Related
I spend days trying to figure out what is going on and I am still getting this error. here is the error I get
ValueError: Variable rnn/multi_rnn_cell/cell_1/basic_lstm_cell/weights
does not exist, or was not created with tf.get_variable(). Did you
mean to set reuse=None in VarScope?
And here is my sample code, does anyone know what I am doing wrong?
x = tf.placeholder(tf.float32,[None,n_steps,n_input])
y = tf.placeholder(tf.float32,[None,n_classes])
weights = {
'out': tf.Variable(tf.random_normal([n_hidden, n_classes]))
}
biases = {
'out': tf.Variable(tf.random_normal([n_classes]))
}
def RNN(x, weights, biases):
x = tf.unstack(x, n_steps, 1)
lstm_cell = rnn.MultiRNNCell([cell() for y in range(2)] , state_is_tuple=True)
# Get lstm cell output
outputs, states = rnn.static_rnn(lstm_cell, x, dtype=tf.float32)
# Linear activation, using rnn inner loop last output
return tf.matmul(outputs[-1], weights['out']) + biases['out']
def cell():
return rnn.BasicLSTMCell(n_hidden,forget_bias=0.1, reuse=True)
pred = RNN(x, weights, biases)
If you don't need to reuse the cell, just use the following,
def cell():
return rnn.BasicLSTMCell(n_hidden,forget_bias=0.1)
Else, If you need to reuse, you can follow this Reuse Reusing Variable of LSTM in Tensorflow post that has a nice explanation.
If you want to reuse the weights, then the easiest way is to create a single cell object and pass it multiple times to MultiRNNCell:
import tensorflow as tf
from tensorflow.contrib import rnn
n_steps = 20
n_input = 10
n_classes = 5
n_hidden = 15
x = tf.placeholder(tf.float32,[None,n_steps,n_input])
y = tf.placeholder(tf.float32,[None,n_classes])
weights = {
'in': tf.Variable(tf.random_normal([n_input, n_hidden])),
'out': tf.Variable(tf.random_normal([n_hidden, n_classes]))
}
biases = {
'in': tf.Variable(tf.random_normal([n_hidden])),
'out': tf.Variable(tf.random_normal([n_classes]))
}
def RNN(x, weights, biases):
# Initial input layer
inp = (tf.matmul(x, weights['in'][tf.newaxis, ...]) +
biases['in'][tf.newaxis, tf.newaxis, ...])
inp = tf.nn.sigmoid(inp)
inp = tf.unstack(inp, axis=-1)
my_cell = cell()
lstm_cell = rnn.MultiRNNCell([my_cell for y in range(2)], state_is_tuple=True)
# Get lstm cell output
outputs, states = rnn.static_rnn(lstm_cell, inp, dtype=tf.float32)
# Linear activation, using rnn inner loop last output
return tf.matmul(outputs[-1], weights['out']) + biases['out']
def cell():
return rnn.BasicLSTMCell(n_hidden,forget_bias=0.1)
pred = RNN(x, weights, biases)
However, you have to make sure that it makes sense to share the variables, dimension-wise, otherwise it will fail. In this case, I have added an additional layer before the LSTM cells to make sure every LSTM input is the same size.
I faced a problem with properly restoring the saved model in tensorflow. I created the Bidirectional RNN model in tensorflow with following code:
batchX_placeholder = tf.placeholder(tf.float32, [None, timesteps, 1],
name="batchX_placeholder")])
batchY_placeholder = tf.placeholder(tf.float32, [None, num_classes],
name="batchY_placeholder")
weights = tf.Variable(np.random.rand(2*STATE_SIZE, num_classes),
dtype=tf.float32, name="weights")
biases = tf.Variable(np.zeros((1, num_classes)), dtype=tf.float32,
name="biases")
logits = BiRNN(batchX_placeholder, weights, biases)
with tf.name_scope("prediction"):
prediction = tf.nn.softmax(logits)
loss_op = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(
logits=logits, labels=batchY_placeholder))
lr = tf.Variable(learning_rate, trainable=False, dtype=tf.float32,
name='lr')
optimizer = tf.train.AdamOptimizer(learning_rate=lr)
train_op = optimizer.minimize(loss_op)
init_op = tf.initialize_all_variables()
saver = tf.train.Saver()
The architecture of BiRNN created with the following function:
def BiRNN(x, weights, biases):
# Unstack to get a list of 'time_steps' tensors of shape (batch_size,
# num_input)
x = tf.unstack(x, time_steps, 1)
# Forward and Backward direction cells
lstm_fw_cell = rnn.BasicLSTMCell(STATE_SIZE, forget_bias=1.0)
lstm_bw_cell = rnn.BasicLSTMCell(STATE_SIZE, forget_bias=1.0)
outputs, _, _ = rnn.static_bidirectional_rnn(lstm_fw_cell,
lstm_bw_cell, x, dtype=tf.float32)
# Linear activation, using rnn inner loop last output
return tf.matmul(outputs[-1], weights) + biases
Then I train a model and save it after each 200 steps:
with tf.Session() as sess:
sess.run(init_op)
current_step = 0
for batch_x, batch_y in get_minibatch():
sess.run(train_op, feed_dict={batchX_placeholder: batch_x,
batchY_placeholder: batch_y})
current_step += 1
if current_step % 200 == 0:
saver.save(sess, os.path.join(model_dir, "model")
To run the saved model in inference mode I use saved tensorflow graph in "model.meta" file:
graph = tf.get_default_graph()
saver = tf.train.import_meta_graph(os.path.join(model_dir, "model.meta"))
sess = tf.Session()
saver.restore(sess, tf.train.latest_checkpoint(model_dir)
weights = graph.get_tensor_by_name("weights:0")
biases = graph.get_tensor_by_name("biases:0")
batchX_placeholder = graph.get_tensor_by_name("batchX_placeholder:0")
batchY_placeholder = graph.get_tensor_by_name("batchY_placeholder:0")
logits = BiRNN(batchX_placeholder, weights, biases)
prediction = graph.get_operation_by_name("prediction/Softmax")
argmax_pred = tf.argmax(prediction, 1)
init = tf.global_variables_initializer()
sess.run(init)
for x_seq, y_gt in get_sequence():
_, y_pred = sess.run([prediction, argmax_pred],
feed_dict={batchX_placeholder: [x_seq]],
batchY_placeholder: [[0.0, 0.0]]})
print("Y ground true: " + str(y_gt) + ", Y pred: " + str(y_pred[0]))
And when I run the code in inference mode, I get different results each time I launch it. It seems that output neurons from the softmax layer randomly bundled with different output classes.
So, my question is: How can I save and then correctly restore the model in tensorflow, so that all neurons properly bundled with corresponding output classes?
There is no need to call tf.global_variables_initializer(), I think that is your problem.
I removed some operations: logits, weights and biases since you don't need them, all those are already loaded, use graph.get_tensor_by_name to get them.
For the prediction, get the tensor instead of the operation. (see this answer):
This is the code:
graph = tf.get_default_graph()
saver = tf.train.import_meta_graph(os.path.join(model_dir, "model.meta"))
sess = tf.Session()
saver.restore(sess, tf.train.latest_checkpoint(model_dir))
batchX_placeholder = graph.get_tensor_by_name("batchX_placeholder:0")
batchY_placeholder = graph.get_tensor_by_name("batchY_placeholder:0")
prediction = graph.get_tensor_by_name("prediction/Softmax:0")
argmax_pred = tf.argmax(prediction, 1)
Edit 1: I notice that I wasn't clear on why you got different results.
And when I run the code in inference mode, I get different results
each time I launch it.
Notice that although you used the weights from the loaded model, you are creating the BiRNN again, and the BasicLSTMCell also have weights and other variables that you don't set from your loaded model, hence they need to be initialized (with new random values) resulting in an untrained model again.
I have built a system that leverages Google ML Engine to train various text classifiers using a simple flat CNN architecture (borrowed from the excellent WildML post). I've also leveraged heavily the ML Engine trainer template which exists here - specifically using the Tensorflow core functions.
My issue is that while the model trains and learns parameters correctly, I cannot get the serialized export in the binary SavedModel format (i.e. - the .pb files) to maintain the learned weights. I can tell this by using the gcloud predict local API on the model export folder and each time it makes randomized predictions - leading me to believe that while the graph structure is being saved to the proto-buf format, the associated weights in the checkpoint file are not being carried over.
Here's the code for my run function:
def run(...):
# ... code to load and transform train/test data
with train_graph.as_default():
with tf.Session(graph=train_graph).as_default() as session:
# Features and label tensors as read using filename queue
features, labels = model.input_fn(
x_train,
y_train,
num_epochs=num_epochs,
batch_size=train_batch_size
)
# Returns the training graph and global step tensor
tf.logging.info("Train vocab size: {:d}".format(vocab_size))
train_op, global_step_tensor, cnn, train_summaries = model.model_fn(
model.TRAIN,
sequence_length,
num_classes,
label_values,
vocab_size,
embedding_size,
filter_sizes,
num_filters
)
tf.logging.info("Created simple training CNN with ({}) filter types".format(filter_sizes))
# Setup writers
train_summary_op = tf.summary.merge(train_summaries)
train_summary_dir = os.path.join(job_dir, "summaries", "train")
# Generate writer
train_summary_writer = tf.summary.FileWriter(train_summary_dir, session.graph)
# Initialize all variables
session.run(tf.global_variables_initializer())
session.run(tf.local_variables_initializer())
model_dir = os.path.abspath(os.path.join(job_dir, "model"))
if not os.path.exists(model_dir):
os.makedirs(model_dir)
saver = tf.train.Saver()
def train_step(x_batch, y_batch):
"""
A single training step
"""
feed_dict = {
cnn.input_x: x_batch,
cnn.input_y: y_batch,
cnn.dropout_keep_prob: 0.5
}
step, _, loss, accuracy = session.run([global_step_tensor, train_op, cnn.loss, cnn.accuracy],
feed_dict=feed_dict)
time_str = datetime.datetime.now().isoformat()
if step % 10 == 0:
tf.logging.info("{}: step {}, loss {:g}, acc {:g}".format(time_str, step, loss, accuracy))
# Return current step
return step
def eval_step(x_batch, y_batch, train_step, total_steps):
"""
Evaluates model on a dev set
"""
feed_dict = {
cnn.input_x: x_batch,
cnn.input_y: y_batch,
cnn.dropout_keep_prob: 1.0
}
step, loss, accuracy, scores, predictions = session.run([global_step_tensor, cnn.loss, cnn.accuracy, cnn.scores, cnn.predictions],
feed_dict=feed_dict)
# Get metrics
y_actual = np.argmax(y_batch, 1)
model_metrics = precision_recall_fscore_support(y_actual, predictions)
#print(scores)
time_str = datetime.datetime.now().isoformat()
print("\n---- EVAULATION ----")
avg_precision = np.mean(model_metrics[0], axis=0)
avg_recall = np.mean(model_metrics[1], axis=0)
avg_f1 = np.mean(model_metrics[2], axis=0)
print("{}: step {}, loss {:g}, acc {:g}, prec {:g}, rec {:g}, f1 {:g}".format(time_str, step, loss, accuracy, avg_precision, avg_recall, avg_f1))
print("Model metrics: ", model_metrics)
print("---- EVALUATION ----\n")
# Generate batches
batches = data_helpers.batch_iter(
list(zip(features, labels)), train_batch_size, num_epochs)
# Training loop. For each batch...
for batch in batches:
x_batch, y_batch = zip(*batch)
current_step = train_step(x_batch, y_batch)
if current_step % 20 == 0 or current_step == 1:
eval_step(x_eval, y_eval, current_step, total_steps)
# Checkpoint directory. Tensorflow assumes this directory already exists so we need to create it
print(model_dir)
trained_model = saver.save(session, os.path.join(job_dir, 'model') + "/model.ckpt", global_step=current_step)
print(trained_model)
print("Saved final model checkpoint to {}".format(trained_model))
# Only perform this if chief
if is_chief:
build_and_run_exports(trained_model, job_dir,
model.SERVING_INPUT_FUNCTIONS[model.TEXT],
sequence_length, num_classes, label_values,
vocab_size, embedding_size, filter_sizes,
num_filters, vocab_processor)
And my build_and_run_exports function:
def build_and_run_exports(...):
# Check if we export already exists - if so delete
export_dir = os.path.join(job_dir, 'export')
if os.path.exists(export_dir):
print("Export currently exists - going to delete:", export_dir)
shutil.rmtree(export_dir)
# Create exporter
exporter = tf.saved_model.builder.SavedModelBuilder(export_dir)
# Restore prediction graph
prediction_graph = tf.Graph()
with prediction_graph.as_default():
with tf.Session(graph=prediction_graph) as session:
# Get training data
features, inputs_dict = serving_input_fn()
# Setup inputs
inputs_info = {
name: tf.saved_model.utils.build_tensor_info(tensor)
for name, tensor in inputs_dict.iteritems()
}
# Load model
cnn = TextCNN(
sequence_length=sequence_length,
num_classes=num_classes,
vocab_size=vocab_size,
embedding_size=embedding_size,
filter_sizes=list(map(int, filter_sizes.split(","))),
num_filters=num_filters,
input_tensor=features)
# Restore model
saver = tf.train.Saver()
saver.restore(session, latest_checkpoint)
# Setup outputs
outputs = {
'logits': cnn.scores,
'probabilities': cnn.probabilities,
'predicted_indices': cnn.predictions
}
# Create output info
output_info = {
name: tf.saved_model.utils.build_tensor_info(tensor)
for name, tensor in outputs.iteritems()
}
# Setup signature definition
signature_def = tf.saved_model.signature_def_utils.build_signature_def(
inputs=inputs_info,
outputs=output_info,
method_name=sig_constants.PREDICT_METHOD_NAME
)
# Create graph export
exporter.add_meta_graph_and_variables(
session,
tags=[tf.saved_model.tag_constants.SERVING],
signature_def_map={
sig_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY: signature_def
},
legacy_init_op=tf.saved_model.main_op.main_op()
)
# Export model
exporter.save()
And last, but not least, the TextCNN model:
class TextCNN(object):
"""
A CNN for text classification.
Uses an embedding layer, followed by a convolutional, max-pooling and softmax layer.
"""
def __init__(
self, sequence_length, num_classes, vocab_size,
embedding_size, filter_sizes, num_filters, l2_reg_lambda=0.0,
dropout_keep_prob=0.5, input_tensor=None):
# Setup input
if input_tensor != None:
self.input_x = input_tensor
self.dropout_keep_prob = tf.constant(1.0)
else:
self.input_x = tf.placeholder(tf.int32, [None, sequence_length], name="input_x")
self.dropout_keep_prob = tf.placeholder(tf.float32, name="dropout_keep_prob")
# Placeholders for input, output and dropout
self.input_y = tf.placeholder(tf.int32, [None, num_classes], name="input_y")
# Keeping track of l2 regularization loss (optional)
l2_loss = tf.constant(0.0)
# Embedding layer
with tf.device('/cpu:0'), tf.name_scope("embedding"):
self.W = tf.Variable(
tf.random_uniform([vocab_size, embedding_size], -1.0, 1.0),
name="W")
self.embedded_chars = tf.nn.embedding_lookup(self.W, self.input_x)
self.embedded_chars_expanded = tf.expand_dims(self.embedded_chars, -1)
# Create a convolution + maxpool layer for each filter size
pooled_outputs = []
for i, filter_size in enumerate(filter_sizes):
with tf.name_scope("conv-maxpool-%s" % filter_size):
# Convolution Layer
filter_shape = [filter_size, embedding_size, 1, num_filters]
W = tf.Variable(tf.truncated_normal(filter_shape, stddev=0.1), name="W")
b = tf.Variable(tf.constant(0.1, shape=[num_filters]), name="b")
conv = tf.nn.conv2d(
self.embedded_chars_expanded,
W,
strides=[1, 1, 1, 1],
padding="VALID",
name="conv")
# Apply nonlinearity
h = tf.nn.relu(tf.nn.bias_add(conv, b), name="relu")
# Maxpooling over the outputs
pooled = tf.nn.max_pool(
h,
ksize=[1, sequence_length - filter_size + 1, 1, 1],
strides=[1, 1, 1, 1],
padding='VALID',
name="pool")
pooled_outputs.append(pooled)
# Combine all the pooled features
num_filters_total = num_filters * len(filter_sizes)
self.h_pool = tf.concat(pooled_outputs, 3)
self.h_pool_flat = tf.reshape(self.h_pool, [-1, num_filters_total])
# Add dropout
with tf.name_scope("dropout"):
self.h_drop = tf.nn.dropout(self.h_pool_flat, self.dropout_keep_prob)
# Final (unnormalized) scores and predictions
with tf.name_scope("output"):
W = tf.get_variable(
"W",
shape=[num_filters_total, num_classes],
initializer=tf.contrib.layers.xavier_initializer())
b = tf.Variable(tf.constant(0.1, shape=[num_classes]), name="b")
l2_loss += tf.nn.l2_loss(W)
l2_loss += tf.nn.l2_loss(b)
self.scores = tf.nn.xw_plus_b(self.h_drop, W, b, name="scores")
self.predictions = tf.argmax(self.scores, 1, name="predictions")
# CalculateMean cross-entropy loss
with tf.name_scope("loss"):
losses = tf.nn.softmax_cross_entropy_with_logits(logits=self.scores, labels=self.input_y)
self.loss = tf.reduce_mean(losses) + l2_reg_lambda * l2_loss
with tf.name_scope("probabilities"):
self.probabilities = tf.nn.softmax(logits=self.scores)
# Accuracy
with tf.name_scope("accuracy"):
correct_predictions = tf.equal(self.predictions, tf.argmax(self.input_y, 1))
self.accuracy = tf.reduce_mean(tf.cast(correct_predictions, "float"), name="accuracy")
I'm hoping I'm just missing something simple in how I'm creating the TF graph / session and restoring stats.
Thank you in advance for your help!
This behavior is caused due to the behavior of tf.saved_model.main_op.main_op() which randomly initializes all of the variables in the graph (code). However, legacy_init_op happens after the variables are restored from the checkpoint (restore happens here followed by legacy_init_op here).
The solution is simply to not re-initialize all of the variables, for example, in your code:
from tensorflow.python.ops import variables
from tensorflow.python.ops import lookup_ops
from tensorflow.python.ops import control_flow_ops
def my_main_op():
init_local = variables.local_variables_initializer()
init_tables = lookup_ops.tables_initializer()
return control_flow_ops.group(init_local, init_tables)
def build_and_run_exports(...):
...
# Create graph export
exporter.add_meta_graph_and_variables(
session,
tags=[tf.saved_model.tag_constants.SERVING],
signature_def_map={
sig_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY: signature_def
},
legacy_init_op=my_main_op()
)
# Export model
exporter.save()
I have built a simple neural network to classify data into only 2 classes
Data is something like this
34.62365962451697,78.0246928153624,0
60.18259938620976,86.30855209546826,1
There are no zero values in data so there's no source of such cost.Cost is zero with adagrad optimiser and nan with gradient descent optimiser
Here's the code
import numpy as ny
import tensorflow as tf
def load():
data = []
for line in open("ex2data1.txt"):
row = line.split(',')
x = ny.array(row, dtype='|S4')
data.append(x.astype(ny.float64))
return ny.array(data)
def multilayer_perceptron(x, weights, biases):
# Hidden layer with ReLU activation
layer_1 = tf.add(tf.matmul(x, weights['h1']), biases['b1'])
layer_1 = tf.nn.relu(layer_1)
# Hidden layer with ReLU activation
layer_2 = tf.add(tf.matmul(layer_1, weights['h2']), biases['b2'])
layer_2 = tf.nn.relu(layer_2)
# Output layer with linear activation
out_layer = tf.matmul(layer_2, weights['out']) + biases['out']
return out_layer
# Store layers weight & bias
weights = {
'h1': tf.Variable(tf.random_normal([2, 15])),
'h2': tf.Variable(tf.random_normal([15, 15])),
'out': tf.Variable(tf.random_normal([15, 1]))
}
biases = {
'b1': tf.Variable(tf.random_normal([15])),
'b2': tf.Variable(tf.random_normal([15])),
'out': tf.Variable(tf.random_normal([1]))
}
data = load()
Xdata = ny.array(data[:, 0:2])
Ydata = ny.array(data[:, 2])
Ydata = ny.array(Ydata.reshape([100, 1]))
# Step 2 - Create input and output placeholders for data
X = tf.placeholder("float", [None, 2], name="X")
Y = tf.placeholder("float", [None, 1], name="Y")
pred = multilayer_perceptron(X, weights, biases)
# Minimize error using cross entropy
with tf.name_scope("cost"):
cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=pred, labels=Y))
optimizer = tf.train.AdagradOptimizer(0.001).minimize(cost)
tf.summary.scalar("cost", cost)
init = tf.global_variables_initializer()
summary_op = tf.summary.merge_all()
with tf.Session() as sess:
sess.run(init)
print(Xdata)
print(Ydata)
# Step 12 train the model
for i in range(1000):
sess.run(optimizer, feed_dict={X: Xdata, Y: Ydata})
if (i % 100 == 0):
print(sess.run(cost, feed_dict={X: Xdata, Y: Ydata}))
With the way your labels are represented you should not use this loss function. I think this is relevant
I'm trying LSTM sequence labeling. Before add embedding layer I put input data directly to LSTM layer. But it gives me GPU memory error even if batch size is 1.
max_length is 330, should I have to change model or adding embedding layer will work? I'm using Titan X GPU with 12 GB RAM.
# tf Graph input
x = tf.placeholder(tf.float32, [None, max_length, num_table])
y = tf.placeholder(tf.float32, [None, max_length, n_classes])
seqlen = tf.placeholder(tf.int32,[None])
# Define weights
weights = {
'out': tf.Variable(tf.random_normal([n_hidden, n_classes]))
}
biases = {
'out': tf.Variable(tf.random_normal([n_classes]))
}
def LSTM(x, seqlen, weights, biases):
# given input: (batch_size, n_step, feature_table )
# required : (n_step, batch_size, feature_table )
x = tf.unpack(tf.transpose(x,perm=[1,0,2]))
lstm_cell = tf.nn.rnn_cell.LSTMCell(n_hidden)
#lstm_cell = tf.nn.rnn_cell.DropoutWrapper(lstm_cell,keep_prob)
outputs, states = tf.nn.rnn(cell=lstm_cell,
dtype=tf.float32,
sequence_length=seqlen,
inputs=x)
# convert to (n_step, batch_size, n_classes)
temp = [tf.matmul(output,weights['out']) + biases['out'] for output in outputs]
# convert to (batch_size, n_step, n_classes)
temp = tf.transpose(tf.pack(temp),perm = [1,0,2])
return temp