def evaluate(args, model, criterion, val_loader):
model.eval()
losses = AverageMeter()
jaccars = AverageMeter()
dices = AverageMeter()
eva = Evaluation()
for i, (images, labels) in enumerate(val_loader):
if args.cuda:
images = images.cuda()
labels = labels.cuda()
criterion = criterion.cuda()
images = Variable(images)
labels = Variable(labels)
outputs = model(images)
loss = criterion(outputs, labels)
losses.update(loss.data.cpu().numpy())
jacc_index = eva.jaccard_similarity_coefficient(
outputs.cpu().data.numpy().squeeze(),
labels.cpu().data.numpy())
dice_index = eva.dice_coefficient(outputs.cpu().data.numpy().squeeze(),
labels.cpu().data.numpy())
jaccars.update(jacc_index)
dices.update(dice_index)
return losses.avg, jaccars.avg, dices.avg
def main():
# Directory Setting
train_dir = "./data/multi_train.csv"
test_dir = "./data/multi_test.csv"
model_dir = "./model_save"
# HyperParameter
epoch = 1
batch = 128
max_len = 50
hidden_units = 64
target_names = ['0', '1', '2', '3']
# Flow
print("0. Setting Environment")
set_env()
print("1. load data")
train_x, train_y, test_x, test_y, val_x, val_y = load_data(
train_dir, test_dir, len(target_names))
print("2. pre processing")
train_x, val_x, test_x = train_x.tolist(), val_x.tolist(), test_x.tolist()
train_x = [' '.join(t.split()[0:max_len]) for t in train_x]
train_x = np.array(train_x, dtype=object)[:, np.newaxis]
val_x = [' '.join(t.split()[0:max_len]) for t in val_x]
val_x = np.array(val_x, dtype=object)[:, np.newaxis]
test_x = [' '.join(t.split()[0:max_len]) for t in test_x]
test_x = np.array(test_x, dtype=object)[:, np.newaxis]
print("3. build model")
model = ELMo(hidden_units=hidden_units,
data_type="multi",
category_size=len(target_names))
model.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy'])
callbacks = create_callbacks(model_dir)
model.fit(x=train_x,
y=train_y,
epochs=epoch,
batch_size=batch,
validation_data=(val_x, val_y),
callbacks=callbacks)
print("4. evaluation")
evaluation = Evaluation(model, test_x, test_y)
accuracy, cf_matrix, report = evaluation.eval_classification(
data_type="multi")
print("## Target Names : ", target_names)
print("## Classification Report \n", report)
print("## Confusion Matrix \n", cf_matrix)
print("## Accuracy \n", accuracy)
def main():
# Directory Setting
train_dir = "./data/multi_train.csv"
test_dir = "./data/multi_test.csv"
model_dir = "./model_save"
embedding_dir = "./glove.6B.50d.txt"
# HyperParameter
epoch = 1
batch = 256
embedding_dim = 50
target_names = ['0', '1', '2', '3']
# Flow
print("0. Setting Environment")
set_env()
print("1. load data")
train_x, train_y, test_x, test_y, val_x, val_y = load_data(
train_dir, test_dir, len(target_names))
print("2. pre processing")
train_x, test_x, val_x, tokenizer = pre_processing(train_x, test_x, val_x)
print("3. text to vector")
embedding_matrix = text_to_vector(tokenizer.word_index,
embedding_dir,
word_dimension=embedding_dim)
print("4. build model")
model = TextCNN(sequence_len=train_x.shape[1],
embedding_matrix=embedding_matrix,
embedding_dim=embedding_dim,
filter_sizes=[3, 4, 5],
flag="pre_training",
data_type="multi",
category_num=len(target_names))
model.compile(optimizer="adam",
loss='categorical_crossentropy',
metrics=['accuracy'])
callbacks = create_callbacks(model_dir)
model.fit(x=train_x,
y=train_y,
epochs=epoch,
batch_size=batch,
validation_data=(val_x, val_y),
callbacks=callbacks)
print("5. evaluation")
evaluation = Evaluation(model, test_x, test_y)
accuracy, cf_matrix, report = evaluation.eval_classification(
data_type="multi")
print("## Target Names : ", target_names)
print("## Classification Report \n", report)
print("## Confusion Matrix \n", cf_matrix)
print("## Accuracy \n", accuracy)
def get_ys(self, tag):
if tag is not None:
y_name = Evaluation.shorten_bioconcept(tag)
ys = [1 if y_col == y_name else 0 for y_col in self.COLUMNS_BY_SOURCE['y']]
else:
ys = np.zeros(len(self.COLUMNS_BY_SOURCE['y']))
return list(ys)
def train(self, df):
metrics_by_y = dict.fromkeys(self.COLUMNS_BY_SOURCE['y'])
for i, this_y in enumerate(self.COLUMNS_BY_SOURCE['y']):
print(f'\n{this_y}\n')
other_ys = [y for y in self.COLUMNS_BY_SOURCE['y'] if y != this_y]
selected_columns = [
c for c in df.columns
if c not in other_ys and c not in self.COLUMNS_TO_EXCLUDE and c not in self.COLUMNS_TO_NOT_ANALYSE]
kingdom = 'plant' if i <= 2 else 'animal'
kingdom_df = df.loc[df.kingdom == kingdom]
if kingdom == 'animal':
self.investigate_shit_features(kingdom_df)
X_train, X_test, y_train, y_test = self.preprocess_dataset_and_split(kingdom_df, selected_columns)
X_train, y_train = self.up_or_down_sample(X_train, y_train)
logreg = LogisticRegression(solver='liblinear')
filtered_X_train = self.auto_or_manual_feature_selection(logreg, X_train, y_train)
logit_model = statsmodels.api.Logit(y_train, filtered_X_train)
result = self.fit_and_check_singular_matrix(logit_model, filtered_X_train, y_train)
print(result.summary2())
print('-------------------------------------------------------------')
logreg.fit(filtered_X_train, y_train)
y_pred = logreg.predict(X_test.loc[:, filtered_X_train.columns])
metrics = Evaluation.calculate_metrics(y_test, y_pred)
metrics_by_y[this_y] = metrics
metrics_df = pandas.DataFrame(metrics_by_y)
average_f1 = np.mean(metrics_df.T.f1) * 100
print(metrics_df.T)
print(f'average f1 score: {average_f1:.0f}%')
def main():
# Directory Setting
train_dir = "../data/binary_train.csv"
test_dir = "../data/binary_test.csv"
model_dir = "./model_save"
# HyperParameter
epoch = 2
batch = 256
# Flow
print("0. Setting Environment")
set_env()
print("1. load data")
train_x, train_y, test_x, test_y, val_x, val_y = load_data(
train_dir, test_dir)
print("2. pre processing")
train_x, test_x, val_x, tokenizer = pre_processing(train_x, test_x, val_x)
print("3. build model")
model = TextCNN(sequence_len=train_x.shape[1],
embedding_matrix=len(tokenizer.word_index) + 1,
embedding_dim=300,
filter_sizes=[3, 4, 5],
flag="self_training",
data_type="binary")
model.compile(loss='binary_crossentropy',
optimizer='adam',
metrics=['accuracy'])
callbacks = create_callbacks(model_dir)
model.fit(x=train_x,
y=train_y,
epochs=epoch,
batch_size=batch,
validation_data=(val_x, val_y),
callbacks=callbacks)
print("4. evaluation")
evaluation = Evaluation(model, test_x, test_y)
accuracy, cf_matrix, report = evaluation.eval_classification(
data_type="binary")
print("## Classification Report \n", report)
print("## Confusion Matrix \n", cf_matrix)
print("## Accuracy \n", accuracy)
def evaluate(args, model, criterion, val_loader):
model.eval()
losses = AverageMeter()
jaccars = AverageMeter()
dices = AverageMeter()
eva = Evaluation()
for i, (images, labels) in enumerate(val_loader):
if args.cuda:
images = images.cuda()
labels = labels.cuda()
criterion = criterion.cuda()
images = Variable(images)
labels = Variable(labels)
outputs = model(images)
# print('images(eval) : ', images.shape)
# print('labels(eval) : ', labels.shape)
# print('outputs(eval) : ', outputs.shape)
outputs = outputs.view(1, -1)
# outputs = outputs.view(1, *outputs.size()[-2:])
# print('output_size_changed(eval) : ', outputs.size())
labels = labels.view(1, -1)
# labels = labels.view(1, *labels.size())
# print('labels.view(changed) : ', labels.size())
jacc_index = eva.jaccard_similarity_coefficient(
outputs.cpu().data.numpy().squeeze(),
labels.cpu().data.numpy())
dice_index = eva.dice_coefficient(outputs.cpu().data.numpy().squeeze(),
labels.cpu().data.numpy())
#loss = criterion(outputs, labels)
crit = criterion(outputs, labels)
loss = crit #- jacc_index
# print('no jacc loss : ', crit)
# print(' jacc loss : ', loss)
losses.update(loss.data.cpu().numpy())
jaccars.update(jacc_index)
dices.update(dice_index)
return losses.avg, jaccars.avg, dices.avg
def main():
# Hyper parameter
batch_size = 64
lr = 0.001
epochs = 3
n_classes = 2
embedding_dim = 300
hidden_dim = 32
use_cuda = torch.cuda.is_available()
device = torch.device("cuda" if use_cuda else "cpu")
# Directory
train_dir = "../data/binary_train_data.csv"
test_dir = "../data/binary_test_data.csv"
model_dir = "snapshot/text_classification.pt"
print("1.Load data")
train_data, valid_data, test_data, text, label = load_data(
train_dir, test_dir)
print("2.Pre processing")
train_iter, val_iter, test_iter, text, label = pre_processing(
train_data, valid_data, test_data, text, label, device, batch_size)
print("3.Build model")
model = BaseModel(hidden_dim=32,
vocab_num=len(text.vocab),
embedding_dim=300,
class_num=len(vars(label.vocab)["itos"])).to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=lr)
print("4.Train")
best_val_loss = None
for e in range(1, epochs + 1):
model = training(model, optimizer, train_iter, device)
val_loss, val_accuracy = Evaluation(model, val_iter, device)
print("[Epoch: %d] val loss : %5.2f | val accuracy : %5.2f" %
(e, val_loss, val_accuracy))
save_model(best_val_loss, val_loss, model, model_dir)
model.load_state_dict(torch.load(model_dir))
test_loss, test_acc = Evaluation(model, test_iter, device)
print('테스트 오차: %5.2f | 테스트 정확도: %5.2f' % (test_loss, test_acc))
def __init__(self, cfg, data, mode):
self.cfg = cfg
self.mode = mode
self.data = data
self.eval_obj = Evaluation()
if mode != 'export':
self.build_model()
self.saver = tf.train.Saver(var_list=tf.global_variables(),
max_to_keep=1000)
def train(trainset, model, optimizer, loss_function, testset):
index = 0
losses = []
acc = []
for epoch in range(6):
total_loss = torch.Tensor([0])
for instance in trainset:
print(index)
index += 1
for question in instance['questions']:
text = stemming(instance['text'])
ques = stemming(question['question'])
for answer in question['answers']:
model.zero_grad()
ans = stemming(answer['answer'])
output = model(text, ques, ans)
if answer['correct'] == 'True':
y = autograd.Variable(torch.FloatTensor([1]))
else:
y = autograd.Variable(torch.FloatTensor([0]))
print("output", output.data[0][0])
# avoid 0 gradient
if output.data[0][0] == 0:
output = output + autograd.Variable(
torch.FloatTensor([0.0001]))
if output.data[0][0] == 1:
output = output - autograd.Variable(
torch.FloatTensor([0.0001]))
loss = loss_function(output, y)
# print('output', output.data[0])
# print('loss', loss.data[0])
loss.backward()
optimizer.step()
# for param in model.parameters():
# print('param', param.grad.data.sum())
total_loss += loss.data[0]
losses.append(total_loss)
y, predicty = test(testset, model)
# print(len(y))
eval = Evaluation()
acc.append(eval.accuracy(y, predicty, data))
return losses, acc
def run_experiments(**exp_params) -> None:
logger = init_logger(exp_params)
trajectory_generator, training_task_descriptors, test_task_confs, controls, dataset, \
session, model, meta_learner, lhs_tasks, test_observations = init_experiments(**exp_params)
meta_learner.train_model()
evaluation = None
if exp_params["evaluation"]:
evaluation = Evaluation(test_task_grid=test_task_confs,
meta_learner=meta_learner,
kwargs=exp_params,
test_observations=test_observations)
evaluation.evaluation_on_test_tasks(dataset=dataset,
test_tasks_params=test_task_confs,
iteration=0,
controls=controls)
for iteration in range(exp_params["task_budget"]):
latent_task_variables_mean, latent_task_variables_var = meta_learner.get_H_space_subset(
end_task_id=exp_params["n_initial_training_envs"] + iteration)
candidates = discretise_region(
latent_task_variables_mean=latent_task_variables_mean,
slack_min_values=exp_params["slack_min_intervals"],
slack_max_values=exp_params["slack_max_intervals"],
grid_resolution=exp_params["candidate_grid_size"])
candidates = filter_candidates(
latent_task_variables_mean=latent_task_variables_mean,
task_configurations=training_task_descriptors,
candidates=candidates,
config_space=exp_params["observed_configuration_space_interval"],
verbose=exp_params["verbose"],
GPModel=model,
session=session)
logger.info(f"Number of candidates: {candidates.shape}")
selected_task_descriptor = acquire_task(
iteration=iteration,
latent_task_variables_mean=latent_task_variables_mean,
latent_task_variables_var=latent_task_variables_var,
discretised_latent_space_region=candidates,
task_descriptors=training_task_descriptors,
meta_learner=meta_learner,
lhs_tasks=lhs_tasks,
model=model,
**exp_params)
logger.info(f"Acquired task configuration: {selected_task_descriptor}")
dataset.add_configuration(new_configuration=selected_task_descriptor)
acquired_task_observations = trajectory_generator.observe_trajectories(
task_configurations=selected_task_descriptor,
controls=controls,
dim_states=exp_params["dim_states"])[0]
meta_learner, training_task_descriptors = add_new_task(
iteration=iteration,
meta_learner=meta_learner,
acquired_task_observations=acquired_task_observations,
controls=controls,
training_task_descriptors=training_task_descriptors,
selected_task_descriptor=selected_task_descriptor,
**exp_params)
meta_learner.train_model()
if exp_params["evaluation"]:
evaluation.evaluation_on_test_tasks(
dataset=dataset,
test_tasks_params=test_task_confs,
iteration=(iteration + 1),
controls=controls)
which_set='train',
image_size=FLAGS.image_size,
shuffle=False)
data_loader_val = DataLoader(FLAGS,
which_set='test',
image_size=FLAGS.image_size,
shuffle=False)
p_data_loader_train = ParallDataWraper(data_loader_train,
batch_size=FLAGS.batch_size,
thread=3)
p_data_loader_val = ParallDataWraper(data_loader_val,
batch_size=FLAGS.batch_size,
thread=3)
## Configure evaluator
evaluator = Evaluation(FLAGS, data_loader_train)
## Configure the CNN model
img_shape = [FLAGS.image_size, FLAGS.image_size]
feed_img_batch = tf.placeholder(tf.float32, shape=[None] + img_shape + [3])
with tf.variable_scope('PRETRAINED_CNN'):
print('--> init the image model ...')
CNN, CNN_all_vars = classifier.get_main_network(
'inception',
input_tensor=feed_img_batch,
num_classes=3,
use_weights=False)
CNN_last_conv = CNN.layers[-3].output
CNN_all_vars = [
var for var in tf.trainable_variables() if 'PRETRAINED_CNN' in var.name
def run(self):
items, _ = self.fit_to_validation()
evaluation = Evaluation(items, verbose=False)
evaluation.run()
def main():
# Directory Setting
train_dir = "../data/multi_train.csv"
test_dir = "../data/multi_test.csv"
model_dir = "./model_save"
# HyperParameter
epoch = 2
batch = 256
max_len = 50
hidden_units = 64
target_names = ['0', '1', '2', '3']
# Flow
print("0. Setting Environment")
set_env()
print("1. load data")
train_x, train_y, test_x, test_y, val_x, val_y = load_data(train_dir, test_dir, len(target_names))
print("2. pre processing")
train_x, val_x, test_x = train_x.tolist(), val_x.tolist(), test_x.tolist()
train_x = [' '.join(t.split()[0:max_len]) for t in train_x]
train_x = np.array(train_x, dtype=object)[:, np.newaxis]
val_x = [' '.join(t.split()[0:max_len]) for t in val_x]
val_x = np.array(val_x, dtype=object)[:, np.newaxis]
test_x = [' '.join(t.split()[0:max_len]) for t in test_x]
test_x = np.array(test_x, dtype=object)[:, np.newaxis]
tokenizer = create_tokenizer_from_hub_module()
train_examples = convert_text_to_examples(train_x, train_y)
val_examples = convert_text_to_examples(val_x, val_y)
test_examples = convert_text_to_examples(test_x, test_y)
train_input_ids, train_input_masks, train_segment_ids, train_labels = convert_examples_to_features(tokenizer, train_examples, max_len)
val_input_ids, val_input_masks, val_segment_ids, val_labels = convert_examples_to_features(tokenizer, val_examples, max_len)
test_input_ids, test_input_masks, test_segment_ids, test_labels = convert_examples_to_features(tokenizer, test_examples, max_len)
print("3. build model")
model = BERT(max_len, data_type="multi", category_size=len(target_names))
model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
initialize_vars(sess)
cp_callback = tf.keras.callbacks.ModelCheckpoint(filepath=model_dir+"/model-weights.{epoch:02d}-{val_acc:.6f}.hdf5", monitor='val_acc', save_best_only=True, verbose=1)
model.fit(
[train_input_ids, train_input_masks, train_segment_ids], train_labels,
validation_data=([val_input_ids, val_input_masks, val_segment_ids], val_labels),
epochs=epoch,
batch_size=batch,
callbacks=[cp_callback]
)
print("4. evaluation")
evaluation = Evaluation(model, [test_input_ids, test_input_masks, test_segment_ids], test_y)
accuracy, cf_matrix, report = evaluation.eval_classification_bert(data_type="multi")
print("## Classification Report \n", report)
print("## Confusion Matrix \n", cf_matrix)
print("## Accuracy \n", accuracy)
def run(self):
model_dir_by_bioconcept = self.train_models_or_get_dirs()
predictions = self.predict_validation_data(model_dir_by_bioconcept)
evaluation = Evaluation(predictions, verbose=False)
evaluation.run()
def run_one_epoch(parameters, config, device, epoch, mode):
model = parameters["model"]
if mode == "train":
model.train()
optimizer = parameters["optimizer"]
elif mode == "valid" or mode == "test":
model.eval()
else:
raise NotImplementedError
dataset = copy.deepcopy(parameters["dataset_{}".format(mode)])
pred = {}
total_loss = 0
evaluation = Evaluation(config)
for step, data in enumerate(dataset):
for key in data:
if isinstance(data[key], torch.Tensor):
data[key] = data[key].to(device)
if mode == "train":
optimizer.zero_grad()
if config.get("model", "model_name") == "Crf":
if mode != "test":
results = model(data=data, mode=mode, crf_mode="train")
loss = results["loss"]
total_loss += loss.item()
results = model(data=data, mode=mode, crf_mode="test")
evaluation.expand(results["prediction"], results["labels"])
else:
results = model(data=data, mode=mode, crf_mode="test")
prediction = results["prediction"]
if not isinstance(prediction, list):
prediction = prediction.cpu().numpy().tolist()
docids = data["docids"]
canids = data["canids"]
for doc, can, pre in zip(docids, canids, prediction):
if doc not in pred.keys():
pred[doc] = []
assert (len(can) == len(pre))
for c, p in zip(can, pre):
if p != "O":
p = p[2:]
assert p in Global.type2id.keys()
pred[doc].append({
"id": c,
"type_id": Global.type2id[p]
})
else:
results = model(data=data, mode=mode)
if mode != "test":
loss = results["loss"]
total_loss += loss.item()
evaluation.expand(results["prediction"], results["labels"])
else:
prediction = results["prediction"].cpu().numpy().tolist()
docids = data["docids"]
canids = data["canids"]
for did, cid, pre in zip(docids, canids, prediction):
if did not in pred.keys():
pred[did] = []
pred[did].append({"id": cid, "type_id": pre})
if mode != "test":
print("\r{}: Epoch {} Step {:0>4d}/{} | Loss = {:.4f}".format(
mode, epoch, step + 1, len(dataset),
round(total_loss / (step + 1), 4)),
end="")
else:
print("\r{}: Epoch {} Step {:0>4d}/{}".format(
mode, epoch, step + 1, len(dataset)),
end="")
if mode == "train":
loss.backward()
optimizer.step()
if mode != "test":
metric = evaluation.get_metric("all")
sys.stdout.write("\r")
print("\r{}: Epoch {} | Metric: {}".format(mode, epoch, metric))
return metric
else:
return pred
# train
rnn = RNN(100, 128, len(vocab))
optimizer = optim.SGD(rnn.parameters(), lr=0.1)
loss_function = nn.BCELoss()
losses, acc = rnn_train(data.trainset, rnn, optimizer, loss_function,
data.testset)
# plt.xlabel("Train epoch")
# plt.ylabel("loss")
# plt.plot([1, 2, 3, 4, 5, 6, 7, 8, 9, 10], losses)
# plt.show()
plt.xlabel("Train epoch")
plt.ylabel("accuracy")
plt.plot([1, 2, 3, 4, 5, 6], acc)
plt.show()
# test
y, predicty = rnn_test(data.testset, rnn)
eval = Evaluation()
eval.accuracy(y, predicty, data)
with open('result_rnn.txt', 'w') as f:
for index, maxd in enumerate(eval.wrong):
f.write("Case #{}: {} ".format(index + 1, maxd) + '\n')
# predicty=[0.1, 0.2, 0.55, 0.51, 0.53, 0.7]
# use common sense
predicty = use_csk(data.testset, predicty, rnn)
# Evaluation
eval = Evaluation()
eval.accuracy(y, predicty, data)
final = time.time()
print("time:", final - begin)
def __init__(self):
self.eval = Evaluation()
self.board = Dashboad(8098)
class Experiments():
def __init__(self):
self.eval = Evaluation()
self.board = Dashboad(8098)
def error_hist(self, gtdir, resdir, imgprefix, plot=False):
listGTFiles = [
k.split('/')[-1].split('.')[0]
for k in glob.glob(os.path.join(gtdir, '*.bmp'))
]
filename_jacc = dict()
filename_dice = dict()
filename_sens = dict()
filename_spec = dict()
for currfile in tqdm(listGTFiles):
if currfile.count('_') == 2:
continue
gt = np.array(Image.open(os.path.join(gtdir,
currfile + '.bmp'))) / 255
res = np.array(
Image.open(
os.path.join(resdir, currfile + '_' + imgprefix + '.bmp')))
res[res > 10] = 255
res /= 255
jacc_index = self.eval.jaccard_similarity_coefficient(
gt.squeeze(), res.squeeze())
dice = self.eval.dice_coefficient(gt.squeeze(), res.squeeze())
spec, sens, _ = self.eval.specificity_sensitivity(
gt.squeeze(), res.squeeze())
filename_jacc[currfile] = jacc_index
filename_dice[currfile] = dice
filename_sens[currfile] = sens
filename_spec[currfile] = spec
if plot:
self.board.metric_bar(filename_jacc.values(),
'Jaccard_' + imgprefix,
nbins=20)
self.board.metric_bar(filename_dice.values(),
'Dice_' + imgprefix,
nbins=20)
self.board.metric_bar(filename_sens.values(),
'Sens_' + imgprefix,
nbins=20)
self.board.metric_bar(filename_spec.values(),
'Spec_' + imgprefix,
nbins=20)
return filename_jacc, filename_dice, filename_sens, filename_spec
def get_failure_cases(self, args, threshold=0.5):
list_dics = []
for m in args.methods:
result, _, _, _ = self.error_hist(args.gtdir,
args.resdir,
m,
plot=False)
list_dics.append(result)
# Remove the images with jaccard greater than 0.5
for d in list_dics:
for k, v in d.items():
if v > threshold:
del d[k]
# Find the failure cases common between all methods
common_failures = set.intersection(*tuple(
set(d.keys()) for d in list_dics))
return common_failures
# TODO Remove val argument in function
def make_grid(self, args, val=True, selected_filenames=None):
bordersize = 2
batch = np.empty((0, 3, 244, 324))
num2sample = 60
if selected_filenames is not None:
filenames = list(selected_filenames)
num2sample = len(filenames)
else:
if val:
filenames = [
k.split('.')[-2].split('/')[-1]
for k in glob.glob(os.path.join(args.imgdir, "val_*"))
]
if args.test:
filenames = [
k.split('.')[-2].split('/')[-1]
for k in glob.glob(os.path.join(args.imgdir, "test_*.bmp"))
]
else:
filenames = [
k.split('.')[-2].split('/')[-1]
for k in glob.glob(os.path.join(args.imgdir, "*.bmp"))
]
train_ind = np.random.choice(np.arange(0, len(filenames)),
num2sample,
replace=False)
for i in range(train_ind.shape[0]):
currfile = filenames[train_ind[i]]
im = np.array(
Image.open(os.path.join(args.imgdir,
currfile + ".bmp")).convert('RGB'))
# Applies a border on the top of the image
im = cv2.copyMakeBorder(im,
top=bordersize,
bottom=bordersize,
left=bordersize,
right=bordersize,
borderType=cv2.BORDER_CONSTANT,
value=[255, 0, 0])
im = cv2.putText(im,
currfile, (10, 20),
fontFace=cv2.FONT_HERSHEY_SIMPLEX,
fontScale=0.75,
color=(255, 255, 255),
thickness=2)
im = im.transpose((2, 0, 1))
batch = np.append(batch, im[np.newaxis, :, :, :], axis=0)
if val:
im = np.array(
Image.open(os.path.join(args.gtdir,
currfile + ".bmp")).convert('L'))
im = np.repeat(im[:, :, np.newaxis], 3, axis=2)
im = cv2.copyMakeBorder(im,
top=bordersize,
bottom=bordersize,
left=bordersize,
right=bordersize,
borderType=cv2.BORDER_CONSTANT,
value=[0, 255, 0])
im = im.transpose((2, 0, 1))
batch = np.append(batch, im[np.newaxis, :, :, :], axis=0)
for m in args.methods:
res = np.array(
Image.open(
os.path.join(args.resdir, currfile + "_" + m +
".bmp")).convert('L'))
res = np.repeat(res[:, :, np.newaxis], 3, axis=2)
res = cv2.copyMakeBorder(res,
top=bordersize,
bottom=bordersize,
left=bordersize,
right=bordersize,
borderType=cv2.BORDER_CONSTANT,
value=[0, 255, 200])
# Writes the name of the models.
res = cv2.putText(res,
m, (10, 20),
fontFace=cv2.FONT_HERSHEY_SIMPLEX,
fontScale=0.75,
color=(255, 255, 255),
thickness=2)
res = res.transpose((2, 0, 1))
batch = np.append(batch, res[np.newaxis, :, :, :], axis=0)
return batch
from opts import *
from TopicMDNet import TopicMDNet as MDNet
# configuration session
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
K.set_session(sess) # set sess before initialize variable...
## Configure data loader
data_loader_val = DataLoader(FLAGS, which_set='train', image_size=FLAGS.image_size, shuffle=False, use_augmentation=False)
p_data_loader_val = ParallDataWraper(data_loader_val, batch_size=FLAGS.batch_size, thread=3)
## Configure evaluator
evaluator = Evaluation(FLAGS, data_loader_val)
## Configure the CNN model
feed_img_batch = tf.placeholder(tf.float32, shape=[None, FLAGS.image_size, FLAGS.image_size, 3])
with tf.variable_scope('PRETRAINED_CNN'):
print ('--> init the image model ...')
CNN, CNN_all_vars = classifier.get_main_network('inception', input_tensor=feed_img_batch, num_classes=3, use_weights=False)
CNN_last_conv = CNN.layers[-3].output
CNN_all_vars = [var for var in tf.trainable_variables() if 'PRETRAINED_CNN' in var.name ]
FLAGS.conv_feat_dim = int(CNN_last_conv.shape[3])
## Configure the language model model
LSTM = MDNet(FLAGS, data_loader_val,
CNN_last_conv,
CNN.output, # classifier logit
CNN.layers[-1].weights[0].op.name, # the classifier weight op name for get_collection in MDNet
def run(self):
results = self.fit_to_validation()
evaluation = Evaluation(results, verbose=False)
evaluation.run()
def train(args, model):
# board = Dashboad(args.visdom_port) #visdom
tr_losses = AverageMeter()
tLoader, vLoader = load_data(args)
criterion = nn.BCELoss()
# criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(model.parameters(),
lr=args.lr,
weight_decay=args.weight_decay,
momentum=0.99)
scheduler = lr_scheduler.StepLR(optimizer, step_size=30, gamma=0.1)
for epoch in range(1, args.num_epochs + 1):
scheduler.step()
if epoch == 1:
tr_loss, _, _ = evaluate(args, model, criterion, tLoader)
vl_loss, vl_jacc, vl_dice = evaluate(args, model, criterion,
vLoader)
# Draw the loss curves
win = None
# win = board.loss_curves([tr_loss, vl_loss], epoch, win=win) #visdom
print('[Initial TrainLoss: {0:.4f}]'
'\t[Initial ValidationLoss: {1:.4f}]'
'\t[Initial ValidationJaccard: {2:.4f}]'
'\t[Initial ValidationDice: {3:.4f}]'.format(
tr_loss, vl_loss, vl_jacc, vl_dice))
print(
'----------------------------------------------------------------------------------------------------'
'--------------')
for step, (images, labels) in enumerate(tLoader):
model.train(True)
if args.cuda:
images = images.cuda()
labels = labels.cuda()
criterion = criterion.cuda()
inputs = Variable(images)
targets = Variable(labels)
optimizer.zero_grad()
outputs = model(inputs)
eva = Evaluation()
# jacc = eva.jaccard_similarity_coefficient(outputs.cpu().data.numpy().squeeze(),
# labels.cpu().data.numpy())
output_np = outputs.cpu().data.numpy().squeeze()
targets_np = targets.cpu().data.numpy()
dice = eva.dice_coefficient(output_np, targets_np)
crit = criterion(outputs, targets)
loss = crit - dice
print('bceloss : ', crit.data)
print('bceloss - dice : ', loss.data)
loss.backward()
optimizer.step()
tr_losses.update(loss.data.cpu().numpy())
if epoch % args.log_step == 0:
vl_loss, vl_jacc, vl_dice = evaluate(args, model, criterion,
vLoader)
print('[Epoch: {0:02}/{1:02}]'
'\t[TrainLoss: {2:.4f}]'
'\t[ValidationLoss: {3:.4f}]'
'\t[ValidationJaccard: {4:.4f}]'
'\t[ValidationDice: {5:.4f}]'.format(epoch, args.num_epochs,
tr_losses.avg, vl_loss,
vl_jacc, vl_dice)),
filename = "weights/{0}-{1:02}.pth".format(args.model, epoch)
torch.save(model.state_dict(), filename)
print(' [Snapshot]')
else:
vl_loss, vl_jacc, vl_dice = evaluate(args, model, criterion,
vLoader)
print('[Epoch: {0:02}/{1:02}]'
'\t[TrainLoss: {2:.4f}]'
'\t[ValidationLoss: {3:.4f}]'
'\t[ValidationJaccard: {4:.4f}]'
'\t[ValidationDice: {5:.4f}]'.format(epoch, args.num_epochs,
tr_losses.avg, vl_loss,
vl_jacc, vl_dice))