def load_my_dataset(filepath, datafolderpath, labelfilepath, test_ratio=0.9):
labellist = load_label(labelfilepath)
with open(filepath, "r") as readfile:
filepathlist = [
os.path.join(datafolderpath, line.strip())
for line in readfile.readlines()
]
datasets = list()
for imagefilepath in filepathlist:
print(imagefilepath)
img = load_data(imagefilepath)
filename = imagefilepath.split(os.path.sep)[-2]
label = filename
index = labellist.index(label)
datasets.append([img, index])
random.shuffle(datasets)
index = int(len(datasets) * test_ratio)
traindatasets = datasets[:index]
testdatasets = datasets[index:]
train_imgs = [train_data[0] for train_data in traindatasets]
train_labels = [train_data[1] for train_data in traindatasets]
test_imgs = [test_data[0] for test_data in testdatasets]
test_labels = [test_data[1] for test_data in testdatasets]
return chainer.datasets.tuple_dataset.TupleDataset(train_imgs, train_labels), \
chainer.datasets.tuple_dataset.TupleDataset(test_imgs, test_labels)
def data_load(self, label):
data_path = os.path.join(self.opt.dataroot, self.opt.cap_scheme,
label) # ex. ../data/min24/synthetic
self.num_train_samples = min(self.opt.train_size,
len(os.listdir(os.path.join(data_path,
'train')))) # number of train data in directory
self.num_test_sample = min(2000,
len(os.listdir(os.path.join(data_path,
'test')))) # number of test data in directory
# load training set
x_train = np.empty((self.num_train_samples, self.opt.loadHeight,
self.opt.loadWidth, 1), dtype='uint8') # initialize
y_train = np.empty(
(self.num_train_samples, self.opt.cap_len * self.opt.char_set_len),
dtype='uint8') # initialize
train_labels = util.load_label(
os.path.join(data_path, label + '_train.txt'))
for i in range(self.num_train_samples):
img_name = os.path.join(data_path, 'train', str(i) + '.png')
x_train[i, :, :, :] = util.load_image(img_name)
try:
y_train[i, :] = self.text2vec(train_labels[i])
except:
print(i)
# load test set
x_test = np.empty(
(self.num_test_sample, self.opt.loadHeight, self.opt.loadWidth, 1),
dtype='uint8')
y_test = np.empty(
(self.num_test_sample, self.opt.cap_len * self.opt.char_set_len),
dtype='uint8')
test_labels = util.load_label(
os.path.join(data_path, label + '_test.txt'))
for i in range(self.num_test_sample):
img_name = os.path.join(data_path, 'test', str(i) + '.png')
x_test[i, :, :, :] = util.load_image(img_name)
try:
y_test[i, :] = self.text2vec(test_labels[i])
except:
print(i)
return (x_train, y_train), (x_test, y_test)
def main():
"""
main
"""
n_units_h1, n_units_h2, n_out,\
modelpath, labelfilepath, imagefilepath = importingargs()
model = L.Classifier(LinearNet(n_units_h1, n_units_h2, n_out))
chainer.serializers.load_npz(modelpath, model)
img = load_data(imagefilepath)
labellist = load_label(labelfilepath)
index, prob, label = predict(model, img, labellist)
print("index: %d, prob: %f, label: %s" % (index, prob, label))
def train_classifier():
X_train, y_train, X_test, y_test = load_label(split=True)
vectorizer = joblib.load(
data_path / "tfidf_vectorizer.pkl"
) # vectorizer dumped on cleaning.py
x = vectorizer.transform(X_train)
classifier = RandomForestClassifier(max_depth=10, random_state=42)
classifier.fit(x, y_train)
joblib.dump(classifier, model_path / "classifier.pkl")
test_array = vectorizer.transform([X_train[1]])
test = classifier.predict(test_array)
print(
accuracy_score(
[classifier.predict(vectorizer.transform([a]).toarray()) for a in X_test],
y_test,
)
)
def load_model(self, checkpoint):
config = self.config
opt = config['opt']
labels = load_label(opt.label_path)
label_size = len(labels)
config['labels'] = labels
self.labels = labels
if config['emb_class'] == 'glove':
if config['enc_class'] == 'gnb':
model = TextGloveGNB(config, opt.embedding_path, label_size)
if config['enc_class'] == 'cnn':
model = TextGloveCNN(config,
opt.embedding_path,
label_size,
emb_non_trainable=True)
if config['enc_class'] == 'densenet-cnn':
model = TextGloveDensenetCNN(config,
opt.embedding_path,
label_size,
emb_non_trainable=True)
if config['enc_class'] == 'densenet-dsa':
model = TextGloveDensenetDSA(config,
opt.embedding_path,
label_size,
emb_non_trainable=True)
else:
from transformers import AutoTokenizer, AutoConfig, AutoModel
bert_config = AutoConfig.from_pretrained(opt.bert_output_dir)
bert_tokenizer = AutoTokenizer.from_pretrained(opt.bert_output_dir)
bert_model = AutoModel.from_config(bert_config)
ModelClass = TextBertCNN
if config['enc_class'] == 'cls': ModelClass = TextBertCLS
model = ModelClass(config, bert_config, bert_model, bert_tokenizer,
label_size)
model.load_state_dict(checkpoint)
model = model.to(opt.device)
logger.info("[Model loaded]")
return model
def load_model(config, checkpoint):
opt = config['opt']
labels = load_label(opt.label_path)
label_size = len(labels)
config['labels'] = labels
if config['emb_class'] == 'glove':
if config['enc_class'] == 'gnb':
model = TextGloveGNB(config, opt.embedding_path, label_size)
if config['enc_class'] == 'cnn':
model = TextGloveCNN(config,
opt.embedding_path,
label_size,
emb_non_trainable=True)
if config['enc_class'] == 'densenet-cnn':
model = TextGloveDensenetCNN(config,
opt.embedding_path,
label_size,
emb_non_trainable=True)
if config['enc_class'] == 'densenet-dsa':
model = TextGloveDensenetDSA(config,
opt.embedding_path,
label_size,
emb_non_trainable=True)
else:
from transformers import AutoTokenizer, AutoConfig, AutoModel
bert_config = AutoConfig.from_pretrained(opt.bert_output_dir)
bert_tokenizer = AutoTokenizer.from_pretrained(opt.bert_output_dir)
bert_model = AutoModel.from_config(bert_config)
ModelClass = TextBertCNN
if config['enc_class'] == 'cls': ModelClass = TextBertCLS
model = ModelClass(config, bert_config, bert_model, bert_tokenizer,
label_size)
if opt.enable_qat:
assert opt.device == 'cpu'
model.qconfig = torch.quantization.get_default_qat_qconfig('fbgemm')
'''
# fuse if applicable
# model = torch.quantization.fuse_modules(model, [['']])
'''
model = torch.quantization.prepare_qat(model)
model.eval()
model.to('cpu')
logger.info("[Convert to quantized model with device=cpu]")
model = torch.quantization.convert(model)
if opt.enable_qat_fx:
import torch.quantization.quantize_fx as quantize_fx
qconfig_dict = {
"": torch.quantization.get_default_qat_qconfig('fbgemm')
}
model = quantize_fx.prepare_qat_fx(model, qconfig_dict)
logger.info("[Convert to quantized model]")
model = quantize_fx.convert_fx(model)
model.load_state_dict(checkpoint)
model = model.to(opt.device)
'''
for name, param in model.named_parameters():
print(name, param.data, param.device, param.requires_grad)
'''
logger.info("[model] :\n{}".format(model.__str__()))
logger.info("[Model loaded]")
return model
parser.add_argument("--label-num", "-l", type=int, default=5)
parser.add_argument("--modelpath", "-mf", help="model path")
parser.add_argument("--labelfilepath", "-lf", help="labelfile")
parser.add_argument("--imagefilepath", "-if",
help="imagefilepath you want to predict")
args = parser.parse_args()
return args.first_hidden_layer_units,\
args.second_hidden_layer_units, args.label_num,\
args.modelpath, args.labelfilepath, args.imagefilepath
if __name__ == "__main__":
n_units_h1, n_units_h2, n_out,\
modelpath, labelfilepath, imagefilepath = importingargs()
model = L.Classifier(LinearNet(n_units_h1, n_units_h2, n_out))
chainer.serializers.load_npz(modelpath, model)
labellist = load_label(labelfilepath)
cascade_path = 'haarcascade_frontalface_alt.xml'
face_cascade = cv2.CascadeClassifier(cascade_path)
cap = cv2.VideoCapture(0)
frame_count = 0
face_count = 0
while True:
frame_count += 1
face_count = 0
# 内蔵カメラから読み込んだキャプチャデータを取得
ret, frame = cap.read()
def load_model(config, checkpoint):
args = config['args']
labels = load_label(args.label_path)
label_size = len(labels)
config['labels'] = labels
if config['emb_class'] == 'glove':
if config['enc_class'] == 'gnb':
model = TextGloveGNB(config, args.embedding_path, label_size)
if config['enc_class'] == 'cnn':
model = TextGloveCNN(config,
args.embedding_path,
label_size,
emb_non_trainable=True)
if config['enc_class'] == 'densenet-cnn':
model = TextGloveDensenetCNN(config,
args.embedding_path,
label_size,
emb_non_trainable=True)
if config['enc_class'] == 'densenet-dsa':
model = TextGloveDensenetDSA(config,
args.embedding_path,
label_size,
emb_non_trainable=True)
else:
if config['emb_class'] == 'bart' and config['use_kobart']:
from transformers import BartModel
from kobart import get_kobart_tokenizer, get_pytorch_kobart_model
bert_tokenizer = get_kobart_tokenizer()
bert_tokenizer.cls_token = '<s>'
bert_tokenizer.sep_token = '</s>'
bert_tokenizer.pad_token = '<pad>'
bert_model = BartModel.from_pretrained(get_pytorch_kobart_model())
bert_config = bert_model.config
elif config['emb_class'] in ['gpt']:
bert_tokenizer = AutoTokenizer.from_pretrained(
args.bert_output_dir)
bert_tokenizer.bos_token = '<|startoftext|>'
bert_tokenizer.eos_token = '<|endoftext|>'
bert_tokenizer.cls_token = '<|startoftext|>'
bert_tokenizer.sep_token = '<|endoftext|>'
bert_tokenizer.pad_token = '<|pad|>'
bert_config = AutoConfig.from_pretrained(args.bert_output_dir)
bert_model = AutoModel.from_pretrained(args.bert_output_dir)
elif config['emb_class'] in ['t5']:
from transformers import T5EncoderModel
bert_tokenizer = AutoTokenizer.from_pretrained(
args.bert_output_dir)
bert_tokenizer.cls_token = '<s>'
bert_tokenizer.sep_token = '</s>'
bert_tokenizer.pad_token = '<pad>'
bert_config = AutoConfig.from_pretrained(args.bert_output_dir)
bert_model = T5EncoderModel(bert_config)
else:
bert_tokenizer = AutoTokenizer.from_pretrained(
args.bert_output_dir)
bert_config = AutoConfig.from_pretrained(args.bert_output_dir)
bert_model = AutoModel.from_config(bert_config)
ModelClass = TextBertCNN
if config['enc_class'] == 'cls': ModelClass = TextBertCLS
if config['enc_class'] == 'densenet-cnn':
ModelClass = TextBertDensenetCNN
model = ModelClass(config, bert_config, bert_model, bert_tokenizer,
label_size)
if args.enable_qat:
assert args.device == 'cpu'
model.qconfig = torch.quantization.get_default_qat_qconfig('fbgemm')
'''
# fuse if applicable
# model = torch.quantization.fuse_modules(model, [['']])
'''
model = torch.quantization.prepare_qat(model)
model.eval()
model.to('cpu')
logger.info("[Convert to quantized model with device=cpu]")
model = torch.quantization.convert(model)
if args.enable_qat_fx:
import torch.quantization.quantize_fx as quantize_fx
qconfig_dict = {
"": torch.quantization.get_default_qat_qconfig('fbgemm')
}
model = quantize_fx.prepare_qat_fx(model, qconfig_dict)
logger.info("[Convert to quantized model]")
model = quantize_fx.convert_fx(model)
if args.enable_diffq:
quantizer = DiffQuantizer(model)
config['quantizer'] = quantizer
quantizer.restore_quantized_state(checkpoint)
else:
model.load_state_dict(checkpoint)
model = model.to(args.device)
'''
for name, param in model.named_parameters():
print(name, param.data, param.device, param.requires_grad)
'''
logger.info("[model] :\n{}".format(model.__str__()))
logger.info("[Model loaded]")
return model
def prepare_model(config, bert_model_name_or_path=None):
args = config['args']
emb_non_trainable = not args.embedding_trainable
labels = load_label(args.label_path)
label_size = len(labels)
config['labels'] = labels
# prepare model
if config['emb_class'] == 'glove':
if config['enc_class'] == 'gnb':
model = TextGloveGNB(config, args.embedding_path, label_size)
if config['enc_class'] == 'cnn':
model = TextGloveCNN(config,
args.embedding_path,
label_size,
emb_non_trainable=emb_non_trainable)
if config['enc_class'] == 'densenet-cnn':
model = TextGloveDensenetCNN(config,
args.embedding_path,
label_size,
emb_non_trainable=emb_non_trainable)
if config['enc_class'] == 'densenet-dsa':
model = TextGloveDensenetDSA(config,
args.embedding_path,
label_size,
emb_non_trainable=emb_non_trainable)
else:
model_name_or_path = args.bert_model_name_or_path
if bert_model_name_or_path:
model_name_or_path = bert_model_name_or_path
if config['emb_class'] == 'bart' and config['use_kobart']:
from transformers import BartModel
from kobart import get_kobart_tokenizer, get_pytorch_kobart_model
bert_tokenizer = get_kobart_tokenizer()
bert_tokenizer.cls_token = '<s>'
bert_tokenizer.sep_token = '</s>'
bert_tokenizer.pad_token = '<pad>'
bert_model = BartModel.from_pretrained(get_pytorch_kobart_model())
elif config['emb_class'] in ['gpt']:
bert_tokenizer = AutoTokenizer.from_pretrained(model_name_or_path)
bert_tokenizer.bos_token = '<|startoftext|>'
bert_tokenizer.eos_token = '<|endoftext|>'
bert_tokenizer.cls_token = '<|startoftext|>'
bert_tokenizer.sep_token = '<|endoftext|>'
bert_tokenizer.pad_token = '<|pad|>'
bert_model = AutoModel.from_pretrained(
model_name_or_path,
from_tf=bool(".ckpt" in model_name_or_path))
# 3 new tokens added
bert_model.resize_token_embeddings(len(bert_tokenizer))
elif config['emb_class'] in ['t5']:
from transformers import T5EncoderModel
bert_tokenizer = AutoTokenizer.from_pretrained(model_name_or_path)
bert_tokenizer.cls_token = '<s>'
bert_tokenizer.sep_token = '</s>'
bert_tokenizer.pad_token = '<pad>'
bert_model = T5EncoderModel.from_pretrained(
model_name_or_path,
from_tf=bool(".ckpt" in model_name_or_path))
else:
bert_tokenizer = AutoTokenizer.from_pretrained(model_name_or_path)
bert_model = AutoModel.from_pretrained(
model_name_or_path,
from_tf=bool(".ckpt" in model_name_or_path))
bert_config = bert_model.config
# bert model reduction
reduce_bert_model(config, bert_model, bert_config)
ModelClass = TextBertCNN
if config['enc_class'] == 'cls': ModelClass = TextBertCLS
if config['enc_class'] == 'densenet-cnn':
ModelClass = TextBertDensenetCNN
model = ModelClass(config,
bert_config,
bert_model,
bert_tokenizer,
label_size,
feature_based=args.bert_use_feature_based,
finetune_last=args.bert_use_finetune_last)
if args.restore_path:
checkpoint = load_checkpoint(args.restore_path)
model.load_state_dict(checkpoint)
if args.enable_qat:
model.qconfig = torch.quantization.get_default_qat_qconfig('fbgemm')
'''
# fuse if applicable
# model = torch.quantization.fuse_modules(model, [['']])
'''
model = torch.quantization.prepare_qat(model)
if args.enable_qat_fx:
import torch.quantization.quantize_fx as quantize_fx
model.train()
qconfig_dict = {
"": torch.quantization.get_default_qat_qconfig('fbgemm')
}
model = quantize_fx.prepare_qat_fx(model, qconfig_dict)
logger.info("[model] :\n{}".format(model.__str__()))
logger.info("[model prepared]")
return model
def prepare_model(config, bert_model_name_or_path=None):
opt = config['opt']
emb_non_trainable = not opt.embedding_trainable
labels = load_label(opt.label_path)
label_size = len(labels)
config['labels'] = labels
# prepare model
if config['emb_class'] == 'glove':
if config['enc_class'] == 'gnb':
model = TextGloveGNB(config, opt.embedding_path, label_size)
if config['enc_class'] == 'cnn':
model = TextGloveCNN(config,
opt.embedding_path,
label_size,
emb_non_trainable=emb_non_trainable)
if config['enc_class'] == 'densenet-cnn':
model = TextGloveDensenetCNN(config,
opt.embedding_path,
label_size,
emb_non_trainable=emb_non_trainable)
if config['enc_class'] == 'densenet-dsa':
model = TextGloveDensenetDSA(config,
opt.embedding_path,
label_size,
emb_non_trainable=emb_non_trainable)
else:
model_name_or_path = opt.bert_model_name_or_path
if bert_model_name_or_path:
model_name_or_path = bert_model_name_or_path
from transformers import AutoTokenizer, AutoConfig, AutoModel
bert_tokenizer = AutoTokenizer.from_pretrained(model_name_or_path)
bert_model = AutoModel.from_pretrained(
model_name_or_path, from_tf=bool(".ckpt" in model_name_or_path))
bert_config = bert_model.config
# bert model reduction
reduce_bert_model(config, bert_model, bert_config)
ModelClass = TextBertCNN
if config['enc_class'] == 'cls': ModelClass = TextBertCLS
model = ModelClass(config,
bert_config,
bert_model,
bert_tokenizer,
label_size,
feature_based=opt.bert_use_feature_based)
if opt.restore_path:
checkpoint = load_checkpoint(opt.restore_path, device=opt.device)
model.load_state_dict(checkpoint)
if opt.enable_qat:
model.qconfig = torch.quantization.get_default_qat_qconfig('fbgemm')
'''
# fuse if applicable
# model = torch.quantization.fuse_modules(model, [['']])
'''
model = torch.quantization.prepare_qat(model)
if opt.enable_qat_fx:
import torch.quantization.quantize_fx as quantize_fx
model.train()
qconfig_dict = {
"": torch.quantization.get_default_qat_qconfig('fbgemm')
}
model = quantize_fx.prepare_qat_fx(model, qconfig_dict)
model.to(opt.device)
logger.info("[model] :\n{}".format(model.__str__()))
logger.info("[model prepared]")
return model