We Keep Coding

how many epochs needed for 68000 input frames for CNN + LSTM acrhitecture?

i'm trying to train my neural network. it aims is to predict video quality. i am using VGG16+LSTM networks, but the VGG16 is not trainable. the total of trainable parameters are 700,000.
i have three questions:
is enough 700,000 trainable parameters for training the 68000 input frames?
is it essential to train vgg16?
how many epochs needed for getting the best resaults?

I haven't been into machine learning in a while, but my understanding is that:
depends, but the only way to find out is to train it and look for over/underfitting
depends on the network layout. It might also be useful to bypass some information around the VGG16, in case the VGG16 hides some of the information you actually need about 'video quality'
depends. You wouldh have to split your data into a training and a test set in order to find that out.
As most things in machine learning and especially deep learning the answers aren't obvious and depend heavily on the problem and the exact network layout. There will be much trial and error involved.
The most important takeaway, I think, is to have two (or even three) different datasets for the training/validation/test step, so you can answer those questions yourself.
For more information, read the wikipedia entry about splitting your datasets.

You start with one and see what impact it had.
Even one epoch will take long and getting the error takes also a bit of time.

Is it possible to validate a deep learning model by training small data subset?

I am looking to train a large model (resnet or vgg) for face identification.
Is it valid strategy to train on few faces (1..3) to validate a model?
In other words - if a model learns one face well - is it evidence that the model is good for the task?
point here is that I don't want to spend a week of GPU expensive time only to find out that my model is no good or data has errors or my TF coding has a bug

Short answer: No, because Deep Learning works well on huge amount of data.
Long answer: No. The problem is that learning only one face could overfit your model on that specific face, without learning features not present in your examples. Because for example, the model has learn to detect your face thanks to a specific, very simple, pattern in that face (that's called overfitting).
Making a stupid simple example, your model has learn to detect that face because there is a mole on your right cheek, and it has learn to identify it
To make your model perform well on the general case, you need an huge amount of data, making your model capable to learn different kind of patterns
Suggestion:
Because the training of a deep neural network is a time consuming task, usually one does not train one single neural network at time, but many neural network are trained in parallel, with different hyperparameters (layers, nodes, activation functions, learning rate, etc).
Edit because of the discussion below:
If your dataset is small is quite impossible to have a good performance on the general case, because the neural network will learn the easiest pattern, which is usually not the general/better one.
Adding data you force the neural network to extract good patterns, that work on the general case.
It's a tradeoff, but usually a training on a small dataset would not lead to a good classifier on the general case
edit2: refrasing everything to make it more clear. A good performance on a small dataset don't tell you if your model when trained on all the dataset is a good model. That's why you train to
the majority of your dataset and test/validate on a smaller dataset

For face recognition, usually a siamese net or triplet loss are used. This is an approach for one-shot learning. Which means it could perform really well given only few examples per class (person face here), but you still need to train it on many examples (different person faces). See for example:
https://towardsdatascience.com/one-shot-learning-with-siamese-networks-using-keras-17f34e75bb3d
You wouldn't train your model from scratch but use a pretrained model anyways and fine-tune it for your task
You could also have a look at pretrained face recognition models for better results like facenet
https://github.com/davidsandberg/facenet

Is it possible to remove categories in a pretrained tensorflow model?

I am currently using Tensorflow Object Detection API for my human detection app.
I tried filtering in the API itself which worked but I am still not contended by it because it's slow. So I'm wondering if I could remove other categories in the model itself to also make it faster.
If it is not possible, can you please give me other suggestions to make the API faster since I will be using two cameras. Thanks in advance and also pardon my english :)

Your questions addresses several topics for using neural network pretrained models.
Theoretical methods
In general, you can always neutralize categories by removing the corresponding neurons in the softmax layer and compute a new softmax layer only with the relevant rows of the matrix.
This method will surely work (maybe that is what you meant by filtering) but will not accelerate the network computation time by much, since most of the flops (multiplications and additions) will remain.
Similar to decision trees, pruning is possible but may reduce performance. I will explain what pruning means, but note that the accuracy over your categories may remain since you are not just trimming, you are predicting less categories as well.
Transfer the learning to your problem. See stanford's course in computer vision here. Most of the times I've seen that works good is by keeping the convolution layers as-is, and preparing a medium-size dataset of the objects you'd like to detect.
I will add more theoretical methods if you request, but the above are the most common and accurate I know.
Practical methods
Make sure you are serving your tensorflow model, and not just using an inference python code. This could significantly accelerate performance.
You can export the parameters of the network and load them in a faster framework such as CNTK or Caffe. These frameworks work in C++/CSharp and can inference much faster. Make sure you load the weights correctly, some frameworks use different order in tensor dimensions when saving/loading (little/big endian-like issues).
If your application perform inference on several images, you can distribute the computation via several GPUs. **This can also be done in tensorflow, see Using GPUs.
Pruning a neural network
Maybe this is the most interesting method of adapting big networks for simple tasks. You can see a beginner's guide here.
Pruning means that you remove parameters from your network, specifically the whole nodes/neurons in a decision tree/neural network (resp). To do that in object detection, you can do as follows (simplest way):
Randomly prune neurons from the fully connected layers.
Train one more epoch (or more) with low learning rate, only on objects you'd like to detect.
(optional) Perform the above several times for validation and choose best network.
The above procedure is the most basic one, but you can find plenty of papers that suggest algorithms to do so. For example
Automated Pruning for Deep Neural Network Compression and An iterative pruning algorithm for feedforward neural networks.

sknn multi layer perceptron classifier

I am using the following neural net classifier in python
from sknn.mlp import Layer,Classifier
nn = mlp.Classifier(
layers=[
mlp.Layer("Tanh", units=n_feat/8),
mlp.Layer("Sigmoid", units=n_feat/16),
mlp.Layer("Softmax", units=n_targets)],
n_iter=50,
n_stable=10,
batch_size=25,
learning_rate=0.002,
learning_rule="momentum",
valid_size=0.1,
verbose=1)
which is working just fine.My question is that how to proceed if I require for example 100,200 or 500 hidden layers? Do I have to specify each layer here manually or someone has better Idea in python for MLP?

You could create some loop-based mechanism to build the list of layers I suppose, but there's a bigger issue here. A standard MLP with hundreds of layers is likely to be extremely expensive to train - both in terms of computational speed as well as memory usage. MLPs typically only have one or two hidden layers, or occasionally a few more. But for problems that can truly benefit from more hidden layers, it becomes important to incorporate some of the lessons learned in the field of deep learning. For example, for object classification on images, using all fully-connected layers is incredibly inefficient, because you're interested in identifying spatially-local patterns, and therefore interactions between spatially-distant pixels or regions is largely noise. (This is a perfect case for using a deep convolutional neural net.)
Although some very deep networks have been created, it's worth pointing out that even Google's very powerful Inception-v3 model is only 42-layers deep. Anyway, if you're interested in building deep models, I'd recommend reading this Deep Learning book. From what I've read of it, it seems to be a very good introduction. Hope that helps!

How to extract feature for an image after neural network training?

Is there a way to learn unsupervised features from set of images. Similar to word2vec or doc2vec, where neural network is learnt and given new document we get its features.
Expecting similar to this example shows that it can load learnt nn-model and predict features for new images.
Is there any simple example how to implement cnn over images and get their features back will help !!
Suppose in this example
If I want to get cnn features for all X_train and X_test ... is there any way?
Also, if we can get weights per layer per image, we can stack them and use as features. In that case is there a way to get the same.
Using those features for unsupervised task would be easier, if we consider them as vectors.

If I correctly understood your question, this task is quite common in a deep learning field. In case of images what I consider the best is a convolutional autoencoder. You may read about this architecture e.g. here
http://people.idsia.ch/~ciresan/data/icann2011.pdf
Previous version of Keras supported this architecture as one of core layers, though from version 1.0 I noticed that it disappeared from documentation. But - it's still quite easy to build it from a scratch :)
In noimage cases there are also another approaches like e.g. Restricted Boltzmann Machines.
UPDATE :
When it comes to what sort of activations are the best for obtaining new features from neural network activations - from my personal experience - it depends on the size of the net which you use. If you use a network which last layer is wide (has a lot of nodes) it might be useful to get only last layer (due to number of parameters if you want to consider also previous layers - it may harm the performance of learning). But - if (like in case of some MNIST networks) your last layer is not sufficient for this task - you may try using also previous layers activation or even all net activity. To be honest - I'm not expecting much of improvement in this case - but you may try. I think that you should use both approaches - starting from taking only last layer activations - and then trying to check the behaviour of your new classifier when you add activations from previous layers.
What I will strongly advise to you is also getting some insights from what sort of features network is learning - using T-SNE embeddings of it activations. In many cases I found it useful - e.g. checking if the size of a layer is sufficient. Using T-SNE you may check if the features obtained from last layer are good discriminators of your classes. It may also give you good insights about your data and what neural networks are really learning (alongside with amazing visualizations :) )