I have recently been looking into incorporating the machine learning release for iOS developers with my app. Since this is my first time ever using anything ML related I was very lost when I started reading the different model descriptions that Apple has made available. They have the same purpose/description, the only difference being the actual file size. What is the difference between these models and how would you know which one is best fit ?
The models Apple makes available are just for simple demo purposes. Most of the time, these models are not sufficient for use in your own app.
The models on Apple's download page are trained for a very specific purpose: image classification on the ImageNet dataset. This means they can take an image and tell you what the "main" object is in the image, but only if it's one of the 1,000 categories from the ImageNet dataset.
Usually, this is not what you want to do in your own apps. If your app wants to do image classification, typically you want to train a model on your own categories (like food or cars or whatever). In that case you can take something like Inception-v3 (the original, not the Core ML version) and re-train it on your own data. That gives you a new model, which you then need to convert to Core ML again.
If your app wants to do something other than image classification, you can use these pretrained models as "feature extractors" in a larger neural network structure. But again this involves training your own model (usually from scratch) and then converting the result to Core ML.
So only in a very specific use case -- image classification using the 1,000 ImageNet categories -- are these Apple-provided models useful to your app.
If you do want to use any of these models, the difference between them is speed vs. accuracy. The smaller models are fastest but also least accurate. (In my opinion, VGG16 shouldn't be used on mobile. It's just too big and it's no more accurate than Inception or even MobileNet.)
SqueezeNets are fully convolutional and use Fire modules which have a squeeze layer of 1x1 convolutions which vastly decreases parameters as it can restrict the number of input channels each layer. This makes SqueezeNets extremely low latency, in addition to the fact they don't have dense layers.
MobileNets utilise depth-wise separable convolutions, very similar to inception towers in inception. These also reduce the number of a parameters and hence latency. MobileNets also have useful model-shrinking parameters than you can call before training to make it exact size you want. The Keras implementation can use ImageNet pre-trained weights too.
The other models are very deep, large models. The reduced number of parameters / style of convolution is not used for low latency but just for the ability to train very deep models, essentially. ResNet introduced residual connections between layers which were originally believed to be key in training very deep models. These aren't seen in the previously mentioned low latency models.
Related
I'm learning transfer learning with some pre-trained models (vgg16, vgg19,…), and I wonder why I need to load pre-trained weight to train my own dataset.
I can understand if the classes in my dataset are included in the dataset that the pre-trained model is trained with. For example, VGG model was trained with 1000 classes in Imagenet dataset, and my model is to classify cat-dog, which are also in the Imagenet dataset. But here the classes in my dataset are not in this dataset. So how the pre-trained weight can help?
You don't have to use a pretrained network in order to train a model for your task. However, in practice using a pretrained network and retraining it to your task/dataset is usually faster and often you end up with better models yielding higher accuracy. This is especially the case if you do not have a lot of training data.
Why faster?
It turns out that (relatively) independent of the dataset and target classes, the first couple of layers converge to similar results. This is due to the fact that low level layers usually act as edge, corner and other simple structure detectors. Check out this example that visualizes the structures that filters of different layers "react" to. Having already trained the lower layers, adapting the higher level layers to your use case is much faster.
Why more accurate?
This question is harder to answer. IMHO it is due to the fact that pretrained models that you use as basis for transfer learning were trained on massive datasets. This means that the knowledge acquired flows into your retrained network and will help you to find a better local minimum of your loss function.
If you are in the compfortable situation that you have a lot of training data you should probably train a model from scratch as the pretained model might "point you in the wrong direction".
In this master thesis you can find a bunch of tasks (small datasets, medium datasets, small semantical gap, large semantical gap) where 3 methods are compared : fine tuning, features extraction + SVM, from scratch. Fine tuning a model pretrained on Imagenet is almost always a better choice.
Closed. This question needs to be more focused. It is not currently accepting answers.
Want to improve this question? Update the question so it focuses on one problem only by editing this post.
Closed 5 years ago.
Improve this question
As I read and searched about the fine tuning of pre-trained network, it is done in following two steps (in short):
freeze the hidden layer and unfreeze the fully connected layer and trained.
unfreeze both the layers and again train.
My questions are:
Whether it is enough to perform only first step?
If I preform only first step, is it not same as network as a feature extractor method?
(The network as a feature extractor method is, to extract the feature using pre-trained network and classify it using tradition machine learning classification algorithm).
If you want more information to clarify the question, please let me know.
There are some issues with your question...
First, you clearly imply a network with only 2 layers, which is rather (very) far from the way fine-tuning is actually used in practice nowadays.
Second, what exactly do you mean by "enough" in your first question (enough for what)?
In fact, there is enough overlapping between the notions of pre-trained models, feature extractors, and fine-tuning, and different people may even use the involved terms in not exactly the same ways. One approach, adopted by the Stanford CNNs for Visual Recognition course, is to consider all these as special cases of something more general called transfer learning; here is a useful excerpt from the respective section of the aforementioned course, which arguably addresses the spirit (if not the letter) of your questions:
The three major Transfer Learning scenarios look as follows:
ConvNet as fixed feature extractor. Take a ConvNet pretrained on ImageNet, remove the last fully-connected layer (this layer’s outputs are the 1000 class scores for a different task like ImageNet), then treat the rest of the ConvNet as a fixed feature extractor for the new dataset. In an AlexNet, this would compute a 4096-D vector for every image that contains the activations of the hidden layer immediately before the classifier. We call these features CNN codes. It is important for performance that these codes are ReLUd (i.e. thresholded at zero) if they were also thresholded during the training of the ConvNet on ImageNet (as is usually the case). Once you extract the 4096-D codes for all images, train a linear classifier (e.g. Linear SVM or Softmax classifier) for the new dataset.
Fine-tuning the ConvNet. The second strategy is to not only replace and retrain the classifier on top of the ConvNet on the new dataset, but to also fine-tune the weights of the pretrained network by continuing the backpropagation. It is possible to fine-tune all the layers of the ConvNet, or it’s possible to keep some of the earlier layers fixed (due to overfitting concerns) and only fine-tune some higher-level portion of the network. This is motivated by the observation that the earlier features of a ConvNet contain more generic features (e.g. edge detectors or color blob detectors) that should be useful to many tasks, but later layers of the ConvNet becomes progressively more specific to the details of the classes contained in the original dataset. In case of ImageNet for example, which contains many dog breeds, a significant portion of the representational power of the ConvNet may be devoted to features that are specific to differentiating between dog breeds.
Pretrained models. Since modern ConvNets take 2-3 weeks to train across multiple GPUs on ImageNet, it is common to see people release their final ConvNet checkpoints for the benefit of others who can use the networks for fine-tuning. For example, the Caffe library has a Model Zoo where people share their network weights.
When and how to fine-tune? How do you decide what type of transfer learning you should perform on a new dataset? This is a function of several factors, but the two most important ones are the size of the new dataset (small or big), and its similarity to the original dataset (e.g. ImageNet-like in terms of the content of images and the classes, or very different, such as microscope images). Keeping in mind that ConvNet features are more generic in early layers and more original-dataset-specific in later layers, here are some common rules of thumb for navigating the 4 major scenarios:
New dataset is small and similar to original dataset. Since the data is small, it is not a good idea to fine-tune the ConvNet due to overfitting concerns. Since the data is similar to the original data, we expect higher-level features in the ConvNet to be relevant to this dataset as well. Hence, the best idea might be to train a linear classifier on the CNN codes.
New dataset is large and similar to the original dataset. Since we have more data, we can have more confidence that we won’t overfit if we were to try to fine-tune through the full network.
New dataset is small but very different from the original dataset. Since the data is small, it is likely best to only train a linear classifier. Since the dataset is very different, it might not be best to train the classifier form the top of the network, which contains more dataset-specific features. Instead, it might work better to train the SVM classifier from activations somewhere earlier in the network.
New dataset is large and very different from the original dataset. Since the dataset is very large, we may expect that we can afford to train a ConvNet from scratch. However, in practice it is very often still beneficial to initialize with weights from a pretrained model. In this case, we would have enough data and confidence to fine-tune through the entire network.
Is it possible to feed image features, say SIFT features, to a convolutional neural network model in Tensorflow? I am trying a tensorflow implementation of this project in which a grayscale image is coloured. Will image features be a better choice than feeding the images as is to the model?
PS. I am a novice to machine learning and is not familiar with creating neural n/w models
You can feed tensorflow neural net almost anything.
If you have extra features for each pixel, then instead of using one channel (intensity) you would use multiple channels.
If you have extra features, which are about whole image, you can make separate input a merge features at some upper layer.
As for the better performance, you should try both approaches.
General intuition is that, extra features help if you don't have many samples and their effect is diminishing if you have many samples and network can learn features by itself.
Also one more point: If you are novice, I strongly recommend using higher level framework like keras.io (which is layer over tensorflow) instead of tensorflow.
I'm trying to utilize a pre-trained model like Inception v3 (trained on the 2012 ImageNet data set) and expand it in several missing categories.
I have TensorFlow built from source with CUDA on Ubuntu 14.04, and the examples like transfer learning on flowers are working great. However, the flowers example strips away the final layer and removes all 1,000 existing categories, which means it can now identify 5 species of flowers, but can no longer identify pandas, for example. https://www.tensorflow.org/versions/r0.8/how_tos/image_retraining/index.html
How can I add the 5 flower categories to the existing 1,000 categories from ImageNet (and add training for those 5 new flower categories) so that I have 1,005 categories that a test image can be classified as? In other words, be able to identify both those pandas and sunflowers?
I understand one option would be to download the entire ImageNet training set and the flowers example set and to train from scratch, but given my current computing power, it would take a very long time, and wouldn't allow me to add, say, 100 more categories down the line.
One idea I had was to set the parameter fine_tune to false when retraining with the 5 flower categories so that the final layer is not stripped: https://github.com/tensorflow/models/blob/master/inception/README.md#how-to-retrain-a-trained-model-on-the-flowers-data , but I'm not sure how to proceed, and not sure if that would even result in a valid model with 1,005 categories. Thanks for your thoughts.
After much learning and working in deep learning professionally for a few years now, here is a more complete answer:
The best way to add categories to an existing models (e.g. Inception trained on the Imagenet LSVRC 1000-class dataset) would be to perform transfer learning on a pre-trained model.
If you are just trying to adapt the model to your own data set (e.g. 100 different kinds of automobiles), simply perform retraining/fine tuning by following the myriad online tutorials for transfer learning, including the official one for Tensorflow.
While the resulting model can potentially have good performance, please keep in mind that the tutorial classifier code is highly un-optimized (perhaps intentionally) and you can increase performance by several times by deploying to production or just improving their code.
However, if you're trying to build a general purpose classifier that includes the default LSVRC data set (1000 categories of everyday images) and expand that to include your own additional categories, you'll need to have access to the existing 1000 LSVRC images and append your own data set to that set. You can download the Imagenet dataset online, but access is getting spotier as time rolls on. In many cases, the images are also highly outdated (check out the images for computers or phones for a trip down memory lane).
Once you have that LSVRC dataset, perform transfer learning as above but including the 1000 default categories along with your own images. For your own images, a minimum of 100 appropriate images per category is generally recommended (the more the better), and you can get better results if you enable distortions (but this will dramatically increase retraining time, especially if you don't have a GPU enabled as the bottleneck files cannot be reused for each distortion; personally I think this is pretty lame and there's no reason why distortions couldn't also be cached as a bottleneck file, but that's a different discussion and can be added to your code manually).
Using these methods and incorporating error analysis, we've trained general purpose classifiers on 4000+ categories to state-of-the-art accuracy and deployed them on tens of millions of images. We've since moved on to proprietary model design to overcome existing model limitations, but transfer learning is a highly legitimate way to get good results and has even made its way to natural language processing via BERT and other designs.
Hopefully, this helps.
Unfortunately, you cannot add categories to an existing graph; you'll basically have to save a checkpoint and train that graph from that checkpoint onward.
In standard cookbook machine learning, we operate on a rectangular matrix; that is, all of our data points have the same number of features. How do we cope with situations in which all of our data points have different numbers of features? For example, if we want to do visual classification but all of our pictures are of different dimensions, or if we want to do sentiment analysis but all of our sentences have different amounts of words, or if we want to do stellar classification but all of the stars have been observed a different number of times, etc.
I think the normal way would be to extract features of regular size from these irregularly sized data. But I attended a talk on deep learning recently where the speaker emphasized that instead of hand-crafting features from data, deep learners are able to learn the appropriate features themselves. But how do we use e.g. a neural network if the input layer is not of a fixed size?
Since you are asking about deep learning, I assume you are more interested in end-to-end systems, rather then feature design. Neural networks that can handle variable data inputs are:
1) Convolutional neural networks with pooling layers. They are usually used in image recognition context, but recently were applied to modeling sentences as well. ( I think they should also be good at classifiying stars ).
2) Recurrent neural networks. (Good for sequential data, like time series,sequence labeling tasks, also good for machine translation).
3) Tree-based autoencoders (also called recursive autoencoders) for data arranged in tree-like structures (can be applied to sentence parse trees)
Lot of papers describing example applications can readily be found by googling.
For uncommon tasks you can select one of these based on the structure of your data, or you can design some variants and combinations of these systems.
You can usually make the number of features the same for all instances quite easily:
if we want to do visual classification but all of our pictures are of different dimensions
Resize them all to a certain dimension / number of pixels.
if we want to do sentiment analysis but all of our sentences have different amounts of words
Keep a dictionary of the k words in all of your text data. Each instance will consist of a boolean vector of size k where the i-th entry is true if word i from the dictionary appears in that instance (this is not the best representation, but many are based on it). See the bag of words model.
if we want to do stellar classification but all of the stars have been observed a different number of times
Take the features that have been observed for all the stars.
But I attended a talk on deep learning recently where the speaker emphasized that instead of hand-crafting features from data deep learners are able to learn the appropriate features themselves.
I think the speaker probably referred to higher level features. For example, you shouldn't manually extract the feature "contains a nose" if you want to detect faces in an image. You should feed it the raw pixels, and the deep learner will learn the "contains a nose" feature somewhere in the deeper layers.