def parse_and_preprocess_src(data_source, corpus_destination, preprocess=True):
if re.search("bundestag", data_source.lower()):
name = "bundestag"
raw_corpus = DataHandler.get_bundestag_speeches(directory=data_source)
elif re.search("sustainability", data_source.lower()):
name = "sustainability"
raw_corpus = DataHandler.get_sustainability_data(path=data_source)
elif re.search("unv1.0-tei", data_source.lower()):
name = "united_nations"
raw_corpus = DataHandler.get_un_texts(directory=data_source)
elif re.search("state_of_the_union", data_source.lower()):
name = "state_of_the_union"
raw_corpus = DataHandler.get_state_of_the_union(directory=data_source)
else:
name = "abstracts"
raw_corpus = DataHandler.get_abstracts(path=data_source)
language = raw_corpus[0].language
print('loaded', len(raw_corpus), 'documents')
if preprocess:
Preprocessor.preprocess(raw_corpus, language=language)
print('preprocessed', len(raw_corpus), 'documents')
corpus = Corpus(source=raw_corpus, language=language, name=name)
print('parsed', len(corpus.get_documents(as_list=True)),
'documents to a Corpus')
corpus.save_corpus(corpus_destination)
def load_portfolio(self) -> dict:
"""
Loads portfolio into the class checking first if the portfolio exists.
:return: dict
"""
if DataHandler.check_portfolio_exists():
return DataHandler.load_portfolio()
else:
return dict({})
def __init__(self, cfg):
self.cfg = cfg
self.cat_features = self.cfg.data.features.cat_features
self.oof = None
self.raw_preds = None
self.weights = []
self.models = []
self.scores = []
self.feature_importance_df = pd.DataFrame(
columns=['feature', 'importance'])
self.dh = DataHandler()
def __getitem__(self, index):
if self.st == 'train' or self.st == 'FTrain':
# index_sec = int(self.rng.rand()*self.len)
# print(os.getpid(), index_sec) # same in diff gpu
# label = np.zeros(self.person)
# label_sec = np.zeros(self.person)
# label[self.y[index]] = 1
# label_sec[self.y[index_sec]] = 1
label = self.y[index]
image = np.array(cv2.imread(self.name[index]), dtype=float)
# image_sec = np.array(cv2.imread(self.name[index_sec]), dtype=float).copy()
# image, label = aug.run2(image, label, image_sec, label_sec)
image = Aug.run(self.rng, image)
image = torch.from_numpy(image).float()
# label = torch.from_numpy(label).float()
return image, label
elif self.st == 'test':
size = (MinS + MaxS) // 2
idx = (size - W) // 2
image = np.array(self.dataset[index], dtype=float).copy()
image = cv2.resize(image, (size, size))
image = image[idx:idx + H, idx:idx + W, :]
image = np.transpose(image, [2, 0, 1])
image_flip = image.copy()[:, :, ::-1]
image = torch.from_numpy((image - 127.5) / 128).float()
image_flip = torch.from_numpy((image_flip - 127.5) / 128).float()
label = np.zeros(self.len)
label[self.y[index]] = 1
label = torch.from_numpy(label).float()
return image, image_flip, label, self.name[index]
else:
exit(-1)
def save_features(df, data_type='train'):
save_path = Path('../configs/feature/all.yml')
dh = DataHandler()
if not save_path.exists():
save_path.touch()
feature_dict = {'features': []}
else:
feature_dict = dh.load(save_path)
new_feature = sorted(set(feature_dict['features'] + df.columns.tolist()))
feature_dict['features'] = new_feature
dh.save(save_path, feature_dict)
for col in df.columns:
df[[col]].reset_index(drop=True).to_feather(f'../features/{col}_{data_type}.feather')
def on_ok(self):
if len(current_portoflio.portfolio) == 6:
self.status_msg.value = "Max 6 stocks allowed. Please purchase premium version!"
else:
last_price = StockDataReader.last_price(
StockDataReader.get_data(self.symbol.value))
validation = DataHandler.validate_entry(self.symbol.value,
self.date.value,
self.amount.value,
self.price.value,
last_price)
if validation != True:
self.status_msg.value = validation
self.display()
else:
current_portoflio.add_stock(self.symbol.value,
float(self.price.value),
float(self.amount.value),
self.date.value)
self.status_msg.value = f"Grabbing data for {self.symbol.value}"
to_main = self.parentApp.getForm('MAIN')
to_main.load_portfolio()
to_main.load_performance()
to_main.top_message.value = "Portfolio:"
self.display()
sleep(1.5)
self.parentApp.switchForm("MAIN")
def __init__(self, params):
# Build data handler
self.data_handler = DataHandler()
self.data_handler.loadData(params['input'])
params['inp_dim'] = self.data_handler.getDataShape()[1]
logging.info("=" * 41)
# Build model
self.model = SPINEModel(params)
self.dtype = torch.FloatTensor
use_cuda = torch.cuda.is_available()
if use_cuda:
self.model.cuda()
self.dtype = torch.cuda.FloatTensor
self.optimizer = torch.optim.SGD(self.model.parameters(), lr=0.1)
logging.info("=" * 41)
def create(self):
# ====================== Init portfolio ========================
self.welcome_msg = self.add(npyscreen.FixedText, value=welcome, rely=2)
self.welcome_msg2 = self.add(npyscreen.FixedText,
value=welcome_line2,
rely=3)
self.top_message = self.add(npyscreen.FixedText,
value="",
relx=8,
rely=4)
# ====================== End init portfolio ====================
# ====================== Menu section ==========================
self.menu = self.new_menu(name="Main Menu", shortcut="m")
self.menu.addItem("1. Add stock to portfolio", self.change_form_add,
"1")
self.menu.addItem("2. Update Portfolio", self.change_form_update, "2")
self.menu.addItem("3. Save performance to PDF",
self.change_form_save_pdf, "3")
self.menu.addItem("4. Delete Stock form portfolio",
self.change_form_delete, "4")
# ====================== End menu section ======================
self.add(npyscreen.FixedText, value="Performance:", relx=8, rely=26)
self.main_ii = self.add(
npyscreen.FixedText,
value=f"Initial Investment: ${self.initial_investment}",
relx=8,
rely=27)
self.main_rd = self.add(
npyscreen.FixedText,
value=f"Returned Amount: ${self.return_dollar}",
relx=8,
rely=28)
self.main_pr = self.add(npyscreen.FixedText,
value=f"Returned Percent: {self.pct_return}%",
relx=8,
rely=29)
self.exit_button = self.add(npyscreen.ButtonPress,
name="Exit",
rely=29,
relx=106,
when_pressed_function=self.exit_press)
if not DataHandler.check_portfolio_exists():
self.top_message.value = no_portfolio_msg
else:
self.top_message.value = "Portfolio:"
self.load_portfolio()
self.load_performance()
def __init__(self, params):
# Build data handler
self.data_handler = DataHandler()
self.data_handler.loadData(params['input'])
params['inp_dim'] = self.data_handler.getDataShape()[1]
logging.info("=" * 41)
# Build model
self.model = SPINEModel(params)
self.dtype = torch.FloatTensor
use_cuda = torch.cuda.is_available()
if use_cuda:
self.model.cuda()
self.dtype = torch.cuda.FloatTensor
# Select optimizer
optim_selected = params['optim']
LR = 0.1
if optim_selected == 'sgd':
self.optimizer = torch.optim.SGD(self.model.parameters(), lr=LR)
elif optim_selected == 'adam':
self.optimizer = torch.optim.Adam(self.model.parameters(), lr=LR)
logging.info("=" * 41)
def __init__(self, params):
# build data handler
self.data_handler = DataHandler()
self.data_handler.loadData(params['input'])
params['inp_dim'] = self.data_handler.getDataShape()[1]
print("=" * 41)
# build model
self.model = SPINEModel(params)
self.dtype = torch.FloatTensor
# check if GPU is available
use_cuda = torch.cuda.is_available()
if use_cuda:
self.model.cuda()
self.dtype = torch.cuda.FloatTensor
# set optimizer
self.optimizer = torch.optim.SGD(self.model.parameters(), lr=0.1)
print("=" * 41)
def __init__(self, net, state, loop_sleep, pause_after_keys, heartbeat_required):
CodependentThread.__init__(self, heartbeat_required)
self.daemon = True
self.net = net
self.input_dims = self.net.blobs['data'].data.shape[2:4] # e.g. (227,227)
self.state = state
self.frames_processed_fwd = 0
self.frames_processed_back = 0
self.loop_sleep = loop_sleep
self.pause_after_keys = pause_after_keys
self.debug_level = 0
self.descriptor = None
self.descriptor_layer_1 = 'conv5'
self.descriptor_layer_2 = 'conv4'
self.descriptor_layers = ['conv5','conv4']
self.net_input_image = None
self.descriptor_handler = DescriptorHandler(self.state.settings.ros_dir + '/models/memory/', self.descriptor_layers)
self.data_handler = DataHandler(self.state.settings.ros_dir)
self.available_layer = ['conv1', 'pool1', 'norm1', 'conv2', 'pool2', 'norm2', 'conv3', 'conv4', 'conv5', 'pool5', 'fc6', 'fc7', 'fc8', 'prob']
#['conv1', 'conv2', 'conv3', 'conv4', 'conv5', 'fc6', 'fc7', 'fc8', 'prob']
# print "layers ", list(self.net._layer_names)
s = rospy.Service('get_cnn_state', GetState, self.handle_get_cnn_state)
def main():
images_dir = 'images'
output_dir = 'output'
debug = True
for image_idx, (reference_image, inspection_image) in enumerate(
DataHandler(images_dir).get()):
detector = DefectDetector(reference_image,
inspection_image,
image_idx,
debug=debug,
output_dir=output_dir)
output_mask = detector.run()
plt.imsave(os.path.join(output_dir,
'output-image-idx-{}.png'.format(image_idx)),
output_mask,
cmap=cm.gray)
import sys
import pandas as pd
from easydict import EasyDict as edict
sys.path.append('../src')
from utils import DataHandler
from feature_utils import save_features
dh = DataHandler()
def get_features(df):
features_df = pd.DataFrame()
features_df['one_hot_sex'] = df['sex'].map({'male': 0, 'female': 1})
age_df = pd.get_dummies(df['age_approx'])
for col in age_df.columns:
features_df[f'one_hot_age_{col}'] = age_df[col]
site_df = pd.get_dummies(df['anatom_site_general_challenge'])
for col in site_df.columns:
features_df[f'one_hot_site_{col.replace("/", "-")}'] = site_df[col]
return features_df
def main():
train_df = dh.load('../data/input/train_concated.csv')
test_df = dh.load('../data/input/test.csv')
from utils import (DataHandler, Notificator, Timer, seed_everything, Git,
Kaggle, reduce_mem_usage)
warnings.filterwarnings('ignore')
# ===============
# Settings
# ===============
parser = argparse.ArgumentParser()
parser.add_argument('--common', default='../configs/common/compe.yml')
parser.add_argument('--notify', default='../configs/common/notify.yml')
parser.add_argument('-m', '--model')
parser.add_argument('-c', '--comment')
options = parser.parse_args()
dh = DataHandler()
cfg = dh.load(options.common)
cfg.update(dh.load(f'../configs/exp/{options.model}.yml'))
notify_params = dh.load(options.notify)
features_params = dh.load(f'../configs/feature/{cfg.data.features.name}.yml')
features = features_params.features
comment = options.comment
model_name = options.model
now = datetime.datetime.now()
if cfg.model.task_type != 'optuna':
run_name = f'{model_name}_{now:%Y%m%d%H%M%S}'
else:
run_name = f'{model_name}_optuna_{now:%Y%m%d%H%M%S}'
def build_data_set(self, root_dir: str):
entries = self.parse_folder(root_dir)
data_handler = DataHandler()
icd10_ontology = data_handler.read_icd10_ontology()
group_codes = set(data_handler.read_reduced_icd10_ontology().keys())
de_entries = dict()
en_entries = dict()
for entry in entries:
if "de_" + entry[0] in de_entries:
continue
icd10_codes = entry[3]
valid_group_codes = set()
main_chapter = ""
for code in icd10_codes:
if code in icd10_chapter_mappings:
code = icd10_chapter_mappings[code]
if not code in icd10_ontology and "." in code:
code = code[:code.index(".")]
if not code in icd10_ontology:
self.logger.error(
f"Can't find code {code} in ICD10 ontology")
continue
path_components = icd10_ontology[code].split("#")
if main_chapter == "":
main_chapter = path_components[0]
for path_comp in path_components:
if path_comp in group_codes:
valid_group_codes.add(path_comp)
valid_group_codes = "|".join(valid_group_codes)
de_entries["de_" + entry[0]] = {
"text": entry[1],
"language": "de",
"all_labels": valid_group_codes,
"main_chapter": main_chapter
}
en_entries["en_" + entry[0]] = {
"text": entry[2],
"language": "en",
"all_labels": valid_group_codes,
"main_chapter": main_chapter
}
de_df = DataFrame.from_dict(de_entries, orient="index")
en_df = DataFrame.from_dict(en_entries, orient="index")
de_train, de_dev = train_test_split(de_df,
train_size=0.8,
stratify=de_df["main_chapter"])
de_train["data_set"] = "train"
de_dev["data_set"] = "dev"
de_df = de_train.append(de_dev)
en_train, en_dev = train_test_split(en_df,
train_size=0.8,
stratify=en_df["main_chapter"])
en_train["data_set"] = "train"
en_dev["data_set"] = "dev"
en_df = en_train.append(en_dev)
full_df = de_df.append(en_df)
full_df = full_df.drop_duplicates()
full_df = full_df[full_df["text"].notna()]
#de_df.to_csv("drks_de.tsv", sep="\t", columns=["data_set", "main_chapter", "all_labels", "text"], index_label="id")
#en_df.to_csv("drks_en.tsv", sep="\t", columns=["data_set", "main_chapter", "all_labels", "text"], index_label="id")
output_dir = "data/drks/prepared"
os.makedirs(output_dir, exist_ok=True)
output_file = os.path.join(output_dir, "drks_full.tsv")
full_df.to_csv(output_file,
sep="\t",
columns=[
"language", "data_set", "main_chapter",
"all_labels", "text"
],
index_label="id")
class Solver:
def __init__(self, params):
# build data handler
self.data_handler = DataHandler()
self.data_handler.loadData(params['input'])
params['inp_dim'] = self.data_handler.getDataShape()[1]
print("=" * 41)
# build model
self.model = SPINEModel(params)
self.dtype = torch.FloatTensor
# check if GPU is available
use_cuda = torch.cuda.is_available()
if use_cuda:
self.model.cuda()
self.dtype = torch.cuda.FloatTensor
# set optimizer
self.optimizer = torch.optim.SGD(self.model.parameters(), lr=0.1)
print("=" * 41)
def train(self, params):
num_epochs, batch_size = params['num_epochs'], params['batch_size'],
optimizer = self.optimizer
dtype = self.dtype
# train for each epoch
for iteration in range(num_epochs):
self.data_handler.shuffleTrain()
num_batches = self.data_handler.getNumberOfBatches(batch_size)
epoch_losses = np.zeros(4) # rl, asl, psl, total
# for each batch
for batch_idx in range(num_batches):
optimizer.zero_grad()
batch_x, batch_y = self.data_handler.getBatch(
batch_idx, batch_size, params['noise_level'],
params['denoising'])
# transform batches into tensors
batch_x = Variable(torch.from_numpy(batch_x),
requires_grad=False).type(dtype)
batch_y = Variable(torch.from_numpy(batch_y),
requires_grad=False).type(dtype)
# calculate losses
out, h, loss, loss_terms = self.model(batch_x, batch_y)
reconstruction_loss, psl_loss, asl_loss = loss_terms
# update loss and optimizer
loss.backward()
optimizer.step()
# assign loss
epoch_losses[0] += reconstruction_loss.data
epoch_losses[1] += asl_loss.data
epoch_losses[2] += psl_loss.data
epoch_losses[3] += loss.data
print("epoch %r: Reconstruction Loss = %.4f, ASL = %.4f, "\
"PSL = %.4f, and total = %.4f"
%(iteration+1, epoch_losses[0], epoch_losses[1], epoch_losses[2], epoch_losses[3]) )
def getSpineEmbeddings(self, batch_size, params):
ret = []
self.data_handler.resetDataOrder()
num_batches = self.data_handler.getNumberOfBatches(batch_size)
# for each batch
for batch_idx in range(num_batches):
batch_x, batch_y = self.data_handler.getBatch(
batch_idx, batch_size, params['noise_level'],
params['denoising'])
# transform batches into tensors
batch_x = Variable(torch.from_numpy(batch_x),
requires_grad=False).type(self.dtype)
batch_y = Variable(torch.from_numpy(batch_y),
requires_grad=False).type(self.dtype)
_, h, _, _ = self.model(batch_x, batch_y)
# append to embeddings
ret.extend(h.cpu().data.numpy())
return np.array(ret)
def getWordsList(self):
return self.data_handler.getWordsList()
class Solver:
def __init__(self, params):
# Build data handler
self.data_handler = DataHandler()
self.data_handler.loadData(params['input'])
params['inp_dim'] = self.data_handler.getDataShape()[1]
logging.info("=" * 41)
# Build model
self.model = SPINEModel(params)
self.dtype = torch.FloatTensor
use_cuda = torch.cuda.is_available()
if use_cuda:
self.model.cuda()
self.dtype = torch.cuda.FloatTensor
# Select optimizer
optim_selected = params['optim']
LR = 0.1
if optim_selected == 'sgd':
self.optimizer = torch.optim.SGD(self.model.parameters(), lr=LR)
elif optim_selected == 'adam':
self.optimizer = torch.optim.Adam(self.model.parameters(), lr=LR)
logging.info("=" * 41)
def train(self, params):
num_epochs, batch_size = params['num_epochs'], params['batch_size'],
optimizer = self.optimizer
dtype = self.dtype
STEP_DENOM = 5
scheduler = StepLR(optimizer,
step_size=math.ceil(num_epochs / STEP_DENOM),
gamma=0.3)
for iteration in range(num_epochs):
# lr adjusting
scheduler.step()
# start epoch
self.data_handler.shuffleTrain()
num_batches = self.data_handler.getNumberOfBatches(batch_size)
epoch_losses = np.zeros(4) # rl, asl, psl, total
for batch_idx in range(num_batches):
optimizer.zero_grad()
batch_x, batch_y = self.data_handler.getBatch(
batch_idx, batch_size, params['noise_level'],
params['denoising'])
batch_x = Variable(torch.from_numpy(batch_x),
requires_grad=False).type(dtype)
batch_y = Variable(torch.from_numpy(batch_y),
requires_grad=False).type(dtype)
out, h, loss, loss_terms = self.model(batch_x, batch_y)
reconstruction_loss, psl_loss, asl_loss = loss_terms
loss.backward()
epoch_losses[0] += reconstruction_loss.data.item()
epoch_losses[1] += asl_loss.data.item()
epoch_losses[2] += psl_loss.data.item()
epoch_losses[3] += loss.data.item()
optimizer.step()
print("After epoch %r, Reconstruction Loss = %.4f, ASL = %.4f, "
"PSL = %.4f, and total = %.4f" %
(iteration + 1, epoch_losses[0], epoch_losses[1],
epoch_losses[2], epoch_losses[3]))
# logging.info("After epoch %r, Sparsity = %.1f"
# %(iteration+1, utils.compute_sparsity(h.cpu().data.numpy())))
# break
# break
def getSpineEmbeddings(self, batch_size, params):
ret = []
self.data_handler.resetDataOrder()
num_batches = self.data_handler.getNumberOfBatches(batch_size)
for batch_idx in range(num_batches):
batch_x, batch_y = self.data_handler.getBatch(
batch_idx, batch_size, params['noise_level'],
params['denoising'])
batch_x = Variable(torch.from_numpy(batch_x),
requires_grad=False).type(self.dtype)
batch_y = Variable(torch.from_numpy(batch_y),
requires_grad=False).type(self.dtype)
_, h, _, _ = self.model(batch_x, batch_y)
ret.extend(h.cpu().data.numpy())
return np.array(ret)
def getWordsList(self):
return self.data_handler.getWordsList()
def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument(
"--data_set",
choices=DataHandler.ALL_DATA_SET_IDS,
required=True,
help=
"The input data dir. Should contain the .tsv files (or other data files) for the task."
)
parser.add_argument(
"--bert_model",
default=None,
type=str,
required=True,
help=
"Bert pre-trained model selected in the list: bert-base-multilingual-uncased, "
"bert-base-multilingual-cased")
parser.add_argument(
"--output_dir",
default=None,
type=str,
required=True,
help=
"The output directory where the model predictions and checkpoints will be written."
)
parser.add_argument(
"--additional_data_set",
choices=DataHandler.ALL_DATA_SET_IDS,
required=False,
help="Additional data set to extend the basic training data set. "
"Only training data will be used! No additional evaluation data!")
## Other parameters
parser.add_argument(
"--cache_dir",
default="_cache",
type=str,
help=
"Where do you want to store the pre-trained models downloaded from s3")
parser.add_argument(
"--max_seq_length",
default=300,
type=int,
help=
"The maximum total input sequence length after WordPiece tokenization. \n"
"Sequences longer than this will be truncated, and sequences shorter \n"
"than this will be padded.")
parser.add_argument("--do_train",
action='store_true',
help="Whether to run training.")
parser.add_argument("--do_eval",
action='store_true',
help="Whether to run eval on the dev set.")
parser.add_argument("--do_test",
action='store_true',
help="Whether to run eval on the test set.")
parser.add_argument(
"--do_lower_case",
action='store_true',
help="Set this flag if you are using an uncased model.")
parser.add_argument("--train_batch_size",
default=32,
type=int,
help="Total batch size for training.")
parser.add_argument("--eval_batch_size",
default=8,
type=int,
help="Total batch size for eval.")
parser.add_argument("--learning_rate",
default=5e-5,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--num_train_epochs",
default=3.0,
type=float,
help="Total number of training epochs to perform.")
parser.add_argument(
"--warmup_proportion",
default=0.1,
type=float,
help=
"Proportion of training to perform linear learning rate warmup for. "
"E.g., 0.1 = 10%% of training.")
parser.add_argument("--no_cuda",
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
parser.add_argument(
'--gradient_accumulation_steps',
type=int,
default=1,
help=
"Number of updates steps to accumulate before performing a backward/update pass."
)
parser.add_argument(
'--fp16',
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument(
'--loss_scale',
type=float,
default=0,
help=
"Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
"0 (default value): dynamic loss scaling.\n"
"Positive power of 2: static loss scaling value.\n")
parser.add_argument('--server_ip',
type=str,
default='',
help="Can be used for distant debugging.")
parser.add_argument('--server_port',
type=str,
default='',
help="Can be used for distant debugging.")
args = parser.parse_args()
if args.server_ip and args.server_port:
# Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
import ptvsd
print("Waiting for debugger attach")
ptvsd.enable_attach(address=(args.server_ip, args.server_port),
redirect_output=True)
ptvsd.wait_for_attach()
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available()
and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(args.local_rank != -1), args.fp16))
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
if not args.do_train and not args.do_eval:
raise ValueError(
"At least one of `do_train` or `do_eval` must be True.")
if os.path.exists(args.output_dir) and os.listdir(
args.output_dir) and args.do_train:
raise ValueError(
"Output directory ({}) already exists and is not empty.".format(
args.output_dir))
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
tokenizer = BertTokenizer.from_pretrained(args.bert_model,
do_lower_case=args.do_lower_case)
data_handler = DataHandler()
train_data, dev_data = data_handler.get_data_set_by_id(args.data_set)
if args.additional_data_set is not None:
logger.info(
f"Extending training data with instances from {args.additional_data_set}"
)
add_train_data, add_dev_data = data_handler.get_data_set_by_id(
args.additional_data_set)
train_data = train_data.append(add_train_data)
train_data = train_data.append(add_dev_data)
logger.info(
f"Data set contains {len(train_data)} training and {len(dev_data)} development instances"
)
test_data = None
if args.do_test:
test_data = data_handler.get_test_data()
processor = DataProcessor(train_data, dev_data, test_data)
label_list = processor.get_labels()
logger.info(f"Labels: {str(label_list)}")
label_encoder = LabelEncoder()
label_encoder.fit(label_list)
num_labels = len(label_encoder.classes_)
logger.info(f"Num labels: {num_labels}")
train_examples = None
num_train_optimization_steps = None
if args.do_train:
train_examples = processor.get_train_instances()
num_train_optimization_steps = int(
len(train_examples) / args.train_batch_size /
args.gradient_accumulation_steps) * args.num_train_epochs
if args.local_rank != -1:
num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size(
)
# Prepare model
cache_dir = args.cache_dir if args.cache_dir else os.path.join(
PYTORCH_PRETRAINED_BERT_CACHE, 'distributed_{}'.format(
args.local_rank))
model = BertForMultiLabelSequenceClassification.from_pretrained(
args.bert_model, cache_dir=cache_dir, num_labels=num_labels)
if args.fp16:
model.half()
model.to(device)
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError \
("Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")
model = DDP(model)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
# Prepare optimizer
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params':
[p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
if args.fp16:
try:
from apex.optimizers import FP16_Optimizer
from apex.optimizers import FusedAdam
except ImportError:
raise ImportError \
("Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
bias_correction=False,
max_grad_norm=1.0)
if args.loss_scale == 0:
optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
else:
optimizer = FP16_Optimizer(optimizer,
static_loss_scale=args.loss_scale)
else:
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
global_step = 0
nb_tr_steps = 0
global_batch_no = 0
tr_loss = 0
if args.do_train:
train_features = convert_examples_to_features(train_examples,
label_encoder,
args.max_seq_length,
tokenizer)
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_examples))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_optimization_steps)
all_input_ids = torch.tensor([f.input_ids for f in train_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in train_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in train_features],
dtype=torch.long)
all_label_ids = torch.tensor([f.label_ids for f in train_features],
dtype=torch.float)
train_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_label_ids)
if args.local_rank == -1:
train_sampler = RandomSampler(train_data)
else:
train_sampler = DistributedSampler(train_data)
train_dataloader = DataLoader(train_data,
sampler=train_sampler,
batch_size=args.train_batch_size)
max_f1 = 0.0
for epoch in trange(int(args.num_train_epochs), desc="Epoch"):
model.train()
tr_loss = 0
nb_tr_examples, nb_tr_steps = 0, 0
for step, batch in enumerate(
tqdm(train_dataloader, desc="Iteration")):
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, segment_ids, label_ids = batch
loss = model(input_ids, segment_ids, input_mask, label_ids)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
optimizer.backward(loss)
else:
loss.backward()
loss_value = loss.item()
tr_loss += loss_value
nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
global_batch_no += 1
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16:
# modify learning rate with special warm up BERT uses
# if args.fp16 is False, BertAdam is used that handles this automatically
lr_this_step = args.learning_rate * warmup_linear(
global_step / num_train_optimization_steps,
args.warmup_proportion)
for param_group in optimizer.param_groups:
param_group['lr'] = lr_this_step
optimizer.step()
optimizer.zero_grad()
global_step += 1
if (epoch + 1) % 2 == 0:
eval_examples = processor.get_dev_instances()
eval_features = convert_examples_to_features(
eval_examples, label_encoder, args.max_seq_length,
tokenizer)
logger.info("***** Running evaluation *****")
logger.info(" Num examples = %d", len(eval_examples))
logger.info(" Batch size = %d", args.eval_batch_size)
all_input_ids = torch.tensor(
[f.input_ids for f in eval_features], dtype=torch.long)
all_input_mask = torch.tensor(
[f.input_mask for f in eval_features], dtype=torch.long)
all_segment_ids = torch.tensor(
[f.segment_ids for f in eval_features], dtype=torch.long)
all_label_ids = torch.tensor(
[f.label_ids for f in eval_features], dtype=torch.float)
eval_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_label_ids)
# Run prediction for full data
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data,
sampler=eval_sampler,
batch_size=args.eval_batch_size)
model.eval()
eval_loss, eval_accuracy = 0, 0
nb_eval_steps, nb_eval_examples = 0, 0
y_dev = None
y_dev_pred = None
y_dev_sigmoid = None
for input_ids, input_mask, segment_ids, label_ids in tqdm(
eval_dataloader, desc="Evaluating"):
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
label_ids = label_ids.to(device)
with torch.no_grad():
tmp_eval_loss = model(input_ids, segment_ids,
input_mask, label_ids)
logits = model(input_ids, segment_ids, input_mask)
# logits = logits.detach().cpu().numpy()
# prediction = np.argmax(logits, axis=1)
logits_cpu = logits.detach().cpu()
logits_sigmoid = logits_cpu.sigmoid()
if y_dev_pred is None:
y_dev_sigmoid = logits_sigmoid
else:
y_dev_sigmoid = np.concatenate(
(y_dev_sigmoid, logits_sigmoid), axis=0)
pred_logits = logits.detach().cpu().numpy()
if y_dev_pred is None:
y_dev_pred = pred_logits
else:
y_dev_pred = np.concatenate((y_dev_pred, pred_logits),
axis=0)
if y_dev is None:
y_dev = label_ids.detach().cpu().numpy()
else:
y_dev = np.concatenate(
(y_dev, label_ids.detach().cpu().numpy()), axis=0)
tmp_eval_accuracy = accuracy_thresh(logits, label_ids)
eval_loss += tmp_eval_loss.mean().item()
eval_accuracy += tmp_eval_accuracy
nb_eval_examples += input_ids.size(0)
nb_eval_steps += 1
eval_loss = eval_loss / nb_eval_steps
eval_accuracy = eval_accuracy / nb_eval_examples
loss = tr_loss / nb_tr_steps if args.do_train else None
i = 0
gold_labels = dict()
pred_labels = dict()
prediction_output = dict()
for example, true_logits, pred_logits in zip(
eval_examples, y_dev, y_dev_sigmoid):
true_indexes = np.argwhere(true_logits > 0.0)
labels = [
label_encoder.inverse_transform(y)[0]
for y in true_indexes
]
pred_indexes = np.argwhere(pred_logits > 0.5)
pred = [
label_encoder.inverse_transform(y)[0]
for y in pred_indexes
]
gold_labels[str(example.guid)] = labels
pred_labels[str(example.guid)] = pred
class_logits = {
label_encoder.inverse_transform([j])[0]:
float(pred_logits[j])
for j in range(len(label_encoder.classes_))
}
prediction_output[example.guid] = class_logits
i += 1
pred_output_file = os.path.join(
args.output_dir, f"prediction_output_{epoch+1}.json")
json.dump(prediction_output,
open(pred_output_file, 'w'),
sort_keys=True,
indent=2)
result = {
'eval_loss': eval_loss,
'eval_accuracy': eval_accuracy,
'global_step': global_step,
'loss': loss,
}
eval_util = EvaluationUtil()
pred_file = os.path.join(args.output_dir,
f"dev_pred_{epoch+1}.txt")
eval_util.save_predictions(pred_labels, pred_file)
clef19_result = eval_util.evaluate(pred_labels, gold_labels)
f1_score = clef19_result["eval_fscore"]
if f1_score > max_f1:
# Save a trained model and the associated configuration
model_to_save = model.module if hasattr(
model,
'module') else model # Only save the model it-self
output_model_file = os.path.join(args.output_dir,
WEIGHTS_NAME)
torch.save(model_to_save.state_dict(), output_model_file)
output_config_file = os.path.join(args.output_dir,
CONFIG_NAME)
with open(output_config_file, 'w') as f:
f.write(model_to_save.config.to_json_string())
max_f1 = f1_score
icd10_ontology = data_handler.read_icd10_ontology()
pred_labels_extended = eval_util.extend_paths(
pred_labels, icd10_ontology)
pred_extended_file = os.path.join(
args.output_dir, f"dev_pred_extended_{epoch+1}.txt")
eval_util.save_predictions(pred_labels_extended,
pred_extended_file)
extended_clef19_result = eval_util.evaluate(
pred_labels_extended, gold_labels)
output_eval_file = os.path.join(args.output_dir,
f"eval_results_{epoch+1}.txt")
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
writer.write("%s = %s\n" % (key, str(result[key])))
clef10_result_str = eval_util.format_result(clef19_result)
logger.info(f"CLEF19 evaluation: {clef10_result_str}")
writer.write("\nResults prediction:\n")
writer.write(clef10_result_str)
extended_clef10_result_str = eval_util.format_result(
extended_clef19_result)
logger.info(
f"CLEF19 evaluation (extended): {extended_clef10_result_str}"
)
writer.write("\n\nResults extended prediction:\n")
writer.write(extended_clef10_result_str)
#if not args.do_train:
# (Re-) Load best model
output_model_file = os.path.join(args.output_dir, WEIGHTS_NAME)
output_config_file = os.path.join(args.output_dir, CONFIG_NAME)
config = BertConfig(output_config_file)
model = BertForMultiLabelSequenceClassification(config,
num_labels=num_labels)
model.load_state_dict(torch.load(output_model_file))
model.to(device)
if args.do_eval and (args.local_rank == -1
or torch.distributed.get_rank() == 0):
eval_examples = processor.get_dev_instances()
eval_features = convert_examples_to_features(eval_examples,
label_encoder,
args.max_seq_length,
tokenizer)
logger.info("***** Running evaluation *****")
logger.info(" Num examples = %d", len(eval_examples))
logger.info(" Batch size = %d", args.eval_batch_size)
all_input_ids = torch.tensor([f.input_ids for f in eval_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in eval_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in eval_features],
dtype=torch.long)
all_label_ids = torch.tensor([f.label_ids for f in eval_features],
dtype=torch.float)
eval_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_label_ids)
# Run prediction for full data
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data,
sampler=eval_sampler,
batch_size=args.eval_batch_size)
model.eval()
eval_loss, eval_accuracy = 0, 0
nb_eval_steps, nb_eval_examples = 0, 0
y_dev = None
y_dev_pred = None
y_dev_sigmoid = None
for input_ids, input_mask, segment_ids, label_ids in tqdm(
eval_dataloader, desc="Evaluating"):
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
label_ids = label_ids.to(device)
with torch.no_grad():
tmp_eval_loss = model(input_ids, segment_ids, input_mask,
label_ids)
logits = model(input_ids, segment_ids, input_mask)
#logits = logits.detach().cpu().numpy()
#prediction = np.argmax(logits, axis=1)
logits_cpu = logits.detach().cpu()
logits_sigmoid = logits_cpu.sigmoid()
if y_dev_sigmoid is None:
y_dev_sigmoid = logits_sigmoid
else:
y_dev_sigmoid = np.concatenate((y_dev_sigmoid, logits_sigmoid),
axis=0)
pred_logits = logits.detach().cpu().numpy()
if y_dev_pred is None:
y_dev_pred = pred_logits
else:
y_dev_pred = np.concatenate((y_dev_pred, pred_logits), axis=0)
if y_dev is None:
y_dev = label_ids.detach().cpu().numpy()
else:
y_dev = np.concatenate(
(y_dev, label_ids.detach().cpu().numpy()), axis=0)
tmp_eval_accuracy = accuracy_thresh(logits, label_ids)
eval_loss += tmp_eval_loss.mean().item()
eval_accuracy += tmp_eval_accuracy
nb_eval_examples += input_ids.size(0)
nb_eval_steps += 1
eval_loss = eval_loss / nb_eval_steps
eval_accuracy = eval_accuracy / nb_eval_examples
loss = tr_loss / nb_tr_steps if args.do_train else None
i = 0
gold_labels = dict()
pred_labels = dict()
prediction_output = dict()
for example, true_logits, pred_logits in zip(eval_examples, y_dev,
y_dev_sigmoid):
true_indexes = np.argwhere(true_logits > 0.0)
labels = [
label_encoder.inverse_transform(y)[0] for y in true_indexes
]
pred_indexes = np.argwhere(pred_logits > 0.5)
pred = [
label_encoder.inverse_transform(y)[0] for y in pred_indexes
]
if i < 2:
logger.info(f"Example: {example.guid}")
logger.info(f"Example labels: {example.labels}")
logger.info(f"True labels: {labels}")
logger.info(f"Pred labels: {pred}")
gold_labels[str(example.guid)] = labels
pred_labels[str(example.guid)] = pred
class_logits = {
label_encoder.inverse_transform([j])[0]: float(pred_logits[j])
for j in range(len(label_encoder.classes_))
}
prediction_output[example.guid] = class_logits
i += 1
pred_output_file = os.path.join(args.output_dir,
"prediction_output.json")
json.dump(prediction_output,
open(pred_output_file, 'w'),
sort_keys=True,
indent=2)
result = {
'eval_loss': eval_loss,
'eval_accuracy': eval_accuracy,
'global_step': global_step,
'loss': loss,
}
eval_util = EvaluationUtil()
pred_file = os.path.join(args.output_dir, "dev_pred.txt")
eval_util.save_predictions(pred_labels, pred_file)
clef19_result = eval_util.evaluate(pred_labels, gold_labels)
icd10_ontology = data_handler.read_icd10_ontology()
pred_labels_extended = eval_util.extend_paths(pred_labels,
icd10_ontology)
pred_extended_file = os.path.join(args.output_dir,
"dev_pred_extended.txt")
eval_util.save_predictions(pred_labels_extended, pred_extended_file)
extended_clef19_result = eval_util.evaluate(pred_labels_extended,
gold_labels)
output_eval_file = os.path.join(args.output_dir, "eval_results.txt")
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
writer.write("%s = %s\n" % (key, str(result[key])))
clef10_result_str = eval_util.format_result(clef19_result)
logger.info(f"CLEF19 evaluation: {clef10_result_str}")
writer.write("\nResults prediction:\n")
writer.write(clef10_result_str)
extended_clef10_result_str = eval_util.format_result(
extended_clef19_result)
logger.info(
f"CLEF19 evaluation (extended): {extended_clef10_result_str}")
writer.write("\n\nResults extended prediction:\n")
writer.write(extended_clef10_result_str)
if args.do_test:
test_examples = processor.get_test_instances()
test_features = convert_examples_to_features(test_examples,
label_encoder,
args.max_seq_length,
tokenizer)
logger.info("***** Running test *****")
logger.info(" Num examples = %d", len(test_examples))
logger.info(" Batch size = %d", args.eval_batch_size)
all_input_ids = torch.tensor([f.input_ids for f in test_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in test_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in test_features],
dtype=torch.long)
test_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids)
# Run prediction for full data
test_sampler = SequentialSampler(test_data)
test_dataloader = DataLoader(test_data,
sampler=test_sampler,
batch_size=args.eval_batch_size)
model.eval()
y_dev_sigmoid = None
for input_ids, input_mask, segment_ids in tqdm(test_dataloader,
desc="Testing"):
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
with torch.no_grad():
logits = model(input_ids, segment_ids, input_mask)
logits_sigmoid = logits.detach().cpu().sigmoid()
if y_dev_sigmoid is None:
y_dev_sigmoid = logits_sigmoid
else:
y_dev_sigmoid = np.concatenate((y_dev_sigmoid, logits_sigmoid),
axis=0)
i = 0
test_labels = dict()
test_output = dict()
for example, pred_logits in zip(test_examples, y_dev_sigmoid):
pred_indexes = np.argwhere(pred_logits > 0.5)
pred = [
label_encoder.inverse_transform(y)[0] for y in pred_indexes
]
if i < 2:
logger.info(f"Example: {example.guid}")
logger.info(f"Pred labels: {pred}")
test_labels[str(example.guid)] = pred
class_logits = {
label_encoder.inverse_transform([j])[0]: float(pred_logits[j])
for j in range(len(label_encoder.classes_))
}
test_output[example.guid] = class_logits
i += 1
pred_output_file = os.path.join(args.output_dir, "test_output.json")
json.dump(test_output,
open(pred_output_file, 'w'),
sort_keys=True,
indent=2)
eval_util = EvaluationUtil()
test_file = os.path.join(args.output_dir, "test_pred.txt")
eval_util.save_predictions(test_labels, test_file)
icd10_ontology = data_handler.read_icd10_ontology()
test_labels_extended = eval_util.extend_paths(test_labels,
icd10_ontology)
test_extended_file = os.path.join(args.output_dir,
"test_pred_extended.txt")
eval_util.save_predictions(test_labels_extended, test_extended_file)
class CaffeProcThread(CodependentThread):
'''Runs Caffe in separate thread.'''
def __init__(self, net, state, loop_sleep, pause_after_keys, heartbeat_required):
CodependentThread.__init__(self, heartbeat_required)
self.daemon = True
self.net = net
self.input_dims = self.net.blobs['data'].data.shape[2:4] # e.g. (227,227)
self.state = state
self.frames_processed_fwd = 0
self.frames_processed_back = 0
self.loop_sleep = loop_sleep
self.pause_after_keys = pause_after_keys
self.debug_level = 0
self.descriptor = None
self.descriptor_layer_1 = 'conv5'
self.descriptor_layer_2 = 'conv4'
self.descriptor_layers = ['conv5','conv4']
self.net_input_image = None
self.descriptor_handler = DescriptorHandler(self.state.settings.ros_dir + '/models/memory/', self.descriptor_layers)
self.data_handler = DataHandler(self.state.settings.ros_dir)
self.available_layer = ['conv1', 'pool1', 'norm1', 'conv2', 'pool2', 'norm2', 'conv3', 'conv4', 'conv5', 'pool5', 'fc6', 'fc7', 'fc8', 'prob']
#['conv1', 'conv2', 'conv3', 'conv4', 'conv5', 'fc6', 'fc7', 'fc8', 'prob']
# print "layers ", list(self.net._layer_names)
s = rospy.Service('get_cnn_state', GetState, self.handle_get_cnn_state)
def handle_get_cnn_state(self, req):
resp = GetStateResponse()
print "get req", req
state = perception_msgs.msg.State()
state.type = req.type
state.name = req.name
if req.name == "None":
state.name = self.data_handler.save_net(self.net)
else:
print "not implemented yet"
resp.state = state
resp.pose = Pose()
print "state", resp
return resp
def mask_out(self, data, mask):
# print "data shape", data.shape
dim = data.shape
for y in range(dim[2]):
for x in range(dim[3]):
if is_masked((dim[2],dim[3]),(x,y),mask):
data[:,:,y,x] = 0
return data
def net_preproc_forward(self, img):
assert img.shape == (227,227,3), 'img is wrong size'
#resized = caffe.io.resize_image(img, net.image_dims) # e.g. (227, 227, 3)
data_blob = self.net.transformer.preprocess('data', img) # e.g. (3, 227, 227), mean subtracted and scaled to [0,255]
data_blob = data_blob[np.newaxis,:,:,:] # e.g. (1, 3, 227, 227)
output = self.net.forward(data=data_blob)
return output
def net_proc_forward_layer(self, img, mask):
assert img.shape == (227,227,3), 'img is wrong size'
#resized = caffe.io.resize_image(img, net.image_dims) # e.g. (227, 227, 3)
data_blob = self.net.transformer.preprocess('data', img) # e.g. (3, 227, 227), mean subtracted and scaled to [0,255]
data_blob = data_blob[np.newaxis,:,:,:] # e.g. (1, 3, 227, 227)
# print "mask", mask.shape
self.net.blobs['data'].data[...] = data_blob
proc_layers = ['conv1', 'conv2', 'conv3', 'conv4', 'conv5', 'prob']
mask_layers = ['']
mode = 0
if mode == 0:
self.net.forward_from_to(start='conv1',end='relu1')
self.mask_out(self.net.blobs['conv1'].data, mask)
self.net.forward_from_to(start='relu1',end='prob')
elif mode == 1:
self.net.forward_from_to(start='conv1',end='relu1')
self.mask_out(self.net.blobs['conv1'].data, mask)
self.net.forward_from_to(start='relu1',end='conv2')
self.net.forward_from_to(start='conv2',end='relu2')
self.mask_out(self.net.blobs['conv2'].data, mask)
self.net.forward_from_to(start='relu2',end='conv3')
self.net.forward_from_to(start='conv3',end='relu3')
self.mask_out(self.net.blobs['conv3'].data, mask)
self.net.forward_from_to(start='relu3',end='conv4')
self.net.forward_from_to(start='conv4',end='relu4')
self.mask_out(self.net.blobs['conv4'].data, mask)
self.net.forward_from_to(start='relu4',end='conv5')
self.net.forward_from_to(start='conv5',end='relu5')
self.mask_out(self.net.blobs['conv5'].data, mask)
self.net.forward_from_to(start='relu5',end='prob')
elif mode == 2:
for idx in range(0,len(proc_layers)-1):
output = self.net.forward_from_to(start=proc_layers[idx],end=proc_layers[idx+1])
self.mask_out(self.net.blobs[proc_layers[idx]].data, mask)
# for idx in range(len(self.available_layer)-1):
# output = self.net.forward(data=data_blob,start=self.available_layer[idx],end=self.available_layer[idx+1])
# if self.available_layer[idx].startswith("conv"):
# new_blob = self.net.blobs[self.available_layer[idx]].data
# new_blob.data = self.mask_out(self.net.blobs[self.available_layer[idx]].data, mask)
# self.net.blobs[self.available_layer[idx]] = new_blob
# print output
# return output
def save_descriptor(self):
print 'save descriptor'
# print type(self.net.blobs[self.descriptor_layer_1])
# descriptor_1 = np.array(caffe.io.blobproto_to_array(self.net.blobs[self.descriptor_layer_1]))
# descriptor_2 = np.array(caffe.io.blobproto_to_array(self.net.blobs[self.descriptor_layer_2]))
# print descriptor
self.time_name = strftime("%d-%m-%Y-%H:%M:%S", gmtime())
desc = self.descriptor_handler.gen_descriptor(self.time_name, self.net.blobs)
self.descriptor_handler.save(self.state.settings.ros_dir + '/models/memory/', desc)
# np.save(self.state.settings.ros_dir + '/models/memory/' + self.time_name + "_" + self.descriptor_layer_1 + '.npy', descriptor_1)
# np.save(self.state.settings.ros_dir + '/models/memory/' + self.time_name + "_" + self.descriptor_layer_2 + '.npy', descriptor_2)
cv2.imwrite(self.state.settings.ros_dir + '/models/memory/' + self.time_name + '.jpg', self.net_input_image[:,:,::-1])
self.state.save_descriptor = False
def run(self):
print 'CaffeProcThread.run called'
frame = None
mask = None
while not self.is_timed_out():
with self.state.lock:
if self.state.quit:
#print 'CaffeProcThread.run: quit is True'
#print self.state.quit
break
#print 'CaffeProcThread.run: caffe_net_state is:', self.state.caffe_net_state
#print 'CaffeProcThread.run loop: next_frame: %s, caffe_net_state: %s, back_enabled: %s' % (
# 'None' if self.state.next_frame is None else 'Avail',
# self.state.caffe_net_state,
# self.state.back_enabled)
frame = None
mask = None
run_fwd = False
run_back = False
if self.state.caffe_net_state == 'free' and time.time() - self.state.last_key_at > self.pause_after_keys:
frame = self.state.next_frame
mask = self.state.mask
self.state.next_frame = None
back_enabled = self.state.back_enabled
back_mode = self.state.back_mode
back_stale = self.state.back_stale
#state_layer = self.state.layer
#selected_unit = self.state.selected_unit
backprop_layer = self.state.backprop_layer
backprop_unit = self.state.backprop_unit
# Forward should be run for every new frame
run_fwd = (frame is not None)
# Backward should be run if back_enabled and (there was a new frame OR back is stale (new backprop layer/unit selected))
run_back = (back_enabled and (run_fwd or back_stale))
self.state.caffe_net_state = 'proc' if (run_fwd or run_back) else 'free'
#print 'run_fwd,run_back =', run_fwd, run_back
if run_fwd:
#print 'TIMING:, processing frame'
self.frames_processed_fwd += 1
self.net_input_image = cv2.resize(frame, self.input_dims)
with WithTimer('CaffeProcThread:forward', quiet = self.debug_level < 1):
print "run forward layer"
self.net_proc_forward_layer(self.net_input_image, mask)
# self.net_preproc_forward(self.net_input_image)
if self.state.save_descriptor:
self.save_descriptor()
# switch descriptor for match and back prop
if self.state.next_descriptor:
print 'load descriptor'
self.descriptor = self.descriptor_handler.get_next()
self.state.next_descriptor = False
if self.state.compare_descriptor:
# find
print 'compare'
desc_current = self.descriptor_handler.gen_descriptor('current', self.net.blobs)
match_file = self.descriptor_handler.get_max_match(desc_current)
print 'match: ' + match_file
self.state.compare_descriptor = False
if run_back:
print "run backward"
# Match to saved descriptor
if self.state.match_descriptor:
print '*'
diffs = self.net.blobs[self.descriptor_layer_1].diff * 0
# zero all diffs if doesn't match
print "shape ", self.net.blobs[self.descriptor_layer_1].data.shape
for unit, response in enumerate(self.net.blobs[self.descriptor_layer_1].data[0]):
if response.max() > 0 and abs(response.max() - self.descriptor.get_sig_list()[0][0][unit].max())/response.max() < 0.2:
diffs[0][unit] = self.net.blobs[self.descriptor_layer_1].data[0][unit]
assert back_mode in ('grad', 'deconv')
if back_mode == 'grad':
with WithTimer('CaffeProcThread:backward', quiet = self.debug_level < 1):
#print '**** Doing backprop with %s diffs in [%s,%s]' % (backprop_layer, diffs.min(), diffs.max())
self.net.backward_from_layer(self.descriptor_layer_1, diffs, zero_higher = True)
else:
with WithTimer('CaffeProcThread:deconv', quiet = self.debug_level < 1):
#print '**** Doing deconv with %s diffs in [%s,%s]' % (backprop_layer, diffs.min(), diffs.max())
self.net.deconv_from_layer(self.descriptor_layer_1, diffs, zero_higher = True)
with self.state.lock:
self.state.back_stale = False
# Filter when back propagating
elif self.state.backprop_filter:
print "run_back"
# print backprop_layer
start_layer_idx = self.available_layer.index(backprop_layer)
idx = start_layer_idx
for current_layer in list(reversed(self.available_layer[0:start_layer_idx+1])):
diffs = self.net.blobs[current_layer].diff * 0
max_response = self.net.blobs[current_layer].data[0].max()
for unit, response in enumerate(self.net.blobs[current_layer].data[0]):
if response.max() > max_response * 0.6:
diffs[0][unit] = self.net.blobs[current_layer].data[0,unit]
assert back_mode in ('grad', 'deconv')
if back_mode == 'grad':
with WithTimer('CaffeProcThread:backward', quiet = self.debug_level < 1):
#print '**** Doing backprop with %s diffs in [%s,%s]' % (backprop_layer, diffs.min(), diffs.max())
self.net.backward_from_to_layer(current_layer, diffs, self.available_layer[idx-1], zero_higher = (idx == start_layer_idx))
# else:
# with WithTimer('CaffeProcThread:deconv', quiet = self.debug_level < 1):
# #print '**** Doing deconv with %s diffs in [%s,%s]' % (backprop_layer, diffs.min(), diffs.max())
# self.net.deconv_from_layer(backprop_layer, diffs, zero_higher = True)
idx -= 1
with self.state.lock:
self.state.back_stale = False
# original approach
else:
diffs = self.net.blobs[backprop_layer].diff * 0
diffs[0][backprop_unit] = self.net.blobs[backprop_layer].data[0,backprop_unit]
assert back_mode in ('grad', 'deconv')
if back_mode == 'grad':
with WithTimer('CaffeProcThread:backward', quiet = self.debug_level < 1):
#print '**** Doing backprop with %s diffs in [%s,%s]' % (backprop_layer, diffs.min(), diffs.max())
self.net.backward_from_layer(backprop_layer, diffs, zero_higher = True)
else:
with WithTimer('CaffeProcThread:deconv', quiet = self.debug_level < 1):
#print '**** Doing deconv with %s diffs in [%s,%s]' % (backprop_layer, diffs.min(), diffs.max())
self.net.deconv_from_layer(backprop_layer, diffs, zero_higher = True)
with self.state.lock:
self.state.back_stale = False
if run_fwd or run_back:
with self.state.lock:
self.state.caffe_net_state = 'free'
self.state.drawing_stale = True
else:
time.sleep(self.loop_sleep)
time.sleep(0.1)
print 'CaffeProcThread.run: finished'
print 'CaffeProcThread.run: processed %d frames fwd, %d frames back' % (self.frames_processed_fwd, self.frames_processed_back)
# crossref=TEXT(stored=True), isbn=TEXT, series=TEXT, school=TEXT,
# chapter=TEXT(stored=True, analyzer=StemmingAnalyzer()),
# publnr=TEXT(stored=True))
# mi serve nel caso in cui non ho fields specificati nella query, vado a cercare in ogni field
schemaFields = ["type","author","title","year", "journal","ee","publisher"]
#se l'indice è già costruito non lo devo rifare
if not os.path.exists("indexdir"):
os.mkdir("indexdir")
ix = create_in("indexdir", schema)
# get the number of processors
nproc = multiprocessing.cpu_count()
writer = ix.writer(procs=nproc, limitmb=512)
parser = xml.sax.make_parser()
handler = DH.DataHandler(writer)
parser.setContentHandler(handler)
parser.parse('../dblp.xml')
writer.commit(optimize=True)
else:
ix = whoosh.index.open_dir("indexdir", schema=schema)
# il searcher va fatto dopo la commit!!
print()
print("scegli il modello di ranking: ")
print("1) BM25F")
print("2) PL2")
choice = int(input())
if choice == 1:
searcher = ix.searcher(weighting=whoosh.scoring.BM25F)
class Trainer:
def __init__(self, cfg):
self.cfg = cfg
self.cat_features = self.cfg.data.features.cat_features
self.oof = None
self.raw_preds = None
self.weights = []
self.models = []
self.scores = []
self.feature_importance_df = pd.DataFrame(
columns=['feature', 'importance'])
self.dh = DataHandler()
def train(self, train_df: pd.DataFrame, target_df: pd.DataFrame,
fold_df: pd.DataFrame):
self.oof = np.zeros(len(train_df))
for fold_, col in enumerate(fold_df.columns):
print(
f'\n========================== FOLD {fold_} ... ==========================\n'
)
logging.debug(
f'\n========================== FOLD {fold_} ... ==========================\n'
)
self._train_fold(train_df, target_df, fold_df[col])
print('\n\n===================================\n')
print(f'CV: {np.mean(self.scores):.4f}')
print('\n===================================\n\n')
logging.debug('\n\n===================================\n')
logging.debug(f'CV: {np.mean(self.scores):.4f}')
logging.debug('\n===================================\n\n')
return np.mean(self.scores)
def _train_fold(self, train_df, target_df, fold):
tr_x, va_x = train_df[fold == 0], train_df[fold > 0]
tr_y, va_y = target_df[fold == 0], target_df[fold > 0]
weight = fold.max()
self.weights.append(weight)
model = factory.get_model(self.cfg.model)
model.fit(tr_x, tr_y, va_x, va_y, self.cat_features)
va_pred = model.predict(va_x, self.cat_features)
if self.cfg.data.target.reconvert_type:
va_y = getattr(np, self.cfg.data.target.reconvert_type)(va_y)
va_pred = getattr(np, self.cfg.data.target.reconvert_type)(va_pred)
va_pred = np.where(va_pred >= 0, va_pred, 0)
self.models.append(model)
self.oof[va_x.index] = va_pred.copy()
score = factory.get_metrics(self.cfg.common.metrics.name)(va_y,
va_pred)
self.scores.append(score)
if self.cfg.model.name in ['lightgbm', 'catboost', 'xgboost']:
importance_fold_df = pd.DataFrame()
fold_importance = model.extract_importances()
importance_fold_df['feature'] = train_df.columns
importance_fold_df['importance'] = fold_importance
self.feature_importance_df = pd.concat(
[self.feature_importance_df, importance_fold_df], axis=0)
def predict(self, test_df):
preds = np.zeros(len(test_df))
for fold_, model in enumerate(self.models):
pred = model.predict(test_df, self.cat_features)
if self.cfg.data.target.reconvert_type:
pred = getattr(np, self.cfg.data.target.reconvert_type)(pred)
pred = np.where(pred >= 0, pred, 0)
preds += pred.copy() * self.weights[fold_]
self.raw_preds = preds.copy()
return preds
def save(self, run_name):
log_dir = Path(f'../logs/{run_name}')
self.dh.save(log_dir / 'oof.npy', self.oof)
self.dh.save(log_dir / 'raw_preds.npy', self.raw_preds)
self.dh.save(log_dir / 'importance.csv', self.feature_importance_df)
self.dh.save(log_dir / 'model_weight.pkl', self.models)
from utils import (DataHandler, Kaggle, Notion, Timer, make_submission,
seed_everything, send_line)
warnings.filterwarnings('ignore')
# ===============
# Settings
# ===============
parser = argparse.ArgumentParser()
parser.add_argument('--common', default='../configs/common/compe.yml')
parser.add_argument('--notify', default='../configs/common/notify.yml')
parser.add_argument('-m', '--model')
parser.add_argument('-c', '--comment')
options = parser.parse_args()
dh = DataHandler()
cfg = dh.load(options.common)
cfg.update(dh.load(f'../configs/exp/{options.model}.yml'))
notify_params = dh.load(options.notify)
comment = options.comment
model_name = options.model
now = datetime.datetime.now()
run_name = f'{model_name}_{now:%Y%m%d%H%M%S}'
logger_path = Path(f'../logs/{run_name}')
# ===============
# Main
class Solver:
def __init__(self, params):
# Build data handler
self.data_handler = DataHandler()
self.data_handler.loadData(params['input'])
params['inp_dim'] = self.data_handler.getDataShape()[
1] # inp_dim = 1000
logging.info("=" * 41)
# Build model
self.model = SPINEModel(params)
self.dtype = torch.FloatTensor
use_cuda = torch.cuda.is_available()
if use_cuda:
self.model.cuda()
self.dtype = torch.cuda.FloatTensor
self.optimizer = torch.optim.SGD(self.model.parameters(), lr=0.1)
logging.info("=" * 41)
def train(self, params):
num_epochs, batch_size = params['num_epochs'], params['batch_size']
optimizer = self.optimizer
dtype = self.dtype
for iteration in range(num_epochs): # for each epoch
self.data_handler.shuffleTrain() # 15000 training data shuffled
num_batches = self.data_handler.getNumberOfBatches(
batch_size) # num_batches = number of iterations
epoch_losses = np.zeros(
5) # rl, asl, psl, total !!!!!!!!!!!!!!!!!!
for batch_idx in range(num_batches): # for each iteration
optimizer.zero_grad()
batch_x, batch_y = self.data_handler.getBatch(
batch_idx, batch_size, params['noise_level'],
params['denoising'])
batch_x = Variable(torch.from_numpy(batch_x),
requires_grad=False).type(dtype)
batch_y = Variable(torch.from_numpy(batch_y),
requires_grad=False).type(
dtype) #dtype = torch.FloatTensor
out, h, loss, loss_terms = self.model(batch_x, batch_y)
reconstruction_loss, psl_loss, asl_loss, local_loss = loss_terms #!!!!!!!!!!!!!!!!!!!!!
loss.backward()
optimizer.step()
#print(local_loss)
epoch_losses[0] += reconstruction_loss.item()
epoch_losses[1] += asl_loss.item()
epoch_losses[2] += psl_loss.item()
epoch_losses[3] += local_loss.item() #!!!!!!!!!!!!!!!!!!!!!!!!
epoch_losses[4] += loss.item()
# epoch_losses[0]+=reconstruction_loss.data[0]
# epoch_losses[1]+=asl_loss.data[0]
# epoch_losses[2]+=psl_loss.data[0]
# epoch_losses[3]+=local_loss.data[0]
# epoch_losses[4]+=loss.data[0]
print("After epoch %r, Reconstruction Loss = %.4f, ASL = %.4f,"\
"PSL = %.4f, Local Loss = %.4f and total = %.4f"
%(iteration+1, epoch_losses[0], epoch_losses[1], epoch_losses[2], epoch_losses[3], epoch_losses[4]) ) #!!!!!!
#logging.info("After epoch %r, Sparsity = %.1f"
# %(iteration+1, utils.compute_sparsity(h.cpu().data.numpy())))
#break
#break
def getSpineEmbeddings(self, batch_size, params):
ret = []
self.data_handler.resetDataOrder()
num_batches = self.data_handler.getNumberOfBatches(batch_size)
for batch_idx in range(num_batches):
batch_x, batch_y = self.data_handler.getBatch(
batch_idx, batch_size, params['noise_level'],
params['denoising'])
batch_x = Variable(torch.from_numpy(batch_x),
requires_grad=False).type(self.dtype)
batch_y = Variable(torch.from_numpy(batch_y),
requires_grad=False).type(self.dtype)
_, h, _, _ = self.model(batch_x, batch_y)
ret.extend(h.cpu().data.numpy())
return np.array(ret)
def getWordsList(self):
return self.data_handler.getWordsList()
class Solver:
def __init__(self, params):
# Build data handler
self.data_handler = DataHandler()
self.data_handler.loadData(params['input'])
params['inp_dim'] = self.data_handler.getDataShape()[1]
logging.info("=" * 41)
# Build model
self.model = SPINEModel(params)
self.dtype = torch.FloatTensor
use_cuda = torch.cuda.is_available()
if use_cuda:
self.model.cuda()
self.dtype = torch.cuda.FloatTensor
self.optimizer = torch.optim.SGD(self.model.parameters(), lr=0.1)
logging.info("=" * 41)
def train(self, params):
num_epochs, batch_size = params['num_epochs'], params['batch_size'],
optimizer = self.optimizer
dtype = self.dtype
for iteration in range(num_epochs):
self.data_handler.shuffleTrain()
num_batches = self.data_handler.getNumberOfBatches(batch_size)
epoch_losses = np.zeros(5) # rl, asl, psl, Siml, total
sparsity = 0
for batch_idx in range(num_batches):
optimizer.zero_grad()
batch_x, batch_y = self.data_handler.getBatch(
batch_idx, batch_size, params['noise_level'],
params['denoising'])
batch_x = Variable(torch.from_numpy(batch_x),
requires_grad=False).type(dtype)
batch_y = Variable(torch.from_numpy(batch_y),
requires_grad=False).type(dtype)
out, h, loss, loss_terms, sparsity_ratio = self.model(
batch_x, batch_y)
reconstruction_loss, psl_loss, asl_loss, siml = loss_terms
loss.backward()
optimizer.step()
epoch_losses[0] += reconstruction_loss.data
epoch_losses[1] += asl_loss.data
epoch_losses[2] += psl_loss.data
epoch_losses[3] += loss.data
epoch_losses[4] += siml.data
sparsity += sparsity_ratio.data
print("After epoch %r, Reconstruction Loss = %.4f, ASL = %.4f,"\
"PSL = %.4f, SimL = %.4f , and total = %.4f, sparsity = %.4f"
%(iteration+1, epoch_losses[0], epoch_losses[1], epoch_losses[2], epoch_losses[4], epoch_losses[3], sparsity/num_batches) )
if iteration % 1000 == 0:
spine_embeddings = self.getSpineEmbeddings(512, params)
utils.dump_vectors(spine_embeddings,
'./' + str(iteration / 1000) + '.txt',
self.getWordsList())
#logging.info("After epoch %r, Sparsity = %.1f"
# %(iteration+1, utils.compute_sparsity(h.cpu().data.numpy())))
#break
#break
def getSpineEmbeddings(self, batch_size, params):
ret = []
self.data_handler.resetDataOrder()
num_batches = self.data_handler.getNumberOfBatches(batch_size)
for batch_idx in range(num_batches):
batch_x, batch_y = self.data_handler.getBatch(
batch_idx, batch_size, params['noise_level'],
params['denoising'])
batch_x = Variable(torch.from_numpy(batch_x),
requires_grad=False).type(self.dtype)
batch_y = Variable(torch.from_numpy(batch_y),
requires_grad=False).type(self.dtype)
_, h, _, _, spars = self.model(batch_x, batch_y)
ret.extend(h.cpu().data.numpy())
return np.array(ret)
def getWordsList(self):
return self.data_handler.getWordsList()
def exit_press(self):
# Save portfolio
DataHandler.save_portfolio(current_portoflio.portfolio)
# Exit
sys.exit(0)
def train(opt, model, input_img):
model.cuda()
handler = DataHandler(opt, input_img)
learning_rate = opt.learning_rate
min_learning_rate = opt.min_learning_rate
learning_rate_change_iter_nums = [0]
mse_steps = []
mse_rec = []
criterionL1 = nn.L1Loss()
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
with tqdm.tqdm(miniters=1, mininterval=0) as progress:
for iter, (hr, lr) in enumerate(handler.preprocess_data()):
lr = lr.cuda()
hr = hr.cuda()
output = model(lr) + lr
loss = criterionL1(output, hr)
cpu_loss = loss.data.cpu().numpy()
model.zero_grad()
optimizer.zero_grad()
progress.set_description("Iteration: {} Loss: {}, Learning rate: {}".format( \
iter, cpu_loss, learning_rate))
progress.update()
if iter > 0 and iter % 10000 == 0:
learning_rate = learning_rate / 10
adjust_learning_rate(optimizer, new_lr=learning_rate)
print("Learning rate reduced to {lr}".format(lr=learning_rate) )
"""
if (not (1 + iter) % opt.learning_rate_policy_check_every
and iter - learning_rate_change_iter_nums[-1] > opt.min_iters):
[slope, _], [[var, _], _] = np.polyfit(mse_steps[-int(opt.learning_rate_slope_range /
opt.run_test_every):],
mse_rec[-int(opt.learning_rate_slope_range /
opt.run_test_every):],
1, cov=True)
std = np.sqrt(var)
if -opt.learning_rate_change_ratio * slope < std:
learning_rate /= 10
learning_rate_change_iter_nums.append(iter)
"""
loss.backward()
optimizer.step()
if learning_rate < min_learning_rate:
print('Done training')
break
