I would like to pad my labels so that they would be of equal length to be passed into the ctc_loss function. Apparently, -1 is not allowed. If I were to apply padding, should the padding value be part of the labels for ctc?
Update
I have this code that converts dense labels into sparse ones to be passed to the ctc_loss function which I think is related to the problem.
def dense_to_sparse(dense_tensor, out_type):
indices = tf.where(tf.not_equal(dense_tensor, tf.constant(0, dense_tensor.dtype)
values = tf.gather_nd(dense_tensor, indices)
shape = tf.shape(dense_tensor, out_type=out_type)
return tf.SparseTensor(indices, values, shape)
Actually, -1 values are allowed to be present in the y_true argument of the ctc_batch_cost with one limitation - they should not appear within the actual label "content" which is specified by label_length (here i-th label "content" would start from the index 0 and end at the index label_length[i]).
So it is perfectly fine to pad labels with -1 so that they would be of equal length, as you intended. The only thing you should take care about is to correctly calculate and pass corresponding label_length values.
Here is the sample code which is a modified version of the test_ctc unit test from keras:
import numpy as np
from tensorflow.keras import backend as K
number_of_categories = 4
number_of_timesteps = 5
labels = np.asarray([[0, 1, 2, 1, 0], [0, 1, 1, 0, -1]])
label_lens = np.expand_dims(np.asarray([5, 4]), 1)
# dimensions are batch x time x categories
inputs = np.zeros((2, number_of_timesteps, number_of_categories), dtype=np.float32)
input_lens = np.expand_dims(np.asarray([5, 5]), 1)
k_labels = K.variable(labels, dtype="int32")
k_inputs = K.variable(inputs, dtype="float32")
k_input_lens = K.variable(input_lens, dtype="int32")
k_label_lens = K.variable(label_lens, dtype="int32")
res = K.eval(K.ctc_batch_cost(k_labels, k_inputs, k_input_lens, k_label_lens))
It runs perfectly fine even with -1 as the last element of the (second) labels sequence because corresponding label_lens item (second) specified that its length is 4.
If we change it to be 5 or if we change some other label value to be -1 then we have the All labels must be nonnegative integers exception that you've mentioned. But this just means that our label_lens is invalid.
Here's how I do it. I have a dense tensor labels that includes padding with -1, so that all targets in a batch have the same length. Then I use
labels_sparse = dense_to_sparse(labels, sparse_val=-1)
where
def dense_to_sparse(dense_tensor, sparse_val=0):
"""Inverse of tf.sparse_to_dense.
Parameters:
dense_tensor: The dense tensor. Duh.
sparse_val: The value to "ignore": Occurrences of this value in the
dense tensor will not be represented in the sparse tensor.
NOTE: When/if later restoring this to a dense tensor, you
will probably want to choose this as the default value.
Returns:
SparseTensor equivalent to the dense input.
"""
with tf.name_scope("dense_to_sparse"):
sparse_inds = tf.where(tf.not_equal(dense_tensor, sparse_val),
name="sparse_inds")
sparse_vals = tf.gather_nd(dense_tensor, sparse_inds,
name="sparse_vals")
dense_shape = tf.shape(dense_tensor, name="dense_shape",
out_type=tf.int64)
return tf.SparseTensor(sparse_inds, sparse_vals, dense_shape)
This creates a sparse tensor of the labels, which is what you need to put into the ctc loss. That is, you call tf.nn.ctc_loss(labels=labels_sparse, ...) The padding (i.e. all values equal to -1 in the dense tensor) is simply not represented in this sparse tensor.
Related
I want to concatenate two tensors checkerboard-ly in tensorflow2, like examples showed below:
example 1:
a = [[1,1],[1,1]]
b = [[0,0],[0,0]]
concated_a_and_b = [[1,0,1,0],[0,1,0,1]]
example 2:
a = [[1,1,1],[1,1,1],[1,1,1]]
b = [[0,0,0],[0,0,0],[0,0,0]]
concated_a_and_b = [[1,0,1,0,1,0],[0,1,0,1,0,1],[1,0,1,0,1,0]]
Is there a decent way in tensorflow2 to concatenate them like this?
A bit of background for this:
I first split a tensor c with a checkerboard mask into two halves a and b. A after some transformation I have to concat them back into oringnal shape and order.
What I mean by checkerboard-ly:
Step 1: Generate a matrix with alternated values
You can do this by first concatenating into [1, 0] pairs, and then by applying a final reshape.
Step 2: Reverse some rows
I split the matrix into two parts, reverse the second part and then rebuild the full matrix by picking alternatively from the first and second part
Code sample:
import math
import numpy as np
import tensorflow as tf
a = tf.ones(shape=(3, 4))
b = tf.zeros(shape=(3, 4))
x = tf.expand_dims(a, axis=-1)
y = tf.expand_dims(b, axis=-1)
paired_ones_zeros = tf.concat([x, y], axis=-1)
alternated_values = tf.reshape(paired_ones_zeros, [-1, a.shape[1] + b.shape[1]])
num_samples = alternated_values.shape[0]
middle = math.ceil(num_samples / 2)
is_num_samples_odd = middle * 2 != num_samples
# Gather first part of the matrix, don't do anything to it
first_elements = tf.gather_nd(alternated_values, [[index] for index in range(middle)])
# Gather second part of the matrix and reverse its elements
second_elements = tf.reverse(tf.gather_nd(alternated_values, [[index] for index in range(middle, num_samples)]), axis=[1])
# Pick alternatively between first and second part of the matrix
indices = np.concatenate([[[index], [index + middle]] for index in range(middle)], axis=0)
if is_num_samples_odd:
indices = indices[:-1]
output = tf.gather_nd(
tf.concat([first_elements, second_elements], axis=0),
indices
)
print(output)
I know this is not a decent way as it will affect time and space complexity. But it solves the above problem
def concat(tf1, tf2):
result = []
for (index, (tf_item1, tf_item2)) in enumerate(zip(tf1, tf2)):
item = []
for (subitem1, subitem2) in zip(tf_item1, tf_item2):
if index % 2 == 0:
item.append(subitem1)
item.append(subitem2)
else:
item.append(subitem2)
item.append(subitem1)
concated_a_and_b.append(item)
return concated_a_and_b
Trying to build loss function which captures the below functionality, which mask the output values once 'end of sequence' is encountered.
Given a tensor of shape [BatchSize,MaxSequenceLenght,OutputNodes]
Consider the below example
batch size = 3
Max Sequence Length=4
OutputNodes = 3
predicted = [[[0.1,0.3,0.2],[0.4,0.6,0.8],[0.5,0.2,0.3],[0.0,0.0,0.99]],
[[0.1,0.3,0.2],[0.4,0.9,0.8],[0.5,0.2,0.9],[0.4,0.6,0.8]],
[[0.1,0.3,0.2],[0.4,0.9,0.8],[0.5,0.2,0.1],[0.4,0.6,0.1]]]
I am dedicating the last output node to symbolise the 'end of sequence(EOS)' here node=2 . Nodes are labelled as (0, 1 and 2)
Based on the predicted value, I have to return a mask which tries to find the first occurrence of EOS.
In the above example,
first row has following sequence (argmax) => 1,2,0,2
Second row has following sequence => 1,1,2,2
Third row has following sequence => 1,1,9,1
So my mask should be
[[1,0,0,0],
[1,1,0,0],
[1,1,1,1]
The mask will ensure, the values post the EOS is ignored or not considered in calculating the loss.
Below is my code snipped I tried
sequence_cluster_asign = keras.backend.argmax(sequence_values,axis=-1)
loss_mask = []
for seq in K.tf.unstack(sequence_cluster_asign):
##appendEOS- To make sure tf.where is not empty
seq = tf.concat([seq,endOfSequenceTensor],axis=0)
endOfSequenceLocation = K.tf.where(K.tf.equal(seq,endOfSequence))[0][0]
loss_mask.append(tf.sequence_mask(endOfSequenceLocation,max_decoder_seq_length,dtype=tf.float32))
final_mask = K.stack(loss_mask)
Error encountered : ValueError: Cannot infer num from shape (?,?)
If you want to get mask in your question, you can use the following method.
import tensorflow as tf
import keras
from keras import backend as K
sequence_values = K.placeholder(shape=(None, 4, 3))
sequence_cluster_asign = keras.backend.argmax(sequence_values,axis=-1)
# keras version
result = K.cast(K.less(sequence_cluster_asign,sequence_values.get_shape().as_list()[-1]-1),dtype='int32')
result = K.cumprod(result,axis=-1)
# tensorflow version
# result = tf.cast(tf.less(sequence_cluster_asign,sequence_values.get_shape().as_list()[-1]-1),dtype=tf.int32)
# result = tf.cumprod(result,axis=-1)
predicted = [[[0.1,0.3,0.2],[0.4,0.6,0.8],[0.5,0.2,0.3],[0.0,0.0,0.99]],
[[0.1,0.3,0.2],[0.4,0.9,0.8],[0.5,0.2,0.9],[0.4,0.6,0.8]],
[[0.1,0.3,0.2],[0.4,0.9,0.8],[0.5,0.2,0.1],[0.4,0.6,0.1]]]
with tf.Session() as sess:
print(result.eval(feed_dict={sequence_values:predicted}))
[[1 0 0 0]
[1 1 0 0]
[1 1 1 1]]
I frequently use tf.add_to_collection to have Tensorflow automatically serialize intermediary results into a checkpoint. I find this the most convenient way to later fetch pointers to interesting tensors when a model was restored from a checkpoint. However, I realized that RNN state tuples cannot easily be added to a graph collection. Consider the following dummy example in TF 1.3:
import tensorflow as tf
import numpy as np
in_ = tf.placeholder(tf.float32, shape=[None, 5, 1])
batch_size = tf.shape(in_)[0]
cell1 = tf.nn.rnn_cell.BasicLSTMCell(num_units=128)
cell2 = tf.nn.rnn_cell.BasicLSTMCell(num_units=256)
cell = tf.nn.rnn_cell.MultiRNNCell([cell1, cell2])
outputs, last_state = tf.nn.dynamic_rnn(cell=cell,
inputs=in_,
initial_state=cell.zero_state(batch_size, dtype=tf.float32))
tf.add_to_collection('states', last_state)
loss = tf.reduce_mean(in_ - outputs)
loss_s = tf.summary.scalar('loss', loss)
writer = tf.summary.FileWriter('.', tf.get_default_graph())
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
l, s = sess.run([loss, loss_s], feed_dict={in_: np.ones([1, 5, 1])})
writer.add_summary(s)
This will produce the following warning:
WARNING:tensorflow:Error encountered when serializing states.
Type is unsupported, or the types of the items don't match field type in CollectionDef.
'tuple' object has no attribute 'name'
It seems that the serialization cannot handle tuples, and of course the last_state variable is a tuple. May be one could loop through the tuple and add each element individually to the collection, but that seems too complicated. What's a better way of handling this? In the end, I would like to access last_state again when the model is restored, ideally without needing access to the original code that created the model.
Actually, looping through every element of the state is not too complicated, and straight-forward to implement:
def add_to_collection_rnn_state(name, rnn_state):
for layer in rnn_state:
tf.add_to_collection(name, layer.c)
tf.add_to_collection(name, layer.h)
And then to load it:
def get_collection_rnn_state(name):
layers = []
coll = tf.get_collection(name)
for i in range(0, len(coll), 2):
state = tf.nn.rnn_cell.LSTMStateTuple(coll[i], coll[i+1])
layers.append(state)
return tuple(layers)
Note that this assumes that one collection only stores on state, i.e. use a different collection for every state you want to store, e.g. like this:
add_to_collection_rnn_state('states', last_state)
add_to_collection_rnn_state('init_state', init_state)
Edit
As pointed out correctly in the comments, the proposed solution only works for LSTMCells (that are represented as tuples as well). A more general solution that can handle GRU cells or potentially custom cells and mixes thereof, could look like this:
import tensorflow as tf
import numpy as np
def add_to_collection_rnn_state(name, rnn_state):
# store the name of each cell type in a different collection
coll_of_names = name + '__names__'
for layer in rnn_state:
n = layer.__class__.__name__
tf.add_to_collection(coll_of_names, n)
try:
for l in layer:
tf.add_to_collection(name, l)
except TypeError:
# layer is not iterable so just add it directly
tf.add_to_collection(name, layer)
def get_collection_rnn_state(name):
layers = []
coll = tf.get_collection(name)
coll_of_names = tf.get_collection(name + '__names__')
idx = 0
for n in coll_of_names:
if 'LSTMStateTuple' in n:
state = tf.nn.rnn_cell.LSTMStateTuple(coll[idx], coll[idx+1])
idx += 2
else: # add more cell types here
state = coll[idx]
idx += 1
layers.append(state)
return tuple(layers)
in_ = tf.placeholder(tf.float32, shape=[None, 5, 1])
batch_size = tf.shape(in_)[0]
cell1 = tf.nn.rnn_cell.BasicLSTMCell(num_units=128)
cell2 = tf.nn.rnn_cell.GRUCell(num_units=256)
cell3 = tf.nn.rnn_cell.BasicRNNCell(num_units=256)
cell = tf.nn.rnn_cell.MultiRNNCell([cell1, cell2, cell3])
outputs, last_state = tf.nn.dynamic_rnn(cell=cell,
inputs=in_,
initial_state=cell.zero_state(batch_size, dtype=tf.float32))
add_to_collection_rnn_state('last_state', last_state)
last_state_r = get_collection_rnn_state('last_state')
Comparing last_state and last_state_r reveals that both are identical (which they should be). Note that I am using a different collection to store the names because tensorflow can only serialize a collection when all elements in the collection are of the same type. E.g. mixing strings with Tensors in the same collection does not work.
I’m running into a number of issues relating to dynamic axes. I am trying to implement a convolutional rnn similar to the of the LSTM() function but handles sequential image input and outputs an image.
I’m able to build the network and pass dummy data through it to produce output, but when I try to compute the error with an input_variable label I consistently see the following error:
RuntimeError: Node '__v2libuid__Input471__v2libname__img_label' (InputValue operation): DataFor: FrameRange's dynamic axis is inconsistent with matrix: {numTimeSteps:1, numParallelSequences:2, sequences:[{seqId:0, s:0, begin:0, end:1}, {seqId:1, s:1, begin:0, end:1}]} vs. {numTimeSteps:2, numParallelSequences:1, sequences:[{seqId:0, s:0, begin:0, end:2}]}`
If I understand this error message correctly, it claims that the value I passed in as the label has inconsistent axes to what is expected with 2 time steps and 1 parallel sequence, when what is desired is 1 time-step and 2 sequences. This makes sense to me, but I’m not sure how the data I’m passing in is not conforming to this. Here are (roughly) the variable declarations and eval statements:
…
img_input = input_variable(shape=img_shape, dtype=np.float32, name="img_input")
convlstm = Recurrence(conv_lstm_cell, initial_state=initial_state)(img_input)
out = select_last(convlstm)
img_label = input_variable(shape=img_shape, dynamic_axes=out.dynamic_axes, dtype=np.float32, name="img_label”)
error = squared_error(out, img_label)
…
dummy_input = np.ones(shape=(2, 3, 3, 32, 32)) # (batch, seq_len, channels, height, width)
dummy_label = np.ones(shape=(2, 3, 32, 32)) # (batch, channels, height, width)
out = error.eval({img_input:dummy_input, img_label:dummy_label})
I believe part of the issue is with the dynamic_axes set when creating the img_label input_variable, I’ve also tried setting it to [Axis.default_batch_axis()] and not setting it at all and either squared error complains about inconsistent axes between out and img_label or I see the same error as above.
The only issue I see with the above setup is that your dummy label should have an explicit dynamic axis so it should be declared as
dummy_label = np.ones(shape=(2, 1, 3, 32, 32))
Assuming your convlstm works similar to an lstm, then the following works without issues for me and it evaluates the loss for two input/output pairs.
x = C.input_variable((3,32,32))
cx = convlstm(x)
lx = C.sequence.last(cx)
y = C.input_variable(lx.shape, dynamic_axes=lx.dynamic_axes)
loss = C.squared_error(y, lx)
x0 = np.arange(2*3*3*32*32,dtype=np.float32).reshape(2,3,3,32,32)
y0 = np.arange(2*1*3*32*32,dtype=np.float32).reshape(2,1,3,32,32)
loss.eval({x:x0, y:y0})
I'm trying to visualize the output of a convolutional layer in tensorflow using the function tf.image_summary. I'm already using it successfully in other instances (e. g. visualizing the input image), but have some difficulties reshaping the output here correctly. I have the following conv layer:
img_size = 256
x_image = tf.reshape(x, [-1,img_size, img_size,1], "sketch_image")
W_conv1 = weight_variable([5, 5, 1, 32])
b_conv1 = bias_variable([32])
h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)
So the output of h_conv1 would have the shape [-1, img_size, img_size, 32]. Just using tf.image_summary("first_conv", tf.reshape(h_conv1, [-1, img_size, img_size, 1])) Doesn't account for the 32 different kernels, so I'm basically slicing through different feature maps here.
How can I reshape them correctly? Or is there another helper function I could use for including this output in the summary?
I don't know of a helper function but if you want to see all the filters you can pack them into one image with some fancy uses of tf.transpose.
So if you have a tensor that's images x ix x iy x channels
>>> V = tf.Variable()
>>> print V.get_shape()
TensorShape([Dimension(-1), Dimension(256), Dimension(256), Dimension(32)])
So in this example ix = 256, iy=256, channels=32
first slice off 1 image, and remove the image dimension
V = tf.slice(V,(0,0,0,0),(1,-1,-1,-1)) #V[0,...]
V = tf.reshape(V,(iy,ix,channels))
Next add a couple of pixels of zero padding around the image
ix += 4
iy += 4
V = tf.image.resize_image_with_crop_or_pad(image, iy, ix)
Then reshape so that instead of 32 channels you have 4x8 channels, lets call them cy=4 and cx=8.
V = tf.reshape(V,(iy,ix,cy,cx))
Now the tricky part. tf seems to return results in C-order, numpy's default.
The current order, if flattened, would list all the channels for the first pixel (iterating over cx and cy), before listing the channels of the second pixel (incrementing ix). Going across the rows of pixels (ix) before incrementing to the next row (iy).
We want the order that would lay out the images in a grid.
So you go across a row of an image (ix), before stepping along the row of channels (cx), when you hit the end of the row of channels you step to the next row in the image (iy) and when you run out or rows in the image you increment to the next row of channels (cy). so:
V = tf.transpose(V,(2,0,3,1)) #cy,iy,cx,ix
Personally I prefer np.einsum for fancy transposes, for readability, but it's not in tf yet.
newtensor = np.einsum('yxYX->YyXx',oldtensor)
anyway, now that the pixels are in the right order, we can safely flatten it into a 2d tensor:
# image_summary needs 4d input
V = tf.reshape(V,(1,cy*iy,cx*ix,1))
try tf.image_summary on that, you should get a grid of little images.
Below is an image of what one gets after following all the steps here.
In case someone would like to "jump" to numpy and visualize "there" here is an example how to display both Weights and processing result. All transformations are based on prev answer by mdaoust.
# to visualize 1st conv layer Weights
vv1 = sess.run(W_conv1)
# to visualize 1st conv layer output
vv2 = sess.run(h_conv1,feed_dict = {img_ph:x, keep_prob: 1.0})
vv2 = vv2[0,:,:,:] # in case of bunch out - slice first img
def vis_conv(v,ix,iy,ch,cy,cx, p = 0) :
v = np.reshape(v,(iy,ix,ch))
ix += 2
iy += 2
npad = ((1,1), (1,1), (0,0))
v = np.pad(v, pad_width=npad, mode='constant', constant_values=p)
v = np.reshape(v,(iy,ix,cy,cx))
v = np.transpose(v,(2,0,3,1)) #cy,iy,cx,ix
v = np.reshape(v,(cy*iy,cx*ix))
return v
# W_conv1 - weights
ix = 5 # data size
iy = 5
ch = 32
cy = 4 # grid from channels: 32 = 4x8
cx = 8
v = vis_conv(vv1,ix,iy,ch,cy,cx)
plt.figure(figsize = (8,8))
plt.imshow(v,cmap="Greys_r",interpolation='nearest')
# h_conv1 - processed image
ix = 30 # data size
iy = 30
v = vis_conv(vv2,ix,iy,ch,cy,cx)
plt.figure(figsize = (8,8))
plt.imshow(v,cmap="Greys_r",interpolation='nearest')
you may try to get convolution layer activation image this way:
h_conv1_features = tf.unpack(h_conv1, axis=3)
h_conv1_imgs = tf.expand_dims(tf.concat(1, h_conv1_features_padded), -1)
this gets one vertical stripe with all images concatenated vertically.
if you want them padded (in my case of relu activations to pad with white line):
h_conv1_features = tf.unpack(h_conv1, axis=3)
h_conv1_max = tf.reduce_max(h_conv1)
h_conv1_features_padded = map(lambda t: tf.pad(t-h_conv1_max, [[0,0],[0,1],[0,0]])+h_conv1_max, h_conv1_features)
h_conv1_imgs = tf.expand_dims(tf.concat(1, h_conv1_features_padded), -1)
I personally try to tile every 2d-filter in a single image.
For doing this -if i'm not terribly mistaken since I'm quite new to DL- I found out that it could be helpful to exploit the depth_to_space function, since it takes a 4d tensor
[batch, height, width, depth]
and produces an output of shape
[batch, height*block_size, width*block_size, depth/(block_size*block_size)]
Where block_size is the number of "tiles" in the output image. The only limitation to this is that the depth should be the square of block_size, which is an integer, otherwise it cannot "fill" the resulting image correctly.
A possible solution could be of padding the depth of the input tensor up to a depth that is accepted by the method, but I sill havn't tried this.
Another way, which I think very easy, is using the get_operation_by_name function. I had hard time visualizing the layers with other methods but this helped me.
#first, find out the operations, many of those are micro-operations such as add etc.
graph = tf.get_default_graph()
graph.get_operations()
#choose relevant operations
op_name = '...'
op = graph.get_operation_by_name(op_name)
out = sess.run([op.outputs[0]], feed_dict={x: img_batch, is_training: False})
#img_batch is a single image whose dimensions are (1,n,n,1).
# out is the output of the layer, do whatever you want with the output
#in my case, I wanted to see the output of a convolution layer
out2 = np.array(out)
print(out2.shape)
# determine, row, col, and fig size etc.
for each_depth in range(out2.shape[4]):
fig.add_subplot(rows, cols, each_depth+1)
plt.imshow(out2[0,0,:,:,each_depth], cmap='gray')
For example below is the input(colored cat) and output of the second conv layer in my model.
Note that I am aware this question is old and there are easier methods with Keras but for people who use an old model from other people (such as me), this may be useful.