[inp.value[:, 1:] != 0] = 1
loss = edf.MeanwithMask(loss, edf.Value(masks))
return loss, score
def BuildModel():
edf.components = []
B, T = inp.value.shape
score = []
loss = None
# Init h_0 with one-hot
vocab_init = np.ones([B])
vocab_init = edf.Value(vocab_init)
h = edf.Embed(vocab_init, C2V)
# Init C_0 to be zero
c = edf.Value(np.zeros([B, hidden_dim]))
for t in range(T):
x_t = edf.Value(inp.value[:, t])
x_t = edf.Embed(x_t, C2V)
h, c = LSTMCell(x_t, h, c)
# Score and loss
pred = edf.SoftMax(edf.VDot(h, V))
if t != T - 1:
score.append(pred)
x_t1 = edf.Value(inp.value[:, t + 1])
else:
x_t1 = edf.Value(np.zeros(B))
loss_t = edf.LogLoss(edf.Aref(pred, x_t1))
if loss is None:
loss = loss_t
else:
loss = edf.Add(loss, loss_t)
loss = edf.Mul(edf.Mean(loss), edf.Value(np.float64(1) / T))
return loss, score
# Load data
data = np.load(fn('inputs/mnist_26k.npz'))
train_im = np.float32(data['im_train'])/255.-0.5
train_im = np.reshape(train_im,[-1,28,28,1])
train_lb = data['lbl_train']
val_im = np.float32(data['im_val'])/255.-0.5
val_im = np.reshape(val_im,[-1,28,28,1])
val_lb = data['lbl_val']
#######################################
# Inputs and parameters
inp = edf.Value()
lab = edf.Value()
K1 = edf.Param()
B1 = edf.Param()
K2 = edf.Param()
B2 = edf.Param()
W3 = edf.Param()
B3 = edf.Param()
# Model
#y = edf.conv2(inp,K1)
#y = edf.down2(y);
import pickle
import os
train_data, trcnt = utils.load_data_onechar('data/ptb.train.txt')
valid_data, vacnt = utils.load_data_onechar('data/ptb.valid.txt')
test_data, tecnt = utils.load_data_onechar('data/ptb.test.txt')
hidden_dim = 200
n_vocab = utils.n_vocab
batch = 50
parameters = []
model = 'model_LSTM.pkl'
eta = 0.5
decay = 0.9
inp = edf.Value()
np.random.seed(0)
edf.params = []
# LSTM parameters
# input embedding
C2V = edf.Param(edf.xavier((n_vocab, hidden_dim)))
# forget gate
Wf = edf.Param(edf.xavier((2 * hidden_dim, hidden_dim)))
bf = edf.Param(np.zeros((hidden_dim)))
# input gate
Wi = edf.Param(edf.xavier((2 * hidden_dim, hidden_dim)))
bi = edf.Param(np.zeros((hidden_dim)))
# carry cell
Wc = edf.Param(edf.xavier((2 * hidden_dim, hidden_dim)))
bc = edf.Param(np.zeros((hidden_dim)))
def MyRMSProp(eta, g, epoch=10):
Log("RMSProp With Learning Rate %.6f Decay Rate:%.4f \n" % (eta, g))
hidden_dim = 200
n_vocab = utils.n_vocab
batch = 50
parameters = []
model = 'Models/RMSProp/model_RMSProp_%.6f_%.4f_.pkl' % (eta, g)
#print(model)
eta = eta
decay = 0.9
inp = edf.Value()
edf.params = []
C2V = edf.Param(edf.xavier((n_vocab, hidden_dim)))
# forget gate
Wf = edf.Param(edf.xavier((2 * hidden_dim, hidden_dim)))
bf = edf.Param(np.zeros((hidden_dim)))
# input gate
Wi = edf.Param(edf.xavier((2 * hidden_dim, hidden_dim)))
bi = edf.Param(np.zeros((hidden_dim)))
# carry cell
Wc = edf.Param(edf.xavier((2 * hidden_dim, hidden_dim)))
bc = edf.Param(np.zeros((hidden_dim)))
# output cell
Wo = edf.Param(edf.xavier((2 * hidden_dim, hidden_dim)))
bo = edf.Param(np.zeros((hidden_dim)))
V = edf.Param(edf.xavier((hidden_dim, n_vocab)))
parameters.extend([C2V, Wf, bf, Wi, bi, Wc, bc, Wo, bo, V])
# load the trained model if exist
if os.path.exists(model):
with open(model, 'rb') as f:
p_value = pickle.load(f)
idx = 0
for p in p_value:
parameters[idx].value = p
idx += 1
def LSTMCell(xt, h, c):
f = edf.Sigmoid(edf.Add(edf.VDot(edf.ConCat(xt, h), Wf), bf))
i = edf.Sigmoid(edf.Add(edf.VDot(edf.ConCat(xt, h), Wi), bi))
o = edf.Sigmoid(edf.Add(edf.VDot(edf.ConCat(xt, h), Wo), bo))
c_hat = edf.Tanh(edf.Add(edf.VDot(edf.ConCat(xt, h), Wc), bc))
c_next = edf.Add(edf.Mul(f, c), edf.Mul(i, c_hat))
h_next = edf.Mul(o, edf.Tanh(c_next))
return h_next, c_next
def BuildModel():
edf.components = []
B = inp.value.shape[0]
T = inp.value.shape[1]
h = edf.Value(np.zeros((B, hidden_dim)))
c = edf.Value(np.zeros((B, hidden_dim)))
score = []
for t in range(T - 1):
wordvec = edf.Embed(edf.Value(inp.value[:, t]), C2V)
xt = edf.Reshape(wordvec, [-1, hidden_dim])
h_next, c_next = LSTMCell(xt, h, c)
p = edf.SoftMax(edf.VDot(h_next, V))
logloss = edf.Reshape(
edf.LogLoss(edf.Aref(p, edf.Value(inp.value[:, t + 1]))),
(B, 1))
if t == 0:
loss = logloss
else:
loss = edf.ConCat(loss, logloss)
score.append(p)
h = h_next
c = c_next
masks = np.zeros((B, T - 1), dtype=np.int32)
masks[inp.value[:, 1:] != 0] = 1
loss = edf.MeanwithMask(loss, edf.Value(masks))
return loss, score
def CalPerp(score):
prob = [p.value for p in score]
prob = np.transpose(np.stack(prob, axis=0), (1, 0, 2))
B = prob.shape[0]
T = prob.shape[1]
V = prob.shape[2]
masks = np.zeros((B, T), dtype=np.int32)
masks[inp.value[:, 1:] != 0] = 1
prob = prob.reshape(-1)
idx = np.int32(inp.value[:, 1:].reshape(-1))
outer_dim = len(idx)
inner_dim = len(prob) / outer_dim
pick = np.int32(np.array(range(outer_dim)) * inner_dim + idx)
prob = prob[pick].reshape(B, T)
return -np.sum(np.log(prob[np.nonzero(prob * masks)]))
def Predict(max_step, prefix):
edf.components = []
T = max_step
h = edf.Value(np.zeros((1, hidden_dim)))
c = edf.Value(np.zeros((1, hidden_dim)))
prediction = []
for t in range(T):
if t < len(prefix):
pred = edf.Value(prefix[t])
prediction.append(pred)
else:
prediction.append(pred)
wordvec = edf.Embed(pred, C2V)
xt = edf.Reshape(wordvec, [-1, hidden_dim])
h_next, c_next = LSTMCell(xt, h, c)
p = edf.SoftMax(edf.VDot(h_next, V))
pred = edf.ArgMax(p)
h = h_next
c = c_next
edf.Forward()
idx = [pred.value for pred in prediction]
stop_idx = utils.to_index('}')
if stop_idx in idx:
return idx[0:idx.index(stop_idx) + 1]
else:
return idx
def Eval(data, cnt):
perp = 0.
avg_loss = 0.
test_batches = range(0, len(data), batch)
test_minbatches = [data[idx:idx + batch] for idx in test_batches]
for minbatch in test_minbatches:
x_padded = utils.make_mask(minbatch)
inp.set(x_padded)
loss, score = BuildModel()
edf.Forward()
avg_loss += loss.value
perp += CalPerp(score)
perp = np.exp(perp / cnt)
avg_loss /= len(test_batches)
return perp, avg_loss
############################################### training loop #####################################################
batches = range(0, len(train_data), batch)
minbatches = [train_data[idx:idx + batch] for idx in batches]
epoch = epoch
# initial Perplexity and loss
#perp, loss = Eval(valid_data, vacnt)
#print("Initial: Perplexity: %0.5f Avg loss = %0.5f" % (perp, loss))
#best_loss = loss
#prefix = 'the agreements bring'
#generation = Predict(400, utils.to_idxs(prefix))
#print("Initial generated sentence ")
#print (utils.to_string(generation))
for ep in range(epoch):
perm = np.random.permutation(len(minbatches)).tolist()
stime = time()
for k in range(len(minbatches)):
minbatch = minbatches[perm[k]]
x_padded = utils.make_mask(minbatch)
inp.set(x_padded)
loss, score = BuildModel()
edf.Forward()
edf.Backward(loss)
edf.GradClip(10)
edf.RMSProp(eta, g)
duration = (time() - stime) / 60.
perp, loss = Eval(valid_data, vacnt)
Log("Epoch %d: Perplexity: %0.5f Avg loss = %0.5f [%.3f mins]" %
(ep, perp, loss, duration))
if (ep == epoch - 1):
# generate some text given the prefix and trained model
prefix = 'the agreements bring'
generation = Predict(400, utils.to_idxs(prefix))
Log("Epoch %d: generated sentence " % ep)
Log(utils.to_string(generation))
#if loss < best_loss:
# save the model
best_loss = loss
f = open(model, 'wb')
p_value = []
for p in parameters:
p_value.append(p.value)
pickle.dump(p_value, f)
#Save the hyperparameters
f_hyper = open("HyperParameters.txt", "a")
f_hyper.write(
"RMSProp LearningRate: %.6f Decay_Rate: %.4f Epoch: %d BestLoss: %0.5f Perplexity: %0.5f\n"
% (eta, g, ep, best_loss, perp))
if (ep == epoch - 1):
f_hyper.write("\n\n")
f_hyper.close()
Log("\n")
