I am currently working on replicating YOLOv2 (not tiny) on iOS (Swift4) using MPS.
A problem is that it is hard for me to implement space_to_depth function (https://www.tensorflow.org/api_docs/python/tf/space_to_depth) and concatenation of two results from convolutions (13x13x256 + 13x13x1024 -> 13x13x1280). Could you give me some advice on making these parts? My codes are below.
...
let conv19 = MPSCNNConvolutionNode(source: conv18.resultImage,
weights: DataSource("conv19", 3, 3, 1024, 1024))
let conv20 = MPSCNNConvolutionNode(source: conv19.resultImage,
weights: DataSource("conv20", 3, 3, 1024, 1024))
let conv21 = MPSCNNConvolutionNode(source: conv13.resultImage,
weights: DataSource("conv21", 1, 1, 512, 64))
/*****
1. space_to_depth with conv21
2. concatenate the result of conv20(13x13x1024) to the result of 1 (13x13x256)
I need your help to implement this part!
******/
I believe space_to_depth can be expressed in form of a convolution:
For instance, for an input with dimension [1,2,2,1], Use 4 convolution kernels that each output one number to one channel, ie. [[1,0],[0,0]] [[0,1],[0,0]] [[0,0],[1,0]] [[0,0],[0,1]], this should put all input numbers from spatial dimension to depth dimension.
MPS actually has a concat node. See here: https://developer.apple.com/documentation/metalperformanceshaders/mpsnnconcatenationnode
You can use it like this:
concatNode = [[MPSNNConcatenationNode alloc] initWithSources:#[layerA.resultImage, layerB.resultImage]];
If you are working with the high level interface and the MPSNNGraph, you should just use a MPSNNConcatenationNode, as described by Tianyu Liu above.
If you are working with the low level interface, manhandling the MPSKernels around yourself, then this is done by:
Create a 1280 channel destination image to hold the result
Run the first filter as normal to produce the first 256 channels of the result
Run the second filter to produce the remaining channels, with the destinationFeatureChannelOffset set to 256.
That should be enough in all cases, except when the data is not the product of a MPSKernel. In that case, you'll need to copy it in yourself or use something like a linear neuron (a=1,b=0) to do it.
Related
I am trying to write a system that takes as input the state of a free body and gives as output the desired state of a revolute joint.
Hence the input port should take a vector of size 13 and the output port should give a vector of size 2.
For now, I just want to extract one value from the input state, so I tried something like this:
ball_state = Variable("ball_state")
desired_theta_system = builder.AddSystem(SymbolicVectorSystem(input=[ball_state], state=[], dynamics=[], output=[ball_state[6], 0]))
However, this did not work, as the ball_state variable is not subscriptable.
How can I do this? Do I need to derive LeafSystem?
Thanks!
You could certainly write a small LeafSystem, but you could accomplish this one with a MatrixGain system (e.g. with D =
[0, ..., 0, 1, 0, ...] ;
[0, ... 0].
So my goal is basically implementing global top-k subsampling. Gradient sparsification is quite simple and I have already done this building on stateful clients example, but now I would like to use encoders as you have recommended here at page 28. Additionally I would like to average only the non-zero gradients, so say we have 10 clients but only 4 have nonzero gradients at a given position for a communication round then I would like to divide the sum of these gradients to 4, not 10. I am hoping to achieve this by summing gradients at numerator and masks, 1s and 0s, at denominator. Also moving forward I will add randomness to gradient selection so it is imperative that I create those masks concurrently with gradient selection. The code I have right now is
import tensorflow as tf
from tensorflow_model_optimization.python.core.internal import tensor_encoding as te
#te.core.tf_style_adaptive_encoding_stage
class GrandienrSparsificationEncodingStage(te.core.AdaptiveEncodingStageInterface):
"""An example custom implementation of an `EncodingStageInterface`.
Note: This is likely not what one would want to use in practice. Rather, this
serves as an illustration of how a custom compression algorithm can be
provided to `tff`.
This encoding stage is expected to be run in an iterative manner, and
alternatively zeroes out values corresponding to odd and even indices. Given
the determinism of the non-zero indices selection, the encoded structure does
not need to be represented as a sparse vector, but only the non-zero values
are necessary. In the decode mehtod, the state (i.e., params derived from the
state) is used to reconstruct the corresponding indices.
Thus, this example encoding stage can realize representation saving of 2x.
"""
ENCODED_VALUES_KEY = 'stateful_topk_values'
INDICES_KEY = 'indices'
SHAPES_KEY = 'shapes'
ERROR_COMPENSATION_KEY = 'error_compensation'
def encode(self, x, encode_params):
shapes_list = [tf.shape(y) for y in x]
flattened = tf.nest.map_structure(lambda y: tf.reshape(y, [-1]), x)
gradients = tf.concat(flattened, axis=0)
error_compensation = encode_params[self.ERROR_COMPENSATION_KEY]
gradients_and_error_compensation = tf.math.add(gradients, error_compensation)
percentage = tf.constant(0.1, dtype=tf.float32)
k_float = tf.multiply(percentage, tf.cast(tf.size(gradients_and_error_compensation), tf.float32))
k_int = tf.cast(tf.math.round(k_float), dtype=tf.int32)
values, indices = tf.math.top_k(tf.math.abs(gradients_and_error_compensation), k = k_int, sorted = False)
indices = tf.expand_dims(indices, 1)
sparse_gradients_and_error_compensation = tf.scatter_nd(indices, values, tf.shape(gradients_and_error_compensation))
new_error_compensation = tf.math.subtract(gradients_and_error_compensation, sparse_gradients_and_error_compensation)
state_update_tensors = {self.ERROR_COMPENSATION_KEY: new_error_compensation}
encoded_x = {self.ENCODED_VALUES_KEY: values,
self.INDICES_KEY: indices,
self.SHAPES_KEY: shapes_list}
return encoded_x, state_update_tensors
def decode(self,
encoded_tensors,
decode_params,
num_summands=None,
shape=None):
del num_summands, decode_params, shape # Unused.
flat_shape = tf.math.reduce_sum([tf.math.reduce_prod(shape) for shape in encoded_tensors[self.SHAPES_KEY]])
sizes_list = [tf.math.reduce_prod(shape) for shape in encoded_tensors[self.SHAPES_KEY]]
scatter_tensor = tf.scatter_nd(
indices=encoded_tensors[self.INDICES_KEY],
updates=encoded_tensors[self.ENCODED_VALUES_KEY],
shape=[flat_shape])
nonzero_locations = tf.nest.map_structure(lambda x: tf.cast(tf.where(tf.math.greater(x, 0), 1, 0), tf.float32) , scatter_tensor)
reshaped_tensor = [tf.reshape(flat_tensor, shape=shape) for flat_tensor, shape in
zip(tf.split(scatter_tensor, sizes_list), encoded_tensors[self.SHAPES_KEY])]
reshaped_nonzero = [tf.reshape(flat_tensor, shape=shape) for flat_tensor, shape in
zip(tf.split(nonzero_locations, sizes_list), encoded_tensors[self.SHAPES_KEY])]
return reshaped_tensor, reshaped_nonzero
def initial_state(self):
return {self.ERROR_COMPENSATION_KEY: tf.constant(0, dtype=tf.float32)}
def update_state(self, state, state_update_tensors):
return {self.ERROR_COMPENSATION_KEY: state_update_tensors[self.ERROR_COMPENSATION_KEY]}
def get_params(self, state):
encode_params = {self.ERROR_COMPENSATION_KEY: state[self.ERROR_COMPENSATION_KEY]}
decode_params = {}
return encode_params, decode_params
#property
def name(self):
return 'gradient_sparsification_encoding_stage'
#property
def compressible_tensors_keys(self):
return False
#property
def commutes_with_sum(self):
return False
#property
def decode_needs_input_shape(self):
return False
#property
def state_update_aggregation_modes(self):
return {}
I have run some simple tests manually following the steps you outlined here at page 45. It works but I have some questions/problems.
When I use list of tensors of same shape (ex:2 2x25 tensors) as input,x, of encode it works without any issues but when I try to use list of tensors of different shapes (2x20 and 6x10) it gives and error saying
InvalidArgumentError: Shapes of all inputs must match: values[0].shape = [2,20] != values1.shape = [6,10] [Op:Pack] name: packed
How can I resolve this issue? As i said I want to use global top-k so it is essential I encode entire trainable model weights at once. Take the cnn model used here, all the tensors have different shapes.
How can I do the averaging I described at the beginning? For example here you have done
mean_factory = tff.aggregators.MeanFactory(
tff.aggregators.EncodedSumFactory(mean_encoder_fn), # numerator
tff.aggregators.EncodedSumFactory(mean_encoder_fn), # denominator )
Is there a way to repeat this with one output of decode going to numerator and other going to denominator? How can I handle dividing 0 by 0? tensorflow has divide_no_nan function, can I use it somehow or do I need to add eps to each?
How is partition handled when I use encoders? Does each client get a unique encoder holding a unique state for it? As you have discussed here at page 6 client states are used in cross-silo settings yet what happens if client ordering changes?
Here you have recommended using stateful clients example. Can you explain this a bit further? I mean in the run_one_round where exactly encoders go and how are they used/combined with client update and aggregation?
I have some additional information such as sparsity I want to pass to encode. What is the suggested method for doing that?
Here are some answers, hope it helps:
If you want to treat all of the aggregated structure just as a single tensor, use concat_factory as the outermost aggregator. That will concatenate entire structure to a rank-1 Tensor at clients, and then unpack back to the original structure at the end. Example use: tff.aggregators.concat_factory(tff.aggregators.MeanFactory(...))
Note the encoding stage objects are meant to work with a single tensor, so what you describe with identical tensors probably works only accidentally.
There are two options.
a. Modify the client training code such that the weights being passed to the weighted aggregator are already what you want it to be (zero/one
mask). In the stateful clients example you link, that would be here. You will then get what you need by default (by summing the numerator).
b. Modify UnweightedMeanFactory to do exactly the variant of averaging you describe and use that. Start would be modifying this
(and 4.) I think that is what you would need to implement. The same way existing client states are initialized in the example here, you would need extend it to contain the aggregator states, and make sure those are sampled together with the clients, as done here. Then, to integrate the aggregators in the example you would need to replace this hard-coded tff.federated_mean. An example of such integration is in the implementation of tff.learning.build_federated_averaging_process, primarily here
I am not sure what the question is. Perhaps get the previous working (seems like a prerequisite to me), and then clarify and ask in a new post?
I am currently planning on training a binary image classification model. The images I want to train on are the difference between two original pictures. In other words, for each data entry, I start out with 2 pictures, take their difference, and the label that difference as a 0 or 1. My question is what is the best way to find this difference. I know about cv2.absdiff and then normal subtraction of images - what is the most effective way to go about this?
About the data: The images I'm training on are screenshots that usually are the same but may have small differences. I found that normal subtraction seems to show the differences less than absdiff.
This is the code I use for absdiff:
diff = cv2.absdiff(img1, img2)
mask = cv2.cvtColor(diff, cv2.COLOR_BGR2GRAY)
th = 1
imask = mask>1
canvas = np.zeros_like(img2, np.uint8)
canvas[imask] = img2[imask]
And then this for normal subtraction:
def extract_diff(self,imageA, imageB, image_name, path):
subtract = imageB.astype(np.float32) - imageA.astype(np.float32)
mask = cv2.inRange(np.abs(subtract),(30,30,30),(255,255,255))
th = 1
imask = mask>1
canvas = np.zeros_like(imageA, np.uint8)
canvas[imask] = imageA[imask]
Thanks!
A difference can be negative or positive.
For some number types, such as uint8 (unsigned 8-bit int), which can't be negative (have no sign), a negative value wraps around and the value would make no sense anymore. Other types can be signed (e.g. floats, signed ints), so a negative value can be represented correctly.
That's why cv.absdiff exists. It always gives you absolute differences, and those are okay to represent in an unsigned type.
Example with numbers: a = 4, b = 6. a-b should be -2, right?
That value, as an uint8, will wrap around to become 0xFE, or 254 in decimal. The 254 value has some relation to the true -2 difference, but it also incorporates the range of values of the data type (8 bits: 256 values), so it's really just "code".
cv.absdiff would give you the absolute of the difference (-2), which is 2.
According to the original paper on Dropout said regularisation method can be applied to convolution layers often improving their performance. TensorFlow function tf.nn.dropout supports that by having a noise_shape parameter to allow the user to choose which parts of the tensors will drop out independently. However, neither the paper nor the documentation give a clear explanation of which dimensions should be kept independently, and the TensorFlow explanation of how noise_shape works is rather unclear.
only dimensions with noise_shape[i] == shape(x)[i] will make independent decisions.
I would assume that for a typical CNN layer output of the shape [batch_size, height, width, channels] we don't want individual rows or columns to drop out by themselves, but rather whole channels (which would be equivalent to a node in a fully connected NN) independently of the examples (i.e. different channels could be dropped for different examples in a batch). Am I correct in this assumption?
If so, how would one go about implementing dropout with such specificity using the noise_shape parameter? Would it be:
noise_shape=[batch_size, 1, 1, channels]
or:
noise_shape=[1, height, width, 1]
from here,
For example, if shape(x) = [k, l, m, n] and noise_shape = [k, 1, 1, n], each batch and channel component will be kept independently and each row and column will be kept or not kept together.
The code may help explain this.
noise_shape = noise_shape if noise_shape is not None else array_ops.shape(x)
# uniform [keep_prob, 1.0 + keep_prob)
random_tensor = keep_prob
random_tensor += random_ops.random_uniform(noise_shape,
seed=seed,
dtype=x.dtype)
# 0. if [keep_prob, 1.0) and 1. if [1.0, 1.0 + keep_prob)
binary_tensor = math_ops.floor(random_tensor)
ret = math_ops.div(x, keep_prob) * binary_tensor
ret.set_shape(x.get_shape())
return ret
the line random_tensor += supports broadcast. When the noise_shape[i] is set to 1, that means all elements in this dimension will add the same random value ranged from 0 to 1. So when noise_shape=[k, 1, 1, n], each row and column in the feature map will be kept or not kept together. On the other hand, each example (batch) or each channel receives different random values and each of them will be kept independently.
I have a convolutional neural network whose output is a 4-channel 2D image. I want to apply sigmoid activation function to the first two channels and then use BCECriterion to computer the loss of the produced images with the ground truth ones. I want to apply squared loss function to the last two channels and finally computer the gradients and do backprop. I would also like to multiply the cost of the squared loss for each of the two last channels by a desired scalar.
So the cost has the following form:
cost = crossEntropyCh[{1, 2}] + l1 * squaredLossCh_3 + l2 * squaredLossCh_4
The way I'm thinking about doing this is as follow:
criterion1 = nn.BCECriterion()
criterion2 = nn.MSECriterion()
error = criterion1:forward(model.output[{{}, {1, 2}}], groundTruth1) + l1 * criterion2:forward(model.output[{{}, {3}}], groundTruth2) + l2 * criterion2:forward(model.output[{{}, {4}}], groundTruth3)
However, I don't think this is the correct way of doing it since I will have to do 3 separate backprop steps, one for each of the cost terms. So I wonder, can anyone give me a better solution to do this in Torch?
SplitTable and ParallelCriterion might be helpful for your problem.
Your current output layer is followed by nn.SplitTable that splits your output channels and converts your output tensor into a table. You can also combine different functions by using ParallelCriterion so that each criterion is applied on the corresponding entry of output table.
For details, I suggest you read documentation of Torch about tables.
After comments, I added the following code segment solving the original question.
M = 100
C = 4
H = 64
W = 64
dataIn = torch.rand(M, C, H, W)
layerOfTables = nn.Sequential()
-- Because SplitTable discards the dimension it is applied on, we insert
-- an additional dimension.
layerOfTables:add(nn.Reshape(M,C,1,H,W))
-- We want to split over the second dimension (i.e. channels).
layerOfTables:add(nn.SplitTable(2, 5))
-- We use ConcatTable in order to create paths accessing to the data for
-- numereous number of criterions. Each branch from the ConcatTable will
-- have access to the data (i.e. the output table).
criterionPath = nn.ConcatTable()
-- Starting from offset 1, NarrowTable will select 2 elements. Since you
-- want to use this portion as a 2 dimensional channel, we need to combine
-- then by using JoinTable. Without JoinTable, the output will be again a
-- table with 2 elements.
criterionPath:add(nn.Sequential():add(nn.NarrowTable(1, 2)):add(nn.JoinTable(2)))
-- SelectTable is simplified version of NarrowTable, and it fetches the desired element.
criterionPath:add(nn.SelectTable(3))
criterionPath:add(nn.SelectTable(4))
layerOfTables:add(criterionPath)
-- Here goes the criterion container. You can use this as if it is a regular
-- criterion function (Please see the examples on documentation page).
criterionContainer = nn.ParallelCriterion()
criterionContainer:add(nn.BCECriterion())
criterionContainer:add(nn.MSECriterion())
criterionContainer:add(nn.MSECriterion())
Since I used almost every possible table operation, it looks a little bit nasty. However, this is the only way I could solve this problem. I hope that it helps you and others suffering from the same problem. This is how the result looks like:
dataOut = layerOfTables:forward(dataIn)
print(dataOut)
{
1 : DoubleTensor - size: 100x2x64x64
2 : DoubleTensor - size: 100x1x64x64
3 : DoubleTensor - size: 100x1x64x64
}