def visualize_model(
live_ims,
dead_ims,
model_file,
output_folder,
num_channels,
smooth_iterations=50,
untargeted=False,
viz='none',
per_timepoint=True):
"""Train an SVM for your dataset on GEDI-model encodings."""
config = GEDIconfig()
if live_ims is None:
raise RuntimeError(
'You need to supply a directory path to the live images.')
if dead_ims is None:
raise RuntimeError(
'You need to supply a directory path to the dead images.')
live_files = glob(os.path.join(live_ims, '*%s' % config.raw_im_ext))
dead_files = glob(os.path.join(dead_ims, '*%s' % config.raw_im_ext))
combined_labels = np.concatenate((
np.zeros(len(live_files)),
np.ones(len(dead_files))))
combined_files = np.concatenate((live_files, dead_files))
if len(combined_files) == 0:
raise RuntimeError('Could not find any files. Check your image path.')
config = GEDIconfig()
model_file_path = os.path.sep.join(model_file.split(os.path.sep)[:-1])
meta_file_pointer = os.path.join(
model_file_path,
'train_maximum_value.npz')
if not os.path.exists(meta_file_pointer):
raise RuntimeError(
'Cannot find the training data meta file: train_maximum_value.npz'
'Closest I could find from directory %s was %s.'
'Download this from the link described in the README.md.'
% (model_file_path, glob(os.path.join(model_file_path, '*.npz'))))
meta_data = np.load(meta_file_pointer)
# Prepare image normalization values
training_max = np.max(meta_data['max_array']).astype(np.float32)
training_min = np.min(meta_data['min_array']).astype(np.float32)
# Find model checkpoints
ds_dt_stamp = re.split('/', model_file)[-2]
out_dir = os.path.join(config.results, ds_dt_stamp)
# Make output directories if they do not exist
dir_list = [config.results, out_dir]
[make_dir(d) for d in dir_list]
# Prepare data on CPU
images = tf.placeholder(
tf.float32,
shape=[None] + config.model_image_size,
name='images')
labels = tf.placeholder(
tf.int64,
shape=[None],
name='labels')
# Prepare model on GPU
with tf.device('/gpu:0'):
with tf.variable_scope('cnn'):
vgg = vgg16.model_struct(
vgg16_npy_path=config.vgg16_weight_path,
fine_tune_layers=config.fine_tune_layers)
vgg.build(
images,
output_shape=config.output_shape)
# Setup validation op
scores = vgg.fc7
preds = tf.argmax(vgg.prob, 1)
activity_pattern = vgg.fc8
if not untargeted:
oh_labels = tf.one_hot(labels, config.output_shape)
activity_pattern *= oh_labels
grad_image = tf.gradients(activity_pattern, images)
# Set up saver
saver = tf.train.Saver(tf.global_variables())
# Loop through each checkpoint then test the entire validation set
ckpts = [model_file]
ckpt_yhat, ckpt_y, ckpt_scores = [], [], []
ckpt_file_array, ckpt_viz_images = [], []
print '-' * 60
print 'Beginning evaluation'
print '-' * 60
if config.validation_batch > len(combined_files):
print 'Trimming validation_batch to %s (same as # of files).' % len(
combined_files)
config.validation_batch = len(combined_files)
count = 0
for idx, c in tqdm(enumerate(ckpts), desc='Running checkpoints'):
dec_scores, yhat, y, file_array, viz_images = [], [], [], [], []
# Initialize the graph
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))
sess.run(
tf.group(
tf.global_variables_initializer(),
tf.local_variables_initializer()))
# Set up exemplar threading
saver.restore(sess, c)
start_time = time.time()
num_batches = np.floor(
len(combined_files) / float(
config.validation_batch)).astype(int)
for image_batch, label_batch, file_batch in tqdm(
image_batcher(
start=0,
num_batches=num_batches,
images=combined_files,
labels=combined_labels,
config=config,
training_max=training_max,
training_min=training_min,
num_channels=num_channels,
per_timepoint=per_timepoint),
total=num_batches):
feed_dict = {
images: image_batch,
labels: label_batch
}
it_grads = np.zeros((image_batch.shape))
sc, tyh = sess.run(
[scores, preds],
feed_dict=feed_dict)
for idx in range(smooth_iterations):
feed_dict = {
images: add_noise(image_batch),
labels: label_batch
}
it_grad = sess.run(
grad_image,
feed_dict=feed_dict)
it_grads += it_grad[0]
it_grads /= smooth_iterations # Mean across iterations
it_grads = visualization_function(it_grads, viz)
# Save each grad individually
for grad_i, pred_i, file_i, label_i in zip(
it_grads, tyh, file_batch, label_batch):
out_pointer = os.path.join(
output_folder,
file_i.split(os.path.sep)[-1])
out_pointer = out_pointer.split('.')[0] + '.png'
f = plt.figure()
plt.imshow(grad_i)
plt.title('Pred=%s, label=%s' % (pred_i, label_batch))
plt.savefig(out_pointer)
plt.close(f)
# Plot a moisaic of the grads
if viz == 'none':
pos_grads = normalize(np.maximum(it_grads, 0))
neg_grads = normalize(np.minimum(it_grads, 0))
alpha_mosaic(
image_batch,
pos_grads,
'pos_batch_%s.pdf' % count,
title='Positive gradient overlays.',
rc=1,
cc=len(image_batch),
cmap=plt.cm.Reds)
alpha_mosaic(
image_batch,
neg_grads,
'neg_batch_%s.pdf' % count,
title='Negative gradient overlays.',
rc=1,
cc=len(image_batch),
cmap=plt.cm.Reds)
else:
alpha_mosaic(
image_batch,
it_grads,
'batch_%s.pdf' % count,
title='Gradient overlays.',
rc=1,
cc=len(image_batch),
cmap=plt.cm.Reds)
count += 1
# Store the results
dec_scores += [sc]
yhat = np.append(yhat, tyh)
y = np.append(y, label_batch)
file_array = np.append(file_array, file_batch)
viz_images += [it_grads]
ckpt_yhat.append(yhat)
ckpt_y.append(y)
ckpt_scores.append(dec_scores)
ckpt_file_array.append(file_array)
ckpt_viz_images.append(viz_images)
print 'Batch %d took %.1f seconds' % (
idx, time.time() - start_time)
sess.close()
# Save everything
np.savez(
os.path.join(out_dir, 'validation_accuracies'),
ckpt_yhat=ckpt_yhat,
ckpt_y=ckpt_y,
ckpt_scores=ckpt_scores,
ckpt_names=ckpts,
combined_files=ckpt_file_array,
ckpt_viz_images=ckpt_viz_images)
def train_vgg16(train_dir=None, validation_dir=None):
config = GEDIconfig()
if train_dir is None: # Use globals
train_data = os.path.join(
config.tfrecord_dir,
config.tf_record_names['train'])
meta_data = np.load(
os.path.join(
config.tfrecord_dir,
'%s_%s' % (config.tvt_flags[0], config.max_file)))
else:
meta_data = np.load(
os.path.join(
train_dir,
'%s_%s' % (config.tvt_flags[0], config.max_file)))
# Prepare image normalization values
if config.max_gedi is None:
max_value = np.nanmax(meta_data['max_array']).astype(np.float32)
if max_value == 0:
max_value = None
print 'Derived max value is 0'
else:
print 'Normalizing with empirical max.'
if 'min_array' in meta_data.keys():
min_value = np.min(meta_data['min_array']).astype(np.float32)
print 'Normalizing with empirical min.'
else:
min_value = None
print 'Not normalizing with a min.'
else:
max_value = config.max_gedi
min_value = config.min_gedi
ratio = meta_data['ratio']
if config.encode_time_of_death:
tod = pd.read_csv(config.encode_time_of_death)
tod_data = tod['dead_tp'].as_matrix()
mask = np.isnan(tod_data).astype(int) + (
tod['plate_well_neuron'] == 'empty').as_matrix().astype(int)
tod_data = tod_data[mask == 0]
tod_data = tod_data[tod_data > config.mask_timepoint_value]
config.output_shape = len(np.unique(tod_data))
ratio = class_weight.compute_class_weight(
'balanced',
np.sort(np.unique(tod_data)),
tod_data)
flip_ratio = False
else:
flip_ratio = True
print 'Ratio is: %s' % ratio
if validation_dir is None: # Use globals
validation_data = os.path.join(
config.tfrecord_dir,
config.tf_record_names['val'])
elif validation_dir is False:
pass # Do not use validation data during training
# Make output directories if they do not exist
dt_stamp = re.split(
'\.', str(datetime.now()))[0].\
replace(' ', '_').replace(':', '_').replace('-', '_')
dt_dataset = config.which_dataset + '_' + dt_stamp + '/'
config.train_checkpoint = os.path.join(
config.train_checkpoint, dt_dataset) # timestamp this run
out_dir = os.path.join(config.results, dt_dataset)
dir_list = [
config.train_checkpoint, config.train_summaries,
config.results, out_dir]
[make_dir(d) for d in dir_list]
# im_shape = get_image_size(config)
im_shape = config.gedi_image_size
print '-'*60
print('Training model:' + dt_dataset)
print '-'*60
# Prepare data on CPU
assert os.path.exists(train_data)
assert os.path.exists(validation_data)
assert os.path.exists(config.vgg16_weight_path)
with tf.device('/cpu:0'):
train_images, train_labels, train_gedi_images = inputs(
train_data,
config.train_batch,
im_shape,
config.model_image_size[:2],
max_value=max_value,
min_value=min_value,
train=config.data_augmentations,
num_epochs=config.epochs,
normalize=config.normalize,
return_gedi=config.include_GEDI_in_tfrecords,
return_extra_gfp=config.extra_image,
return_GEDI_derivative=True)
val_images, val_labels, val_gedi_images = inputs(
validation_data,
config.validation_batch,
im_shape,
config.model_image_size[:2],
max_value=max_value,
min_value=min_value,
num_epochs=config.epochs,
normalize=config.normalize,
return_gedi=config.include_GEDI_in_tfrecords,
return_extra_gfp=config.extra_image,
return_GEDI_derivative=True)
if config.include_GEDI_in_tfrecords:
extra_im_name = 'GEDI at current timepoint'
else:
extra_im_name = 'next gfp timepoint'
tf.summary.image('train images', train_images)
tf.summary.image('validation images', val_images)
tf.summary.image('train %s' % extra_im_name, train_gedi_images)
tf.summary.image('validation %s' % extra_im_name, val_gedi_images)
# Prepare model on GPU
with tf.device('/gpu:0'):
with tf.variable_scope('cnn') as scope:
if config.ordinal_classification is None:
vgg_output = 2 # Sign of derivative (inf norm)
train_labels = tf.cast(tf.sign(train_labels), tf.int32)
val_labels = tf.cast(tf.sign(val_labels), tf.int32)
elif config.ordinal_classification == 'regression':
vgg_output = 1
else:
raise RuntimeError(
'config.ordinal_classification must be sign or regression.'
)
vgg = vgg16.model_struct()
train_mode = tf.get_variable(name='training', initializer=True)
# Mask NAN images from loss
image_nan = tf.reduce_sum(
tf.cast(tf.is_nan(train_images), tf.float32),
reduction_indices=[1, 2, 3])
gedi_nan = tf.reduce_sum(
tf.cast(tf.is_nan(train_gedi_images), tf.float32),
reduction_indices=[1, 2, 3],
keep_dims=True)
image_mask = tf.cast(tf.equal(image_nan, 0.), tf.float32)
gedi_nan = tf.cast(tf.equal(gedi_nan, 0.), tf.float32)
train_images = tf.where(
tf.is_nan(train_images),
tf.zeros_like(train_images),
train_images)
train_gedi_images = tf.where(
tf.is_nan(train_gedi_images),
tf.zeros_like(train_gedi_images),
train_gedi_images)
train_images = tf.concat([train_images, train_images, train_images], axis=3)
val_images = tf.concat([val_images, val_images, val_images], axis=3)
vgg.build(
train_images, output_shape=vgg_output,
train_mode=train_mode, batchnorm=config.batchnorm_layers)
# Prepare the cost function
if config.ordinal_classification is None:
# Encode y w/ k-hot and yhat w/ sigmoid ce. units capture dist.
cost = softmax_cost(
vgg.fc8,
train_labels,
mask=image_mask)
elif config.ordinal_classification == 'regression':
cost = tf.nn.l2_loss(tf.squeeze(vgg.fc8) - train_labels)
class_loss = cost
tf.summary.scalar("cce cost", cost)
# Weight decay
if config.wd_layers is not None:
_, l2_wd_layers = fine_tune_prepare_layers(
tf.trainable_variables(), config.wd_layers)
l2_wd_layers = [
x for x in l2_wd_layers if 'biases' not in x.name]
if len(l2_wd_layers) > 0:
cost += (config.wd_penalty * tf.add_n(
[tf.nn.l2_loss(x) for x in l2_wd_layers]))
# Optimize
train_op = tf.train.AdamOptimizer(config.new_lr).minimize(cost)
if config.ordinal_classification is None:
train_accuracy = class_accuracy(
vgg.prob, train_labels) # training accuracy
elif config.ordinal_classification == 'regression':
train_accuracy = tf.nn.l2_loss(
tf.squeeze(vgg.fc8) - train_labels)
tf.summary.scalar("training accuracy", train_accuracy)
# Setup validation op
if validation_data is not False:
scope.reuse_variables()
# Validation graph is the same as training except no batchnorm
val_vgg = vgg16.model_struct(
fine_tune_layers=config.fine_tune_layers)
val_vgg.build(val_images, output_shape=vgg_output)
# Calculate validation accuracy
if config.ordinal_classification is None:
val_accuracy = class_accuracy(val_vgg.prob, val_labels)
elif config.ordinal_classification == 'regression':
val_accuracy = tf.nn.l2_loss(tf.squeeze(val_vgg.fc8) - val_labels)
tf.summary.scalar("validation accuracy", val_accuracy)
# Set up summaries and saver
saver = tf.train.Saver(
tf.global_variables(), max_to_keep=config.keep_checkpoints)
summary_op = tf.summary.merge_all()
# Initialize the graph
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))
# Need to initialize both of these if supplying num_epochs to inputs
sess.run(tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer()))
summary_dir = os.path.join(
config.train_summaries, config.which_dataset + '_' + dt_stamp)
summary_writer = tf.summary.FileWriter(summary_dir, sess.graph)
# Set up exemplar threading
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
# Restore model if requested
if config.restore_path is not None:
print '-' * 60
print 'Restoring from a previous model: %s' % config.restore_path
print '-' * 60
saver.restore(sess, config.restore_path)
# Start training loop
np.save(out_dir + 'meta_info', config)
step, losses = 0, [] # val_max = 0
try:
# print response
while not coord.should_stop():
start_time = time.time()
_, loss_value, train_acc = sess.run(
[train_op, cost, train_accuracy])
losses.append(loss_value)
duration = time.time() - start_time
if np.isnan(loss_value).sum():
import ipdb;ipdb.set_trace()
assert not np.isnan(loss_value), 'Model diverged with loss = NaN'
if step % config.validation_steps == 0:
if validation_data is not False:
_, val_acc = sess.run([train_op, val_accuracy])
else:
val_acc -= 1 # Store every checkpoint
# Summaries
summary_str = sess.run(summary_op)
summary_writer.add_summary(summary_str, step)
# Training status and validation accuracy
format_str = (
'%s: step %d, loss = %.2f (%.1f examples/sec; '
'%.3f sec/batch) | Training accuracy = %s | '
'Training %s = %s | Training class loss = %s | '
'Validation accuracy = %s | Validation %s = %s | '
'logdir = %s')
print (format_str % (
datetime.now(), step, loss_value,
config.train_batch / duration, float(duration),
train_acc, extra_im_name, 0.,
0., val_acc, extra_im_name,
0., summary_dir))
# Save the model checkpoint if it's the best yet
if 1: # val_acc >= val_max:
saver.save(
sess, os.path.join(
config.train_checkpoint,
'model_' + str(step) + '.ckpt'), global_step=step)
# Store the new max validation accuracy
# val_max = val_acc
else:
# Training status
format_str = ('%s: step %d, loss = %.2f (%.1f examples/sec; '
'%.3f sec/batch) | Training accuracy = %s | '
'Training %s = %s | Training class loss = %s')
print (format_str % (datetime.now(), step, loss_value,
config.train_batch / duration,
float(duration), train_acc,
extra_im_name, 0.,
0.))
# End iteration
step += 1
except tf.errors.OutOfRangeError:
print('Done training for %d epochs, %d steps.' % (config.epochs, step))
finally:
coord.request_stop()
np.save(os.path.join(config.tfrecord_dir, 'training_loss'), losses)
coord.join(threads)
sess.close()
def test_vgg16(image_dir,
model_file,
output_csv='prediction_file',
training_max=None):
print(image_dir)
# tf.set_random_seed(0)
config = GEDIconfig()
if image_dir is None:
raise RuntimeError(
'You need to supply a directory path to the images.')
combined_files = np.asarray(
glob(os.path.join(image_dir, '*%s' % config.raw_im_ext)))
if len(combined_files) == 0:
raise RuntimeError('Could not find any files. Check your image path.')
config = GEDIconfig()
model_file_path = os.path.sep.join(model_file.split(os.path.sep)[:-1])
print('model file path', model_file_path)
meta_file_pointer = os.path.join(model_file_path,
'train_maximum_value.npz')
if not os.path.exists(meta_file_pointer):
raise RuntimeError(
'Cannot find the training data meta file: train_maximum_value.npz'
'Closest I could find from directory %s was %s.'
'Download this from the link described in the README.md.' %
(model_file_path, glob(os.path.join(model_file_path, '*.npz'))))
meta_data = np.load(meta_file_pointer)
# Prepare image normalization values
if training_max is None:
training_max = np.max(meta_data['max_array']).astype(np.float32)
training_min = np.min(meta_data['min_array']).astype(np.float32)
# Find model checkpoints
ds_dt_stamp = re.split('/', model_file)[-2]
out_dir = os.path.join(config.results, ds_dt_stamp)
print('out_dir', out_dir)
# Make output directories if they do not exist
dir_list = [config.results, out_dir]
[make_dir(d) for d in dir_list]
# Prepare data on CPU
if config.model_image_size[-1] < 3:
print('*' * 60)
print('Warning: model is expecting a H/W/1 image. '
'Do you mean to set the last dimension of '
'config.model_image_size to 3?')
print('*' * 60)
images = tf.placeholder(tf.float32,
shape=[None] + config.model_image_size,
name='images')
# Prepare model on GPU
with tf.device('/gpu:0'):
with tf.variable_scope('cnn'):
vgg = vgg16.model_struct(vgg16_npy_path=config.vgg16_weight_path,
fine_tune_layers=config.fine_tune_layers)
vgg.build(images, output_shape=config.output_shape)
# Setup validation op
scores = vgg.prob
preds = tf.argmax(vgg.prob, 1)
# Set up saver
saver = tf.train.Saver(tf.global_variables())
# Loop through each checkpoint then test the entire validation set
ckpts = [model_file]
ckpt_yhat, ckpt_y, ckpt_scores, ckpt_file_array = [], [], [], []
print('-' * 60)
print('Beginning evaluation')
print('-' * 60)
if config.validation_batch > len(combined_files):
print('Trimming validation_batch size to %s (same as # of files).' %
len(combined_files))
config.validation_batch = len(combined_files)
for idx, c in tqdm(enumerate(ckpts), desc='Running checkpoints'):
dec_scores, yhat, file_array = [], [], []
# Initialize the graph
# sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))
with tf.Session(config=tf.ConfigProto(
allow_soft_placement=True)) as sess:
sess.run(
tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer()))
# Set up exemplar threading
saver.restore(sess, c)
start_time = time.time()
num_batches = np.floor(
len(combined_files) /
float(config.validation_batch)).astype(int)
for image_batch, file_batch in tqdm(image_batcher(
start=0,
num_batches=num_batches,
images=combined_files,
config=config,
training_max=training_max,
training_min=training_min),
total=num_batches):
feed_dict = {images: image_batch}
sc, tyh = sess.run([scores, preds], feed_dict=feed_dict)
dec_scores = np.append(dec_scores, sc)
yhat = np.append(yhat, tyh)
file_array = np.append(file_array, file_batch)
ckpt_yhat.append(yhat)
ckpt_scores.append(dec_scores)
ckpt_file_array.append(file_array)
print('Batch %d took %.1f seconds' %
(idx, time.time() - start_time))
# sess.close()
# Save everything
print('Save npz.')
print(os.path.join(out_dir, 'validation_accuracies'))
np.savez(os.path.join(out_dir, 'validation_accuracies'),
ckpt_yhat=ckpt_yhat,
ckpt_scores=ckpt_scores,
ckpt_names=ckpts,
combined_files=ckpt_file_array)
# Also save a csv with item/guess pairs
try:
dec_scores = np.asarray(dec_scores)
yhat = np.asarray(yhat)
df = pd.DataFrame(np.hstack(
(np.asarray(ckpt_file_array).reshape(-1, 1), yhat.reshape(-1, 1),
dec_scores.reshape(dec_scores.shape[0] // 2, 2))),
columns=[
'files', 'live_guesses', 'classifier score dead',
'classifier score live'
])
output_name = image_dir.split('/')[-1]
if output_name is None or len(output_name) == 0:
output_name = 'output'
df.to_csv(os.path.join(out_dir, '%s.csv' % output_name))
print('Saved csv to: %s' %
os.path.join(out_dir, '%s.csv' % output_name))
except:
print('X' * 60)
print('Could not save a spreadsheet of file info')
print('X' * 60)
# Plot everything
try:
plot_accuracies(ckpt_y, ckpt_yhat, config, ckpts,
os.path.join(out_dir, 'validation_accuracies.png'))
plot_std(ckpt_y, ckpt_yhat, ckpts,
os.path.join(out_dir, 'validation_stds.png'))
plot_cms(ckpt_y, ckpt_yhat, config,
os.path.join(out_dir, 'confusion_matrix.png'))
plot_pr(ckpt_y, ckpt_yhat, ckpt_scores,
os.path.join(out_dir, 'precision_recall.png'))
# plot_cost(
# os.path.join(out_dir, 'training_loss.npy'), ckpts,
# os.path.join(out_dir, 'training_costs.png'))
except:
print('X' * 60)
print('Could not locate the loss numpy')
print('X' * 60)
def train_model(train_dir=None, validation_dir=None):
config = GEDIconfig()
if train_dir is None: # Use globals
train_data = os.path.join(config.tfrecord_dir,
config.tf_record_names['train'])
meta_data = np.load(
os.path.join(config.tfrecord_dir,
'%s_%s' % (config.tvt_flags[0], config.max_file)))
else:
meta_data = np.load(
os.path.join(train_dir,
'%s_%s' % (config.tvt_flags[0], config.max_file)))
# Prepare image normalization values
if config.max_gedi is None:
max_value = np.nanmax(meta_data['max_array']).astype(np.float32)
if max_value == 0:
max_value = None
print 'Derived max value is 0'
else:
print 'Normalizing with empirical max.'
if 'min_array' in meta_data.keys():
min_value = np.min(meta_data['min_array']).astype(np.float32)
print 'Normalizing with empirical min.'
else:
min_value = None
print 'Not normalizing with a min.'
else:
max_value = config.max_gedi
min_value = config.min_gedi
ratio = meta_data['ratio']
print 'Ratio is: %s' % ratio
if validation_dir is None: # Use globals
validation_data = os.path.join(config.tfrecord_dir,
config.tf_record_names['val'])
elif validation_dir is False:
pass # Do not use validation data during training
# Make output directories if they do not exist
dt_stamp = re.split(
'\.', str(datetime.now()))[0].\
replace(' ', '_').replace(':', '_').replace('-', '_')
dt_dataset = config.which_dataset + '_' + dt_stamp + '/'
config.train_checkpoint = os.path.join(config.train_checkpoint,
dt_dataset) # timestamp this run
out_dir = os.path.join(config.results, dt_dataset)
dir_list = [
config.train_checkpoint, config.train_summaries, config.results,
out_dir
]
[make_dir(d) for d in dir_list]
# im_shape = get_image_size(config)
im_shape = config.gedi_image_size
print '-' * 60
print('Training model:' + dt_dataset)
print '-' * 60
# Prepare data on CPU
assert os.path.exists(train_data)
assert os.path.exists(validation_data)
assert os.path.exists(config.vgg16_weight_path)
with tf.device('/cpu:0'):
train_images_0, train_images_1, train_labels, train_times = inputs(
train_data,
config.train_batch,
im_shape,
config.model_image_size[:2],
max_value=max_value,
min_value=min_value,
train=config.data_augmentations,
num_epochs=config.epochs,
normalize=config.normalize,
return_filename=True)
val_images_0, val_images_1, val_labels, val_times = inputs(
validation_data,
config.validation_batch,
im_shape,
config.model_image_size[:2],
max_value=max_value,
min_value=min_value,
num_epochs=config.epochs,
normalize=config.normalize,
return_filename=True)
tf.summary.image('train image frame 0', train_images_0)
tf.summary.image('train image frame 1', train_images_1)
tf.summary.image('validation image frame 0', val_images_0)
tf.summary.image('validation image frame 1', val_images_1)
# Prepare model on GPU
with tf.device('/gpu:0'):
with tf.variable_scope('gedi'):
# Build training GEDI model for frame 0
vgg_train_mode = tf.get_variable(name='vgg_training',
initializer=False)
gedi_model_0 = vgg16.model_struct(
vgg16_npy_path=config.gedi_weight_path, trainable=False)
gedi_model_0.build(prep_images_for_gedi(train_images_0),
output_shape=2,
train_mode=vgg_train_mode)
gedi_scores_0 = gedi_model_0.fc7
with tf.variable_scope('match'):
# Build matching model for frame 0
model_0 = matching_gedi.model_struct()
model_0.build(train_images_0)
# Build frame 0 vector
frame_0 = tf.concat([gedi_scores_0, model_0.output], axis=-1)
# Build output layer
if config.matching_combine == 'concatenate':
output_shape = [int(frame_0.get_shape()[-1]) * 2, 2]
elif config.matching_combine == 'subtract':
output_shape = [int(frame_0.get_shape()[-1]), 2]
else:
raise RuntimeError
# Build GEDI model for frame 1
with tf.variable_scope('gedi', reuse=True):
gedi_model_1 = vgg16.model_struct(
vgg16_npy_path=config.gedi_weight_path, trainable=False)
gedi_model_1.build(prep_images_for_gedi(train_images_1),
output_shape=2,
train_mode=vgg_train_mode)
gedi_scores_1 = gedi_model_1.fc7
with tf.variable_scope('match', reuse=True):
# Build matching model for frame 1
model_1 = matching_gedi.model_struct()
model_1.build(train_images_1)
# Build frame 0 and frame 1 vectors
frame_1 = tf.concat([gedi_scores_1, model_1.output], axis=-1)
with tf.variable_scope('output'):
# Concatenate or subtract
if config.matching_combine == 'concatenate':
output_scores = tf.concat([frame_0, frame_1], axis=-1)
elif config.matching_combine == 'subtract':
output_scores = frame_0 - frame_1
else:
raise NotImplementedError
# Build output layer
output_shape = [int(output_scores.get_shape()[-1]), 2]
output_weights = tf.get_variable(
name='output_weights',
shape=output_shape,
initializer=tf.contrib.layers.xavier_initializer(
uniform=False))
output_bias = tf.get_variable(name='output_bias',
initializer=tf.truncated_normal(
[output_shape[-1]], .0, .001))
decision_logits = tf.nn.bias_add(
tf.matmul(output_scores, output_weights), output_bias)
train_soft_decisions = tf.nn.softmax(decision_logits)
cost = softmax_cost(decision_logits, train_labels)
tf.summary.scalar("cce loss", cost)
cost += tf.nn.l2_loss(output_weights)
# Weight decay
if config.wd_layers is not None:
_, l2_wd_layers = fine_tune_prepare_layers(
tf.trainable_variables(), config.wd_layers)
l2_wd_layers = [
x for x in l2_wd_layers if 'biases' not in x.name
]
if len(l2_wd_layers) > 0:
cost += (config.wd_penalty *
tf.add_n([tf.nn.l2_loss(x)
for x in l2_wd_layers]))
# Optimize
train_op = tf.train.AdamOptimizer(config.new_lr).minimize(cost)
train_accuracy = class_accuracy(train_soft_decisions,
train_labels) # training accuracy
tf.summary.scalar("training accuracy", train_accuracy)
# Setup validation op
if validation_data is not False:
with tf.variable_scope('gedi', reuse=tf.AUTO_REUSE): # FIX THIS
# Validation graph is the same as training except no batchnorm
val_gedi_model_0 = vgg16.model_struct(
vgg16_npy_path=config.gedi_weight_path)
val_gedi_model_0.build(prep_images_for_gedi(val_images_0),
output_shape=2,
train_mode=vgg_train_mode)
val_gedi_scores_0 = val_gedi_model_0.fc7
# Build GEDI model for frame 1
val_gedi_model_1 = vgg16.model_struct(
vgg16_npy_path=config.gedi_weight_path)
val_gedi_model_1.build(prep_images_for_gedi(val_images_1),
output_shape=2,
train_mode=vgg_train_mode)
val_gedi_scores_1 = val_gedi_model_1.fc7
with tf.variable_scope('match', reuse=tf.AUTO_REUSE):
# Build matching model for frame 0
val_model_0 = matching_gedi.model_struct()
val_model_0.build(val_images_0)
# Build matching model for frame 1
val_model_1 = matching_gedi.model_struct()
val_model_1.build(val_images_1)
# Build frame 0 and frame 1 vectors
val_frame_0 = tf.concat([val_gedi_scores_0, val_model_0.output],
axis=-1)
val_frame_1 = tf.concat([val_gedi_scores_1, val_model_1.output],
axis=-1)
# Concatenate or subtract
if config.matching_combine == 'concatenate':
val_output_scores = tf.concat([val_frame_0, val_frame_1],
axis=-1)
elif config.matching_combine == 'subtract':
val_output_scores = val_frame_0 - val_frame_1
else:
raise NotImplementedError
with tf.variable_scope('output', reuse=tf.AUTO_REUSE):
# Build output layer
val_output_weights = tf.get_variable(
name='val_output_weights',
shape=output_shape,
trainable=False,
initializer=tf.contrib.layers.xavier_initializer(
uniform=False))
val_output_bias = tf.get_variable(
name='output_bias',
trainable=False,
initializer=tf.truncated_normal([output_shape[-1]], .0,
.001))
val_decision_logits = tf.nn.bias_add(
tf.matmul(val_output_scores, val_output_weights),
val_output_bias)
val_soft_decisions = tf.nn.softmax(val_decision_logits)
# Calculate validation accuracy
val_accuracy = class_accuracy(val_soft_decisions, val_labels)
tf.summary.scalar("validation accuracy", val_accuracy)
# Set up summaries and saver
saver = tf.train.Saver(tf.global_variables(),
max_to_keep=config.keep_checkpoints)
summary_op = tf.summary.merge_all()
# Initialize the graph
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))
# Need to initialize both of these if supplying num_epochs to inputs
sess.run(
tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer()))
summary_dir = os.path.join(config.train_summaries,
config.which_dataset + '_' + dt_stamp)
summary_writer = tf.summary.FileWriter(summary_dir, sess.graph)
# Set up exemplar threading
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
# Start training loop
np.save(out_dir + 'meta_info', config)
step, losses = 0, [] # val_max = 0
try:
# print response
while not coord.should_stop():
start_time = time.time()
_, loss_value, train_acc, val_acc = sess.run(
[train_op, cost, train_accuracy, val_accuracy])
losses += [loss_value]
duration = time.time() - start_time
if np.isnan(loss_value).sum():
assert not np.isnan(loss_value), 'Model loss = NaN'
if step % config.validation_steps == 0:
if validation_data is not False:
val_acc = sess.run(val_accuracy)
else:
val_acc -= 1 # Store every checkpoint
# Summaries
summary_str = sess.run(summary_op)
summary_writer.add_summary(summary_str, step)
# Training status and validation accuracy
format_str = ('%s: step %d, loss = %.2f (%.1f examples/sec; '
'%.3f sec/batch) | Training accuracy = %s | '
'Validation accuracy = %s | '
'logdir = %s')
print(format_str %
(datetime.now(), step,
loss_value, config.train_batch / duration,
float(duration), train_acc, val_acc, summary_dir))
# Save the model checkpoint if it's the best yet
if 1: # val_acc >= val_max:
saver.save(sess,
os.path.join(config.train_checkpoint,
'model_' + str(step) + '.ckpt'),
global_step=step)
# Store the new max validation accuracy
# val_max = val_acc
else:
# Training status
format_str = ('%s: step %d, loss = %.2f (%.1f examples/sec; '
'%.3f sec/batch) | Training accuracy = %s | '
'Training loss = %s')
print(format_str %
(datetime.now(), step, loss_value, config.train_batch /
duration, float(duration), loss_value))
# End iteration
step += 1
except tf.errors.OutOfRangeError:
print('Done training for %d epochs, %d steps.' % (config.epochs, step))
finally:
coord.request_stop()
np.save(os.path.join(config.tfrecord_dir, 'training_loss'), losses)
coord.join(threads)
sess.close()
def test_vgg16(model_file,
trained_svm,
ims,
dead_ims=None,
output_csv='prediction_file',
training_max=None,
C=1e-3,
k_folds=10):
"""Test an SVM you've trained on a new dataset."""
config = GEDIconfig()
if ims is None:
raise RuntimeError(
'You need to supply a directory path to the images.')
if dead_ims is None:
print 'Assuming all of your images are in the ims folder' + \
'-- will not derive labels to calculate accuracy.'
# if not os.path.exists(trained_svm):
# raise RuntimeError(
# 'Cannot find the trained svm model. Check the path you passed.')
try:
clf = cPickle.load(open(trained_svm, 'rb'))
# clf = model_dict['clf']
# mu = model_dict['mu']
# sd = model_dict['sd']
except:
raise RuntimeError('Cannot find SVM file: %s' % trained_svm)
if dead_ims is not None:
live_files = glob(os.path.join(ims, '*%s' % config.raw_im_ext))
dead_files = glob(os.path.join(dead_ims, '*%s' % config.raw_im_ext))
combined_labels = np.concatenate(
(np.zeros(len(live_files)), np.ones(len(dead_files))))
combined_files = np.concatenate((live_files, dead_files))
else:
live_files = glob(os.path.join(ims, '*%s' % config.raw_im_ext))
combined_labels = None
combined_files = np.asarray(live_files)
if len(combined_files) == 0:
raise RuntimeError('Could not find any files. Check your image path.')
config = GEDIconfig()
model_file_path = os.path.sep.join(model_file.split(os.path.sep)[:-1])
meta_file_pointer = os.path.join(model_file_path,
'train_maximum_value.npz')
if not os.path.exists(meta_file_pointer):
raise RuntimeError(
'Cannot find the training data meta file: train_maximum_value.npz'
'Closest I could find from directory %s was %s.'
'Download this from the link described in the README.md.' %
(model_file_path, glob(os.path.join(model_file_path, '*.npz'))))
meta_data = np.load(meta_file_pointer)
# Prepare image normalization values
if training_max is None:
training_max = np.max(meta_data['max_array']).astype(np.float32)
training_min = np.min(meta_data['min_array']).astype(np.float32)
# Find model checkpoints
ds_dt_stamp = re.split('/', model_file)[-2]
out_dir = os.path.join(config.results, ds_dt_stamp)
# Make output directories if they do not exist
dir_list = [config.results, out_dir]
[make_dir(d) for d in dir_list]
# Prepare data on CPU
images = tf.placeholder(tf.float32,
shape=[None] + config.model_image_size,
name='images')
# Prepare model on GPU
with tf.device('/gpu:0'):
with tf.variable_scope('cnn'):
vgg = vgg16.model_struct(vgg16_npy_path=config.vgg16_weight_path,
fine_tune_layers=config.fine_tune_layers)
vgg.build(images, output_shape=config.output_shape)
# Setup validation op
scores = vgg.fc7
preds = tf.argmax(vgg.prob, 1)
# Set up saver
saver = tf.train.Saver(tf.global_variables())
# Loop through each checkpoint then test the entire validation set
ckpts = [model_file]
ckpt_yhat, ckpt_y, ckpt_scores, ckpt_file_array = [], [], [], []
print '-' * 60
print 'Beginning evaluation'
print '-' * 60
if config.validation_batch > len(combined_files):
print 'Trimming validation_batch size to %s (same as # of files).' % len(
combined_files)
config.validation_batch = len(combined_files)
for idx, c in tqdm(enumerate(ckpts), desc='Running checkpoints'):
dec_scores, yhat, y, file_array = [], [], [], []
# Initialize the graph
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))
sess.run(
tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer()))
# Set up exemplar threading
saver.restore(sess, c)
start_time = time.time()
num_batches = np.floor(
len(combined_files) / float(config.validation_batch)).astype(int)
for image_batch, label_batch, file_batch in tqdm(image_batcher(
start=0,
num_batches=num_batches,
images=combined_files,
labels=combined_labels,
config=config,
training_max=training_max,
training_min=training_min),
total=num_batches):
feed_dict = {images: image_batch}
sc, tyh = sess.run([scores, preds], feed_dict=feed_dict)
dec_scores += [sc]
yhat = np.append(yhat, tyh)
y = np.append(y, label_batch)
file_array = np.append(file_array, file_batch)
ckpt_yhat.append(yhat)
ckpt_y.append(y)
ckpt_scores.append(dec_scores)
ckpt_file_array.append(file_array)
print 'Batch %d took %.1f seconds' % (idx, time.time() - start_time)
sess.close()
# Save everything
new_dt_string = re.split('\.', str(datetime.now()))[0].\
replace(' ', '_').replace(':', '_').replace('-', '_')
np.savez(os.path.join(out_dir, '%s_validation_accuracies' % new_dt_string),
ckpt_yhat=ckpt_yhat,
ckpt_y=ckpt_y,
ckpt_scores=ckpt_scores,
ckpt_names=ckpts,
combined_files=ckpt_file_array)
# Run SVM
all_scores = np.concatenate(dec_scores)
# all_scores = (all_scores - mu) / sd
predictions = clf.predict(all_scores)
if dead_ims is not None:
mean_acc = np.mean(predictions == y)
p_value = randomization_test(y=y, yhat=predictions)
print 'SVM performance: mean accuracy = %s%%, p = %.5f' % (mean_acc,
p_value)
df_col_label = 'true label'
else:
mean_acc, p_value = None, None
y = np.copy(yhat)
df_col_label = 'Dummy column (no labels supplied)'
np.savez(os.path.join(out_dir, '%s_svm_test_data' % new_dt_string),
yhat=yhat,
y=y,
scores=dec_scores,
ckpts=ckpts,
p_value=p_value)
# Also save a csv with item/guess pairs
trimmed_files = np.asarray([
x.split(os.path.sep)[-1] for x in np.asarray(ckpt_file_array).ravel()
])
yhat = np.asarray(yhat)
df = pd.DataFrame(
np.hstack((trimmed_files.reshape(-1, 1), yhat.reshape(-1, 1))),
# y.reshape(-1, 1))),
columns=['files', 'guesses']) # , df_col_label])
df.to_csv(os.path.join(out_dir, 'prediction_file.csv'))
print 'Saved csv to: %s' % out_dir
def test_vgg16(live_ims,
dead_ims,
model_file,
svm_model='svm_model',
output_csv='prediction_file',
training_max=None,
C=1e-3,
k_folds=10):
"""Train an SVM for your dataset on GEDI-model encodings."""
config = GEDIconfig()
if live_ims is None:
raise RuntimeError(
'You need to supply a directory path to the live images.')
if dead_ims is None:
raise RuntimeError(
'You need to supply a directory path to the dead images.')
live_files = glob(os.path.join(live_ims, '*%s' % config.raw_im_ext))
dead_files = glob(os.path.join(dead_ims, '*%s' % config.raw_im_ext))
combined_labels = np.concatenate(
(np.zeros(len(live_files)), np.ones(len(dead_files))))
combined_files = np.concatenate((live_files, dead_files))
if len(combined_files) == 0:
raise RuntimeError('Could not find any files. Check your image path.')
config = GEDIconfig()
model_file_path = os.path.sep.join(model_file.split(os.path.sep)[:-1])
meta_file_pointer = os.path.join(model_file_path,
'train_maximum_value.npz')
if not os.path.exists(meta_file_pointer):
raise RuntimeError(
'Cannot find the training data meta file: train_maximum_value.npz'
'Closest I could find from directory %s was %s.'
'Download this from the link described in the README.md.' %
(model_file_path, glob(os.path.join(model_file_path, '*.npz'))))
meta_data = np.load(meta_file_pointer)
# Prepare image normalization values
if training_max is None:
training_max = np.max(meta_data['max_array']).astype(np.float32)
training_min = np.min(meta_data['min_array']).astype(np.float32)
# Find model checkpoints
ds_dt_stamp = re.split('/', model_file)[-2]
out_dir = os.path.join(config.results, ds_dt_stamp)
# Make output directories if they do not exist
dir_list = [config.results, out_dir]
[make_dir(d) for d in dir_list]
# Prepare data on CPU
images = tf.placeholder(tf.float32,
shape=[None] + config.model_image_size,
name='images')
# Prepare model on GPU
with tf.device('/gpu:0'):
with tf.variable_scope('cnn'):
vgg = vgg16.model_struct(vgg16_npy_path=config.vgg16_weight_path,
fine_tune_layers=config.fine_tune_layers)
vgg.build(images, output_shape=config.output_shape)
# Setup validation op
scores = vgg.fc7
preds = tf.argmax(vgg.prob, 1)
# Set up saver
saver = tf.train.Saver(tf.global_variables())
# Loop through each checkpoint then test the entire validation set
ckpts = [model_file]
ckpt_yhat, ckpt_y, ckpt_scores, ckpt_file_array = [], [], [], []
print '-' * 60
print 'Beginning evaluation'
print '-' * 60
if config.validation_batch > len(combined_files):
print 'Trimming validation_batch size to %s (same as # of files).' % len(
combined_files)
config.validation_batch = len(combined_files)
for idx, c in tqdm(enumerate(ckpts), desc='Running checkpoints'):
dec_scores, yhat, y, file_array = [], [], [], []
# Initialize the graph
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))
sess.run(
tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer()))
# Set up exemplar threading
saver.restore(sess, c)
start_time = time.time()
num_batches = np.floor(
len(combined_files) / float(config.validation_batch)).astype(int)
for image_batch, label_batch, file_batch in tqdm(image_batcher(
start=0,
num_batches=num_batches,
images=combined_files,
labels=combined_labels,
config=config,
training_max=training_max,
training_min=training_min),
total=num_batches):
feed_dict = {images: image_batch}
sc, tyh = sess.run([scores, preds], feed_dict=feed_dict)
dec_scores += [sc]
yhat = np.append(yhat, tyh)
y = np.append(y, label_batch)
file_array = np.append(file_array, file_batch)
ckpt_yhat.append(yhat)
ckpt_y.append(y)
ckpt_scores.append(dec_scores)
ckpt_file_array.append(file_array)
print 'Batch %d took %.1f seconds' % (idx, time.time() - start_time)
sess.close()
# Save everything
np.savez(os.path.join(out_dir, 'validation_accuracies'),
ckpt_yhat=ckpt_yhat,
ckpt_y=ckpt_y,
ckpt_scores=ckpt_scores,
ckpt_names=ckpts,
combined_files=ckpt_file_array)
# Run SVM
svm = LinearSVC(C=C, dual=False, class_weight='balanced')
clf = make_pipeline(preprocessing.StandardScaler(), svm)
predictions = cross_val_predict(clf,
np.concatenate(dec_scores),
y,
cv=k_folds)
cv_performance = metrics.accuracy_score(predictions, y)
p_value = randomization_test(y=y, yhat=predictions)
clf.fit(np.concatenate(dec_scores), y)
# mu = dec_scores.mean(0)
# sd = dec_scores.std(0)
print '%s-fold SVM performance: accuracy = %s%% , p = %.5f' % (
k_folds, np.mean(cv_performance * 100), p_value)
np.savez(
os.path.join(out_dir, 'svm_data'),
yhat=predictions,
y=y,
scores=dec_scores,
ckpts=ckpts,
cv_performance=cv_performance,
p_value=p_value,
k_folds=k_folds,
# mu=mu,
# sd=sd,
C=C)
# Also save a csv with item/guess pairs
try:
trimmed_files = [re.split('/', x)[-1] for x in combined_files]
trimmed_files = np.asarray(trimmed_files)
dec_scores = np.asarray(dec_scores)
yhat = np.asarray(yhat)
df = pd.DataFrame(np.hstack(
(trimmed_files.reshape(-1, 1), yhat.reshape(-1,
1), y.reshape(-1, 1))),
columns=['files', 'guesses', 'true label'])
df.to_csv(os.path.join(out_dir, 'prediction_file.csv'))
print 'Saved csv to: %s' % out_dir
except:
print 'X' * 60
print 'Could not save a spreadsheet of file info'
print 'X' * 60
# save the classifier
with open('%s.pkl' % svm_model, 'wb') as fid:
# model_dict = {
# 'model': clf,
# 'mu': mu,
# 'sd': sd
# }
cPickle.dump(clf, fid)
print 'Saved svm model to: %s.pkl' % svm_model
def test_vgg16(image_dir,
model_file,
autopsy_csv=None,
autopsy_path=None,
output_csv='prediction_file',
target_layer='fc7',
save_npy=False,
shuffle_images=True,
embedding_type='PCA'):
"""Testing function for pretrained vgg16."""
assert autopsy_csv is not None, 'You must pass an autopsy file name.'
assert autopsy_path is not None, 'You must pass an autopsy path.'
# Load autopsy information
autopsy_data = pd.read_csv(os.path.join(autopsy_path, autopsy_csv))
# Load config and begin preparing data
config = GEDIconfig()
if image_dir is None:
raise RuntimeError(
'You need to supply a directory path to the images.')
combined_files = np.asarray(
glob(os.path.join(image_dir, '*%s' % config.raw_im_ext)))
if shuffle_images:
combined_files = combined_files[np.random.permutation(
len(combined_files))]
if len(combined_files) == 0:
raise RuntimeError('Could not find any files. Check your image path.')
config = GEDIconfig()
meta_file_pointer = os.path.join(
model_file.split('/model')[0], 'train_maximum_value.npz')
if not os.path.exists(meta_file_pointer):
raise RuntimeError(
'Cannot find the training data meta file.'
'Download this from the link described in the README.md.')
meta_data = np.load(meta_file_pointer)
# Prepare image normalization values
training_max = np.max(meta_data['max_array']).astype(np.float32)
training_min = np.min(meta_data['min_array']).astype(np.float32)
# Find model checkpoints
ds_dt_stamp = re.split('/', model_file)[-2]
out_dir = os.path.join(config.results, ds_dt_stamp)
# Make output directories if they do not exist
dir_list = [config.results, out_dir]
[make_dir(d) for d in dir_list]
# Prepare data on CPU
images = tf.placeholder(tf.float32,
shape=[None] + config.model_image_size,
name='images')
# Prepare model on GPU
with tf.device('/gpu:0'):
with tf.variable_scope('cnn'):
vgg = vgg16.model_struct(vgg16_npy_path=config.vgg16_weight_path,
fine_tune_layers=config.fine_tune_layers)
vgg.build(images, output_shape=config.output_shape)
# Setup validation op
scores = vgg[target_layer]
preds = tf.argmax(vgg.prob, 1)
# Derive pathologies from file names
pathologies = []
for f in combined_files:
sf = f.split('/')[-1].split('_')
sf = '_'.join(sf[1:4])
it_path = autopsy_data[autopsy_data['plate_well_neuron'] ==
sf]['disease']
if not len(it_path):
it_path = 'Absent'
else:
it_path = it_path.as_matrix()[0]
pathologies += [it_path]
pathologies = np.asarray(pathologies)
# Set up saver
saver = tf.train.Saver(tf.global_variables())
# Loop through each checkpoint then test the entire validation set
ckpts = [model_file]
ckpt_yhat, ckpt_scores, ckpt_file_array = [], [], []
print '-' * 60
print 'Beginning evaluation'
print '-' * 60
if config.validation_batch > len(combined_files):
print 'Trimming validation_batch size to %s.' % len(combined_files)
config.validation_batch = len(combined_files)
for idx, c in tqdm(enumerate(ckpts), desc='Running checkpoints'):
dec_scores, yhat, file_array = [], [], []
# Initialize the graph
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))
sess.run(
tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer()))
# Set up exemplar threading
saver.restore(sess, c)
start_time = time.time()
num_batches = np.floor(
len(combined_files) / float(config.validation_batch)).astype(int)
for image_batch, file_batch in tqdm(image_batcher(
start=0,
num_batches=num_batches,
images=combined_files,
config=config,
training_max=training_max,
training_min=training_min),
total=num_batches):
feed_dict = {images: image_batch}
sc, tyh = sess.run([scores, preds], feed_dict=feed_dict)
dec_scores += [sc]
yhat += [tyh]
file_array += [file_batch]
ckpt_yhat.append(yhat)
ckpt_scores.append(dec_scores)
ckpt_file_array.append(file_array)
print 'Batch %d took %.1f seconds' % (idx, time.time() - start_time)
sess.close()
# Create and plot an embedding
im_path_map = pathologies[:num_batches * config.validation_batch]
dec_scores = np.concatenate(dec_scores)
mu = dec_scores.mean(0)[None, :]
sd = dec_scores.std(0)[None, :]
dec_scores = (dec_scores - mu) / sd
yhat = np.concatenate(yhat)
file_array = np.concatenate(file_array)
if embedding_type == 'TSNE' or embedding_type == 'tsne':
emb = manifold.TSNE(n_components=2, init='pca', random_state=0)
elif embedding_type == 'PCA' or embedding_type == 'pca':
emb = PCA(n_components=2, svd_solver='randomized', random_state=0)
elif embedding_type == 'spectral':
emb = manifold.SpectralEmbedding(n_components=2, random_state=0)
y = emb.fit_transform(dec_scores)
# Ouput csv
df = pd.DataFrame(np.hstack(
(y, im_path_map.reshape(-1, 1), file_array.reshape(-1, 1))),
columns=['D1', 'D2', 'pathology', 'filename'])
out_name = os.path.join(out_dir, 'embedding.csv')
df.to_csv(out_name)
print 'Saved csv to: %s' % out_name
# Create plot
f, ax = plt.subplots()
unique_cats = np.unique(im_path_map)
h = []
for idx, cat in enumerate(unique_cats):
h += [
plt.scatter(y[im_path_map == cat, 0],
y[im_path_map == cat, 1],
c=plt.cm.Spectral(idx * 1000))
]
plt.legend(h, unique_cats)
plt.axis('tight')
plt.show()
plt.savefig('embedding.png')
plt.close(f)
# Save everything
if save_npy:
np.savez(os.path.join(out_dir, 'validation_accuracies'),
ckpt_yhat=ckpt_yhat,
ckpt_scores=ckpt_scores,
ckpt_names=ckpts,
combined_files=ckpt_file_array)
