def eval(split, train_dir):
""" Evaluates the given network on a subset of the data-split (1000 images) and prints the top-5 and top-1 Accuracy
Args:
split: Chooses split of flower dataset to test the network
train_dir: Directory in which checkpoints get loaded
Returns:
-
"""
with tf.Graph().as_default():
tf.logging.set_verbosity(tf.logging.INFO)
dataset = dataVisualisation.get_split(split_name=split,
dataset_dir=flowers_data_dir,
label_type="one")
images, _, labels = load_batch(dataset,
batch_size=100,
height=image_size,
width=image_size)
logits = my_cnn(images, is_training=False)
predictions = tf.argmax(logits, 1)
# Define the metrics:
names_to_values, names_to_updates = slim.metrics.aggregate_metric_map({
'eval/Accuracy':
slim.metrics.streaming_accuracy(predictions, labels),
'eval/Recall@5':
slim.metrics.streaming_sparse_recall_at_k(logits, labels, 5)
})
op = tf.summary.scalar('top5percentError',
names_to_values['eval/Recall@5'],
collections=[])
tf.add_to_collection(tf.GraphKeys.SUMMARIES, op)
op = tf.summary.scalar('top1percentError',
names_to_values['eval/Accuracy'],
collections=[])
tf.add_to_collection(tf.GraphKeys.SUMMARIES, op)
print('Running evaluation Loop...')
checkpoint_path = tf.train.latest_checkpoint(train_dir)
metric_values = slim.evaluation.evaluate_once(
master='',
checkpoint_path=checkpoint_path,
logdir=train_dir,
num_evals=30, # TODO Adjust number of evaluations
summary_op=tf.summary.merge_all(),
eval_op=list(names_to_updates.values()),
final_op=list(names_to_values.values()))
print("Evaluation 'eval' started.")
names_to_values = dict(zip(names_to_values.keys(), metric_values))
for name in names_to_values:
print('%s: %f' % (name, names_to_values[name]))
return names_to_values['eval/Accuracy'], names_to_values[
'eval/Recall@5']
def train(split, train_steps, train_dir):
with tf.Graph().as_default():
tf.logging.set_verbosity(tf.logging.INFO) # showing INFO logs
dataset = dataVisualisation.get_split(split, flowers_data_dir)
images, _, labels = load_batch(dataset, height=image_size, width=image_size, is_training=True)
one_hot_labels = slim.one_hot_encoding(labels, num_classes)
# Forward pass with non-flipped images
logits = my_cnn(images, is_training=True)
tf.losses.softmax_cross_entropy(one_hot_labels, logits)
total_loss = tf.losses.get_total_loss()
# Create some summaries to visualize the training process:
tf.summary.scalar('losses/Total_Loss', total_loss)
# Learning rate decay
global_step = variables.get_or_create_global_step()
# Piecewise constant from boundaries and interval values.
boundaries = [tf.constant(100000, dtype= "int64"), tf.constant(200000, dtype= "int64"), tf.constant(300000, dtype= "int64")]
values = [0.001, 0.0001, 0.00001, 0.000001] # 20.000 steps lr 0.001 -> 0.0005 # 35000 steps lr 0.0005 -> 0.001 # 40.000 steps lr 0.001 -> 0.0005 # 55.000 lr 0.0005 -> 0.0001
#boundaries = [tf.constant(10000, dtype= "int64"), tf.constant(100000, dtype= "int64"), tf.constant(200000, dtype= "int64"), tf.constant(300000, dtype= "int64")]
#values = [0.001, 0.0005, 0.0001, 0.00001, 0.000001] # 20.000 steps lr 0.001 -> 0.0005 # 35000 steps lr 0.0005 -> 0.001 # 40.000 steps lr 0.001 -> 0.0005 # 55.000 lr 0.0005 -> 0.0001
#boundaries = [tf.constant(100000, dtype= "int64"), tf.constant(200000, dtype= "int64"), tf.constant(300000, dtype= "int64"), tf.constant(400000, dtype= "int64")]
#values = [0.001, 0.0001, 0.00001, 0.000001, 0.000001] # 20.000 steps lr 0.001 -> 0.0005 # 35000 steps lr 0.0005 -> 0.001 # 40.000 steps lr 0.001 -> 0.0005 # 55.000 lr 0.0005 -> 0.0001
my_learning_rate = tf.train.piecewise_constant(global_step, boundaries, values)
# Specify the optimizer and create the train op:
optimizer = tf.train.MomentumOptimizer(learning_rate=my_learning_rate, momentum = 0.9)
train_op = slim.learning.create_train_op(total_loss, optimizer)
#TODO Choose init_fn or get_init_fn()
#checkpoint_path = tf.train.latest_checkpoint(train_dir)
#init_fn = slim.assign_from_checkpoint_fn(checkpoint_path, slim.get_variables_to_restore())
# Run the training:
final_loss = slim.learning.train(
train_op,
logdir=train_dir,
log_every_n_steps=1,
init_fn= get_init_fn(),
number_of_steps=train_steps,
global_step = global_step)
print('Finished training. Last batch loss %f' % final_loss)
def eval(split, train_dir):
# This might take a few minutes.
with tf.Graph().as_default():
tf.logging.set_verbosity(tf.logging.INFO)
dataset = dataVisualisation.get_split(split, flowers_data_dir)
images, _, labels = load_batch(dataset, batch_size = 100, height=image_size, width=image_size) # TODO load_batch really necessary? Processing all! At least adjust batch_size
logits = my_cnn(images, is_training=False)
predictions = tf.argmax(logits, 1)
# Define the metrics:
names_to_values, names_to_updates = slim.metrics.aggregate_metric_map({
'eval/Accuracy': slim.metrics.streaming_accuracy(predictions, labels),
'eval/Recall@5': slim.metrics.streaming_sparse_recall_at_k(logits, labels, 5)#,
#'eval/Recall@1000': slim.metrics.streaming_sparse_recall_at_k(logits, labels, 1000) # should be 1
})
op = tf.summary.scalar('top5percentError', names_to_values['eval/Recall@5'], collections=[])
tf.add_to_collection(tf.GraphKeys.SUMMARIES, op)
op = tf.summary.scalar('top1percentError', names_to_values['eval/Accuracy'], collections=[])
tf.add_to_collection(tf.GraphKeys.SUMMARIES, op)
# for name, value in names_to_values.iteritems():
# summary_name = 'eval/%s' % name
# op = tf.summary.scalar(summary_name, value, collections=[])
# op = tf.Print(op, [value], summary_name)
# tf.add_to_collection(tf.GraphKeys.SUMMARIES, op)
print('Running evaluation Loop...')
checkpoint_path = tf.train.latest_checkpoint(train_dir)
#print_tensors_in_checkpoint_file(checkpoint_path, tensor_name ='my_fc_1/weights', all_tensors=False)
#print_tensors_in_checkpoint_file(checkpoint_path, tensor_name ='resnet_v2_50/block1/unit_1/bottleneck_v2/conv3/weights', all_tensors=False)
metric_values = slim.evaluation.evaluate_once(
master='',
checkpoint_path=checkpoint_path,
logdir=train_dir,
num_evals = 10, # TODO Adjust number of evaluations
summary_op=tf.summary.merge_all(),
eval_op=list(names_to_updates.values()),
final_op=list(names_to_values.values()))
print("Evaluation 'eval' started.")
names_to_values = dict(zip(names_to_values.keys(), metric_values))
for name in names_to_values:
print('%s: %f' % (name, names_to_values[name]))
def eval(split, eval_steps, train_dir, fc_after, level):
""" Evaluates the given neural network, which is saved in train_dir
Args:
split: Chooses split of flower dataset to train the network
eval_steps: Number of steps to evaluate network
train_dir: Directory in which checkpoints sare found
fc_after: where to cut of the network
level: which class do you want to eval on? (family, genus, species, organ)
Returns:
-
"""
with tf.Graph().as_default():
tf.logging.set_verbosity(tf.logging.INFO)
dataset = dataVisualisation.get_split(split_name = split, dataset_dir = flowers_data_dir, label_type="multiple")
images, _, label_species, labels_genus, labels_family, labels_organ= load_batch_intermediate(dataset,height=224, width=224, is_training=False, batch_size=100)
#print(family, genus, species)
abstraction_levels = {"family":labels_family, "genus":labels_genus, "species":label_species, "organs":labels_organ}
levels_length = {"family":124, "genus":516, "species":1000, "organs":7}
labels = tf.stack(abstraction_levels.get(level, label_species))
one_hot_labels = slim.one_hot_encoding(labels, levels_length.get(level, 1000))
logits,_ = my_intermediate_cnn(images, is_training=True, fc_after=fc_after,num_classes = levels_length.get(level, 1000))
logits = slim.softmax(logits)
#logits = tf.cast(logits, tf.int64)
predictions = tf.argmax(logits, 1)
predictions = tf.cast(predictions, tf.int64)
# Define the metrics:
names_to_values, names_to_updates = slim.metrics.aggregate_metric_map({
'eval/Accuracy': slim.metrics.streaming_accuracy(predictions, labels),
'eval/Recall@5': slim.metrics.streaming_sparse_recall_at_k(logits, labels, 5)
})
op = tf.summary.scalar('top5percentError', names_to_values['eval/Recall@5'], collections=[])
tf.add_to_collection(tf.GraphKeys.SUMMARIES, op)
op = tf.summary.scalar('top1percentError', names_to_values['eval/Accuracy'], collections=[])
tf.add_to_collection(tf.GraphKeys.SUMMARIES, op)
print('Running evaluation Loop...')
checkpoint_path = tf.train.latest_checkpoint(train_dir)
metric_values = slim.evaluation.evaluate_once(
master='',
checkpoint_path=checkpoint_path,
logdir=train_dir,
num_evals = eval_steps, # TODO Adjust number of evaluations
summary_op=tf.summary.merge_all(),
eval_op=list(names_to_updates.values()),
final_op=list(names_to_values.values()))
print("Evaluation 'eval' started.")
names_to_values = dict(zip(names_to_values.keys(), metric_values))
for name in names_to_values:
print('%s: %f' % (name, names_to_values[name]))
return names_to_values['eval/Accuracy'], names_to_values['eval/Recall@5']
def train(split, train_steps, train_dir, fc_after, level, checkpoints_dir = checkpoints_dir,checkpoint = 'model.ckpt-150000'):
""" Trains the given neural network and saves the weights and summary information into a new checkpoint file in the train_dir
Args:
split: Chooses split of flower dataset to train the network
train_steps: Number of steps to train network
train_dir: Directory in which checkpoints should be stored, and old checkpoints get loaded
Returns:
-
"""
with tf.Graph().as_default():
tf.logging.set_verbosity(tf.logging.INFO) # showing INFO logs
dataset = dataVisualisation.get_split(split_name = split, dataset_dir = flowers_data_dir, label_type="multiple")
images, _, label_species, labels_genus, labels_family, labels_organ= load_batch_intermediate(dataset,height=224, width=224, is_training=True, batch_size=100)
#print(family, genus, species)
abstraction_levels = {"family":labels_family, "genus":labels_genus, "species":label_species, "organs":labels_organ}
levels_length = {"family":124, "genus":516, "species":1000, "organs":7}
labels = tf.stack(abstraction_levels.get(level, label_species))
one_hot_labels = slim.one_hot_encoding(labels, levels_length.get(level, 1000))
# Forward pass with non-flipped images
logits,_ = my_intermediate_cnn(images, is_training=True, fc_after=fc_after, num_classes = levels_length.get(level, 1000))
#print(logits, one_hot_labels)
#with tf.Session() as sess:
# print(sess.run(tf.shape(logits)))
tf.losses.softmax_cross_entropy(one_hot_labels, logits)
total_loss = tf.losses.get_total_loss()
tf.summary.scalar('losses/Total_Loss', total_loss)
# Learning rate decay
global_step = variables.get_or_create_global_step()
boundaries = [tf.constant(100000, dtype= "int64"), tf.constant(200000, dtype= "int64"), tf.constant(300000, dtype= "int64")]
values = [0.001, 0.0001, 0.00001, 0.000001]
my_learning_rate = tf.train.piecewise_constant(global_step, boundaries, values)
#for v in tf.trainable_variables():
# print(v)
#for v in slim.get_variables(scope="resnet_v2_50/fc_intermediate/"):
# print(v)
# Specify the optimizer and create the train op:
optimizer = tf.train.MomentumOptimizer(learning_rate=my_learning_rate, momentum = 0.9)
train_op = slim.learning.create_train_op(total_loss=total_loss, optimizer=optimizer, variables_to_train=slim.get_variables(scope="fc_intermediate"))
saver = tf.train.Saver(max_to_keep=1)
# Run the training:
final_loss = slim.learning.train(
train_op,
logdir=train_dir,
log_every_n_steps=50,
init_fn= get_init_fn(fc_after),
number_of_steps=train_steps,
global_step = global_step,
saver = saver)
print('Finished training. Last batch loss %f' % final_loss)
def train(split, train_steps, train_dir):
""" Trains the given neural network and saves the weights and summary information into a new checkpoint file in the train_dir
Args:
split: Chooses split of flower dataset to train the network
train_steps: Number of steps to train network
train_dir: Directory in which checkpoints should be stored, and old checkpoints get loaded
Returns:
-
"""
with tf.Graph().as_default():
tf.logging.set_verbosity(tf.logging.INFO) # showing INFO logs
dataset = dataVisualisation.get_split(split_name=split,
dataset_dir=flowers_data_dir,
label_type="one")
images, _, labels = load_batch(dataset,
height=image_size,
width=image_size,
is_training=True)
one_hot_labels = slim.one_hot_encoding(labels, num_classes)
# Forward pass with non-flipped images
logits = my_cnn(images, is_training=True)
tf.losses.softmax_cross_entropy(one_hot_labels, logits)
total_loss = tf.losses.get_total_loss()
# Create some summaries to visualize the training process:
tf.summary.scalar('losses/Total_Loss', total_loss)
# Learning rate decay
global_step = variables.get_or_create_global_step()
# Piecewise constant from boundaries and interval values.
boundaries = [
tf.constant(100000, dtype="int64"),
tf.constant(200000, dtype="int64"),
tf.constant(300000, dtype="int64")
]
values = [0.001, 0.0001, 0.00001, 0.000001]
my_learning_rate = tf.train.piecewise_constant(global_step, boundaries,
values)
# Specify the optimizer and create the train op:
optimizer = tf.train.MomentumOptimizer(learning_rate=my_learning_rate,
momentum=0.9)
train_op = slim.learning.create_train_op(total_loss, optimizer)
# Run the training:
final_loss = slim.learning.train(train_op,
logdir=train_dir,
log_every_n_steps=1,
init_fn=get_init_fn(),
number_of_steps=train_steps,
global_step=global_step)
print('Finished training. Last batch loss %f' % final_loss)
def final_evaluation(label_dir, dataset_dir, filename="predictions.txt"):
number_images = 100 #8000
correct = 0 # TODO REMOVE
#[ 2.49181414 -0.33259141 -1.27059281 -1.52844727 2.57813239 -1.20948148
checkpoint_paths = [
"mydata/resnet_finetuned_plantclef2015_1/model.ckpt-145552",
"mydata/resnet_finetuned_plantclef2015_1/model.ckpt-145552",
"mydata/resnet_finetuned_plantclef2015_1/model.ckpt-145552"
]
output_list = []
labels_list = []
for index in range(len(checkpoint_paths)):
with tf.Graph().as_default():
dataset = dataVisualisation.get_split('train_set3', dataset_dir)
data_provider = slim.dataset_data_provider.DatasetDataProvider(
dataset,
shuffle=False,
common_queue_capacity=8000,
common_queue_min=0)
image_raw, label = data_provider.get(['image', 'label'])
"""
imaget = my_resnet_preprocessing.preprocess_image(image_raw, 224, 224, is_training=False)
image,augmented_image1,augmented_image2, augmented_image3, augmented_image4,augmented_image5, labels = tf.train.batch([imaget,imaget,imaget,imaget,imaget,imaget, label], batch_size=1,
num_threads=1,
capacity=12 * 1)
"""
image, augmented_image1, augmented_image2, augmented_image3, augmented_image4, augmented_image5 = my_resnet_preprocessing.preprocess_for_final_run(
image_raw, 224, 224)
image, augmented_image1, augmented_image2, augmented_image3, augmented_image4, augmented_image5, labels = tf.train.batch(
[
image, augmented_image1, augmented_image2,
augmented_image3, augmented_image4, augmented_image5, label
],
batch_size=1,
num_threads=1,
capacity=2 * 1)
logits1 = resNetClassifier.my_cnn(image,
is_training=False,
dropout_rate=1)
logits2 = resNetClassifier.my_cnn(augmented_image1,
is_training=False,
dropout_rate=1)
logits3 = resNetClassifier.my_cnn(augmented_image2,
is_training=False,
dropout_rate=1)
logits4 = resNetClassifier.my_cnn(augmented_image3,
is_training=False,
dropout_rate=1)
logits5 = resNetClassifier.my_cnn(augmented_image4,
is_training=False,
dropout_rate=1)
logits6 = resNetClassifier.my_cnn(augmented_image5,
is_training=False,
dropout_rate=1)
total_output = np.empty([number_images * 1, dataset.num_classes])
total_labels = np.empty([number_images * 1], dtype=np.int32)
offset = 0
#init_fn = slim.assign_from_checkpoint_fn(checkpoint_paths[index], slim.get_model_variables())
with tf.Session() as sess:
coord = tf.train.Coordinator()
saver = tf.train.Saver()
saver.restore(sess, checkpoint_paths[index])
#init_fn(sess)
merged = tf.summary.merge_all()
test_writer = tf.summary.FileWriter(
'/home/lolek/FlowerIdentificationM/models/slim/mydata/test/'
+ '/train')
visualize_writer = tf.summary.FileWriter(
'/home/lolek/FlowerIdentificationM/models/slim/mydata/test/'
+ '/visualize')
#for v in tf.trainable_variables():
# print(v.name)
#print_tensors_in_checkpoint_file(checkpoint_paths[index], tensor_name =None, all_tensors=False)
#var = [v for v in tf.trainable_variables() if v.name == "my_fc_1/weights:0"]
#print("my_fc_1/weights \n", sess.run(var))
#print_tensors_in_checkpoint_file(checkpoint_paths[index], tensor_name ='my_fc_1/weights', all_tensors=False)
#var2 = [v for v in tf.trainable_variables() if v.name == "resnet_v2_50/block1/unit_1/bottleneck_v2/conv3/weights:0"]
#print("resnet_v2_50/block1/unit_1/bottleneck_v2/conv3/weights \n", sess.run(var2))
#print_tensors_in_checkpoint_file(checkpoint_paths[index], tensor_name ='resnet_v2_50/block1/unit_1/bottleneck_v2/conv3/weights', all_tensors=False)
# Visualize Kernels
tf.get_variable_scope().reuse_variables()
weights = tf.get_variable("resnet_v2_50/conv1/weights")
grid = kernel_visualization.put_kernels_on_grid(weights)
sum1 = tf.summary.image('conv1/kernels', grid, max_outputs=1)
_, summary1 = sess.run([merged, sum1])
visualize_writer.add_summary(summary1, 2)
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
for i in range(number_images):
print('step: %d/%d' % (i + 1, number_images))
"""
image_t1, image_t2 = sess.run([image, augmented_image1])
plt.figure()
plt.imshow(image_t1[0, :, :, :].astype(np.uint8))
plt.title('title')
plt.axis('off')
plt.show()
plt.figure()
plt.imshow(image_t2[0, :, :, :].astype(np.uint8))
plt.title('title')
plt.axis('off')
plt.show()
"""
sum1 = tf.summary.image('final_eval_whole_image1', image)
sum2 = tf.summary.image('final_eval_image_center1',
augmented_image1)
sum3 = tf.summary.image('final_eval_top_left1',
augmented_image2)
sum4 = tf.summary.image('final_eval_bottom_left1',
augmented_image3)
sum5 = tf.summary.image('final_eval_top_right1',
augmented_image4)
sum6 = tf.summary.image('final_eval_bottom_right1',
augmented_image5)
_, summary1, summary2, summary3, summary4, summary5, summary6 = sess.run(
[merged, sum1, sum2, sum3, sum4, sum5, sum6])
test_writer.add_summary(summary1, 1)
test_writer.add_summary(summary2, 1)
test_writer.add_summary(summary3, 1)
test_writer.add_summary(summary4, 1)
test_writer.add_summary(summary5, 1)
test_writer.add_summary(summary6, 1)
logit1, logit2, logit3, logit4, logit5, logit6, media_id = sess.run(
[
logits1, logits2, logits3, logits4, logits5,
logits6, labels
])
print(media_id, " ", np.argmax(logit1[0]), " ",
np.argmax(logit2[0]), " ", np.argmax(logit3[0]), " ",
np.argmax(logit4[0]), " ", np.argmax(logit5[0]), " ",
np.argmax(logit6[0]))
#print(np.amax(logit1[0]), " ",np.amax(logit2[0]), " ",np.amax(logit3[0]), " ",np.amax(logit4[0]), " ",np.amax(logit5[0]), " ",np.amax(logit6[0]))
#print(len(logit1[0]))
media_id = media_id[0]
logits = tuple(
max(i, j) for i, j in zip(logit1[0], logit2[0]))
logits = tuple(
max(i, j) for i, j in zip(logits, logit3[0]))
logits = tuple(
max(i, j) for i, j in zip(logits, logit4[0]))
logits = tuple(
max(i, j) for i, j in zip(logits, logit5[0]))
logits = tuple(
max(i, j) for i, j in zip(logits, logit6[0]))
"""
logits = tuple(i + j for i, j in zip(logit1[0], logit2[0]))
logits = tuple(i + j for i, j in zip(logits, logit3[0]))
logits = tuple(i + j for i, j in zip(logits, logit4[0]))
logits = tuple(i + j for i, j in zip(logits, logit5[0]))
logits = tuple(i + j for i, j in zip(logits, logit6[0]))
logits = [x / 6 for x in logits]
"""
logits = numpy_softmax(logits)
#print(np.argmax(logits))
#print(len(logits))
#first_prediction = np.argmax(logits)
#print(first_prediction," ", media_id)
#print(np.argmax(logits))
#logits = logit1
#print(np.amax(logits))
#pred = [np.argmax(logit1[0]),np.argmax(logit2[0]),np.argmax(logit3[0]), np.argmax(logit4[0]), np.argmax(logit5[0]), np.argmax(logit6[0] )]
if media_id == np.argmax(logits): # TODO REMOVE
correct = correct + 1 # TODO REMOVE
#print(correct) # TODO REMOVE
#print(media_id, " ", np.argmax(logit1)," ",np.argmax(logit2)," ",np.argmax(logit3), " ",np.argmax(logit4), " ",np.argmax(logit5), " ",np.argmax(logit6) )
total_output[offset:offset + 1] = logits
total_labels[offset:offset + 1] = media_id
offset += 1
coord.request_stop()
coord.join(threads)
output_list.append(total_output)
labels_list.append(total_labels)
print(correct) # TODO REMOVE
for i in range(len(output_list)):
logits = tf.cast(tf.constant(output_list[i]), dtype=tf.float32)
predictions = tf.nn.softmax(logits)
labels = tf.constant(labels_list[i])
top1_op = tf.nn.in_top_k(predictions, labels, 1)
top5_op = tf.nn.in_top_k(predictions, labels, 5)
with tf.Session() as sess:
top1, top5 = sess.run([top1_op, top5_op])
print('Top 1 accuracy: %f' % (np.sum(top1) / float(number_images)))
print('Top 5 accuracy: %f' % (np.sum(top5) / float(number_images)))
for i in range(number_images):
image_id = labels_list[0][i]
prediction1 = np.amax(output_list[0][i])
prediction2 = np.amax(output_list[1][i])
prediction3 = np.amax(output_list[2][i])
print(prediction1, " ", prediction2, " ", prediction3)
# Find best class with highest prediction (softmax) score
if prediction1 > prediction2:
if prediction1 > prediction3:
prediction = np.argmax(output_list[0][i])
probability = prediction1
else:
prediction = np.argmax(output_list[2][i])
probability = prediction3
else:
if prediction2 > prediction3:
prediction = np.argmax(output_list[1][i])
probability = prediction2
else:
prediction = np.argmax(output_list[2][i])
probability = prediction3
class_id = dataset_utils.read_label_file(label_dir)[prediction]
image_id = dataset_utils.read_label_file(label_dir)[
image_id] # TODO REMOVE!!!
print('<%s;%s;%f>\n' % (image_id, class_id, probability))
# Save the predictions
labels_filename = os.path.join(label_dir, filename)
with tf.gfile.Open(labels_filename, 'a') as f:
f.write('<%s;%s;%f>\n' %
(image_id, class_id,
probability)) # <ImageId;ClassId;Probability>
def final_evaluation_generic(label_dir, dataset_dir, checkpoint_paths, preprocessing_methods, filename="predictions.txt"):
"""
Evaulates a CNN on the test-set and saves the predictions in the form <ImageId;ClassId;Probability> into a txt-file (filename)
Can use multiple models, is not limited to 3
Args:
label_dir: Directory where labels dictionary can be found (mapping from one-hot encodings to class_id)
dataset_dir: Directory where test dataset can be found
checkpoint_paths: checkpoints of the used models
preprocessing_methods: corresponding preprocessing methods for the models
filename: filename of txt-file to save predictions
Returns:
Saves the predictions into "filename"
"""
number_images =8000
output_list = []
labels_list = []
for checkpoint, preprocessing_method in zip(checkpoint_paths, preprocessing_methods):
with tf.Graph().as_default() as graph:
dataset = dataVisualisation.get_split('test_set', dataset_dir,label_type="one")
data_provider = slim.dataset_data_provider.DatasetDataProvider(dataset,
shuffle=False,
common_queue_capacity=8000,
common_queue_min=0)
image_raw, label = data_provider.get(['image', 'label'])
# Preprocessing return original image, center_crop and 4 corner crops with adjusted color values
image = preprocessing_method(image_raw, 224, 224)
images, labels = tf.train.batch([image, label],
batch_size=1,
num_threads=1,
capacity=2 * 1)
logits1 = resNetClassifier.my_cnn(images, is_training = False, dropout_rate =1)
total_output = np.empty([number_images * 1, dataset.num_classes])
total_labels = np.empty([number_images * 1], dtype=np.int32)
offset = 0
with tf.Session() as sess:
coord = tf.train.Coordinator()
saver = tf.train.Saver()
saver.restore(sess, checkpoint)
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
for i in range(number_images):
#print('step: %d/%d' % (i+1, number_images))
logit1, media_id = sess.run([logits1, labels])
media_id = media_id[0]
# Passing logits through softmax function to receive "probabilities"
logits = my_functions.numpy_softmax(logit1)
total_output[offset:offset + 1] = logits
total_labels[offset:offset + 1] = media_id
offset += 1
coord.request_stop()
coord.join(threads)
output_list.append(total_output)
labels_list.append(total_labels)
with tf.gfile.Open(filename, 'a') as f:
for i in range(number_images):
image_id = labels_list[0][i]
for p in range(1000):
predictions = []
for index in range(len(output_list)):
predictions.append(output_list[index][i][p])
probability = np.sum(predictions)/len(predictions) #np.amax(predictions)
class_id = dataset_utils.read_label_file(label_dir)[p]
f.write('%s;%s;%f\n' % (image_id, class_id, probability)) # <ImageId;ClassId;Probability>
def final_evaluation(label_dir, dataset_dir, filename="predictions.txt", visualize_kernel = False):
"""
Evaulates a CNN on the test-set and saves the predictions in the form <ImageId;ClassId;Probability> into a txt-file (filename)
Args:
label_dir: Directory where labels dictionary can be found (mapping from one-hot encodings to class_id)
dataset_dir: Directory where test dataset can be found
filename: filename of txt-file to save predictions
visualize_kernel: Do you want to visualize the first layer of convolutions?
Returns:
Saves the predictions into "filename"
"""
number_images =8000
# Choose the three networks to evaluate on
checkpoint_paths = [ "mydata/resnet_finetuned_plantclef2015_5/model.ckpt-150000", "mydata/resnet_finetuned_plantclef2015_6/model.ckpt-150000","mydata/resnet_finetuned_plantclef2015_7/model.ckpt-102500"]
output_list = []
labels_list = []
for index in range(len(checkpoint_paths)):
with tf.Graph().as_default() as graph:
dataset = dataVisualisation.get_split('test_set', dataset_dir,label_type="one")
data_provider = slim.dataset_data_provider.DatasetDataProvider(dataset,
shuffle=False,
common_queue_capacity=8000,
common_queue_min=0)
image_raw, label = data_provider.get(['image', 'label'])
# Preprocessing return original image, center_crop and 4 corner crops with adjusted color values
image, augmented_image1, augmented_image2, augmented_image3, augmented_image4, augmented_image5 = my_resnet_preprocessing.preprocess_for_final_run2(image_raw, 224, 224)
image,augmented_image1,augmented_image2, augmented_image3, augmented_image4,augmented_image5, labels = tf.train.batch([image, augmented_image1, augmented_image2, augmented_image3, augmented_image4, augmented_image5, label],
batch_size=1,
num_threads=1,
capacity=2 * 1)
logits1 = resNetClassifier.my_cnn(image, is_training = False, dropout_rate =1)
logits2 = resNetClassifier.my_cnn(augmented_image1, is_training = False, dropout_rate =1)
logits3 = resNetClassifier.my_cnn(augmented_image2, is_training = False, dropout_rate =1)
logits4 = resNetClassifier.my_cnn(augmented_image3, is_training = False, dropout_rate =1)
logits5 = resNetClassifier.my_cnn(augmented_image4, is_training = False, dropout_rate =1)
logits6 = resNetClassifier.my_cnn(augmented_image5, is_training = False, dropout_rate =1)
total_output = np.empty([number_images * 1, dataset.num_classes])
total_labels = np.empty([number_images * 1], dtype=np.int32)
offset = 0
with tf.Session() as sess:
coord = tf.train.Coordinator()
saver = tf.train.Saver()
saver.restore(sess, checkpoint_paths[index])
if visualize_kernel:
tf.get_variable_scope().reuse_variables()
#for v in tf.global_variables():
# print(v.name)
weights = tf.get_variable("resnet_v2_50/conv1/weights")
print(weights.get_shape()[0].value, weights.get_shape()[1].value, weights.get_shape()[2].value, weights.get_shape()[3].value)
weights = tf.slice(weights,[0,0,0,1] , [weights.get_shape()[0].value, weights.get_shape()[1].value, weights.get_shape()[2].value, 2])
grid = kernel_visualization.put_kernels_on_grid (weights)
sum1 = tf.summary.image('conv1/kernels', grid, max_outputs=1)
_, summary1, img = sess.run([merged, sum1, tf.squeeze(grid)])
visualize_writer.add_summary(summary1,2)
fig = plt.figure()
plt.imshow(img)
plt.savefig("images/kernelsOne_%s.png" % (index))
#plt.show()
plt.close(fig)
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
for i in range(number_images):
#print('step: %d/%d' % (i+1, number_images))
logit1, logit2, logit3, logit4,logit5, logit6, media_id = sess.run([logits1, logits2, logits3, logits4, logits5, logits6, labels])
media_id = media_id[0]
# Use Average for voting of logits
logits = tuple(i + j for i, j in zip(logit1[0], logit2[0]))
logits = tuple(i + j for i, j in zip(logits, logit3[0]))
logits = tuple(i + j for i, j in zip(logits, logit4[0]))
logits = tuple(i + j for i, j in zip(logits, logit5[0]))
logits = tuple(i + j for i, j in zip(logits, logit6[0]))
logits = [x / 6 for x in logits]
# Passing logits through softmax function to receive "probabilities"
logits = my_functions.numpy_softmax(logits)
total_output[offset:offset + 1] = logits
total_labels[offset:offset + 1] = media_id
offset += 1
coord.request_stop()
coord.join(threads)
output_list.append(total_output)
labels_list.append(total_labels)
prediction_filename = filename
#os.remove(prediction_filename)
for i in range(number_images):
image_id = labels_list[0][i]
for p in range(1000):
p1 = output_list[0][i][p]
p2 = output_list[1][i][p]
p3 = output_list[2][i][p]
probability = np.amax([p1, p2, p3])
class_id = dataset_utils.read_label_file(label_dir)[p]
with tf.gfile.Open(prediction_filename, 'a') as f:
f.write('%s;%s;%f\n' % (image_id, class_id, probability)) # <ImageId;ClassId;Probability>