copy construct from a tensor: USER WARNING - neural-network

I am creating a random tensor from normal distribution and since this tensor is served as the weight in the NN, to add requires_grad attributes, I use torch.tensor() as below:
import torch
input_dim, hidden_dim = 3, 5
norm = torch.distributions.normal.Normal(loc=0, scale=0.01)
W = norm.sample((input_dim, hidden_dim))
W = torch.tensor(W, requires_grad=True)
I am getting user warning error as below:
UserWarning: To copy construct from a tensor,
it is recommended to use sourceTensor.clone().detach() or
sourceTensor.clone().detach().requires_grad_(True),
rather than torch.tensor(sourceTensor).
Is there an alternative way to achieve the above? Thanks

You can just set W.requires_grad to True
import torch
input_dim, hidden_dim = 3, 5
norm = torch.distributions.normal.Normal(loc=0, scale=0.01)
W = norm.sample((input_dim, hidden_dim))
W.requires_grad = True

Related

Seemingly inconsistent tensor sizes in pytorch

I'm building a convolutional autoencoder, but want the encoding to be in a linear form so I can more easily feed it as input into an MLP. I have two convolutional layers on the encoder along with a linear inner layer to reduce dimension. This encoding is then fed into the corresponding decoder.
When I flatten the output of the second convolutional layer, based on my calculation (using the standard formula: Calculate the Output size in Convolution layer) should come out to a 1x100352 rank 1 tensor. However, when I set the input dimension of the linear layer to be 100352, the flattened rank 1 tensor has dimension 1x50176. Then comes the weird part.
I tried changing the input dimension of the linear layer to be 50176, assuming I had miscalculated. When I do this, the reshaped rank 1 tensor confusingly becomes 1x100352, and then the aforementioned weight matrix becomes 50176x256 as expected.
This response to modifying the linear layer's input dimension doesn't make sense to me. That hyperparameter controls the weight matrix correctly, but I guess I'm uncertain why it has any bearing on the linear layer's input since that's just a reshaped tensor output from a convolutional layer whose hyperparameters are unrelated to the hyperparameter in question.
I apologize if I'm just missing something obvious. I'm very new to pytorch, and I couldn't find any other posts which discussed this sort of issue.
Here's what I believe to be the minimal reproducible example:
import os
import torch
import torchvision
import torchvision.transforms as transforms
import matplotlib.pyplot as plt
import torch.autograd as autograd
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torchvision import datasets
from torch.utils.data import DataLoader
from torchvision.utils import save_image
class convAutoEncoder(nn.Module):
def __init__(self,**kwargs):
super().__init__()
#Creating network structure
#Encoder portion of autoencoder
self.enc1 = nn.Conv2d(in_channels = kwargs["inputChannels"], out_channels = kwargs["channelsEncoderMid"], kernel_size = kwargs["kernelSize"])
self.enc2 = nn.Conv2d(in_channels = kwargs["channelsEncoderMid"], out_channels = kwargs["channelsEncoderInner"], kernel_size = kwargs["kernelSize"])
self.enc3 = nn.Linear(in_features = kwargs["intoLinear"], out_features = kwargs["linearEncoded"])
#Decoder portion of autoencoder
self.dec1 = nn.Linear(in_features = kwargs["linearEncoded"], out_features = kwargs["intoLinear"])
self.dec2 = nn.ConvTranspose2d(in_channels = kwargs["channelsEncoderInner"], out_channels = kwargs["channelsDecoderMid"], kernel_size = kwargs["kernelSize"])
self.dec3 = nn.ConvTranspose2d(in_channels = kwargs["channelsDecoderMid"], out_channels = kwargs["inputChannels"], kernel_size = kwargs["kernelSize"])
def forward(self,x):
#Encoding
x = F.relu(self.enc1(x))
x = F.relu(self.enc2(x))
x = x.reshape(1,-1)
x = x.squeeze()
x = F.relu(self.enc3(x))
#Decoding
x = F.relu(self.dec1(x))
x = x.reshape([32,4,28,28])
x = F.relu(self.dec2(x))
x = F.relu(self.dec3(x))
return x
def encodeDecodeConv(numEpochs = 20, input_Channels = 3, batchSize = 32,
channels_Encoder_Inner = 4, channels_Encoder_Mid = 8, into_Linear = 100352,
linear_Encoded = 256, channels_Decoder_Mid = 8, kernel_Size = 3,
learningRate = 1e-3):
#Pick a device. If GPU available, use that. Otherwise, use CPU.
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
#Define data transforms
transform = transforms.Compose([transforms.ToTensor(),transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])
#Define training dataset
trainSet = datasets.CIFAR10(root = './data', train = True, download = True, transform = transform)
#Define testing dataset
testSet = datasets.CIFAR10(root = './data', train = False, download = True, transform = transform)
#Define data loaders
trainLoader = DataLoader(trainSet, batch_size = batchSize, shuffle = True)
testLoader = DataLoader(testSet, batch_size = batchSize, shuffle = True)
#Initialize neural network
model = convAutoEncoder(inputChannels = input_Channels, channelsEncoderMid = channels_Encoder_Mid, channelsEncoderInner = channels_Encoder_Inner, intoLinear = into_Linear, linearEncoded = linear_Encoded, channelsDecoderMid = channels_Decoder_Mid, kernelSize = kernel_Size)
#Optimization setup
criterion = nn.MSELoss()
optimizer = optim.Adam(model.parameters(),lr = learningRate)
lossTracker = []
for epoch in range(numEpochs):
loss = 0
for data,_ in trainLoader:
data = data.to(device)
optimizer.zero_grad()
outputs = model(data)
train_loss = criterion(outputs,data)
train_loss.backward()
optimizer.step()
loss += train_loss.item()
loss = loss/len(trainLoader)
print('Epoch {} of {}, Train loss: {:.3f}'.format(epoch+1,numEpochs,loss))
encodeDecodeConv()
Edit2: Somewhere in the CIFAR10 dataset, the data appears to change dimension. After playing around with print statements more, I discovered that setting the relevant hyperparameter to 100352 works great for many entries, but then seemingly one image pops up that has a different size. Not sure why that would occur, though.

gaussian process regression in multiple dimensions with GPflow

I would like to perform some multivariant regression using gaussian process regression as implemented in GPflow using version 2.
Installed with pip install gpflow==2.0.0rc1
Below is some example code that generates some 2D data and then attempts to fit it with using GPR and the finally computes the difference
between the true input data and the GPR prediction.
Eventually I would like to extend to higher dimensions
and do tests against a validation set to check for over-fitting
and experiment with other kernels and "Automatic Relevance Determination"
but understanding how to get this to work is the first step.
Thanks!
The following code snippet will work in a jupyter notebook.
import gpflow
import numpy as np
import matplotlib
from gpflow.utilities import print_summary
%matplotlib inline
matplotlib.rcParams['figure.figsize'] = (12, 6)
import matplotlib.pyplot as plt
from mpl_toolkits.mplot3d import Axes3D
def gen_data(X, Y):
"""
make some fake data.
X, Y are np.ndarrays with shape (N,) where
N is the number of samples.
"""
ys = []
for x0, x1 in zip(X,Y):
y = x0 * np.sin(x0*10)
y = x1 * np.sin(x0*10)
y += 1
ys.append(y)
return np.array(ys)
# generate some fake data
x = np.linspace(0, 1, 20)
X, Y = np.meshgrid(x, x)
X = X.ravel()
Y = Y.ravel()
z = gen_data(X, Y)
#note X.shape, Y.shape and z.shape
#are all (400,) for this case.
# if you would like to plot the data you can do the following
fig = plt.figure()
ax = Axes3D(fig)
ax.scatter(X, Y, z, s=100, c='k')
# had to set this
# to avoid the following error
# tensorflow.python.framework.errors_impl.InvalidArgumentError: Cholesky decomposition was not successful. The input might not be valid. [Op:Cholesky]
gpflow.config.set_default_positive_minimum(1e-7)
# setup the kernel
k = gpflow.kernels.Matern52()
# set up GPR model
# I think the shape of the independent data
# should be (400, 2) for this case
XY = np.column_stack([[X, Y]]).T
print(XY.shape) # this will be (400, 2)
m = gpflow.models.GPR(data=(XY, z), kernel=k, mean_function=None)
# optimise hyper-parameters
opt = gpflow.optimizers.Scipy()
def objective_closure():
return - m.log_marginal_likelihood()
opt_logs = opt.minimize(objective_closure,
m.trainable_variables,
options=dict(maxiter=100)
)
# predict training set
mean, var = m.predict_f(XY)
print(mean.numpy().shape)
# (400, 400)
# I would expect this to be (400,)
# If it was then I could compute the difference
# between the true data and the GPR prediction
# `diff = mean - z`
# but because the shape is not as expected this of course
# won't work.
The shape of z must be (N, 1), whereas in your case it is (N,). However, this is a missing check in GPflow and not your fault.

GP + Tensorflow training

I am trying to train a GPR model and a tensorflow model together. The training part has no issue. But for prediction using the trained model I receive a type error in a tf.placeholder op.
pred, uncp=sess.run([my, yv], feed_dict={X:xtr})
The code is similar to the 2nd example from https://gpflow.readthedocs.io/en/master/notebooks/advanced_usage.html
import numpy as np
import tensorflow as tf
import gpflow
float_type = gpflow.settings.float_type
gpflow.reset_default_graph_and_session()
def cnn_fn(x, output_dim):
out= tf.layers.dense(inputs=x, units=output_dim, activation=tf.nn.relu)
print(out)
return out
N = 150
xtr = np.random.rand(N,1)
ytr = np.sin(12*xtr) + 0.66*np.cos(25*xtr) + np.random.randn(N,1)*0.1 + 3
xtr = np.random.rand(N,28)
print(xtr.shape, ytr.shape)
nepoch=50
gp_dim=xtr.shape[1]
print(gp_dim)
minibatch_size = 16
X = tf.placeholder(tf.float32, [None, gp_dim])
Y = tf.placeholder(tf.float32, [None, 1])
with tf.variable_scope('cnn'):
f_X = tf.cast(cnn_fn(X, gp_dim), dtype=float_type)
k = gpflow.kernels.Matern52(gp_dim)
gp_model = gpflow.models.GPR(f_X, tf.cast(Y, dtype=float_type), k)
loss = -gp_model.likelihood_tensor
m, v = gp_model._build_predict(f_X)
my, yv = gp_model.likelihood.predict_mean_and_var(m, v)
with tf.variable_scope('adam'):
opt_step = tf.train.AdamOptimizer(0.001).minimize(loss)
tf_vars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='adam')
tf_vars += tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='cnn')
## initialize
sess = tf.Session()
sess.run(tf.variables_initializer(var_list=tf_vars))
gp_model.initialize(session=sess)
for i in range(nepoch):
shind=np.array(range(len(xtr)))
np.random.shuffle(shind)
for j in range(int(len(xtr)/minibatch_size)):
ind=shind[j*minibatch_size: (j+1)*minibatch_size]
sess.run(opt_step, feed_dict={X:xtr[ind], Y:ytr[ind]})
Executing the code above runs fine. But adding the following line gives an error:
pred, uncp=sess.run([my, yv], feed_dict={X:xtr})
with the following error:
<ipython-input-25-269715087df2> in <module>
----> 1 pred, uncp=sess.run([my, yv], feed_dict={X:xtr})
[...]
InvalidArgumentError: You must feed a value for placeholder tensor 'Placeholder_1' with dtype float and shape [?,1]
[[node Placeholder_1 (defined at <ipython-input-24-39ccf45cd248>:2) = Placeholder[dtype=DT_FLOAT, shape=[?,1], _device="/job:localhost/replica:0/task:0/device:GPU:0"]()]]
The reason your code fails is because you are not actually feeding in a value to one of the placeholders. This is easier to spot if you actually give them names:
X = tf.placeholder(tf.float32, [None, gp_dim], name='X')
Y = tf.placeholder(tf.float32, [None, 1], name='Y')
Tensorflow requires the entire compute graph to be well-defined, and the GPR model you are using depends on both X and Y. If you run the following line, it works fine:
pred, uncp = sess.run([my, yv], feed_dict={X: xtr, Y: ytr})
Update: as user1018464 pointed out in the comment, you are using the GPR model, in which the predictions directly depend on the training data (e.g. see equations (2.22) and (2.23) on page 16 of http://www.gaussianprocess.org/gpml/chapters/RW2.pdf). Hence you will need to pass in both xtr and ytr to make predictions.
Other models such as SVGP represent the function through "inducing features", commonly "pseudo input/output" pairs that summarise the data, in which case you won't need to feed in the original input values at all (I got it wrong when I first answered).
You could set up the model as follows:
gp_model = gpflow.models.SVGP(f_X, tf.cast(Y, dtype=float_type), k, gpflow.likelihoods.Gaussian(), xtr.copy(), num_data=N)
Then pred, uncp=sess.run([my, yv], feed_dict={X:xtr}) works as expected.
If you want to predict at different points Xtest, you need to set up a separate placeholder, and reuse the cnn (note the reuse=True in the variable_scope with the same name), as in example 2 of the notebook:
Xtest = tf.placeholder(tf.float32, [None, Mnist.input_dim], name='Xtest')
with tf.variable_scope('cnn', reuse=True):
f_Xtest = tf.cast(cnn_fn(Xtest, gp_dim), dtype=float_type)
Set up the model as before using f_X, but use f_Xtest in the call to _build_predict:
m, v = gp_model._build_predict(f_Xtest)
my, yv = gp_model.likelihood.predict_mean_and_var(m, v)
Now you need to pass in both X, Y, and Xtest into the session's run():
pred, uncp = sess.run([my, yv], feed_dict={X: xtr, Y: Ytr, Xtest: xtest})
where xtest is the numpy array with points at which you want to predict.
The GPflow manages TensorFlow sessions for you and you don't need to create your own TF session, when you use GPflow alone. In your case, tf.layers.dense makes
new variables, which should be initialized and I would advise to use a session which were created by GPflow. Essentially, you need to replace these lines
## initialize
sess = tf.Session()
sess.run(tf.variables_initializer(var_list=tf_vars))
gp_model.initialize(session=sess)
with:
sess = gpflow.get_default_session()
sess.run(tf.variables_initializer(var_list=tf_vars)
or wrap your code with default session context:
with tf.Session() as session:
... build and run

Merging two tensors by convolution in Keras

I'm trying to convolve two 1D tensors in Keras.
I get two inputs from other models:
x - of length 100
ker - of length 5
I would like to get the 1D convolution of x using the kernel ker.
I wrote a Lambda layer to do it:
import tensorflow as tf
def convolve1d(x):
y = tf.nn.conv1d(value=x[0], filters=x[1], padding='VALID', stride=1)
return y
x = Input(shape=(100,))
ker = Input(shape=(5,))
y = Lambda(convolve1d)([x,ker])
model = Model([x,ker], [y])
I get the following error:
ValueError: Shape must be rank 4 but is rank 3 for 'lambda_67/conv1d/Conv2D' (op: 'Conv2D') with input shapes: [?,1,100], [1,?,5].
Can anyone help me understand how to fix it?
It was much harder than I expected because Keras and Tensorflow don't expect any batch dimension in the convolution kernel so I had to write the loop over the batch dimension myself, which requires to specify batch_shape instead of just shape in the Input layer. Here it is :
import numpy as np
import tensorflow as tf
import keras
from keras import backend as K
from keras import Input, Model
from keras.layers import Lambda
def convolve1d(x):
input, kernel = x
output_list = []
if K.image_data_format() == 'channels_last':
kernel = K.expand_dims(kernel, axis=-2)
else:
kernel = K.expand_dims(kernel, axis=0)
for i in range(batch_size): # Loop over batch dimension
output_temp = tf.nn.conv1d(value=input[i:i+1, :, :],
filters=kernel[i, :, :],
padding='VALID',
stride=1)
output_list.append(output_temp)
print(K.int_shape(output_temp))
return K.concatenate(output_list, axis=0)
batch_input_shape = (1, 100, 1)
batch_kernel_shape = (1, 5, 1)
x = Input(batch_shape=batch_input_shape)
ker = Input(batch_shape=batch_kernel_shape)
y = Lambda(convolve1d)([x,ker])
model = Model([x, ker], [y])
a = np.ones(batch_input_shape)
b = np.ones(batch_kernel_shape)
c = model.predict([a, b])
In the current state :
It doesn't work for inputs (x) with multiple channels.
If you provide several filters, you get as many outputs, each being the convolution of the input with the corresponding kernel.
From given code it is difficult to point out what you mean when you say
is it possible
But if what you mean is to merge two layers and feed merged layer to convulation, yes it is possible.
x = Input(shape=(100,))
ker = Input(shape=(5,))
merged = keras.layers.concatenate([x,ker], axis=-1)
y = K.conv1d(merged, 'same')
model = Model([x,ker], y)
EDIT:
#user2179331 thanks for clarifying your intention. Now you are using Lambda Class incorrectly, that is why the error message is showing.
But what you are trying to do can be achieved using keras.backend layers.
Though be noted that when using lower level layers you will lose some higher level abstraction. E.g when using keras.backend.conv1d you need to have input shape of (BATCH_SIZE,width, channels) and kernel with shape of (kernel_size,input_channels,output_channels). So in your case let as assume the x has channels of 1(input channels ==1) and y also have the same number of channels(output channels == 1).
So your code now can be refactored as follows
from keras import backend as K
def convolve1d(x,kernel):
y = K.conv1d(x,kernel, padding='valid', strides=1,data_format="channels_last")
return y
input_channels = 1
output_channels = 1
kernel_width = 5
input_width = 100
ker = K.variable(K.random_uniform([kernel_width,input_channels,output_channels]),K.floatx())
x = Input(shape=(input_width,input_channels)
y = convolve1d(x,ker)
I guess I have understood what you mean. Given the wrong example code below:
input_signal = Input(shape=(L), name='input_signal')
input_h = Input(shape=(N), name='input_h')
faded= Lambda(lambda x: tf.nn.conv1d(input, x))(input_h)
You want to convolute each signal vector with different fading coefficients vector.
The 'conv' operation in TensorFlow, etc. tf.nn.conv1d, only support a fixed value kernel. Therefore, the code above can not run as you want.
I have no idea, too. The code you given can run normally, however, it is too complex and not efficient. In my idea, another feasible but also inefficient way is to multiply with the Toeplitz matrix whose row vector is the shifted fading coefficients vector. When the signal vector is too long, the matrix will be extremely large.

Python Neural Network Prediction

I am working on a project , I have two various numbers :
1- First number maximum value is 1500 and minimum is 200.
2- Second number maximum value is 200 and minimum is 10.
3- I want to create neural network , adding samples and train the network to predict the last number, for example :
900,67 equals 87
870,99 equals 100
1000,50 equals ?
What's type of neural networks can work with my project?
To this example, you input two values and get one.
import numpy as np
import keras
from keras import Sequential
from keras.layers import Dense
import matplotlib.pyplot as plt
MIN = np.random.rand(100)*500
MAX = np.random.rand(100)*500 + 500
x = np.concatenate((MIN.reshape(-1,1),MAX.reshape(-1,1)),axis = 1)
y = np.sin(x[:,0])*500 + np.cos(x[:,1])*500
x_max = x.max()
y_max = y.max()
x = x/x_max
y = (y-y.min())/(y_max-y.min())
model = Sequential()
model.add(Dense(200,input_dim = 2, activation = 'relu'))
model.add(Dense(100, activation = 'sigmoid'))
model.add(Dense(100, activation = 'sigmoid'))
model.add(Dense(1,activation = 'relu'))
opt = keras.optimizers.Adam(learning_rate=0.001, beta_1=0.6, beta_2=0.97, amsgrad=False)
model.compile(loss='mean_squared_error',optimizer=opt , metrics=['mse'])
model.fit(x, y, epochs=10000, batch_size=2)
y_hat = model.predict(x)
plt.figure(figsize=(10,5))
plt.plot(y)
plt.plot(y_hat.reshape(-1))
This is the result:
You will need to make pre and post processing, normalizing inputs and reescaling outputs from the neural network. This is the input:
use example:
In [10]: model.predict(np.array([0.234,0.567]).reshape(-1,2))
Out[10]: array([[0.61975896]], dtype=float32)