def get_initial_weights(self, model_type):
tf.reset_default_graph()
if model_type == "perceptron":
m = Perceptron()
inputs = tf.placeholder(tf.float32, shape=(None, 28 * 28))
_ = m.get_model(features={"x": inputs},
labels=None,
mode='predict',
params=None)
elif model_type == 'cnn-mnist':
m = CNN()
inputs = tf.placeholder(tf.float32, shape=(None, 28, 28, 1))
_ = m.get_model(features={"x": inputs},
labels=None,
mode='predict',
params=None)
elif model_type == 'cnn-cifar10':
m = CNN()
inputs = tf.placeholder(tf.float32, shape=(None, 32, 32, 3))
_ = m.get_model(features={"x": inputs},
labels=None,
mode='predict',
params=None)
else:
raise ValueError(
"Model {model_type} not supported.".format(model_type))
with tf.Session().as_default() as sess:
sess.run(tf.global_variables_initializer())
collection = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)
weights = {tensor.name: sess.run(tensor) for tensor in collection}
tf.reset_default_graph()
return weights
def test(args):
np.random.seed(1234)
train, _, test = load_dataset(args.data_dir, valid=args.validation, dataset_seed=args.dataset_seed)
print("N_train:{}".format(train.N))
enc = CNN(n_outputs=args.n_categories, dropout_rate=args.dropout_rate)
chainer.serializers.load_npz(args.trained_model_path, enc)
if args.gpu > -1:
chainer.cuda.get_device(args.gpu).use()
enc.to_gpu()
print("Finetune")
for i in range(args.finetune_iter):
x,_ = train.get(args.batchsize_finetune, gpu=args.gpu)
enc(x)
acc_sum = 0
test_x, test_t = test.get()
N_test = test.N
for i in range(0, N_test, args.batchsize_eval):
x = test_x[i:i + args.batchsize_eval]
t = test_t[i:i + args.batchsize_eval]
if args.gpu > -1:
x, t = cuda.to_gpu(x, device=args.gpu), cuda.to_gpu(t, device=args.gpu)
logit = enc(x, train=False)
acc = F.accuracy(logit, t).data
acc_sum += acc * x.shape[0]
acc_test = acc_sum / N_test
print("test acc: ", acc_test)
def get_predictions(self, frames, scope):
frames = self._reshape_to_conv(frames)
cnn = CNN()
if self.operation == 'training':
cnn_output = cnn.create_model(frames,
cnn.conv_filters,
keep_prob=self.keep_prob)
else:
cnn_output = cnn.create_model(frames,
cnn.conv_filters,
keep_prob=1.0)
cnn_output = self._reshape_to_rnn(cnn_output)
rnn = RNN()
rnn_output = rnn.create_model(cnn_output, scope + '_rnn')
if self.is_attention:
attention = Attention(self.batch_size)
attention_output = attention.create_model(rnn_output,
scope + '_attention')
fc = FC(self.num_classes)
outputs = fc.create_model(attention_output, scope + '_fc')
else:
rnn_output = rnn_output[:, -1, :]
fc = FC(self.num_classes)
outputs = fc.create_model(rnn_output, scope + '_fc')
return outputs
def GetDefaultCNN():
cnn = CNN( in_features=(32,32,3),
out_features=10,
conv_filters=[32,32,64,64],
conv_kernel_size=[3,3,3,3],
conv_strides=[1,1,1,1],
conv_pad=[0,0,0,0],
max_pool_kernels=[None, (2,2), None, (2,2)],
max_pool_strides=[None,2,None,2],
use_dropout=False,
use_batch_norm=False,
actv_func=[None, "relu", None, "relu"],
device=device
)
# Create MLP
# Calculate the input shape
s = cnn.GetCurShape()
in_features = s[0]*s[1]*s[2]
mlp = MLP(in_features,
10,
[],
[],
use_batch_norm=False,
use_dropout=False,
use_softmax=False,
device=device)
# mlp = DefaultCifar10MLP(device=device, in_features=in_features)
cnn.AddMLP(mlp)
return cnn
def setup_model(self, model_type):
self.model_type = model_type
if model_type == "perceptron":
self.model = Perceptron()
elif model_type == "cnn-mnist":
self.model = CNN()
elif model_type == "cnn-cifar10":
self.model = CNN()
else:
raise ValueError("Model {0} not supported.".format(model_type))
def __init__(self, config, pad_idx, train_iter=None, valid_iter=None, test_iter=None):
self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
self.config = config
self.pad_idx = pad_idx
# Train mode
if self.config.mode == 'train':
self.train_iter = train_iter
self.valid_iter = valid_iter
# Test mode
else:
self.test_iter = test_iter
model_type = {
'vanilla_rnn': RNN(self.config, self.pad_idx),
'bidirectional_lstm': BidirectionalLSTM(self.config, self.pad_idx),
'cnn': CNN(self.config),
}
self.model = model_type[self.config.model]
self.model.to(self.device)
# SGD updates all parameters with the 'same' learning rate
# Adam adapts learning rate for each parameter
optim_type = {
'SGD': optim.SGD(self.model.parameters(), lr=self.config.lr),
'Adam': optim.Adam(self.model.parameters()),
}
self.optimizer = optim_type[self.config.optim]
# BCEWithLogitsLoss carries out both the sigmoid and the binary cross entropy steps.
self.criterion = nn.BCEWithLogitsLoss()
self.criterion.to(self.device)
def main(args):
# Get dataset
datasource = DataSource(None, [args.input_image],
shape=tuple(args.eye_shape),
data_format=args.data_format,
heatmap_scale=args.heatmap_scale)
# Get model
learning_schedule = [{
'loss_terms_to_optimize': {
'heatmaps_mse': ['hourglass'],
'radius_mse': ['radius'],
},
'learning_rate': 1e-3,
}]
model = CNN(datasource.tensors,
datasource.x_shape,
learning_schedule,
data_format=args.data_format)
# Get evaluator
evaluator = Trainer(model, model_checkpoint=args.model_checkpoint)
# Predict
output, losses = evaluator.run_predict(datasource)
input_data = util.load_pickle(args.input_image)
print('Losses', losses)
util.plot_predictions(output, input_data, tuple(args.eye_shape))
def __init__(self, args, device, rel2id, word_emb=None):
lr = args.lr
lr_decay = args.lr_decay
self.cpu = torch.device('cpu')
self.device = device
self.args = args
self.max_grad_norm = args.max_grad_norm
if args.model == 'pa_lstm':
self.model = PositionAwareLSTM(args, rel2id, word_emb)
elif args.model == 'bgru':
self.model = BGRU(args, rel2id, word_emb)
elif args.model == 'cnn':
self.model = CNN(args, rel2id, word_emb)
elif args.model == 'pcnn':
self.model = PCNN(args, rel2id, word_emb)
elif args.model == 'lstm':
self.model = LSTM(args, rel2id, word_emb)
else:
raise ValueError
self.model.to(device)
self.criterion = nn.CrossEntropyLoss()
self.parameters = [
p for p in self.model.parameters() if p.requires_grad
]
# self.parameters = self.model.parameters()
self.optimizer = torch.optim.SGD(self.parameters, lr)
def __init__(self, args, device, rel2id, word_emb=None):
lr = args.lr
lr_decay = args.lr_decay
self.cpu = torch.device('cpu')
self.device = device
self.args = args
self.rel2id = rel2id
self.max_grad_norm = args.max_grad_norm
if args.model == 'pa_lstm':
self.model = PositionAwareRNN(args, rel2id, word_emb)
elif args.model == 'bgru':
self.model = BGRU(args, rel2id, word_emb)
elif args.model == 'cnn':
self.model = CNN(args, rel2id, word_emb)
elif args.model == 'pcnn':
self.model = PCNN(args, rel2id, word_emb)
elif args.model == 'lstm':
self.model = LSTM(args, rel2id, word_emb)
else:
raise ValueError
self.model.to(device)
self.criterion = nn.CrossEntropyLoss()
if args.fix_bias:
self.model.flinear.bias.requires_grad = False
self.parameters = [
p for p in self.model.parameters() if p.requires_grad
]
# self.parameters = self.model.parameters()
self.optimizer = torch.optim.SGD(self.parameters, lr)
self.scheduler = lr_scheduler.ReduceLROnPlateau(self.optimizer,
'min',
patience=3,
factor=lr_decay)
def main(args):
# Get dataset
test_files = glob.glob(os.path.join(args.test_path, '*.pickle'))
datasource = DataSource(None,
test_files,
shape=tuple(args.eye_shape),
data_format=args.data_format,
heatmap_scale=args.heatmap_scale)
# Get model
learning_schedule = [{
'loss_terms_to_optimize': {
'heatmaps_mse': ['hourglass'],
'radius_mse': ['radius'],
},
'learning_rate': 1e-3,
}]
model = CNN(datasource.tensors, datasource.x_shape, learning_schedule)
# Get evaluator
evaluator = Trainer(model, model_checkpoint=args.model_checkpoint)
# Evaluate
avg_losses = evaluator.run_eval(datasource)
print('Avarage Losses', avg_losses)
def setup(self):
# Load a Tensorflow model into memory.
# If needed froze the graph to get better performance.
data_format = self.args.data_format
shape = tuple(self.args.eye_shape)
preprocessor = ImgPreprocessor(data_format)
datasource = ImgDataSource(shape=shape,
data_format=data_format)
# Get model
model = CNN(datasource.tensors, datasource.x_shape, None,
data_format=data_format, predict_only=True)
# Start session
saver = tf.train.Saver()
sess = tf.Session()
# Init variables
init = tf.global_variables_initializer()
init_l = tf.local_variables_initializer()
sess.run(init)
sess.run(init_l)
# Restore model checkpoint
saver.restore(sess, self.args.model_checkpoint)
return datasource, preprocessor, sess, model
def main(args):
# Get dataset
train_files = glob.glob(os.path.join(args.train_path, '*.pickle'))
eval_files = glob.glob(os.path.join(args.eval_path, '*.pickle'))
datasource = DataSource(train_files,
eval_files,
shape=tuple(args.eye_shape),
batch_size=args.batch_size,
data_format=args.data_format,
heatmap_scale=args.heatmap_scale)
# Get model
learning_schedule = [
{
'loss_terms_to_optimize': {
'heatmaps_mse': ['hourglass'],
'radius_mse': ['radius'],
},
'learning_rate': 1e-3,
},
]
model = CNN(datasource.tensors, datasource.x_shape, learning_schedule)
# Get trainer
trainer = Trainer(model, eval_steps=args.eval_steps)
# Train for 10000 steps
return trainer.run_training(datasource, args.steps)
def multi_get_attention(self, frames):
frames = self._reshape_to_conv(frames)
cnn = CNN()
cnn_output = cnn.create_model(frames, cnn.conv_filters)
cnn_output = self._reshape_to_rnn(cnn_output)
rnn = RNN()
rnn_output = rnn.create_model(cnn_output)
if self.is_attention:
attention = Attention(self.batch_size)
attention_output = attention.attention_analysis(rnn_output)
return attention_output
else:
rnn_output = rnn_output[:, -1, :]
fc = FC(self.num_classes)
outputs = fc.create_model(rnn_output)
return outputs
def main(args):
model = CNN()
chainer.serializers.load_npz(args.init_weights, model)
model = chainer.links.Classifier(model)
paths = glob.glob(os.path.join(args.dataset_dir, '**/*.wav'))
testset = ESDataset(paths, label_index=args.label_index)
y_targs = []
y_preds = []
for data in testset:
x, y = data
y_targs.append(int(y))
y_preds.append(int(predict(model, x)))
class_names = ['中立', '穏やか', '幸せ', '悲しみ', '怒り', '恐怖', '嫌悪', '驚き']
accuracy = accuracy_score(y_targs, y_preds)
plot_confusion_matrix(y_targs,
y_preds,
classes=class_names,
normalize=True,
title='{0} 精度:{1:.1%}'.format(args.title, accuracy))
plt.savefig(os.path.join(args.out_dir, 'confusion_matrix.png'))
class CNNPredictor:
def __init__(self, name= "WikiContrvCNN", input_name=None):
if input_name is None:
input_name = name
self.hp = hyperparams.HPCNN()
self.sess = init_session()
self.sess.run(tf.global_variables_initializer())
self.dropout_prob = tf.placeholder(tf.float32, name="dropout_keep_prob")
self.seq_max = self.hp.seq_max
self.word2idx = cache.load_cache(input_name+".voca")
init_emb = cache.load_cache("init_emb_word2vec")
self.model = CNN("controv", self.seq_max, 2, [2, 3, 4], 128,
init_emb, self.hp.embedding_size, self.dropout_prob)
self.input_text = tf.placeholder(tf.int32,
shape=[None, self.seq_max],
name="comment_input")
self.sout = self.model.network(self.input_text)
self.tokenize = lambda x: tokenize(x, set(), False)
loader = tf.train.Saver()
loader.restore(self.sess, cpath.get_model_full_path(name))
def encode(self, docs):
def convert(word):
if word in self.word2idx:
return self.word2idx[word]
else:
return OOV
data = []
for doc in docs:
entry = []
for token in self.tokenize(doc):
entry.append(convert(token))
entry = entry[:self.seq_max]
while len(entry) < self.seq_max:
entry.append(PADDING)
data.append((entry, 0))
return data
def score(self, docs):
inputs = self.encode(docs)
def forward_run(inputs):
batches = get_batches_ex(inputs, self.hp.batch_size, 2)
logit_list = []
for batch in batches:
x, y, = batch
logits, = self.sess.run([self.sout, ],
feed_dict={
self.input_text: x,
self.dropout_prob: 1.0,
})
logit_list.append(logits)
return np.concatenate(logit_list)
output = forward_run(inputs)[:,1]
return output
def get_model(args):
mean = np.array([640., 476.23620605, 88.2875590389])
std = np.array([227.59802246, 65.00177002, 52.7303319245])
if "scale" not in args.model:
mean, std = mean[:2], std[:2]
logger.info("Mean: {}, std: {}".format(mean, std))
if args.model == "cnn" or args.model == "cnn_scale":
model = CNN(mean, std, args.gpu, args.channel_list, args.deconv_list, args.ksize_list,
args.dc_ksize_list, args.inter_list, args.last_list, args.pad_list)
elif args.model == "cnn_pose" or args.model == "cnn_pose_scale":
model = CNN_Pose(mean, std, args.gpu, args.channel_list, args.deconv_list, args.ksize_list,
args.dc_ksize_list, args.inter_list, args.last_list, args.pad_list)
elif args.model == "cnn_ego" or args.model == "cnn_ego_scale":
model = CNN_Ego(mean, std, args.gpu, args.channel_list, args.deconv_list, args.ksize_list,
args.dc_ksize_list, args.inter_list, args.last_list, args.pad_list, args.ego_type)
elif args.model == "cnn_ego_pose" or args.model == "cnn_ego_pose_scale":
model = CNN_Ego_Pose(mean, std, args.gpu, args.channel_list, args.deconv_list, args.ksize_list,
args.dc_ksize_list, args.inter_list, args.last_list, args.pad_list, args.ego_type)
else:
logger.info("Invalid argument: model={}".format(args.model))
exit(1)
if args.resume != "":
serializers.load_npz(args.resume, model)
if args.gpu >= 0:
model.to_gpu(args.gpu)
return model
def models(m):
if m == 'rnn':
return RNN(1, opt.hidden_size, opt.num_layers, 1, opt.cuda)
elif m == 'lstm':
return LSTM(1, opt.hidden_size, opt.num_layers, 1, opt.cuda)
elif m == 'qrnn':
return QRNN(1, opt.hidden_size, opt.num_layers, 1, opt.cuda)
elif m == 'cnn':
return CNN(1, opt.hidden_size, 1, opt.cuda)
def get_multi_predictions(self, frames):
frames = self._reshape_to_conv(frames)
cnn = CNN()
if self.operation == 'training':
cnn_output = cnn.create_model(frames,
cnn.conv_filters,
keep_prob=self.keep_prob)
else:
cnn_output = cnn.create_model(frames,
cnn.conv_filters,
keep_prob=1.0)
cnn_output = self._reshape_to_rnn(cnn_output)
rnn = RNN()
arousal_rnn_output = rnn.create_model(cnn_output, 'arousal_rnn')
valence_rnn_output = rnn.create_model(cnn_output, 'valence_rnn')
dominance_rnn_output = rnn.create_model(cnn_output, 'dominance_rnn')
if self.is_attention:
attention = Attention(self.batch_size)
arousal_attention_output = attention.create_model(
arousal_rnn_output, 'arousal_attention')
valence_attention_output = attention.create_model(
valence_rnn_output, 'valence_attention')
dominance_attention_output = attention.create_model(
dominance_rnn_output, 'dominance_attention')
fc = FC(self.num_classes)
arousal_fc_outputs = fc.create_model(arousal_attention_output,
'arousal_fc')
valence_fc_outputs = fc.create_model(valence_attention_output,
'valence_fc')
dominance_fc_outputs = fc.create_model(dominance_attention_output,
'dominance_fc')
else:
arousal_rnn_output = arousal_rnn_output[:, -1, :]
valence_rnn_output = valence_rnn_output[:, -1, :]
dominance_rnn_output = dominance_rnn_output[:, -1, :]
fc = FC(self.num_classes)
arousal_fc_outputs = fc.create_model(arousal_rnn_output,
'arousal_fc')
valence_fc_outputs = fc.create_model(valence_rnn_output,
'valence_fc')
dominance_fc_outputs = fc.create_model(dominance_rnn_output,
'dominance_fc')
return arousal_fc_outputs, valence_fc_outputs, dominance_fc_outputs
def __init__(self, filed=80):
super(TD_LSTM, self).__init__()
self.filed = filed
self.cnn_l = CNN(self.filed)
self.rnn_l = nn.LSTM(input_size=40,
hidden_size=64,
num_layers=4,
batch_first=True)
self.cnn_r = CNN(self.filed)
self.rnn_r = nn.LSTM(input_size=40,
hidden_size=64,
num_layers=4,
batch_first=True)
self.linear = nn.Sequential(
nn.Linear(128, 64),
nn.Linear(64, 2),
)
def __init__(self, name= "WikiContrvCNN", input_name=None):
if input_name is None:
input_name = name
self.hp = hyperparams.HPCNN()
self.sess = init_session()
self.sess.run(tf.global_variables_initializer())
self.dropout_prob = tf.placeholder(tf.float32, name="dropout_keep_prob")
self.seq_max = self.hp.seq_max
self.word2idx = cache.load_cache(input_name+".voca")
init_emb = cache.load_cache("init_emb_word2vec")
self.model = CNN("controv", self.seq_max, 2, [2, 3, 4], 128,
init_emb, self.hp.embedding_size, self.dropout_prob)
self.input_text = tf.placeholder(tf.int32,
shape=[None, self.seq_max],
name="comment_input")
self.sout = self.model.network(self.input_text)
self.tokenize = lambda x: tokenize(x, set(), False)
loader = tf.train.Saver()
loader.restore(self.sess, cpath.get_model_full_path(name))
def __init__(self, filed=80):
super(MemNet, self).__init__()
self.filed = filed
self.cnn_l = CNN(filed=self.filed)
self.attention = Attention(40, score_function='mlp')
self.x_linear = nn.Sequential(nn.Linear(40, 40), )
self.linear = nn.Sequential(
nn.Linear(40, 64),
nn.Linear(64, 2),
)
def get_predictor():
dropout_keep_prob = tf.placeholder(tf.float32, name="dropout_keep_prob")
cnn = CNN("agree",
sequence_length=FLAGS.comment_length,
num_classes=3,
filter_sizes=[1, 2, 3],
num_filters=64,
init_emb=load_local_pickle("init_embedding"),
embedding_size=FLAGS.embedding_size,
dropout_prob=dropout_keep_prob)
input_comment = tf.placeholder(tf.int32,
shape=[None, FLAGS.comment_length],
name="comment_input")
#sout = model.cnn.network(input_comment)
sout = cnn.network(input_comment)
sess = init_session()
batch_size = 512
path = os.path.join(model_path, "runs", "agree", "model-36570")
variables = tf.contrib.slim.get_variables_to_restore()
for v in variables:
print(v.name)
loader = tf.train.Saver(variables)
loader.restore(sess, path)
def predict(comments):
batches = get_batches_ex(comments, batch_size, 1)
all_scores = []
ticker = TimeEstimator(len(batches))
for batch in batches:
scores, = sess.run([sout],
feed_dict={
input_comment: batch[0],
dropout_keep_prob: 1.0,
})
all_scores.append(scores)
ticker.tick()
return np.concatenate(all_scores)
return predict
def __init__(self, prior, init_emb):
self.comment_length = FLAGS.comment_length
self.comment_count = FLAGS.comment_count
self.embedding_size = FLAGS.embedding_size
self.prior = prior
self.dropout_keep_prob = tf.placeholder(tf.float32,
name="dropout_keep_prob")
self.input_comment = tf.placeholder(
tf.int32,
shape=[None, self.comment_count, self.comment_length],
name="input_reaction")
self.input_comment_y = tf.placeholder(
tf.int32, shape=[None, self.comment_count],
name="input_y_comment") # agree label for comments
self.input_y = tf.placeholder(tf.int32, shape=[
None,
], name="input_y") # Controversy Label
self.cnn = CNN("agree",
sequence_length=self.comment_length,
num_classes=3,
filter_sizes=[1, 2, 3],
num_filters=64,
init_emb=init_emb,
embedding_size=self.embedding_size,
dropout_prob=self.dropout_keep_prob)
self.score = self.controversy(self.input_comment)
self.acc = accuracy(self.score, self.input_y, axis=1)
self.agree_logit = self.predict_2d(self.input_comment)
self.agree_acc = accuracy(self.agree_logit,
self.input_comment_y,
axis=2)
self.agree_loss = tf.reduce_mean(
tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=self.agree_logit, labels=self.input_comment_y))
def __init__(self, args):
super(CNN_MLP, self).__init__()
self.args = args
self.device = args["device"]
self.batch_size = args["batch_size"]
self.input_dim = args["input_dim"]
self.obs_len = args["input_length"]
self.kernel_size = args["kernel_size"]
self.nb_conv = args["nb_conv"]
self.nb_kernel = args["nb_kernel"]
self.cnn_feat_size = args["cnn_feat_size"]
self.mlp_layers = args["mlp_layers"]
self.output_size = args["output_size"]
self.cnn = CNN(num_inputs=self.input_dim,
nb_kernel=self.nb_kernel,
cnn_feat_size=self.cnn_feat_size,
obs_len=self.obs_len,
kernel_size=self.kernel_size,
nb_conv=self.nb_conv)
# self.cnn = nn.Sequential()
# padding = int((self.kernel_size-1)/2.0)
# for i in range(self.nb_conv):
# conv = nn.Conv1d(self.nb_kernel , self.nb_kernel , self.kernel_size, padding=padding)
# if i == 0:
# conv = nn.Conv1d(self.coord_embedding_size, self.nb_kernel , self.kernel_size, padding=padding)
# self.cnn.add_module("conv0",conv)
# self.project_cnn = nn.Linear(self.obs_len*self.nb_kernel,self.cnn_feat_size)
self.mlp = nn.Sequential()
self.mlp.add_module("layer0",
nn.Linear(self.cnn_feat_size, self.mlp_layers[0]))
self.mlp.add_module("relu0", nn.ReLU())
for i in range(1, len(self.mlp_layers)):
self.mlp.add_module(
"layer{}".format(i),
nn.Linear(self.mlp_layers[i - 1], self.mlp_layers[i]))
self.mlp.add_module("relu{}".format(i), nn.ReLU())
self.mlp.add_module("layer{}".format(len(self.mlp_layers)),
nn.Linear(self.mlp_layers[-1], self.output_size))
def __init__(self, filed=80):
super(ATAE_LSTM, self).__init__()
self.filed = filed
self.cnn_l = CNN(filed=self.filed)
self.rnn_l = nn.LSTM(input_size=110,
hidden_size=64,
num_layers=4,
batch_first=True)
self.attention = NoQueryAttention(128, score_function='bi_linear')
self.linear = nn.Sequential(
nn.Linear(64, 64),
nn.Linear(64, 2),
)
def GetCNN():
cnn = CNN(
in_features=(32, 32, 3),
out_features=10,
conv_filters=[32, 32, 64, 64],
conv_kernel_size=[3, 3, 3, 3],
conv_strides=[1, 1, 1, 1],
conv_pad=[0, 0, 0, 0],
max_pool_kernels=[None, (2, 2), None, (2, 2)],
max_pool_strides=[None, 2, None, 2],
use_dropout=False,
use_batch_norm=True, #False
actv_func=["relu", "relu", "relu", "relu"],
device=device)
return cnn
def setup_model(self, model_type):
self.model_type = model_type
if model_type == "perceptron":
self.model = Perceptron()
self.weights_metadata = self.model.get_weights_shape()
elif model_type == "cnn":
#TODO: Support CNN
self.model = CNN()
elif model_type == "lstm":
#TODO: Support LSTM
self.model = LSTM()
elif model_type == "gan":
self.model = ConversationalNetwork()
self.model.build_model(is_retraining=True)
else:
raise ValueError("Model {0} not supported.".format(model_type))
def __init__(self, filed = 80):
super(IAN_LSTM, self).__init__()
self.filed = filed
self.cnn_l = CNN(filed=self.filed)
self.rnn_l = nn.LSTM(
input_size=55,
hidden_size=64,
num_layers=4,
batch_first=True)
self.attention_aspect = Attention(64, score_function='bi_linear')
self.attention_context = Attention(64, score_function='bi_linear')
self.linear = nn.Sequential(
nn.Linear(128, 64),
nn.Linear(64, 2),
)
def get_model(args, parallel=True, ckpt_path=False):
if args.clf == 'fcn':
print('Initializing FCN...')
model = FCN(args.input_size, args.output_size)
elif args.clf == 'mlp':
print('Initializing MLP...')
model = MLP(args.input_size, args.output_size)
elif args.clf == 'svm':
print('Initializing SVM...')
model = SVM(args.input_size, args.output_size)
elif args.clf == 'cnn':
print('Initializing CNN...')
model = CNN(nc=args.num_channels, fs=args.cnn_view)
elif args.clf == 'resnet18':
print('Initializing ResNet18...')
model = resnet.resnet18(num_channels=args.num_channels,
num_classes=args.output_size)
elif args.clf == 'vgg19':
print('Initializing VGG19...')
model = VGG(vgg_name=args.clf,
num_channels=args.num_channels,
num_classes=args.output_size)
elif args.clf == 'unet':
print('Initializing UNet...')
model = UNet(in_channels=args.num_channels,
out_channels=args.output_size)
num_params, num_layers = get_model_size(model)
print("# params: {}\n# layers: {}".format(num_params, num_layers))
if ckpt_path:
model.load_state_dict(torch.load(ckpt_path))
print('Load init: {}'.format(ckpt_path))
if parallel:
model = nn.DataParallel(model.to(get_device(args)),
device_ids=args.device_id)
else:
model = model.to(get_device(args))
loss_type = 'hinge' if args.clf == 'svm' else args.loss_type
print("Loss: {}".format(loss_type))
return model, loss_type
def predict_cnn(config):
# load tokenizer and torchtext Field
pickle_tokenizer = open('pickles/tokenizer.pickle', 'rb')
cohesion_scores = pickle.load(pickle_tokenizer)
tokenizer = LTokenizer(scores=cohesion_scores)
pickle_vocab = open('pickles/text.pickle', 'rb')
text = pickle.load(pickle_vocab)
model = CNN(config)
model.load_state_dict(torch.load(config.save_model))
model.to(device)
model.eval()
tokenized = tokenizer.tokenize(config.input)
min_len = config.filter_sizes[-1]
# if user's input sentence is shorter than the largest filter size, add pad tokens to input sentence
if len(tokenized) < min_len:
tokenized += ['<pad>'] * (min_len - len(tokenized))
indexed = [text.vocab.stoi[token] for token in tokenized]
length = [len(indexed)]
tensor = torch.LongTensor(indexed).to(device)
tensor = tensor.unsqueeze(1)
length_tensor = torch.LongTensor(length)
prediction = torch.sigmoid(model(tensor, length_tensor))
label = torch.round(prediction)
if label == 1:
label = 'Positive'
else:
label = 'Negative'
sentiment_percent = prediction.item()
print(f'[in] >> {config.input}')
print(f'[out] >> {sentiment_percent*100:.2f} % : {label}')
train_dataset, train_dataset_args = read_data_args(train_data_spec)
valid_dataset, valid_dataset_args = read_data_args(valid_data_spec)
# reading data
train_sets, train_xy, train_x, train_y = read_dataset(train_dataset, train_dataset_args)
valid_sets, valid_xy, valid_x, valid_y = read_dataset(valid_dataset, valid_dataset_args)
numpy_rng = numpy.random.RandomState(89677)
theano_rng = RandomStreams(numpy_rng.randint(2 ** 30))
log('> ... building the model')
# construct the cnn architecture
cnn = CNN(numpy_rng=numpy_rng, theano_rng = theano_rng,
batch_size = batch_size, n_outs=n_outs,
conv_layer_configs = conv_layer_configs,
hidden_layers_sizes = hidden_layers_sizes,
conv_activation = conv_activation,
full_activation = full_activation,
use_fast = use_fast, update_layers = update_layers)
total_layer_number = len(cnn.layers)
if full_ptr_layer_number > 0:
_file2nnet(cnn.layers[len(conv_layer_configs):total_layer_number], set_layer_num = full_ptr_layer_number, filename = full_ptr_file, withfinal=False)
if conv_ptr_layer_number > 0:
_file2cnn(cnn.layers[0:len(conv_layer_configs)], filename=conv_ptr_file)
# get the training, validation and testing function for the model
log('> ... getting the finetuning functions')
train_fn, valid_fn = cnn.build_finetune_functions(
(train_x, train_y), (valid_x, valid_y),
batch_size=batch_size)
if arguments.has_key('ptr_file') and arguments.has_key('ptr_layer_number'):
ptr_file = arguments['ptr_file']
ptr_layer_number = int(arguments['ptr_layer_number'])
# check working dir to see whether it's resuming training
resume_training = False
if os.path.exists(wdir + '/nnet.tmp') and os.path.exists(wdir + '/training_state.tmp'):
resume_training = True
cfg.lrate = _file2lrate(wdir + '/training_state.tmp')
log('> ... found nnet.tmp and training_state.tmp, now resume training from epoch ' + str(cfg.lrate.epoch))
numpy_rng = numpy.random.RandomState(89677)
theano_rng = RandomStreams(numpy_rng.randint(2 ** 30))
log('> ... initializing the model')
# construct the cnn architecture
cnn = CNN(numpy_rng=numpy_rng, theano_rng = theano_rng, cfg = cfg)
# load the pre-training networks, if any, for parameter initialization
if (ptr_layer_number > 0) and (resume_training is False):
_file2nnet(cnn.layers, set_layer_num = ptr_layer_number, filename = ptr_file)
if resume_training:
_file2nnet(cnn.layers, filename = wdir + '/nnet.tmp')
# get the training, validation and testing function for the model
log('> ... getting the finetuning functions')
train_fn, valid_fn = cnn.build_finetune_functions(
(cfg.train_x, cfg.train_y), (cfg.valid_x, cfg.valid_y),
batch_size=cfg.batch_size)
log('> ... finetunning the model')
while (cfg.lrate.get_rate() != 0):
# one epoch of sgd training
