Can we permanently prune Tensorflow low magnitude weights? - tensorflow

I have a TensorFlow model where I can apply the pruner.prune_low_magnitude layer to the output of my Dense layers. This seems to work according to the instructions, and I get almost the same results down to 95% sparsity. The Processing time in GPU and CPU seems to be the same. It seems the pruning layer is calculating all the weights, then setting the smallest weights to zero to fit the sparsity figure. Is this right?
It would be nice to get a speed-up, of course, but for my present purposes, this is fine.
I would like to prune the weights and have them stay zero thereafter. I would prune some weights; then continue training to allow the model to recover from the pruning; then prune a bit more. I feel this should be a bit closer to what real neurones do. Is there some way of doing this?
My solution (which does not work yet) is to add a custom layer with Trainable=false. This has a mask array that starts off as all ones, and is set to zero if the corresponding weight is zero. The layer multiplies the weights by this mask array, so once a weight goes to zero, it will stay zero. Should this work? Is there a better way?

It would be nice to get a speed-up, of course, but for my present
purposes, this is fine.
To get a reduction in inference time, the op (Dense in your example) implementation needs to be able to take advantage of the sparsity and the particular hardware on which it runs.
The Tensorflow runtime on CPU/GPU does not support this yet, but TFLite does. The
Pruning for on-device inference w/ XNNPACK tutorial demonstrates that.
I would like to prune the weights and have them stay zero thereafter.
I would prune some weights; then continue training to allow the model
to recover from the pruning; then prune a bit more. I feel this should
be a bit closer to what real neurones do. Is there some way of doing
this?
This is the responsibility of the pruning schedule that is passed as argument to prune_low_magnitude().
Existing implementations, such as ConstantSparsity or PolynomialDecay perform what you described: prune only at some steps and let the model recover in between. The begin_step, end_step, and frequency arguments let you control when and how frequently the pruning is applied during training.
My solution (which does not work yet) is to add a custom layer with
Trainable=false. This has a mask array that starts off as all ones,
and is set to zero if the corresponding weight is zero. The layer
multiplies the weights by this mask array, so once a weight goes to
zero, it will stay zero. Should this work? Is there a better way?
This is basically how Tensorflow Model Optimization does it under the hood: pruning_impl.py. You just need to apply prune_low_magnitude() as per the Pruning with Keras tutorial.

Related

Does changing a token name in an image caption model affect performance?

If I train an image caption model then stop to rename a few tokens:
Should I train the model from scratch?
Or can I reload the model and continue training from the last epoch with the updated vocabulary?
Will either approach effect model accuracy/performance differently?
I would go for option 2.
When training the model from scratch, you are initializing the model's weights randomly and then you fit them based on your problem. However, if, instead of using random weights, you use weights that have already been trained for a similar problem, you may decrease the convergence time. This option is kind similar to the idea of transfer learning.
Just to give the other team a voice: So what is actually the difference between training from scratch and reloading a model and continuing training?
(2) will converge faster, (1) will probably have a better performance and should thus be chosen. Do we actually care about training times when we trade them off with performance - do you really? See you do not.
The further your model is already converged to a specific problem, the harder it gets to get it back into another optimum. Now you might be lucky and the chance, that you are going down the right rabid hole, rises with similar tasks and similar data. Yet with a change in your setup this can not be guaranteed.
Initializing a few epochs on other than your target domain, definitely makes sense and is beneficial, yet the question arises why you would not train on your target domain from the very beginning.
Note: For a more substantial read I'd like to refer you to this paper, where they explain in more depth why domain is of the essence and transfer learning could mess with your final performance.
It depends on the number of tokens being relabeled compared to the total amount. Just because you mentioned there are few of them, then the optimal solution in my opinion is clear.
You should start the training from scratch but initialize the weights with the values they had from wherever the previous training stopped (again mentioning that it is crucial that the samples that are being re-labeled are not of substantial amount). This way, the model will likely converge faster than starting with random weights and also better than trying to re-fit ("forget") what it managed to learn from the previous training.
Topologically speaking you are initializing in a position where the model is closer to a global minimum but has not made any steps towards a local minimum.
Hope this helps.

Strategies for pre-training models for use in tfjs

This is a more general version of a question I've already asked: Significant difference between outputs of deep tensorflow keras model in Python and tensorflowjs conversion
As far as I can tell, the layers of a tfjs model when run in the browser (so far only tested in Chrome and Firefox) will have small numerical differences in the output values when compared to the same model run in Python or Node. The cumulative effect of these small differences across all the layers of the model can cause fairly significant differences in the output. See here for an example of this.
This means a model trained in Python or Node will not perform as well in terms of accuracy when run in the browser. And the deeper your model, the worse it will get.
Therefore my question is, what is the best way to train a model to use with tfjs in the browser? Is there a way to ensure the output will be identical? Or do you just have to accept that there will be small numerical differences and, if so, are there any methods that can be used to train a model to be more resilient to this?
This answer is based on my personal observations. As such, it is debatable and not backed by much evidence. Some things that I follow to get accuracy of 16-bit models close to 32 bit models are:
Avoid using activations that have small upper and lower bounds, such as sigmoid or tanh, for hidden layers. These activations cause the weights of the next layer to become very sensitive to small values, and hence, small changes. I prefer using ReLU for such models. Since it is now the standard activation for hidden layers in most models, you should be using it in any case.
Avoid weight decay and L1/L2 regularizations on weights while training (the kernel_regularizer parameter in keras), since these increase sensitivity of weights. Use Dropout instead, I didn't observe a major drop in performance on TFLite when using it instead of numerical regularizers.

Memory requirements for back propagation - why not use the mean activation?

I need help understanding the memory requirements of a neural network and their differences between training and evaluation processes. More specifically, the memory requirements of the training process (I'm using a Keras API running on top of TensorFlow).
For a CNN that contains N weights, when using a batch of size x, there is a constant amount of memory required for the weights themselves and the input data. During the forward pass the GPU needs additional x*N units of memory (the specific required amount is not crucial to the question) for passing all the samples simultaneously and calculating the activation of each neuron.
My question is regarding the back propagation process, it seems that the process requires additional x*N units of memory(*) for the specific gradient of every weight for every sample. According to my understanding, it means that the algorithm calculates the specific gradients of each sample and then sums them up for the back-propagation to the previous layer.
Q. Since there is only a single update step per batch, why isn't the gradient calculation performed on the mean activation of each neuron? That way the additional required memory for training will only be (x+1)*N and not 2*x*N.
(*) This is according to my own little experiment of the maximal allowed batch size during evaluation (~4200) and training (~1200). Obviously it is a very simplified way of looking at the memory requirments
The short answer is: that is just the way the mini-batch SGD back-propagation algorithm works.
Looking back at its origins and difference between using the standard SGD and mini-batch SGD it is clearer why.
The standard stochastic gradient decent algorithm passes a single sample thru the model, then back-propagates its gradients and updates model weights before repeating the process with the next sample. The main downside is that it is a serial process (can't run samples simultaneously because the each sample needs to run on a model that was already updated by the previous sample), so it is very computationally expensive. In addition using just a single sample for each update results in a very noisy gradient.
The mini-batch SGD utilizes the same principle, with one difference - the gradients are accumulated from multiple samples and an update is only performed once every x samples. This helps to get a smooth gradient during training and enables passing multiple samples thru the model in parallel. This is the algorithm which is used when training with keras/tensorflow in mini-batches (commonly called batches but that term actually means using the batch gradient decent which is slightly different algorithm).
I haven't found any work regarding using the mean of the gradients in each layer for the update. It is interesting to check the results of such an algorithm. It would be more memory efficient however it is likely that it will also be less capable of reaching good minimum points.

tf-slim batch norm: different behaviour between training/inference mode

I'm attempting to train a tensorflow model based on the popular slim implementation of mobilenet_v2 and am observing behaviour I cannot explain related (I think) to batch normalization.
Problem Summary
Model performance in inference mode improves initially but starts producing trivial inferences (all near-zeros) after a long period. Good performance continues when run in training mode, even on the evaluation dataset. Evaluation performance is impacted by batch normalization decay/momentum rate... somehow.
More extensive implementation details below, but I'll probably lose most of you with the wall of text, so here are some pictures to get you interested.
The curves below are from a model which I tweaked the bn_decay parameter of while training.
0-370k: bn_decay=0.997 (default)
370k-670k: bn_decay=0.9
670k+: bn_decay=0.5
Loss for (orange) training (in training mode) and (blue) evaluation (in inference mode). Low is good.
Evaluation metric of model on evaluation dataset in inference mode. High is good.
I have attempted to produce a minimal example which demonstrates the issue - classification on MNIST - but have failed (i.e. classification works well and the problem I experience is not exhibited). My apologies for not being able to reduce things further.
Implementation Details
My problem is 2D pose estimation, targeting Gaussians centered at the joint locations. It is essentially the same as semantic segmentation, except rather than using a softmax_cross_entropy_with_logits(labels, logits) I use tf.losses.l2_loss(sigmoid(logits) - gaussian(label_2d_points)) (I use the term "logits" to describe unactivated output of my learned model, though this probably isn't the best term).
Inference Model
After preprocessing my inputs, my logits function is a scoped call to the base mobilenet_v2 followed by a single unactivated convolutional layer to make the number of filters appropriate.
from slim.nets.mobilenet import mobilenet_v2
def get_logtis(image):
with mobilenet_v2.training_scope(
is_training=is_training, bn_decay=bn_decay):
base, _ = mobilenet_v2.mobilenet(image, base_only=True)
logits = tf.layers.conv2d(base, n_joints, 1, 1)
return logits
Training Op
I have experimented with tf.contrib.slim.learning.create_train_op as well as a custom training op:
def get_train_op(optimizer, loss):
global_step = tf.train.get_or_create_global_step()
opt_op = optimizer.minimize(loss, global_step)
update_ops = set(tf.get_collection(tf.GraphKeys.UPDATE_OPS))
update_ops.add(opt_op)
return tf.group(*update_ops)
I'm using tf.train.AdamOptimizer with learning rate=1e-3.
Training Loop
I'm using the tf.estimator.Estimator API for training/evaluation.
Behaviour
Training initially goes well, with an expected sharp increase in performance. This is consistent with my expectations, as the final layer is rapidly trained to interpret the high-level features output by the pretrained base model.
However, after a long period (60k steps with batch_size 8, ~8 hours on a GTX-1070) my model begins to output near-zero values (~1e-11) when run in inference mode, i.e. is_training=False. The exact same model continues to improve when run in *training mode, i.e.is_training=True`, even on the valuation set. I have visually verified this is.
After some experimentation I changed the bn_decay (batch normalization decay/momentum rate) from the default 0.997 to 0.9 at ~370k steps (also tried 0.99, but that didn't make much of a difference) and observed an immdeiate improvement in accuracy. Visual inspection of the inference in inference mode showed clear peaks in the inferred values of order ~1e-1 in the expected places, consistent with the location of peaks from training mode (though values much lower). This is why the accuracy increases significantly, but the loss - while more volative - does not improve much.
These effects dropped off after more training and reverted to all zero inference.
I further dropped the bn_decay to 0.5 at step ~670k. This resulted in improvements to both loss and accuracy. I'll likely have to wait until tomorrow to see the long-term effect.
Loss and an evaluation metric plots given below. Note the evaluation metric is based on the argmax of the logits and high is good. Loss is based on the actual values, and low is good. Orange uses is_training=True on the training set, while blue uses is_training=False on the evaluation set. The loss of around 8 is consistent with all zero outputs.
Other notes
I have also experimented with turning off dropout (i.e. always running the dropout layers with is_training=False), and observed no difference.
I have experimented with all versions of tensorflow from 1.7 to 1.10. No difference.
I have trained models from the pretrained checkpoint using bn_decay=0.99 from the start. Same behaviour as using default bn_decay.
Other experiments with a batch size of 16 result in qualitatively identical behaviour (though I can't evaluate and train simultaneously due to memory constraints, hence quantitatively analysing on batch size of 8).
I have trained different models using the same loss and using tf.layers API and trained from scratch. They have worked fine.
Training from scratch (rather than using pretrained checkpoints) results in similar behaviour, though takes longer.
Summary/my thoughts:
I am confident this is not an overfitting/dataset problem. The model makes sensible inferences on the evaluation set when run with is_training=True, both in terms of location of peaks and magnitude.
I am confident this is not a problem with not running update ops. I haven't used slim before, but apart from the use of arg_scope it doesn't look too much different to the tf.layers API which I've used extensively. I can also inspect the moving average values and observe that they are changing as training progresses.
Chaning bn_decay values significantly effected the results temporarily. I accept that a value of 0.5 is absurdly low, but I'm running out of ideas.
I have tried swapping out slim.layers.conv2d layers for tf.layers.conv2d with momentum=0.997 (i.e. momentum consistent with default decay value) and behaviour was the same.
Minimal example using pretrained weights and Estimator framework worked for classification of MNIST without modification to bn_decay parameter.
I've looked through issues on both the tensorflow and models github repositories but haven't found much apart from this. I'm currently experimenting with a lower learning rate and a simpler optimizer (MomentumOptimizer), but that's more because I'm running out of ideas rather than because I think that's where the problem lies.
Possible Explanations
The best explanation I have is that my model parameters are rapidly cycling in a manner such that the moving statistics are unable to keep up with the batch statistics. I've never heard of such behaviour, and it doesn't explain why the model reverts to poor behaviour after more time, but it's the best explanation I have.
There may be a bug in the moving average code, but it has worked perfectly for me in every other case, including a simple classification task. I don't want to file an issue until I can produce a simpler example.
Anyway, I'm running out of ideas, the debug cycle is long, and I've already spent too much time on this. Happy to provide more details or run experiments on demand. Also happy to post more code, though I'm worried that'll scare more people off.
Thanks in advance.
Both lowering the learning rate to 1e-4 with Adam and using Momentum optimizer (with learning_rate=1e-3 and momentum=0.9) resolved this issue. I also found this post which suggests the problem spans multiple frameworks and is an undocumented pathology of some networks due to the interaction between optimizer and batch-normalization. I do not believe it is a simple case of the optimizer failing to find a suitable minimum due to the learning rate being too high (otherwise performance in training mode would be poor).
I hope that helps others experiencing the same issue, but I'm a long way from satisfied. I'm definitely happy to hear other explanations.

Neural network weights explode in linear unit

I am currently implementing a simple neural network and the backprop algorithm in Python with numpy. I have already tested my backprop method using central differences and the resulting gradient is equal.
However, the network fails to approximate a simple sine curve. The network hast one hidden layer (100 neurons) with tanh activation functions and a output layer with a linear activation function. Each unit hast also a bias input. The training is done by simple gradient descent with a learning rate of 0.2.
The problem arises from the gradient, which gets with every epoch larger, but I don't know why? Further, the problem is unchanged, if I decrease the learning rate.
EDIT: I have uploaded the code to pastebin: http://pastebin.com/R7tviZUJ
There are two things you can try, maybe in combination:
Use a smaller learning rate. If it is too high, you may be overshooting the minimum in the current direction by a lot, and so your weights will keep getting larger.
Use smaller initial weights. This is related to the first item. A smaller learning rate would fix this as well.
I had a similar problem (with a different library, DL4J), even in the case of extremely simple target functions. In my case, the issue turned out to be the cost function. When I changed from negative log likelihood to Poisson or L2, I started to get decent results. (And my results got MUCH better once I added exponential learning rate decay.)
Looks like you dont use regularization. If you train your network long enough it will start to learn the excact data rather than abstract pattern.
There are a couple of method to regularize your network like: stopped training, put a high cost to large gradients or more complex like e.g.g drop out. If you search web/books you probably will find many options for this.
A too big learning rate can fail to converge, and even DIVERGE, that is the point.
The gradient could diverge for this reason: when exceeding the position of the minima, the resulting point could not only be a bit further, but could even be at a greater distance than initially, but the other side. Repeat the process, and it will continue to diverge. in other words, the variation rate around the optimal position could be just to big compared to the learning rate.
Source: my understanding of the following video (watch near 7:30).
https://www.youtube.com/watch?v=Fn8qXpIcdnI&list=PLLH73N9cB21V_O2JqILVX557BST2cqJw4&index=10