def SentimentNet(word_to_idx):
"""Construct a RNN model for sentiment analysis
# Arguments:
word_to_idx: A dictionary giving the vocabulary. It contains V entries,
and maps each string to a unique integer in the range [0, V).
# Returns
model: the constructed model
"""
vocab_size = len(word_to_idx)
model = Model()
model.add(
FCLayer(vocab_size,
200,
name='embedding',
initializer=Gaussian(std=0.01)))
model.add(
BidirectionalRNN(in_features=200,
units=50,
initializer=Gaussian(std=0.01)))
model.add(FCLayer(100, 32, name='fclayer1',
initializer=Gaussian(std=0.01)))
model.add(TemporalPooling()) # defined in layers.py
model.add(FCLayer(32, 2, name='fclayer2', initializer=Gaussian(std=0.01)))
return model
def MyModel_SentimentNet(word_to_idx):
vocab_size = len(word_to_idx)
model = Model()
model.add(Linear2D(vocab_size, 200, name='embedding', initializer=Gaussian(std=0.01)))
model.add(BiRNN(in_features=200, units=50, initializer=Gaussian(std=0.01)))
model.add(Linear2D(100, 50, name='linear1', initializer=Gaussian(std=0.01)))
model.add(TemporalPooling()) # defined in layers.py
model.add(Linear2D(50, 2, name='linear2', initializer=Gaussian(std=0.01)))
# model.add(Linear2D(vocab_size, 200, name='embedding', initializer=Gaussian(std=0.01)))
# model.add(GRU(in_features=200, units=50, initializer=Gaussian(std=0.01)))
# model.add(Linear2D(50, 50, name='embedding', initializer=Gaussian(std=0.01)))
# model.add(GRU(in_features=200, units=50, initializer=Gaussian(std=0.01)))
# model.add(TemporalPooling())
# model.add(Linear(200,2,initializer=Gaussian(std=0.01)))
return model
def load_model(config, n_classes=2):
model = Model()
X = model.add(input([config.SEQUENCE_LEN], dtype='int32', name="input"))
if config.is_use_embedding():
embedding = model.add(
embeddings(X,
config.WORD_COUNT,
config.EMBEDDING_DIM,
weights=config.EMBEDDING_MATRIX,
input_length=config.SEQUENCE_LEN,
frozen=config.is_embedding_trainable()))
else:
embedding = model.add(
embeddings(X,
config.WORD_COUNT,
config.EMBEDDING_DIM,
input_length=config.SEQUENCE_LEN,
frozen=config.is_embedding_trainable()))
dropout_1 = model.add(dropout(embedding, config.DROPOUT_LIST[0]))
conv_list = []
for k_size, n_C, k_pool in zip(config.FILTER_SIZE_LIST,
config.FILTERS_PER_LAYER,
config.POOL_SIZE_LIST):
c = conv1d(dropout_1, k_size, n_C, nonlin='relu')
p = maxpool(c, k_pool)
conv_list.append(flatten(p))
if len(conv_list) > 1:
conv_out = model.add(concat(conv_list))
else:
conv_out = model.add(conv_list[0])
dense_1 = model.add(dense(conv_out, 150, nonlin='relu'))
dropout_2 = model.add(dropout(dense_1, config.DROPOUT_LIST[1]))
out = model.add(dense(dropout_2, n_classes, nonlin='softmax'))
model.compile(optimizer='rmsprop',
loss='softmax_entropy',
learning_rate=config.LEARNING_RATE,
ckpt_file=config.CKPT_PATH,
device=config.DEVICE)
return model
def main():
args = parse_args()
# dataset
X_train, y_train, X_val, y_val, X_test, y_test = load_dataset(flatten=True)
# model layer dimensions
input_dim = X_train.shape[1]
num_classes = 10
# create model
model = Model()
model.add(Dense(input_dim, 100), activation='relu')
model.add(Dense(100, 200), activation='relu')
model.add(Dense(200, 200), activation='relu')
model.add(Dense(200, num_classes))
# train model
model.fit(X_train,
y_train,
val_data=(X_val, y_val),
verbose=True,
epochs=args.epochs,
batch_size=args.batch_size,
lr=args.lr)
# evaluate model
model.eval(X_test, y_test, verbose=True)
from nn.layer import FC, Softmax, ReLU, Dropout from nn.optimizer import Adam, SGD from nn.loss import cross_entropy from nn.metrix import accuracy (X_train, y_train), (X_test, y_test) = load_mnist() X_train = X_train.reshape((X_train.shape[0], -1)) / 255 X_test = X_test.reshape((X_test.shape[0], -1)) / 255 transformer = MakeOneHot() y_train = transformer.fit_transform(y_train) y_test = transformer.transform(y_test) model = Model() model.add(FC(500, input_shape=784)) model.add(ReLU()) model.add(Dropout(0.5)) model.add(FC(150)) model.add(ReLU()) model.add(Dropout(0.5)) model.add(FC(50)) model.add(ReLU()) model.add(Dropout(0.5)) model.add(FC(10)) model.add(Softmax()) model.compile(Adam(eta=0.01), cross_entropy, accuracy) model.fit(X_train, y_train, max_iter=10, batch_size=2000)
def MNISTNet():
conv1_params = {
'kernel_h': 3,
'kernel_w': 3,
'pad': 0,
'stride': 1,
'in_channel': 1,
'out_channel': 6
}
conv2_params = {
'kernel_h': 3,
'kernel_w': 3,
'pad': 0,
'stride': 1,
'in_channel': 6,
'out_channel': 16
}
pool1_params = {
'pool_type': 'max',
'pool_height': 2,
'pool_width': 2,
'stride': 2,
'pad': 0
}
pool2_params = {
'pool_type': 'max',
'pool_height': 3,
'pool_width': 3,
'stride': 2,
'pad': 0
}
model = Model()
model.add(
Conv2D(conv1_params, name='conv1', initializer=Gaussian(std=0.001)))
model.add(ReLU(name='relu1'))
model.add(Pool2D(pool1_params, name='pooling1'))
model.add(
Conv2D(conv2_params, name='conv2', initializer=Gaussian(std=0.001)))
model.add(ReLU(name='relu2'))
model.add(Pool2D(pool2_params, name='pooling2'))
# model.add(Dropout(ratio=0.25, name='dropout1'))
model.add(Flatten(name='flatten'))
model.add(Linear(400, 256, name='fclayer1',
initializer=Gaussian(std=0.01)))
model.add(ReLU(name='relu3'))
# model.add(Dropout(ratio=0.5))
model.add(Linear(256, 10, name='fclayer2', initializer=Gaussian(std=0.01)))
return model
def MyFashMNIST_CNN():
conv1_params = {
'kernel_h': 3,
'kernel_w': 3,
'pad': 2,
'stride': 1,
'in_channel': 1,
'out_channel': 32
}
conv2_params = {
'kernel_h': 3,
'kernel_w': 3,
'pad': 2,
'stride': 1,
'in_channel': 32,
'out_channel': 64
}
conv3_params = {
'kernel_h': 3,
'kernel_w': 3,
'pad': 2,
'stride': 1,
'in_channel': 64,
'out_channel': 64
}
pool1_params = {
'pool_type': 'max',
'pool_height': 2,
'pool_width': 2,
'stride': 2,
'pad': 2
}
pool2_params = {
'pool_type': 'max',
'pool_height': 3,
'pool_width': 3,
'stride': 2,
'pad': 2
}
model = Model()
model.add(
Conv2D(conv1_params, name='conv1', initializer=Gaussian(std=0.001)))
model.add(ReLU(name='relu1'))
#model.add(Pool2D(pool1_params, name='pooling1'))
model.add(
Conv2D(conv2_params, name='conv2', initializer=Gaussian(std=0.001)))
model.add(ReLU(name='relu2'))
model.add(Pool2D(pool1_params, name='pooling1'))
model.add(Dropout(rate=0.25, name='dropout1'))
model.add(
Conv2D(conv3_params, name='conv3', initializer=Gaussian(std=0.001)))
model.add(ReLU(name='relu2'))
model.add(Pool2D(pool1_params, name='pooling2'))
model.add(Dropout(rate=0.25, name='dropout2'))
model.add(Flatten(name='flatten'))
model.add(
Linear(4096, 1024, name='fclayer1', initializer=Gaussian(std=0.01)))
model.add(ReLU(name='relu3'))
model.add(Dropout(rate=0.5))
model.add(
Linear(1024, 256, name='fclayer2', initializer=Gaussian(std=0.01)))
model.add(Dropout(rate=0.5))
model.add(Linear(256, 10, name='fclayer3', initializer=Gaussian(std=0.01)))
return model
def FashionMNIST_CNN():
conv1_params = {
'kernel_h': 3,
'kernel_w': 3,
'pad': 0,
'stride': 1,
'in_channel': 1,
'out_channel': 32
}
conv2_params = {
'kernel_h': 3,
'kernel_w': 3,
'pad': 0,
'stride': 1,
'in_channel': 32,
'out_channel': 64
}
pool1_params = {
'pool_type': 'max',
'pool_height': 2,
'pool_width': 2,
'stride': 2,
'pad': 0
}
model = Model()
model.add(
Conv2D(conv1_params, name='conv1', initializer=Gaussian(std=0.001)))
model.add(ReLU(name='relu1'))
model.add(
Conv2D(conv2_params, name='conv2', initializer=Gaussian(std=0.001)))
model.add(ReLU(name='relu2'))
model.add(Pool2D(pool1_params, name='pooling1'))
model.add(Dropout(rate=0.25))
model.add(Flatten(name='flatten'))
model.add(
Linear(9216, 128, name='fclayer1', initializer=Gaussian(std=0.01)))
model.add(ReLU(name='relu3'))
model.add(Dropout(rate=0.25))
model.add(Linear(128, 10, name='fclayer2', initializer=Gaussian(std=0.01)))
return model
def MyModel_FashionMNIST():
conv1_params = {
'kernel_h': 3,
'kernel_w': 3,
'pad': 2,
'stride': 1,
'in_channel': 1,
'out_channel': 6
}
conv2_params = {
'kernel_h': 3,
'kernel_w': 3,
'pad': 2,
'stride': 1,
'in_channel': 6,
'out_channel': 16
}
conv3_params = {
'kernel_h': 3,
'kernel_w': 3,
'pad': 2,
'stride': 1,
'in_channel': 16,
'out_channel': 16
}
pool1_params = {
'pool_type': 'max',
'pool_height': 2,
'pool_width': 2,
'stride': 2,
'pad': 0
}
pool2_params = {
'pool_type': 'max',
'pool_height': 2,
'pool_width': 2,
'stride': 2,
'pad': 0
}
model = Model()
model.add(
Conv2D(conv1_params, name='conv1', initializer=Gaussian(std=0.001)))
model.add(ReLU(name='relu1'))
# model.add(Pool2D(pool1_params, name='pooling1'))
model.add(
Conv2D(conv2_params, name='conv2', initializer=Gaussian(std=0.001)))
model.add(ReLU(name='relu2'))
model.add(Pool2D(pool2_params, name='pooling2'))
# model.add(Conv2D(conv3_params, name='conv3',
# initializer=Gaussian(std=0.001)))
# model.add(ReLU(name='relu1'))
# model.add(Pool2D(pool1_params, name='pooling1'))
model.add(Flatten(name='flatten'))
model.add(
Linear(
3136,
100,
name='fclayer1', #10816
initializer=Gaussian(std=0.01)))
model.add(ReLU(name='relu3'))
model.add(Linear(100, 10, name='fclayer2', initializer=Gaussian(std=0.01)))
return model
def MyFashionModel_CNN():
conv1_params = {
'kernel_h': 3,
'kernel_w': 3,
'pad': 2, #same
'stride': 1,
'in_channel': 1,
'out_channel': 32
}
conv2_params = {
'kernel_h': 3,
'kernel_w': 3,
'pad': 2, #same
'stride': 1,
'in_channel': 32,
'out_channel': 32
}
conv3_params = {
'kernel_h': 3,
'kernel_w': 3,
'pad': 2, #same
'stride': 1,
'in_channel': 32,
'out_channel': 64
}
conv4_params = {
'kernel_h': 3,
'kernel_w': 3,
'pad': 2, #same
'stride': 1,
'in_channel': 64,
'out_channel': 128
}
pool1_params = {
'pool_type': 'max',
'pool_height': 2,
'pool_width': 2,
'stride': 2,
'pad': 0
}
model = Model()
model.add(Conv2D(conv1_params, name='conv1',initializer=Gaussian(std=0.001)))
model.add(ReLU(name='relu1'))
# model.add(Dropout(rate=0.25, name='dropout1'))
model.add(Conv2D(conv2_params, name='conv2',
initializer=Gaussian(std=0.001)))
model.add(ReLU(name='relu2'))
model.add(Dropout(rate=0.25, name='dropout2'))
model.add(Conv2D(conv3_params, name='conv3',
initializer=Gaussian(std=0.001)))
model.add(ReLU(name='relu3'))
model.add(Pool2D(pool1_params, name='pooling1'))
model.add(Dropout(rate=0.25, name='dropout3'))
model.add(Conv2D(conv4_params, name='conv4',initializer=Gaussian(std=0.001)))
model.add(ReLU(name='relu4'))
model.add(Dropout(rate=0.25, name='dropout4'))
model.add(Flatten(name='flatten'))
model.add(Linear(25088, 512, name='fclayer1', #512
initializer=Gaussian(std=0.01)))
model.add(ReLU(name='relu5'))
model.add(Dropout(rate=0.5, name='dropout5'))
model.add(Linear(512, 256, name='fclayer1', #128
initializer=Gaussian(std=0.01)))
model.add(ReLU(name='relu5'))
model.add(Dropout(rate=0.5, name='dropout5'))
model.add(Linear(256, 10, name='fclayer2',
initializer=Gaussian(std=0.01)))
return model
