def _load_model(sess, model_folder):
"""
Loads the model for inference, if the given folder contains a protobuffer loads the serialized version
:param sess: TF session
:model_folder: The model folder
:return: tuple with TF placeholders (image_input, logits, keep_prob)
"""
if helper.is_model_serialized(model_folder):
graph = helper.load_serialized_model(sess, MODEL_NAME, _model_folder())
image_input = graph.get_tensor_by_name('image_input:0')
keep_prob = graph.get_tensor_by_name('keep_prob:0')
model_output = graph.get_tensor_by_name('model_output:0')
else:
vgg_path = helper.maybe_download_pretrained_vgg(FLAGS.data_dir)
image_input, keep_prob, layer3, layer4, layer7 = load_vgg(
sess, vgg_path)
model_output = layers(layer3, layer4, layer7, CLASSES_N)
helper.load_model(sess, _model_folder())
logits = tf.reshape(model_output, (-1, CLASSES_N))
return image_input, logits, keep_prob
def __init__(self):
self.detection_graph = None
self.tl_map = {
1: TrafficLight.GREEN,
2: TrafficLight.RED,
3: TrafficLight.YELLOW,
4: TrafficLight.UNKNOWN
}
self.detection_graph = tf.Graph()
model_path = rospy.get_param('~pb_path')
load_model(model_path, self.detection_graph)
# Model warmup - run a dummy classification against a random image
rospy.loginfo("Model warmup starting")
with tf.Session(graph=self.detection_graph) as sess:
image_tensor = sess.graph.get_tensor_by_name('image_tensor:0')
detection_boxes = sess.graph.get_tensor_by_name(
'detection_boxes:0')
detection_scores = sess.graph.get_tensor_by_name(
'detection_scores:0')
detection_classes = sess.graph.get_tensor_by_name(
'detection_classes:0')
#
gen_image = np.uint8(np.random.randn(1, 600, 800, 3))
sess.run([detection_boxes, detection_scores, detection_classes],
feed_dict={image_tensor: gen_image})
#
rospy.loginfo("Model warmup done")
def main():
ap = argparse.ArgumentParser(description='predict.py')
ap.add_argument('image_path',
nargs='*',
action="store",
default="./flowers",
type=str)
ap.add_argument('checkpoint', action="store", type=str)
ap.add_argument('--top_k',
dest="top_k",
action="store",
type=int,
default=1)
ap.add_argument('--category_names',
dest="category_names",
action="store",
type=str)
ap.add_argument('--gpu', dest="gpu", action='store_true')
args = ap.parse_args()
if args.category_names:
with open(args.category_names, 'r') as f:
cat_to_name = json.load(f)
else:
cat_to_name = None
model = helper.load_model(args.checkpoint)
ps, cl = helper.predict(args.image_path[0], model, args.top_k, args.gpu)
helper.show_results(ps, cl, args.top_k, cat_to_name)
def main(image_path, saved_model_path, top_k=1, json_path=None):
'''
Input: image_path, saved_model_path, top_k, json_path
Output: plot with the photo provided with the predicted class as title, together with a histogram
with top_k most likely classes and their probabilities.
'''
model = load_model(saved_model_path)
probs, class_probs, image = predict(image_path, model, top_k)
plot(image, probs, class_probs, json_path, top_k)
def evaluate_model(args, params):
loader_params = {
'batch_size': params['batch_size'],
'shuffle': False,
'num_workers': params['num_workers']
}
batch_size = params['batch_size']
_, _, test_loader = data_handler.init_data_loaders(params, loader_params)
total_predicted = np.zeros((batch_size, 14))
total_labels = np.zeros((batch_size, 14))
path_to_load = helper.compute_paths(
args, params)['save_path'] # compute path from args and params
net = networks.init_unified_net(args.model, params)
net, _ = helper.load_model(path_to_load,
args.epoch,
net,
optimizer=None,
resume_train=False)
print(
f'In [evaluate_model]: loading the model done from: "{path_to_load}"')
print(
f'In [evaluate_model]: starting evaluation with {len(test_loader)} batches'
)
with torch.no_grad():
for i_batch, batch in enumerate(test_loader):
img_batch = batch['image'].to(device).float()
label_batch = batch['label'].to(device).float()
pred = net(img_batch)
if i_batch > 0:
total_predicted = np.append(total_predicted,
pred.cpu().detach().numpy(),
axis=0)
total_labels = np.append(total_labels,
label_batch.cpu().detach().numpy(),
axis=0)
else:
total_predicted = pred.cpu().detach().numpy()
total_labels = label_batch.cpu().detach().numpy()
if i_batch % 50 == 0:
print(
f'In [evaluate_model]: prediction done for batch {i_batch}'
)
results_path = helper.compute_paths(args, params)['results_path']
helper.make_dir_if_not_exists(results_path)
# plot roc
print(f'In [evaluate_model]: starting plotting ROC...')
plotting.plot_roc(total_predicted, total_labels, pathology_names,
results_path)
def main():
stop_words = get_stop_words(STOP_WORDS_PATH)
data = Initialize_Data();
visualizer = Visualize();
data.initialize_twitter_posts(TWITTER_POSTS_CSV, TWITTER_DATA_DIR)
data.initialize_facebook_posts(FACEBOOK_POSTS_CSV, FACEBOOK_DATA_DIR)
# Visalize daya
df = np.array(data.posts);
lf= np.array(data.labels);
pos_ind = lf == "positive";
neg_ind = lf == "negative"
pos = df[pos_ind]
neg = df[neg_ind]
visualizer.plot_data_distibution([pos.shape[0], neg.shape[0]], ["positive", "negative"], "Training set distribution")
# Cleanup posts
text_Cleanuper = Posts_Cleansing(data)
text_Cleanuper.cleanup(Text_Cleanuper())
# Train and Test Model
clf = train_test_model(create_ngram_model(frozenset(stop_words)), np.array(data.posts), np.array(data.labels) == "positive")
# Find best Model params and train
clf = grid_search_model(create_ngram_model, np.array(data.posts), np.array(data.labels) == "positive", frozenset(stop_words))
print('Saving model')
save_model(clf, NAIVE_BAYES_MODEL_PATH);
print('Loading model')
trained_model = load_model(NAIVE_BAYES_MODEL_PATH)
train_test_model(trained_model, np.array(data.posts), np.array(data.labels) == "positive")
importance = get_most_important_features(trained_model.named_steps['vect'].vocabulary_.items(), trained_model.named_steps['clf'], 10)
top_scores = [a[0] for a in importance[0]['tops']]
top_words = [a[1] for a in importance[0]['tops']]
bottom_scores = [a[0] for a in importance[0]['bottom']]
bottom_words = [a[1] for a in importance[0]['bottom']]
visualizer.plot_important_words(top_scores, top_words, bottom_scores, bottom_words, "Most important words for relevance")
Y_predicted_word2vec = trained_model.predict(["Նա վատ աղջիկ է"])
print(Y_predicted_word2vec)
def main():
res = load_model("bert_en_cased_L-12_H-768_A-12-v3", verbose=1)
pathname, output, inputs, outputs, onnx_inputs = res
output_names = outputs
structured_outputs = [
"answer_types", "tf_op_layer_end_logits", "tf_op_layer_start_logits",
"unique_ids"
]
perf_iter = 5
rtol = 0.01
atol = 0.0001
print("[main] testing ONNX %r" % output)
m = rt.InferenceSession(output)
results_onnx = m.run(output_names.split(','), onnx_inputs)
print("[main] got results, testing perf")
start = time.time()
for _ in range(perf_iter):
_ = m.run(output_names.split(','), onnx_inputs)
onnx_runtime_ms = (time.time() - start) / perf_iter * 1000
print("[main] ONNX perf:", onnx_runtime_ms)
print("[main] loading TF")
imported = tf.saved_model.load(".", tags=['serve'])
concrete_func = imported.signatures["serving_default"]
tf_inputs = {}
for k, v in onnx_inputs.items():
tf_inputs[k.split(":")[0]] = tf.constant(v)
tf_func = tf.function(concrete_func)
print("[main] running TF")
tf_results_d = tf_func(**tf_inputs)
#results_tf = [tf_results_d[output].numpy() for output in structured_outputs]
print("[main] got results, testing perf")
start = time.time()
for _ in range(perf_iter):
_ = concrete_func(**tf_inputs)
tf_runtime_ms = (time.time() - start) / perf_iter * 1000
print("[main] TF perf:", tf_runtime_ms)
# for tf_res, onnx_res in zip(results_tf, results_onnx):
# np.testing.assert_allclose(tf_res, onnx_res, rtol=rtol, atol=atol)
print("[main] Results match")
print('[main] device', rt.get_device(), rt.__version__, rt.__file__)
print("[main] TF perf, ONNX perf, ratio")
print("[main]", tf_runtime_ms, onnx_runtime_ms,
tf_runtime_ms / onnx_runtime_ms)
def __init__(self):
root_path = os.path.dirname(os.path.dirname(__file__))
if torch.cuda.is_available():
self.device = torch.device("cuda")
gpu_ids = list(range(torch.cuda.device_count()))
else:
self.device = torch.device("cpu")
gpu_ids = []
state_dict = torch.load(os.path.join(root_path, "models",
"trained_models",
"densenet201.pth"),
map_location=self.device)
self.model = load_model(25, state_dict)
self.model = self.model.to(self.device)
def classify_image(path_to_image, checkpoint, top_k, category_names, gpu):
if not torch.cuda.is_available() and gpu:
raise (
"No gpu available to train the network. Please remove the --gpu argument to train using the cpu"
)
device = ('cuda' if gpu else 'cpu')
model = helper.load_model(checkpoint)
image_tensor = torch.tensor(helper.load_image(path_to_image))
(probs, classes) = helper.predict(image_tensor, model, top_k, device)
if category_names != None:
#convert the classes array to hold the string representation of the category
with open(category_names, 'r') as f:
cat_to_name = json.load(f)
classes = [cat_to_name[class_] for class_ in classes]
return (classes, probs)
def test_gray_scale():
# path = 'data_big_original/extracted/images'
path = 'data_big'
h5_file = 'chest_xray.h5'
partition, labels, labels_hot = \
data_handler.read_already_partitioned(h5_file)
# just one image that exists now
partition = {'train': [], 'validation': [], 'test': ['00000001_000.png']}
# labels, labels_hot = None, None
preprocess = helper.preprocess_fn(no_crop=True) # does not crop the images
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
loader_params = {'batch_size': 1, 'shuffle': False, 'num_workers': 0}
_, _, test_loader = data_handler.create_data_loaders(partition,
labels,
labels_hot,
path,
preprocess,
device,
loader_params,
scale='gray')
# model_path = "models/max_epochs=30_batch_size=256_pool_mode=max_lr=5e-05_no_crop=True/unified_net_epoch_23.pt"
# model_path = 'models/max_epochs=30_batch_size=256_pool_mode=max_lr=0.0001_no_crop=True/unified_net_epoch_8.pt'
model_path = "models/max_epochs=50_batch_size=128_pool_mode=max_lr=0.0001_no_crop=True_es=True_26253292/unified_net_epoch_1.pt"
# reading the other params from the JSON file
with open('params.json', 'r') as f:
params = json.load(f)
net = helper.load_model(model_path, device, params['transition_params'],
'resnet34')
for _, batch in enumerate(test_loader):
image = batch['image'].to(device).float()
print(f'image shape: {image.shape}')
model_name = model_path.split("/")[1]
helper.plot_heatmaps(image,
net,
resize_dim=(256, 256),
save_path=f'figures/{model_name}/',
show_or_save='show')
if __name__ == "__main__":
dict_path = model_path = args.output_base_path + args.task+'/'
dict_path += 'dictionary.p'
model_path += args.model_file_name #'model_best.pth.tar'
dictionary = helper.load_object(dict_path)
embeddings_index = helper.load_word_embeddings(args.word_vectors_directory, args.word_vectors_file,
dictionary.word2idx)
model = BCN(dictionary, embeddings_index, args)
if args.cuda:
torch.cuda.set_device(args.gpu)
model = model.cuda()
print('loading model')
helper.load_model(model, model_path, 'state_dict', args.cuda)
print('vocabulary size = ', len(dictionary))
task_names = ['snli', 'multinli'] if args.task == 'allnli' else [args.task]
for task in task_names:
test_corpus = data.Corpus(args.tokenize)
if 'IMDB' in args.task:
###############################################################################
# Load Learning to Skim paper's Pickle file
###############################################################################
train_d, dev_d, test_d = helper.get_splited_imdb_data(args.output_base_path+task+'/'+'imdb.p')
test_corpus.parse(test_d, task, args.max_example)
# test_corpus.parse(args.output_base_path + task + '/' + args.test + '.txt', 'RT', args.max_example) #although IMDB but selected text saved by budget model from theano in 'RT' format
def train():
# 配置 Tensorboard，每次训练的结果保存在以日期时间命名的文件夹中。
print("Configuring TensorBoard and Saver...")
tensorboard_dir = 'tensorboard/textcnn' + '/' + time.strftime(
'%Y%m%d%H%M', time.localtime(time.time()))
if not os.path.exists(tensorboard_dir):
os.makedirs(tensorboard_dir)
tf.summary.scalar("loss", model.loss)
tf.summary.scalar("accuracy", model.acc)
merged_summary = tf.summary.merge_all()
writer = tf.summary.FileWriter(tensorboard_dir)
# 载入训练集与验证集
print("Loading training and validation data...")
start_time = time.time()
x_train, y_train, x_val, y_val = load_data(temp_dir, train_dir, val_dir,
word_to_id, cat_to_id,
config.seq_length)
time_dif = get_time_dif(start_time)
print("Time usage:", time_dif)
total_batch = tf.Variable(0, trainable=False) # 总批次，不可训练的变量
# 创建session
session = tf.Session()
# 导入权重
saver = load_model(session, save_dir)
# 图写入tensorboard
writer.add_graph(session.graph)
print('Training and evaluating...')
start_time = time.time()
best_acc_val = 0.0 # 最佳验证集准确率
last_improved = session.run(total_batch) # 记录上一次提升批次
require_improvement = 5000 # 如果超过1000轮未提升，提前结束训练
flag = False
for epoch in range(config.num_epochs):
print('Epoch:', epoch + 1)
batch_train = batch_iter(x_train, y_train, config.batch_size)
for x_batch, y_batch in batch_train:
feed_dict = feed_data(model, x_batch, y_batch,
config.dropout_keep_prob)
if session.run(total_batch) % config.save_per_batch == 0:
# 每多少轮次将训练结果写入tensorboard scalar
s = session.run(merged_summary, feed_dict=feed_dict)
writer.add_summary(s, session.run(total_batch))
if session.run(total_batch) % config.print_per_batch == 0:
# 每多少轮次输出在训练集和验证集上的性能
loss_train, F1_train, _, _ = evaluate(session, model, x_train,
y_train)
loss_val, F1_val, _, _ = evaluate(session, model, x_val,
y_val) # todo
if F1_val > best_acc_val:
# 保存最好结果
best_acc_val = F1_val
last_improved = session.run(total_batch)
saver.save(sess=session, save_path=save_path)
improved_str = '*'
else:
improved_str = ''
time_dif = get_time_dif(start_time)
msg = 'Iter: {0:>6}, Train Loss: {1:>6.2}, Train F1: {2:>7.2%},' \
+ ' Val Loss: {3:>6.2}, Val F1: {4:>7.2%}, Time: {5} {6}'
print(
msg.format(session.run(total_batch), loss_train, F1_train,
loss_val, F1_val, time_dif, improved_str))
session.run(model.optim, feed_dict=feed_dict) # 运行优化
session.run(tf.assign(
total_batch, total_batch +
1)) # 用tf.assign迭代total_batch可以在saver中记录total_batch的变化
if session.run(total_batch) - last_improved > require_improvement:
# 验证集正确率长期不提升，提前结束训练
print("No optimization for a long time, auto-stopping...")
flag = True
break # 跳出循环
if flag: # 同上
break
session.close()
if l2_status[i]:
name += '_l2'
if bn_status[i]:
name += '_bn'
if noise:
name += '_noise'
if adv_status[n]:
name += '_adv'
print(name)
model_path = utils.make_directory(params_path, model_name)
file_path = os.path.join(model_path, name)
# load model parameters
model_layers, optimization, _ = helper.load_model(
model_name, input_shape, dropout_status[i], l2_status[i],
bn_status[i])
# build neural network class
nnmodel = nn.NeuralNet(seed=247)
nnmodel.build_layers(model_layers,
optimization,
supervised=True)
nntrainer = nn.NeuralTrainer(nnmodel,
save='best',
file_path=file_path)
# initialize session
sess = utils.initialize_session()
def train_nn(sess,
global_step,
epochs,
batch_size,
get_batches_fn,
batches_n,
train_op,
cross_entropy_loss,
prediction_op,
metrics,
metrics_reset_op,
image_input,
labels,
keep_prob,
learning_rate,
save_model_freq=None,
tensorboard_freq=None):
"""
Train neural network and print out the loss during training.
:param sess: TF Session
:param global_step: TF Placeholder containing the global step
:param epochs: Number of epochs
:param batch_size: Batch size
:param get_batches_fn: Function to get batches of training data. Call using get_batches_fn(batch_size)
:param batches_n: Number of batches to cover all the samples
:param train_op: TF Operation to train the neural network
:param cross_entropy_loss: TF Tensor for the amount of loss
:param prediction_op: TF Tensor for the prediction class (index)
:param metrics: Dictionary with the evaluation metrics
:param metric_reset_op: TF Tensor used to reset the metrics counters
:param image_input: TF Placeholder for input images
:param labels: TF Placeholder for label images
:param keep_prob: TF Placeholder for dropout keep probability
:param learning_rate: TF Placeholder for learning rate
:param save_model_freq: The frequency to save the model to disk, None to disable
:param tensorboard_freq: The frequency to push the summaries to tensorboard, None to disable
"""
model_folder = _model_folder()
if save_model_freq and helper.checkpoint_exists(model_folder):
print(
'Checkpoint exists, restoring model from {}'.format(model_folder))
helper.load_model(sess, model_folder)
else:
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
if save_model_freq:
saver = tf.train.Saver(max_to_keep=MODELS_LIMIT)
iou_mean, iou_op = metrics['iou']
acc_mean, acc_op = metrics['acc']
# Evaluate current step
step = global_step.eval(session=sess)
start_step = step
if tensorboard_freq:
# Creates the tensorboard writer
train_writer = _summary_writer(sess, model_folder)
# Gets the batch of images/labels to feed to the image summary op
summary_images, summary_labels = helper.image_summary_batch(
os.path.join(FLAGS.data_dir, 'data_road', 'training'), IMAGE_SHAPE,
TENSORBOARD_MAX_IMG)
# Setup the summary ops
summary_op, image_summary_op = _setup_summaries(
sess, train_writer, image_input, labels, keep_prob,
cross_entropy_loss, prediction_op, iou_mean, acc_mean,
summary_images, summary_labels, step, CLASSES_N)
training_log = []
print('Model folder: {}'.format(model_folder))
print(
'Training (First batch: {}, Epochs: {}, Batch Size: {}, Learning Rate: {}, Dropout: {}, L2 Reg: {}, Scaling: {})'
.format(step + 1, FLAGS.epochs, FLAGS.batch_size, FLAGS.learning_rate,
FLAGS.dropout, FLAGS.l2_reg, 'ON' if FLAGS.scale else 'OFF'))
best_loss = 9999
ep_loss_incr = 0
start = time.time()
for epoch in range(epochs):
total_loss = 0
mean_loss = 9999
mean_acc = 0
mean_iou = 0
images_n = 0
# Resets the metrics variables at the beginning of the epoch
sess.run(metrics_reset_op)
batches = tqdm(
get_batches_fn(batch_size),
desc=
'Epoch {}/{} (Step: {}, Samples: N/A, Loss: N/A, Acc: N/A, IoU: N/A)'
.format(epoch + 1, epochs, step),
unit='batches',
total=batches_n)
for batch_images, batch_labels in batches:
feed_dict = {
image_input: batch_images,
labels: batch_labels,
keep_prob: (1.0 - FLAGS.dropout),
learning_rate: FLAGS.learning_rate
}
# Train
_ = sess.run(train_op, feed_dict=feed_dict)
images_n += len(batch_images)
# Evaluate
loss, _, mean_iou, _, mean_acc = sess.run(
[cross_entropy_loss, iou_op, iou_mean, acc_op, acc_mean],
feed_dict={
image_input: batch_images,
labels: batch_labels,
keep_prob: 1.0
})
step = global_step.eval(session=sess)
total_loss += loss * len(batch_images)
mean_loss = total_loss / images_n
# Saves metrics for tensorboard
if tensorboard_freq:
# Updates the summary according to frequency
if step % tensorboard_freq == 0:
training_summary = sess.run(summary_op,
feed_dict={
image_input: batch_images,
labels: batch_labels,
keep_prob: 1.0
})
train_writer.add_summary(training_summary,
global_step=step)
# Writes the image every epoch
if step % batches_n == 0:
image_pred_summary = sess.run(image_summary_op,
feed_dict={
image_input:
summary_images,
labels: summary_labels,
keep_prob: 1.0
})
train_writer.add_summary(image_pred_summary,
global_step=step)
train_writer.flush()
batches.set_description(
'Epoch {}/{} (Step: {}, Samples: {}, Loss: {:.4f}, Acc: {:.4f}, IoU: {:.4f})'
.format(epoch + 1, epochs, step, images_n, mean_loss, mean_acc,
mean_iou))
training_log.append((mean_loss, mean_acc, mean_iou))
if mean_loss < best_loss:
ep_loss_incr = 0
best_loss = mean_loss
else:
ep_loss_incr += 1
if FLAGS.early_stopping is not None and ep_loss_incr >= FLAGS.early_stopping:
print(
'Early Stopping Triggered (Loss not decreasing in the last {} epochs)'
.format(ep_loss_incr))
break
if save_model_freq and (epoch + 1) % save_model_freq == 0:
helper.save_model(sess, saver, MODEL_NAME, model_folder,
global_step)
log_data = _to_log_data(training_log, start_step, step, batches_n)
helper.save_log(log_data, model_folder)
helper.plot_log(log_data, model_folder)
elapsed = time.time() - start
print(
'Training Completed ({:.1f} s): Last batch: {}, Loss: {:.4f}, Acc: {:.4f}, IoU: {:.4f}'
.format(elapsed, step, mean_loss, mean_acc, mean_iou))
if save_model_freq:
helper.save_model(sess, saver, MODEL_NAME, model_folder, global_step)
log_data = _to_log_data(training_log, start_step, step, batches_n)
helper.save_log(log_data, model_folder)
helper.plot_log(log_data, model_folder)
from torch import randn from torch import onnx from linknet_batch import linknet_batch_model from helper import load_model exp_model = linknet_batch_model() exp_model = load_model(exp_model, model_dir="linknet_10epch.pt") dummy_input = randn(1, 3, 512, 512) onnx.export(exp_model, dummy_input, "linknet.onnx")
if __name__ == '__main__':
parMapNet = ParametersMapNet()
parIL = Parameters_IL()
action_list = np.asarray(parMapNet.action_list)
if parIL.use_predefined_test_set:
# Open predefined initial configurations and load them in avd
avd = AVD_online(par=parIL, nStartPos=0, scene_list=parIL.predefined_test_scenes,
action_list=action_list, init_configs=parIL.predefined_confs_file)
else:
# sample random starting positions and targets from the AVD_online class
avd = AVD_online(par=parIL, nStartPos=10, scene_list=["Home_001_1"], action_list=action_list)
test_ids = list(range(len(avd)))
# Need to load the trained MapNet
if parIL.finetune_mapNet: # choose whether to use a finetuned mapNet model or not
mapNet_model = hl.load_model(model_dir=parIL.model_dir, model_name="MapNet", test_iter=parIL.test_iters)
else:
mapNet_model = hl.load_model(model_dir=parIL.mapNet_model_dir, model_name="MapNet", test_iter=parIL.mapNet_iters)
if parIL.use_ego_obsv:
ego_encoder = Encoder()
ego_encoder.cuda()
ego_encoder.eval()
else:
ego_encoder = None
evaluate_NavNet(parIL, parMapNet, mapNet_model, ego_encoder, test_iter=parIL.test_iters,
test_ids=test_ids, test_data=avd, action_list=action_list)
def main():
stop_words = get_stop_words(STOP_WORDS_PATH)
data = Initialize_Data()
visualizer = Visualize()
data.initialize_twitter_posts(TWITTER_POSTS_CSV, TWITTER_DATA_DIR)
data.initialize_facebook_posts(FACEBOOK_POSTS_CSV, FACEBOOK_DATA_DIR)
# Cleanup posts
text_Cleanuper = Posts_Cleansing(data)
text_Cleanuper.cleanup(Text_Cleanuper())
# Divide data into test and train set
X_train, X_test, Y_train, Y_test = train_test_split(data.posts,
data.labels,
test_size=0.2,
random_state=40)
# Bag of Words model vectorization
bag_of_words_model = Bag_Of_Words(X_train)
bag_of_words_model.build_vectorizer(stop_words)
X_train_counts = bag_of_words_model.data_counts
X_test_counts = bag_of_words_model.vectorizer.transform(X_test)
forest = RandomForestClassifier(n_estimators=100)
forest = forest.fit(X_train_counts, Y_train)
y_predicted_counts_train = forest.predict(X_train_counts)
accuracy, precision, recall, f1 = get_metrics(Y_train,
y_predicted_counts_train)
print("Train accuracy = %.3f, precision = %.3f, recall = %.3f, f1 = %.3f" %
(accuracy, precision, recall, f1))
y_predicted_counts = forest.predict(X_test_counts)
accuracy, precision, recall, f1 = get_metrics(Y_test, y_predicted_counts)
print("Test accuracy = %.3f, precision = %.3f, recall = %.3f, f1 = %.3f" %
(accuracy, precision, recall, f1))
# Find best hyperparams
# Number of trees in random forest
n_estimators = [int(x) for x in np.linspace(start=200, stop=2000, num=10)]
# Number of features to consider at every split
max_features = ['auto', 'sqrt']
# Maximum number of levels in tree
max_depth = [int(x) for x in np.linspace(10, 110, num=11)]
max_depth.append(None)
# Minimum number of samples required to split a node
min_samples_split = [2, 5, 10]
# Minimum number of samples required at each leaf node
min_samples_leaf = [1, 2, 4]
# Method of selecting samples for training each tree
bootstrap = [True, False]
# Create the random grid
random_grid = {
'n_estimators': n_estimators,
'max_features': max_features,
'max_depth': max_depth,
'min_samples_split': min_samples_split,
'min_samples_leaf': min_samples_leaf,
'bootstrap': bootstrap
}
# First create the model to tune
rf = RandomForestClassifier()
rf_random = RandomizedSearchCV(estimator=rf,
param_distributions=random_grid,
n_iter=100,
cv=3,
verbose=2,
random_state=42,
n_jobs=-1)
# Fit the random search model
rf_random.fit(X_train_counts, Y_train)
print('Get Best Params')
print(rf_random.best_params_)
print('Saving model')
save_model(rf_random, RANDOM_FOREST_MODEL_PATH)
print('Load model')
trained_model = load_model(RANDOM_FOREST_MODEL_PATH)
y_predicted_counts_train = trained_model.predict(X_train_counts)
accuracy, precision, recall, f1 = get_metrics(Y_train,
y_predicted_counts_train)
print(
"Train accuracy = %.3f, precisionս = %.3f, recall = %.3f, f1 = %.3f" %
(accuracy, precision, recall, f1))
y_predicted_counts = trained_model.predict(X_test_counts)
accuracy, precision, recall, f1 = get_metrics(Y_test, y_predicted_counts)
print("Test accuracy = %.3f, precision = %.3f, recall = %.3f, f1 = %.3f" %
(accuracy, precision, recall, f1))
# ---
# # 检查点
#
# 通过运行上面的训练单元，你的模型已经以`trained_rnn`名字存储，如果你存储了你的notebook， **你可以在之后的任何时间来访问你的代码和结果**. 下述代码可以帮助你重载你的结果!
# In[26]:
"""
DON'T MODIFY ANYTHING IN THIS CELL
"""
import torch
import helper
import problem_unittests as tests
_, vocab_to_int, int_to_vocab, token_dict = helper.load_preprocess()
trained_rnn = helper.load_model('./save/trained_rnn')
# ## 生成电视剧剧本
# 你现在可以生成你的“假”电视剧剧本啦！
#
# ### 生成文字
# 你的神经网络会不断重复生成一个单词，直到生成满足你要求长度的剧本。使用 `generate` 函数来完成上述操作。首先，使用 `prime_id` 来生成word id，之后确定生成文本长度 `predict_len`。同时， topk 采样来引入文字选择的随机性!
# In[27]:
"""
DON'T MODIFY ANYTHING IN THIS CELL THAT IS BELOW THIS LINE
"""
import torch.nn.functional as F
def generate(rnn,
n_layers,
dropout=0.5)
if train_on_gpu:
rnn.cuda()
# training the model
trained_rnn = train_rnn(rnn, batch_size, optimizer, criterion, num_epochs,
show_every_n_batches)
# saving the trained model
helper.save_model('./modl/trained_rnn', trained_rnn)
print('Model Trained and Saved')
_, vocab_to_int, int_to_vocab, token_dict = helper.load_preprocess()
trained_rnn = helper.load_model('./model/trained_rnn')
# run the cell multiple times to get different results!
gen_length = 400 # modify the length to your preference
prime_word = 'george' # name for starting the script
pad_word = helper.SPECIAL_WORDS['PADDING']
generated_script = generate(trained_rnn, vocab_to_int[prime_word + ':'],
int_to_vocab, token_dict,
vocab_to_int[pad_word], gen_length)
print(generated_script)
# save script to a text file
f = open("generated_script_george.txt", "w")
f.write(generated_script)
f.close()
import torch
import helper
import numpy as np
import torch.nn.functional as F
from tv_scripts_rnn import RNN
train_on_gpu = torch.cuda.is_available()
_, vocab_to_int, int_to_vocab, token_dict = helper.load_preprocess()
trained_rnn = helper.load_model('trained_rnn', is_gpu=train_on_gpu)
def generate(rnn,
prime_id,
int_to_vocab,
token_dict,
pad_value,
predict_len=100):
"""
Generate text using the neural network
:param decoder: The PyTorch Module that holds the trained neural network
:param prime_id: The word id to start the first prediction
:param int_to_vocab: Dict of word id keys to word values
:param token_dict: Dict of puncuation tokens keys to puncuation values
:param pad_value: The value used to pad a sequence
:param predict_len: The length of text to generate
:return: The generated text
"""
rnn.eval()
sequence_length = 10
# create a sequence (batch_size=1) with the prime_id
results = {}
for model_name in model_names:
for activation in activations:
base_name = model_name+'_'+activation
print(base_name)
results[base_name] = {}
trial_roc_mean = []
trial_roc_std = []
trial_pr_mean = []
trial_pr_std = []
for trial in range(num_trials):
keras.backend.clear_session()
# load model
model = helper.load_model(model_name, activation=activation)
name = base_name+'_'+str(trial)
print('model: ' + name)
# compile model
helper.compile_model(model)
# setup callbacks
callbacks = helper.get_callbacks(monitor='val_auroc', patience=20,
decay_patience=5, decay_factor=0.2)
# fit model
history = model.fit(x_train, y_train,
epochs=100,
batch_size=100,
shuffle=True,
import sys
import pandas as pd
from tqdm import tqdm
sys.path.append(os.getcwd())
import helper
from Config import UPLOAD_FOLDER, FN_DF_TRANSFORMED
app = Flask(__name__)
app.secret_key = "secret key"
app.config['UPLOAD_FOLDER'] = UPLOAD_FOLDER
app.config['MAX_CONTENT_LENGTH'] = 16 * 1024 * 1024
ALLOWED_EXTENSIONS = set(['pdf', 'png', 'jpg', 'jpeg', 'gif'])
trained_densenet_model = helper.load_model()
def allowed_file(filename):
return '.' in filename and filename.rsplit(
'.', 1)[1].lower() in ALLOWED_EXTENSIONS
@app.route('/')
def upload_form():
return render_template('upload.html')
@app.route('/', methods=['POST'])
def upload_file():
if request.method == 'POST':
from torchvision import utils
from linknet import linknet_model
from helper import load_model, make_mask_overlay, get_ids_from_file_in_list
import time
import cv2
import numpy as np
thrs = 0.45400802
upper = 1
lower = 0
start_time = time.time()
segm_model = linknet_model()
segm_model = load_model(segm_model, model_dir="linknet_10epch.pt")
direc = "./TestImages/"
saved_dir = "./TestResultsImagesLink/"
list_ids = get_ids_from_file_in_list(direc)
for img_id in list_ids:
img = Image.open(direc + img_id)
trf = transforms.Compose([transforms.Resize([512, 512]), transforms.ToTensor()])
img_input = trf(img)
img_input = img_input.unsqueeze(dim=0)
output = segm_model(img_input)
output = torch.sigmoid(output)
inference_time = time.time() - start_time
print("Time to do inference was: {:.0f} seconds and {:.5f} micro seconds ".format(inference_time, inference_time % 60))
output = output.squeeze(0)
num_epochs = adv[0]
num_clean_epochs = adv[1]
prob_clean = adv[2]
tf.reset_default_graph()
# compile neural trainer
name = model_name+'_adv' + str(idx)
print('model: ' + name)
file_path = os.path.join(model_path, name)
# load model parameters
model_layers, optimization, _ = helper.load_model(model_name,
input_shape,
output_shape)
# build neural network class
nnmodel = nn.NeuralNet()
nnmodel.build_layers(model_layers, optimization, supervised=True)
grad_tensor = tf.gradients(nnmodel.mean_loss, nnmodel.placeholders['inputs'])[0]
xx = nnmodel.placeholders['inputs']
yy = nnmodel.placeholders['targets']
is_train = nnmodel.placeholders['is_training']
loss = nnmodel.mean_loss
# nnmodel.inspect_layers()
performance = nn.MonitorPerformance('train', optimization['objective'], verbose)
performance.set_start_time(start_time = time.time())
train_y, dev_y, test_y = datasets_y
hp.saveLogMsg("#Train={}, #Dev={}, #Test={}".format(len(train_x), len(dev_x),
len(test_x)))
# Finding Labels in Dataset
hp.saveLogMsg("\nFinding labels...")
labels = [each_y for sample_y in train_y for each_y in sample_y]
labels = list(set(labels))
labels.remove('O')
hp.saveLogMsg("#Labels={}\n".format(len(labels) + 1))
# Run Model
handler = CRFHandler(labels)
model = None
if cf.MODE == "test" and os.path.exists(cf.MODEL_PATH):
model = hp.load_model()
hp.saveLogMsg("\nLoading best model from {}".format(cf.MODEL_PATH))
else:
model = handler.train(train_x, train_y)
hp.save_model(model)
hp.saveLogMsg("\nSaving best model at {}".format(cf.MODEL_PATH))
assert model is not None
# Eval Model
if cf.TEST_LABELED:
acc_score, clf_report = handler.evaluate(model, dev_x, dev_y)
hp.saveLogMsg('\n[DEV] Accuracy Score: {}'.format(acc_score))
hp.saveLogMsg('\n[DEV] Classification Report: \n{}'.format(clf_report))
else:
handler.predict(model, test_x)
hp.saveLogMsg('\nSaving prediction at {}'.format(cf.PREDICT_PATH))
print(
'number of trainable parameters = ',
numpy.sum(list(param_dict_selector.values())),
numpy.sum(list(param_dict.values())),
numpy.sum(list(param_dict.values())) +
numpy.sum(list(param_dict_selector.values())))
if args.cuda:
torch.cuda.set_device(args.gpu)
selector = selector.cuda()
model = model.cuda()
if args.load_model == 0 or args.load_model == 2:
print('loading selector')
helper.load_model(
selector,
args.output_base_path + args.task + '/' + args.selector_file_name,
'selector', args.cuda)
if args.load_model == 1 or args.load_model == 2:
print('loading classifier')
helper.load_model(
model,
args.output_base_path + args.task + '/' + args.classifier_file_name,
'state_dict', args.cuda)
if args.resume:
if os.path.isfile(args.resume):
print("=> loading checkpoint '{}'".format(args.resume))
checkpoint = helper.load_checkpoint(args.resume)
args.start_epoch = checkpoint['epoch']
best_acc = checkpoint['best_acc']
selector.load_state_dict(checkpoint['selector'])
def evaluate_iobb(args, params, img_name=None, img_disease=None):
"""
Goes through every image in the BBox csv file, and calculates IoBB.
Results are stored in data/iobb.npy together with image name and disease type.
:param img_name: If given, only load one image and plot bbox and heatmap.
:param img_disease: Which disease to plot bbox for, used together with img_name.
"""
pathologies = {
'Atelectasis': 0,
'Cardiomegaly': 1,
'Effusion': 4,
'Infiltrate': 8,
'Mass': 9,
'Nodule': 10,
'Pneumonia': 12,
'Pneumothorax': 13
}
############################################
### Load model checkpoint and get heatmap.
############################################
path_to_load = helper.compute_paths(args, params)['save_path']
net = networks.init_unified_net(args.model, params)
net, _ = helper.load_model(path_to_load,
args.epoch,
net,
optimizer=None,
resume_train=False)
# Get bbox_data from csv file.
f = open('../data/BBox_List_2017.csv', 'rt')
reader = csv.reader(f)
rows = list(reader)[1:] # ignoring the first row because it is the titles
# A list of tuples (img_name, disease_index, iobb, num_bboxes)
results = []
for i, img_data in enumerate(rows):
# Make sure image exists.
file_path = f'../data/extracted/images/{img_data[0]}'
if not os.path.isfile(file_path):
continue
# If only loading one image, check if this row contains the img and correct disease, otherwise continue.
if img_name is not None:
if img_data[0] != img_name:
continue
if img_disease is not None and pathologies[
img_data[1]] != img_disease:
continue
############################################
### Load image and turn into tensor.
############################################
xray_img = Image.open(file_path)
rgb = Image.new('RGB', xray_img.size)
rgb.paste(xray_img)
preprocess = transforms.Compose([
transforms.Resize(256),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
])
img_tensor = preprocess(rgb)
img_tensor.unsqueeze_(0)
# Get heatmap for correct disease and turn into numpy array.
heatmap = net.forward(img_tensor, return_cam=True)
disease = pathologies[img_data[1]]
heatmap = heatmap[0, disease].numpy()
ground_truth = BBox(float(img_data[2]), float(img_data[3]),
float(img_data[4]), float(img_data[5]))
# Save results if evaluating all images, not just plotting one.
if img_name is None:
iobb, num = evaluate_single_bbox(ground_truth, heatmap, iobb=True)
results.append((img_data[0], disease, iobb, num))
print(f'#{i}, n:{num}, {iobb}')
else:
iobb, _ = evaluate_single_bbox(ground_truth,
heatmap,
iobb=True,
xray_img=xray_img)
print(f'iobb: {iobb}')
break
if img_name is None:
# Order results by IoBB value.
results = sorted(results, key=lambda x: x[2], reverse=True)
results = np.array(results)
# Save as numpy array.
np.save(f'../data/iou_blob_{args.model}.npy', results)
# Save as txt
with open(f'../data/iou_blob_{args.model}.txt', 'w') as txt:
for i in range(results.shape[0]):
txt.write(
f'{float(results[i][2])}, {int(results[i][1])}, {int(results[i][3])}, {results[i][0]}\n'
)
f.close() # Close csv file.
f.write('%s\t%s\t%s\n'%('model', 'ave roc', 'ave pr'))
results = {}
for model_name in model_names:
results[model_name] = {}
for activation in activations:
trial_roc_mean = []
trial_roc_std = []
trial_pr_mean = []
trial_pr_std = []
for trial in range(num_trials):
keras.backend.clear_session()
# load model
model = helper.load_model(model_name,
activation=activation,
input_shape=200)
name = model_name+'_'+activation+'_'+str(trial)
print('model: ' + name)
# compile model
helper.compile_model(model)
# setup callbacks
callbacks = helper.get_callbacks(monitor='val_aupr', patience=20,
decay_patience=5, decay_factor=0.2)
# fit model
history = model.fit(x_train, y_train,
epochs=100,
batch_size=100,
def evaluate_NavNet(parIL, parMapNet, mapNet, ego_encoder, test_iter, test_ids, test_data, action_list):
print("\nRunning validation on NavNet!")
with torch.no_grad():
policy_net = hl.load_model(model_dir=parIL.model_dir, model_name="ILNet", test_iter=test_iter)
acc, epi_length, path_ratio = 0, 0, 0
episode_results, episode_count = {}, 0 # store predictions
for i in test_ids:
test_ex = test_data[i]
# Get all info for the starting position
mapNet_input_start = prepare_mapNet_input(ex=test_ex)
target_lbl = test_ex["target_lbl"]
im_obsv = test_ex['image_obsv'].cuda()
dets_obsv = test_ex['dets_obsv'].cuda()
tvec = torch.zeros(1, parIL.nTargets).float().cuda()
tvec[0,target_lbl] = 1
# We need to keep other info to allow us to do the steps later
image_name, scene, scale = [], [], []
image_name.append(test_ex['image_name'])
scene.append(test_ex['scene'])
scale.append(test_ex['scale'])
shortest_path_length = test_ex['path_length']
if parIL.use_p_gt:
# get the ground-truth pose, which is the relative pose with respect to the first image
info, annotations, _ = dh.load_scene_info(parIL.avd_root, scene[0])
im_names_all = info['image_name'] # info 0 # list of image names in the scene
im_names_all = np.hstack(im_names_all) # flatten the array
start_abs_pose = dh.get_image_poses(info, im_names_all, image_name, scale[0]) # init pose of the episode # 1 x 3
# Get state from mapNet
p_, map_ = mapNet.forward_single_step(local_info=mapNet_input_start, t=0,
input_flags=parMapNet.input_flags, update_type=parMapNet.update_type)
collision_ = torch.tensor([0], dtype=torch.float32).cuda() # collision indicator is 0
if parIL.use_ego_obsv:
enc_in = torch.cat((im_obsv, dets_obsv), 0).unsqueeze(0)
ego_obsv_feat = ego_encoder(enc_in) # 1 x 512 x 1 x 1
state = (map_, p_, tvec, collision_, ego_obsv_feat)
else:
state = (map_, p_, tvec, collision_)
current_im = image_name[0]
done=0
image_seq, action_seq = [], []
image_seq.append(current_im)
policy_net.hidden = policy_net.init_hidden(batch_size=1, state_items=len(state)-1)
for t in range(1, parIL.max_steps+1):
pred_costs = policy_net(state, parIL.use_ego_obsv) # apply policy for single step
pred_costs = pred_costs.view(-1).cpu().numpy()
# choose the action with a certain prob
pred_probs = softmax(-pred_costs)
pred_label = np.random.choice(len(action_list), 1, p=pred_probs)[0]
pred_action = action_list[pred_label]
# get the next image, check collision and goal
next_im = test_data.scene_annotations[scene[0]][current_im][pred_action]
if next_im=='':
image_seq.append(current_im)
else:
image_seq.append(next_im)
action_seq.append(pred_action)
print(t, current_im, pred_action, next_im)
if not(next_im==''): # not collision case
collision = 0
# check for goal
path_dist = len(nx.shortest_path(test_data.graphs_dict[target_lbl][scene[0]], next_im, "goal")) - 2
if path_dist <= parIL.steps_from_goal: # GOAL!
acc += 1
epi_length += t
path_ratio += t/float(shortest_path_length) # ratio of estimated path towards shortest path
done=1
break
# get next state from mapNet
batch_next, obsv_batch_next = test_data.get_step_data(next_ims=[next_im], scenes=scene, scales=scale)
if parIL.use_p_gt:
next_im_abs_pose = dh.get_image_poses(info, im_names_all, [next_im], scale[0])
abs_poses = np.concatenate((start_abs_pose, next_im_abs_pose), axis=0)
rel_poses = dh.relative_poses(poses=abs_poses)
next_im_rel_pose = np.expand_dims(rel_poses[1,:], axis=0)
p_gt = dh.build_p_gt(parMapNet, pose_gt_batch=np.expand_dims(next_im_rel_pose, axis=1)).squeeze(1)
p_next, map_next = mapNet.forward_single_step(local_info=batch_next, t=t, input_flags=parMapNet.input_flags,
map_previous=state[0], p_given=p_gt, update_type=parMapNet.update_type)
else:
p_next, map_next = mapNet.forward_single_step(local_info=batch_next, t=t,
input_flags=parMapNet.input_flags, map_previous=state[0], update_type=parMapNet.update_type)
if parIL.use_ego_obsv:
enc_in = torch.cat(obsv_batch_next, 1)
ego_obsv_feat = ego_encoder(enc_in) # b x 512 x 1 x 1
state = (map_next, p_next, tvec, torch.tensor([collision], dtype=torch.float32).cuda(), ego_obsv_feat)
else:
state = (map_next, p_next, tvec, torch.tensor([collision], dtype=torch.float32).cuda())
current_im = next_im
else: # collision case
collision = 1
if parIL.stop_on_collision:
break
if parIL.use_ego_obsv:
state = (state[0], state[1], state[2], torch.tensor([collision], dtype=torch.float32).cuda(), state[4])
else:
state = (state[0], state[1], state[2], torch.tensor([collision], dtype=torch.float32).cuda())
episode_results[episode_count] = (image_seq, action_seq, parIL.lbl_to_cat[target_lbl], done)
episode_count+=1
# store the episodes
episode_results_path = parIL.model_dir+'episode_results_eval_'+str(test_iter)+'.pkl'
with open(episode_results_path, 'wb') as f:
pickle.dump(episode_results, f)
success_rate = acc / float(len(test_ids))
if acc > 0:
mean_epi_length = epi_length / float(acc)
avg_path_length_ratio = path_ratio / float(acc)
else:
mean_epi_length = 0
avg_path_length_ratio = 0
print("Test iter:", test_iter, "Success rate:", success_rate)
print("Mean epi length:", mean_epi_length, "Avg path length ratio:", avg_path_length_ratio)
import json
import pandas
import helper
model_r = helper.load_model("data/model_target_r")
model_g = helper.load_model("data/model_target_g")
model_b = helper.load_model("data/model_target_b")
def do(features):
"""
Do the inference with a loaded model on specified features.
Parameters
----------
features : list[dict[str, str]]
Feature to process.
It a list of dict containing feature names as key and feature abstracted values as values.
Returns
-------
A list of dict.
All dict must have all three R, G, and B target. If one is missing, it will be considered as `null`.
Notes
-----
The order must be kept for a response to be considered valid by the datacrunch arena.
If more or less lines are returned, everything will be considered as invalid.
