Numpy `Broadcast` array - numpy

I would like to do a transformation like dimshuffle in theano using numpy.
Example input:
np.array([[1, 0, 0], [1, 0, 0]])
Example output:
np.array([
[[1, 0, 0], [1, 0, 0]],
[[1, 0, 0], [1, 0, 0]],
[[1, 0, 0], [1, 0, 0]]
])


I don't know what dimshuffle does, but the output can be produced with repeat
In [319]: np.repeat(np.array([[1, 0, 0], [1, 0, 0]])[None,:,:],3,axis=0)
Out[319]:
array([[[1, 0, 0],
[1, 0, 0]],
[[1, 0, 0],
[1, 0, 0]],
[[1, 0, 0],
[1, 0, 0]]])
The input is 2d (2,3), so I have to add an axis - output is (3,2,3). tile would work, so would indexing, or even:
A=np.array([[1, 0, 0], [1, 0, 0]])
np.array([A,A,A])

Related

What is the purpose of rotating filters while building convolutions with scipy signal?

I recently came across a bit of python code (shown below) which does 2d convolution with scipy signal.
x = np.array([[1, 1, 1, 0, 0],
[0, 1, 1, 1, 0],
[0, 0, 1, 1, 1],
[0, 0, 1, 1, 0],
[0, 1, 1, 0, 0]],
dtype='float')
w_k = np.array([[1, 0, 1],
[0, 1, 0],
[1, 0, 1],],
dtype='float')
w_k = np.rot90(w_k, 2)
f = signal.convolve2d(x, w_k, 'valid')
Right before the convolve2d operation, the filter was rotated. What is the purpose of that?

Tensorflow.js: tf.pad results in TypeError: t.map is not a function

This code is from the TF API docs:
let t = tf.tensor([[1, 2, 3], [4, 5, 6]])
let padding = tf.tensor([[1, 1,], [2, 2]])
When I execute it:
tf.pad(t, padding, "CONSTANT")
I get:
TypeError: t.map is not a function
I'm using the latest version of tfjs.

padding is a normal js array of tuples ( array of arrray) and not a tensor.
As for now, the version 1.3.1, only the CONSTANT mode is supported. Here is the way to go:
let t = tf.tensor([[1, 2, 3], [4, 5, 6]])
let padding = [[2, 2,], [1, 1]]
tf.pad(t, padding).print()
// [[0, 0, 0, 0, 0],
// [0, 0, 0, 0, 0],
// [0, 1, 2, 3, 0],
// [0, 4, 5, 6, 0],
// [0, 0, 0, 0, 0],
// [0, 0, 0, 0, 0]]

Numpy check if a matrix can be transformed to another matrix by swaping columns

Say we have matrix A and B as follow
>>> A
matrix([[0, 0, 0, 1],
[1, 0, 0, 0],
[1, 0, 0, 0]])
>>> B
matrix([[0, 1, 0, 0],
[0, 0, 1, 0],
[0, 0, 1, 0]])
Clearly we can "transform" matrix A to B by column swapping. Is there an efficient algorithm to check whether two (potentially large) matrices can be transformed to each other in this way?

Here is a simple function. For very large matrix, it is possible that (A==B).all() is slower than np.array_equal(A,B).
import numpy as np
A = np.array([[0, 0, 0, 1],
[1, 0, 0, 0],
[1, 0, 0, 0]])
B = np.array([[0, 1, 0, 0],
[0, 0, 1, 0],
[0, 0, 1, 0]])
def isSwaping(a, b):
count = 0
for i, c in enumerate(a.T): # transpose of a
for d in b.T:
if (c == d).all():
count += 1
break
if count == i : # then it is uncessary to continue
return False
return True
print isSwaping(A, B)

tensorflow manipulate labels vector into "multiple hot encoder"

is it possible (in a nice way, that is) in tensorflow to achieve the next functionality:
assume we have a dense vector of tags
labels = [0,3,1,2,0]
I need to make a "multiple hot encoder" of it. meaning, for each row I need 1's up to the index of the label minus 1
so the required result will be
[[0, 0, 0],
[1, 1, 1],
[0, 0, 1],
[0, 1, 1],
[0, 0, 0]]
thanks

You could do this using tf.nn.embeddings_lookup as shown here:
embeddings = tf.constant([[0,0,0], [0,0,1], [0,1,1], [1,1,1]])
labels = [0,3,1,2,0]
encode_tensors = tf.nn.embedding_lookup(embeddings,labels)
Output of sess.run(encode_tensors) :
array([[0, 0, 0],
[1, 1, 1],
[0, 0, 1],
[0, 1, 1],
[0, 0, 0]], dtype=int32)
Hope this helps !

for completion:
it's also possible to use:
In [397]: labels = np.array([1, 2, 0, 3, 0])
In [398]: sess.run(tf.sequence_mask(labels, 3, dtype=tf.int8))
Out[398]:
array([[1, 0, 0],
[1, 1, 0],
[0, 0, 0],
[1, 1, 1],
[0, 0, 0]], dtype=int8)
the result matrix will be reversed from what I asked though

TensorFlow: Unpooling

Is there TensorFlow native function that does unpooling for Deconvolutional Networks ?
I have written this in normal python, but it is getting complicated when want to translate it to TensorFlow as it's objects does not even support item assignment at the moment, and I think this is a great inconvenience with TF.

I don't think there is an official unpooling layer yet which is frustrating because you have to use image resize (bilinear interpolation or nearest neighbor) which is like an average unpooling operation and it's reaaaly slow. Look at the tf api in the section 'image' and you will find it.
Tensorflow has a maxpooling_with_argmax thing where you get you maxpooled output as well as the activation map which is nice as you could use it in an unpooling layer to preserve the 'lost' spacial information but it seems as there isn't such an unpooling operation that does it. I guess that they are planning to add it ... soon.
Edit: I found some guy on google discuss a week ago who seems to have implemented something like this but I personally haven't tried it yet.
https://github.com/ppwwyyxx/tensorpack/blob/master/tensorpack/models/pool.py#L66

There is a couple of tensorflow implementations here pooling.py
Namely:
1) unpool operation (source) that utilizes output of tf.nn.max_pool_with_argmax. Although please notice, that as of tensorflow 1.0 tf.nn.max_pool_with_argmax is GPU-only
2) upsample operation that mimics inverse of max-pooling by filling positions of unpooled region with either zeros or copies of max element.
Comparing to tensorpack it allows copies of elements instead of zeros and supports strides other than [2, 2].
No recompile, back-prop friendly.
Illustration:

I was searching for a maxunpooling operation and tried implementing it. I came up with some kind of hacky implementation for the gradient, as I was struggling with CUDA.
The code is here, you will need to build it from source with GPU support.
Below is a demo application. No warranties, though!
There also exists an open issue for this operation.
import tensorflow as tf
import numpy as np
def max_pool(inp, k=2):
return tf.nn.max_pool_with_argmax_and_mask(inp, ksize=[1, k, k, 1], strides=[1, k, k, 1], padding="SAME")
def max_unpool(inp, argmax, argmax_mask, k=2):
return tf.nn.max_unpool(inp, argmax, argmax_mask, ksize=[1, k, k, 1], strides=[1, k, k, 1], padding="SAME")
def conv2d(inp, name):
w = weights[name]
b = biases[name]
var = tf.nn.conv2d(inp, w, [1, 1, 1, 1], padding='SAME')
var = tf.nn.bias_add(var, b)
var = tf.nn.relu(var)
return var
def conv2d_transpose(inp, name, dropout_prob):
w = weights[name]
b = biases[name]
dims = inp.get_shape().dims[:3]
dims.append(w.get_shape()[-2]) # adpot channels from weights (weight definition for deconv has switched input and output channel!)
out_shape = tf.TensorShape(dims)
var = tf.nn.conv2d_transpose(inp, w, out_shape, strides=[1, 1, 1, 1], padding="SAME")
var = tf.nn.bias_add(var, b)
if not dropout_prob is None:
var = tf.nn.relu(var)
var = tf.nn.dropout(var, dropout_prob)
return var
weights = {
"conv1": tf.Variable(tf.random_normal([3, 3, 3, 16])),
"conv2": tf.Variable(tf.random_normal([3, 3, 16, 32])),
"conv3": tf.Variable(tf.random_normal([3, 3, 32, 32])),
"deconv2": tf.Variable(tf.random_normal([3, 3, 16, 32])),
"deconv1": tf.Variable(tf.random_normal([3, 3, 1, 16])) }
biases = {
"conv1": tf.Variable(tf.random_normal([16])),
"conv2": tf.Variable(tf.random_normal([32])),
"conv3": tf.Variable(tf.random_normal([32])),
"deconv2": tf.Variable(tf.random_normal([16])),
"deconv1": tf.Variable(tf.random_normal([ 1])) }
## Build Miniature CEDN
x = tf.placeholder(tf.float32, [12, 20, 20, 3])
y = tf.placeholder(tf.float32, [12, 20, 20, 1])
p = tf.placeholder(tf.float32)
conv1 = conv2d(x, "conv1")
maxp1, maxp1_argmax, maxp1_argmax_mask = max_pool(conv1)
conv2 = conv2d(maxp1, "conv2")
maxp2, maxp2_argmax, maxp2_argmax_mask = max_pool(conv2)
conv3 = conv2d(maxp2, "conv3")
maxup2 = max_unpool(conv3, maxp2_argmax, maxp2_argmax_mask)
deconv2 = conv2d_transpose(maxup2, "deconv2", p)
maxup1 = max_unpool(deconv2, maxp1_argmax, maxp1_argmax_mask)
deconv1 = conv2d_transpose(maxup1, "deconv1", None)
## Optimizing Stuff
loss = tf.reduce_sum(tf.nn.sigmoid_cross_entropy_with_logits(deconv1, y))
optimizer = tf.train.AdamOptimizer(learning_rate=1).minimize(loss)
## Test Data
np.random.seed(123)
batch_x = np.where(np.random.rand(12, 20, 20, 3) > 0.5, 1.0, -1.0)
batch_y = np.where(np.random.rand(12, 20, 20, 1) > 0.5, 1.0, 0.0)
prob = 0.5
with tf.Session() as session:
tf.set_random_seed(123)
session.run(tf.initialize_all_variables())
print "\n\n"
for i in range(10):
session.run(optimizer, feed_dict={x: batch_x, y: batch_y, p: prob})
print "step", i + 1
print "loss", session.run(loss, feed_dict={x: batch_x, y: batch_y, p: 1.0}), "\n\n"
Edit 29.11.17
Some time back, I reimplemented it in a clean fashion against TensorFlow 1.0, the forward operations are also available as CPU-version. You can find it in this branch, I recommend you looking up the last few commits if you want to use it.

Nowadays there's a Tensorflow Addon MaxUnpooling2D:
Unpool the outputs of a maximum pooling operation.
tfa.layers.MaxUnpooling2D(
pool_size: Union[int, Iterable[int]] = (2, 2),
strides: Union[int, Iterable[int]] = (2, 2),
padding: str = 'SAME',
**kwargs
)
This class can e.g. be used as
import tensorflow as tf
import tensorflow_addons as tfa
pooling, max_index = tf.nn.max_pool_with_argmax(input, 2, 2, padding='SAME')
unpooling = tfa.layers.MaxUnpooling2D()(pooling, max_index)

I checked this which shagas mentioned here and it is working.
x = [[[[1, 1, 2,2, 3, 3],
[1, 1, 2,2, 3, 3],
[1, 1, 2,2, 3, 3],
[1, 1, 2,2, 3, 3],
[1, 1, 2,2, 3, 3],
[1, 1, 2,2, 3, 3]],
[[1, 1, 2,2, 3, 3],
[1, 1, 2,2, 3, 3],
[1, 1, 2,2, 3, 3],
[1, 1, 2,2, 3, 3],
[1, 1, 2,2, 3, 3],
[1, 1, 2,2, 3, 3]],
[[1, 1, 2,2, 3, 3],
[1, 1, 2,2, 3, 3],
[1, 1, 2,2, 3, 3],
[1, 1, 2,2, 3, 3],
[1, 1, 2,2, 3, 3],
[1, 1, 2,2, 3, 3]]]]
x = np.array(x)
inp = tf.convert_to_tensor(x)
out = UnPooling2x2ZeroFilled(inp)
out
Out[19]:
<tf.Tensor: id=36, shape=(1, 6, 12, 6), dtype=int64, numpy=
array([[[[1, 1, 2, 2, 3, 3],
[0, 0, 0, 0, 0, 0],
[1, 1, 2, 2, 3, 3],
[0, 0, 0, 0, 0, 0],
[1, 1, 2, 2, 3, 3],
[0, 0, 0, 0, 0, 0],
[1, 1, 2, 2, 3, 3],
[0, 0, 0, 0, 0, 0],
[1, 1, 2, 2, 3, 3],
[0, 0, 0, 0, 0, 0],
[1, 1, 2, 2, 3, 3],
[0, 0, 0, 0, 0, 0]],
[[0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0]],
[[1, 1, 2, 2, 3, 3],
[0, 0, 0, 0, 0, 0],
[1, 1, 2, 2, 3, 3],
[0, 0, 0, 0, 0, 0],
[1, 1, 2, 2, 3, 3],
[0, 0, 0, 0, 0, 0],
[1, 1, 2, 2, 3, 3],
[0, 0, 0, 0, 0, 0],
[1, 1, 2, 2, 3, 3],
[0, 0, 0, 0, 0, 0],
[1, 1, 2, 2, 3, 3],
[0, 0, 0, 0, 0, 0]],
[[0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0]],
[[1, 1, 2, 2, 3, 3],
[0, 0, 0, 0, 0, 0],
[1, 1, 2, 2, 3, 3],
[0, 0, 0, 0, 0, 0],
[1, 1, 2, 2, 3, 3],
[0, 0, 0, 0, 0, 0],
[1, 1, 2, 2, 3, 3],
[0, 0, 0, 0, 0, 0],
[1, 1, 2, 2, 3, 3],
[0, 0, 0, 0, 0, 0],
[1, 1, 2, 2, 3, 3],
[0, 0, 0, 0, 0, 0]],
[[0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0]]]])>
out1 = tf.keras.layers.MaxPool2D()(out)
out1
Out[37]:
<tf.Tensor: id=118, shape=(1, 3, 6, 6), dtype=int64, numpy=
array([[[[1, 1, 2, 2, 3, 3],
[1, 1, 2, 2, 3, 3],
[1, 1, 2, 2, 3, 3],
[1, 1, 2, 2, 3, 3],
[1, 1, 2, 2, 3, 3],
[1, 1, 2, 2, 3, 3]],
[[1, 1, 2, 2, 3, 3],
[1, 1, 2, 2, 3, 3],
[1, 1, 2, 2, 3, 3],
[1, 1, 2, 2, 3, 3],
[1, 1, 2, 2, 3, 3],
[1, 1, 2, 2, 3, 3]],
[[1, 1, 2, 2, 3, 3],
[1, 1, 2, 2, 3, 3],
[1, 1, 2, 2, 3, 3],
[1, 1, 2, 2, 3, 3],
[1, 1, 2, 2, 3, 3],
[1, 1, 2, 2, 3, 3]]]])>
If you need max unpooling then you can use (though I didn't check it) this one

Here it is my implementation. You should apply the max-pooling using tf.nn.max_pool_with_argmax and then pass the argmax result of tf.nn.max_pool_with_argmax
def unpooling(inputs, output_shape, argmax):
"""
Performs unpooling, as explained in:
https://www.oreilly.com/library/view/hands-on-convolutional-neural/9781789130331/6476c4d5-19f2-455f-8590-c6f99504b7a5.xhtml
:param inputs: Input Tensor.
:param output_shape: Desired output shape. For example, on 2D unpooling, this should be 4D (because of number of samples and channels).
:param argmax: Result argmax from tf.nn.max_pool_with_argmax
https://www.tensorflow.org/api_docs/python/tf/nn/max_pool_with_argmax
"""
flat_output_shape = tf.cast(tf.reduce_prod(output_shape), tf.int64)
updates = tf.reshape(inputs, [-1])
indices = tf.expand_dims(tf.reshape(argmax, [-1]), axis=-1)
ret = tf.scatter_nd(indices, updates, shape=[flat_output_shape])
ret = tf.reshape(ret, output_shape)
return ret
This has a small bug/feature that is that if argmax has a repeated value it will perform an addition instead of just putting the value once. Beware of this if stride is 1. I don't know, however, if this is desired or not.