def create_data_processor(is_undersampled, is_oversampled, seed=0):
global DataProcessor
if is_undersampled:
DataProcessor = data_processor.DataProcessor(
TrainData).with_scaling().with_undersampling(seed)
else:
if is_oversampled:
DataProcessor = data_processor.DataProcessor(
TrainData).with_scaling().with_oversampling(seed)
else:
DataProcessor = data_processor.DataProcessor(
TrainData).with_scaling()
print("Batch size: " + str(batch_size))
print("Epochs: " + str(num_epochs))
print("Evaluate every steps: " + str(evaluate_steps))
print("Early Stopping: " + str(early_stopping))
print("Patience: " + str(patience))
# Data set parameters
vocab_size = experiment_params['vocab_size']
max_seq_length = experiment_params['max_seq_length']
to_tokens = experiment_params['to_tokens']
embedding_dim = experiment_params['embedding_dim']
embedding_type = experiment_params['embedding_type']
embedding_source = experiment_params['embedding_source']
# Initialize the dataset and embedding processor
data_set = data_processor.DataProcessor(task_name, dataset_dir, max_seq_length, vocab_size=vocab_size, to_tokens=to_tokens)
embedding = embedding_processor.get_embedding(embedding_type)
# If dataset folder is empty get the metadata and datasets
if not os.listdir(dataset_dir):
data_set.get_dataset()
# Load the metadata
vocabulary, labels = data_set.load_metadata()
# Generate the embedding matrix
embedding_matrix = embedding.get_embedding_matrix(embeddings_dir, embedding_source, embedding_dim, vocabulary)
# Loop over each experiment in the optimiser
for experiment in model_optimiser.get_experiments(project_name=experiment_params['project_name'], workspace="nathanduran", auto_output_logging='simple'):
import torch.optim as optim
import torch.utils.data as torch_data
from torch.autograd import Variable
import matplotlib.pyplot as plt
import time
import numpy as np
import global_define
import network as nw
import data_processor
import torch
import torch.nn as nn
IfInitial = False
dp = data_processor.DataProcessor()
if IfInitial:
dp.init_data()
print('initial done!')
# get training data and test data
train_set_np = np.load("data/train_set.npy")
train_set_label_np = np.load("data/train_set_label.npy")
test_set_np = np.load("data/test_set.npy")
test_set_label_np = np.load("data/test_set_label.npy")
# network
G = nw.Generator().cuda()
D = nw.Discriminator().cuda()
G.weight_init(mean=0, std=0.01)
D.weight_init(mean=0, std=0.01)
import data_processor as dp import recommendation as rec import numpy as np processor = dp.DataProcessor() print processor.data print processor.simi(0, 1) print processor.simi(3, 1) print processor.simi(4, 1) recomend = rec.Recommendation() print recomend.get_non_watched_movies(1)
def train(self, epochs, batch_size, g_iter, d_iter, model_dir, model_name):
print('Training . . .')
dp = data_processor.DataProcessor()
sample_noise = self.get_noise(5 * 5)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
ckpt = tf.train.get_checkpoint_state(model_dir)
print(ckpt)
if ckpt and tf.train.checkpoint_exists(ckpt.model_checkpoint_path):
print('Reloading model parameters...')
self.saver.restore(sess, ckpt.model_checkpoint_path)
for epoch in range(epochs):
d_loss, g_loss = 0, 0
batched_img, batched_label, batched_wrong_label = dp.get_batch(
batch_size, d_iter)
dataset = list(
zip(batched_img, batched_label, batched_wrong_label))
for b, (img, label, wrong_label) in enumerate(dataset):
z = self.get_noise(batch_size)
for it in range(d_iter):
_, d_loss, d_real, d_fake = sess.run(
[
self.d_optim, self.d_loss, self.d_real_loss,
self.d_fake_loss
],
feed_dict={
self.g_inputs: z,
self.d_inputs: img,
self.label: label,
self.wrong_label: wrong_label
})
# print('Epoch [{:>2}/{:>2}] | Batch [{:>3}/{:>3}] | Iter [{:>2}/{:>2}] | W_dist: {:.6f}'
# .format(epoch+1, end_epoch, b+1, len(batched_img), it+1, d_iter, w_distance), end='\r')
z = self.get_noise(batch_size)
for it in range(g_iter):
_, g_loss = sess.run([self.g_optim, self.g_loss],
feed_dict={
self.g_inputs: z,
self.label: label
})
# print('Epoch [{:>2}/{:>2}] | Batch [{:>3}/{:>3}] | Iter [{:>2}/{:>2}] | G_loss: {:.6f}'
# .format(epoch+1, end_epoch, b+1, len(batched_img), it+1, g_iter, g_loss), end='\r')
print(
'Epoch [{:>2}/{:>2}] | Batch [{:>4}/{:>4}] | D_loss: {:.6f} | G_loss: {:.6f} | d_real: {:.6f} | d_fake: {:.6f}'
.format(epoch + 1, epochs, b + 1, len(batched_img),
d_loss, g_loss, d_real, d_fake))
if (epoch + 1) % self.display_step == 0:
samples = sess.run(self.g_infer,
feed_dict={
self.g_inputs: sample_noise,
self.label: self.test_labels
})
samples = samples / 2 + 0.5
fig = self.visualize_result(samples)
plt.savefig(self.pic_path +
'{}.png'.format(str(epoch + 1).zfill(4)),
bbox_inches='tight')
plt.close(fig)
if model_name != '':
# if os.path.exists('./model')
self.saver.save(sess,
'./model/{}/{}'.format(model_name, model_name))
print('Model saved at \'{}\''.format('./model/' + model_name))
vocab_size = experiment_params['vocab_size']
max_seq_length = experiment_params['max_seq_length']
to_tokens = experiment_params['to_tokens']
to_lower = experiment_params['to_lower']
use_punct = experiment_params['use_punct']
train_embeddings = experiment_params['train_embeddings']
embedding_dim = experiment_params['embedding_dim']
embedding_type = experiment_params['embedding_type']
embedding_source = experiment_params['embedding_source']
use_crf = model_params['use_crf'] if 'use_crf' in model_params else False
# Initialize the dataset processor
data_set = data_processor.DataProcessor(task_name,
dataset_dir,
max_seq_length,
vocab_size=vocab_size,
to_tokens=to_tokens,
to_lower=to_lower,
use_punct=use_punct)
# If dataset folder is empty get the metadata and datasets to .npz files
if not os.listdir(dataset_dir):
data_set.get_dataset()
# Load the metadata
vocabulary, labels = data_set.load_metadata()
# Generate the embedding matrix
embeddings = embedding_processor.get_embedding(embeddings_dir,
embedding_type,
embedding_source,
def train(self, start_epoch, epochs, batch_size, g_iter, d_iter,
model_dir):
dp = data_processor.DataProcessor()
sample_noise = np.random.normal(0, 1, (5 * 5, self.noise_dim))
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
ckpt = tf.train.get_checkpoint_state(model_dir)
print(ckpt)
if ckpt and tf.train.checkpoint_exists(ckpt.model_checkpoint_path):
print('Reloading model parameters...')
self.saver.restore(sess, ckpt.model_checkpoint_path)
end_epoch = start_epoch + epochs
for epoch in range(start_epoch, end_epoch):
w_distance, g_loss = 0, 0
batched_img = dp.get_batch(batch_size, d_iter)
for b, img in enumerate(batched_img):
for it in range(d_iter):
z = np.random.normal(0, 1,
(batch_size, self.noise_dim))
_, _, w_distance, d_real, d_fake = sess.run(
[
self.d_optim, self.clip_D, self.w_distance,
tf.reduce_mean(self.d_real),
tf.reduce_mean(self.d_fake)
],
feed_dict={
self.g_inputs: z,
self.d_inputs: img
})
# print('Epoch [{:>2}/{:>2}] | Batch [{:>3}/{:>3}] | Iter [{:>2}/{:>2}] | W_dist: {:.6f}'
# .format(epoch+1, end_epoch, b+1, len(batched_img), it+1, d_iter, w_distance), end='\r')
for it in range(g_iter):
z = np.random.normal(0, 1,
(batch_size, self.noise_dim))
_, g_loss = sess.run([self.g_optim, self.g_loss],
feed_dict={self.g_inputs: z})
# print('Epoch [{:>2}/{:>2}] | Batch [{:>3}/{:>3}] | Iter [{:>2}/{:>2}] | G_loss: {:.6f}'
# .format(epoch+1, end_epoch, b+1, len(batched_img), it+1, g_iter, g_loss), end='\r')
print(
'Epoch [{:>2}/{:>2}] | W_dist: {:.6f} | G_loss: {:.6f} | d_real: {:.6f} | d_fake: {:.6f}'
.format(epoch + 1, end_epoch, w_distance, g_loss,
d_real, d_fake))
if (epoch + 1) % self.display_step == 0:
samples = sess.run(self.g_sample,
feed_dict={self.g_inputs: sample_noise})
samples = samples / 2 + 0.5
# print(samples[0])
fig = self.visualize_result(samples)
plt.savefig(self.pic_path +
'{}.png'.format(str(epoch + 1).zfill(4)),
bbox_inches='tight')
plt.close(fig)
self.saver.save(
sess,
'./model/model_' + str(end_epoch) + '/model_' + str(end_epoch))
