Related
I would like to re-create the following keras model in PyTorch.
vocab_size = 22
maxlen = 200
embed_dim = 256
num_heads = 2
feed_forward_dim = 256
batch_size = 128
decoders = 5
def create_model():
inputs = layers.Input(shape=(maxlen,), dtype=tf.int32)
embedding_layer = TokenAndPositionEmbedding(maxlen, vocab_size, embed_dim)
x = embedding_layer(inputs)
decoder_blocks = []
for i in range(decoders):
decoder_blocks.append(DecoderBlock(embed_dim, num_heads, feed_forward_dim))
for i in range(len(decoder_blocks)):
x = decoder_blocks[i](x)
outputs = layers.Dense(vocab_size)(x)
model = keras.Model(inputs=inputs, outputs=[outputs, x])
loss_fn = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
model.compile(
optimizer=keras.optimizers.Adam(learning_rate=lr_schedule),
loss=[loss_fn, None],
)
return model
model = create_model()
Here are the Decoder and the TokenAndPositionEmbedding layers along with the Causal Attention Mask
def causal_attention_mask(batch_size, n_dest, n_src, dtype):
i = tf.range(n_dest)[:, None]
j = tf.range(n_src)
m = i >= j - n_src + n_dest
mask = tf.cast(m, dtype)
mask = tf.reshape(mask, [1, n_dest, n_src])
mult = tf.concat(
[tf.expand_dims(batch_size, -1), tf.constant([1, 1], dtype=tf.int32)], 0
)
return tf.tile(mask, mult)
class DecoderBlock(layers.Layer):
def __init__(self, embed_dim, num_heads, ff_dim, rate=0.1):
super(DecoderBlock, self).__init__()
self.att = layers.MultiHeadAttention(num_heads, embed_dim)
self.ffn = keras.Sequential(
[layers.Dense(ff_dim, activation="relu"), layers.Dense(embed_dim),]
)
self.layernorm1 = layers.LayerNormalization(epsilon=1e-6)
self.layernorm2 = layers.LayerNormalization(epsilon=1e-6)
self.dropout1 = layers.Dropout(rate)
self.dropout2 = layers.Dropout(rate)
def call(self, inputs):
input_shape = tf.shape(inputs)
batch_size = input_shape[0]
seq_len = input_shape[1]
causal_mask = causal_attention_mask(batch_size, seq_len, seq_len, tf.bool)
attention_output = self.att(inputs, inputs, attention_mask=causal_mask)
attention_output = self.dropout1(attention_output)
out1 = self.layernorm1(inputs + attention_output)
ffn_output = self.ffn(out1)
ffn_output = self.dropout2(ffn_output)
return self.layernorm2(out1 + ffn_output)
class TokenAndPositionEmbedding(layers.Layer):
def __init__(self, maxlen, vocab_size, embed_dim):
super(TokenAndPositionEmbedding, self).__init__()
self.token_emb = layers.Embedding(input_dim=vocab_size, output_dim=embed_dim)
self.pos_emb = layers.Embedding(input_dim=maxlen, output_dim=embed_dim)
def call(self, x):
maxlen = tf.shape(x)[-1]
positions = tf.range(start=0, limit=maxlen, delta=1)
positions = self.pos_emb(positions)
x = self.token_emb(x)
return x + positions
For reference, this code is copied directly from: https://keras.io/examples/generative/text_generation_with_miniature_gpt/
I have tried to create equivalent architecture in PyTorch using nn.TransformerDecoderLayer. Apologies for not including my own code, but I have been completely unsuccessful.
When I use linear or No activation in the last Discriminator layer using weight clipping Discriminator accuracy goes to 1 and Generator goes to 0. In case when I remove weight clipping, Generator accuracy goes to 1 and discriminator goes to 0 around 300 iterations. But when I use sigmoid activation as the last layer in the discriminator with clipping Generator accuracy goes to 1 and without clipping the generator loss get stuck while accuracies going as they should around 0.5.
NOTE - in all cases, results are produced and all of the show WARNING:tensorflow:Discrepancy between trainable weights and collected trainable weights, did you set model.trainable without calling model.compile after ?
Code is given here, please do not mind the indentation on copying and pasting it's everywhere -
class WGAN():
def __init__(self,
input_dim,
disc_filter,
disc_kernel,
disc_strides,
disc_dropout,
disc_lr,
gen_filter,
gen_kernel,
gen_strides,
gen_upsample,
gen_lr,
z_dim,
batch_size):
self.input_dim = input_dim
self.disc_filter = disc_filter
self.disc_kernel = disc_kernel
self.disc_strides = disc_strides
self.disc_dropout = disc_dropout
self.disc_lr = disc_lr
self.gen_filter = gen_filter
self.gen_kernel = gen_kernel
self.gen_strides = gen_strides
self.gen_upsample = gen_upsample
self.gen_lr = gen_lr
self.z_dim = z_dim
self.batch_size = batch_size
self.weight_init = RandomNormal(mean=0., stddev=0.02)
self.d_losses = []
self.g_losses = []
self.epoch = 0
self.Discriminator()
self.Generator()
self.full_model()
def wasserstein(self, y_true, y_pred):
return -K.mean(y_true * y_pred)
def Discriminator(self):
disc_input = Input(shape=self.input_dim, name='discriminator_input')
x = disc_input
for i in range(len(self.disc_filter)):
x = Conv2D(filters=self.disc_filter[i], kernel_size=self.disc_kernel[i], strides=self.disc_strides[i], padding='same', name='disc_'+str(i))(x)
x = LeakyReLU()(x)
x = Dropout(self.disc_dropout)(x)
x = BatchNormalization()(x)
x = Flatten()(x)
disc_output = Dense(1, activation='sigmoid', kernel_initializer = self.weight_init)(x)
self.discriminator = Model(disc_input, disc_output)
def Generator(self):
gen_input = Input(shape=(self.z_dim,), name='generator_input')
x = gen_input
x = Dense(7*7*self.batch_size, kernel_initializer = self.weight_init)(x)
x = LeakyReLU()(x)
x = BatchNormalization()(x)
x = Reshape(target_shape=(7,7,self.batch_size))(x)
for i in range(len(self.gen_filter)):
if self.gen_upsample[i]==2:
x = UpSampling2D(size=self.gen_upsample[i], name='upsample_'+str(i/2))(x)
x = Conv2D(filters=self.gen_filter[i], kernel_size=self.gen_kernel[i], strides=self.gen_strides[i], padding='same', name='gen_'+str(i))(x)
else:
x = Conv2DTranspose(filters=self.gen_filter[i], kernel_size=self.gen_kernel[i], strides=self.gen_strides[i], padding='same', name='gen_'+str(i))(x)
if i<len(self.gen_filter)-1:
x = BatchNormalization()(x)
x = LeakyReLU()(x)
else:
x = Activation("tanh")(x)
gen_output = x
self.generator = Model(gen_input, gen_output)
def set_trainable(self, model, val):
model.trainable=val
for l in model.layers:
l.trainable=val
def full_model(self):
### COMPILE DISCRIMINATOR
self.discriminator.compile(optimizer= Adam(self.disc_lr), loss = self.wasserstein, metrics=['accuracy'])
### COMPILE THE FULL GAN
self.set_trainable(self.discriminator, False)
self.discriminator.compile(optimizer= Adam(self.disc_lr), loss = self.wasserstein, metrics=['accuracy'])
model_input = Input(shape=(self.z_dim,), name='model_input')
model_output = self.discriminator(self.generator(model_input))
self.model = Model(model_input, model_output)
self.model.compile(optimizer= Adam(self.disc_lr), loss = self.wasserstein, metrics=['accuracy'])
self.set_trainable(self.discriminator, True)
def train_generator(self, batch_size):
valid = np.ones((batch_size,1))
noise = np.random.normal(0, 1, (batch_size, self.z_dim))
return self.model.train_on_batch(noise, valid)
def train_discriminator(self, x_train, batch_size, using_generator):
valid = np.ones((batch_size,1))
fake = np.zeros((batch_size,1))
if using_generator:
true_imgs = next(x_train)[0]
if true_imgs.shape[0] != batch_size:
true_imgs = next(x_train)[0]
else:
idx = np.random.randint(0, x_train.shape[0], batch_size)
true_imgs = x_train[idx]
noise = np.random.normal(0, 1, (batch_size, self.z_dim))
gen_imgs = self.generator.predict(noise)
d_loss_real, d_acc_real = self.discriminator.train_on_batch(true_imgs, valid)
d_loss_fake, d_acc_fake = self.discriminator.train_on_batch(gen_imgs, fake)
d_loss = 0.5 * (d_loss_real + d_loss_fake)
d_acc = 0.5 * (d_acc_real + d_acc_fake)
for l in self.discriminator.layers:
weights = l.get_weights()
weights = [np.clip(w, -0.01, 0.01) for w in weights]
l.set_weights(weights)
return [d_loss, d_loss_real, d_loss_fake, d_acc, d_acc_real, d_acc_fake]
def train(self, x_train, batch_size, epochs, print_every_n_batches = 50, using_generator = False):
for epoch in range(self.epoch, self.epoch + epochs):
d = self.train_discriminator(x_train, batch_size, using_generator)
g = self.train_generator(batch_size)
if self.epoch % print_every_n_batches == 0:
print ("%d [D loss: (%.3f)(R %.3f, F %.3f)] [D acc: (%.3f)(%.3f, %.3f)] [G loss: %.3f] [G acc: %.3f]" % (epoch, d[0], d[1], d[2], d[3], d[4], d[5], g[0], g[1]))
self.d_losses.append(d)
self.g_losses.append(g)
self.epoch+=1
So I tried implementing a Convolutional Neural Network on MNIST dataset in a similar fashion as this: https://github.com/tensorflow/tensorflow/blob/r1.1/tensorflow/examples/tutorials/mnist/mnist_with_summaries.py
However, on doing that, I noticed that for some reason my second max_pool is not happening. Also, I don't understand how the code in the above link works, more specifically, how the nn_layer method can be reused as the weights exist only in that scope and calling it twice would change them?
My code:
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
import os
from tensorflow.contrib.tensorboard.plugins import projector
current_path = os.path.dirname(os.path.realpath(__file__))
current_path = current_path+"/logs"
def train():
mnist = input_data.read_data_sets("MNIST_data", one_hot = True)
def initializer(shape):
initial = tf.truncated_normal(shape, stddev=0.1)
return tf.Variable(initial)
def conv2d(x,W):
return tf.nn.conv2d(x , W , [1,1,1,1] , padding="SAME")
def max_pool(x):
return tf.nn.max_pool(x , [1,2,2,1] , [1,2,2,1] , padding="SAME")
def conv_layer(x,length,width,input_channels,output_channels,layer_name,act=tf.nn.relu):
with tf.name_scope(layer_name):
with tf.name_scope('weights'):
weights = initializer([length,width,input_channels,output_channels])
tf.summary.histogram(layer_name+"_weights",weights)
with tf.name_scope('biases'):
biases = initializer([output_channels])
tf.summary.histogram(layer_name+"_biases",biases)
with tf.name_scope('activations'):
activations = act(conv2d(x,weights) + biases)
activations = max_pool(activations)
tf.summary.histogram(layer_name+"_activations",activations)
return activations
def dense_layer(x,input_size,output_size,layer_name,act=tf.nn.relu):
with tf.name_scope(layer_name):
with tf.name_scope('weights'):
weights = initializer([input_size,output_size])
tf.summary.histogram(layer_name+"_weights",weights)
with tf.name_scope('biases'):
biases = initializer([output_size])
tf.summary.histogram(layer_name+"_biases",biases)
with tf.name_scope('activations'):
activations = act(tf.matmul(x,weights) + biases)
tf.summary.histogram(layer_name+"_activations",activations)
return activations
def dropout(x,keep_prob):
with tf.name_scope('Dropout'):
dropped =tf.nn.dropout(x,keep_prob)
return dropped
with tf.name_scope('input'):
x = tf.placeholder(tf.float32, [None,784],name='image_inputs')
y = tf.placeholder(tf.float32, [None,10],name='image_labels')
keep_prob = tf.placeholder(tf.float32,name='keep_probability')
with tf.name_scope('input_reshape'):
x_image = tf.reshape(x , [-1,28,28,1])
tf.summary.image('input',x_image,50)
h1 = conv_layer(x_image,3,3,1,32,"first_convolution_layer")
h2 = conv_layer(h1,3,3,32,64,"second_convolution_layer")
h2 = tf.reshape(h1,[-1,7*7*64])
h2 = dropout(h2,keep_prob)
h3 = dense_layer(h2,7*7*64,1024,"first_dense_layer")
h3 = dropout(h3,keep_prob)
h4 = dense_layer(h3,1024,1024,"second_dense_layer")
h4 = dropout(h4,keep_prob)
h_out = dense_layer(h4,1024,10,"output_dense_layer",act=tf.nn.sigmoid)
with tf.name_scope("Loss"):
cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y, logits=h_out))
tf.summary.scalar('Loss',cost)
train = tf.train.AdamOptimizer().minimize(cost)
with tf.name_scope("Accuracy"):
correct_pred = tf.equal(tf.argmax(h_out, 1), tf.argmax(y, 1))
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
summary = tf.summary.merge_all()
init = tf.global_variables_initializer()
sess = tf.InteractiveSession()
sess.run(init)
saver = tf.train.Saver()
summary_writer = tf.summary.FileWriter(current_path, sess.graph)
for i in range(500):
batch = mnist.train.next_batch(500)
if(i%100 == 0):
summary_str = sess.run(summary,feed_dict={x:batch[0], y:batch[1], keep_prob:1.0})
summary_writer.add_summary(summary_str, i)
summary_writer.flush()
train_accuracy = accuracy.eval(feed_dict={x:batch[0], y:batch[1], keep_prob:1.0})
saver.save(sess, os.path.join(current_path,'model.ckpt'), i)
print("Step %d Training Accuracy: %f" %((i/100 + 1), train_accuracy))
train.run(feed_dict={x:batch[0], y:batch[1], keep_prob:0.5})
sum=0.0
for i in range(10):
batch_x = mnist.test.images[(i*1000):((i+1)*1000)-1]
batch_y = mnist.test.labels[(i*1000):((i+1)*1000)-1]
sum = sum + accuracy.eval(feed_dict={x:batch_x, y:batch_y, keep_prob:1.0})
print("Test Accuracy: %f" %(sum/10.0))
if tf.gfile.Exists(current_path):
tf.gfile.DeleteRecursively(current_path)
tf.gfile.MakeDirs(current_path)
train()
This is a simple typo.
Change this
h2 = tf.reshape(h1,[-1,7*7*64])
to this
h2 = tf.reshape(h2,[-1,7*7*64])
The error
InvalidArgumentError (see above for traceback): logits and labels must be same size: logits_size=[1000,10] labels_size=[500,10]
[[Node: Loss/SoftmaxCrossEntropyWithLogits = SoftmaxCrossEntropyWithLogits[T=DT_FLOAT, _device="/job:localhost/replica:0/task:0/cpu:0"](Loss/Reshape, Loss/Reshape_1)]]
went away.
I am using a machine which has 2 GPUs Titan Black to train my deep learning model which has 3 layers (3x3, 3x3 and 5x5).
The training runs pretty well but when I watch nvidia-smi (watch every 1 sec), I realized that my program uses only one GPU for computation, the second one always 0% even when the first one reach 100%.
I am trying to use tf.device to assign specific tasks for each of them but then they run one-by-one, not in parallel, and the total time was even increased, not reduced (I guess because 2 GPUs had to exchange values with each other)
Below is my program. It is quite messy, maybe you just need to pay attention at the graph where I use tf.device is enough...
Thank you so much!
import tensorflow as tf
import numpy as np
from six.moves import cPickle as pickle
import matplotlib.pyplot as plt
from os import listdir, sys
from os.path import isfile, join
from time import gmtime, strftime
import time
def validatePath(path):
path = path.replace("\\","/")
if (path[len(path)-1] != "/"):
path = path + "/"
return path
hidden_size_default = np.array([16, 32, 64, 32])
cnn1_default = 3
cnn2_default = 3
cnn3_default = 5
SIZE_BATCH_VALID = 200
input_path = 'ARCHIVES-sub-dataset'
output_path = 'ARCHIVES-model'
log_address = "trainlog.txt"
tf.app.flags.DEFINE_integer('h0', hidden_size_default[0], 'Size of hidden layer 0th')
tf.app.flags.DEFINE_integer('h1', hidden_size_default[1], 'Size of hidden layer 1st')
tf.app.flags.DEFINE_integer('h2', hidden_size_default[2], 'Size of hidden layer 2nd')
tf.app.flags.DEFINE_integer('h3', hidden_size_default[3], 'Size of hidden layer 3rd')
tf.app.flags.DEFINE_integer('k1', cnn1_default , 'Size of kernel 1st')
tf.app.flags.DEFINE_integer('k2', cnn2_default , 'Size of kernel 2nd')
tf.app.flags.DEFINE_integer('k3', cnn3_default , 'Size of kernel 3rd')
tf.app.flags.DEFINE_string('input_path', input_path, 'The parent directory which contains 2 directories: dataset and label')
tf.app.flags.DEFINE_string('output_path', output_path, 'The directory which will store models (you have to create)')
tf.app.flags.DEFINE_string('log_address', log_address, 'The file name which will store the log')
FLAGS = tf.app.flags.FLAGS
load_path = FLAGS.input_path
save_model_path = FLAGS.output_path
log_addr = FLAGS.log_address
load_path = validatePath(load_path)
save_model_path = validatePath(save_model_path)
cnn1 = FLAGS.k1
cnn2 = FLAGS.k2
cnn3 = FLAGS.k3
hidden_size = np.array([FLAGS.h0, FLAGS.h1, FLAGS.h2, FLAGS.h3])
# Shuffle the dataset and its label
def randomize(dataset, labels):
permutation = np.random.permutation(labels.shape[0])
shuffled_dataset = dataset[permutation,:]
shuffled_labels = labels[permutation]
return shuffled_dataset, shuffled_labels
def writemyfile(mystring):
with open(log_addr, "a") as myfile:
myfile.write(str(mystring + "\n"))
num_labels = 5
def accuracy(predictions, labels):
return (100.0 * np.sum(np.argmax(predictions, 1) == np.argmax(labels, 1))/ predictions.shape[0])
def weight_variable(shape):
initial = tf.truncated_normal(shape, stddev=0.1)
return tf.Variable(initial)
def bias_variable(shape):
initial = tf.constant(0.1, shape=shape)
return tf.Variable(initial)
def conv2d(x, W):
return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')
def max_pool_2x2(x):
return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
def DivideSets(input_set):
length_set = input_set.shape[0]
index_70 = int(length_set*0.7)
index_90 = int(length_set*0.9)
set_train = input_set[0:index_70]
set_valid = input_set[index_70:index_90]
set_test = input_set[index_90:length_set]
return np.float32(set_train), np.float32(set_valid), np.float32(set_test)
# from 1-value labels to 5 values of (0 and 1)
def LabelReconstruct(label_set):
label_set = label_set.astype(int)
new_label_set = np.zeros(shape=(len(label_set),num_labels))
for i in range(len(label_set)):
new_label_set[i][label_set[i]] = 1
return new_label_set.astype(int)
def LoadDataSet(load_path):
list_data = [f for f in listdir(load_path + "dataset/") if isfile(join(load_path + "dataset/", f))]
list_label = [f for f in listdir(load_path + "label/") if isfile(join(load_path + "dataset/", f))]
if list_data.sort() == list_label.sort():
return list_data
else:
print("data and labels are not suitable")
return 0
# load, randomize, normalize images and reconstruct labels
def PrepareData(*arg):
filename = arg[0]
loaded_dataset = pickle.load( open( load_path + "dataset/" + filename, "rb" ))
loaded_labels = pickle.load( open( load_path + "label/" + filename, "rb" ))
if len(arg) == 1:
datasize = len(loaded_labels)
elif len(arg) == 2:
datasize = int(arg[1])
else:
print("not more than 2 arguments please!")
dataset_full,labels_full = randomize(loaded_dataset[0:datasize], loaded_labels[0:datasize])
return NormalizeData(dataset_full), LabelReconstruct(labels_full)
def NormalizeData(dataset):
dataset = dataset - (dataset.mean())
dataset = dataset / (dataset.std())
return dataset
### LOAD DATA
listfiles = LoadDataSet(load_path)
# divide
listfiles_train = listfiles[0:15]
listfiles_valid = listfiles[15:25]
listfiles_test = listfiles[25:len(listfiles)]
graphCNN = tf.Graph()
with graphCNN.as_default():
with tf.device('/gpu:0'):
x = tf.placeholder(tf.float32, shape=(None, 224,224,3)) # X
y_ = tf.placeholder(tf.float32, shape=(None, num_labels)) # Y_
dropout = tf.placeholder(tf.float32)
if dropout == 1.0:
keep_prob = tf.constant([0.2, 0.3, 0.5], dtype=tf.float32)
else:
keep_prob = tf.constant([1.0, 1.0, 1.0], dtype=tf.float32)
weights_1 = weight_variable([cnn1,cnn1,3, hidden_size[0]])
biases_1 = bias_variable([hidden_size[0]])
weights_2 = weight_variable([cnn2,cnn2,hidden_size[0], hidden_size[1]])
biases_2 = bias_variable([hidden_size[1]])
weights_3 = weight_variable([cnn3,cnn3,hidden_size[1], hidden_size[2]])
biases_3 = bias_variable([hidden_size[2]])
weights_4 = weight_variable([56 * 56 * hidden_size[2], hidden_size[3]])
biases_4 = bias_variable([hidden_size[3]])
weights_5 = weight_variable([hidden_size[3], num_labels])
biases_5 = bias_variable([num_labels])
def model(data):
with tf.device('/gpu:1'):
train_hidden_1 = tf.nn.relu(conv2d(data, weights_1) + biases_1)
train_hidden_2 = max_pool_2x2(tf.nn.relu(conv2d(train_hidden_1, weights_2) + biases_2))
train_hidden_2_drop = tf.nn.dropout(train_hidden_2, keep_prob[0])
train_hidden_3 = max_pool_2x2(tf.nn.relu(conv2d(train_hidden_2_drop, weights_3) + biases_3))
train_hidden_3_drop = tf.nn.dropout(train_hidden_3, keep_prob[1])
train_hidden_3_drop = tf.reshape(train_hidden_3_drop,[-1, 56 * 56 * hidden_size[2]])
train_hidden_4 = tf.nn.relu(tf.matmul(train_hidden_3_drop, weights_4) + biases_4)
train_hidden_4_drop = tf.nn.dropout(train_hidden_4, keep_prob[2])
logits = tf.matmul(train_hidden_4_drop, weights_5) + biases_5
return logits
t_train_labels = tf.argmax(y_, 1) # From one-hot (one and zeros) vectors to values
loss = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(logits=model(x), labels=t_train_labels))
optimizer = tf.train.AdamOptimizer(0.01).minimize(loss)
y = tf.nn.softmax(model(x))
### RUNNING
print("log address: %s" % (log_addr))
#num_steps = 10001
times_repeat = 20 # number of epochs
batch_size = 100
with tf.Session(graph=graphCNN,config=tf.ConfigProto(log_device_placement=True)) as session:
tf.initialize_all_variables().run()
saver = tf.train.Saver(max_to_keep=0)
writemyfile("---ARCHIVES_M1----")
mytime = strftime("%Y-%m-%d %H:%M:%S", time.localtime())
writemyfile(str("\nTime: %s \nLayers: %d,%d,%d \epochs: %d" % (mytime,cnn1,cnn2,cnn3,times_repeat)))
writemyfile("Train files:" + str(listfiles_train))
writemyfile("Valid files:" + str(listfiles_valid))
writemyfile("Test files:" + str(listfiles_test))
print("Model will be saved in file: %s" % save_model_path)
writemyfile(str("Model will be saved in file: %s" % save_model_path))
### TRAINING & VALIDATION
valid_accuracies_epochs = np.array([])
for time_repeat in range(times_repeat):
print("- time_repeat:",time_repeat)
writemyfile("- time_repeat:"+str(time_repeat))
for file_train in listfiles_train:
file_train_id = int(file_train[0:len(file_train)-4])
time_start_this_file = time.time()
#LOAD DATA
print("- - file:",file_train_id, end=' ')
writemyfile("- - file:" + str(file_train_id))
Data_train, Label_train= PrepareData(file_train)
for step in range(0,len(Data_train)-batch_size,batch_size):
batch_data = Data_train[step:step+batch_size]
batch_labels = Label_train[step:step+batch_size]
feed_dict = {x : batch_data, y_ : batch_labels, dropout: 1.0}
opti, l, predictions = session.run([optimizer, loss, y], feed_dict=feed_dict)
train_accuracies = np.array([])
for index_tr_accu in range(0,len(Data_train)-SIZE_BATCH_VALID,SIZE_BATCH_VALID):
current_predictions = y.eval(feed_dict={x: Data_train[index_tr_accu:index_tr_accu+SIZE_BATCH_VALID],dropout: 0.0})
current_accuracy = accuracy(current_predictions, Label_train[index_tr_accu:index_tr_accu+SIZE_BATCH_VALID])
train_accuracies = np.r_[train_accuracies,current_accuracy]
train_accuracy = train_accuracies.mean()
print("batch accu: %.2f%%" %(train_accuracy),end=" | ")
writemyfile("batch accu: %.2f%%" %(train_accuracy))
time_done_this_file = time.time() - time_start_this_file
print("time: %.2fs" % (time_done_this_file))
writemyfile("time: %.2fs" % (time_done_this_file))
# save model
model_addr = save_model_path + "model335" + "-epoch-" + str(time_repeat) + ".ckpt"
save_path = saver.save(session, model_addr,) # max_to_keep default was 5
mytime = strftime("%Y-%m-%d %H:%M:%S", time.localtime())
print("epoch finished at %s \n model address: %s" % (mytime,model_addr))
writemyfile("epoch finished at %s \n model address: %s" % (mytime,model_addr))
# validation
valid_accuracies = np.array([])
for file_valid in listfiles_valid:
file_valid_id = int(file_valid[0:len(file_valid)-4])
Data_valid, Label_valid = PrepareData(file_valid)
for index_vl_accu in range(0,len(Data_valid)-SIZE_BATCH_VALID,SIZE_BATCH_VALID):
current_predictions = y.eval(feed_dict={x: Data_valid[index_vl_accu:index_vl_accu+SIZE_BATCH_VALID],dropout: 0.0})
current_accuracy = accuracy(current_predictions, Label_valid[index_vl_accu:index_vl_accu+SIZE_BATCH_VALID])
valid_accuracies = np.r_[valid_accuracies,current_accuracy]
valid_accuracy = valid_accuracies.mean()
print("epoch %d - valid accu: %.2f%%" %(time_repeat,valid_accuracy))
writemyfile("epoch %d - valid accu: %.2f%%" %(time_repeat,valid_accuracy))
valid_accuracies_epochs = np.hstack([valid_accuracies_epochs,valid_accuracy])
print('Done!!')
writemyfile(str('Done!!'))
session.close()
Update: I found cifar10_multi_gpu_train.py seems to be a good example for training with multi GPUs, but honestly I don't know how to apply on my case.
I think you need to change
def model(data):
with tf.device('/gpu:1'):
to:
def model(data):
for d in ['/gpu:0', '/gpu:1']:
with tf.device(d):
and ditch the line with tf.device('/gpu:0'):
Since at the first with tf.device... you are only doing initiation
of variables and then you are resetting your devices with the next with tf.device.
Let me know if this works since I can't test it.
My RNN for language modelling is predicting only "the" "and" and "unknown" what's wrong with my code?
Here I define the hyper parameters:
num_epochs = 300
total_series_length = len(uniqueSentence) - 4
truncated_backprop_length = 30
state_size = 100
num_classes = NUM_MEANINGFUL + 1
echo_step = 1
batch_size = 32
vocab_length = len(decoder)
num_batches = total_series_length//batch_size//truncated_backprop_length
learning_rate = 0.01
old_perplexity = 0
Here I generate the data (my input is given by word embeddings long 100 calculated with Word2Vec):
def generateData():
uniqueSent = uniqueSentence[0 : len(uniqueSentence) - 4]
x_tr = np.array([model_ted[word] for words in uniqueSent])
#Roll array elements along a given axis.
#Elements that roll beyond the last position are re-introduced at the first.
x_tr = x_tr.reshape((100, batch_size, -1)) # The first index changing slowest, subseries as rows
x = x_tr.transpose((1, 2, 0))
print("hi")
new_y = indexList[1: len(indexList)- 4]
new_y.append(indexList[len(indexList)-3])
y = np.array(new_y)
print(len(y))
y = y.reshape((batch_size, -1))
return (x, y)
Define the placeholders:
batchX_placeholder = tf.placeholder(tf.float32, [batch_size, truncated_backprop_length, 100])
batchY_placeholder = tf.placeholder(tf.int32, [batch_size, truncated_backprop_length])
W = tf.Variable(np.random.rand(state_size, num_classes),dtype=tf.float32)
b = tf.Variable(np.zeros((batch_size, num_classes)), dtype=tf.float32)
W2 = tf.Variable(np.random.rand(state_size, num_classes),dtype=tf.float32)
b2 = tf.Variable(np.zeros((batch_size, num_classes)), dtype=tf.float32)
Inputs and desired outputs:
labels_series = tf.transpose(batchY_placeholder)
labels_series = tf.unstack(batchY_placeholder, axis=1)
inputs_series = batchX_placeholder
Forward pass:
from tensorflow.contrib.rnn.python.ops import core_rnn_cell_impl
print(tf.__version__)
#cell = tf.contrib.rnn.BasicRNNCell(state_size)
cell = tf.contrib.rnn.BasicLSTMCell(state_size, state_is_tuple = False)
print(cell.state_size)
init_state = tf.zeros([batch_size, cell.state_size])
outputs, current_state = tf.nn.dynamic_rnn(cell, inputs_series, initial_state = init_state)
iterable_outputs = tf.unstack(outputs, axis = 1)
logits_series = [tf.matmul(state, W2) + b2 for state in iterable_outputs] #Broadcasted addition
predictions_series = [tf.nn.softmax(logits) for logits in logits_series]
losses = [tf.losses.sparse_softmax_cross_entropy(labels, logits)
for logits, labels in zip(logits_series, labels_series)]
total_loss = tf.add_n(losses)
train_step = tf.train.AdamOptimizer(learning_rate).minimize(total_loss)
x,y = generateData()
del(model_ted)
Training:
with tf.Session(config=tf.ConfigProto(log_device_placement=True)) as sess:
sess.run(tf.initialize_all_variables())
loss_list = []
print("start")
_current_state = np.zeros((batch_size, 2*state_size))
#avevo genrateData fuori e -currentstate dentro
for epoch_idx in range(num_epochs):
print("New data, epoch", epoch_idx)
for batch_idx in range(num_batches):
start_idx = batch_idx * truncated_backprop_length
end_idx = start_idx + truncated_backprop_length
batchX = x[:,start_idx:end_idx,:]
batchY = y[:,start_idx:end_idx]
_total_loss, _train_step, _current_state, _predictions_series = sess.run(
[total_loss, train_step, current_state, predictions_series],
feed_dict={
batchX_placeholder:batchX,
batchY_placeholder:batchY,
init_state:_current_state
})
loss_list.append(_total_loss)
del(batchX)
del(batchY)
perplexity = 2 ** (_total_loss/truncated_backprop_length )
print(perplexity)
del(perplexity)
_predictions_series = np.array(_predictions_series)
pr = _predictions_series.transpose([1, 0, 2])
pr_ind = []
for line in pr[0]:
pr_ind.append(np.argmax(line))
for index in pr_ind:
print(decoder[index], end = " " )
del(pr_ind)
print("\n learning rate: ", end = " ")
print(learning_rate)