def test(model_name, threshold=0.5, save=True, verbose=True, refine=False):
classifications = np.array([0, 0, 0, 0])
results_folder = os.path.join(model_name, 'results')
if not os.path.exists(results_folder): os.mkdir(results_folder)
_, test_set = get_dataset_split()
if refine: test_set = sort_imgs(test_set)
prediction = None
model = keras.models.load_model(os.path.join(model_name,
model_name + '.h5'),
custom_objects=get_custom_objects())
for i in range(len(test_set)):
if verbose: display_progress(i / len(test_set))
img_path, gt_path = test_set[i].replace('\n', '').split(',')
img = read_image(img_path, pad=(4, 4))
img = normalise_img(img)
ground_truth = read_gt(gt_path)
ground_truth = np.squeeze(ground_truth)
if refine:
prediction = ground_truth if prediction is None else prediction
pmap = create_map(prediction > 0.5, 1)
prob = get_prob_map(pmap)
prediction = model.predict(img)
prediction = np.squeeze(prediction)
prediction = prediction[4:-4, ...]
prediction = (prediction > threshold).astype(np.uint8)
if refine: prediction = prediction * prob
classifications += getPixels(prediction, ground_truth, 0.5)
if save:
save_image(prediction,
os.path.join(results_folder, ntpath.basename(img_path)))
save_image(
ground_truth,
os.path.join(
results_folder,
ntpath.basename(img_path).replace('.png', '_gt.png')))
# if refine:
# prob = prob.astype(np.uint8)
# save_image(os.path.join(results_folder,
# ntpath.basename(img_path).replace('.png', '_prob.png')), prob)
print(model_name, threshold)
printMetrics(getMetrics(classifications))
def main():
saver = tf.train.Saver()
if os.path.isdir('model'):
shutil.rmtree('model')
#Add labels to saved model as json file for classification
os.makedirs('model')
with open(utils.train_data_path, 'w') as file:
train_data = {
'labels': utils.labels
}
file.write(json.dumps(train_data))
#Loading dataset and displaying information about it
print('Count of images:')
dataset_count = 0
for labelID in range(len(utils.dataset)):
label_count = len(utils.dataset[labelID])
dataset_count += label_count
print('{0}: {1}'.format(utils.labels[labelID], label_count))
print('Total: {0}'.format(dataset_count))
print('Loading test batch...')
test_images, test_labels = utils.random_batch(utils.config['training']['batch_size_test'], True)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
last_acc = 0
for i in range(utils.config['training']['n_epoch']+1):
batch_X, batch_Y = utils.random_batch()
# the back-propagation training step
sess.run(net.train_step, feed_dict={net.X: batch_X, net.Y_: batch_Y, net.pkeep: utils.config['training']['pkeep']})
acc, loss = sess.run([net.accuracy, net.cross_entropy], feed_dict={net.X: test_images, net.Y_: test_labels, net.pkeep: 1.0})
#Model accuracy has improved since last epoch
if acc > last_acc:
last_acc = acc
#Remove older save
for file_name in os.listdir(os.path.join(utils.root_dir, 'model')):
file_path = os.path.join(utils.root_dir, 'model', file_name)
if os.path.isfile(file_path) and file_name != utils.train_data_name:
os.unlink(file_path)
#Save new model
save_path = saver.save(sess, utils.model_dir)
#Display information obout progress
utils.display_progress(i, acc, loss)
headers = utils.get_headers(token)
username = raw_input('Username: '******'next']
# measures progress by page, not by result
current_page += 1
total_pages = (data['count'] / 20.0)
utils.display_progress(current_page, total_pages)
for result in data['results']:
request_id = result['id']
print "Working on request " + str(request_id)
# get request first
request_url = api_url + 'foia/%d/' % request_id
print request_url
request = requests.get(request_url, headers=headers)
print request
request_data = request.json()
# get agency second
agency_url = api_url + 'agency/%d/' % request_data['agency']
agency = requests.get(agency_url , headers=headers)
agency_data = agency.json()
# get communications third
]
# forward pass
logits = classifier(trn_imgs_batch_mixup_device)
losses = criterion(logits, trn_lbls_oh_batch_mixup_device)
total_trn_loss = .8 * losses[0] + .1 * losses[1] + .1 * losses[2]
total_trn_loss.backward()
optimizer.step()
# record
epoch_trn_loss.append(total_trn_loss.item())
utils.display_progress(len(training_loader), j + 1,
{'training_loss': epoch_trn_loss[-1]})
# validation
classifier.eval()
with torch.no_grad():
for k, (vld_imgs_batch,
vld_lbls_batch) in enumerate(validation_loader):
# move to device
vld_imgs_batch_device = vld_imgs_batch.cuda()
vld_lbls_batch_device = [l.cuda() for l in vld_lbls_batch]
vld_lbls_batch_numpy = [
l.detach().cpu().numpy() for l in vld_lbls_batch
]
url = utils.API_URL
token = utils.get_api_key()
headers = utils.get_headers(token)
username = raw_input('Username: '******'next']
# measures progress by page, not by result
current_page += 1
total_pages = (data['count'] / 20.0)
utils.display_progress(current_page, total_pages)
for request in data['results']:
request_id = request['id']
request_url = 'https://www.muckrock.com/api_v1/foia/%s/' % str(
request_id)
# Removes the embargo date to make sure it permanently embargos.
data = json.dumps({
'embargo': True,
'date_embargo': None,
'permanent_embargo': True
})
http_request.patch(request_url, headers=headers, data=data)
