def __init__(self, text_layer, image_layer, n_texts=100, n_images=10):
"""
Constructor
:param text_layer: Pre-trained text layer
:param image_layer: Pre-trained image layer
:param n_texts: Number of text in the inputs
:param n_images: Number of images in the inputs
"""
# Properties
self.text_layer = text_layer
self.image_layer = image_layer
self.n_texts = n_texts
self.n_images = n_images
# Number of features of the final fully connected layer of the text pre-trained network
text_num_ftrs = self.text_layer.fc.in_features
self.text_layer.fc = etnn.Identity()
# Number of features of the final fully connected layer of the image
image_num_ftrs = self.image_layer.fc.in_features
self.image_layer.fc = etnn.Identity()
# Linear layer
self.linear_size = text_num_ftrs * n_texts + image_num_ftrs * n_images
self.linear = nn.Linear(self.linear_size, 2)
def load_cgfs(fold=0, dataset_size=100, dataset_start=0, use_cuda=False):
"""
Load CGFS
:param fold:
:return:
"""
# Path
path = os.path.join("feature_selectors", "cgfs", "c3",
str(int(dataset_size)), str(int(dataset_start)),
"cgfs.{}.p".format(fold))
# print(path)
# CNN Glove Feature Selector
cgfs = models.CGFS(n_gram=3, n_features=settings.cgfs_output_dim['c3'])
if use_cuda:
cgfs.cuda()
# end if
# Eval
cgfs.eval()
# Load states
# print(u"Load {}".format(path))
state_dict = torch.load(open(path, 'rb'))
# Load dict
cgfs.load_state_dict(state_dict)
# Remove last linear layer
cgfs.linear2 = etnn.Identity()
cgfs.linear3 = etnn.Identity()
# Transformer
if use_cuda:
transformer = torchlanguage.transforms.Compose([
torchlanguage.transforms.GloveVector(),
torchlanguage.transforms.ToNGram(n=3, overlapse=True),
torchlanguage.transforms.Reshape(
(-1, 1, 3, settings.cgfs_input_dim)),
torchlanguage.transforms.ToCUDA(),
torchlanguage.transforms.FeatureSelector(
cgfs, settings.cgfs_output_dim['c3'], to_variable=True),
torchlanguage.transforms.Reshape(
(-1, settings.cgfs_output_dim['c3'])),
# torchlanguage.transforms.Normalize(mean=settings.cgfs_mean, std=settings.cgfs_std, input_dim=90)
])
else:
transformer = torchlanguage.transforms.Compose([
torchlanguage.transforms.GloveVector(),
torchlanguage.transforms.ToNGram(n=3, overlapse=True),
torchlanguage.transforms.Reshape(
(-1, 1, 3, settings.cgfs_input_dim)),
torchlanguage.transforms.FeatureSelector(
cgfs, settings.cgfs_output_dim['c3'], to_variable=True),
torchlanguage.transforms.Reshape(
(-1, settings.cgfs_output_dim['c3'])),
# torchlanguage.transforms.Normalize(mean=settings.cgfs_mean, std=settings.cgfs_std, input_dim=90)
])
# end if
return cgfs, transformer
def create_wenc_transformer(wenc_model, cgfs_info, use_cuda=True):
# Remove last linear layer
wenc_model.linear2 = etnn.Identity()
# Eval
wenc_model.eval()
# Transformer
if use_cuda:
transformer = torchlanguage.transforms.Compose([
torchlanguage.transforms.GloveVector(),
torchlanguage.transforms.ToNGram(n=3, overlapse=True),
torchlanguage.transforms.Reshape(
(-1, 1, 3, settings.cgfs_input_dim)),
torchlanguage.transforms.ToCUDA(),
torchlanguage.transforms.FeatureSelector(wenc_model,
settings.wenc_output_dim,
to_variable=True),
torchlanguage.transforms.Reshape((-1, settings.wenc_output_dim)),
])
else:
transformer = torchlanguage.transforms.Compose([
torchlanguage.transforms.GloveVector(),
torchlanguage.transforms.ToNGram(n=3, overlapse=True),
torchlanguage.transforms.Reshape(
(-1, 1, 3, settings.cgfs_input_dim)),
torchlanguage.transforms.FeatureSelector(wenc_model,
settings.wenc_output_dim,
to_variable=True),
torchlanguage.transforms.Reshape((-1, settings.wenc_output_dim)),
])
# end if
return transformer
def create_cenc_transformer(cenc_model, cenc_voc, use_cuda=True):
# Remove last linear layer
cenc_model.linear2 = etnn.Identity()
# Transformer
if use_cuda:
transformer = torchlanguage.transforms.Compose([
torchlanguage.transforms.Character(),
torchlanguage.transforms.ToIndex(token_to_ix=cenc_voc),
torchlanguage.transforms.ToNGram(n=settings.cenc_text_length,
overlapse=True),
torchlanguage.transforms.Reshape((-1, settings.cenc_text_length)),
torchlanguage.transforms.ToCUDA(),
torchlanguage.transforms.FeatureSelector(cenc_model,
settings.cenc_output_dim,
to_variable=True),
torchlanguage.transforms.Reshape((-1, settings.cenc_output_dim))
])
else:
transformer = torchlanguage.transforms.Compose([
torchlanguage.transforms.Character(),
torchlanguage.transforms.ToIndex(token_to_ix=cenc_voc),
torchlanguage.transforms.ToNGram(n=settings.cenc_text_length,
overlapse=True),
torchlanguage.transforms.Reshape((-1, settings.cenc_text_length)),
torchlanguage.transforms.FeatureSelector(cenc_model,
settings.cenc_output_dim,
to_variable=True),
torchlanguage.transforms.Reshape((-1, settings.cenc_output_dim))
])
# end if
return transformer
def __init__(self,
reservoir_dim,
sfa_dim,
ica_dim,
pretrained=False,
feature_selector='resnet18'):
"""
Constructor
"""
# Upper class
super(NilsNet, self).__init__()
# ResNet
if feature_selector == 'resnet18':
self.feature_selector = torchvision.models.resnet18(
pretrained=True)
elif feature_selector == 'resnet34':
self.feature_selector = torchvision.models.resnet34(
pretrained=True)
elif feature_selector == 'resnet50':
self.feature_selector = torchvision.models.resnet50(
pretrained=True)
elif feature_selector == 'alexnet':
self.feature_selector = torchvision.models.alexnet(pretrained=True)
# end if
# Skip last layer
self.reservoir_input_dim = self.feature_selector.fc.in_features
self.feature_selector.fc = ecnn.Identity()
def load_cenc35(use_cuda=False):
"""
Character encoder
:param fold:
:return:
"""
# Path
path = os.path.join("feature_selectors", "cenc", "cenc35.pth")
voc_path = os.path.join("feature_selectors", "cenc", "cenc35.voc.pth")
# print(path)
# print(voc_path)
# cEnc
cenc = torchlanguage.models.cEnc(text_length=settings.ccsaa_text_length,
vocab_size=settings.cenc_voc_size,
embedding_dim=settings.cenc_embedding_dim,
n_classes=settings.n_authors,
n_features=300)
if use_cuda:
cenc.cuda()
# end if
# Eval
cenc.eval()
# Load dict and voc
cenc.load_state_dict(torch.load(open(path, 'rb')))
voc = torch.load(open(voc_path, 'rb'))
# Remove last linear layer
cenc.linear2 = etnn.Identity()
# Transformer
if use_cuda:
transformer = torchlanguage.transforms.Compose([
torchlanguage.transforms.Character(),
torchlanguage.transforms.ToIndex(token_to_ix=voc),
torchlanguage.transforms.ToNGram(n=settings.cenc_text_length,
overlapse=False),
torchlanguage.transforms.Reshape((-1, settings.cenc_text_length)),
torchlanguage.transforms.ToCUDA(),
torchlanguage.transforms.FeatureSelector(cenc,
settings.cenc_output_dim,
to_variable=True),
torchlanguage.transforms.Reshape((-1, settings.cenc_output_dim))
])
else:
transformer = torchlanguage.transforms.Compose([
torchlanguage.transforms.Character(),
torchlanguage.transforms.ToIndex(token_to_ix=voc),
torchlanguage.transforms.ToNGram(n=settings.cenc_text_length,
overlapse=False),
torchlanguage.transforms.Reshape((-1, settings.cenc_text_length)),
torchlanguage.transforms.FeatureSelector(cenc,
settings.cenc_output_dim,
to_variable=True),
torchlanguage.transforms.Reshape((-1, settings.cenc_output_dim))
])
# end if
return cenc, transformer
def load_ccsaa35(use_cuda=False):
"""
Load CNN Character Selector for Authorship Attribution
:param fold:
:return:
"""
# Path
path = os.path.join("feature_selectors", "ccsaa", "ccsaa35.pth")
voc_path = os.path.join("feature_selectors", "ccsaa", "ccsaa35.voc.pth")
print(path)
print(voc_path)
# CNN Character Selector for Authorship Attribution
ccsaa = models.CCSAA(text_length=settings.ccsaa_text_length,
vocab_size=settings.ccsaa_pretrain_voc_size,
n_classes=settings.n_pretrain_authors,
n_features=settings.ccsaa_output_dim)
if use_cuda:
ccsaa.cuda()
# end if
# Eval
ccsaa.eval()
# Load dict and voc
ccsaa.load_state_dict(torch.load(open(path, 'rb')))
voc = torch.load(open(voc_path, 'rb'))
# Remove last linear layer
ccsaa.linear2 = etnn.Identity()
# Transformer
if use_cuda:
transformer = torchlanguage.transforms.Compose([
torchlanguage.transforms.Character(),
torchlanguage.transforms.ToIndex(token_to_ix=voc),
torchlanguage.transforms.ToNGram(n=settings.ccsaa_text_length,
overlapse=True),
torchlanguage.transforms.Reshape((-1, settings.ccsaa_text_length)),
torchlanguage.transforms.ToCUDA(),
torchlanguage.transforms.FeatureSelector(ccsaa,
settings.ccsaa_output_dim,
to_variable=True),
torchlanguage.transforms.Reshape((-1, settings.ccsaa_output_dim))
])
else:
transformer = torchlanguage.transforms.Compose([
torchlanguage.transforms.Character(),
torchlanguage.transforms.ToIndex(token_to_ix=voc),
torchlanguage.transforms.ToNGram(n=settings.ccsaa_text_length,
overlapse=True),
torchlanguage.transforms.Reshape((-1, settings.ccsaa_text_length)),
torchlanguage.transforms.FeatureSelector(ccsaa,
settings.ccsaa_output_dim,
to_variable=True),
torchlanguage.transforms.Reshape((-1, settings.ccsaa_output_dim))
])
# end if
return ccsaa, transformer
def load_cgfs35(use_cuda=False):
"""
Load CGFS
:param fold:
:return:
"""
# Path
path = os.path.join("feature_selectors", "cgfs", "c3", "cgfs35.p")
# print(path)
# CNN Glove Feature Selector
cgfs = models.CGFS(n_gram=3,
n_features=settings.cgfs_output_dim['c3'],
n_authors=settings.n_pretrain_authors)
if use_cuda:
cgfs.cuda()
# end if
# Eval
cgfs.eval()
# Load states
# print(u"Load {}".format(path))
state_dict = torch.load(open(path, 'rb'))
# Load dict
cgfs.load_state_dict(state_dict)
# Remove last linear layer
cgfs.linear3 = etnn.Identity()
# Transformer
if use_cuda:
transformer = torchlanguage.transforms.Compose([
torchlanguage.transforms.GloveVector(),
torchlanguage.transforms.ToNGram(n=3, overlapse=True),
torchlanguage.transforms.Reshape(
(-1, 1, 3, settings.cgfs_input_dim)),
torchlanguage.transforms.ToCUDA(),
torchlanguage.transforms.FeatureSelector(
cgfs, settings.cgfs_output_dim['c3'], to_variable=True),
torchlanguage.transforms.Reshape(
(-1, settings.cgfs_output_dim['c3'])),
])
else:
transformer = torchlanguage.transforms.Compose([
torchlanguage.transforms.GloveVector(),
torchlanguage.transforms.ToNGram(n=3, overlapse=True),
torchlanguage.transforms.Reshape(
(-1, 1, 3, settings.cgfs_input_dim)),
torchlanguage.transforms.FeatureSelector(
cgfs, settings.cgfs_output_dim['c3'], to_variable=True),
torchlanguage.transforms.Reshape(
(-1, settings.cgfs_output_dim['c3'])),
])
# end if
return cgfs, transformer
def load_wenc(fold=0, dataset_size=100, dataset_start=0, use_cuda=False):
"""
Load wEnc
:param fold:
:return:
"""
# Path
path = os.path.join("feature_selectors", "wenc", str(int(dataset_size)),
str(int(dataset_start)), "wenc.{}.p".format(fold))
# wEnc
wenc = torchlanguage.models.wEnc(n_classes=settings.n_authors, n_gram=3)
if use_cuda:
wenc.cuda()
# end if
# Eval
wenc.eval()
# Load states
# print(u"Load {}".format(path))
state_dict = torch.load(open(path, 'rb'))
# Load dict
wenc.load_state_dict(state_dict)
# Remove last linear layer
wenc.linear2 = etnn.Identity()
# Transformer
if use_cuda:
transformer = torchlanguage.transforms.Compose([
torchlanguage.transforms.GloveVector(),
torchlanguage.transforms.ToNGram(n=3, overlapse=True),
torchlanguage.transforms.Reshape(
(-1, 1, 3, settings.cgfs_input_dim)),
torchlanguage.transforms.ToCUDA(),
torchlanguage.transforms.FeatureSelector(wenc,
settings.wenc_output_dim,
to_variable=True),
torchlanguage.transforms.Reshape((-1, settings.wenc_output_dim)),
])
else:
transformer = torchlanguage.transforms.Compose([
torchlanguage.transforms.GloveVector(),
torchlanguage.transforms.ToNGram(n=3, overlapse=True),
torchlanguage.transforms.Reshape(
(-1, 1, 3, settings.cgfs_input_dim)),
torchlanguage.transforms.FeatureSelector(wenc,
settings.wenc_output_dim,
to_variable=True),
torchlanguage.transforms.Reshape((-1, settings.wenc_output_dim)),
])
# end if
return wenc, transformer
def __init__(self, image_model, n_images=10):
"""
Constructor
:param image_model: Single image model
:param n_images: Number of images
"""
# Super
super(ImageModel, self).__init__()
# Properties
self.image_model = image_model
self.n_images = n_images
# Number of features of the final fully connected layer of the image
image_num_ftrs = self.image_model.fc.in_features
self.image_model.fc = etnn.Identity()
# Linear layer
self.linear_size = image_num_ftrs * n_images
# self.linear1 = nn.Linear(self.linear_size, image_num_ftrs)
# self.linear2 = nn.Linear(image_num_ftrs, 2)
self.linear = nn.Linear(self.linear_size, 2)
def create_cgfs_transformer(cgfs_model, cgfs_info, use_cuda=True):
# Remove last linear layer
cgfs_model.linear3 = etnn.Identity()
# Eval
cgfs_model.eval()
# Transformer
if use_cuda:
transformer = torchlanguage.transforms.Compose([
torchlanguage.transforms.GloveVector(),
torchlanguage.transforms.ToNGram(n=3, overlapse=True),
torchlanguage.transforms.Reshape(
(-1, 1, 3, settings.cgfs_input_dim)),
torchlanguage.transforms.ToCUDA(),
torchlanguage.transforms.FeatureSelector(
cgfs_model, settings.cgfs_output_dim['c3'], to_variable=True),
torchlanguage.transforms.Reshape(
(-1, settings.cgfs_output_dim['c3'])),
# torchlanguage.transforms.Normalize(mean=cgfs_info['mean'], std=cgfs_info['std'], input_dim=90)
# torchlanguage.transforms.Normalize(mean=settings.cgfs_mean, std=settings.cgfs_std, input_dim=90)
])
else:
transformer = torchlanguage.transforms.Compose([
torchlanguage.transforms.GloveVector(),
torchlanguage.transforms.ToNGram(n=3, overlapse=True),
torchlanguage.transforms.Reshape(
(-1, 1, 3, settings.cgfs_input_dim)),
torchlanguage.transforms.FeatureSelector(
cgfs_model, settings.cgfs_output_dim['c3'], to_variable=True),
torchlanguage.transforms.Reshape(
(-1, settings.cgfs_output_dim['c3'])),
# torchlanguage.transforms.Normalize(mean=cgfs_info['mean'], std=cgfs_info['std'], input_dim=90)
# torchlanguage.transforms.Normalize(mean=settings.cgfs_mean, std=settings.cgfs_std, input_dim=90)
])
# end if
return transformer
input_sparsity = 0.1
w_sparsity = 0.1
input_scaling = 0.5
n_test = 5
n_samples = 1
# Argument
args = tools.functions.argument_parser_training_model()
# Load feature selector
feature_selector = torchlanguage.models.CCSAA(text_length=20,
vocab_size=84,
embedding_dim=50,
n_classes=15)
feature_selector.load_state_dict(torch.load(open(args.feature_selector, 'rb')))
feature_selector.linear = etnn.Identity()
if args.cuda:
feature_selector.cuda()
# end if
feature_selector_voc = torch.load(open(args.feature_selector_voc, 'rb'))
# Transforms
transform = transforms.Compose([
transforms.Character(),
transforms.ToIndex(token_to_ix=feature_selector_voc),
transforms.MaxIndex(max_id=83),
transforms.ToNGram(n=20, overlapse=True),
transforms.Reshape((-1, 20)),
transforms.ToCUDA(),
transforms.FeatureSelector(model=feature_selector,
n_features=150,