I've setup a print statement and I've noticed that for the first batch when feeding an RNN, the embeddings exist, but after the second batch they don't and I get the following error:
ValueError: Variable RNNLM/RNNLM/Embedding/Adam_2/ does not exist, or was not created with tf.get_variable(). Did you mean to set reuse=None in VarScope?
Here is my code for generating the embeddings:
def add_embedding(self):
with tf.device('/gpu:0'):
embedding = tf.get_variable("Embedding", [len(self.vocab), self.config.embed_size])
e_x = tf.nn.embedding_lookup(embedding, self.input_placeholder)
inputs = [tf.squeeze(s, [1]) for s in tf.split(1, self.config.num_steps, e_x)]
return inputs
Here is how the model is seutp, this is where I suspect the problem lies
def model(self, inputs):
with tf.variable_scope("input_drop"):
inputs_drop = [tf.nn.dropout(i, self.dropout_placeholder) for i in inputs]
with tf.variable_scope("RNN") as scope:
self.initial_state = tf.zeros([self.config.batch_size, self.config.hidden_size], tf.float32)
state = self.initial_state
states = []
for t, e in enumerate(inputs_drop):
print "t is {0}".format(t)
if t > 0:
scope.reuse_variables()
H = tf.get_variable("Hidden", [self.config.hidden_size, self.config.hidden_size])
I = tf.get_variable("I", [self.config.embed_size, self.config.hidden_size])
b_1 = tf.get_variable("b_1", (self.config.hidden_size,))
state = tf.sigmoid(tf.matmul(state, H) + tf.matmul(e, I) + b_1)
states.append(state)
with tf.variable_scope("output_dropout"):
rnn_outputs = [tf.nn.dropout(o, self.dropout_placeholder) for o in states]
return rnn_outputs
The issue arises when I get to the loss function, defined as follows
def add_training_op(self, loss):
opt = tf.train.AdamOptimizer(self.config.lr)
train_op = opt.minimize(loss)
return train_op
EDIT: Here is some updated code to help everyone out
def __init__(self, config):
self.config = config
self.load_data(debug=False)
self.add_placeholders()
self.inputs = self.add_embedding()
self.rnn_outputs = self.add_model(self.inputs)
self.outputs = self.add_projection(self.rnn_outputs)
self.predictions = [tf.nn.softmax(tf.cast(o, 'float64')) for o in self.outputs]
output = tf.reshape(tf.concat(1, self.outputs), [-1, len(self.vocab)])
self.calculate_loss = self.add_loss_op(output)
self.train_step = self.add_training_op(self.calculate_loss)
Here are the other methods here, pertaining to add_projection and calculate_loss so we can rule them out.
def add_loss_op(self, output):
weights = tf.ones([self.config.batch_size * self.config.num_steps], tf.int32)
seq_loss = tf.python.seq2seq.sequence_loss(
[output],
tf.reshape(self.labels_placeholder, [-1]),
weights
)
tf.add_to_collection('total_loss', seq_loss)
loss = tf.add_n(tf.get_collection('total_loss'))
return loss
def add_projection(self, rnn_outputs):
with tf.variable_scope("Projection", initializer=tf.contrib.layers.xavier_initializer()) as scope:
U = tf.get_variable("U", [self.config.hidden_size, len(self.vocab)])
b_2 = tf.get_variable("b_2", [len(self.vocab)])
outputs = [tf.matmul(x, U) + b_2 for x in rnn_outputs]
return outputs
def train_RNNLM():
config = Config()
gen_config = deepcopy(config)
gen_config.batch_size = gen_config.num_steps = 1
with tf.variable_scope('RNNLM') as scope:
model = RNNLM_Model(config)
# This instructs gen_model to reuse the same variables as the model above
scope.reuse_variables()
gen_model = RNNLM_Model(gen_config)
init = tf.initialize_all_variables()
saver = tf.train.Saver()
with tf.Session() as session:
best_val_pp = float('inf')
best_val_epoch = 0
session.run(init)
for epoch in xrange(config.max_epochs):
print 'Epoch {}'.format(epoch)
start = time.time()
###
train_pp = model.run_epoch(
session, model.encoded_train,
train_op=model.train_step)
valid_pp = model.run_epoch(session, model.encoded_valid)
print 'Training perplexity: {}'.format(train_pp)
print 'Validation perplexity: {}'.format(valid_pp)
if valid_pp < best_val_pp:
best_val_pp = valid_pp
best_val_epoch = epoch
saver.save(session, './ptb_rnnlm.weights')
if epoch - best_val_epoch > config.early_stopping:
break
print 'Total time: {}'.format(time.time() - start)
Seems that the code is trying to create a new Adam Variable in each batch.
Possible that the add_training_op is called twice?
Also, the snippet of def add_training_op is incomplete since there is no return statement.
The problem turned out to be the following line of code:
model = RNNLM_Model(config)
# This instructs gen_model to reuse the same variables as the model above
scope.reuse_variables()
gen_model = RNNLM_Model(gen_config)
It turns out that the second model was an issue by using reuse_variables(). By removing this line by issues went away.
Related
I build an Inverse Compositional CNN but it reports errors as bellows:
ValueError: Variable left_src_tgt_warp/ICSTN/icnv1/weight already exists, disallowed. Did you mean to set reuse=True in VarScope? Originally defined at:
I find that using tf.reset_default_graph() can solve this problem. But I do not know where I should add it.
for l in range(opt.warpN):
with tf.variable_scope("ICSTN", reuse=l > 0) as sc:
end_points_collection = sc.original_name_scope + '_end_points'
with slim.arg_scope([slim.conv2d, slim.conv2d_transpose],
normalizer_fn=slim.batch_norm,
weights_regularizer=slim.l2_regularizer(0.05),
normalizer_params=batch_norm_params,
activation_fn=tf.nn.relu,
outputs_collections=end_points_collection):
imageWarp = inverse_warp(
inputImage,
depth,
pM,
intrinsics,
intrinsics_inv)
imageWarpAll.append(imageWarp)
feat = tf.reshape(imageWarp, [batch_size, H, W, C])
print('feat shape:', feat.get_shape())
print('pM_ini:', pM.get_shape())
with tf.variable_scope("icnv1"):
feat = conv2Layer(opt, feat, 4)
feat = tf.nn.relu(feat)
with tf.variable_scope("icnv2"):
feat = conv2Layer(opt, feat, 8)
feat = tf.nn.relu(feat)
feat = tf.nn.max_pool(feat, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding="VALID")
feat = tf.reshape(feat, [opt.batch_size, -1])
with tf.variable_scope("fc3"):
feat = linearLayer(opt, feat, 48)
feat = tf.nn.relu(feat)
with tf.variable_scope("fc4"):
feat = linearLayer(opt, feat, 6, final=True)
dp = tf.reshape(feat, [-1, 6])
print('dp: ', dp.get_shape())
dpM = pose_vec2mat(dp)
pM = tf.matmul(dpM, pM)
imageWarp = inverse_warp(
inputImage,
depth,
pM,
intrinsics,
intrinsics_inv)
imageWarpAll.append(imageWarp)
return imageWarpAll, pM
def build_train_graph():
with tf.name_scope("cnn1"):...
with tf.name_scope("cnn2"):...
with tf.name_scope("Inverse Compositional CNN"):...
def train(self, opt):
with tf.variable_scope(tf.get_variable_scope()):
for i in range(opt.num_gpus):
print('gpu:', i)
with tf.device('/gpu:%d' % i):
self.build_train_graph(L_img_splits[i], R_img_splits[i], L_cam2pix_splits[i], L_pix2cam_splits[i],
R_cam2pix_splits[i], R_pix2cam_splits[i], L_sca_splits[i], R_sca_splits[i],
reuse_variables)
self.collect_summaries(i)
tower_losses.append(self.total_loss)
reuse_variables = True
grads = opt_step.compute_gradients(self.total_loss)
tower_grads.append(grads)
grads = average_gradients(tower_grads)
apply_gradient_op = opt_step.apply_gr`enter code here`adients(grads, global_step=global_step)
incr_global_step = tf.assign(global_step, global_step + 1)
total_loss = tf.reduce_mean(tower_losses)
tf.summary.scalar('learning_rate', learning_rate, ['model_0'])
tf.summary.scalar('total_loss', total_loss, ['model_0'])
summary_op = tf.summary.merge_all('model_0')
# self.collect_summaries()
# SESSION
config = tf.ConfigProto(allow_soft_placement=True)
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
# SAVER
summary_writer = tf.summary.FileWriter(
opt.checkpoint_dir + '/s%.1d_%.3d/' % (opt.seq_length, opt.img_height) + opt.model_name, sess.graph)
self.saver = tf.train.Saver()
# COUNT PARAM
total_num_parameters = 0
for variable in tf.trainable_variables():
total_num_parameters += np.array(variable.get_shape().as_list()).prod()
print("number of trainable parameters: {}".format(total_num_parameters))
# INIT
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
coordinator = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coordinator)
# LOAD CHECKPOINT IF SET
if opt.continue_train:
print("Resume training from previous checkpoint")
checkpoint = tf.train.latest_checkpoint(
os.path.join(opt.checkpoint_dir, 's%.1d_%.3d' % (opt.seq_length, opt.img_height), opt.model_name))
self.saver.restore(sess, checkpoint)
if opt.re_train:
sess.run(global_step.assign(0))
It's because of the for loop in the first part of the code that's possibly missing function name.
The loop tries to create the left_src_tgt_warp/ICSTN/icnv1/weight (same for icnv2 and so on):
def foo(num_layers):
opt = tf.placeholder(tf.float32, [None, 64])
for i in range(num_layers):
with tf.variable_scope("icnv1"):
feat = tf.layers.dense(opt, units=1, activation=tf.nn.relu)
foo(5)
ValueError: Variable icnv1/dense/kernel already exists, disallowed. Did you mean to set reuse=True
You need distinct names for Variables. One way to achieve this is like this:
def foo(num_layers):
opt = tf.placeholder(tf.float32, [None, 64])
for i in range(num_layers):
with tf.variable_scope("icnv1_layer_{}".format(i)):
feat = tf.layers.dense(opt, units=1, activation=tf.nn.relu)
we now have distinct names for each, icnv1_layer_1, icnv1_layer_2, etc. Depending on the depth.
Unless of course, you want a shared weights (e.g. it's the same layer, updates as one). In that case just set:
with tf.variable_scope("icnv1", reuse=tf.AUTO_REUSE):
I'm using Jupyter Notebooks to run my models and I recently realize that this error was occuring due to the fact that variables of my model was been saved on an "outter context". So when I restart kernel (and thus clean all my workspace variable) and run all cells the error gone away.
I am currently playing around and learning about distributed tensorflow.
I recently created a cluster with One GPU server(two cards) - One CPU server
I was browsing through various articles and in the TensorFlow distributed guide I saw that distribution happened across cards by explicitly calling them with names.
https://github.com/tensorflow/models/blob/master/tutorials/image/cifar10/cifar10_multi_gpu_train.py
but here no cluster are being created.
Can i create a TensorFlow cluster and then specify which card the code should run on?
If yes, does the below look correct?
In one github question who's link i dont have right now but the code below, the card is specified under with tf.device(replica_device_setter) but when i try to do that my code throws an error stating "Cannot assign a device for operation 'dummy_queue_Close_1': Could not satisfy explicit device specification '/job:ps/task:0/device:GPU:0' because no supported kernel for GPU devices is available."
Is this because i am assinging tasks which were supposed to happen on a CPU but instead as i gave with tf.device('/gpu:0/') it throws the error ?
Also I cant share my official code but it looks very similar to the below code which i took for reference.
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import numpy
import tensorflow as tf
tf.app.flags.DEFINE_string("ps_hosts", "localhost:2222", "...")
tf.app.flags.DEFINE_string("worker_hosts", "localhost:2223", "...")
tf.app.flags.DEFINE_string("job_name", "", "...")
tf.app.flags.DEFINE_integer("task_index", 0, "...")
tf.app.flags.DEFINE_integer('gpu_cards', 4, 'Number of GPU cards in a machine to use.')
FLAGS = tf.app.flags.FLAGS
def dense_to_one_hot(labels_dense, num_classes = 10) :
"""Convert class labels from scalars to one-hot vectors."""
num_labels = labels_dense.shape[0]
index_offset = numpy.arange(num_labels) * num_classes
labels_one_hot = numpy.zeros((num_labels, num_classes))
labels_one_hot.flat[index_offset + labels_dense.ravel()] = 1
return labels_one_hot
def run_training(server, cluster_spec, num_workers) :
is_chief = (FLAGS.task_index == 0)
with tf.Graph().as_default():
with tf.device(tf.train.replica_device_setter(cluster = cluster_spec)) :
with tf.device('/cpu:0') :
global_step = tf.get_variable('global_step', [],
initializer = tf.constant_initializer(0), trainable = False)
with tf.device('/gpu:%d' % (FLAGS.task_index % FLAGS.gpu_cards)) :
# Create the model
x = tf.placeholder("float", [None, 784])
W = tf.Variable(tf.zeros([784, 10]))
b = tf.Variable(tf.zeros([10]))
y = tf.nn.softmax(tf.matmul(x, W) + b)
# Define loss and optimizer
y_ = tf.placeholder("float", [None, 10])
cross_entropy = -tf.reduce_sum(y_ * tf.log(y))
opt = tf.train.GradientDescentOptimizer(0.01)
opt = tf.train.SyncReplicasOptimizer(opt, replicas_to_aggregate = num_workers,
replica_id = FLAGS.task_index, total_num_replicas = num_workers)
train_step = opt.minimize(cross_entropy, global_step = global_step)
# Test trained model
correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(y_, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))
init_token_op = opt.get_init_tokens_op()
chief_queue_runner = opt.get_chief_queue_runner()
init = tf.initialize_all_variables()
sv = tf.train.Supervisor(is_chief = is_chief,
init_op = init,
global_step = global_step)
# Create a session for running Ops on the Graph.
config = tf.ConfigProto(allow_soft_placement = True)
sess = sv.prepare_or_wait_for_session(server.target, config = config)
if is_chief:
sv.start_queue_runners(sess, [chief_queue_runner])
sess.run(init_token_op)
for i in range(100000):
source_data = numpy.random.normal(loc = 0.0, scale = 1.0, size = (100, 784))
labels_dense = numpy.clip(numpy.sum(source_data, axis = 1) / 5 + 5, 0, 9).astype(int)
labels_one_hot = dense_to_one_hot(labels_dense)
_, cost, acc, step = sess.run([train_step, cross_entropy, accuracy, global_step], feed_dict = { x: source_data, y_ : labels_one_hot })
print("[%d]: cost=%.2f, accuracy=%.2f" % (step, cost, acc))
def main(_) :
ps_hosts = FLAGS.ps_hosts.split(",")
worker_hosts = FLAGS.worker_hosts.split(",")
num_workers = len(worker_hosts)
print("gup_cards=%d; num_worders=%d" % (FLAGS.gpu_cards, num_workers))
cluster_spec = tf.train.ClusterSpec({ "ps":ps_hosts, "worker" : worker_hosts })
server = tf.train.Server(cluster_spec, job_name = FLAGS.job_name, task_index = FLAGS.task_index)
if FLAGS.job_name == "ps":
server.join()
elif FLAGS.job_name == "worker" :
run_training(server, cluster_spec, num_workers)
if __name__ == '__main__' :
tf.app.run()
I found a way to do this it sounds very simple, and it is very simple.
I created a TensorFlow cluster in the same way and passed the n_workers parameter to the cluster, and I called different instances of the code with an extra parameter for CUDA_VISIBLE_DEVICES.
CUDA_VISIBLE_DEVICES is an environment variable which can be used to restrict the vision of TensorFlow or any DL framework to a limited number of cards.
CUDA_VISIBLE_DEVICES value can range from -1 to n (where n is the number of GPUs).
-1 indicates no cards to use
n indicates nth card to use
I hope someone who is looking for a similar answer can find this useful.
I have been using TensorFlow for a reasonable length of time now. and believed I had a thorough understanding of how a TensorFlow graph works and executes within a session. However, I have written all of my TensorFlow models in a script-like fashion as such:
import tensorflow as tf
import DataWorker
import Constants
x = tf.placeholder(tf.float32, [None, Constants.sequenceLength, DataWorker.numFeatures])
y = tf.placeholder(tf.float32, [None, 1])
xTensors = tf.unstack(x, axis=1) # [seqLength tensors of shape (batchSize, numFeatures)]
W = tf.Variable(tf.random_normal([Constants.numHidden, 1])) # Weighted matrix
b = tf.Variable(tf.random_normal([1])) # Bias
cell = tf.contrib.rnn.BasicLSTMCell(Constants.numHidden, forget_bias=Constants.forgetBias)
outputs, finalState = tf.nn.static_rnn(cell, xTensors, dtype=tf.float32)
# predictions = [tf.add(tf.matmul(output, W), b) for output in outputs] # List of predictions after each time step
prediction = tf.add(tf.matmul(outputs[-1], W), b) # Prediction after final time step
prediction = tf.tanh(prediction) # Activation
mse = tf.losses.mean_squared_error(predictions=prediction, labels=y) # Mean loss over entire batch
accuracy = tf.reduce_mean(1 - (tf.abs(y - prediction) / DataWorker.labelRange)) # Accuracy over entire batch
optimiser = tf.train.AdamOptimizer(Constants.learningRate).minimize(mse) # Backpropagation
with tf.Session() as session:
session.run(tf.global_variables_initializer())
# #############################################
# TRAINING
# #############################################
for epoch in range(Constants.numEpochs):
print("***** EPOCH:", epoch + 1, "*****\n")
IDPointer, TSPointer = 0, 0 # Pointers to current ID and timestamp
epochComplete = False
batchNum = 0
while not epochComplete:
batchNum += 1
batchX, batchY, IDPointer, TSPointer, epochComplete = DataWorker.generateBatch(IDPointer, TSPointer, isTraining=True)
dict = {x: batchX, y: batchY}
session.run(optimiser, dict)
if batchNum % 1000 == 0 or epochComplete:
batchLoss = session.run(mse, dict)
batchAccuracy = session.run(accuracy, dict)
print("Iteration:", batchNum)
print(batchLoss)
print(str("%.2f" % (batchAccuracy * 100) + "%\n"))
# #############################################
# TESTING
# #############################################
testX, testY, _, _, _ = DataWorker.generateBatch(0, 0, isTraining=False)
testAccuracy = session.run(accuracy, {x: testX, y: testY})
print("Testing Accuracy:", str("%.2f" % (testAccuracy * 100) + "%"))
But now, for practicality and readability, I want to implement my model as a class, but have encountered many problems with initializing my variables, etc.
This is the closest I have got to implementing the above example using my own LSTM class
Model.py
import tensorflow as tf
import Constants
import DataWorker # Remove this dependency
class LSTM():
"""docstring."""
def __init__(self,
inputDimensionList,
outputDimensionList,
numLayers=Constants.numLayers,
numHidden=Constants.numHidden,
learningRate=Constants.learningRate,
forgetBias=Constants.forgetBias
):
"""docstring."""
self.batchInputs = tf.placeholder(tf.float32, [None] + inputDimensionList)
self.batchLabels = tf.placeholder(tf.float32, [None] + outputDimensionList)
self.weightedMatrix = tf.Variable(tf.random_normal([numHidden] + outputDimensionList))
self.biasMatrix = tf.Variable(tf.random_normal(outputDimensionList))
self.cell = tf.contrib.rnn.BasicLSTMCell(numHidden, forget_bias=forgetBias)
self.numLayers = numLayers
self.numHidden = numHidden
self.learningRate = learningRate
self.forgetBias = forgetBias
self.batchDict = {}
self.batchInputTensors = None
self.batchOutputs = None # All needed as instance variables?
self.batchFinalStates = None
self.batchPredictions = None
self.batchLoss = None
self.batchAccuracy = None
self.initialised = False
self.session = tf.Session()
# Take in activation, loss and optimiser FUNCTIONS as args
def execute(self, command):
"""docstring."""
return self.session.run(command, self.batchDict)
def setBatchDict(self, inputs, labels):
"""docstring."""
self.batchDict = {self.batchInputs: inputs, self.batchLabels: labels}
self.batchInputTensors = tf.unstack(self.batchInputs, axis=1)
def processBatch(self):
"""docstring."""
self.batchOutputs, self.batchFinalState = tf.nn.static_rnn(self.cell, self.batchInputTensors, dtype=tf.float32)
pred = tf.tanh(tf.add(tf.matmul(self.batchOutputs[-1], self.weightedMatrix), self.biasMatrix))
mse = tf.losses.mean_squared_error(predictions=pred, labels=self.batchLabels)
optimiser = tf.train.AdamOptimizer(self.learningRate).minimize(mse)
if not self.initialised:
self.session.run(tf.global_variables_initializer())
self.initialised = True
with tf.variable_scope("model") as scope:
if self.initialised:
scope.reuse_variables()
self.execute(optimiser)
self.batchPredictions = self.execute(pred)
self.batchLoss = self.execute(tf.losses.mean_squared_error(predictions=self.batchPredictions, labels=self.batchLabels))
self.batchAccuracy = self.execute(tf.reduce_mean(1 - (tf.abs(self.batchLabels - self.batchPredictions) / DataWorker.labelRange)))
return self.batchPredictions, self.batchLabels, self.batchLoss, self.batchAccuracy
def kill(self):
"""docstring."""
self.session.close()
This class is quite messy, especially processBatch() as I have just been trying to get it to work before refining it.
I then run my model here:
Main.py
import DataWorker
import Constants
from Model import LSTM
inputDim = [Constants.sequenceLength, DataWorker.numFeatures]
outputDim = [1]
lstm = LSTM(inputDimensionList=inputDim, outputDimensionList=outputDim)
# #############################################
# TRAINING
# #############################################
for epoch in range(Constants.numEpochs):
print("***** EPOCH:", epoch + 1, "*****\n")
IDPointer, TSPointer = 0, 0 # Pointers to current ID and timestamp
epochComplete = False
batchNum = 0
while not epochComplete:
batchNum += 1
batchX, batchY, IDPointer, TSPointer, epochComplete = DataWorker.generateBatch(IDPointer, TSPointer, isTraining=True)
lstm.setBatchDict(batchX, batchY)
batchPredictions, batchLabels, batchLoss, batchAccuracy = lstm.runBatch()
if batchNum % 1000 == 0 or epochComplete:
print("Iteration:", batchNum)
print("Pred:", batchPredictions[-1], "\tLabel:", batchLabels[-1])
print("Loss:", batchLoss)
print("Accuracy:", str("%.2f" % (batchAccuracy * 100) + "%\n"))
# #############################################
# TESTING
# #############################################
testX, testY, _, _, _ = DataWorker.generateBatch(0, 0, isTraining=False)
lstm.setBatchDict(testX, testY)
_, _, _, testAccuracy = lstm.runBatch()
print("Testing Accuracy:", str("%.2f" % (testAccuracy * 100) + "%"))
lstm.kill()
A single passthrough of the graph is executed fine, when all the variables are initialized, but it is on the second iteration where I get the error
ValueError: Variable rnn/basic_lstm_cell/kernel/Adam/ already exists, disallowed. Did you mean to set reuse=True in VarScope? Originally defined at:
optimiser = tf.train.AdamOptimizer(self.learningRate).minimize(mse)
I Googled this problem and learned that using scope.reuse_variables() should stop it trying to initialize the AdamOptimizer a second time, but cleary this isn't working how I have implemented it. How can I fix this issue?
As a side note, is my method of creating the TensorFlow session as an instance variable within my LSTM class acceptable, or should I create the session in Main and then pass it into the LSTM instance?
In general I wrap anything that creates variables under the hood with tf.make_template when doing object oriented model building.
However, you should avoid adding ops to the graph in a training loop, which looks like it's happening here. They will build up and cause problems, and likely give you incorrect results. Instead, define the graph (with inputs from tf.data, placeholders, or queues) and only loop over a session.run call. Even better, structure your code as an Estimator and this will be enforced.
This is part of my current python code for NN training in python using CNTK module
batch_axis = C.Axis.default_batch_axis()
input_seq_axis = C.Axis.default_dynamic_axis()
input_dynamic_axes = [batch_axis, input_seq_axis]
input_dynamic_axes2 = [batch_axis, input_seq_axis]
input = C.input_variable(n_ins, dynamic_axes=input_dynamic_axes, dtype=numpy.float32)
output = C.input_variable(n_outs, dynamic_axes=input_dynamic_axes2, dtype=numpy.float32)
dnn_model = cntk_model.create_model(input, hidden_layer_type, hidden_layer_size, n_outs)
loss = C.squared_error(dnn_model, output)
error = C.squared_error(dnn_model, output)
lr_schedule = C.learning_rate_schedule(current_finetune_lr, C.UnitType.minibatch)
momentum_schedule = C.momentum_schedule(current_momentum)
learner = C.adam(dnn_model.parameters, lr_schedule, momentum_schedule, unit_gain = False, l1_regularization_weight=l1_reg, l2_regularization_weight= l2_reg)
trainer = C.Trainer(dnn_model, (loss, error), [learner])
And here is code for creating NN model
def create_model(features, hidden_layer_type, hidden_layer_size, n_out):
logger.debug('Creating cntk model')
assert len(hidden_layer_size) == len(hidden_layer_type)
n_layers = len(hidden_layer_size)
my_layers = list()
for i in xrange(n_layers):
if(hidden_layer_type[i] == 'TANH'):
my_layers.append(C.layers.Dense(hidden_layer_size[i], activation=C.tanh, init=C.layers.glorot_uniform()))
elif (hidden_layer_type[i] == 'LSTM'):
my_layers.append(C.layers.Recurrence(C.layers.LSTM(hidden_layer_size[i])))
else:
raise Exception('Unknown hidden layer type')
my_layers.append(C.layers.Dense(n_out, activation=None))
my_model = C.layers.Sequential([my_layers])
my_model = my_model(features)
return my_model
Now, I would like to change a backpropagation, so when the error is calculated not direct network output is used, but the output after some additional calculation. I tried to define something like this
def create_error_function(self, prediction, target):
prediction_denorm = C.element_times(prediction, self.std_vector)
prediction_denorm = C.plus(prediction_denorm, self.mean_vector)
prediction_denorm_rounded = C.round(C.element_times(prediction_denorm[0:5], C.round(prediction_denorm[5])))
prediction_denorm_rounded = C.element_divide(prediction_denorm_rounded, C.round(prediction_denorm[5]))
prediction_norm = C.minus(prediction_denorm_rounded, self.mean_vector[0:5])
prediction_norm = C.element_divide(prediction_norm, self.std_vector[0:5])
first = C.squared_error(prediction_norm, target[0:5])
second = C.minus(C.round(prediction_denorm[5]), self.mean_vector[5])
second = C.element_divide(second, self.std_vector[5])
return C.plus(first, C.squared_error(second, target[5]))
and use it instead standard squared_error.
And the part for NN training
dnn_model = cntk_model.create_model(input, hidden_layer_type, hidden_layer_size, n_outs)
error_function = cntk_model.ErrorFunction(cmp_mean_vector, cmp_std_vector)
loss = error_function.create_error_function(dnn_model, output)
error = error_function.create_error_function(dnn_model, output)
lr_schedule = C.learning_rate_schedule(current_finetune_lr, C.UnitType.minibatch)
momentum_schedule = C.momentum_schedule(current_momentum)
learner = C.adam(dnn_model.parameters, lr_schedule, momentum_schedule, unit_gain = False, l1_regularization_weight=l1_reg,
l2_regularization_weight= l2_reg)
trainer = C.Trainer(dnn_model, (loss, error), [learner])
trainer.train_minibatch({input: temp_train_x, output: temp_train_y})
But after two epochs I start gettting always the same average loss, as my network is not learning
Every time you want to change how backprop works, you need to use stop_gradient. This is the only function whose gradient is different from the gradient of the operation of the forward pass. In the forward pass stop_gradient acts as identity. In the backward pass it blocks the gradient from propagating.
To do an operation f(x) on some x in the forward pass and pretend as if it never happened in the backward pass you need to do something like:
C.stop_gradient(f(x) - x) + x. In your case that would be
norm_features = C.stop_gradient(features/normalization - features) + features
I'm trying out Tensorflow's rnn example.
With some problems at the start I could run the example in order to train the ptb and now I have a model trained.
How do I exactly use the model now to create sentences without having to train every time again?
I'm running it with a command like python ptb_word_lm.py --data_path=/home/data/ --model medium --save_path=/home/medium
Is there a example somewhere on how to use the trained model to make sentences?
1.Add the following code at the last line of PTBModel:__init__() function:
self._output_probs = tf.nn.softmax(logits)
2.Add the following function in PTBModel:
#property
def output_probs(self):
return self._output_probs
3.Try to run the following code:
raw_data = reader.ptb_raw_data(FLAGS.data_path)
train_data, valid_data, test_data, vocabulary, word_to_id, id_to_word = raw_data
eval_config = get_config()
eval_config.batch_size = 1
eval_config.num_steps = 1
sess = tf.Session()
initializer = tf.random_uniform_initializer(-eval_config.init_scale,
eval_config.init_scale)
with tf.variable_scope("model", reuse=None, initializer=initializer):
mtest = PTBModel(is_training=False, config=eval_config)
sess.run(tf.initialize_all_variables())
saver = tf.train.Saver()
ckpt = tf.train.get_checkpoint_state('/home/medium') # __YOUR__MODEL__SAVE__PATH__
if ckpt and gfile.Exists(ckpt.model_checkpoint_path):
msg = 'Reading model parameters from %s' % ckpt.model_checkpoint_path
print(msg)
saver.restore(sess, ckpt.model_checkpoint_path)
def pick_from_weight(weight, pows=1.0):
weight = weight**pows
t = np.cumsum(weight)
s = np.sum(weight)
return int(np.searchsorted(t, np.random.rand(1) * s))
while True:
number_of_sentences = 10 # generate 10 sentences one time
sentence_cnt = 0
text = '\n'
end_of_sentence_char = word_to_id['<eos>']
input_char = np.array([[end_of_sentence_char]])
state = sess.run(mtest.initial_state)
while sentence_cnt < number_of_sentences:
feed_dict = {mtest.input_data: input_char,
mtest.initial_state: state}
probs, state = sess.run([mtest.output_probs, mtest.final_state],
feed_dict=feed_dict)
sampled_char = pick_from_weight(probs[0])
if sampled_char == end_of_sentence_char:
text += '.\n'
sentence_cnt += 1
else:
text += ' ' + id_to_word[sampled_char]
input_char = np.array([[sampled_char]])
print(text)
raw_input('press any key to continue ...')
This website has an answer with an alteration on the PTB word script, working currently with Tensorflow version 1.0
http://deeplearningathome.com/2016/10/Text-generation-using-deep-recurrent-neural-networks.html