def test_1(self):
# single passthrough
(train_data_node,
train_labels_node,
validation_data_node,
test_data_node,
# predictions
train_prediction,
validation_prediction,
test_prediction,
# weights
conv1_weights,
conv2_weights,
fc1_weights,
fc2_weights,
# training
optimizer,
loss,
learning_rate,
summaries) = mnist.build_model()
with tf.Session() as s:
tf.initialize_all_variables().run()
print('Variables initialized')
step=0
with np.load(data_file) as data:
feed_dict = {
train_data_node: data['batch_data'],
train_labels_node: data['batch_labels']
}
evals=[train_prediction]
status,result=tdb.debug(evals, feed_dict=feed_dict, breakpoints=None, break_immediately=False, session=s)
self.assertEqual(status,tdb.FINISHED)
def test_1(self):
# single passthrough
(train_data_node,
train_labels_node,
validation_data_node,
test_data_node,
# predictions
train_prediction,
validation_prediction,
test_prediction,
# weights
conv1_weights,
conv2_weights,
fc1_weights,
fc2_weights,
# training
optimizer,
loss,
learning_rate,
summaries) = mnist.build_model()
with tf.Session() as s:
tf.initialize_all_variables().run()
print('Variables initialized')
step=0
with np.load("mnist_0.npz") as data:
feed_dict = {
train_data_node: data['batch_data'],
train_labels_node: data['batch_labels']
}
evals=[train_prediction]
status,result=tdb.debug(evals, feed_dict=feed_dict, breakpoints=None, break_immediately=False, session=s)
self.assertEqual(status,tdb.FINISHED)
def test_2(self):
"""
mnist with plotting
"""
(train_data_node,
train_labels_node,
validation_data_node,
test_data_node,
# predictions
train_prediction,
validation_prediction,
test_prediction,
# weights
conv1_weights,
conv2_weights,
fc1_weights,
fc2_weights,
# training
optimizer,
loss,
learning_rate,
summaries) = mnist.build_model()
s=tf.InteractiveSession()
tf.initialize_all_variables().run()
# use the same input every time for this test
with np.load(data_file) as data:
a=data['batch_data']
b=data['batch_labels']
feed_dict = {
train_data_node: a,
train_labels_node: b
}
# pdb.set_trace()
# result=s.run(optimizer,feed_dict)
# pdb.set_trace()
# tmp
# return
evals=[optimizer,loss,train_prediction,conv1_weights,conv2_weights,fc1_weights,fc2_weights]
# define some plotting functions
# use one debugSession per run
# attach plot nodes
g=tf.get_default_graph()
p1=tdb.plot_op(viz.viz_conv_weights,inputs=[g.as_graph_element(conv1_weights)])
p2=tdb.plot_op(viz.viz_conv_weights,inputs=[g.as_graph_element(conv2_weights)])
p3=tdb.plot_op(viz.viz_fc_weights,inputs=[g.as_graph_element(fc1_weights)])
p4=tdb.plot_op(viz.viz_fc_weights,inputs=[g.as_graph_element(fc2_weights)])
# get the plot op by name and
evals=[optimizer, loss, learning_rate, train_prediction, p1,p2,p3,p4]
status,result=tdb.debug(evals, feed_dict=feed_dict, session=s)
def test_2(self):
"""
mnist with plotting
"""
(train_data_node,
train_labels_node,
validation_data_node,
test_data_node,
# predictions
train_prediction,
validation_prediction,
test_prediction,
# weights
conv1_weights,
conv2_weights,
fc1_weights,
fc2_weights,
# training
optimizer,
loss,
learning_rate,
summaries) = mnist.build_model()
s=tf.InteractiveSession()
tf.initialize_all_variables().run()
# use the same input every time for this test
with np.load("mnist_0.npz") as data:
a=data['batch_data']
b=data['batch_labels']
feed_dict = {
train_data_node: a,
train_labels_node: b
}
# pdb.set_trace()
# result=s.run(optimizer,feed_dict)
# pdb.set_trace()
# tmp
# return
evals=[optimizer,loss,train_prediction,conv1_weights,conv2_weights,fc1_weights,fc2_weights]
# define some plotting functions
# use one debugSession per run
# attach plot nodes
g=tf.get_default_graph()
p1=tdb.plot_op(viz.viz_conv_weights,inputs=[g.as_graph_element(conv1_weights)])
p2=tdb.plot_op(viz.viz_conv_weights,inputs=[g.as_graph_element(conv2_weights)])
p3=tdb.plot_op(viz.viz_fc_weights,inputs=[g.as_graph_element(fc1_weights)])
p4=tdb.plot_op(viz.viz_fc_weights,inputs=[g.as_graph_element(fc2_weights)])
# get the plot op by name and
evals=[optimizer, loss, learning_rate, train_prediction, p1,p2,p3,p4]
status,result=tdb.debug(evals, feed_dict=feed_dict, session=s)
def test_2(self): """ See TestDebuggingTF.test_2 """ a,b,c,c2,d=build_graph_ht() status,result=tdb.debug([c], feed_dict=None, breakpoints=None, break_immediately=True) self.assertEqual(status, tdb.PAUSED) status,result=tdb.c() # continue self.assertEqual(status, tdb.FINISHED) self.assertEqual(result[0],5) # check that c = 5
def test_1(self): """ ht->tf """ a=tf.constant(2) b=tf.constant(3) c=tdb.python_op(myadd,inputs=[a,b],outputs=[tf.placeholder(tf.int32)]) # a+b d=tf.neg(c) status,result=tdb.debug([d], feed_dict=None, breakpoints=None, break_immediately=False) self.assertEqual(status, tdb.FINISHED) self.assertEqual(result[0],-5)
def test_3(self): """ with breakpoints """ # construct TensorFlow graph as usual a,b,c,c2,d=build_graph_tf() status,result=tdb.debug(d, feed_dict=None, breakpoints=[c], break_immediately=False) self.assertEqual(status, tdb.PAUSED) self.assertEqual(result, None) status,result=tdb.c() self.assertEqual(status, tdb.FINISHED) self.assertEqual(result[0],-5)
def test_2(self): """ single eval of the pentultimate node breka immediately. verify that the execution order does NOT contain d or c2 """ a,b,c,c2,d=build_graph_tf() status,result=tdb.debug([c], feed_dict=None, breakpoints=None, break_immediately=True) self.assertEqual(status, tdb.PAUSED) status,result=tdb.c() # continue self.assertEqual(status, tdb.FINISHED) self.assertEqual(result[0],5) # check that c = 5
def test_2(self): """ tf -> ht """ a=tf.constant(2) b=tf.constant(3) c=tf.add(a,b) d=tdb.python_op(myneg,inputs=[c],outputs=[tf.placeholder(tf.int32)]) status,result=tdb.debug([d], feed_dict=None, breakpoints=None, break_immediately=False) self.assertEqual(status, tdb.FINISHED) self.assertEqual(result[0],-5) self.assertEqual(tdb.get_value(d),-5) self.assertEqual(tdb.get_value(d.name),-5)
def test_1(self): """ See TestDebuggingTF.test_1 """ # construct TensorFlow graph as usual a,b,c,c2,d=build_graph_ht() evals=[a,b,c,c2,d] status,result=tdb.debug(evals, feed_dict=None, breakpoints=None, break_immediately=False) self.assertEqual(status, tdb.FINISHED) self.assertEqual(result[0],2) # a = 2 self.assertEqual(result[1],3) # b = 3 self.assertEqual(result[2],5) # c = 5 self.assertEqual(result[4],-5) # c2 = -5 self.assertEqual(result[3],6) # d = 6
def test_3(self):
"""
with breakpoints
"""
# construct TensorFlow graph as usual
a, b, c, c2, d = build_graph_tf()
status, result = tdb.debug(d,
feed_dict=None,
breakpoints=[c],
break_immediately=False)
self.assertEqual(status, tdb.PAUSED)
self.assertEqual(result, None)
status, result = tdb.c()
self.assertEqual(status, tdb.FINISHED)
self.assertEqual(result[0], -5)
def test_2(self):
"""
single eval of the pentultimate node
breka immediately.
verify that the execution order does NOT contain d or c2
"""
a, b, c, c2, d = build_graph_tf()
status, result = tdb.debug([c],
feed_dict=None,
breakpoints=None,
break_immediately=True)
self.assertEqual(status, tdb.PAUSED)
status, result = tdb.c() # continue
self.assertEqual(status, tdb.FINISHED)
self.assertEqual(result[0], 5) # check that c = 5
def test_1(self): """ test debugging of a pure TensorFlow graph no breakpoints, all nodes evaluated this should automatically build an InteractiveSession for us and create a HyperTree """ # construct TensorFlow graph as usual a,b,c,c2,d=build_graph_tf() evals=[a,b,c,c2,d] status,result=tdb.debug(evals, feed_dict=None, breakpoints=None, break_immediately=False) self.assertEqual(status, tdb.FINISHED) self.assertEqual(result[0],2) # a = 2 self.assertEqual(result[1],3) # b = 3 self.assertEqual(result[2],5) # c = 5 self.assertEqual(result[4],-5) # c2 = 6 self.assertEqual(result[3],6) # d = -5
def test_1(self):
"""
ht->tf
"""
a = tf.constant(2)
b = tf.constant(3)
c = tdb.python_op(myadd,
inputs=[a, b],
outputs=[tf.placeholder(tf.int32)]) # a+b
d = tf.neg(c)
status, result = tdb.debug([d],
feed_dict=None,
breakpoints=None,
break_immediately=False)
self.assertEqual(status, tdb.FINISHED)
self.assertEqual(result[0], -5)
def test_2(self):
"""
tf -> ht
"""
a = tf.constant(2)
b = tf.constant(3)
c = tf.add(a, b)
d = tdb.python_op(myneg,
inputs=[c],
outputs=[tf.placeholder(tf.int32)])
status, result = tdb.debug([d],
feed_dict=None,
breakpoints=None,
break_immediately=False)
self.assertEqual(status, tdb.FINISHED)
self.assertEqual(result[0], -5)
self.assertEqual(tdb.get_value(d), -5)
self.assertEqual(tdb.get_value(d.name), -5)
def test_1(self):
"""
test debugging of a pure TensorFlow graph
no breakpoints, all nodes evaluated
this should automatically build an InteractiveSession for us and create a HyperTree
"""
# construct TensorFlow graph as usual
a, b, c, c2, d = build_graph_tf()
evals = [a, b, c, c2, d]
status, result = tdb.debug(evals,
feed_dict=None,
breakpoints=None,
break_immediately=False)
self.assertEqual(status, tdb.FINISHED)
self.assertEqual(result[0], 2) # a = 2
self.assertEqual(result[1], 3) # b = 3
self.assertEqual(result[2], 5) # c = 5
self.assertEqual(result[4], -5) # c2 = 6
self.assertEqual(result[3], 6) # d = -5
# for f in files:
# print(f)
# urllib.urlretrieve(base_url+f, download_dir+f)
# return the TF nodes corresponding to graph input placeholders
(train_data, train_labels, validation_data, validation_labels, test_data,
test_labels) = mnist.get_data(download_dir)
# start the TensorFlow session that will be used to evaluate the graph
s = tf.InteractiveSession()
tf.initialize_all_variables().run()
BATCH_SIZE = 64
NUM_EPOCHS = 5
TRAIN_SIZE = 10000
for step in range(NUM_EPOCHS * TRAIN_SIZE // BATCH_SIZE):
offset = (step * BATCH_SIZE) % (TRAIN_SIZE - BATCH_SIZE)
batch_data = train_data[offset:(offset + BATCH_SIZE), :, :, :]
batch_labels = train_labels[offset:(offset + BATCH_SIZE)]
feed_dict = {train_data_node: batch_data, train_labels_node: batch_labels}
# run training node and visualization node
status, result = tdb.debug([optimizer, p0], feed_dict=feed_dict, session=s)
if step % 10 == 0:
status, result = tdb.debug([loss, p1, p2, p3, p4, ploss],
feed_dict=feed_dict,
breakpoints=None,
break_immediately=False,
session=s)
print('loss: %f' % (result[0]))