def create_model(self):
input_dim = self.input_dim
x = self.x
y = self.y
p = self.p
mask = self.mask
hidden_dim = self.hidden_dim
embedding_dim = self.embedding_dim
lookup = LookupTable(self.dict_size,
embedding_dim,
weights_init=IsotropicGaussian(0.001),
name='LookupTable')
x_to_h = Linear(embedding_dim,
hidden_dim * 4,
name='x_to_h',
weights_init=IsotropicGaussian(0.001),
biases_init=Constant(0.0))
lstm = LSTM(hidden_dim,
name='lstm',
weights_init=IsotropicGaussian(0.001),
biases_init=Constant(0.0))
h_to_o = MLP([Logistic()], [hidden_dim, 1],
weights_init=IsotropicGaussian(0.001),
biases_init=Constant(0),
name='h_to_o')
lookup.initialize()
x_to_h.initialize()
lstm.initialize()
h_to_o.initialize()
embed = lookup.apply(x).reshape(
(x.shape[0], x.shape[1], self.embedding_dim))
embed.name = "embed_vec"
x_transform = x_to_h.apply(embed.transpose(1, 0, 2))
x_transform.name = "Transformed X"
self.lookup = lookup
self.x_to_h = x_to_h
self.lstm = lstm
self.h_to_o = h_to_o
#if mask is None:
h, c = lstm.apply(x_transform)
#else:
#h, c = lstm.apply(x_transform, mask=mask)
h.name = "hidden_state"
c.name = "cell state"
# only values of hidden units of the last timeframe are used for
# the classification
indices = T.sum(mask, axis=0) - 1
rel_hid = h[indices, T.arange(h.shape[1])]
out = self.h_to_o.apply(rel_hid)
probs = out
return probs
def __init__(self, input1_size, input2_size, lookup1_dim=200, lookup2_dim=200, hidden_size=512):
self.hidden_size = hidden_size
self.input1_size = input1_size
self.input2_size = input2_size
self.lookup1_dim = lookup1_dim
self.lookup2_dim = lookup2_dim
x1 = tensor.lmatrix('durations')
x2 = tensor.lmatrix('syllables')
y = tensor.lmatrix('pitches')
lookup1 = LookupTable(dim=self.lookup1_dim, length=self.input1_size, name='lookup1',
weights_init=initialization.Uniform(width=0.01),
biases_init=Constant(0))
lookup1.initialize()
lookup2 = LookupTable(dim=self.lookup2_dim, length=self.input2_size, name='lookup2',
weights_init=initialization.Uniform(width=0.01),
biases_init=Constant(0))
lookup2.initialize()
merge = Merge(['lookup1', 'lookup2'], [self.lookup1_dim, self.lookup2_dim], self.hidden_size,
weights_init=initialization.Uniform(width=0.01),
biases_init=Constant(0))
merge.initialize()
recurrent_block = LSTM(dim=self.hidden_size, activation=Tanh(),
weights_init=initialization.Uniform(width=0.01)) #RecurrentStack([LSTM(dim=self.hidden_size, activation=Tanh())] * 3)
recurrent_block.initialize()
linear = Linear(input_dim=self.hidden_size, output_dim=self.input1_size,
weights_init=initialization.Uniform(width=0.01),
biases_init=Constant(0))
linear.initialize()
softmax = NDimensionalSoftmax()
l1 = lookup1.apply(x1)
l2 = lookup2.apply(x2)
m = merge.apply(l1, l2)
h = recurrent_block.apply(m)
a = linear.apply(h)
y_hat = softmax.apply(a, extra_ndim=1)
# ValueError: x must be 1-d or 2-d tensor of floats. Got TensorType(float64, 3D)
self.Cost = softmax.categorical_cross_entropy(y, a, extra_ndim=1).mean()
self.ComputationGraph = ComputationGraph(self.Cost)
self.Model = Model(y_hat)
def construct_model(vocab_size, embedding_dim, ngram_order, hidden_dims,
activations):
# Construct the model
x = tensor.lmatrix('features')
y = tensor.lvector('targets')
lookup = LookupTable(length=vocab_size, dim=embedding_dim, name='lookup')
hidden = MLP(activations=activations + [None],
dims=[ngram_order * embedding_dim] + hidden_dims +
[vocab_size])
embeddings = lookup.apply(x)
embeddings = embeddings.flatten(ndim=2) # Concatenate embeddings
activations = hidden.apply(embeddings)
cost = Softmax().categorical_cross_entropy(y, activations)
# Initialize parameters
lookup.weights_init = IsotropicGaussian(0.001)
hidden.weights_init = IsotropicGaussian(0.01)
hidden.biases_init = Constant(0.001)
lookup.initialize()
hidden.initialize()
return cost
def construct_model(vocab_size, embedding_dim, ngram_order, hidden_dims,
activations):
# Construct the model
x = tensor.lmatrix('features')
y = tensor.lvector('targets')
lookup = LookupTable(length=vocab_size, dim=embedding_dim, name='lookup')
hidden = MLP(activations=activations + [None],
dims=[ngram_order * embedding_dim] + hidden_dims +
[vocab_size])
embeddings = lookup.apply(x)
embeddings = embeddings.flatten(ndim=2) # Concatenate embeddings
activations = hidden.apply(embeddings)
cost = Softmax().categorical_cross_entropy(y, activations)
# Initialize parameters
lookup.weights_init = IsotropicGaussian(0.001)
hidden.weights_init = IsotropicGaussian(0.01)
hidden.biases_init = Constant(0.001)
lookup.initialize()
hidden.initialize()
return cost
def __init__(self, config, vocab_size):
question = tensor.imatrix('question')
question_mask = tensor.imatrix('question_mask')
context = tensor.imatrix('context')
context_mask = tensor.imatrix('context_mask')
answer = tensor.imatrix('answer')
answer_mask = tensor.imatrix('answer_mask')
ans_indices = tensor.imatrix('ans_indices') # n_steps * n_samples
ans_indices_mask = tensor.imatrix('ans_indices_mask')
bricks = []
question = question.dimshuffle(1, 0)
question_mask = question_mask.dimshuffle(1, 0)
context = context.dimshuffle(1, 0)
context_mask = context_mask.dimshuffle(1, 0)
answer = answer.dimshuffle(1, 0)
answer_mask = answer_mask.dimshuffle(1, 0)
ans_indices = ans_indices.dimshuffle(1, 0)
ans_indices_mask = ans_indices_mask.dimshuffle(1, 0)
# Embed questions and context
embed = LookupTable(vocab_size, config.embed_size, name='embed')
#embed.weights_init = IsotropicGaussian(0.01)
embed.weights_init = Constant(
init_embedding_table(filename='embeddings/vocab_embeddings.txt'))
# one directional LSTM encoding
q_lstm_ins = Linear(input_dim=config.embed_size,
output_dim=4 * config.pre_lstm_size,
name='q_lstm_in')
q_lstm = LSTM(dim=config.pre_lstm_size,
activation=Tanh(),
name='q_lstm')
c_lstm_ins = Linear(input_dim=config.embed_size,
output_dim=4 * config.pre_lstm_size,
name='c_lstm_in')
c_lstm = LSTM(dim=config.pre_lstm_size,
activation=Tanh(),
name='c_lstm')
bricks += [q_lstm, c_lstm, q_lstm_ins, c_lstm_ins]
q_tmp = q_lstm_ins.apply(embed.apply(question))
c_tmp = c_lstm_ins.apply(embed.apply(context))
q_hidden, _ = q_lstm.apply(q_tmp,
mask=question_mask.astype(
theano.config.floatX)) # lq, bs, dim
c_hidden, _ = c_lstm.apply(c_tmp,
mask=context_mask.astype(
theano.config.floatX)) # lc, bs, dim
# Attention mechanism Bilinear question
attention_question = Linear(input_dim=config.pre_lstm_size,
output_dim=config.pre_lstm_size,
name='att_question')
bricks += [attention_question]
att_weights_question = q_hidden[
None, :, :, :] * attention_question.apply(
c_hidden.reshape(
(c_hidden.shape[0] * c_hidden.shape[1],
c_hidden.shape[2]))).reshape(
(c_hidden.shape[0], c_hidden.shape[1],
c_hidden.shape[2]))[:,
None, :, :] # --> lc,lq,bs,dim
att_weights_question = att_weights_question.sum(
axis=3) # sum over axis 3 -> dimensions --> lc,lq,bs
att_weights_question = att_weights_question.dimshuffle(
0, 2, 1) # --> lc,bs,lq
att_weights_question = att_weights_question.reshape(
(att_weights_question.shape[0] * att_weights_question.shape[1],
att_weights_question.shape[2])) # --> lc*bs,lq
att_weights_question = tensor.nnet.softmax(
att_weights_question
) # softmax over axis 1 -> length of question # --> lc*bs,lq
att_weights_question = att_weights_question.reshape(
(c_hidden.shape[0], q_hidden.shape[1],
q_hidden.shape[0])) # --> lc,bs,lq
att_weights_question = att_weights_question.dimshuffle(
0, 2, 1) # --> lc,lq,bs
attended_question = tensor.sum(
q_hidden[None, :, :, :] * att_weights_question[:, :, :, None],
axis=1) # sum over axis 1 -> length of question --> lc,bs,dim
attended_question.name = 'attended_question'
# Match LSTM
cqembed = tensor.concatenate([c_hidden, attended_question], axis=2)
mlstms, mhidden_list = make_bidir_lstm_stack(
cqembed, 2 * config.pre_lstm_size,
context_mask.astype(theano.config.floatX), config.match_lstm_size,
config.match_skip_connections, 'match')
bricks = bricks + mlstms
if config.match_skip_connections:
menc_dim = 2 * sum(config.match_lstm_size)
menc = tensor.concatenate(mhidden_list, axis=2)
else:
menc_dim = 2 * config.match_lstm_size[-1]
menc = tensor.concatenate(mhidden_list[-2:], axis=2)
menc.name = 'menc'
# Attention mechanism MLP start
attention_mlp_start = MLP(
dims=config.attention_mlp_hidden + [1],
activations=config.attention_mlp_activations[1:] + [Identity()],
name='attention_mlp_start')
attention_clinear_start = Linear(
input_dim=menc_dim,
output_dim=config.attention_mlp_hidden[0],
name='attm_start') # Wym
bricks += [attention_mlp_start, attention_clinear_start]
layer1_start = Tanh(name='layer1_start')
layer1_start = layer1_start.apply(
attention_clinear_start.apply(
menc.reshape(
(menc.shape[0] * menc.shape[1], menc.shape[2]))).reshape(
(menc.shape[0], menc.shape[1],
config.attention_mlp_hidden[0])))
att_weights_start = attention_mlp_start.apply(
layer1_start.reshape(
(layer1_start.shape[0] * layer1_start.shape[1],
layer1_start.shape[2])))
att_weights_start = att_weights_start.reshape(
(layer1_start.shape[0], layer1_start.shape[1]))
att_weights_start = tensor.nnet.softmax(att_weights_start.T).T
attended = tensor.sum(menc * att_weights_start[:, :, None], axis=0)
attended.name = 'attended'
# Attention mechanism MLP end
attention_mlp_end = MLP(
dims=config.attention_mlp_hidden + [1],
activations=config.attention_mlp_activations[1:] + [Identity()],
name='attention_mlp_end')
attention_qlinear_end = Linear(
input_dim=menc_dim,
output_dim=config.attention_mlp_hidden[0],
name='atts_end') #Wum
attention_clinear_end = Linear(
input_dim=menc_dim,
output_dim=config.attention_mlp_hidden[0],
use_bias=False,
name='attm_end') # Wym
bricks += [
attention_mlp_end, attention_qlinear_end, attention_clinear_end
]
layer1_end = Tanh(name='layer1_end')
layer1_end = layer1_end.apply(
attention_clinear_end.apply(
menc.reshape((menc.shape[0] * menc.shape[1], menc.shape[2]
))).reshape((menc.shape[0], menc.shape[1],
config.attention_mlp_hidden[0])) +
attention_qlinear_end.apply(attended)[None, :, :])
att_weights_end = attention_mlp_end.apply(
layer1_end.reshape((layer1_end.shape[0] * layer1_end.shape[1],
layer1_end.shape[2])))
att_weights_end = att_weights_end.reshape(
(layer1_end.shape[0], layer1_end.shape[1]))
att_weights_end = tensor.nnet.softmax(att_weights_end.T).T
att_weights_start = tensor.dot(
tensor.le(
tensor.tile(
theano.tensor.arange(context.shape[0])[None, :],
(context.shape[0], 1)),
tensor.tile(
theano.tensor.arange(context.shape[0])[:, None],
(1, context.shape[0]))), att_weights_start)
att_weights_end = tensor.dot(
tensor.ge(
tensor.tile(
theano.tensor.arange(context.shape[0])[None, :],
(context.shape[0], 1)),
tensor.tile(
theano.tensor.arange(context.shape[0])[:, None],
(1, context.shape[0]))), att_weights_end)
# add attention from left and right
att_weights = att_weights_start * att_weights_end
#att_weights = tensor.minimum(att_weights_start, att_weights_end)
att_target = tensor.zeros((ans_indices.shape[1], context.shape[0]),
dtype=theano.config.floatX)
att_target = tensor.set_subtensor(
att_target[tensor.arange(ans_indices.shape[1]), ans_indices], 1)
att_target = att_target.dimshuffle(1, 0)
#att_target = tensor.eq(tensor.tile(answer[None,:,:], (context.shape[0], 1, 1)),
# tensor.tile(context[:,None,:], (1, answer.shape[0], 1))).sum(axis=1).clip(0,1)
self.predictions = tensor.gt(att_weights, 0.25) * context
att_target = att_target / (att_target.sum(axis=0) + 0.00001)
#att_weights = att_weights / (att_weights.sum(axis=0) + 0.00001)
cost = (tensor.nnet.binary_crossentropy(att_weights, att_target) *
context_mask).sum() / context_mask.sum()
# Apply dropout
cg = ComputationGraph([cost])
if config.w_noise > 0:
noise_vars = VariableFilter(roles=[WEIGHT])(cg)
cg = apply_noise(cg, noise_vars, config.w_noise)
if config.dropout > 0:
cg = apply_dropout(cg, mhidden_list, config.dropout)
[cost_reg] = cg.outputs
# Other stuff
cost.name = 'cost'
cost_reg.name = 'cost_reg'
att_weights_start.name = 'att_weights_start'
att_weights_end.name = 'att_weights_end'
att_weights.name = 'att_weights'
att_target.name = 'att_target'
self.predictions.name = 'pred'
self.sgd_cost = cost_reg
self.monitor_vars = [[cost_reg]]
self.monitor_vars_valid = [[cost_reg]]
self.analyse_vars = [
cost, self.predictions, att_weights_start, att_weights_end,
att_weights, att_target
]
# Initialize bricks
embed.initialize()
for brick in bricks:
brick.weights_init = config.weights_init
brick.biases_init = config.biases_init
brick.initialize()
def create_model(self, symbols_num = 500):
# Hyperparameters
# The dimension of the hidden state of the GRUs in each direction.
hidden_states = self.args.encoder_hidden_dims
# Dimension of the word-embedding space
embedding_dims = self.args.source_embeddings_dim
###################
# Declaration of the Theano variables that come from the data stream
###################
# The context document.
context_bt = tt.lmatrix('context')
# Context document mask used to distinguish real symbols from the sequence and padding symbols that are at the end
context_mask_bt = tt.matrix('context_mask')
# The question
question_bt = tt.lmatrix('question')
question_mask_bt = tt.matrix('question_mask')
# The correct answer
y = tt.lmatrix('answer')
y = y[:,0] # originally answers are in a 2d matrix, here we convert it to a vector
# The candidates among which the answer is selected
candidates_bi = tt.lmatrix("candidates")
candidates_bi_mask = tt.matrix("candidates_mask")
###################
# Network's components
###################
# Lookup table with randomly initialized word embeddings
lookup = LookupTable(symbols_num, embedding_dims, weights_init=Uniform(width=0.2))
# bidirectional encoder that translates context
context_encoder = self.create_bidi_encoder("context_encoder", embedding_dims, hidden_states)
# bidirectional encoder for question
question_encoder = self.create_bidi_encoder("question_encoder", embedding_dims, hidden_states)
# Initialize the components (where not done upon creation)
lookup.initialize()
###################
# Wiring the components together
#
# Where present, the 3 letters at the end of the variable name identify its dimensions:
# b ... position of the example within the batch
# t ... position of the word within the document/question
# f ... features of the embedding vector
###################
### Read the context document
# Map token indices to word embeddings
context_embedding_tbf = lookup.apply(context_bt.T)
# Read the embedded context document using the bidirectional GRU and produce the contextual embedding of each word
memory_encoded_btf = context_encoder.apply(context_embedding_tbf, context_mask_bt.T).dimshuffle(1,0,2)
memory_encoded_btf.name = "memory_encoded_btf"
### Correspondingly, read the query
x_embedded_tbf = lookup.apply(question_bt.T)
x_encoded_btf = question_encoder.apply(x_embedded_tbf, question_mask_bt.T).dimshuffle(1,0,2)
# The query encoding is a concatenation of the final states of the forward and backward GRU encoder
x_forward_encoded_bf = x_encoded_btf[:,-1,0:hidden_states]
x_backward_encoded_bf = x_encoded_btf[:,0,hidden_states:hidden_states*2]
query_representation_bf = tt.concatenate([x_forward_encoded_bf,x_backward_encoded_bf],axis=1)
# Compute the attention on each word in the context as a dot product of its contextual embedding and the query
mem_attention_presoft_bt = tt.batched_dot(query_representation_bf, memory_encoded_btf.dimshuffle(0,2,1))
# TODO is this pre-masking necessary?
mem_attention_presoft_masked_bt = tt.mul(mem_attention_presoft_bt,context_mask_bt)
# Normalize the attention using softmax
mem_attention_bt = SoftmaxWithMask(name="memory_query_softmax").apply(mem_attention_presoft_masked_bt,context_mask_bt)
if self.args.weighted_att:
# compute weighted attention over original word vectors
att_weighted_responses_bf = theano.tensor.batched_dot(mem_attention_bt, context_embedding_tbf.dimshuffle(1,0,2))
# compare desired response to all candidate responses
# select relevant candidate answer words
candidates_embeddings_bfi = lookup.apply(candidates_bi).dimshuffle(0,2,1)
# convert it to output symbol probabilities
y_hat_presoft = tt.batched_dot(att_weighted_responses_bf, candidates_embeddings_bfi)
y_hat = SoftmaxWithMask(name="output_softmax").apply(y_hat_presoft,candidates_bi_mask)
else:
# Sum the attention of each candidate word across the whole context document,
# this is the key innovation of the model
# TODO: Get rid of sentence-by-sentence processing?
# TODO: Rewrite into matrix notation instead of scans?
def sum_prob_of_word(word_ix, sentence_ixs, sentence_attention_probs):
word_ixs_in_sentence = tt.eq(sentence_ixs,word_ix).nonzero()[0]
return sentence_attention_probs[word_ixs_in_sentence].sum()
def sum_probs_single_sentence(candidate_indices_i, sentence_ixs_t, sentence_attention_probs_t):
result, updates = theano.scan(
fn=sum_prob_of_word,
sequences=[candidate_indices_i],
non_sequences=[sentence_ixs_t, sentence_attention_probs_t])
return result
def sum_probs_batch(candidate_indices_bt,sentence_ixs_bt, sentence_attention_probs_bt):
result, updates = theano.scan(
fn=sum_probs_single_sentence,
sequences=[candidate_indices_bt, sentence_ixs_bt, sentence_attention_probs_bt],
non_sequences=None)
return result
# Sum the attention of each candidate word across the whole context document
y_hat = sum_probs_batch(candidates_bi, context_bt, mem_attention_bt)
y_hat.name = "y_hat"
# We use the convention that ground truth is always at index 0, so the following are the target answers
y = y.zeros_like()
# We use Cross Entropy as the training objective
cost = CategoricalCrossEntropy().apply(y.flatten(), y_hat)
cost.name = "cost"
predicted_response_index = tt.argmax(y_hat,axis=1)
accuracy = tt.eq(y,predicted_response_index).mean()
accuracy.name = "accuracy"
return cost, accuracy, mem_attention_bt, y_hat, context_bt, candidates_bi, candidates_bi_mask, y, context_mask_bt, question_bt, question_mask_bt
def __init__(self, config, vocab_size):
question = tensor.imatrix('question')
question_mask = tensor.imatrix('question_mask')
context = tensor.imatrix('context')
context_mask = tensor.imatrix('context_mask')
answer = tensor.imatrix('answer')
answer_mask = tensor.imatrix('answer_mask')
bricks = []
question = question.dimshuffle(1, 0)
question_mask = question_mask.dimshuffle(1, 0)
context = context.dimshuffle(1, 0)
context_mask = context_mask.dimshuffle(1, 0)
answer = answer.dimshuffle(1, 0)
answer_mask = answer_mask.dimshuffle(1, 0)
# Embed questions and context
embed = LookupTable(vocab_size,
config.embed_size,
name='question_embed')
embed.weights_init = IsotropicGaussian(0.01)
# Calculate question encoding (concatenate layer1)
qembed = embed.apply(question)
qlstms, qhidden_list = make_bidir_lstm_stack(
qembed, config.embed_size,
question_mask.astype(theano.config.floatX),
config.question_lstm_size, config.question_skip_connections, 'q')
bricks = bricks + qlstms
if config.question_skip_connections:
qenc_dim = 2 * sum(config.question_lstm_size)
qenc = tensor.concatenate([h[-1, :, :] for h in qhidden_list],
axis=1)
else:
qenc_dim = 2 * config.question_lstm_size[-1]
qenc = tensor.concatenate([h[-1, :, :] for h in qhidden_list[-2:]],
axis=1)
qenc.name = 'qenc'
# Calculate context encoding (concatenate layer1)
cembed = embed.apply(context)
cqembed = tensor.concatenate([
cembed,
tensor.extra_ops.repeat(qenc[None, :, :], cembed.shape[0], axis=0)
],
axis=2)
clstms, chidden_list = make_bidir_lstm_stack(
cqembed, config.embed_size + qenc_dim,
context_mask.astype(theano.config.floatX), config.ctx_lstm_size,
config.ctx_skip_connections, 'ctx')
bricks = bricks + clstms
if config.ctx_skip_connections:
cenc_dim = 2 * sum(config.ctx_lstm_size) #2 : fw & bw
cenc = tensor.concatenate(chidden_list, axis=2)
else:
cenc_dim = 2 * config.question_lstm_size[-1]
cenc = tensor.concatenate(chidden_list[-2:], axis=2)
cenc.name = 'cenc'
# Attention mechanism Bilinear
attention_clinear_1 = Linear(input_dim=cenc_dim,
output_dim=qenc_dim,
name='attc_1')
bricks += [attention_clinear_1]
att_start = qenc[None, :, :] * attention_clinear_1.apply(
cenc.reshape(
(cenc.shape[0] * cenc.shape[1], cenc.shape[2]))).reshape(
(cenc.shape[0], cenc.shape[1], cenc.shape[2]))
att_start = att_start.sum(axis=2)
att_start = tensor.nnet.softmax(att_start.T).T
attention_clinear_2 = Linear(input_dim=cenc_dim,
output_dim=qenc_dim,
name='attc_2')
bricks += [attention_clinear_2]
att_end = qenc[None, :, :] * attention_clinear_2.apply(
cenc.reshape(
(cenc.shape[0] * cenc.shape[1], cenc.shape[2]))).reshape(
(cenc.shape[0], cenc.shape[1], cenc.shape[2]))
att_end = att_end.sum(axis=2)
att_end = tensor.nnet.softmax(att_end.T).T
att_start = tensor.dot(
tensor.le(
tensor.tile(
theano.tensor.arange(context.shape[0])[None, :],
(context.shape[0], 1)),
tensor.tile(
theano.tensor.arange(context.shape[0])[:, None],
(1, context.shape[0]))), att_start)
att_end = tensor.dot(
tensor.ge(
tensor.tile(
theano.tensor.arange(context.shape[0])[None, :],
(context.shape[0], 1)),
tensor.tile(
theano.tensor.arange(context.shape[0])[:, None],
(1, context.shape[0]))), att_end)
# add attention from left and right
att_weights = att_start * att_end
att_target = tensor.eq(
tensor.tile(answer[None, :, :], (context.shape[0], 1, 1)),
tensor.tile(context[:, None, :],
(1, answer.shape[0], 1))).sum(axis=1).clip(0, 1)
self.predictions = tensor.gt(att_weights, 0.25) * context
att_target = att_target / (att_target.sum(axis=0) + 0.00001)
att_weights = att_weights / (att_weights.sum(axis=0) + 0.00001)
#cost = (tensor.nnet.binary_crossentropy(att_weights, att_target) * context_mask).sum() / context_mask.sum()
cost = (((att_weights - att_target)**2) *
context_mask).sum() / context_mask.sum()
# Apply dropout
cg = ComputationGraph([cost])
if config.w_noise > 0:
noise_vars = VariableFilter(roles=[WEIGHT])(cg)
cg = apply_noise(cg, noise_vars, config.w_noise)
if config.dropout > 0:
cg = apply_dropout(cg, qhidden_list + chidden_list, config.dropout)
[cost_reg] = cg.outputs
# Other stuff
cost.name = 'cost'
cost_reg.name = 'cost_reg'
att_start.name = 'att_start'
att_end.name = 'att_end'
att_weights.name = 'att_weights'
att_target.name = 'att_target'
self.predictions.name = 'pred'
self.sgd_cost = cost_reg
self.monitor_vars = [[cost_reg]]
self.monitor_vars_valid = [[cost_reg]]
self.analyse_vars = [
cost, self.predictions, att_start, att_end, att_weights, att_target
]
# Initialize bricks
embed.initialize()
for brick in bricks:
brick.weights_init = config.weights_init
brick.biases_init = config.biases_init
brick.initialize()
class NeuralLM:
def __init__(self, x, y, vocab_size, hidden_size, num_layers, pretrained_embeds=None):
"""
Implements a neural language model using an LSTM.
Word y_n+1 ~ Softmax(U * h_n)
:param x A minibatch: each row is an instance (a sequence),
with batch_size rows
:param y x shifted by 1, which are the target words to predict
for the language modeling objective based on the hidden LSTM
state
:param vocab_size The number of types in the training data
:param hidden_size The dimensionality of the word embeddings
:param pretrained_embeds Pretrained embeddings for initailization as an ND array
"""
self.vocab_size = vocab_size
self.hidden_size = hidden_size
self.num_layers = num_layers
# Initialize the word embedding table. If we have pretrained embeddings, we use those
self.word_embedding_lookup = LookupTable(length=vocab_size, dim=hidden_size, name="word_embeddings")
if pretrained_embeds is None:
initialize(self.word_embedding_lookup, 0.8)
else:
assert pretrained_embeds.shape[0] == vocab_size and pretrained_embeds.shape[1] == hidden_size
self.word_embedding_lookup.weights_init = Constant(pretrained_embeds)
self.word_embedding_lookup.biases_init = Constant(0)
self.word_embedding_lookup.initialize()
self.word_embeddings = self.word_embedding_lookup.W
self.y_hat, self.cost, self.cells = self.nn_fprop(x, y, num_layers)
def lstm_layer(self, h, n):
"""
Performs the LSTM update for a batch of word sequences
:param h The word embeddings for this update
:param n The number of layers of the LSTM
"""
# Maps the word embedding to a dimensionality to be used in the LSTM
linear = Linear(input_dim=self.hidden_size, output_dim=self.hidden_size * 4, name='linear_lstm' + str(n))
initialize(linear, sqrt(6.0 / (5 * self.hidden_size)))
lstm = LSTM(dim=self.hidden_size, name='lstm' + str(n))
initialize(lstm, 0.08)
return lstm.apply(linear.apply(h))
def softmax_layer(self, h, y):
"""
Perform Softmax over the hidden state in order to
predict the next word in the sequence and compute
the loss.
:param h The hidden state sequence
:param y The target words
"""
hidden_to_output = Linear(name='hidden_to_output', input_dim=self.hidden_size,
output_dim=self.vocab_size)
initialize(hidden_to_output, sqrt(6.0 / (self.hidden_size + self.vocab_size)))
linear_output = hidden_to_output.apply(h)
linear_output.name = 'linear_output'
softmax = NDimensionalSoftmax(name="lm_softmax")
y_hat = softmax.log_probabilities(linear_output, extra_ndim=1)
y_hat.name = 'y_hat'
cost = softmax.categorical_cross_entropy(y, linear_output, extra_ndim=1).mean()
cost.name = 'cost'
return y_hat, cost
def nn_fprop(self, x, y, num_layers):
h = T.nnet.sigmoid(self.word_embedding_lookup.apply(x)) # constrain the word embeddings
cells = []
for i in range(num_layers):
h, c = self.lstm_layer(h, i)
cells.append(c)
return self.softmax_layer(h, y) + (cells, )
@property
def cost(self):
return self.cost
@property
def embeddings(self):
return self.word_embeddings
def _build_lookup(self, name, word_num, dim=1, *args, **kwargs):
lookup = LookupTable(length=word_num, dim=dim, name=name)
lookup.weights_init = Constant(1. / word_num**0.25)
lookup.initialize()
return lookup
def run(epochs=1, corpus="data/", HIDDEN_DIMS=100, path="./"):
brown = BrownDataset(corpus)
INPUT_DIMS = brown.get_vocabulary_size()
OUTPUT_DIMS = brown.get_vocabulary_size()
# These are theano variables
x = tensor.lmatrix('context')
y = tensor.ivector('output')
# Construct the graph
input_to_hidden = LookupTable(name='input_to_hidden', length=INPUT_DIMS,
dim=HIDDEN_DIMS)
# Compute the weight matrix for every word in the context and then compute
# the average.
h = tensor.mean(input_to_hidden.apply(x), axis=1)
hidden_to_output = Linear(name='hidden_to_output', input_dim=HIDDEN_DIMS,
output_dim=OUTPUT_DIMS)
y_hat = Softmax().apply(hidden_to_output.apply(h))
# And initialize with random varibales and set the bias vector to 0
weights = IsotropicGaussian(0.01)
input_to_hidden.weights_init = hidden_to_output.weights_init = weights
input_to_hidden.biases_init = hidden_to_output.biases_init = Constant(0)
input_to_hidden.initialize()
hidden_to_output.initialize()
# And now the cost function
cost = CategoricalCrossEntropy().apply(y, y_hat)
cg = ComputationGraph(cost)
W1, W2 = VariableFilter(roles=[WEIGHT])(cg.variables)
cost = cost + 0.01 * (W1 ** 2).sum() + 0.01 * (W2 ** 2).sum()
cost.name = 'cost_with_regularization'
mini_batch = SequentialScheme(brown.num_instances(), 512)
data_stream = DataStream.default_stream(brown, iteration_scheme=mini_batch)
# Now we tie up lose ends and construct the algorithm for the training
# and define what happens in the main loop.
algorithm = GradientDescent(cost=cost, parameters=cg.parameters,
step_rule=Scale(learning_rate=0.1))
extensions = [
ProgressBar(),
FinishAfter(after_n_epochs=epochs),
Printing(),
# TrainingDataMonitoring(variables=[cost]),
SaveWeights(layers=[input_to_hidden, hidden_to_output],
prefixes=['%sfirst' % path, '%ssecond' % path]),
# Plot(
# 'Word Embeddings',
# channels=[
# [
# 'cost_with_regularization'
# ]
# ])
]
logger.info("Starting main loop...")
main = MainLoop(data_stream=data_stream,
algorithm=algorithm,
extensions=extensions)
main.run()
pickle.dump(cg, open('%scg.pickle' % path, 'wb'))
def __init__(self, config, vocab_size):
context = tensor.imatrix('context')
context_mask = tensor.imatrix('context_mask')
answer = tensor.imatrix('answer')
answer_mask = tensor.imatrix('answer_mask')
bricks = []
context = context.dimshuffle(1, 0)
context_mask = context_mask.dimshuffle(1, 0)
answer = answer.dimshuffle(1, 0)
answer_mask = answer_mask.dimshuffle(1, 0)
context_bag = to_bag(context, vocab_size)
# Embed questions and context
embed = LookupTable(vocab_size, config.embed_size, name='embed')
embed.weights_init = IsotropicGaussian(0.01)
#embeddings_initial_value = init_embedding_table(filename='embeddings/vocab_embeddings.txt')
#embed.weights_init = Constant(embeddings_initial_value)
# Calculate context encoding (concatenate layer1)
cembed = embed.apply(context)
clstms, chidden_list = make_bidir_lstm_stack(
cembed, config.embed_size,
context_mask.astype(theano.config.floatX), config.ctx_lstm_size,
config.ctx_skip_connections, 'ctx')
bricks = bricks + clstms
if config.ctx_skip_connections:
cenc_dim = 2 * sum(config.ctx_lstm_size) #2 : fw & bw
cenc = tensor.concatenate(chidden_list, axis=2)
else:
cenc_dim = 2 * config.ctx_lstm_size[-1]
cenc = tensor.concatenate(chidden_list[-2:], axis=2)
cenc.name = 'cenc'
# Build the encoder bricks
transition = GatedRecurrent(activation=Tanh(),
dim=config.generator_lstm_size,
name="transition")
attention = SequenceContentAttention(
state_names=transition.apply.states,
attended_dim=cenc_dim,
match_dim=config.generator_lstm_size,
name="attention")
readout = Readout(readout_dim=vocab_size,
source_names=[
transition.apply.states[0],
attention.take_glimpses.outputs[0]
],
emitter=MaskedSoftmaxEmitter(context_bag=context_bag,
name='emitter'),
feedback_brick=LookupFeedback(
vocab_size, config.feedback_size),
name="readout")
generator = SequenceGenerator(readout=readout,
transition=transition,
attention=attention,
name="generator")
cost = generator.cost(answer,
answer_mask.astype(theano.config.floatX),
attended=cenc,
attended_mask=context_mask.astype(
theano.config.floatX),
name="cost")
self.predictions = generator.generate(
n_steps=7,
batch_size=config.batch_size,
attended=cenc,
attended_mask=context_mask.astype(theano.config.floatX),
iterate=True)[1]
# Apply dropout
cg = ComputationGraph([cost])
if config.w_noise > 0:
noise_vars = VariableFilter(roles=[WEIGHT])(cg)
cg = apply_noise(cg, noise_vars, config.w_noise)
if config.dropout > 0:
cg = apply_dropout(cg, chidden_list, config.dropout)
[cost_reg] = cg.outputs
# Other stuff
cost.name = 'cost'
cost_reg.name = 'cost_reg'
self.sgd_cost = cost_reg
self.monitor_vars = [[cost_reg]]
self.monitor_vars_valid = [[cost_reg]]
# initialize new stuff manually (change!)
generator.weights_init = IsotropicGaussian(0.01)
generator.biases_init = Constant(0)
generator.push_allocation_config()
generator.push_initialization_config()
transition.weights_init = Orthogonal()
generator.initialize()
# Initialize bricks
embed.initialize()
for brick in bricks:
brick.weights_init = config.weights_init
brick.biases_init = config.biases_init
brick.initialize()
def create_model(self):
input_dim = self.input_dim
x = self.x
y = self.y
p = self.p
mask = self.mask
hidden_dim = self.hidden_dim
embedding_dim = self.embedding_dim
lookup = LookupTable(self.dict_size,
embedding_dim,
weights_init=IsotropicGaussian(0.001),
name='LookupTable')
x_to_h = Linear(embedding_dim,
hidden_dim * 4,
name='x_to_h',
weights_init=IsotropicGaussian(0.001),
biases_init=Constant(0.0))
lstm = LSTM(hidden_dim,
name='lstm',
weights_init=IsotropicGaussian(0.001),
biases_init=Constant(0.0))
h_to_o = MLP([Logistic()], [hidden_dim, 1],
weights_init=IsotropicGaussian(0.001),
biases_init=Constant(0),
name='h_to_o')
lookup.initialize()
x_to_h.initialize()
lstm.initialize()
h_to_o.initialize()
embed = lookup.apply(x).reshape(
(x.shape[0], x.shape[1], self.embedding_dim))
embed.name = "embed_vec"
x_transform = x_to_h.apply(embed.transpose(1, 0, 2))
x_transform.name = "Transformed X"
self.lookup = lookup
self.x_to_h = x_to_h
self.lstm = lstm
self.h_to_o = h_to_o
#if mask is None:
h, c = lstm.apply(x_transform)
#else:
#h, c = lstm.apply(x_transform, mask=mask)
h.name = "hidden_state"
c.name = "cell state"
# only values of hidden units of the last timeframe are used for
# the classification
indices = T.sum(mask, axis=0) - 1
rel_hid = h[indices, T.arange(h.shape[1])]
out = self.h_to_o.apply(rel_hid)
probs = 1 - out
probs.name = "probability"
y = y.dimshuffle(0, 'x')
# Create the if-else cost function
pos_ex = (y * probs) / p
neg_ex = (1 - y) * (1 - probs) / np.float32(1 - p)
reward = pos_ex + neg_ex
cost = reward # Negative of reward
cost.name = "cost"
return cost
x = tensor.imatrix('features')
y = tensor.ivector('targets')
v = dataset.get_vocab_size()
input_to_hidden = LookupTable(name='input_to_hidden', length=v, dim=hidden_size)
h = tensor.mean(input_to_hidden.apply(x), axis=1)
hidden_to_output = Linear(name='hidden_to_output', input_dim=hidden_size, output_dim=v)
y_hat = Softmax().apply(hidden_to_output.apply(h))
input_to_hidden.weights_init = hidden_to_output.weights_init = IsotropicGaussian(0.01)
input_to_hidden.biases_init = hidden_to_output.biases_init = Constant(0)
input_to_hidden.initialize()
hidden_to_output.initialize()
cost = CategoricalCrossEntropy().apply(y, y_hat)
cg = ComputationGraph(cost)
W1, W2 = VariableFilter(roles=[WEIGHT])(cg.variables)
cost = cost + 0.005 * (abs(W1)).sum() + 0.005 * (abs(W2)).sum()
cost.name = 'cost'
algorithm = GradientDescent(cost=cost,
parameters=cg.parameters,
step_rule=Scale(learning_rate=0.1))
def main(model_path, recurrent_type):
dataset_options = dict(dictionary=char2code, level="character",
preprocess=_lower)
dataset = OneBillionWord("training", [99], **dataset_options)
data_stream = dataset.get_example_stream()
data_stream = Filter(data_stream, _filter_long)
data_stream = Mapping(data_stream, _make_target,
add_sources=('target',))
data_stream = Batch(data_stream, iteration_scheme=ConstantScheme(100))
data_stream = Padding(data_stream)
data_stream = Mapping(data_stream, _transpose)
features = tensor.lmatrix('features')
features_mask = tensor.matrix('features_mask')
target = tensor.lmatrix('target')
target_mask = tensor.matrix('target_mask')
dim = 100
lookup = LookupTable(len(all_chars), dim,
weights_init=IsotropicGaussian(0.01),
biases_init=Constant(0.))
if recurrent_type == 'lstm':
rnn = LSTM(dim / 4, Tanh(),
weights_init=IsotropicGaussian(0.01),
biases_init=Constant(0.))
elif recurrent_type == 'simple':
rnn = SimpleRecurrent(dim, Tanh())
rnn = Bidirectional(rnn,
weights_init=IsotropicGaussian(0.01),
biases_init=Constant(0.))
else:
raise ValueError('Not known RNN type')
rnn.initialize()
lookup.initialize()
y_hat = rnn.apply(lookup.apply(features), mask=features_mask)
print len(all_chars)
linear = Linear(2 * dim, len(all_chars),
weights_init=IsotropicGaussian(0.01),
biases_init=Constant(0.))
linear.initialize()
y_hat = linear.apply(y_hat)
seq_lenght = y_hat.shape[0]
batch_size = y_hat.shape[1]
y_hat = Softmax().apply(y_hat.reshape((seq_lenght * batch_size, -1))).reshape(y_hat.shape)
cost = CategoricalCrossEntropy().apply(
target.flatten(),
y_hat.reshape((-1, len(all_chars)))) * seq_lenght * batch_size
cost.name = 'cost'
cost_per_character = cost / features_mask.sum()
cost_per_character.name = 'cost_per_character'
cg = ComputationGraph([cost, cost_per_character])
model = Model(cost)
algorithm = GradientDescent(step_rule=Adam(), cost=cost,
params=cg.parameters)
train_monitor = TrainingDataMonitoring(
[cost, cost_per_character], prefix='train',
after_batch=True)
extensions = [train_monitor, Printing(every_n_batches=40),
Dump(model_path, every_n_batches=200),
#Checkpoint('rnn.pkl', every_n_batches=200)
]
main_loop = MainLoop(model=model, algorithm=algorithm,
data_stream=data_stream, extensions=extensions)
main_loop.run()
def main(config):
vocab_src, _ = text_to_dict([config['train_src'],
config['dev_src'], config['test_src']])
vocab_tgt, cabvo = text_to_dict([config['train_tgt'],
config['dev_tgt']])
# Create Theano variables
logger.info('Creating theano variables')
source_sentence = tensor.lmatrix('source')
source_sentence_mask = tensor.matrix('source_mask')
target_sentence = tensor.lmatrix('target')
target_sentence_mask = tensor.matrix('target_mask')
source_sentence.tag.test_value = [[13, 20, 0, 20, 0, 20, 0],
[1, 4, 8, 4, 8, 4, 8],]
source_sentence_mask.tag.test_value = [[0, 1, 0, 1, 0, 1, 0],
[1, 0, 1, 0, 1, 0, 1],]
target_sentence.tag.test_value = [[0,1,1,5],
[2,0,1,0],]
target_sentence_mask.tag.test_value = [[0,1,1,0],
[1,1,1,0],]
logger.info('Building RNN encoder-decoder')
### Building Encoder
embedder = LookupTable(
length=len(vocab_src),
dim=config['embed_src'],
weights_init=IsotropicGaussian(),
biases_init=Constant(0.0),
name='embedder')
transformer = Linear(
config['embed_src'],
config['hidden_src']*4,
weights_init=IsotropicGaussian(),
biases_init=Constant(0.0),
name='transformer')
lstminit = np.asarray([0.0,]*config['hidden_src']+[0.0,]*config['hidden_src']+[1.0,]*config['hidden_src']+[0.0,]*config['hidden_src'])
encoder = Bidirectional(
LSTM(
dim=config['hidden_src'],
weights_init=IsotropicGaussian(0.01),
biases_init=Constant(lstminit)),
name='encoderBiLSTM'
)
encoder.prototype.weights_init = Orthogonal()
### Building Decoder
lstminit = np.asarray([0.0,]*config['hidden_tgt']+[0.0,]*config['hidden_tgt']+[1.0,]*config['hidden_tgt']+[0.0,]*config['hidden_tgt'])
transition = LSTM2GO(
attended_dim=config['hidden_tgt'],
dim=config['hidden_tgt'],
weights_init=IsotropicGaussian(0.01),
biases_init=Constant(lstminit),
name='decoderLSTM')
attention = SequenceContentAttention(
state_names=transition.apply.states, # default activation is Tanh
state_dims=[config['hidden_tgt']],
attended_dim=config['hidden_src']*2,
match_dim=config['hidden_tgt'],
name="attention")
readout = Readout(
source_names=['states',
'feedback',
attention.take_glimpses.outputs[0]],
readout_dim=len(vocab_tgt),
emitter = SoftmaxEmitter(
name='emitter'),
feedback_brick = LookupFeedback(
num_outputs=len(vocab_tgt),
feedback_dim=config['embed_tgt'],
name='feedback'),
post_merge=InitializableFeedforwardSequence([
Bias(dim=config['hidden_tgt'],
name='softmax_bias').apply,
Linear(input_dim=config['hidden_tgt'],
output_dim=config['embed_tgt'],
use_bias=False,
name='softmax0').apply,
Linear(input_dim=config['embed_tgt'],
name='softmax1').apply]),
merged_dim=config['hidden_tgt'])
decoder = SequenceGenerator(
readout=readout,
transition=transition,
attention=attention,
weights_init=IsotropicGaussian(0.01),
biases_init=Constant(0),
name="generator",
fork=Fork(
[name for name in transition.apply.sequences if name != 'mask'],
prototype=Linear()),
add_contexts=True)
decoder.transition.weights_init = Orthogonal()
#printchildren(encoder, 1)
# Initialize model
logger.info('Initializing model')
embedder.initialize()
transformer.initialize()
encoder.initialize()
decoder.initialize()
# Apply model
embedded = embedder.apply(source_sentence)
tansformed = transformer.apply(embedded)
encoded = encoder.apply(tansformed)[0]
generated = decoder.generate(
n_steps=2*source_sentence.shape[1],
batch_size=source_sentence.shape[0],
attended = encoded.dimshuffle(1,0,2),
attended_mask=tensor.ones(source_sentence.shape).T
)
print 'Generated: ', generated
# generator_generate_outputs
#samples = generated[1] # For GRU
samples = generated[2] # For LSTM
samples.name = 'samples'
#samples_cost = generated[4] # For GRU
samples_cost = generated[5] # For LSTM
samples_cost = 'sampling_cost'
cost = decoder.cost(
mask = target_sentence_mask.T,
outputs = target_sentence.T,
attended = encoded.dimshuffle(1,0,2),
attended_mask = source_sentence_mask.T)
cost.name = 'target_cost'
cost.tag.aggregation_scheme = TakeLast(cost)
model = Model(cost)
logger.info('Creating computational graph')
cg = ComputationGraph(cost)
# apply dropout for regularization
if config['dropout'] < 1.0: # dropout is applied to the output of maxout in ghog
logger.info('Applying dropout')
dropout_inputs = [x for x in cg.intermediary_variables if x.name == 'maxout_apply_output']
cg = apply_dropout(cg, dropout_inputs, config['dropout'])
########
# Print shapes
shapes = [param.get_value().shape for param in cg.parameters]
logger.info("Parameter shapes: ")
for shape, count in Counter(shapes).most_common():
logger.info(' {:15}: {}'.format(shape, count))
logger.info("Total number of parameters: {}".format(len(shapes)))
printchildren(embedder, 1)
printchildren(transformer, 1)
printchildren(encoder, 1)
printchildren(decoder, 1)
# Print parameter names
# enc_dec_param_dict = merge(Selector(embedder).get_parameters(), Selector(encoder).get_parameters(), Selector(decoder).get_parameters())
# enc_dec_param_dict = merge(Selector(decoder).get_parameters())
# logger.info("Parameter names: ")
# for name, value in enc_dec_param_dict.items():
# logger.info(' {:15}: {}'.format(value.get_value().shape, name))
# logger.info("Total number of parameters: {}".format(len(enc_dec_param_dict)))
##########
# Training data
train_stream = get_train_stream(config,
[config['train_src'],], [config['train_tgt'],],
vocab_src, vocab_tgt)
dev_stream = get_dev_stream(
[config['dev_src'],], [config['dev_tgt'],],
vocab_src, vocab_tgt)
test_stream = get_test_stream([config['test_src'],], vocab_src)
# Set extensions
logger.info("Initializing extensions")
extensions = [
FinishAfter(after_n_batches=config['finish_after']),
ProgressBar(),
TrainingDataMonitoring([cost],
prefix="tra",
after_batch=True),
DataStreamMonitoring(variables=[cost],
data_stream=dev_stream,
prefix="dev",
after_batch=True),
Sampler(
model=Model(samples),
data_stream=dev_stream,
vocab=cabvo,
saveto=config['saveto']+'dev',
every_n_batches=config['save_freq']),
Sampler(
model=Model(samples),
data_stream=test_stream,
vocab=cabvo,
saveto=config['saveto']+'test',
after_n_batches=1,
on_resumption=True,
before_training=True),
Plotter(saveto=config['saveto'], after_batch=True),
Printing(after_batch=True),
Checkpoint(
path=config['saveto'],
parameters = cg.parameters,
save_main_loop=False,
every_n_batches=config['save_freq'])]
if BOKEH_AVAILABLE:
Plot('Training cost', channels=[['target_cost']], after_batch=True)
if config['reload']:
extensions.append(Load(path=config['saveto'],
load_iteration_state=False,
load_log=False))
else:
with open(config['saveto']+'.txt', 'w') as f:
pass
# Set up training algorithm
logger.info("Initializing training algorithm")
algorithm = GradientDescent(cost=cost,
parameters=cg.parameters,
step_rule=CompositeRule([StepClipping(config['step_clipping']),
eval(config['step_rule'])()])
)
# Initialize main loop
logger.info("Initializing main loop")
main_loop = MainLoop(
model=model,
algorithm=algorithm,
data_stream=train_stream,
extensions=extensions)
main_loop.run()
def __init__(self,
input1_size,
input2_size,
lookup1_dim=200,
lookup2_dim=200,
hidden_size=512):
self.hidden_size = hidden_size
self.input1_size = input1_size
self.input2_size = input2_size
self.lookup1_dim = lookup1_dim
self.lookup2_dim = lookup2_dim
x1 = tensor.lmatrix('durations')
x2 = tensor.lmatrix('syllables')
y = tensor.lmatrix('pitches')
lookup1 = LookupTable(dim=self.lookup1_dim,
length=self.input1_size,
name='lookup1',
weights_init=initialization.Uniform(width=0.01),
biases_init=Constant(0))
lookup1.initialize()
lookup2 = LookupTable(dim=self.lookup2_dim,
length=self.input2_size,
name='lookup2',
weights_init=initialization.Uniform(width=0.01),
biases_init=Constant(0))
lookup2.initialize()
merge = Merge(['lookup1', 'lookup2'],
[self.lookup1_dim, self.lookup2_dim],
self.hidden_size,
weights_init=initialization.Uniform(width=0.01),
biases_init=Constant(0))
merge.initialize()
recurrent_block = LSTM(
dim=self.hidden_size,
activation=Tanh(),
weights_init=initialization.Uniform(width=0.01)
) #RecurrentStack([LSTM(dim=self.hidden_size, activation=Tanh())] * 3)
recurrent_block.initialize()
linear = Linear(input_dim=self.hidden_size,
output_dim=self.input1_size,
weights_init=initialization.Uniform(width=0.01),
biases_init=Constant(0))
linear.initialize()
softmax = NDimensionalSoftmax()
l1 = lookup1.apply(x1)
l2 = lookup2.apply(x2)
m = merge.apply(l1, l2)
h = recurrent_block.apply(m)
a = linear.apply(h)
y_hat = softmax.apply(a, extra_ndim=1)
# ValueError: x must be 1-d or 2-d tensor of floats. Got TensorType(float64, 3D)
self.Cost = softmax.categorical_cross_entropy(y, a,
extra_ndim=1).mean()
self.ComputationGraph = ComputationGraph(self.Cost)
self.Model = Model(y_hat)
def main(num_epochs=100):
x = tensor.matrix('features')
m = tensor.matrix('features_mask')
y = tensor.imatrix('targets')
x_int = x.astype(dtype='int32').T
train_dataset = IMDB()
idx_sort = numpy.argsort(
[len(s) for s in
train_dataset.indexables[
train_dataset.sources.index('features')]]
)
n_voc = len(train_dataset.dict.keys())
for idx in xrange(len(train_dataset.sources)):
train_dataset.indexables[idx] = train_dataset.indexables[idx][idx_sort]
n_h = 100
linear_embedding = LookupTable(
length=n_voc,
dim=4 * n_h,
weights_init=Uniform(std=0.01),
biases_init=Constant(0.)
)
linear_embedding.initialize()
lstm_biases = numpy.zeros(4 * n_h).astype(dtype=theano.config.floatX)
lstm_biases[n_h:(2 * n_h)] = 4.
rnn = LSTM(
dim=n_h,
weights_init=Uniform(std=0.01),
biases_init=Constant(0.)
)
rnn.initialize()
score_layer = Linear(
input_dim=n_h,
output_dim=1,
weights_init=Uniform(std=0.01),
biases_init=Constant(0.)
)
score_layer.initialize()
embedding = linear_embedding.apply(x_int) * tensor.shape_padright(m.T)
rnn_out = rnn.apply(embedding)
rnn_out_mean_pooled = rnn_out[0][-1]
probs = Sigmoid().apply(
score_layer.apply(rnn_out_mean_pooled))
cost = - (y * tensor.log(probs)
+ (1 - y) * tensor.log(1 - probs)
).mean()
cost.name = 'cost'
misclassification = (y * (probs < 0.5)
+ (1 - y) * (probs > 0.5)
).mean()
misclassification.name = 'misclassification'
cg = ComputationGraph([cost])
params = cg.parameters
algorithm = GradientDescent(
cost=cost,
params=params,
step_rule=CompositeRule(
components=[StepClipping(threshold=10.),
Adam()
]
)
)
n_train = int(numpy.floor(.8 * train_dataset.num_examples))
n_valid = int(numpy.floor(.1 * train_dataset.num_examples))
train_data_stream = Padding(
data_stream=DataStream(
dataset=train_dataset,
iteration_scheme=BatchwiseShuffledScheme(
examples=range(n_train),
batch_size=10,
)
),
mask_sources=('features',)
)
valid_data_stream = Padding(
data_stream=DataStream(
dataset=train_dataset,
iteration_scheme=BatchwiseShuffledScheme(
examples=range(n_train, n_train + n_valid),
batch_size=10,
)
),
mask_sources=('features',)
)
test_data_stream = Padding(
data_stream=DataStream(
dataset=train_dataset,
iteration_scheme=BatchwiseShuffledScheme(
examples=range(n_train + n_valid,
train_dataset.num_examples),
batch_size=10,
)
),
mask_sources=('features',)
)
model = Model(cost)
extensions = []
extensions.append(Timing())
extensions.append(FinishAfter(after_n_epochs=num_epochs))
extensions.append(DataStreamMonitoring(
[cost, misclassification],
test_data_stream,
prefix='test'))
extensions.append(DataStreamMonitoring(
[cost, misclassification],
valid_data_stream,
prefix='valid'))
extensions.append(TrainingDataMonitoring(
[cost, misclassification],
prefix='train',
after_epoch=True))
plotters = []
plotters.append(Plotter(
channels=[['train_cost', 'train_misclassification',
'valid_cost', 'valid_misclassification']],
titles=['Costs']))
extensions.append(PlotManager('IMDB classification example',
plotters=plotters,
after_epoch=True,
after_training=True))
extensions.append(Printing())
main_loop = MainLoop(model=model,
data_stream=train_data_stream,
algorithm=algorithm,
extensions=extensions)
main_loop.run()
def _embed(self, sample_num, dim, name, *args, **kwargs):
embed = LookupTable(sample_num, dim, name=name)
embed.weights_init = IsotropicGaussian(std=1 / numpy.sqrt(dim))
embed.initialize()
return embed
def main(mode, save_path, num_batches, from_dump):
if mode == "train":
# Experiment configuration
dimension = 100
readout_dimension = len(char2code)
# Data processing pipeline
data_stream = DataStreamMapping(
mapping=lambda data: tuple(array.T for array in data),
data_stream=PaddingDataStream(
BatchDataStream(
iteration_scheme=ConstantScheme(10),
data_stream=DataStreamMapping(
mapping=reverse_words,
add_sources=("targets", ),
data_stream=DataStreamFilter(
predicate=lambda data: len(data[0]) <= 100,
data_stream=OneBillionWord(
"training", [99],
char2code,
level="character",
preprocess=str.lower).get_default_stream())))))
# Build the model
chars = tensor.lmatrix("features")
chars_mask = tensor.matrix("features_mask")
targets = tensor.lmatrix("targets")
targets_mask = tensor.matrix("targets_mask")
encoder = Bidirectional(GatedRecurrent(dim=dimension,
activation=Tanh()),
weights_init=Orthogonal())
encoder.initialize()
fork = Fork([
name
for name in encoder.prototype.apply.sequences if name != 'mask'
],
weights_init=IsotropicGaussian(0.1),
biases_init=Constant(0))
fork.input_dim = dimension
fork.fork_dims = {name: dimension for name in fork.fork_names}
fork.initialize()
lookup = LookupTable(readout_dimension,
dimension,
weights_init=IsotropicGaussian(0.1))
lookup.initialize()
transition = Transition(activation=Tanh(),
dim=dimension,
attended_dim=2 * dimension,
name="transition")
attention = SequenceContentAttention(
state_names=transition.apply.states,
match_dim=dimension,
name="attention")
readout = LinearReadout(readout_dim=readout_dimension,
source_names=["states"],
emitter=SoftmaxEmitter(name="emitter"),
feedbacker=LookupFeedback(
readout_dimension, dimension),
name="readout")
generator = SequenceGenerator(readout=readout,
transition=transition,
attention=attention,
weights_init=IsotropicGaussian(0.1),
biases_init=Constant(0),
name="generator")
generator.push_initialization_config()
transition.weights_init = Orthogonal()
generator.initialize()
bricks = [encoder, fork, lookup, generator]
# Give an idea of what's going on
params = Selector(bricks).get_params()
logger.info("Parameters:\n" +
pprint.pformat([(key, value.get_value().shape)
for key, value in params.items()],
width=120))
# Build the cost computation graph
batch_cost = generator.cost(
targets,
targets_mask,
attended=encoder.apply(**dict_union(fork.apply(
lookup.lookup(chars), return_dict=True),
mask=chars_mask)),
attended_mask=chars_mask).sum()
batch_size = named_copy(chars.shape[1], "batch_size")
cost = aggregation.mean(batch_cost, batch_size)
cost.name = "sequence_log_likelihood"
logger.info("Cost graph is built")
# Fetch variables useful for debugging
max_length = named_copy(chars.shape[0], "max_length")
cost_per_character = named_copy(
aggregation.mean(batch_cost, batch_size * max_length),
"character_log_likelihood")
cg = ComputationGraph(cost)
energies = unpack(VariableFilter(application=readout.readout,
name="output")(cg.variables),
singleton=True)
min_energy = named_copy(energies.min(), "min_energy")
max_energy = named_copy(energies.max(), "max_energy")
(activations, ) = VariableFilter(
application=generator.transition.apply,
name="states")(cg.variables)
mean_activation = named_copy(activations.mean(), "mean_activation")
# Define the training algorithm.
algorithm = GradientDescent(cost=cost,
step_rule=CompositeRule([
GradientClipping(10.0),
SteepestDescent(0.01)
]))
observables = [
cost, min_energy, max_energy, mean_activation, batch_size,
max_length, cost_per_character, algorithm.total_step_norm,
algorithm.total_gradient_norm
]
for name, param in params.items():
observables.append(named_copy(param.norm(2), name + "_norm"))
observables.append(
named_copy(algorithm.gradients[param].norm(2),
name + "_grad_norm"))
main_loop = MainLoop(
model=bricks,
data_stream=data_stream,
algorithm=algorithm,
extensions=([LoadFromDump(from_dump)] if from_dump else []) + [
Timing(),
TrainingDataMonitoring(observables, after_every_batch=True),
TrainingDataMonitoring(
observables, prefix="average", every_n_batches=10),
FinishAfter(after_n_batches=num_batches).add_condition(
"after_batch", lambda log: math.isnan(
log.current_row.total_gradient_norm)),
Plot(os.path.basename(save_path),
[["average_" + cost.name],
["average_" + cost_per_character.name]],
every_n_batches=10),
SerializeMainLoop(save_path,
every_n_batches=500,
save_separately=["model", "log"]),
Printing(every_n_batches=1)
])
main_loop.run()
elif mode == "test":
with open(save_path, "rb") as source:
encoder, fork, lookup, generator = dill.load(source)
logger.info("Model is loaded")
chars = tensor.lmatrix("features")
generated = generator.generate(
n_steps=3 * chars.shape[0],
batch_size=chars.shape[1],
attended=encoder.apply(**dict_union(
fork.apply(lookup.lookup(chars), return_dict=True))),
attended_mask=tensor.ones(chars.shape))
sample_function = ComputationGraph(generated).get_theano_function()
logging.info("Sampling function is compiled")
while True:
# Python 2-3 compatibility
line = input("Enter a sentence\n")
batch_size = int(input("Enter a number of samples\n"))
encoded_input = [
char2code.get(char, char2code["<UNK>"])
for char in line.lower().strip()
]
encoded_input = ([char2code['<S>']] + encoded_input +
[char2code['</S>']])
print("Encoder input:", encoded_input)
target = reverse_words((encoded_input, ))[0]
print("Target: ", target)
states, samples, glimpses, weights, costs = sample_function(
numpy.repeat(numpy.array(encoded_input)[:, None],
batch_size,
axis=1))
messages = []
for i in range(samples.shape[1]):
sample = list(samples[:, i])
try:
true_length = sample.index(char2code['</S>']) + 1
except ValueError:
true_length = len(sample)
sample = sample[:true_length]
cost = costs[:true_length, i].sum()
message = "({})".format(cost)
message += "".join(code2char[code] for code in sample)
if sample == target:
message += " CORRECT!"
messages.append((cost, message))
messages.sort(key=lambda tuple_: -tuple_[0])
for _, message in messages:
print(message)
def __init__(self, config, vocab_size):
question = tensor.imatrix('question')
question_mask = tensor.imatrix('question_mask')
context = tensor.imatrix('context')
context_mask = tensor.imatrix('context_mask')
answer = tensor.imatrix('answer')
answer_mask = tensor.imatrix('answer_mask')
ans_indices = tensor.imatrix('ans_indices') # n_steps * n_samples
ans_indices_mask = tensor.imatrix('ans_indices_mask')
bricks = []
question = question.dimshuffle(1, 0)
question_mask = question_mask.dimshuffle(1, 0)
context = context.dimshuffle(1, 0)
context_mask = context_mask.dimshuffle(1, 0)
answer = answer.dimshuffle(1, 0)
answer_mask = answer_mask.dimshuffle(1, 0)
ans_indices = ans_indices.dimshuffle(1, 0)
ans_indices_mask = ans_indices_mask.dimshuffle(1, 0)
# Embed questions and context
embed = LookupTable(vocab_size, config.embed_size, name='embed')
embed.weights_init = IsotropicGaussian(0.01)
# embed.weights_init = Constant(init_embedding_table(filename='embeddings/vocab_embeddings.txt'))
# one directional LSTM encoding
q_lstm_ins = Linear(input_dim=config.embed_size,
output_dim=4 * config.pre_lstm_size,
name='q_lstm_in')
q_lstm = LSTM(dim=config.pre_lstm_size,
activation=Tanh(),
name='q_lstm')
c_lstm_ins = Linear(input_dim=config.embed_size,
output_dim=4 * config.pre_lstm_size,
name='c_lstm_in')
c_lstm = LSTM(dim=config.pre_lstm_size,
activation=Tanh(),
name='c_lstm')
bricks += [q_lstm, c_lstm, q_lstm_ins, c_lstm_ins]
q_tmp = q_lstm_ins.apply(embed.apply(question))
c_tmp = c_lstm_ins.apply(embed.apply(context))
q_hidden, _ = q_lstm.apply(q_tmp,
mask=question_mask.astype(
theano.config.floatX)) # lq, bs, dim
c_hidden, _ = c_lstm.apply(c_tmp,
mask=context_mask.astype(
theano.config.floatX)) # lc, bs, dim
# Attention mechanism Bilinear question
attention_question = Linear(input_dim=config.pre_lstm_size,
output_dim=config.pre_lstm_size,
name='att_question')
bricks += [attention_question]
att_weights_question = q_hidden[
None, :, :, :] * attention_question.apply(
c_hidden.reshape(
(c_hidden.shape[0] * c_hidden.shape[1],
c_hidden.shape[2]))).reshape(
(c_hidden.shape[0], c_hidden.shape[1],
c_hidden.shape[2]))[:,
None, :, :] # --> lc,lq,bs,dim
att_weights_question = att_weights_question.sum(
axis=3) # sum over axis 3 -> dimensions --> lc,lq,bs
att_weights_question = att_weights_question.dimshuffle(
0, 2, 1) # --> lc,bs,lq
att_weights_question = att_weights_question.reshape(
(att_weights_question.shape[0] * att_weights_question.shape[1],
att_weights_question.shape[2])) # --> lc*bs,lq
att_weights_question = tensor.nnet.softmax(
att_weights_question
) # softmax over axis 1 -> length of question # --> lc*bs,lq
att_weights_question = att_weights_question.reshape(
(c_hidden.shape[0], q_hidden.shape[1],
q_hidden.shape[0])) # --> lc,bs,lq
att_weights_question = att_weights_question.dimshuffle(
0, 2, 1) # --> lc,lq,bs
question_context_attention = att_weights_question.dimshuffle(2, 1, 0)
question_context_attention.name = "question_context_attention"
self.analyse_vars = [question_context_attention]
attended_question = tensor.sum(
q_hidden[None, :, :, :] * att_weights_question[:, :, :, None],
axis=1) # sum over axis 1 -> length of question --> lc,bs,dim
attended_question.name = 'attended_question'
# Match LSTM
cqembed = tensor.concatenate([c_hidden, attended_question], axis=2)
mlstms, mhidden_list = make_bidir_lstm_stack(
cqembed, 2 * config.pre_lstm_size,
context_mask.astype(theano.config.floatX), config.match_lstm_size,
config.match_skip_connections, 'match')
bricks = bricks + mlstms
if config.match_skip_connections:
menc_dim = 2 * sum(config.match_lstm_size)
menc = tensor.concatenate(mhidden_list, axis=2)
else:
menc_dim = 2 * config.match_lstm_size[-1]
menc = tensor.concatenate(mhidden_list[-2:], axis=2)
menc.name = 'menc'
#pointer networks decoder LSTM and Attention parameters
params = init_params(data_dim=config.decoder_data_dim,
lstm_dim=config.decoder_lstm_output_dim)
tparams = init_tparams(params)
self.theano_params = []
add_role(tparams['lstm_de_W'], WEIGHT)
add_role(tparams['lstm_de_U'], WEIGHT)
add_role(tparams['lstm_de_b'], BIAS)
add_role(tparams['ptr_b1'], BIAS)
add_role(tparams['ptr_b2'], BIAS)
add_role(tparams['ptr_v'], WEIGHT)
add_role(tparams['ptr_W1'], WEIGHT)
add_role(tparams['ptr_W2'], WEIGHT)
self.theano_params = tparams.values()
#n_steps = length , n_samples = batch_size
n_steps = ans_indices.shape[0]
n_samples = ans_indices.shape[1]
preds, generations = ptr_network(
tparams, cqembed, context_mask.astype(theano.config.floatX),
ans_indices, ans_indices_mask.astype(theano.config.floatX),
config.decoder_lstm_output_dim, menc)
self.generations = generations
idx_steps = tensor.outer(tensor.arange(n_steps, dtype='int64'),
tensor.ones((n_samples, ), dtype='int64'))
idx_samples = tensor.outer(tensor.ones((n_steps, ), dtype='int64'),
tensor.arange(n_samples, dtype='int64'))
probs = preds[idx_steps, ans_indices, idx_samples]
# probs *= y_mask
off = 1e-8
if probs.dtype == 'float16':
off = 1e-6
# probs += (1 - y_mask) # change unmasked position to 1, since log(1) = 0
probs += off
# probs_printed = theano.printing.Print('this is probs')(probs)
cost = -tensor.log(probs)
cost *= ans_indices_mask
cost = cost.sum(axis=0) / ans_indices_mask.sum(axis=0)
cost = cost.mean()
# Apply dropout
cg = ComputationGraph([cost])
if config.w_noise > 0:
noise_vars = VariableFilter(roles=[WEIGHT])(cg)
cg = apply_noise(cg, noise_vars, config.w_noise)
if config.dropout > 0:
cg = apply_dropout(cg, mhidden_list, config.dropout)
[cost_reg] = cg.outputs
# Other stuff
cost.name = 'cost'
cost_reg.name = 'cost_reg'
# self.predictions.name = 'pred'
self.sgd_cost = cost_reg
self.monitor_vars = [[cost_reg]]
self.monitor_vars_valid = [[cost_reg]]
# self.analyse_vars= [cost, self.predictions, att_weights_start, att_weights_end, att_weights, att_target]
# Initialize bricks
embed.initialize()
for brick in bricks:
brick.weights_init = config.weights_init
brick.biases_init = config.biases_init
brick.initialize()
source_path = 'dataset/normalized_syllables_rhythm_notes.json-seqlen-30.hdf5'
train_dataset = T_H5PYDataset(source_path, which_sets=('train', ))
hidden_layer_dim = 1000
x = tensor.lmatrix('syllables')
y = tensor.lmatrix('durations')
lookup_input = LookupTable(name='lookup_input',
length=train_dataset.syllables_vocab_size() + 1,
dim=hidden_layer_dim,
weights_init=initialization.Uniform(width=0.01),
biases_init=Constant(0))
lookup_input.initialize()
linear_input = Linear(name='linear_input',
input_dim=hidden_layer_dim,
output_dim=hidden_layer_dim,
weights_init=initialization.Uniform(width=0.01),
biases_init=Constant(0))
linear_input.initialize()
rnn = SimpleRecurrent(name='hidden',
dim=hidden_layer_dim,
activation=Tanh(),
weights_init=initialization.Uniform(width=0.01))
rnn.initialize()
linear_output = Linear(name='linear_output',
def main(num_epochs=100):
x = tensor.matrix('features')
m = tensor.matrix('features_mask')
y = tensor.imatrix('targets')
x_int = x.astype(dtype='int32').T - 2
train_dataset = IMDB()
idx_sort = numpy.argsort(
[len(s) for s in
train_dataset.indexables[
train_dataset.sources.index('features')]]
)
n_voc = len(train_dataset.dict.keys())
for idx in xrange(len(train_dataset.sources)):
train_dataset.indexables[idx] = train_dataset.indexables[idx][idx_sort]
n_h = 10
linear_embedding = LookupTable(
length=n_voc,
dim=4 * n_h,
weights_init=Uniform(std=0.01),
biases_init=Constant(0.)
)
linear_embedding.initialize()
lstm_biases = numpy.zeros(4 * n_h).astype(dtype=theano.config.floatX)
lstm_biases[n_h:(2 * n_h)] = 4.
rnn = Bidirectional(LSTM(
dim=n_h,
weights_init=Uniform(std=0.01),
biases_init=Constant(0.)
))
rnn.initialize()
score_layer = Linear(
input_dim=2*n_h,
output_dim=1,
weights_init=Uniform(std=0.01),
biases_init=Constant(0.)
)
score_layer.initialize()
embedding = linear_embedding.apply(x_int) * tensor.shape_padright(m.T)
rnn_out = rnn.apply(embedding)
rnn_out_mean_pooled = tensor.mean(rnn_out[0], axis=0)
probs = Sigmoid().apply(
score_layer.apply(rnn_out_mean_pooled))
cost = - (y * tensor.log(probs)
+ (1 - y) * tensor.log(1 - probs)
).mean()
cost.name = 'cost'
misclassification = (y * (probs < 0.5)
+ (1 - y) * (probs > 0.5)
).mean()
misclassification.name = 'misclassification'
cg = ComputationGraph([cost])
params = cg.parameters
algorithm = GradientDescent(
cost=cost,
params=params,
step_rule=CompositeRule(
components=[StepClipping(threshold=10.),
Adam()
]
)
)
n_train = int(numpy.floor(.8 * train_dataset.num_examples))
n_valid = int(numpy.floor(.1 * train_dataset.num_examples))
train_data_stream = Padding(
data_stream=DataStream(
dataset=train_dataset,
iteration_scheme=BatchwiseShuffledScheme(
examples=range(100),
batch_size=10,
)
),
mask_sources=('features',)
)
valid_data_stream = Padding(
data_stream=DataStream(
dataset=train_dataset,
iteration_scheme=BatchwiseShuffledScheme(
examples=range(100, 110),
batch_size=10,
)
),
mask_sources=('features',)
)
test_data_stream = Padding(
data_stream=DataStream(
dataset=train_dataset,
iteration_scheme=BatchwiseShuffledScheme(
examples=range(110, 120),
batch_size=10,
)
),
mask_sources=('features',)
)
model = Model(cost)
extensions = []
extensions.append(Timing())
extensions.append(FinishAfter(after_n_epochs=num_epochs))
extensions.append(DataStreamMonitoring(
[cost, misclassification],
test_data_stream,
prefix='test'))
extensions.append(DataStreamMonitoring(
[cost, misclassification],
valid_data_stream,
prefix='valid'))
extensions.append(TrainingDataMonitoring(
[cost, misclassification],
prefix='train',
after_epoch=True))
plotters = []
plotters.append(Plotter(
channels=[['train_cost', 'train_misclassification',
'valid_cost', 'valid_misclassification']],
titles=['Costs']))
extensions.append(PlotManager('IMDB classification example',
plotters=plotters,
after_epoch=True,
after_training=True))
extensions.append(Printing())
main_loop = MainLoop(model=model,
data_stream=train_data_stream,
algorithm=algorithm,
extensions=extensions)
main_loop.run()
def main(num_epochs=100):
x = tensor.matrix('features')
m = tensor.matrix('features_mask')
x_int = x.astype(dtype='int32').T
train_dataset = TextFile('inspirational.txt')
train_dataset.indexables[0] = numpy.array(sorted(
train_dataset.indexables[0], key=len
))
n_voc = len(train_dataset.dict.keys())
init_probs = numpy.array(
[sum(filter(lambda idx:idx == w,
[s[0] for s in train_dataset.indexables[
train_dataset.sources.index('features')]]
)) for w in xrange(n_voc)],
dtype=theano.config.floatX
)
init_probs = init_probs / init_probs.sum()
n_h = 100
linear_embedding = LookupTable(
length=n_voc,
dim=n_h,
weights_init=Uniform(std=0.01),
biases_init=Constant(0.)
)
linear_embedding.initialize()
lstm_biases = numpy.zeros(4 * n_h).astype(dtype=theano.config.floatX)
lstm_biases[n_h:(2 * n_h)] = 4.
rnn = SimpleRecurrent(
dim=n_h,
activation=Tanh(),
weights_init=Uniform(std=0.01),
biases_init=Constant(0.)
)
rnn.initialize()
score_layer = Linear(
input_dim=n_h,
output_dim=n_voc,
weights_init=Uniform(std=0.01),
biases_init=Constant(0.)
)
score_layer.initialize()
embedding = (linear_embedding.apply(x_int[:-1])
* tensor.shape_padright(m.T[1:]))
rnn_out = rnn.apply(inputs=embedding, mask=m.T[1:])
probs = softmax(
sequence_map(score_layer.apply, rnn_out, mask=m.T[1:])[0]
)
idx_mask = m.T[1:].nonzero()
cost = CategoricalCrossEntropy().apply(
x_int[1:][idx_mask[0], idx_mask[1]],
probs[idx_mask[0], idx_mask[1]]
)
cost.name = 'cost'
misclassification = MisclassificationRate().apply(
x_int[1:][idx_mask[0], idx_mask[1]],
probs[idx_mask[0], idx_mask[1]]
)
misclassification.name = 'misclassification'
cg = ComputationGraph([cost])
params = cg.parameters
algorithm = GradientDescent(
cost=cost,
params=params,
step_rule=Adam()
)
train_data_stream = Padding(
data_stream=DataStream(
dataset=train_dataset,
iteration_scheme=BatchwiseShuffledScheme(
examples=train_dataset.num_examples,
batch_size=10,
)
),
mask_sources=('features',)
)
model = Model(cost)
extensions = []
extensions.append(Timing())
extensions.append(FinishAfter(after_n_epochs=num_epochs))
extensions.append(TrainingDataMonitoring(
[cost, misclassification],
prefix='train',
after_epoch=True))
batch_size = 10
length = 30
trng = MRG_RandomStreams(18032015)
u = trng.uniform(size=(length, batch_size, n_voc))
gumbel_noise = -tensor.log(-tensor.log(u))
init_samples = (tensor.log(init_probs).dimshuffle(('x', 0))
+ gumbel_noise[0]).argmax(axis=-1)
init_states = rnn.initial_state('states', batch_size)
def sampling_step(g_noise, states, samples_step):
embedding_step = linear_embedding.apply(samples_step)
next_states = rnn.apply(inputs=embedding_step,
states=states,
iterate=False)
probs_step = softmax(score_layer.apply(next_states))
next_samples = (tensor.log(probs_step)
+ g_noise).argmax(axis=-1)
return next_states, next_samples
[_, samples], _ = theano.scan(
fn=sampling_step,
sequences=[gumbel_noise[1:]],
outputs_info=[init_states, init_samples]
)
sampling = theano.function([], samples.owner.inputs[0].T)
plotters = []
plotters.append(Plotter(
channels=[['train_cost', 'train_misclassification']],
titles=['Costs']))
extensions.append(PlotManager('Language modelling example',
plotters=plotters,
after_epoch=True,
after_training=True))
extensions.append(Printing())
extensions.append(PrintSamples(sampler=sampling,
voc=train_dataset.inv_dict))
main_loop = MainLoop(model=model,
data_stream=train_data_stream,
algorithm=algorithm,
extensions=extensions)
main_loop.run()
train_dataset = MyDataset(source_path, which_sets=('train',))
hidden_layer_dim = 1000
x = tensor.lmatrix('x')
y = tensor.lmatrix('y')
lookup_input = LookupTable(
name='lookup_input',
length=charset_size+1,
dim=hidden_layer_dim,
weights_init=initialization.Uniform(width=0.01),
biases_init=Constant(0))
lookup_input.initialize()
linear_input = Linear(
name='linear_input',
input_dim=hidden_layer_dim,
output_dim=hidden_layer_dim,
weights_init=initialization.Uniform(width=0.01),
biases_init=Constant(0))
linear_input.initialize()
rnn = SimpleRecurrent(
name='hidden',
dim=hidden_layer_dim,
activation=Tanh(),
weights_init=initialization.Uniform(width=0.01))
rnn.initialize()
def create_model(self, symbols_num = 500):
hidden_states = self.args.encoder_hidden_dims
embedding_dims = self.args.source_embeddings_dim
# dimensions of sequence embeddings that are created bz bidir net, so the dimensionality is two times dim of a single net
thought_dim = hidden_states * 2
#query_dims = self.args.recurrent_stack_depth * self.args.encoder_hidden_dims
# batch X input symbols
context = tt.lmatrix('context')
context_mask = tt.matrix('context_mask')
context_mask = decorate(context_mask, "context_mask",level=1)
# batch X output symbols
x = tt.lmatrix('question')
x_mask = tt.matrix('question_mask')
# answer ix for each example in the batch
y = tt.lmatrix('answer')
# candidate answer words for each example, batch X candidate words (10 per each example)
candidates_bi = tt.lmatrix("candidates")
candidates_bi_mask = tt.matrix("candidates_mask")
# TODO y can contain long sequences, here we use just the first symbol of each answer (that is possibly longer)
# this have to be adjusted when response can be a sequence and not only a symbol
y = decorate(y, "output")
y = y[:,0]
###################
# create model parts
###################
lookup = LookupTable(symbols_num, embedding_dims, weights_init=Uniform(width=0.2))
context_encoder = self.create_bidi_encoder("context_encoder", embedding_dims, hidden_states)
question_encoder = self.create_bidi_encoder("question_encoder", embedding_dims, hidden_states)
# inits
lookup.initialize()
#rnn.initialize()
###################
# wire the model together
###################
context = decorate(context, "CONTEXT",1)
context_embedding_tbf = lookup.apply(context.T)
#memory_encoded_btf = rnn.apply(context_embedding_tbf[:,0,:])[1] # use cells
memory_encoded_btf = context_encoder.apply(context_embedding_tbf.T,context_mask).dimshuffle(1,0,2)
memory_encoded_btf.name = "memory_encoded_btf"
memory_encoded_btf = decorate(memory_encoded_btf,"MEM ENC")
# batch X features
x = decorate(x,"X")
x_embedded_btf = lookup.apply(x.T)
x_embedded_btf = decorate(x_embedded_btf,"QUESTION EMB")
x_encoded_btf = question_encoder.apply(x_embedded_btf.T, x_mask).dimshuffle(1,0,2)
x_last = x_encoded_btf[-1]
# extract forward rnn that is the first in bidir encoder
x_encoded_btf = decorate(x_encoded_btf,"QUESTION ENC")
x_forward_encoded_bf = x_encoded_btf[:,-1,0:hidden_states]
x_backward_encoded_bf = x_encoded_btf[:,0,hidden_states:hidden_states*2]
query_representation_bf = tt.concatenate([x_forward_encoded_bf,x_backward_encoded_bf],axis=1)
# bidirectional representation of question is used as the search key
search_key = query_representation_bf
#search_key = x_last
#search_key = W_um.apply(x_encoded)
search_key = decorate(search_key,"SEARCH KEY")
mem_attention_pre = tt.batched_dot(search_key, memory_encoded_btf.dimshuffle(0,2,1))
mem_attention_pre = decorate(mem_attention_pre,"ATT presoftmax")
# use masking on attention, this might be unnecessary but we do it just to be sure
mem_attention_pre_masked_bt = tt.mul(mem_attention_pre,context_mask)
mem_attention_pre_masked_bt = decorate(mem_attention_pre_masked_bt,"ATT presoftmax masked")
#mem_attention_bt = Softmax(name="memory_query_softmax").apply(mem_attention_pre_masked_bt)
mem_attention_bt = SoftmaxWithMask(name="memory_query_softmax").apply(mem_attention_pre_masked_bt,context_mask)
mem_attention_bt = decorate(mem_attention_bt,"ATT",level=2)
# compute weighted attention over original word vectors
att_weighted_responses_bf = theano.tensor.batched_dot(mem_attention_bt, context_embedding_tbf.dimshuffle(1,0,2))
#use mask to remove the probability mass from the unmasked candidates
#word_probs_bi = word_probs_bi * candidates_bi_mask
# compare desired response to all candidate responses
# select relevant candidate answer words
candidates_embeddings_bfi = lookup.apply(candidates_bi).dimshuffle(0,2,1)
# convert it to output symbol probabilities
y_hat_presoft = tt.batched_dot(att_weighted_responses_bf, candidates_embeddings_bfi)
y_hat = SoftmaxWithMask(name="output_softmax").apply(y_hat_presoft,candidates_bi_mask)
y_hat.name = "y_hat"
y_hat = decorate(y_hat,"y_hat",level=2)
# the correct answer is always the first among the candidates, so we can use zeros as index of ground truth
y = y.zeros_like()
# cost associated with prediction error
cost_prediction = CategoricalCrossEntropy().apply(y.flatten(), y_hat)
cost_prediction.name = "cost_prediction"
cost = cost_prediction
attention_cost_weight = None
cost_attention = None
cost.name = "cost"
predicted_response_index = tt.argmax(y_hat,axis=1)
accuracy = tt.eq(y,predicted_response_index).mean()
accuracy.name = "accuracy"
return cost, accuracy, mem_attention_bt, y_hat, attention_cost_weight, cost_prediction, cost_attention, context, candidates_bi, candidates_bi_mask, y, context_mask, x, x_mask
def __init__(self, config, vocab_size):
question = tensor.imatrix('question')
question_mask = tensor.imatrix('question_mask')
answer = tensor.imatrix('answer')
answer_mask = tensor.imatrix('answer_mask')
better = tensor.imatrix('better')
better_mask = tensor.imatrix('better_mask')
worse = tensor.imatrix('worse')
worse_mask = tensor.imatrix('worse_mask')
b_left = tensor.imatrix('b_left')
b_left_mask = tensor.imatrix('b_left_mask')
b_right = tensor.imatrix('b_right')
b_right_mask = tensor.imatrix('b_right_mask')
w_left = tensor.imatrix('w_left')
w_left_mask = tensor.imatrix('w_left_mask')
w_right = tensor.imatrix('w_right')
w_right_mask = tensor.imatrix('w_right_mask')
bricks = []
question = question.dimshuffle(1, 0)
question_mask = question_mask.dimshuffle(1, 0)
better = better.dimshuffle(1, 0)
better_mask = better_mask.dimshuffle(1, 0)
worse = worse.dimshuffle(1, 0)
worse_mask = worse_mask.dimshuffle(1, 0)
b_left = b_left.dimshuffle(1, 0)
b_left_mask = b_left_mask.dimshuffle(1, 0)
b_right = b_right.dimshuffle(1, 0)
b_right_mask = b_right_mask.dimshuffle(1, 0)
w_left = w_left.dimshuffle(1, 0)
w_left_mask = w_left_mask.dimshuffle(1, 0)
w_right = w_right.dimshuffle(1, 0)
w_right_mask = w_right_mask.dimshuffle(1, 0)
answer = answer.dimshuffle(1, 0)
answer_mask = answer_mask.dimshuffle(1, 0)
# Embed questions and context
embed = LookupTable(vocab_size, config.embed_size, name='question_embed')
embed.weights_init = IsotropicGaussian(0.01)
# Calculate question encoding (concatenate layer1)
qembed = embed.apply(question)
qlstms, qhidden_list = make_bidir_lstm_stack(qembed, config.embed_size, question_mask.astype(theano.config.floatX),
config.question_lstm_size, config.question_skip_connections, 'q')
bricks = bricks + qlstms
if config.question_skip_connections:
qenc_dim = 2*sum(config.question_lstm_size)
qenc = tensor.concatenate([h[-1,:,:] for h in qhidden_list], axis=1)
else:
qenc_dim = 2*config.question_lstm_size[-1]
qenc = tensor.concatenate([h[-1,:,:] for h in qhidden_list[-2:]], axis=1)
qenc.name = 'qenc'
# candidate encoders
candidates_hidden_list = []
candidate_fwd_lstm_ins = Linear(input_dim=config.embed_size, output_dim=4*config.ctx_lstm_size[0], name='candidate_fwd_lstm_in_0_0')
candidate_fwd_lstm = LSTM(dim=config.ctx_lstm_size[0], activation=Tanh(), name='candidate_fwd_lstm_0')
candidate_bwd_lstm_ins = Linear(input_dim=config.embed_size, output_dim=4*config.ctx_lstm_size[0], name='candidate_bwd_lstm_in_0_0')
candidate_bwd_lstm = LSTM(dim=config.ctx_lstm_size[0], activation=Tanh(), name='candidate_bwd_lstm_0')
#adding encoding bricks for initialization
bricks = bricks + [candidate_fwd_lstm, candidate_bwd_lstm, candidate_fwd_lstm_ins, candidate_bwd_lstm_ins]
#computing better encoding
better_embed = embed.apply(better)
better_fwd_tmp = candidate_fwd_lstm_ins.apply(better_embed)
better_bwd_tmp = candidate_bwd_lstm_ins.apply(better_embed)
better_fwd_hidden, _ = candidate_fwd_lstm.apply(better_fwd_tmp, mask=better_mask.astype(theano.config.floatX))
better_bwd_hidden, _ = candidate_bwd_lstm.apply(better_bwd_tmp[::-1], mask=better_mask.astype(theano.config.floatX)[::-1])
better_hidden_list = [better_fwd_hidden, better_bwd_hidden]
better_enc_dim = 2*sum(config.ctx_lstm_size)
better_enc = tensor.concatenate([h[-1,:,:] for h in better_hidden_list], axis=1) #concating last state of fwd and bwd LSTMs 2*dim * batch_size
better_enc.name = 'better_enc'
candidates_hidden_list = candidates_hidden_list + [better_fwd_hidden, better_bwd_hidden]
#computing worse encoding
worse_embed = embed.apply(worse)
worse_fwd_tmp = candidate_fwd_lstm_ins.apply(worse_embed)
worse_bwd_tmp = candidate_bwd_lstm_ins.apply(worse_embed)
worse_fwd_hidden, _ = candidate_fwd_lstm.apply(worse_fwd_tmp, mask=worse_mask.astype(theano.config.floatX))
worse_bwd_hidden, _ = candidate_bwd_lstm.apply(worse_bwd_tmp[::-1], mask=worse_mask.astype(theano.config.floatX)[::-1])
worse_hidden_list = [worse_fwd_hidden, worse_bwd_hidden]
worse_enc_dim = 2*sum(config.ctx_lstm_size)
worse_enc = tensor.concatenate([h[-1,:,:] for h in worse_hidden_list], axis=1)
worse_enc.name = 'worse_enc'
candidates_hidden_list = candidates_hidden_list + [worse_fwd_hidden, worse_bwd_hidden]
#left encoders
left_context_hidden_list = []
left_context_fwd_lstm_ins = Linear(input_dim=config.embed_size, output_dim=4*config.ctx_lstm_size[0], name='left_context_fwd_lstm_in_0_0')
left_context_fwd_lstm = LSTM(dim=config.ctx_lstm_size[0], activation=Tanh(), name='left_context_fwd_lstm_0')
left_context_bwd_lstm_ins = Linear(input_dim=config.embed_size, output_dim=4*config.ctx_lstm_size[0], name='left_context_bwd_lstm_in_0_0')
left_context_bwd_lstm = LSTM(dim=config.ctx_lstm_size[0], activation=Tanh(), name='left_context_bwd_lstm_0')
#adding encoding bricks for initialization
bricks = bricks + [left_context_fwd_lstm, left_context_bwd_lstm, left_context_fwd_lstm_ins, left_context_bwd_lstm_ins]
#right encoders
right_context_hidden_list = []
right_context_fwd_lstm_ins = Linear(input_dim=config.embed_size, output_dim=4*config.ctx_lstm_size[0], name='right_context_fwd_lstm_in_0_0')
right_context_fwd_lstm = LSTM(dim=config.ctx_lstm_size[0], activation=Tanh(), name='right_context_fwd_lstm_0')
right_context_bwd_lstm_ins = Linear(input_dim=config.embed_size, output_dim=4*config.ctx_lstm_size[0], name='right_context_bwd_lstm_in_0_0')
right_context_bwd_lstm = LSTM(dim=config.ctx_lstm_size[0], activation=Tanh(), name='right_context_bwd_lstm_0')
#adding encoding bricks for initialization
bricks = bricks + [right_context_fwd_lstm, right_context_bwd_lstm, right_context_fwd_lstm_ins, right_context_bwd_lstm_ins]
#left half encodings
better_left_embed = embed.apply(b_left)
better_left_fwd_tmp = left_context_fwd_lstm_ins.apply(better_left_embed)
better_left_bwd_tmp = left_context_bwd_lstm_ins.apply(better_left_embed)
better_left_fwd_hidden, _ = left_context_fwd_lstm.apply(better_left_fwd_tmp, mask=b_left_mask.astype(theano.config.floatX))
better_left_bwd_hidden, _ = left_context_bwd_lstm.apply(better_left_bwd_tmp[::-1], mask=b_left_mask.astype(theano.config.floatX)[::-1])
better_left_hidden_list = [better_left_fwd_hidden, better_left_bwd_hidden]
better_left_enc_dim = 2*sum(config.ctx_lstm_size)
better_left_enc = tensor.concatenate([h[-1,:,:] for h in better_left_hidden_list], axis=1) #concating last state of fwd and bwd LSTMs 2*dim * batch_size
better_left_enc.name = 'better_left_enc'
left_context_hidden_list = left_context_hidden_list + [better_left_fwd_hidden, better_left_bwd_hidden]
worse_left_embed = embed.apply(w_left)
worse_left_fwd_tmp = left_context_fwd_lstm_ins.apply(worse_left_embed)
worse_left_bwd_tmp = left_context_bwd_lstm_ins.apply(worse_left_embed)
worse_left_fwd_hidden, _ = left_context_fwd_lstm.apply(worse_left_fwd_tmp, mask=w_left_mask.astype(theano.config.floatX))
worse_left_bwd_hidden, _ = left_context_bwd_lstm.apply(worse_left_bwd_tmp[::-1], mask=w_left_mask.astype(theano.config.floatX)[::-1])
worse_left_hidden_list = [worse_left_fwd_hidden, worse_left_bwd_hidden]
worse_left_enc_dim = 2*sum(config.ctx_lstm_size)
worse_left_enc = tensor.concatenate([h[-1,:,:] for h in worse_left_hidden_list], axis=1) #concating last state of fwd and bwd LSTMs 2*dim * batch_size
worse_left_enc.name = 'worse_left_enc'
left_context_hidden_list = left_context_hidden_list + [worse_left_fwd_hidden, worse_left_bwd_hidden]
#right half encoding
better_right_embed = embed.apply(b_right)
better_right_fwd_tmp = right_context_fwd_lstm_ins.apply(better_right_embed)
better_right_bwd_tmp = right_context_bwd_lstm_ins.apply(better_right_embed)
better_right_fwd_hidden, _ = right_context_fwd_lstm.apply(better_right_fwd_tmp, mask=b_right_mask.astype(theano.config.floatX))
better_right_bwd_hidden, _ = right_context_bwd_lstm.apply(better_right_bwd_tmp[::-1], mask=b_right_mask.astype(theano.config.floatX)[::-1])
better_right_hidden_list = [better_right_fwd_hidden, better_right_bwd_hidden]
better_right_enc_dim = 2*sum(config.ctx_lstm_size)
better_right_enc = tensor.concatenate([h[-1,:,:] for h in better_right_hidden_list], axis=1) #concating last state of fwd and bwd LSTMs 2*dim * batch_size
better_right_enc.name = 'better_right_enc'
right_context_hidden_list = right_context_hidden_list + [better_right_fwd_hidden, better_right_bwd_hidden]
worse_right_embed = embed.apply(w_right)
worse_right_fwd_tmp = right_context_fwd_lstm_ins.apply(worse_right_embed)
worse_right_bwd_tmp = right_context_bwd_lstm_ins.apply(worse_right_embed)
worse_right_fwd_hidden, _ = right_context_fwd_lstm.apply(worse_right_fwd_tmp, mask=w_right_mask.astype(theano.config.floatX))
worse_right_bwd_hidden, _ = right_context_bwd_lstm.apply(worse_right_bwd_tmp[::-1], mask=w_right_mask.astype(theano.config.floatX)[::-1])
worse_right_hidden_list = [worse_right_fwd_hidden, worse_right_bwd_hidden]
worse_right_enc_dim = 2*sum(config.ctx_lstm_size)
worse_right_enc = tensor.concatenate([h[-1,:,:] for h in worse_right_hidden_list], axis=1) #concating last state of fwd and bwd LSTMs 2*dim * batch_size
worse_right_enc.name = 'worse_right_enc'
right_context_hidden_list = right_context_hidden_list + [worse_right_fwd_hidden, worse_right_bwd_hidden]
# F1 prediction MLP
prediction_mlp = MLP(dims=config.prediction_mlp_hidden + [1],
activations=config.prediction_mlp_activations[1:] + [Identity()],
name='prediction_mlp')
prediction_qlinear = Linear(input_dim=qenc_dim, output_dim=config.prediction_mlp_hidden[0]/4.0, name='preq')
prediction_cand_linear = Linear(input_dim=worse_enc_dim, output_dim=config.prediction_mlp_hidden[0]/4.0, use_bias=False, name='precand')
prediction_left_half_linear = Linear(input_dim=better_left_enc_dim, output_dim=config.prediction_mlp_hidden[0]/4.0, use_bias=False, name='preleft')
prediction_right_half_linear = Linear(input_dim=better_right_enc_dim, output_dim=config.prediction_mlp_hidden[0]/4.0, use_bias=False, name='preright')
bricks += [prediction_mlp, prediction_qlinear, prediction_cand_linear, prediction_left_half_linear, prediction_right_half_linear]
better_layer1 = Tanh('tan1').apply(tensor.concatenate([prediction_cand_linear.apply(better_enc), prediction_qlinear.apply(qenc), prediction_left_half_linear.apply(better_left_enc), prediction_right_half_linear.apply(better_right_enc)],axis=1))
better_layer1.name = 'better_layer1'
worse_layer1 = Tanh('tan2').apply(tensor.concatenate([prediction_cand_linear.apply(worse_enc), prediction_qlinear.apply(qenc), prediction_left_half_linear.apply(worse_left_enc), prediction_right_half_linear.apply(worse_right_enc)],axis=1))
worse_layer1.name = 'worse_layer1'
better_pred_weights = Tanh('rec1').apply(prediction_mlp.apply(better_layer1)) #batch_size
worse_pred_weights = Tanh('rec2').apply(prediction_mlp.apply(worse_layer1)) #batch_size
# numpy.set_printoptions(edgeitems=500)
# better_pred_weights = theano.printing.Print('better')(better_pred_weights)
# worse_pred_weights = theano.printing.Print('better')(worse_pred_weights)
# #cost : max(0,- score-better + score-worse + margin)
margin = config.margin
conditions = tensor.lt(better_pred_weights, worse_pred_weights + margin).astype(theano.config.floatX)
self.predictions = conditions
cost = (-better_pred_weights + worse_pred_weights + margin) * conditions
cost = cost.mean()
# Apply dropout
cg = ComputationGraph([cost])
if config.w_noise > 0:
noise_vars = VariableFilter(roles=[WEIGHT])(cg)
cg = apply_noise(cg, noise_vars, config.w_noise)
if config.dropout > 0:
cg = apply_dropout(cg, qhidden_list + candidates_hidden_list, config.dropout)
[cost_reg] = cg.outputs
# Other stuff
cost.name = 'cost'
cost_reg.name = 'cost_reg'
self.sgd_cost = cost_reg
self.monitor_vars = [[cost_reg]]
self.monitor_vars_valid = [[cost_reg]]
# Initialize bricks
embed.initialize()
for brick in bricks:
brick.weights_init = config.weights_init
brick.biases_init = config.biases_init
brick.initialize()
VOCAB_DIM = dataset.vocabulary_size
print "vocab size:", VOCAB_DIM
EMBEDDING_DIM = 100
Xs = tensor.imatrix("context")
y = tensor.ivector('center')
w1 = LookupTable(name="w1", length=VOCAB_DIM, dim=EMBEDDING_DIM)
w2 = Linear(name='w2', input_dim=EMBEDDING_DIM, output_dim=VOCAB_DIM)
hidden = tensor.mean(w1.apply(Xs), axis=1)
y_hat = Softmax().apply(w2.apply(hidden))
w1.weights_init = w2.weights_init = IsotropicGaussian(0.01)
w1.biases_init = w2.biases_init = Constant(0)
w1.initialize()
w2.initialize()
cost = CategoricalCrossEntropy().apply(y, y_hat)
cg = ComputationGraph(cost)
W1, W2 = VariableFilter(roles=[WEIGHT])(cg.variables)
cost = cost + 0.005 * (W1**2).sum() + 0.005 * (W2**2).sum()
cost.name = "loss"
#
# the actual training of the model
#
main = MainLoop(
data_stream=DataStream.default_stream(dataset,
def __init__(self, config, vocab_size):
question = tensor.imatrix('question')
question_mask = tensor.imatrix('question_mask')
context = tensor.imatrix('context')
context_mask = tensor.imatrix('context_mask')
answer = tensor.imatrix('answer')
answer_mask = tensor.imatrix('answer_mask')
ans_indices = tensor.imatrix('ans_indices') # n_steps * n_samples
ans_indices_mask = tensor.imatrix('ans_indices_mask')
context_bag = tensor.eq(context[:, :, None],
tensor.arange(vocab_size)).sum(axis=1).clip(
0, 1)
bricks = []
question = question.dimshuffle(1, 0)
question_mask = question_mask.dimshuffle(1, 0)
context = context.dimshuffle(1, 0)
context_mask = context_mask.dimshuffle(1, 0)
answer = answer.dimshuffle(1, 0)
answer_mask = answer_mask.dimshuffle(1, 0)
ans_indices = ans_indices.dimshuffle(1, 0)
ans_indices_mask = ans_indices_mask.dimshuffle(1, 0)
# Embed questions and context
embed = LookupTable(vocab_size,
config.embed_size,
name='question_embed')
embed.weights_init = IsotropicGaussian(0.01)
# embeddings_initial_value = init_embedding_table(filename='embeddings/vocab_embeddings.txt')
# embed.weights_init = Constant(embeddings_initial_value)
# Calculate question encoding (concatenate layer1)
qembed = embed.apply(question)
qlstms, qhidden_list = make_bidir_lstm_stack(
qembed, config.embed_size,
question_mask.astype(theano.config.floatX),
config.question_lstm_size, config.question_skip_connections, 'q')
bricks = bricks + qlstms
if config.question_skip_connections:
qenc_dim = 2 * sum(config.question_lstm_size)
qenc = tensor.concatenate([h[-1, :, :] for h in qhidden_list],
axis=1)
else:
qenc_dim = 2 * config.question_lstm_size[-1]
qenc = tensor.concatenate([h[-1, :, :] for h in qhidden_list[-2:]],
axis=1)
qenc.name = 'qenc'
#embed size: 200, lstm_size = 256
#qenc: length * batch_size * (2*lstm_size)
# Calculate context encoding (concatenate layer1)
cembed = embed.apply(context)
cqembed = tensor.concatenate(
[
cembed,
tensor.extra_ops.repeat(
qenc[None, :, :], cembed.shape[0], axis=0)
],
axis=2
) #length * batch_size * (embed+2*lstm_size) this is what goes into encoder
clstms, chidden_list = make_bidir_lstm_stack(
cqembed, config.embed_size + qenc_dim,
context_mask.astype(theano.config.floatX), config.ctx_lstm_size,
config.ctx_skip_connections, 'ctx')
bricks = bricks + clstms
if config.ctx_skip_connections:
cenc_dim = 2 * sum(config.ctx_lstm_size) #2 : fw & bw
cenc = tensor.concatenate(chidden_list, axis=2)
else:
cenc_dim = 2 * config.question_lstm_size[-1]
cenc = tensor.concatenate(chidden_list[-2:], axis=2)
cenc.name = 'cenc'
#cenc: length * batch_size * (2*lstm_size)
#pointer networks decoder LSTM and Attention parameters
params = init_params(data_dim=config.decoder_data_dim,
lstm_dim=config.decoder_lstm_output_dim)
tparams = init_tparams(params)
self.theano_params = []
add_role(tparams['lstm_de_W'], WEIGHT)
add_role(tparams['lstm_de_U'], WEIGHT)
add_role(tparams['lstm_de_b'], BIAS)
add_role(tparams['ptr_v'], WEIGHT)
add_role(tparams['ptr_W1'], WEIGHT)
add_role(tparams['ptr_W2'], WEIGHT)
self.theano_params = tparams.values()
# for p in tparams.values():
# add_role(p, WEIGHT)
# self.theano_params.append(p)
#n_steps = length , n_samples = batch_size
n_steps = ans_indices.shape[0]
n_samples = ans_indices.shape[1]
preds, generations = ptr_network(
tparams, cqembed, context_mask.astype(theano.config.floatX),
ans_indices, ans_indices_mask.astype(theano.config.floatX),
config.decoder_lstm_output_dim, cenc)
self.generations = generations
idx_steps = tensor.outer(tensor.arange(n_steps, dtype='int64'),
tensor.ones((n_samples, ), dtype='int64'))
idx_samples = tensor.outer(tensor.ones((n_steps, ), dtype='int64'),
tensor.arange(n_samples, dtype='int64'))
probs = preds[idx_steps, ans_indices, idx_samples]
# probs *= y_mask
off = 1e-8
if probs.dtype == 'float16':
off = 1e-6
# probs += (1 - y_mask) # change unmasked position to 1, since log(1) = 0
probs += off
# probs_printed = theano.printing.Print('this is probs')(probs)
cost = -tensor.log(probs)
cost *= ans_indices_mask
cost = cost.sum(axis=0) / ans_indices_mask.sum(axis=0)
cost = cost.mean()
# Apply dropout
cg = ComputationGraph([cost])
if config.w_noise > 0:
noise_vars = VariableFilter(roles=[WEIGHT])(cg)
cg = apply_noise(cg, noise_vars, config.w_noise)
if config.dropout > 0:
cg = apply_dropout(cg, qhidden_list + chidden_list, config.dropout)
[cost_reg] = cg.outputs
# Other stuff
cost.name = 'cost'
cost_reg.name = 'cost_reg'
self.sgd_cost = cost_reg
self.monitor_vars = [[cost_reg]]
self.monitor_vars_valid = [[cost_reg]]
# Initialize bricks
embed.initialize()
for brick in bricks:
brick.weights_init = config.weights_init
brick.biases_init = config.biases_init
brick.initialize()
[attention.take_glimpses.outputs[0]],
emitter=emitter,
name="readout")
generator = SequenceGenerator(readout=readout,
attention=attention,
transition=transition,
name="generator")
generator.weights_init = IsotropicGaussian(0.01)
generator.biases_init = Constant(0.001)
generator.push_initialization_config()
lookup.weights_init = IsotropicGaussian(0.01)
lookup.biases_init = Constant(0.001)
lookup.initialize()
#generator.transition.weights_init = initialization.Identity(0.98)
#generator.transition.biases_init = IsotropicGaussian(0.01,0.9)
generator.transition.push_initialization_config()
generator.initialize()
cost_matrix = generator.cost_matrix(x,
x_mask,
attended=embed,
attended_mask=context_mask)
cost = cost_matrix.sum(axis=0).mean()
cost.name = "nll"
cg = ComputationGraph(cost)
model = Model(cost)
def __init__(self, config, vocab_size):
question = tensor.imatrix('question')
question_mask = tensor.imatrix('question_mask')
context = tensor.imatrix('context')
context_mask = tensor.imatrix('context_mask')
answer = tensor.imatrix('answer')
answer_mask = tensor.imatrix('answer_mask')
bricks = []
question = question.dimshuffle(1, 0)
question_mask = question_mask.dimshuffle(1, 0)
context = context.dimshuffle(1, 0)
context_mask = context_mask.dimshuffle(1, 0)
answer = answer.dimshuffle(1, 0)
answer_mask = answer_mask.dimshuffle(1, 0)
# Embed questions and context
embed = LookupTable(vocab_size,
config.embed_size,
name='question_embed')
embed.weights_init = IsotropicGaussian(0.01)
# Calculate question encoding (concatenate layer1)
qembed = embed.apply(question)
qlstms, qhidden_list = make_bidir_lstm_stack(
qembed, config.embed_size,
question_mask.astype(theano.config.floatX),
config.question_lstm_size, config.question_skip_connections, 'q')
bricks = bricks + qlstms
if config.question_skip_connections:
qenc_dim = 2 * sum(config.question_lstm_size)
qenc = tensor.concatenate([h[-1, :, :] for h in qhidden_list],
axis=1)
else:
qenc_dim = 2 * config.question_lstm_size[-1]
qenc = tensor.concatenate([h[-1, :, :] for h in qhidden_list[-2:]],
axis=1)
qenc.name = 'qenc'
# Calculate context encoding (concatenate layer1)
cembed = embed.apply(context)
clstms, chidden_list = make_bidir_lstm_stack(
cembed, config.embed_size,
context_mask.astype(theano.config.floatX), config.ctx_lstm_size,
config.ctx_skip_connections, 'ctx')
bricks = bricks + clstms
if config.ctx_skip_connections:
cenc_dim = 2 * sum(config.ctx_lstm_size) #2 : fw & bw
cenc = tensor.concatenate(chidden_list, axis=2)
else:
cenc_dim = 2 * config.question_lstm_size[-1]
cenc = tensor.concatenate(chidden_list[-2:], axis=2)
cenc.name = 'cenc'
# Attention mechanism MLP fwd
attention_mlp_fwd = MLP(
dims=config.attention_mlp_hidden + [1],
activations=config.attention_mlp_activations[1:] + [Identity()],
name='attention_mlp_fwd')
attention_qlinear_fwd = Linear(
input_dim=qenc_dim,
output_dim=config.attention_mlp_hidden[0],
name='attq_fwd')
attention_clinear_fwd = Linear(
input_dim=cenc_dim / 2,
output_dim=config.attention_mlp_hidden[0],
use_bias=False,
name='attc_fwd')
bricks += [
attention_mlp_fwd, attention_qlinear_fwd, attention_clinear_fwd
]
layer1_fwd = Tanh(name='tanh_fwd')
layer1_fwd = layer1_fwd.apply(
attention_clinear_fwd.apply(cenc[:, :, :cenc_dim / 2].reshape(
(cenc.shape[0] * cenc.shape[1], cenc.shape[2] /
2))).reshape((cenc.shape[0], cenc.shape[1],
config.attention_mlp_hidden[0])) +
attention_qlinear_fwd.apply(qenc)[None, :, :])
att_weights_fwd = attention_mlp_fwd.apply(
layer1_fwd.reshape((layer1_fwd.shape[0] * layer1_fwd.shape[1],
layer1_fwd.shape[2])))
att_weights_fwd = att_weights_fwd.reshape(
(layer1_fwd.shape[0], layer1_fwd.shape[1]))
att_weights_fwd = tensor.nnet.softmax(att_weights_fwd.T)
att_weights_fwd.name = 'att_weights_fwd'
attended_fwd = tensor.sum(cenc[:, :, :cenc_dim / 2] *
att_weights_fwd.T[:, :, None],
axis=0)
attended_fwd.name = 'attended_fwd'
# Attention mechanism MLP bwd
attention_mlp_bwd = MLP(
dims=config.attention_mlp_hidden + [1],
activations=config.attention_mlp_activations[1:] + [Identity()],
name='attention_mlp_bwd')
attention_qlinear_bwd = Linear(
input_dim=qenc_dim,
output_dim=config.attention_mlp_hidden[0],
name='attq_bwd')
attention_clinear_bwd = Linear(
input_dim=cenc_dim / 2,
output_dim=config.attention_mlp_hidden[0],
use_bias=False,
name='attc_bwd')
bricks += [
attention_mlp_bwd, attention_qlinear_bwd, attention_clinear_bwd
]
layer1_bwd = Tanh(name='tanh_bwd')
layer1_bwd = layer1_bwd.apply(
attention_clinear_bwd.apply(cenc[:, :, cenc_dim / 2:].reshape(
(cenc.shape[0] * cenc.shape[1], cenc.shape[2] /
2))).reshape((cenc.shape[0], cenc.shape[1],
config.attention_mlp_hidden[0])) +
attention_qlinear_bwd.apply(qenc)[None, :, :])
att_weights_bwd = attention_mlp_bwd.apply(
layer1_bwd.reshape((layer1_bwd.shape[0] * layer1_bwd.shape[1],
layer1_bwd.shape[2])))
att_weights_bwd = att_weights_bwd.reshape(
(layer1_bwd.shape[0], layer1_bwd.shape[1]))
att_weights_bwd = tensor.nnet.softmax(att_weights_bwd.T)
att_weights_bwd.name = 'att_weights_bwd'
attended_bwd = tensor.sum(cenc[:, :, cenc_dim / 2:] *
att_weights_bwd.T[:, :, None],
axis=0)
attended_bwd.name = 'attended_bwd'
ctx_question = tensor.concatenate([attended_fwd, attended_bwd, qenc],
axis=1)
ctx_question.name = 'ctx_question'
answer_bag = to_bag(answer, vocab_size)
answer_bag = tensor.set_subtensor(answer_bag[:, 0:3], 0)
relevant_items = answer_bag.sum(axis=1, dtype=theano.config.floatX)
def createSequences(j, index, c_enc, c_enc_dim, c_context,
c_window_size):
sequence = tensor.concatenate([
c_context[j:j + index, :],
tensor.zeros((c_window_size - index, c_context.shape[1]))
],
axis=0)
enc = tensor.concatenate([
c_enc[j + index - 1, :, :], c_enc[j, :, :-1],
tensor.tile(c_window_size[None, None], (c_enc.shape[1], 1))
],
axis=1)
return enc, sequence
def createTargetValues(j, index, c_context, c_vocab_size):
sequence_bag = to_bag(c_context[j:j + index, :], c_vocab_size)
sequence_bag = tensor.set_subtensor(sequence_bag[:, 0:3], 0)
selected_items = sequence_bag.sum(axis=1,
dtype=theano.config.floatX)
tp = (sequence_bag * answer_bag).sum(axis=1,
dtype=theano.config.floatX)
precision = tp / (selected_items + 0.00001)
recall = tp / (relevant_items + 0.00001)
#precision = tensor.set_subtensor(precision[tensor.isnan(precision)], 0.0)
#recall = tensor.set_subtensor(recall[tensor.isnan(recall)], 1.0)
macroF1 = (2 *
(precision * recall)) / (precision + recall + 0.00001)
#macroF1 = tensor.set_subtensor(macroF1[tensor.isnan(macroF1)], 0.0)
return macroF1
window_size = 3
senc = []
sequences = []
pred_targets = []
for i in range(1, window_size + 1):
(all_enc, all_sequence), _ = theano.scan(
fn=createSequences,
sequences=tensor.arange(cenc.shape[0] - i + 1),
non_sequences=[i, cenc, cenc_dim, context, window_size])
(all_macroF1), _ = theano.scan(
fn=createTargetValues,
sequences=tensor.arange(cenc.shape[0] - i + 1),
non_sequences=[i, context, vocab_size])
senc.append(all_enc)
sequences.append(all_sequence)
pred_targets.append(all_macroF1)
senc = tensor.concatenate(senc, axis=0)
sequences = tensor.concatenate(sequences, axis=0)
pred_targets = tensor.concatenate(pred_targets, axis=0)
# F1 prediction Bilinear
prediction_linear = Linear(input_dim=2 * cenc_dim,
output_dim=cenc_dim + qenc_dim,
name='pred_linear')
bricks += [prediction_linear]
pred_weights = ctx_question[None, :, :] * prediction_linear.apply(
senc.reshape(
(senc.shape[0] * senc.shape[1], senc.shape[2]))).reshape(
(senc.shape[0], senc.shape[1], senc.shape[2]))
pred_weights = pred_weights.sum(axis=2)
pred_weights = tensor.nnet.sigmoid(pred_weights.T).T
pred_weights.name = 'pred_weights'
pred_targets = pred_targets / (pred_targets.sum(axis=0) + 0.00001)
pred_weights = pred_weights / (pred_weights.sum(axis=0) + 0.00001)
#numpy.set_printoptions(edgeitems=500)
#pred_targets = theano.printing.Print('pred_targets')(pred_targets)
#pred_weights = theano.printing.Print('pred_weights')(pred_weights)
cost = tensor.nnet.binary_crossentropy(pred_weights,
pred_targets).mean()
self.predictions = sequences[pred_weights.argmax(axis=0), :,
tensor.arange(sequences.shape[2])].T
# Apply dropout
cg = ComputationGraph([cost])
if config.w_noise > 0:
noise_vars = VariableFilter(roles=[WEIGHT])(cg)
cg = apply_noise(cg, noise_vars, config.w_noise)
if config.dropout > 0:
cg = apply_dropout(cg, qhidden_list + chidden_list, config.dropout)
[cost_reg] = cg.outputs
# Other stuff
cost.name = 'cost'
cost_reg.name = 'cost_reg'
self.sgd_cost = cost_reg
self.monitor_vars = [[cost_reg]]
self.monitor_vars_valid = [[cost_reg]]
# Initialize bricks
embed.initialize()
for brick in bricks:
brick.weights_init = config.weights_init
brick.biases_init = config.biases_init
brick.initialize()
def main():
x = T.imatrix('features')
m = T.matrix('features_mask')
y = T.imatrix('targets')
#x_int = x.astype(dtype='int32').T
x_int = x.T
train_dataset = IMDB('train')
n_voc = len(train_dataset.dict.keys())
n_h = 2
lookup = LookupTable(
length=n_voc+2,
dim = n_h*4,
weights_init = Uniform(std=0.01),
biases_init = Constant(0.)
)
lookup.initialize()
#rnn = SimpleRecurrent(
#dim = n_h,
#activation=Tanh(),
#weights_init = Uniform(std=0.01),
#biases_init = Constant(0.)
#)
rnn = LSTM(
dim = n_h,
activation=Tanh(),
weights_init = Uniform(std=0.01),
biases_init = Constant(0.)
)
rnn.initialize()
score_layer = Linear(
input_dim = n_h,
output_dim = 1,
weights_init = Uniform(std=0.01),
biases_init = Constant(0.))
score_layer.initialize()
embedding = lookup.apply(x_int) * T.shape_padright(m.T)
#embedding = lookup.apply(x_int) + m.T.mean()*0
#embedding = lookup.apply(x_int) + m.T.mean()*0
rnn_states = rnn.apply(embedding, mask=m.T)
#rnn_states, rnn_cells = rnn.apply(embedding)
rnn_out_mean_pooled = rnn_states[-1]
#rnn_out_mean_pooled = rnn_states.mean()
probs = Sigmoid().apply(
score_layer.apply(rnn_out_mean_pooled))
cost = - (y * T.log(probs) + (1-y) * T.log(1 - probs)).mean()
cost.name = 'cost'
misclassification = (y * (probs < 0.5) + (1-y) * (probs > 0.5)).mean()
misclassification.name = 'misclassification'
# =================
cg = ComputationGraph([cost])
params = cg.parameters
algorithm = GradientDescent(
cost = cost,
params=params,
step_rule = CompositeRule([
StepClipping(threshold=10),
Adam(),
#AdaDelta(),
])
)
# ========
test_dataset = IMDB('test')
batch_size = 64
n_train = train_dataset.num_examples
train_stream = DataStream(
dataset=train_dataset,
iteration_scheme=ShuffledScheme(
examples=n_train,
batch_size=batch_size)
)
train_padded = Padding(
data_stream=train_stream,
mask_sources=('features',)
#mask_sources=[]
)
test_stream = DataStream(
dataset=test_dataset,
iteration_scheme=ShuffledScheme(
examples=n_train,
batch_size=batch_size)
)
test_padded = Padding(
data_stream=test_stream,
mask_sources=('features',)
#mask_sources=[]
)
#import ipdb
#ipdb.set_trace()
#======
model = Model(cost)
extensions = []
extensions.append(EpochProgress(batch_per_epoch=train_dataset.num_examples // batch_size + 1))
extensions.append(TrainingDataMonitoring(
[cost, misclassification],
prefix='train',
after_epoch=True
))
extensions.append(DataStreamMonitoring(
[cost, misclassification],
data_stream=test_padded,
prefix='test',
after_epoch=True
))
extensions.append(Timing())
extensions.append(Printing())
main_loop = MainLoop(
model=model,
data_stream=train_padded,
algorithm=algorithm,
extensions=extensions)
main_loop.run()
def __init__(self, config, vocab_size):
question = tensor.imatrix('question')
question_mask = tensor.imatrix('question_mask')
context = tensor.imatrix('context')
context_mask = tensor.imatrix('context_mask')
answer = tensor.imatrix('answer')
answer_mask = tensor.imatrix('answer_mask')
bricks = []
question = question.dimshuffle(1, 0)
question_mask = question_mask.dimshuffle(1, 0)
context = context.dimshuffle(1, 0)
context_mask = context_mask.dimshuffle(1, 0)
answer = answer.dimshuffle(1, 0)
answer_mask = answer_mask.dimshuffle(1, 0)
# Embed questions and context
embed = LookupTable(vocab_size,
config.embed_size,
name='question_embed')
embed.weights_init = IsotropicGaussian(0.01)
# Calculate question encoding (concatenate layer1)
qembed = embed.apply(question)
qlstms, qhidden_list = make_bidir_lstm_stack(
qembed, config.embed_size,
question_mask.astype(theano.config.floatX),
config.question_lstm_size, config.question_skip_connections, 'q')
bricks = bricks + qlstms
if config.question_skip_connections:
qenc_dim = 2 * sum(config.question_lstm_size)
qenc = tensor.concatenate([h[-1, :, :] for h in qhidden_list],
axis=1)
else:
qenc_dim = 2 * config.question_lstm_size[-1]
qenc = tensor.concatenate([h[-1, :, :] for h in qhidden_list[-2:]],
axis=1)
qenc.name = 'qenc'
# Calculate context encoding (concatenate layer1)
cembed = embed.apply(context)
clstms, chidden_list = make_bidir_lstm_stack(
cembed, config.embed_size,
context_mask.astype(theano.config.floatX), config.ctx_lstm_size,
config.ctx_skip_connections, 'ctx')
bricks = bricks + clstms
if config.ctx_skip_connections:
cenc_dim = 2 * sum(config.ctx_lstm_size) #2 : fw & bw
cenc = tensor.concatenate(chidden_list, axis=2)
else:
cenc_dim = 2 * config.question_lstm_size[-1]
cenc = tensor.concatenate(chidden_list[-2:], axis=2)
cenc.name = 'cenc'
# Attention mechanism MLP
attention_mlp = MLP(dims=config.attention_mlp_hidden + [1],
activations=config.attention_mlp_activations[1:] +
[Identity()],
name='attention_mlp')
attention_qlinear = Linear(input_dim=qenc_dim,
output_dim=config.attention_mlp_hidden[0],
name='attq')
attention_clinear = Linear(input_dim=cenc_dim,
output_dim=config.attention_mlp_hidden[0],
use_bias=False,
name='attc')
bricks += [attention_mlp, attention_qlinear, attention_clinear]
layer1 = Tanh().apply(
attention_clinear.apply(
cenc.reshape((cenc.shape[0] * cenc.shape[1], cenc.shape[2]
))).reshape((cenc.shape[0], cenc.shape[1],
config.attention_mlp_hidden[0])) +
attention_qlinear.apply(qenc)[None, :, :])
layer1.name = 'layer1'
att_weights = attention_mlp.apply(
layer1.reshape(
(layer1.shape[0] * layer1.shape[1], layer1.shape[2])))
att_weights = att_weights.reshape((layer1.shape[0], layer1.shape[1]))
att_weights = tensor.nnet.sigmoid(att_weights.T).T
att_weights.name = 'att_weights'
att_target = tensor.eq(
tensor.tile(answer[None, :, :], (context.shape[0], 1, 1)),
tensor.tile(context[:, None, :],
(1, answer.shape[0], 1))).sum(axis=1).clip(0, 1)
cost = (tensor.nnet.binary_crossentropy(att_weights, att_target) *
context_mask).sum() / context_mask.sum()
self.predictions = tensor.gt(att_weights, 0.1) * context
# Apply dropout
cg = ComputationGraph([cost])
if config.w_noise > 0:
noise_vars = VariableFilter(roles=[WEIGHT])(cg)
cg = apply_noise(cg, noise_vars, config.w_noise)
if config.dropout > 0:
cg = apply_dropout(cg, qhidden_list + chidden_list, config.dropout)
[cost_reg] = cg.outputs
# Other stuff
cost.name = 'cost'
cost_reg.name = 'cost_reg'
self.sgd_cost = cost_reg
self.monitor_vars = [[cost_reg]]
self.monitor_vars_valid = [[cost_reg]]
# Initialize bricks
embed.initialize()
for brick in bricks:
brick.weights_init = config.weights_init
brick.biases_init = config.biases_init
brick.initialize()
hidden_layer_size = 200
layer1 = LookupTable(name='layer1',
length=train_dataset.words_bag_size,
dim=hidden_layer_size,
weights_init=Uniform(mean=0, std=0.01),
biases_init=Constant(0))
act1_mean = tensor.mean(layer1.apply(x), axis=1)
layer2 = Linear(name='layer2',
input_dim=layer1.output_dim,
output_dim=train_dataset.words_bag_size,
weights_init=Uniform(mean=0, std=0.01),
biases_init=Constant(0))
act2_softmax = Softmax().apply(layer2.apply(act1_mean))
layer1.initialize()
layer2.initialize()
missclass = MisclassificationRate().apply(y, act2_softmax)
cost = CategoricalCrossEntropy().apply(y, act2_softmax)
cg = ComputationGraph([cost])
W1, W2 = VariableFilter(roles=[WEIGHT])(cg.variables)
cost = cost + 0.00001 * (W1**2).sum() + 0.00005 * (W2**2).sum()
cost.name = 'cost'
from blocks.algorithms import GradientDescent, Scale
VOCAB_DIM = dataset.vocabulary_size
print "vocab size:", VOCAB_DIM
EMBEDDING_DIM = 100
Xs = tensor.imatrix("context")
y = tensor.ivector('center')
w1 = LookupTable(name="w1", length=VOCAB_DIM, dim=EMBEDDING_DIM)
w2 = Linear(name='w2', input_dim=EMBEDDING_DIM, output_dim=VOCAB_DIM)
hidden = tensor.mean(w1.apply(Xs), axis=1)
y_hat = Softmax().apply(w2.apply(hidden))
w1.weights_init = w2.weights_init = IsotropicGaussian(0.01)
w1.biases_init = w2.biases_init = Constant(0)
w1.initialize()
w2.initialize()
cost = CategoricalCrossEntropy().apply(y, y_hat)
cg = ComputationGraph(cost)
W1, W2 = VariableFilter(roles=[WEIGHT])(cg.variables)
cost = cost + 0.005 * (W1 ** 2).sum() + 0.005 * (W2 ** 2).sum()
cost.name = "loss"
#
# the actual training of the model
#
main = MainLoop(data_stream = DataStream.default_stream(
def __init__(self, config, vocab_size):
unsorted = tensor.imatrix('unsorted')
unsorted_mask = tensor.imatrix('unsorted_mask')
answer = tensor.imatrix('answer')
answer_mask = tensor.imatrix('answer_mask')
bricks = []
unsorted = unsorted.dimshuffle(1, 0)
unsorted_mask = unsorted_mask.dimshuffle(1, 0)
answer = answer.dimshuffle(1, 0)
answer_mask = answer_mask.dimshuffle(1, 0)
# Embed unsorted sequence
embed = LookupTable(vocab_size, config.embed_size, name='embed')
embed.weights_init = IsotropicGaussian(0.01)
#make_bidir_lstm_stack(seq, seq_dim, mask, sizes, skip=True, name=''):
unsorted_embed = embed.apply(unsorted)
unsorted_lstms, unsorted_hidden_list = make_bidir_lstm_stack(
unsorted_embed, config.embed_size,
unsorted_mask.astype(theano.config.floatX), config.lstm_size,
config.match_skip_connections, 'u') #lu,bs,lstm_dim
bricks = bricks + unsorted_lstms
unsorted_enc_dim = 2 * sum(config.lstm_size)
unsorted_enc = tensor.concatenate(
unsorted_hidden_list,
axis=2) #concatenate fwd & bwd lstm hidden states
unsorted_enc.name = 'unsorted_enc'
#pointer networks decoder LSTM and Attention parameters
params = init_params(data_dim=config.decoder_data_dim,
lstm_dim=config.decoder_lstm_output_dim)
tparams = init_tparams(params)
add_role(tparams['lstm_de_W'], WEIGHT)
add_role(tparams['lstm_de_U'], WEIGHT)
add_role(tparams['lstm_de_b'], BIAS)
add_role(tparams['ptr_b1'], BIAS)
add_role(tparams['ptr_b2'], BIAS)
add_role(tparams['ptr_v'], WEIGHT)
add_role(tparams['ptr_W1'], WEIGHT)
add_role(tparams['ptr_W2'], WEIGHT)
self.theano_params = tparams.values()
#n_steps = length , n_samples = batch_size
n_steps = answer.shape[0]
n_samples = answer.shape[1]
preds, generations = ptr_network(
tparams, unsorted_embed,
unsorted_mask.astype(theano.config.floatX), answer,
answer_mask.astype(theano.config.floatX),
config.decoder_lstm_output_dim, unsorted_enc)
self.generations = generations
idx_steps = tensor.outer(tensor.arange(n_steps, dtype='int64'),
tensor.ones((n_samples, ), dtype='int64'))
idx_samples = tensor.outer(tensor.ones((n_steps, ), dtype='int64'),
tensor.arange(n_samples, dtype='int64'))
probs = preds[idx_steps, answer, idx_samples]
# probs *= y_mask
off = 1e-8
if probs.dtype == 'float16':
off = 1e-6
probs += off
# probs_printed = theano.printing.Print('probs')(probs)
cost = -tensor.log(probs)
cost *= answer_mask
cost = cost.sum(axis=0) / answer_mask.sum(axis=0)
cost = cost.mean()
# Apply dropout
cg = ComputationGraph([cost])
if config.w_noise > 0:
noise_vars = VariableFilter(roles=[WEIGHT])(cg)
cg = apply_noise(cg, noise_vars, config.w_noise)
if config.dropout > 0:
cg = apply_dropout(cg, unsorted_hidden_list, config.dropout)
[cost_reg] = cg.outputs
# Other stuff
cost.name = 'cost'
cost_reg.name = 'cost_reg'
self.sgd_cost = cost_reg
self.monitor_vars = [[cost_reg]]
self.monitor_vars_valid = [[cost_reg]]
# Initialize bricks
embed.initialize()
for brick in bricks:
brick.weights_init = config.weights_init
brick.biases_init = config.biases_init
brick.initialize()
)
generator = SequenceGenerator(readout=readout, transition=transition, attention=attention, name="generator")
generator.weights_init = IsotropicGaussian(0.01)
generator.biases_init = Constant(0.0)
generator.push_initialization_config()
generator.transition.biases_init = IsotropicGaussian(0.01, 1)
generator.transition.push_initialization_config()
generator.initialize()
lookup.weights_init = IsotropicGaussian(0.001)
lookup.biases_init = Constant(0.0)
lookup.initialize()
# states = {}
states = [state for state in generator.transition.apply.outputs if state != "step"]
# ipdb.set_trace()
states = {name: shared_floatx_zeros((batch_size, hidden_size_recurrent)) for name in states}
cost_matrix = generator.cost_matrix(x, attended=context, **states)
cost = cost_matrix.mean() + 0.0 * start_flag
cost.name = "nll"
cg = ComputationGraph(cost)
def main():
x = T.imatrix('features')
m = T.matrix('features_mask')
y = T.imatrix('targets')
#x_int = x.astype(dtype='int32').T
x_int = x.T
train_dataset = IMDB('train')
n_voc = len(train_dataset.dict.keys())
n_h = 2
lookup = LookupTable(length=n_voc + 2,
dim=n_h * 4,
weights_init=Uniform(std=0.01),
biases_init=Constant(0.))
lookup.initialize()
#rnn = SimpleRecurrent(
#dim = n_h,
#activation=Tanh(),
#weights_init = Uniform(std=0.01),
#biases_init = Constant(0.)
#)
rnn = LSTM(dim=n_h,
activation=Tanh(),
weights_init=Uniform(std=0.01),
biases_init=Constant(0.))
rnn.initialize()
score_layer = Linear(input_dim=n_h,
output_dim=1,
weights_init=Uniform(std=0.01),
biases_init=Constant(0.))
score_layer.initialize()
embedding = lookup.apply(x_int) * T.shape_padright(m.T)
#embedding = lookup.apply(x_int) + m.T.mean()*0
#embedding = lookup.apply(x_int) + m.T.mean()*0
rnn_states = rnn.apply(embedding, mask=m.T)
#rnn_states, rnn_cells = rnn.apply(embedding)
rnn_out_mean_pooled = rnn_states[-1]
#rnn_out_mean_pooled = rnn_states.mean()
probs = Sigmoid().apply(score_layer.apply(rnn_out_mean_pooled))
cost = -(y * T.log(probs) + (1 - y) * T.log(1 - probs)).mean()
cost.name = 'cost'
misclassification = (y * (probs < 0.5) + (1 - y) * (probs > 0.5)).mean()
misclassification.name = 'misclassification'
# =================
cg = ComputationGraph([cost])
params = cg.parameters
algorithm = GradientDescent(
cost=cost,
params=params,
step_rule=CompositeRule([
StepClipping(threshold=10),
Adam(),
#AdaDelta(),
]))
# ========
test_dataset = IMDB('test')
batch_size = 64
n_train = train_dataset.num_examples
train_stream = DataStream(dataset=train_dataset,
iteration_scheme=ShuffledScheme(
examples=n_train, batch_size=batch_size))
train_padded = Padding(data_stream=train_stream,
mask_sources=('features', )
#mask_sources=[]
)
test_stream = DataStream(dataset=test_dataset,
iteration_scheme=ShuffledScheme(
examples=n_train, batch_size=batch_size))
test_padded = Padding(data_stream=test_stream,
mask_sources=('features', )
#mask_sources=[]
)
#import ipdb
#ipdb.set_trace()
#======
model = Model(cost)
extensions = []
extensions.append(
EpochProgress(
batch_per_epoch=train_dataset.num_examples // batch_size + 1))
extensions.append(
TrainingDataMonitoring([cost, misclassification],
prefix='train',
after_epoch=True))
extensions.append(
DataStreamMonitoring([cost, misclassification],
data_stream=test_padded,
prefix='test',
after_epoch=True))
extensions.append(Timing())
extensions.append(Printing())
main_loop = MainLoop(model=model,
data_stream=train_padded,
algorithm=algorithm,
extensions=extensions)
main_loop.run()
