Related
I am learning neural network modeling and its uses in time series prediction.
First, thank you for reading this post and for your help :)
On this page there are various NN models (LSTM, CNN etc.) for predicting "traffic volume":
https://michael-fuchs-python.netlify.app/2020/11/01/time-series-analysis-neural-networks-for-forecasting-univariate-variables/#train-validation-split
I got inspired and decided to use/shorten/adapt the code in there for a problem of my own: predicting the bitcoin price.
I have the bitcoin daily prices starting 1.1.2017
in total 2024 daily prices
I use the first 85% of the data for the training data, and the rest as the validation (except the last 10 observation, which I would like to use as test data to see how good my model is)
I would like to use a Feedforward model
My goal is merely having a code that runs.
I have managed so far to have most of my code run. However, I get a strange format for my test forecast results: It should be simply an array of 10 numbers (i.e. predicted prices corresponding to the 10 day at the end of my data). To my surprise what is printed out is a long list of numbers. I need help to find out what changes I need to make to make to the code to make it run.
Thank you for helping me :)
The code is pasted down there, followed by the error:
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from sklearn import preprocessing #import MinMaxScaler
from sklearn import metrics #import mean_squared_error
import seaborn as sns
sns.set()
import tensorflow as tf
from tensorflow import keras
from keras.layers import Input, Dense, Flatten
from keras.optimizers import Adam
from keras.models import Sequential
from keras.callbacks import EarlyStopping
tf.__version__
df = pd.read_csv('/content/BTC-USD.csv')
def mean_absolute_percentage_error_func(y_true, y_pred):
y_true, y_pred = np.array(y_true), np.array(y_pred)
return np.mean(np.abs((y_true - y_pred) / y_true)) * 100
def timeseries_evaluation_metrics_func(y_true, y_pred):
print('Evaluation metric results: ')
print(f'MSE is : {metrics.mean_squared_error(y_true, y_pred)}')
print(f'MAE is : {metrics.mean_absolute_error(y_true, y_pred)}')
print(f'RMSE is : {np.sqrt(metrics.mean_squared_error(y_true, y_pred))}')
print(f'MAPE is : {mean_absolute_percentage_error_func(y_true, y_pred)}')
print(f'R2 is : {metrics.r2_score(y_true, y_pred)}',end='\n\n')
def univariate_data_prep_func(dataset, start, end, window, horizon):
X = []
y = []
start = start + window
if end is None:
end = len(dataset) - horizon
for i in range(start, end):
indicesx = range(i-window, i)
X.append(np.reshape(dataset[indicesx], (window, 1)))
indicesy = range(i,i+horizon)
y.append(dataset[indicesy])
return np.array(X), np.array(y)
# Generating the test set
test_data = df['close'].tail(10)
df = df.drop(df['close'].tail(10).index)
df.shape
# Defining the target variable
uni_data = df['close']
uni_data.index = df['formatted_date']
uni_data.head()
#scaling
from sklearn import preprocessing
uni_data = uni_data.values
scaler_x = preprocessing.MinMaxScaler()
x_scaled = scaler_x.fit_transform(uni_data.reshape(-1, 1))
# Single Step Style (sss) modeling
univar_hist_window_sss = 50
horizon_sss = 1
# 2014 observations in total
# 2014*0.85=1710 should be part of the training (304 validation)
train_split_sss = 1710
x_train_uni_sss, y_train_uni_sss = univariate_data_prep_func(x_scaled, 0, train_split_sss,
univar_hist_window_sss, horizon_sss)
x_val_uni_sss, y_val_uni_sss = univariate_data_prep_func(x_scaled, train_split_sss, None,
univar_hist_window_sss, horizon_sss)
print ('Length of first Single Window:')
print (len(x_train_uni_sss[0]))
print()
print ('Target horizon:')
print (y_train_uni_sss[0])
BATCH_SIZE_sss = 32
BUFFER_SIZE_sss = 150
train_univariate_sss = tf.data.Dataset.from_tensor_slices((x_train_uni_sss, y_train_uni_sss))
train_univariate_sss = train_univariate_sss.cache().shuffle(BUFFER_SIZE_sss).batch(BATCH_SIZE_sss).repeat()
validation_univariate_sss = tf.data.Dataset.from_tensor_slices((x_val_uni_sss, y_val_uni_sss))
validation_univariate_sss = validation_univariate_sss.batch(BATCH_SIZE_sss).repeat()
n_steps_per_epoch = 55
n_validation_steps = 10
n_epochs = 100
#FFNN architecture
model = tf.keras.models.Sequential([
tf.keras.layers.Dense(8, input_shape=x_train_uni_sss.shape[-2:]),
tf.keras.layers.Dense(units=horizon_sss)])
model.compile(loss='mse',
optimizer='adam')
#fit the model
model_path = '/content/FFNN_model_sss.h5'
keras_callbacks = [tf.keras.callbacks.EarlyStopping(monitor='val_loss',
min_delta=0, patience=10,
verbose=1, mode='min'),
tf.keras.callbacks.ModelCheckpoint(model_path,monitor='val_loss',
save_best_only=True,
mode='min', verbose=0)]
history = model.fit(train_univariate_sss, epochs=n_epochs, steps_per_epoch=n_steps_per_epoch,
validation_data=validation_univariate_sss, validation_steps=n_validation_steps, verbose =1,
callbacks = keras_callbacks)
#validation
loss = history.history['loss']
val_loss = history.history['val_loss']
epochs = range(1, len(loss) + 1)
plt.plot(epochs, loss, 'r', label='Training loss')
plt.plot(epochs, val_loss, 'b', label='Validation loss')
plt.title('Training and validation loss')
plt.legend()
plt.show()
# Testing our model
trained_ffnn_model_sss = tf.keras.models.load_model(model_path)
df_temp = df['close']
test_horizon = df_temp.tail(univar_hist_window_sss)
test_history = test_horizon.values
result = []
# Define Forecast length here
window_len = len(test_data)
test_scaled = scaler_x.fit_transform(test_history.reshape(-1, 1))
for i in range(1, window_len+1):
test_scaled = test_scaled.reshape((1, test_scaled.shape[0], 1))
# Inserting the model
predicted_results = trained_ffnn_model_sss.predict(test_scaled)
print(f'predicted : {predicted_results}')
result.append(predicted_results[0])
test_scaled = np.append(test_scaled[:,1:],[[predicted_results]])
result_inv_trans = scaler_x.inverse_transform(result)
result_inv_trans
I believe the problem might have to do with the shapes of data. How exactly I do not yet know.
Data:
click here
Traceback:
click here
I'm trying to build a Pytorch network for image captioning.
Currently I have a working network of Encoder and Decoder, and I want to add nn.MultiheadAttnetion layer to it (to be used as self attention).
Currently my decode looks like this:
class Decoder(nn.Module):
def __init__(self, hidden_size, embed_dim, vocab_size, layers = 1):
super(Decoder, self).__init__()
self.embed_dim = embed_dim
self.vocab_size = vocab_size
self.layers = layers
self.hidden_size = hidden_size
self.embedding = nn.Embedding(vocab_size, embed_dim, padding_idx=0)
self.lstm = nn.LSTM(input_size = embed_dim, hidden_size = hidden_size, batch_first = True, num_layers = layers)
#self.attention = nn.MultiheadAttention(hidden_size, num_heads=1, batch_first= True)
self.fc = nn.Linear(hidden_size, self.vocab_size)
def init_hidden(self, batch_size):
h = torch.zeros(self.layers, batch_size, self.hidden_size).to(device)
c = torch.zeros(self.layers, batch_size, self.hidden_size).to(device)
return h,c
def forward(self, features, caption):
batch_size = caption.size(0)
caption_size = caption.size(1)
h,c = self.init_hidden(batch_size)
embeddings = self.embedding(caption)
lstm_input = torch.cat((features.unsqueeze(1), embeddings[:,:-1,:]), dim=1)
output, (h,c) = self.lstm(lstm_input, (h,c))
#output, _ = self.attention(output, output, output)
output = self.fc(output)
return output
def generate_caption(self, features, max_caption_size = MAX_LEN):
h,c = self.init_hidden(1)
caption = ""
embeddings = features.unsqueeze(1)
for i in range(max_caption_size):
output, (h, c) = self.lstm(embeddings, (h,c))
#output, _ = self.attention(output, output, output)
output = self.fc(output)
_, word_index = torch.max(output, dim=2) # take the word with highest probability
if word_index == vocab.get_index(END_WORD):
break
caption += vocab.get_word(word_index) + " "
embeddings = self.embedding(torch.LongTensor([word_index]).view(1,-1).to(device))
return caption
and it gives relatively good results for image captioning.
I want to add the commented out lines so the model will use Attention. But- when I do that- the model breaks, although the loss becomes extremely low (decreasing from 2.7 to 0.2 during training instead of 2.7 to 1 without the attention) - the caption generation is not really working (predicts the same word over and over again).
My questions are:
Am I using the nn.MultiheadAttention correctly? it is very weird to me that it should be used after the LSTM, but I saw this online, and it works from dimension sizes perspective
Any idea why my model breaks when I use Attention?
EDIT: I also tried to put the Attention before the LSTM, and it didn't work as well (network predicted the same caption for every picture)
I want to train a model in distributed system. I have found a code in github for distributed training where the worker node send gradient to the parameter server and the parameter server sends the average gradient to the workers. But in client/worker side code, i couldn't understand where the received gradient updates the weights and biases.
Here is client/worker side the code, it receives initial gradients from the parameter server and then calculates loss, gradients and sends the gradient value to the server again.
from __future__ import division
from __future__ import print_function
import numpy as np
import sys
import pickle as pickle
import socket
from datetime import datetime
import time
import tensorflow as tf
import cifar10
TCP_IP = 'some IP'
TCP_PORT = 5014
port = 0
port_main = 0
s = 0
FLAGS = tf.app.flags.FLAGS
tf.app.flags.DEFINE_string('train_dir', '/home/ubuntu/cifar10_train',
"""Directory where to write event logs """
"""and checkpoint.""")
tf.app.flags.DEFINE_integer('max_steps', 5000,
"""Number of batches to run.""")
tf.app.flags.DEFINE_boolean('log_device_placement', False,
"""Whether to log device placement.""")
tf.app.flags.DEFINE_integer('log_frequency', 10,
"""How often to log results to the console.""")
#gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.30)
def safe_recv(size, server_socket):
data = ""
temp = ""
data = bytearray()
recv_size = 0
while 1:
try:
temp = server_socket.recv(size-len(data))
data.extend(temp)
recv_size = len(data)
if recv_size >= size:
break
except:
print("Error")
data = bytes(data)
return data
def train():
"""Train CIFAR-10 for a number of steps."""
g1 = tf.Graph()
with g1.as_default():
global_step = tf.Variable(-1, name='global_step',
trainable=False, dtype=tf.int32)
increment_global_step_op = tf.assign(global_step, global_step+1)
# Get images and labels for CIFAR-10.
images, labels = cifar10.distorted_inputs()
# Build a Graph that computes the logits predictions from the
# inference model.
logits = cifar10.inference(images)
# Calculate loss.
loss = cifar10.loss(logits, labels)
grads = cifar10.train_part1(loss, global_step)
only_gradients = [g for g, _ in grads]
class _LoggerHook(tf.train.SessionRunHook):
"""Logs loss and runtime."""
def begin(self):
self._step = -1
self._start_time = time.time()
def before_run(self, run_context):
self._step += 1
return tf.train.SessionRunArgs(loss) # Asks for loss value.
def after_run(self, run_context, run_values):
if self._step % FLAGS.log_frequency == 0:
current_time = time.time()
duration = current_time - self._start_time
self._start_time = current_time
loss_value = run_values.results
examples_per_sec = FLAGS.log_frequency * FLAGS.batch_size / duration
sec_per_batch = float(duration / FLAGS.log_frequency)
format_str = ('%s: step %d, loss = %.2f (%.1f examples/sec; %.3f '
'sec/batch)')
print(format_str % (datetime.now(), self._step, loss_value,
examples_per_sec, sec_per_batch))
with tf.train.MonitoredTrainingSession(
checkpoint_dir=FLAGS.train_dir,
hooks=[tf.train.StopAtStepHook(last_step=FLAGS.max_steps),
tf.train.NanTensorHook(loss),
_LoggerHook()],
config=tf.ConfigProto(
# log_device_placement=FLAGS.log_device_placement, gpu_options=gpu_options)) as mon_sess:
log_device_placement=FLAGS.log_device_placement)) as mon_sess:
global port
s = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
s.connect((TCP_IP, port_main))
recv_size = safe_recv(17, s)
recv_size = pickle.loads(recv_size)
recv_data = safe_recv(recv_size, s)
var_vals = pickle.loads(recv_data)
s.close()
feed_dict = {}
i = 0
for v in tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES):
feed_dict[v] = var_vals[i]
i = i+1
print("Received variable values from ps")
# Opening the socket and connecting to server
s = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
s.connect((TCP_IP, port))
while not mon_sess.should_stop():
gradients, step_val = mon_sess.run(
[only_gradients, increment_global_step_op], feed_dict=feed_dict)
# sending the gradients
send_data = pickle.dumps(gradients, pickle.HIGHEST_PROTOCOL)
to_send_size = len(send_data)
send_size = pickle.dumps(to_send_size, pickle.HIGHEST_PROTOCOL)
s.sendall(send_size)
s.sendall(send_data)
# receiving the variable values
recv_size = safe_recv(17, s)
recv_size = pickle.loads(recv_size)
recv_data = safe_recv(recv_size, s)
var_vals = pickle.loads(recv_data)
feed_dict = {}
i = 0
for v in tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES):
feed_dict[v] = var_vals[i]
i = i+1
s.close()
def main(argv=None): # pylint: disable=unused-argument
global port
global port_main
global s
if(len(sys.argv) != 3):
print("<port> <worker-id> required")
sys.exit()
port = int(sys.argv[1]) + int(sys.argv[2])
port_main = int(sys.argv[1])
print("Connecting to port ", port)
cifar10.maybe_download_and_extract()
if tf.gfile.Exists(FLAGS.train_dir):
tf.gfile.DeleteRecursively(FLAGS.train_dir)
tf.gfile.MakeDirs(FLAGS.train_dir)
total_start_time = time.time()
train()
print("--- %s seconds ---" % (time.time() - total_start_time))
if __name__ == '__main__':
tf.app.run()
EDIT:
Here is the train_part1() code:
def train_part1(total_loss, global_step):
"""Train CIFAR-10 model.
Create an optimizer and apply to all trainable variables. Add moving
average for all trainable variables.
Args:
total_loss: Total loss from loss().
global_step: Integer Variable counting the number of training steps
processed.
Returns:
train_op: op for training.
"""
# Variables that affect learning rate.
num_batches_per_epoch = NUM_EXAMPLES_PER_EPOCH_FOR_TRAIN / FLAGS.batch_size
decay_steps = int(num_batches_per_epoch * NUM_EPOCHS_PER_DECAY)
# Decay the learning rate exponentially based on the number of steps.
lr = tf.train.exponential_decay(INITIAL_LEARNING_RATE,
global_step,
decay_steps,
LEARNING_RATE_DECAY_FACTOR,
staircase=True)
tf.summary.scalar('learning_rate', lr)
# Generate moving averages of all losses and associated summaries.
loss_averages_op = _add_loss_summaries(total_loss)
# Compute gradients.
with tf.control_dependencies([loss_averages_op]):
opt = tf.train.GradientDescentOptimizer(lr)
grads = opt.compute_gradients(total_loss)
return grads
To me it seems that line
gradients, step_val = mon_sess.run(
[only_gradients, increment_global_step_op], feed_dict=feed_dict)
receieves new values for variables in feed_dict, assign these values to variables, and makes a training step, during which it only calculates and returns the gradients, that are later sent to the parameter server. I would expect cifar10.train_part1 (the one that returns only_gradients) to depend on variable values and define the update.
Update: I looked into the code and changed my mind. Had to google and found next answer that shed some light on what is happening.
Gradients are actually not applied in this code anywhere implicitly. Instead, gradients are sent to the parameter server, parameter server averages gradients and applies them to weights, it returns the weights to the local worker, * recieved weights are used instead of local weights during session run through feed_dict* i.e. local weights are never actually updated and do not actually matter at all. The key, is that feed_dict allows to rewrite any tensor output of the session run and this code rewrites variables.
I am working on Keras in Python and I have a neural network (see code below).
Currently it works with only a ReLu activation.
For experimental reasons I would like to have some neurons on ReLu and some on softmax (or any other activation function). for example in a Layer with 20 neurons, I would like to have 10 with ReLu and 10 with Softmax.
I have tried some different ways, but always failed to get an output.
Would you know how I should do this?
# - Libraries
from keras.layers import Dense
from keras.models import Sequential
from keras.callbacks import EarlyStopping
early_spotting_monitor = EarlyStopping(patience=2)
layers = 4
neurons = 20
act = "ReLu"
# - Create Neural Network
model = Sequential()
model.add(Dense(neurons,activation=act,input_dim=X_train.shape[1]))
layers -= 1
while layers > 0:
model.add(Dense(neurons,activation=act))
layers -= 1
model.add(Dense(n_months))
model.compile(optimizer="adam",loss="mean_absolute_error")
model.fit(X_train,Y_train,validation_split=0.10,epochs=13,callbacks=[early_spotting_monitor])
EDIT: this is my (working) code now:
# - Libraries
from keras.callbacks import EarlyStopping
early_spotting_monitor = EarlyStopping(patience=2)
from keras.layers import Input, Dense
from keras.models import Model
from keras.layers.merge import concatenate
# input layer
visible = Input(shape=(X_train.shape[1],))
hidden11 = Dense(14, activation='relu')(visible)
hidden12 = Dense(3, activation='softplus')(visible)
hidden13 = Dense(2, activation='linear')(visible)
hidden13 = Dense(2, activation='selu')(visible)
merge1 = concatenate([hidden11, hidden12, hidden13])
hidden21 = Dense(14, activation='relu')(merge1)
hidden22 = Dense(3, activation='softplus')(merge1)
hidden23 = Dense(2, activation='linear')(merge1)
hidden13 = Dense(2, activation='selu')(visible)
merge2 = concatenate([hidden21, hidden22, hidden23])
hidden3 = Dense(20, activation='relu')(merge2)
output = Dense(Y_train.shape[1],activation="linear")(hidden3)
model = Model(inputs=visible, outputs=output)
model.compile(optimizer="adam",loss="mean_absolute_error")
model.fit(X_train,Y_train,validation_split=0.10,epochs=13,callbacks=[early_spotting_monitor]) # starts training
return model
You have to use the Functional API to do this, for example:
input = Input(shape = (X_train.shape[1]))
branchA = Dense(neuronsA, activation = "relu")(input)
branchB = Dense(neuronsB, activation = "sigmoid")(input)
out = concatenate([branchA, branchB])
You cannot do it with the Sequential API, so I recommend you move your code to the functional API.
So this is something I have been trying to do recently and so far this is what I have done. I think it's working, but I would appreciate if anyone tells me what I'm doing wrong here. I'm doing this only on the output layer and my output layer has two units:
def activations(l):
l_0 = tf.keras.activations.exponential(l[...,0])
l_1 = tf.keras.activations.elu(l[...,1])
lnew = tf.stack([l_0, l_1], axis = 1)
return lnew
model = tf.keras.Sequential([..., Dense(2, activation = activations)])
I followed the tutorial about Reading data with TF and made some tries myself. Now, the problem is that my tests show duplicate data in the batches I created when reading data from a CSV.
My code looks like this:
# -*- coding: utf-8 -*-
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import os
import collections
import numpy as np
from six.moves import xrange # pylint: disable=redefined-builtin
import tensorflow as tf
class XICSDataSet:
def __init__(self, height=20, width=195, batch_size=1000, noutput=15):
self.depth = 1
self.height = height
self.width = width
self.batch_size = batch_size
self.noutput = noutput
def trainingset_files_reader(self, data_dir, nfiles):
fnames = [os.path.join(data_dir, "test%d"%i) for i in range(nfiles)]
filename_queue = tf.train.string_input_producer(fnames, shuffle=False)
reader = tf.TextLineReader()
key, value = reader.read(filename_queue)
record_defaults = [[.0],[.0],[.0],[.0],[.0]]
data_tuple = tf.decode_csv(value, record_defaults=record_defaults, field_delim = ' ')
features = tf.pack(data_tuple[:-self.noutput])
label = tf.pack(data_tuple[-self.noutput:])
depth_major = tf.reshape(features, [self.height, self.width, self.depth])
min_after_dequeue = 100
capacity = min_after_dequeue + 30 * self.batch_size
example_batch, label_batch = tf.train.shuffle_batch([depth_major, label], batch_size=self.batch_size, capacity=capacity,
min_after_dequeue=min_after_dequeue)
return example_batch, label_batch
with tf.Graph().as_default():
ds = XICSDataSet(2, 2, 3, 1)
im, lb = ds.trainingset_files_reader(filename, 1)
sess = tf.Session()
init = tf.initialize_all_variables()
sess.run(init)
tf.train.start_queue_runners(sess=sess)
for i in range(1000):
lbs = sess.run([im, lb])[1]
_, nu = np.unique(lbs, return_counts=True)
if np.array_equal(nu, np.array([1, 1, 1])) == False:
print('Not unique elements found in a batch!')
print(lbs)
I tried with different batch sizes, different number of files, different values of capacity and min_after_dequeue, but I always get the problem. In the end, I would like to be able to read data from only one file, creating batches and shuffling the examples.
My files, created ad hoc for this test, have 5 lines each representing samples, and 5 columns. The last column is meant to be the label for that sample. These are just random numbers. I'm using only 10 files just to test this out.
The default behavior for tf.train.string_input_producer(fnames) is to produce an infinite number of copies of the elements in fnames. Therefore, since your tf.train.shuffle_batch() capacity is larger than the total number of elements in your input files (5 elements per file * 10 files = 50 elements), and the min_after_dequeue is also larger than the number of elements, the queue will contain at least two full copies of the input data before the first batch is produced. As a result, it is likely that some batches will contain duplicate data.
If you only want to process each example once, you can set an explicit num_epochs=1 when creating the tf.train.string_input_producer(). For example:
def trainingset_files_reader(self, data_dir, nfiles):
fnames = [os.path.join(data_dir, "test%d" % i) for i in range(nfiles)]
filename_queue = tf.train.string_input_producer(
fnames, shuffle=False, num_epochs=1)
# ...