def generate_lstm(net, step, lstm_params, lstm_out, dropout_ratio):
"""Takes the parameters to create the lstm, concatenates the lstm input
with the previous hidden state, runs the lstm for the current timestep
and then applies dropout to the output hidden state."""
hidden_bottom = lstm_out[0]
mem_bottom = lstm_out[1]
num_cells = lstm_params[0]
filler = lstm_params[1]
net.f(Concat("concat%d" % step, bottoms=["lstm_input", hidden_bottom]))
try:
lstm_unit = LstmUnit("lstm%d" % step, num_cells,
weight_filler=filler, tie_output_forget=True,
param_names=["input_value", "input_gate",
"forget_gate", "output_gate"],
bottoms=["concat%d" % step, mem_bottom],
tops=["lstm_hidden%d" % step, "lstm_mem%d" % step])
except:
# Old version of Apollocaffe sets tie_output_forget=True by default
lstm_unit = LstmUnit("lstm%d" % step, num_cells,
weight_filler=filler,
param_names=["input_value", "input_gate",
"forget_gate", "output_gate"],
bottoms=["concat%d" % step, mem_bottom],
tops=["lstm_hidden%d" % step, "lstm_mem%d" % step])
net.f(lstm_unit)
net.f(Dropout("dropout%d" % step, dropout_ratio,
bottoms=["lstm_hidden%d" % step]))
def forward(net, sentence_batches):
net.clear_forward()
batch = next(sentence_batches)
sentence_batch = pad_batch(batch)
length = min(sentence_batch.shape[1], 100)
assert length > 0
net.f(NumpyData('lstm_seed', np.zeros((batch_size, dimension))))
for step in range(length):
if step == 0:
prev_hidden = 'lstm_seed'
prev_mem = 'lstm_seed'
word = np.zeros(sentence_batch[:, 0].shape)
else:
prev_hidden = 'lstm%d_hidden' % (step - 1)
prev_mem = 'lstm%d_mem' % (step - 1)
word = sentence_batch[:, step - 1]
net.f(NumpyData('word%d' % step, word))
net.f(
Wordvec('wordvec%d' % step,
dimension,
vocab_size,
bottoms=['word%d' % step],
param_names=['wordvec_param']))
net.f(
Concat('lstm_concat%d' % step,
bottoms=[prev_hidden, 'wordvec%d' % step]))
net.f(
LstmUnit('lstm%d' % step,
bottoms=['lstm_concat%d' % step, prev_mem],
param_names=[
'lstm_input_value', 'lstm_input_gate',
'lstm_forget_gate', 'lstm_output_gate'
],
tops=['lstm%d_hidden' % step,
'lstm%d_mem' % step],
num_cells=dimension))
net.f(
Dropout('dropout%d' % step, 0.16,
bottoms=['lstm%d_hidden' % step]))
net.f(NumpyData('label%d' % step, sentence_batch[:, step]))
net.f(
InnerProduct('ip%d' % step,
vocab_size,
bottoms=['dropout%d' % step],
param_names=['softmax_ip_weights',
'softmax_ip_bias']))
net.f(
SoftmaxWithLoss('softmax_loss%d' % step,
ignore_label=zero_symbol,
bottoms=['ip%d' % step,
'label%d' % step]))
def eval_forward(net):
net.clear_forward()
output_words = []
net.f(NumpyData('lstm_hidden_prev', np.zeros((1, dimension))))
net.f(NumpyData('lstm_mem_prev', np.zeros((1, dimension))))
length = 150
for step in range(length):
net.clear_forward()
net.f(NumpyData('word', [0]))
prev_hidden = 'lstm_hidden_prev'
prev_mem = 'lstm_mem_prev'
if step == 0:
output = ord('.')
else:
output = softmax_choice(net.blobs['softmax'].data)
output_words.append(output)
net.blobs['word'].data[0] = output
net.f(
Wordvec('wordvec',
dimension,
vocab_size,
bottoms=['word'],
param_names=['wordvec_param']))
net.f(Concat('lstm_concat', bottoms=[prev_hidden, 'wordvec']))
net.f(
LstmUnit('lstm',
dimension,
bottoms=['lstm_concat', prev_mem],
param_names=[
'lstm_input_value', 'lstm_input_gate',
'lstm_forget_gate', 'lstm_output_gate'
],
tops=['lstm_hidden_next', 'lstm_mem_next']))
net.f(Dropout('dropout', 0.16, bottoms=['lstm_hidden_next']))
net.f(
InnerProduct('ip',
vocab_size,
bottoms=['dropout'],
param_names=['softmax_ip_weights',
'softmax_ip_bias']))
net.blobs['ip'].data[:] *= i_temperature
net.f(Softmax('softmax', bottoms=['ip']))
net.blobs['lstm_hidden_prev'].data_tensor.copy_from(
net.blobs['lstm_hidden_next'].data_tensor)
net.blobs['lstm_mem_prev'].data_tensor.copy_from(
net.blobs['lstm_mem_next'].data_tensor)
print ''.join([chr(x) for x in output_words])
def evaluate_forward(net, net_config):
net.clear_forward()
length = 20
net.f(NumpyData("prev_hidden", np.zeros((1, net_config["mem_cells"]))))
net.f(NumpyData("prev_mem", np.zeros((1, net_config["mem_cells"]))))
filler = Filler("uniform", net_config["init_range"])
predictions = []
value = 0.5
for _ in range(length):
# We'll be updating values in place for efficient memory usage. This
# will break backprop and cause warnings. Use clear_forward to suppress.
net.clear_forward()
# Add 0.5 to the sum at each step
net.f(NumpyData("value", data=np.array(value).reshape((1, 1))))
prev_hidden = "prev_hidden"
prev_mem = "prev_mem"
net.f(Concat("lstm_concat", bottoms=[prev_hidden, "value"]))
net.f(
LstmUnit("lstm",
net_config["mem_cells"],
bottoms=["lstm_concat", prev_mem],
param_names=[
"input_value", "input_gate", "forget_gate",
"output_gate"
],
weight_filler=filler,
tops=["next_hidden", "next_mem"]))
net.f(InnerProduct("ip", 1, bottoms=["next_hidden"]))
predictions.append(float(net.blobs["ip"].data.flatten()[0]))
# set up for next prediction by copying LSTM outputs back to inputs
net.blobs["prev_hidden"].data_tensor.copy_from(
net.blobs["next_hidden"].data_tensor)
net.blobs["prev_mem"].data_tensor.copy_from(
net.blobs["next_mem"].data_tensor)
targets = np.cumsum([value for _ in predictions])
residuals = [x - y for x, y in zip(predictions, targets)]
return targets, predictions, residuals
def generate_lstm(net, step, num_cells, hidden_bottom, mem_bottom, filler,
dropout_ratio):
"""Takes the parameters to create the lstm, concatenates the lstm input
with the previous hidden state, runs the lstm for the current timestep and then
applies dropout to the output hidden state."""
net.f(Concat("concat%d" % step, bottoms=["lstm_input", hidden_bottom]))
net.f(
LstmUnit("lstm%d" % step,
num_cells,
weight_filler=filler,
param_names=[
"input_value", "input_gate", "forget_gate", "output_gate"
],
bottoms=["concat%d" % step, mem_bottom],
tops=["lstm_hidden%d" % step,
"lstm_mem%d" % step]))
net.f(
Dropout("dropout%d" % step,
dropout_ratio,
bottoms=["lstm_hidden%d" % step]))
def evaluate_forward(net, net_config, feat,scene_feat):
net.clear_forward()
feat_dim=feat.shape[1]
net.f(NumpyData("prev_hidden", np.zeros((1, net_config["mem_cells"]))))
net.f(NumpyData("prev_mem", np.zeros((1, net_config["mem_cells"]))))
filler = Filler("uniform", net_config["init_range"])
length = feat.shape[0]+1
for step in range(length):
net.clear_forward()
if step==0:
value=scene_feat.reshape(1,feat_dim)
else:
value = feat[step-1,:].reshape(1,feat_dim)
net.f(NumpyData("value", data=value ))
prev_hidden = "prev_hidden"
prev_mem = "prev_mem"
net.f(Concat("lstm_concat", bottoms=[prev_hidden, "value"]))
net.f(LstmUnit("lstm", net_config["mem_cells"],
bottoms=["lstm_concat", prev_mem],
param_names=[
"input_value", "input_gate", "forget_gate", "output_gate"],
weight_filler=filler,
tops=["next_hidden", "next_mem"]))
net.f(InnerProduct("ip", 1, bottoms=["next_hidden"]))
net.blobs["prev_hidden"].data_tensor.copy_from(
net.blobs["next_hidden"].data_tensor)
net.blobs["prev_mem"].data_tensor.copy_from(
net.blobs["next_mem"].data_tensor)
for i in xrange(6):
net.f(InnerProduct("ip%d"%i, 2, bottoms=["next_hidden"]))
net.f(Softmax("prob%d"%i, bottoms=["ip%d"%i]))
predictions=[]
for i in xrange(6):
predictions.append(float(net.blobs["prob%d"%i].data.flatten()[1]))
return predictions
def lstm_layers(net, step, filler, net_config):
layer_list = []
if step == 0:
prev_hidden = "lstm_seed"
prev_mem = "lstm_seed"
else:
prev_hidden = "lstm_hidden%d" % (step - 1)
prev_mem = "lstm_mem%d" % (step - 1)
# Concatenate the hidden output with the next input value
layer_list.append(
Concat("lstm_concat%d" % step, bottoms=[prev_hidden,
"value%d" % step]))
# Run the LSTM for one more step
layer_list.append(
LstmUnit("lstm%d" % step,
net_config["mem_cells"],
bottoms=["lstm_concat%d" % step, prev_mem],
param_names=[
"input_value", "input_gate", "forget_gate", "output_gate"
],
tops=["lstm_hidden%d" % step,
"lstm_mem%d" % step],
weight_filler=filler))
return layer_list
def forward(net, input_data, net_config, deploy=False):
"""Defines and creates the ReInspect network given the net, input data
and configurations."""
net.clear_forward()
net.f(
NumpyData("wordvec_layer",
data=np.array(input_data["wordvec_layer"]))) # 128*38*100*1
net.f(NumpyData("target_words",
data=np.array(input_data["target_words"]))) # 128*100*1*1
tops = []
slice_point = []
for i in range(net_config['max_len']):
tops.append('label%d' % i)
if i != 0:
slice_point.append(i)
net.f(
Slice("label_slice_layer",
slice_dim=1,
bottoms=["target_words"],
tops=tops,
slice_point=slice_point))
tops = []
slice_point = []
for i in range(net_config['max_len']):
tops.append('target_wordvec%d_4d' % i)
if i != 0:
slice_point.append(i)
net.f(
Slice("wordvec_slice_layer",
slice_dim=2,
bottoms=['wordvec_layer'],
tops=tops,
slice_point=slice_point))
for i in range(net_config["max_len"]): # 128*38*1*1 -> 128*38
net.f("""
name: "target_wordvec%d"
type: "Reshape"
bottom: "target_wordvec%d_4d"
top: "target_wordvec%d"
reshape_param {
shape {
dim: 0 # copy the dimension from below
dim: -1
}
}
""" % (i, i, i))
#net.f(Reshape('target_wordvec%d'%i, bottoms = ['target_wordvec%d_4d'%i], shape = [0,-1]))
filler = Filler("uniform", net_config["init_range"])
for i in range(net_config['max_len']):
if i == 0:
net.f(
NumpyData(
"dummy_layer",
np.zeros((net_config["batch_size"],
net_config["lstm_num_cells"]))))
net.f(
NumpyData(
"dummy_mem_cell",
np.zeros((net_config["batch_size"],
net_config["lstm_num_cells"]))))
for j in range(net_config['lstm_num_stacks']):
bottoms = []
if j == 0:
bottoms.append('target_wordvec%d' % i)
if j >= 1:
bottoms.append('dropout%d_%d' % (j - 1, i))
if i == 0:
bottoms.append("dummy_layer")
else:
bottoms.append('lstm%d_hidden%d' % (j, i - 1))
net.f(Concat('concat%d_layer%d' % (j, i), bottoms=bottoms))
param_names = []
for k in range(4):
param_names.append('lstm%d_param_%d' % (j, k))
bottoms = ['concat%d_layer%d' % (j, i)]
if i == 0:
bottoms.append('dummy_mem_cell')
else:
bottoms.append('lstm%d_mem_cell%d' % (j, i - 1))
net.f(
LstmUnit('lstm%d_layer%d' % (j, i),
net_config["lstm_num_cells"],
weight_filler=filler,
param_names=param_names,
bottoms=bottoms,
tops=[
'lstm%d_hidden%d' % (j, i),
'lstm%d_mem_cell%d' % (j, i)
]))
net.f(
Dropout('dropout%d_%d' % (j, i),
net_config["dropout_ratio"],
bottoms=['lstm%d_hidden%d' % (j, i)]))
bottoms = []
for i in range(net_config['max_len']):
bottoms.append('dropout%d_%d' % (net_config['lstm_num_stacks'] - 1, i))
net.f(Concat('hidden_concat', bottoms=bottoms, concat_dim=0))
net.f(
InnerProduct("inner_product",
net_config['vocab_size'],
bottoms=["hidden_concat"],
weight_filler=filler))
bottoms = []
for i in range(net_config['max_len']):
bottoms.append('label%d' % i)
net.f(Concat('label_concat', bottoms=bottoms, concat_dim=0))
if deploy:
net.f(Softmax("word_probs", bottoms=["inner_product"]))
else:
net.f(
SoftmaxWithLoss("word_loss",
bottoms=["inner_product", "label_concat"],
ignore_label=net_config['zero_symbol']))
def forward(net, input_data, net_config, phase='train', deploy=False):
"""Defines and creates the ReInspect network given the net, input data
and configurations."""
net.clear_forward()
batch_ws_i = input_data["ws_i"]
batch_stop_i = [net_config['max_len']] * net_config['batch_size']
wordvec_layer = input_data["wordvec_layer"] # 128*38*100*1
net.f(NumpyData("target_words",
data=np.array(input_data["target_words"]))) # 128*100*1*1
tops = []
slice_point = []
for i in range(net_config['max_len']):
tops.append('label%d' % i)
if i != 0:
slice_point.append(i)
net.f(
Slice("label_slice_layer",
slice_dim=1,
bottoms=["target_words"],
tops=tops,
slice_point=slice_point))
net.f(NumpyData("target_wordvec%d" % 0,
data=wordvec_layer[:, :, 0, 0])) # start symbol, 128*38
filler = Filler("uniform", net_config["init_range"])
for i in range(net_config['max_len']):
if i == 0:
net.f(
NumpyData(
"dummy_layer",
np.zeros((net_config["batch_size"],
net_config["lstm_num_cells"]))))
net.f(
NumpyData(
"dummy_mem_cell",
np.zeros((net_config["batch_size"],
net_config["lstm_num_cells"]))))
for j in range(net_config['lstm_num_stacks']):
bottoms = []
if j == 0:
bottoms.append('target_wordvec%d' % i)
if j >= 1:
bottoms.append('dropout%d_%d' % (j - 1, i))
if i == 0:
bottoms.append("dummy_layer")
else:
bottoms.append('lstm%d_hidden%d' % (j, i - 1))
net.f(Concat('concat%d_layer%d' % (j, i), bottoms=bottoms))
param_names = []
for k in range(4):
param_names.append('lstm%d_param_%d' % (j, k))
bottoms = ['concat%d_layer%d' % (j, i)]
if i == 0:
bottoms.append('dummy_mem_cell')
else:
bottoms.append('lstm%d_mem_cell%d' % (j, i - 1))
net.f(
LstmUnit('lstm%d_layer%d' % (j, i),
net_config["lstm_num_cells"],
weight_filler=filler,
param_names=param_names,
bottoms=bottoms,
tops=[
'lstm%d_hidden%d' % (j, i),
'lstm%d_mem_cell%d' % (j, i)
]))
net.f(
Dropout('dropout%d_%d' % (j, i),
net_config["dropout_ratio"],
bottoms=['lstm%d_hidden%d' % (j, i)]))
net.f(
InnerProduct("ip%d" % i,
net_config['vocab_size'],
bottoms=[
'dropout%d_%d' %
(net_config['lstm_num_stacks'] - 1, i)
],
weight_filler=filler))
if i < net_config['max_len'] - 1:
tar_wordvec = np.array(wordvec_layer[:, :, i + 1, 0]) # 128*38
if phase == 'test':
net.f(Softmax("word_probs%d" % i, bottoms=["ip%d" % i]))
probs = net.blobs["word_probs%d" % i].data
for bi in range(net_config['batch_size']):
if i >= batch_ws_i[bi] and i < batch_stop_i[bi]:
vec = [0] * net_config["vocab_size"]
peakIndex = np.argmax(probs[bi, :])
if peakIndex == net_config['whitespace_symbol']:
batch_stop_i[bi] = i + 1
vec[peakIndex] = 1
tar_wordvec[bi, :] = vec
net.f(NumpyData("target_wordvec%d" % (i + 1), data=tar_wordvec))
bottoms = []
for i in range(net_config['max_len']):
bottoms.append("ip%d" % i)
net.f(Concat('ip_concat', bottoms=bottoms, concat_dim=0))
bottoms = []
for i in range(net_config['max_len']):
bottoms.append('label%d' % i)
net.f(Concat('label_concat', bottoms=bottoms, concat_dim=0))
if deploy:
net.f(Softmax("word_probs", bottoms=["ip_concat"]))
net.f(
SoftmaxWithLoss("word_loss",
bottoms=["ip_concat", "label_concat"],
ignore_label=net_config['zero_symbol']))