def build_forward(self, _input):
output = _input # [batch_size, num_steps, rnn_units]
self.feature_dim = int(output.get_shape()[2]) # rnn_units
output = tf.reshape(output, [-1, self.feature_dim]) # [batch_size * num_steps, rnn_units]
final_activation = 'sigmoid' if self.out_dim == 1 else 'softmax'
if self.net_type == 'simple':
net_config = [] if self.net_config is None else self.net_config
with tf.variable_scope('wider_actor'):
for layer in net_config:
units, activation = layer.get('units'), layer.get('activation', 'relu')
output = BasicModel.fc_layer(output, units, use_bias=True)
output = BasicModel.activation(output, activation)
logits = BasicModel.fc_layer(output, self.out_dim, use_bias=True) # [batch_size * num_steps, out_dim]
probs = BasicModel.activation(logits, final_activation) # [batch_size * num_steps, out_dim]
probs_dim = self.out_dim
if self.out_dim == 1:
probs = tf.concat([1 - probs, probs], axis=1)
probs_dim = 2
self.q_values = tf.reshape(BasicModel.fc_layer(output, probs_dim, use_bias=True), [-1, self.num_steps, probs_dim])
# [batch_size, num_steps, out_dim]
self.decision = tf.multinomial(tf.log(probs), 1) # [batch_size * num_steps, 1]
self.decision = tf.reshape(self.decision, [-1, self.num_steps]) # [batch_size, num_steps]
self.probs = tf.reshape(probs, [-1, self.num_steps, probs_dim]) # [batch_size, num_steps, out_dim]
self.values = tf.reduce_sum(tf.multiply(self.q_values, self.probs), axis=-1) # [batch_size, num_steps]
self.selected_prob = tf.reduce_sum(tf.one_hot(self.decision, probs_dim) * self.probs, axis=-1)
self.selected_q = tf.reduce_sum(tf.one_hot(self.decision, probs_dim) * self.q_values, axis=-1)
else:
raise ValueError('Do not support %s' % self.net_type)
def build_forward(self, encoder_output, encoder_state, is_training, decision_trajectory):
self._define_input()
self.decision, self.probs, self.selected_prob, self.q_values, self.selected_q, self.values = [], [], [], [], [], []
batch_size = array_ops.shape(encoder_output)[0]
if self.attention_config is None:
cell = self.build_decoder_cell(encoder_state)
cell_state = encoder_state
cell_input = tf.zeros(shape=[batch_size], dtype=tf.int32)
with tf.variable_scope('deeper_actor'):
for _i in range(self.decision_num):
cell_input_embed = embedding(cell_input, 1 if _i == 0 else self.out_dims[_i - 1],
self.embedding_dim, name='deeper_actor_embedding_%d' % _i)
with tf.variable_scope('rnn', reuse=(_i > 0)):
cell_output, cell_state = cell(cell_input_embed, cell_state)
with tf.variable_scope('classifier_%d' % _i):
logits_i = BasicModel.fc_layer(cell_output, self.out_dims[_i], use_bias=True) # [batch_size, out_dim_i]
with tf.variable_scope('q_value_%d' % _i):
qv = BasicModel.fc_layer(cell_output, self.out_dims[_i], use_bias=True) # [batch_size, out_dim_i]
act_i = 'softmax'
probs_i = BasicModel.activation(logits_i, activation=act_i) # [batch_size, out_dim_i]
if _i == 1:
# determine the layer index for deeper actor
# require mask
one_hot_block_decision = tf.one_hot(cell_input, depth=self.out_dims[0], dtype=tf.int32)
max_layer_num = tf.multiply(self.block_layer_num, one_hot_block_decision)
max_layer_num = tf.reduce_max(max_layer_num, axis=1) # [batch_size]
layer_mask = tf.sequence_mask(max_layer_num, self.out_dims[1], dtype=tf.float32)
probs_i = tf.multiply(probs_i, layer_mask)
# rescale the sum to 1
probs_i = tf.divide(probs_i, tf.reduce_sum(probs_i, axis=1, keep_dims=True))
decision_i = tf.multinomial(tf.log(probs_i), 1) # [batch_size, 1]
decision_i = tf.cast(decision_i, tf.int32)
decision_i = tf.reshape(decision_i, shape=[-1]) # [batch_size]
cell_input = tf.cond(
is_training,
lambda: decision_trajectory[:, _i],
lambda: decision_i,
)
self.q_values.append(qv)
self.decision.append(decision_i)
self.probs.append(probs_i)
self.values.append(tf.reduce_sum(tf.multiply(qv, probs_i), axis=-1))
sq = tf.reduce_sum(tf.one_hot(decision_i, self.out_dims[_i]) * qv, axis=-1)
self.selected_q.append(sq)
sp = tf.reduce_sum(tf.one_hot(decision_i, self.out_dims[_i]) * probs_i, axis=-1)
self.selected_prob.append(sp)
self.decision = tf.stack(self.decision, axis=1) # [batch_size, decision_num]
self.values = tf.stack(self.values, axis=1) # [batch_size, decision_num]
self.selected_q = tf.stack(self.selected_q, axis=1)
self.selected_prob = tf.stack(self.selected_prob, axis=1)
else:
raise NotImplementedError
def test_train_with_stages(self):
model = BasicModel({
"stage": "second"
})
model._first_stage = MagicMock()
model._second_stage = MagicMock()
model._third_stage = MagicMock()
model._save = MagicMock()
model.stages = [
("first", model._first_stage),
("second", model._second_stage),
("third", model._third_stage)
]
process = Mock(
first=model._first_stage,
second=model._second_stage,
third=model._third_stage,
save=model._save
)
model.train()
process.assert_has_calls([
call.second(),
call.save(),
call.third(),
call.save()
])
model._first_stage.assert_not_called()
def build(self, _input, net, store_output_op=False):
assert (isinstance(net, BasicModel))
output = _input
if not self.ready:
return output
with tf.variable_scope(self._id):
self._scope = tf.get_variable_scope().name
param_initializer = self.param_initializer
if self.pre_activation:
# batch normalization
if self.use_bn:
output = BasicModel.batch_norm(
output,
net.is_training,
net.net_config.bn_epsilon,
net.net_config.bn_decay,
param_initializer=param_initializer)
# activation
output = BasicModel.activation(output, self.activation)
# Pooling
if self._type == 'avg':
output = BasicModel.avg_pool(output,
k=self.kernel_size,
s=self.strides)
elif self._type == 'max':
output = BasicModel.max_pool(output,
k=self.kernel_size,
s=self.strides)
else:
raise ValueError('Do not support the pooling type: %s' %
self._type)
else:
# Pooling
if self._type == 'avg':
output = BasicModel.avg_pool(output,
k=self.kernel_size,
s=self.strides)
elif self._type == 'max':
output = BasicModel.max_pool(output,
k=self.kernel_size,
s=self.strides)
else:
raise ValueError('Do not support the pooling type: %s' %
self._type)
# batch normalization
if self.use_bn:
output = BasicModel.batch_norm(
output,
net.is_training,
net.net_config.bn_epsilon,
net.net_config.bn_decay,
param_initializer=param_initializer)
# activation
output = BasicModel.activation(output, self.activation)
# dropout
output = BasicModel.dropout(output, self.keep_prob,
net.is_training)
if store_output_op:
self.output_op = output
return output
def test_submit(self):
model = BasicModel({
"results_folder": "./"
})
test_X = pd.DataFrame([[0, 1], [1, 0]], columns=["y", "n"])
test_y = pd.DataFrame([[0, 1], [1, 1]], columns=["id", "prediction"])
model._predict = MagicMock(return_value=[1, 1])
model.submit(test_X)
results = pd.read_csv("./submission.csv")
self.assertSequenceEqual(results.to_dict(), test_y.to_dict())
def get_user_list(cls, group_id=None):
sql = ''
if group_id == None:
sql = f'select * from {cls.prefix}user_info_tbl where stat<3 order by id'
else:
sql = f'select * from {cls.prefix}user_info_tbl where stat<3 and group_id={group_id} order by id'
return BasicModel.query_all(sql)
def build(self, _input, net, store_output_op=False):
assert (isinstance(net, BasicModel))
output = _input
if not self.ready:
return output
with tf.variable_scope(self._id):
self._scope = tf.get_variable_scope().name
param_initializer = self.param_initializer
# flatten if not
output = BasicModel.flatten(output)
if self.pre_activation:
# batch normalization
if self.use_bn:
output = BasicModel.batch_norm(
output,
net.is_training,
net.net_config.bn_epsilon,
net.net_config.bn_decay,
param_initializer=param_initializer)
# activation
output = BasicModel.activation(output, self.activation)
# FC
output = BasicModel.fc_layer(
output,
self.units,
self.use_bias,
param_initializer=param_initializer)
else:
# FC
output = BasicModel.fc_layer(
output,
self.units,
self.use_bias,
param_initializer=param_initializer)
# batch normalization
if self.use_bn:
output = BasicModel.batch_norm(
output,
net.is_training,
net.net_config.bn_epsilon,
net.net_config.bn_decay,
param_initializer=param_initializer)
# activation
output = BasicModel.activation(output, self.activation)
# dropout
output = BasicModel.dropout(output, self.keep_prob,
net.is_training)
if store_output_op:
self.output_op = output
return output
def test_preprocess_train_data(self):
train_df = pd.DataFrame(
[["text", "additional", 1]],
columns=["text", "additional", "label"]
)
train_df.to_csv("train.csv")
config = {
"data_folder": "./",
"train_data_path": "train.csv"
}
model = BasicModel(config)
def remove_additional(df):
return df.drop(columns=["additional"])
model._preprocess_data = MagicMock(side_effect=remove_additional)
model._preprocess_and_save_data(config["train_data_path"])
preprocessed_train_df = pd.read_csv("./preprocessed_train.csv")
assert "text" in preprocessed_train_df.columns
assert "label" in preprocessed_train_df.columns
assert "additional" not in preprocessed_train_df.columns
def make_model(model_name):
if model_name == available_models[0]:
return AvgSeqDenseModel()
elif model_name == available_models[1]:
return StackedBiLSTMDenseModel()
elif model_name == available_models[2]:
return TransformerEncoderBiLSTMDenseModel()
elif model_name == available_models[3]:
return RNMTPlusEncoderBiLSTMDenseModel()
elif model_name == available_models[4]:
return MultiHeadAttnAvgDenseModel()
else:
return BasicModel()
def test_load_parameters(self):
config = {
"data_folder": "folder",
"train_data_path": "train",
"test_data_path": "test",
"unlabeled_data_path": "unlabeled",
"results_folder": "results",
"stage": "stage"
}
model = BasicModel(config)
assert model.train_data_path == "train"
assert model.test_data_path == "test"
assert model.unlabeled_data_path == "unlabeled"
assert model.data_folder == "folder"
assert model.results_folder == "results"
assert model.stage == "stage"
def build(self, _input, net, store_output_op=False):
assert (isinstance(net, BasicModel))
output = _input
if not self.ready:
return output
with tf.variable_scope(self._id):
self._scope = tf.get_variable_scope().name
param_initializer = self.param_initializer
if self.pre_activation:
# batch normalization
if self.use_bn:
output = BasicModel.batch_norm(
output,
net.is_training,
net.net_config.bn_epsilon,
net.net_config.bn_decay,
param_initializer=param_initializer)
# activation
output = BasicModel.activation(output, self.activation)
# convolutional
output = BasicModel.conv2d(output,
self.filter_num,
self.kernel_size,
self.strides,
param_initializer=param_initializer)
else:
# convolutional
output = BasicModel.conv2d(output,
self.filter_num,
self.kernel_size,
self.strides,
param_initializer=param_initializer)
# batch normalization
if self.use_bn:
output = BasicModel.batch_norm(
output,
net.is_training,
net.net_config.bn_epsilon,
net.net_config.bn_decay,
param_initializer=param_initializer)
# activation
output = BasicModel.activation(output, self.activation)
# dropout
output = BasicModel.dropout(output, self.keep_prob,
net.is_training)
if store_output_op:
self.output_op = output
return output
def test_init_model(self):
model = BasicModel({
"train_data_path": "train",
"unlabeled_data_path": "unlabeled",
"test_data_path": "test"
})
model._preprocess_and_save_data = MagicMock()
model._load = MagicMock()
model.init()
model._preprocess_and_save_data.assert_any_call("train")
model._preprocess_and_save_data.assert_any_call("unlabeled")
model._preprocess_and_save_data.assert_any_call("test")
model._load.assert_called_with()
def build_training_process(self):
c = tf.constant(0.8, dtype=tf.float32)
# if self.wider_seg_deepr > 0, then get wide_side_obj, else wider_entropy = 0
wider_side_obj, wider_entropy = tf.cond(
tf.greater(self.wider_seg_deeper,
0), lambda: self.get_wider_side_obj(), lambda:
(tf.constant(0.0, dtype=tf.float32),
tf.constant(0.0, dtype=tf.float32)))
wider_side_obj = tf.reduce_sum(wider_side_obj *
tf.minimum(c, self.wider_rho))
batch_size = array_ops.shape(self.reward)[0]
deeper_side_obj, deeper_entropy = tf.cond(
self.has_deeper, lambda: self.get_deeper_side_obj(), lambda:
(tf.constant(0.0, dtype=tf.float32),
tf.constant(0.0, dtype=tf.float32)))
deeper_side_obj = tf.reduce_sum(deeper_side_obj *
tf.minimum(c, self.deeper_rho))
self.obj = wider_side_obj + deeper_side_obj
entropy_term = wider_entropy * tf.cast(self.wider_seg_deeper, tf.float32) + \
deeper_entropy * tf.cast(batch_size - self.wider_seg_deeper, tf.float32)
entropy_term /= tf.cast(batch_size, tf.float32)
g = -self.obj - self.entropy_penalty * entropy_term
optimizer = BasicModel.build_optimizer(self.learning_rate,
self.opt_config[0],
self.opt_config[1])
self.train_step = [optimizer.minimize(g)]
wq = tf.reshape(self.wider_actor.selected_q, [-1])
w_loss = tf.losses.mean_squared_error(self.wider_qrets, wq)
self.update_wider_q = tf.train.AdamOptimizer(
learning_rate=self.learning_rate).minimize(loss=w_loss)
dq = tf.reshape(self.deeper_actor.selected_q, [-1])
d_loss = tf.losses.mean_squared_error(self.deeper_qrets, dq)
self.update_deeper_q = tf.train.AdamOptimizer(
learning_rate=self.learning_rate).minimize(loss=d_loss)
# add baseline to training step
if self.baseline_actor is not None:
self.build_baseline_network()
def build_training_process(self):
wider_side_obj, wider_entropy = tf.cond(
tf.greater(self.wider_seg_deeper,
0), lambda: self.get_wider_side_obj(), lambda:
(tf.constant(0.0, dtype=tf.float32),
tf.constant(0.0, dtype=tf.float32)))
batch_size = array_ops.shape(self.reward)[0]
deeper_side_obj, deeper_entropy = tf.cond(
self.has_deeper, lambda: self.get_deeper_side_obj(), lambda:
(tf.constant(0.0, dtype=tf.float32),
tf.constant(0.0, dtype=tf.float32)))
self.obj = wider_side_obj + deeper_side_obj
entropy_term = wider_entropy * tf.cast(self.wider_seg_deeper, tf.float32) + \
deeper_entropy * tf.cast(batch_size - self.wider_seg_deeper, tf.float32)
entropy_term /= tf.cast(batch_size, tf.float32)
optimizer = BasicModel.build_optimizer(self.learning_rate,
self.opt_config[0],
self.opt_config[1])
# print "in build, reward = {}".format(self.reward)
self.train_step = optimizer.minimize(-self.obj - self.entropy_penalty *
entropy_term)
def build_training_process(self):
# if self.wider_seg_deepr > 0, then get wide_side_obj, else wider_entropy = 0
wider_side_obj, wider_entropy = tf.cond(
tf.greater(self.wider_seg_deeper, 0),
lambda: self.get_wider_side_obj(),
lambda: (tf.constant(0.0, dtype=tf.float32), tf.constant(0.0, dtype=tf.float32))
)
batch_size = array_ops.shape(self.reward)[0]
deeper_side_obj, deeper_entropy = tf.cond(
self.has_deeper,
lambda: self.get_deeper_side_obj(),
lambda: (tf.constant(0.0, dtype=tf.float32), tf.constant(0.0, dtype=tf.float32))
)
self.obj = wider_side_obj + deeper_side_obj
entropy_term = wider_entropy * tf.cast(self.wider_seg_deeper, tf.float32) + \
deeper_entropy * tf.cast(batch_size - self.wider_seg_deeper, tf.float32)
entropy_term /= tf.cast(batch_size, tf.float32)
optimizer = BasicModel.build_optimizer(self.learning_rate, self.opt_config[0], self.opt_config[1])
self.train_step = [optimizer.minimize(- self.obj - self.entropy_penalty * entropy_term)]
# add baseline to training step
if self.baseline_actor is not None:
self.build_baseline_network()
args = parser.parse_args()
return args
if __name__ == '__main__':
# get arguments
args = parse()
# set seeds
torch.manual_seed(args.seed)
if use_cuda:
torch.cuda.manual_seed_all(args.seed)
#model = VGGModel(vgg_name='VGG13')
#model = resnet(model_name='resnet18', pretrained=False, num_classes=31)
model_empty_vs_all = BasicModel(nlabels=2)
if use_cuda:
model_empty_vs_all.cuda()
model_parameters = filter(lambda p: p.requires_grad, model_empty_vs_all.parameters())
params = sum([np.prod(p.size()) for p in model_parameters])
print('Total number of parameters: {}\n'.format(params))
optimizer = optim.Adam(params=model_empty_vs_all.parameters(), lr=args.learning_rate,
weight_decay=args.weight_decay)
scheduler = StepLR(optimizer, step_size=20, gamma=0.8)
if args.load_model is None:
epoch_start = 0
best_valid_acc = 0
pred[np.intersect1d(zero_idx, zero_idx_p)] = 21
pred[pred>=21] += 1
pdb.set_trace()
correct += float(pred.eq(target.data).sum())
del output, data, target
state['{}_acc'.format(mode)] = correct / len(data_loader.dataset)
if __name__ == '__main__':
# get arguments
args = parse()
model_empty_vs_all = BasicModel(nlabels=2)
if use_cuda:
model_empty_vs_all.cuda()
model_parameters = filter(lambda p: p.requires_grad, model_empty_vs_all.parameters())
params = sum([np.prod(p.size()) for p in model_parameters])
print('Total number of parameters: {}\n'.format(params))
model = AttentionLocalizationModel(nlabels=30)
if use_cuda:
model.cuda()
model_parameters = filter(lambda p: p.requires_grad, model.parameters())
params = sum([np.prod(p.size()) for p in model_parameters])
print('Total number of parameters: {}\n'.format(params))
import tensorflow as tf
from models.basic_model import BasicModel
from models.convolutional_model import ConvModel
from dataset import MnistDataset
epochs = 5
batch_size = 4
train_set = MnistDataset('data/train')
n_batches = len(train_set) // batch_size
model = BasicModel(resolution=[28, 28], channels=1)
# model = ConvModel(resolution=[28, 28], channels=1)
saver = tf.train.Saver(
) # We use this to save the model. Instantiate it after all Variables have been created
label_placeholder = tf.placeholder(tf.float32, shape=[batch_size, 10])
loss = tf.losses.softmax_cross_entropy(label_placeholder, model.predictions)
update = tf.train.AdamOptimizer(learning_rate=0.001).minimize(loss)
top_predictions = tf.argmax(model.predictions,
axis=1) # probabilities -> top prediction
top_labels = tf.argmax(label_placeholder, axis=1) # one_hot -> number
correct = tf.equal(top_predictions, top_labels) # bool Tensor
accuracy = tf.reduce_mean(tf.cast(correct,
tf.float32)) # Average correct guesses
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
def get_model(self):
model = gconfig.get_config_param("model", "str")
if model == "basic_conv_graph":
model = BasicModel()
elif model == "conv_graph_dgcnn_fast_conv":
model = BasicModel()
model.set_conv_segment(FastConvSegment())
model.set_graph_segment(DgcnnSegment())
elif model == "conv_graph_garnet_fast_conv":
model = BasicModel()
model.set_conv_segment(FastConvSegment())
model.set_graph_segment(GarNetSegment())
elif model == "conv_fcnn_fast_conv":
model = BasicModel()
model.set_conv_segment(FastConvSegment())
model.set_graph_segment(FcnnSegment())
elif model == "conv_grav_net_fast_conv":
model = BasicModel()
model.set_conv_segment(FastConvSegment())
model.set_graph_segment(GravnetSegment())
else:
return ModelInterface() # TODO: Fix this
return model
graph_encoder = GraphEncoder()
with open('basic_model_graph_encoder.pickle', 'wb') as out:
pickle.dump(graph_encoder, out)
np.savez_compressed(data_file, data=data, no_labels=no_labels, all_labels=all_labels)
model_attributes = {
'afm': 30,
'bfm': 8,
'mfm': 30,
'adj': 1,
'out': 120,
'classification_output': 2
}
model = nn.Sequential(
GraphWrapper(BasicModel(model_attributes['afm'], model_attributes['bfm'], model_attributes['mfm'],
model_attributes['adj'], model_attributes['out'])),
nn.BatchNorm1d(model_attributes['out']),
nn.Linear(model_attributes['out'], 60),
nn.ReLU(),
nn.Linear(60, 30),
nn.ReLU(),
nn.Linear(30, 15),
nn.ReLU(),
nn.Linear(15, model_attributes['classification_output'])
) #
# selected_label = np.random.choice(np.arange(no_labels))
selected_label = 243
print "Target selected: {}".format(selected_label)
for graph in data:
def update_user_stat_by_id(cls, id, stat):
sql = f'update {cls.prefix}user_info_tbl set stat={stat} where id={id}'
return BasicModel.update(sql)
def get_user_info_by_id(cls, id):
sql = f'select * from {cls.prefix}user_info_tbl where id={id}'
return BasicModel.query_all(sql)[0]