def test_base_fcn(net_name, img_fnames, weight_fname):
"Quick effort to test the base fcns and figure out what preprocessing is appropriate--isn't well documented."
inputs = tf.placeholder(tf.float32, [None, None, None, 3])
inputs = _vgg_preprocess(inputs)
if net_name == 'resnet_v1_152':
with slim.arg_scope(resnet_v1.resnet_arg_scope()) as scope:
base_net, _ = resnet_v1.resnet_v1_152(inputs,
is_training=False,
num_classes=1000)
else:
raise Exception('net_name not recognized.')
pred = tf.argmax(base_net, -1)
saver = tf.train.Saver()
sess = tf.InteractiveSession()
#tf.global_variables_initializer().run()
saver.restore(sess, weight_fname)
for img_fname in img_fnames:
img = cv2.imread(img_fname)
prediction = sess.run(pred,
feed_dict={inputs: img[np.newaxis, :, :, :]})
print('pred for %s:' % img_fname)
print(imagenet_names[int(prediction)])
print('=========')
def _load_tf_model(checkpoint_file):
# Placeholder
input_tensor = tf.placeholder(tf.float32,
shape=(None, 224, 224, 3),
name='input_image')
# Load the model
sess = tf.Session()
arg_scope = resnet_v1.resnet_arg_scope()
with tf.contrib.slim.arg_scope(arg_scope):
logits, _ = resnet_v1.resnet_v1_152(input_tensor,
num_classes=1000,
is_training=False,
reuse=tf.AUTO_REUSE)
probabilities = tf.nn.softmax(logits)
saver = tf.train.Saver()
saver.restore(sess, checkpoint_file)
def predict_for(image):
pred, pred_proba = sess.run([logits, probabilities],
feed_dict={input_tensor: image})
return pred_proba
return predict_for
def _load_tf_model(checkpoint_file):
# Placeholder
input_tensor = tf.compat.v1.placeholder(tf.float32,
shape=(None, 224, 224, 3),
name='input_image')
# Make the Tensorflow warnings go away
# Load the model
sess = tf.compat.v1.Session()
arg_scope = resnet_v1.resnet_arg_scope()
with tf.contrib.slim.arg_scope(arg_scope):
logits, _ = resnet_v1.resnet_v1_152(input_tensor,
num_classes=1000,
is_training=False,
reuse=tf.AUTO_REUSE)
probabilities = tf.nn.softmax(logits)
saver = tf.train.Saver()
try:
saver.restore(sess, checkpoint_file)
except ValueError:
raise mlhub.utils.DataResourceNotFoundException(checkpoint_file)
return sess, logits, probabilities, input_tensor
def _build_graph(self, inputs):
orig_image = inputs[0]
mean = tf.get_variable('resnet_v1_'+str(args.depth)+'/mean_rgb', shape=[3])
with tp.symbolic_functions.guided_relu():
with slim.arg_scope(resnet_v1.resnet_arg_scope(is_training=False)):
image = tf.expand_dims(orig_image - mean, 0)
if args.depth == 50:
logits, _ = resnet_v1.resnet_v1_50(image, 1000)
elif args.depth == 101:
logits, _ = resnet_v1.resnet_v1_101(image, 1000)
else:
logits, _ = resnet_v1.resnet_v1_152(image, 1000)
tp.symbolic_functions.saliency_map(logits, orig_image, name="saliency")
def _build_base_net(inputs, net_opts, is_training, reuse=None):
"return a net from slim research nets. is_training must be an actual boolean, not a tensor, which is why we're abstracting this out."
if net_opts['base_net'] == 'resnet_v1_152':
with slim.arg_scope(resnet_v1.resnet_arg_scope()) as scope:
inputs = _vgg_preprocess(inputs)
base_net, _ = resnet_v1.resnet_v1_152(
inputs,
is_training=False
if net_opts['is_batchnorm_fixed'] else is_training,
global_pool=False,
reuse=reuse)
if DEBUG:
print('resnet_out:')
print(base_net.shape.as_list())
elif net_opts['base_net'] == 'resnet_v2_152':
with slim.arg_scope(resnet_v2.resnet_arg_scope()) as scope:
inputs = _vgg_preprocess(inputs)
base_net, _ = resnet_v2.resnet_v2_152(
inputs,
is_training=False
if net_opts['is_batchnorm_fixed'] else is_training,
global_pool=False,
reuse=reuse)
if DEBUG:
print('resnet_out:')
print(base_net.shape.as_list())
elif net_opts['base_net'] == 'inception_v3':
#WARNING: is_train for inception controls not just batch norm but dropout. So it's a little awkward. We may need more functionality here later TODO
'''
TODO add inception preprocessing: https://github.com/tensorflow/models/blob/master/research/slim/preprocessing/preprocessing_factory.py
#WARNING untested. Not sure I fully understand slim scopes yet
base_scope = 'InceptionV3'
with slim.arg_scope(inception.inception_v3_arg_scope()) as scope:
with slim.variable_scope(scope, base_scope, [inputs, None], reuse=False) as scope:
with slim.arg_scope([layers_lib.batch_norm, layers_lib.dropout], is_training=is_training):
base_net, _ = inception_v3_base(inputs,scope=scope)
'''
elif net_opts['base_net'] == 'nothing':
# a nothing for debugging purposes
base_net = inputs
else:
raise Exception("basenet name not recognized")
return base_net
def _get_endpoints(model_name, img_tensor):
if model_name == "res50":
with slim.arg_scope(resnet_v1.resnet_arg_scope()):
_, end_points = resnet_v1.resnet_v1_50(img_tensor,
1000,
is_training=False)
return end_points["predictions"]
elif model_name == "res152":
with slim.arg_scope(resnet_v1.resnet_arg_scope()):
_, end_points = resnet_v1.resnet_v1_152(img_tensor,
1000,
is_training=False)
return end_points["predictions"]
elif model_name.startswith("mobilenet"):
with tf.contrib.slim.arg_scope(
mobilenet_v2.training_scope(is_training=False)):
_, endpoints = mobilenet_v2.mobilenet(img_tensor)
return endpoints["Predictions"]
def main(_): reader = init_reader(FLAGS.data_dir) batch_xs, batch_ys = reader.make_one_shot_iterator().get_next() # param batch_xs: shape [-1, 512, 512, 3] type tf.float32 # param batch_ys: shape [-1] type tf.int32 off_ws = [0, 0, 0, 0, 256, 256, 256, 256] off_hs = [0, 128, 256, 384, 0, 128, 256, 384] x_img_cuts = [tf.image.crop_to_bounding_box(batch_xs, hs, ws, 128, 256)\ for hs, ws in zip(off_hs, off_ws)] batch_xs = tf.reshape(tf.concat(x_img_cuts, axis=0), [FLAGS.batch_size*8, 128, 256, 3]) batch_ys = tf.reshape(batch_ys, [FLAGS.batch_size * 8]) with slim.arg_scope(resnet_v1.resnet_arg_scope()): probs, end_points = resnet_v1.resnet_v1_152(batch_xs, num_classes=6, is_training=False) probs = tf.reshape(probs, [-1, 6]) prediction = tf.argmax(probs, axis=-1) loss, binary = init_loss(probs, batch_ys, end_points, losstype=FLAGS.losstype) config = tf.ConfigProto() config.gpu_options.allow_growth = True sess = tf.InteractiveSession(config=config) counter = get_counter(FLAGS.model_dir) learning_rate = tf.train.exponential_decay(FLAGS.learning_rate, counter, 100, 0.98, staircase=True) if 'SGD' in FLAGS.optimizer: optim = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss, global_step=tf.Variable(counter)) elif 'Adam' in FLAGS.optimizer: optim = tf.train.AdamOptimizer(learning_rate).minimize(loss, global_step=tf.Variable(counter)) else: raise NotImplementedError # Initialize all variables # Load the pretrained model and initialize the global counter sess.run(tf.global_variables_initializer()) counter = load(sess, FLAGS.model_basedir) confusion_matrix = Confusion(sess, prediction, binary, batch_ys) print(confusion_matrix)
def main(_):
reader = init_reader(FLAGS.data_dir, batch_size=FLAGS.batch_size)
batch_xs, batch_ys = reader.make_one_shot_iterator().get_next()
# param batch_xs: shape [-1, 512, 512, 3] type tf.float32
# param batch_ys: shape [-1] type tf.int32
off_ws = [0, 0, 0, 0, 256, 256, 256, 256]
off_hs = [0, 128, 256, 384, 0, 128, 256, 384]
x_img_cuts = [tf.image.crop_to_bounding_box(batch_xs, hs, ws, 128, 256)\
for hs, ws in zip(off_hs, off_ws)]
batch_xs = tf.reshape(tf.concat(x_img_cuts, axis=0), [FLAGS.batch_size*8, 128, 256, 3])
batch_ys = tf.reshape(batch_ys, [FLAGS.batch_size * 8])
if FLAGS.is_training:
with slim.arg_scope(resnet_v1.resnet_arg_scope()):
logits, end_points = resnet_v1.resnet_v1_152(batch_xs, num_classes=6,
is_training=True)
logits = tf.reshape(logits, [-1, 6], name='logits_2d')
prediction = tf.argmax(logits, axis=-1, output_type=tf.int32)
mAP = tf.reduce_mean(tf.cast(tf.equal(prediction, batch_ys),
dtype=tf.float32))
loss, _ = init_loss(logits, batch_ys, end_points=end_points, losstype=FLAGS.losstype)
mAP_sum = tf.summary.scalar('mAP', mAP)
loss_sum = tf.summary.scalar('loss', loss)
summaries = tf.summary.merge([mAP_sum, loss_sum])
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.InteractiveSession(config=config)
counter = get_counter(FLAGS.model_dir)
# Exponential decay learning rate and optimizer configurations
learning_rate = tf.train.exponential_decay(FLAGS.learning_rate, counter,
100, 0.98, staircase=True)
if 'SGD' in FLAGS.optimizer:
optim = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss,
global_step=tf.Variable(counter))
elif 'Adam' in FLAGS.optimizer:
optim = tf.train.AdamOptimizer(learning_rate).minimize(loss, global_step=tf.Variable(counter))
else:
optim = None
raise NotImplementedError
sess.run(tf.global_variables_initializer())
if FLAGS.pretrained:
# Load the pretrained model given by TensorFlow official
exclusions = ['resnet_v1_152/logits', 'predictions']
resnet_except_logits = slim.get_variables_to_restore(exclude=exclusions)
init_fn = slim.assign_from_checkpoint_fn(FLAGS.pretrain_dir, resnet_except_logits,
ignore_missing_vars=True)
init_fn(sess)
print('Model successfully loaded')
else:
# Load the model trained by ourselves
counter = load(sess, FLAGS.model_basedir)
# Ready to train
train(sess, optim, loss, summaries, FLAGS.loops, counter=counter)
print('Training finished')
else:
with slim.arg_scope(resnet_v1.resnet_arg_scope()):
probs, end_points = resnet_v1.resnet_v1_152(batch_xs, num_classes=6,
is_training=False)
prediction = tf.argmax(tf.reshape(probs, [-1, 6]), axis=-1)
accuracy, update_acc = tf.metrics.accuracy(batch_ys, arg_ys)
_, binary = init_loss(probs, batch_ys, end_points, loss_type=FLAGS.losstype)
# TODO: Consider the binary classification...
# which is even more tricky to be implemented...
print('The model accuracy is {}'.format(evaluate(sess, accuracy, update_acc)))