print("testing_set size",len(testing_set)*(input_size/1.0e9))
print("making model")
#training_set = training_set[:5000]
#testing_set = testing_set[:1000]
#make_X_y_convnet(images,w,h,offset=None,size=None)
X_train,y_train = utils.make_X_y_convnet(training_set,w,h,offset=[offset_w,offset_h],size=[target_w,target_h])
X_test,y_test = utils.make_X_y_convnet(testing_set,w,h,offset=[offset_w,offset_h],size=[target_w,target_h])
#model = convnet.make_model_1(3,target_w,target_h,nb_filters = 16,nb_conv = 10,nb_classes=nb_classes,dropout=0.5)
model = convnet.make_model_3(3,target_w,target_h,nb_classes)
score,max_value, mean_value = convnet.train(model, X_train,X_test,y_train,y_test,nb_classes,batch_size,nb_epoch)
score = convnet.test(model,max_value, mean_value, X_test,y_test,nb_classes)
print('Cross Validation result at:', i, 'Test score:', score[0], 'Test accuracy:', score[1])
convnet.save_model(model,"model-convnet-{0}.json".format(i),"model-convnet-{0}.h5".format(i))
#for epoch in range(nb_epoch):
#print("processing epoch",epoch)
##process batches
#for start_index, end_index in make_batches(len(training_set), batch_size):
##print("processing batch",start_index, end_index)
def train():
with tf.Graph().as_default():
global_step = tf.Variable(0, trainable=False)
# Get images and labels.
images, labels = convnet.inputs()
# Build a Graph that computes the logits predictions from the
# inference model.
logits = convnet.inference(images)
# Calculate loss.
loss = convnet.loss(logits, labels)
# Build a Graph that trains the model with one batch of examples and
# updates the model parameters.
train_op = convnet.train(loss, global_step)
# Create a saver.
saver = tf.train.Saver(tf.all_variables())
# Build the summary operation based on the TF collection of Summaries.
summary_op = tf.merge_all_summaries()
# Build an initialization operation to run below.
init = tf.initialize_all_variables()
# Start running operations on the Graph.
sess = tf.Session(config=tf.ConfigProto(log_device_placement=FLAGS.log_device_placement))
sess.run(init)
# Start the queue runners.
tf.train.start_queue_runners(sess=sess)
summary_writer = tf.train.SummaryWriter(FLAGS.train_dir, sess.graph)
for step in xrange(FLAGS.max_steps):
start_time = time.time()
_, loss_value = sess.run([train_op, loss])
duration = time.time() - start_time
assert not np.isnan(loss_value), 'Model diverged with loss = NaN'
if step % 10 == 0:
num_examples_per_step = FLAGS.batch_size
examples_per_sec = num_examples_per_step / duration
sec_per_batch = float(duration)
format_str = ('%s: step %d, loss = %.2f (%.1f examples/sec; %.3f ''sec/batch)')
print (format_str % (datetime.now(), step, loss_value, examples_per_sec, sec_per_batch))
if step % 100 == 0:
summary_str = sess.run(summary_op)
summary_writer.add_summary(summary_str, step)
# Save the model checkpoint periodically.
if step % 1000 == 0 or (step + 1) == FLAGS.max_steps:
checkpoint_path = os.path.join(FLAGS.train_dir, 'model.ckpt')
saver.save(sess, checkpoint_path, global_step=step)
if __name__ == '__main__':
train()
target = target[index]
n = len(data)
data.resize(n, 28 * 28)
target.resize(n, 1)
y = np.zeros((n, 10))
for i in range(n):
y[i][int(target[i])] = 1
alpha = 0.00001
k = 10000
batch = 600
data.resize(n, 1, 28, 28)
data = data * 1.01
#data = data / (np.std(data, axis = 0) + 1.01)
print (data)
model = convnet.modelPre()
for i in range(k):
pos = i % batch
batch_index = range(pos * 100, pos * 100 + 100, 1)
data_batch = data[batch_index]
target_batch = target[batch_index]
y_batch = y[batch_index]
cost, rate = convnet.train(data_batch, target_batch, y_batch, model)
print ("iteration : ", i , "rate : ", rate, "cost: " , cost)
while (go):
min_val_err = 20
for epoch in range(num_epochs):
# Full pass over training data:
train_err = 0
train_batches = 0
start_time = time.time()
for batch in iterate_minibatches(x_train, labels_train,
oneshot_indices_train,
oneshot_class, batch_size,
shuffle=True):
inputs, targets = batch
print('\rTraining phase {:6.1f}%'.format(train_batches * batch_size / train_size * 100), end="");sys.stdout.flush()
train_err += convnet.train(inputs, targets)
train_batches += 1
print('\rTraining phase {:6.1f}%'.format(100))
# Full pass over validation data:
val_err = 0
val_acc = 0
val_batches = 0
for batch in iterate_minibatches(x_validate, labels_validate,
oneshot_indices_validate,
oneshot_class, batch_size,
shuffle=False):
inputs, targets = batch
print('\rValidation phase {:6.1f}%'.format(val_batches * batch_size / validate_size * 100), end="");sys.stdout.flush()
err, acc = convnet.validate(inputs, targets)
val_err += err
def train():
with tf.Graph().as_default():
global_step = tf.Variable(0, trainable=False)
# Get images and labels.
images, labels = convnet.inputs(eval_data=False)
# Build a Graph that computes the logits predictions from the
# inference model.
logits = convnet.inference(images)
# Calculate loss.
loss = convnet.loss(logits, labels)
# Build a Graph that trains the model with one batch of examples and
# updates the model parameters.
train_op = convnet.train(loss, global_step)
# Create a saver.
saver = tf.train.Saver(tf.all_variables())
# Build the summary operation based on the TF collection of Summaries.
summary_op = tf.merge_all_summaries()
# Build an initialization operation to run below.
init = tf.initialize_all_variables()
# Start running operations on the Graph.
sess = tf.Session(config=tf.ConfigProto(
log_device_placement=FLAGS.log_device_placement))
sess.run(init)
# Load previously stored model from checkpoint
ckpt = tf.train.get_checkpoint_state(FLAGS.train_dir)
if ckpt and ckpt.model_checkpoint_path:
# Restores from checkpoint
saver.restore(sess, ckpt.model_checkpoint_path)
# Assuming model_checkpoint_path looks something like:
# /my-favorite-path/cifar10_train/model.ckpt-0,
# extract global_step from it.
global_step = ckpt.model_checkpoint_path.split('/')[-1].split(
'-')[-1]
print("Loading from checkpoint.Global step %s" % global_step)
else:
print("No checkpoint file found...Creating a new model...")
stepfile = "/home/soms/EmotionMusic/Model1/stepfile.txt"
if not os.path.exists(stepfile):
print("No step file found.")
step = 0
else:
f = open(stepfile, "r")
step = int(f.readlines()[0])
print("Step file step %d" % step)
# Start the queue runners.
tf.train.start_queue_runners(sess=sess)
summary_writer = tf.train.SummaryWriter(FLAGS.train_dir, sess.graph)
while step < FLAGS.max_steps:
start_time = time.time()
_, loss_value = sess.run([train_op, loss])
duration = time.time() - start_time
def signal_handler(signal, frame):
f = open(stepfile, 'w')
f.write(str(step))
print("Step file written to.")
sys.exit(0)
signal.signal(signal.SIGINT, signal_handler)
assert not np.isnan(loss_value), 'Model diverged with loss = NaN'
if step % 10 == 0:
num_examples_per_step = FLAGS.batch_size
examples_per_sec = num_examples_per_step / duration
sec_per_batch = float(duration)
format_str = (
'%s: step %d, loss = %.2f (%.1f examples/sec; %.3f '
'sec/batch)')
print(format_str % (datetime.now(), step, loss_value,
examples_per_sec, sec_per_batch))
if step % 100 == 0:
summary_str = sess.run(summary_op)
summary_writer.add_summary(summary_str, step)
# Save the model checkpoint periodically.
if step % 500 == 0 or (step + 1) == FLAGS.max_steps:
checkpoint_path = os.path.join(FLAGS.train_dir, 'model.ckpt')
saver.save(sess, checkpoint_path, global_step=step)
step += 1
modelcheckpoint = SaverCallback(i+1)
#training_set = training_set[:5000]
#testing_set = testing_set[:1000]
#make_X_y_convnet(images,w,h,offset=None,size=None)
#model = convnet.make_model_1(3,target_w,target_h,nb_filters = 16,nb_conv = 10,nb_classes=nb_classes,dropout=0.5)
print("making model")
model = convnet.make_window_model((3,target_h,target_w),nb_classes=nb_classes)
score = convnet.train(model, X_train,X_test,y_train,y_test,nb_classes,batch_size,nb_epoch,callbacks=[modelcheckpoint],generator=None)
#print('After training at:', i, 'Test score:', score[0], 'Test accuracy:', score[1])
#score = convnet.test(model,max_value,mean_value,X_test,y_test,nb_classes)
print('Cross Validation result at:', i, 'Test score:', score[0], 'Test accuracy:', score[1])
convnet.save_model(model,"model-convnet-{0}.json".format(i),"model-convnet-{0}.h5".format(i))
#print("mean",mean_value,"max",max_value)
#for epoch in range(nb_epoch):
#print("processing epoch",epoch)
##process batches
#for start_index, end_index in make_batches(len(training_set), batch_size):
##print("processing batch",start_index, end_index)
alpha = 0.0001
k = 100000
model = convnet.modelPre()
#sgd with momentum
mu = 0.1
v = {}
for key in model:
v[key] = np.zeros_like(model[key])
for i in range(k):
#spark driver is acting as parameter server -- gather gradients then broadcast updated weights
gradientsByPartition = data.map(
lambda x: convnet.train(x, i, model, alpha))
gradients, cost, rate = gradientsByPartition.reduce(add)
for key in model:
if (key.startswith("b")):
model[key] -= alpha * gradients[key]
else:
v[key] = mu * v[key] - alpha * (gradients[key] +
decayRate * model[key])
model[key] += v[key]
rate_test = 0.00
#estimate test rate
if (i % 300 == 1):
count = testData.map(
lambda x: convnet.getTestRightCount(x, model)).reduce(addNum)
