monk1337

torch.cuda.is_available()

#output
# False

#use of cuda
print(torch.Tensor(1,2).cuda())

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)

monk1337

x=torch.rand(1,2)
print(x)
y=torch.rand(1,2)
print(y)

# output:
# tensor([[ 0.9781,  0.0128]])
# tensor([[ 0.9404,  0.6528]])

monk1337

#Inplace / Out-of-place
value_1=0
value_2=2

#creating new object
value_3=value_1+value_2
print(value_2)

#inplace
value_2+=value_1

monk1337

a=torch.rand(3)
print(a)
a.fill_(2.0)  #inplace
print(a)
b=a.add(3) #new_object creating
print(b)

# output:
# tensor([ 0.0025,  0.9584,  0.7258])
# tensor([ 2.,  2.,  2.])

monk1337

#converting torch to numpy

a=torch.Tensor(1,5)
print(a)
print(a.numpy())

# output:
# tensor(1.00000e-19 *
#        [[ 0.0000,  1.0842,  0.0000,  1.0842,  0.0000]])
# [[  0.00000000e+00   1.08420217e-19   0.00000000e+00   1.08420217e-19

monk1337

#cool experiment
import numpy as np

var_1=np.random.randint(1,3,[2,3])
new_= torch.from_numpy(var_1)
print(var_1)
print(new_)
np.add(var_1,1,out=var_1)
print(var_1)
print(new_) #change automatically

monk1337

a=torch.Tensor(3,8)
print(a)
b=a.view(-1,2)
print(b)

# output:
# tensor([[ 0.0000e+00,  1.0842e-19,  5.1128e+35,  3.6902e+19,  3.3927e-26,
#           1.4013e-45,  3.2906e-26,  1.4013e-45],
#         [ 3.2668e-26,  1.4013e-45,  3.3719e-26,  1.4013e-45, -1.7837e-37,
#          -1.6074e+07,  0.0000e+00,  0.0000e+00],

monk1337

def multilayer_perceptron(x, weights, biases, keep_prob):
    layer_1 = tf.add(tf.matmul(x, weights['h1']), biases['b1'])
    layer_1 = tf.nn.relu(layer_1)
    layer_1 = tf.nn.dropout(layer_1, keep_prob)
    out_layer = tf.matmul(layer_1, weights['out']) + biases['out']
    return out_layer
    
n_hidden_1 = 38
n_input = train_x.shape[1]
n_classes = train_y.shape[1]

monk1337

In Neural network we need normalizing data (features) when they have different ranges, 
for example one of them ranges from (1000-30000) while another feature ranges from (0.01 - 0.99). 
We cast both of them in one unified range for example (-1 to +1) or (0 to 1).. why we do that ? Two reasons, 
first to eliminate the influence of one factor over another (i.e. to give features equal chances), 
second reason is that the gradient descent with momentom GDM algorithm which is used for backpropagation converges 
faster with normalized data than with un-normalized data.   So, if you have different features have same range of data
then you don't need normalization.
Read the link I provided, it contains the equations required for normalization for both [0,1] and [-1, +1] ranges.
Standardization is changing data in such away that the new set has mean=0 and standard deviation =1. This kind of 
scaling is useful when the set of the data contains outliers (anomalies), because it has no boundaries like normalization.  

monk1337

c = np.array([[3.,4], [5.,6], [6.,7]])
print(np.mean(c,1))
print((5.+6)/2)
print([float(sum(l))/len(l) for l in zip(*c)])

+------------+---------+--------+
|            |  A      |  B     |
+------------+---------+---------
|      0     | 0.626386| 1.52325|----axis=1----->
+------------+---------+--------+