Picture Classification on Small Datasets with Keras

Coaching a convnet with a small dataset

Having to coach an image-classification mannequin utilizing little or no information is a typical state of affairs, which you’ll doubtless encounter in observe in case you ever do pc imaginative and prescient in knowledgeable context. A “few” samples can imply wherever from just a few hundred to some tens of 1000’s of pictures. As a sensible instance, we’ll give attention to classifying pictures as canine or cats, in a dataset containing 4,000 footage of cats and canine (2,000 cats, 2,000 canine). We’ll use 2,000 footage for coaching – 1,000 for validation, and 1,000 for testing.

In Chapter 5 of the Deep Studying with R ebook we evaluation three strategies for tackling this downside. The primary of those is coaching a small mannequin from scratch on what little information you’ve got (which achieves an accuracy of 82%). Subsequently we use function extraction with a pretrained community (leading to an accuracy of 90%) and fine-tuning a pretrained community (with a closing accuracy of 97%). On this submit we’ll cowl solely the second and third strategies.

The relevance of deep studying for small-data issues

You’ll typically hear that deep studying solely works when a lot of information is offered. That is legitimate partly: one basic attribute of deep studying is that it might probably discover attention-grabbing options within the coaching information by itself, with none want for handbook function engineering, and this will solely be achieved when a lot of coaching examples can be found. That is very true for issues the place the enter samples are very high-dimensional, like pictures.

However what constitutes a lot of samples is relative – relative to the scale and depth of the community you’re attempting to coach, for starters. It isn’t potential to coach a convnet to unravel a fancy downside with only a few tens of samples, however just a few hundred can doubtlessly suffice if the mannequin is small and nicely regularized and the duty is straightforward. As a result of convnets study native, translation-invariant options, they’re extremely information environment friendly on perceptual issues. Coaching a convnet from scratch on a really small picture dataset will nonetheless yield cheap outcomes regardless of a relative lack of knowledge, with out the necessity for any customized function engineering. You’ll see this in motion on this part.

What’s extra, deep-learning fashions are by nature extremely repurposable: you’ll be able to take, say, an image-classification or speech-to-text mannequin educated on a large-scale dataset and reuse it on a considerably completely different downside with solely minor modifications. Particularly, within the case of pc imaginative and prescient, many pretrained fashions (normally educated on the ImageNet dataset) at the moment are publicly out there for obtain and can be utilized to bootstrap highly effective imaginative and prescient fashions out of little or no information. That’s what you’ll do within the subsequent part. Let’s begin by getting your arms on the information.

Downloading the information

The Canines vs. Cats dataset that you simply’ll use isn’t packaged with Keras. It was made out there by Kaggle as a part of a computer-vision competitors in late 2013, again when convnets weren’t mainstream. You may obtain the unique dataset from https://www.kaggle.com/c/dogs-vs-cats/information (you’ll must create a Kaggle account in case you don’t have already got one – don’t fear, the method is painless).

The photographs are medium-resolution colour JPEGs. Listed below are some examples:

Unsurprisingly, the dogs-versus-cats Kaggle competitors in 2013 was gained by entrants who used convnets. The perfect entries achieved as much as 95% accuracy. Beneath you’ll find yourself with a 97% accuracy, although you’ll practice your fashions on lower than 10% of the information that was out there to the rivals.

This dataset accommodates 25,000 pictures of canine and cats (12,500 from every class) and is 543 MB (compressed). After downloading and uncompressing it, you’ll create a brand new dataset containing three subsets: a coaching set with 1,000 samples of every class, a validation set with 500 samples of every class, and a take a look at set with 500 samples of every class.

Following is the code to do that:

original_dataset_dir <- "~/Downloads/kaggle_original_data"

base_dir <- "~/Downloads/cats_and_dogs_small"
dir.create(base_dir)

train_dir <- file.path(base_dir, "practice")
dir.create(train_dir)
validation_dir <- file.path(base_dir, "validation")
dir.create(validation_dir)
test_dir <- file.path(base_dir, "take a look at")
dir.create(test_dir)

train_cats_dir <- file.path(train_dir, "cats")
dir.create(train_cats_dir)

train_dogs_dir <- file.path(train_dir, "canine")
dir.create(train_dogs_dir)

validation_cats_dir <- file.path(validation_dir, "cats")
dir.create(validation_cats_dir)

validation_dogs_dir <- file.path(validation_dir, "canine")
dir.create(validation_dogs_dir)

test_cats_dir <- file.path(test_dir, "cats")
dir.create(test_cats_dir)

test_dogs_dir <- file.path(test_dir, "canine")
dir.create(test_dogs_dir)

fnames <- paste0("cat.", 1:1000, ".jpg")
file.copy(file.path(original_dataset_dir, fnames), 
          file.path(train_cats_dir)) 

fnames <- paste0("cat.", 1001:1500, ".jpg")
file.copy(file.path(original_dataset_dir, fnames), 
          file.path(validation_cats_dir))

fnames <- paste0("cat.", 1501:2000, ".jpg")
file.copy(file.path(original_dataset_dir, fnames),
          file.path(test_cats_dir))

fnames <- paste0("canine.", 1:1000, ".jpg")
file.copy(file.path(original_dataset_dir, fnames),
          file.path(train_dogs_dir))

fnames <- paste0("canine.", 1001:1500, ".jpg")
file.copy(file.path(original_dataset_dir, fnames),
          file.path(validation_dogs_dir)) 

fnames <- paste0("canine.", 1501:2000, ".jpg")
file.copy(file.path(original_dataset_dir, fnames),
          file.path(test_dogs_dir))

Utilizing a pretrained convnet

A standard and extremely efficient strategy to deep studying on small picture datasets is to make use of a pretrained community. A pretrained community is a saved community that was beforehand educated on a big dataset, usually on a large-scale image-classification process. If this unique dataset is massive sufficient and normal sufficient, then the spatial hierarchy of options discovered by the pretrained community can successfully act as a generic mannequin of the visible world, and therefore its options can show helpful for a lot of completely different computer-vision issues, although these new issues could contain utterly completely different lessons than these of the unique process. For example, you would possibly practice a community on ImageNet (the place lessons are principally animals and on a regular basis objects) after which repurpose this educated community for one thing as distant as figuring out furnishings gadgets in pictures. Such portability of discovered options throughout completely different issues is a key benefit of deep studying in comparison with many older, shallow-learning approaches, and it makes deep studying very efficient for small-data issues.

On this case, let’s take into account a big convnet educated on the ImageNet dataset (1.4 million labeled pictures and 1,000 completely different lessons). ImageNet accommodates many animal lessons, together with completely different species of cats and canine, and you may thus anticipate to carry out nicely on the dogs-versus-cats classification downside.

You’ll use the VGG16 structure, developed by Karen Simonyan and Andrew Zisserman in 2014; it’s a easy and broadly used convnet structure for ImageNet. Though it’s an older mannequin, removed from the present cutting-edge and considerably heavier than many different latest fashions, I selected it as a result of its structure is much like what you’re already aware of and is simple to grasp with out introducing any new ideas. This can be your first encounter with one among these cutesy mannequin names – VGG, ResNet, Inception, Inception-ResNet, Xception, and so forth; you’ll get used to them, as a result of they are going to come up continuously in case you preserve doing deep studying for pc imaginative and prescient.

There are two methods to make use of a pretrained community: function extraction and fine-tuning. We’ll cowl each of them. Let’s begin with function extraction.

Characteristic extraction consists of utilizing the representations discovered by a earlier community to extract attention-grabbing options from new samples. These options are then run via a brand new classifier, which is educated from scratch.

As you noticed beforehand, convnets used for picture classification comprise two components: they begin with a sequence of pooling and convolution layers, they usually finish with a densely linked classifier. The primary half known as the convolutional base of the mannequin. Within the case of convnets, function extraction consists of taking the convolutional base of a beforehand educated community, operating the brand new information via it, and coaching a brand new classifier on high of the output.

Why solely reuse the convolutional base? May you reuse the densely linked classifier as nicely? Basically, doing so ought to be prevented. The reason being that the representations discovered by the convolutional base are prone to be extra generic and subsequently extra reusable: the function maps of a convnet are presence maps of generic ideas over an image, which is prone to be helpful whatever the computer-vision downside at hand. However the representations discovered by the classifier will essentially be particular to the set of lessons on which the mannequin was educated – they are going to solely comprise details about the presence chance of this or that class in all the image. Moreover, representations present in densely linked layers not comprise any details about the place objects are positioned within the enter picture: these layers do away with the notion of area, whereas the item location continues to be described by convolutional function maps. For issues the place object location issues, densely linked options are largely ineffective.

Word that the extent of generality (and subsequently reusability) of the representations extracted by particular convolution layers is dependent upon the depth of the layer within the mannequin. Layers that come earlier within the mannequin extract native, extremely generic function maps (resembling visible edges, colours, and textures), whereas layers which can be increased up extract more-abstract ideas (resembling “cat ear” or “canine eye”). So in case your new dataset differs loads from the dataset on which the unique mannequin was educated, you might be higher off utilizing solely the primary few layers of the mannequin to do function extraction, slightly than utilizing all the convolutional base.

On this case, as a result of the ImageNet class set accommodates a number of canine and cat lessons, it’s prone to be useful to reuse the knowledge contained within the densely linked layers of the unique mannequin. However we’ll select to not, as a way to cowl the extra normal case the place the category set of the brand new downside doesn’t overlap the category set of the unique mannequin.

Let’s put this in observe through the use of the convolutional base of the VGG16 community, educated on ImageNet, to extract attention-grabbing options from cat and canine pictures, after which practice a dogs-versus-cats classifier on high of those options.

The VGG16 mannequin, amongst others, comes prepackaged with Keras. Right here’s the record of image-classification fashions (all pretrained on the ImageNet dataset) which can be out there as a part of Keras:

• Xception
• Inception V3
• ResNet50
• VGG16
• VGG19
• MobileNet

Let’s instantiate the VGG16 mannequin.

library(keras)

conv_base <- application_vgg16(
  weights = "imagenet",
  include_top = FALSE,
  input_shape = c(150, 150, 3)
)

You move three arguments to the perform:

• weights specifies the burden checkpoint from which to initialize the mannequin.
• include_top refers to together with (or not) the densely linked classifier on high of the community. By default, this densely linked classifier corresponds to the 1,000 lessons from ImageNet. Since you intend to make use of your individual densely linked classifier (with solely two lessons: cat and canine), you don’t want to incorporate it.
• input_shape is the form of the picture tensors that you simply’ll feed to the community. This argument is solely non-obligatory: in case you don’t move it, the community will be capable to course of inputs of any dimension.

Right here’s the element of the structure of the VGG16 convolutional base. It’s much like the easy convnets you’re already aware of:

Layer (sort)                     Output Form          Param #  
================================================================
input_1 (InputLayer)             (None, 150, 150, 3)   0       
________________________________________________________________
block1_conv1 (Convolution2D)     (None, 150, 150, 64)  1792     
________________________________________________________________
block1_conv2 (Convolution2D)     (None, 150, 150, 64)  36928    
________________________________________________________________
block1_pool (MaxPooling2D)       (None, 75, 75, 64)    0        
________________________________________________________________
block2_conv1 (Convolution2D)     (None, 75, 75, 128)   73856    
________________________________________________________________
block2_conv2 (Convolution2D)     (None, 75, 75, 128)   147584   
________________________________________________________________
block2_pool (MaxPooling2D)       (None, 37, 37, 128)   0        
________________________________________________________________
block3_conv1 (Convolution2D)     (None, 37, 37, 256)   295168   
________________________________________________________________
block3_conv2 (Convolution2D)     (None, 37, 37, 256)   590080   
________________________________________________________________
block3_conv3 (Convolution2D)     (None, 37, 37, 256)   590080   
________________________________________________________________
block3_pool (MaxPooling2D)       (None, 18, 18, 256)   0        
________________________________________________________________
block4_conv1 (Convolution2D)     (None, 18, 18, 512)   1180160  
________________________________________________________________
block4_conv2 (Convolution2D)     (None, 18, 18, 512)   2359808  
________________________________________________________________
block4_conv3 (Convolution2D)     (None, 18, 18, 512)   2359808  
________________________________________________________________
block4_pool (MaxPooling2D)       (None, 9, 9, 512)     0        
________________________________________________________________
block5_conv1 (Convolution2D)     (None, 9, 9, 512)     2359808  
________________________________________________________________
block5_conv2 (Convolution2D)     (None, 9, 9, 512)     2359808  
________________________________________________________________
block5_conv3 (Convolution2D)     (None, 9, 9, 512)     2359808  
________________________________________________________________
block5_pool (MaxPooling2D)       (None, 4, 4, 512)     0        
================================================================
Complete params: 14,714,688
Trainable params: 14,714,688
Non-trainable params: 0

The ultimate function map has form (4, 4, 512). That’s the function on high of which you’ll stick a densely linked classifier.

At this level, there are two methods you can proceed:

• Operating the convolutional base over your dataset, recording its output to an array on disk, after which utilizing this information as enter to a standalone, densely linked classifier much like these you noticed partly 1 of this ebook. This answer is quick and low-cost to run, as a result of it solely requires operating the convolutional base as soon as for each enter picture, and the convolutional base is by far the costliest a part of the pipeline. However for a similar purpose, this method gained’t will let you use information augmentation.

• Extending the mannequin you’ve got (conv_base) by including dense layers on high, and operating the entire thing finish to finish on the enter information. This can will let you use information augmentation, as a result of each enter picture goes via the convolutional base each time it’s seen by the mannequin. However for a similar purpose, this method is much dearer than the primary.

On this submit we’ll cowl the second method intimately (within the ebook we cowl each). Word that this method is so costly that it is best to solely try it you probably have entry to a GPU – it’s completely intractable on a CPU.

As a result of fashions behave identical to layers, you’ll be able to add a mannequin (like conv_base) to a sequential mannequin identical to you’ll add a layer.

mannequin <- keras_model_sequential() %>% 
  conv_base %>% 
  layer_flatten() %>% 
  layer_dense(items = 256, activation = "relu") %>% 
  layer_dense(items = 1, activation = "sigmoid")

That is what the mannequin seems like now:

Layer (sort)                     Output Form          Param #  
================================================================
vgg16 (Mannequin)                    (None, 4, 4, 512)     14714688                                     
________________________________________________________________
flatten_1 (Flatten)              (None, 8192)          0        
________________________________________________________________
dense_1 (Dense)                  (None, 256)           2097408  
________________________________________________________________
dense_2 (Dense)                  (None, 1)             257      
================================================================
Complete params: 16,812,353
Trainable params: 16,812,353
Non-trainable params: 0

As you’ll be able to see, the convolutional base of VGG16 has 14,714,688 parameters, which may be very massive. The classifier you’re including on high has 2 million parameters.

Earlier than you compile and practice the mannequin, it’s crucial to freeze the convolutional base. Freezing a layer or set of layers means stopping their weights from being up to date throughout coaching. For those who don’t do that, then the representations that have been beforehand discovered by the convolutional base can be modified throughout coaching. As a result of the dense layers on high are randomly initialized, very massive weight updates can be propagated via the community, successfully destroying the representations beforehand discovered.

In Keras, you freeze a community utilizing the freeze_weights() perform:

size(mannequin$trainable_weights)

[1] 30

freeze_weights(conv_base)
size(mannequin$trainable_weights)

[1] 4

With this setup, solely the weights from the 2 dense layers that you simply added can be educated. That’s a complete of 4 weight tensors: two per layer (the primary weight matrix and the bias vector). Word that to ensure that these modifications to take impact, you could first compile the mannequin. For those who ever modify weight trainability after compilation, it is best to then recompile the mannequin, or these modifications can be ignored.

Utilizing information augmentation

Overfitting is brought on by having too few samples to study from, rendering you unable to coach a mannequin that may generalize to new information. Given infinite information, your mannequin can be uncovered to each potential facet of the information distribution at hand: you’ll by no means overfit. Knowledge augmentation takes the strategy of producing extra coaching information from present coaching samples, by augmenting the samples by way of plenty of random transformations that yield believable-looking pictures. The objective is that at coaching time, your mannequin won’t ever see the very same image twice. This helps expose the mannequin to extra points of the information and generalize higher.

In Keras, this may be completed by configuring plenty of random transformations to be carried out on the pictures learn by an image_data_generator(). For instance:

train_datagen = image_data_generator(
  rescale = 1/255,
  rotation_range = 40,
  width_shift_range = 0.2,
  height_shift_range = 0.2,
  shear_range = 0.2,
  zoom_range = 0.2,
  horizontal_flip = TRUE,
  fill_mode = "nearest"
)

These are only a few of the choices out there (for extra, see the Keras documentation). Let’s shortly go over this code:

• rotation_range is a price in levels (0–180), a variety inside which to randomly rotate footage.
• width_shift and height_shift are ranges (as a fraction of complete width or peak) inside which to randomly translate footage vertically or horizontally.
• shear_range is for randomly making use of shearing transformations.
• zoom_range is for randomly zooming inside footage.
• horizontal_flip is for randomly flipping half the pictures horizontally – related when there aren’t any assumptions of horizontal asymmetry (for instance, real-world footage).
• fill_mode is the technique used for filling in newly created pixels, which might seem after a rotation or a width/peak shift.

Now we are able to practice our mannequin utilizing the picture information generator:

# Word that the validation information should not be augmented!
test_datagen <- image_data_generator(rescale = 1/255)  

train_generator <- flow_images_from_directory(
  train_dir,                  # Goal listing  
  train_datagen,              # Knowledge generator
  target_size = c(150, 150),  # Resizes all pictures to 150 × 150
  batch_size = 20,
  class_mode = "binary"       # binary_crossentropy loss for binary labels
)

validation_generator <- flow_images_from_directory(
  validation_dir,
  test_datagen,
  target_size = c(150, 150),
  batch_size = 20,
  class_mode = "binary"
)

mannequin %>% compile(
  loss = "binary_crossentropy",
  optimizer = optimizer_rmsprop(lr = 2e-5),
  metrics = c("accuracy")
)

historical past <- mannequin %>% fit_generator(
  train_generator,
  steps_per_epoch = 100,
  epochs = 30,
  validation_data = validation_generator,
  validation_steps = 50
)

Let’s plot the outcomes. As you’ll be able to see, you attain a validation accuracy of about 90%.

High quality-tuning

One other broadly used method for mannequin reuse, complementary to function extraction, is fine-tuning
High quality-tuning consists of unfreezing just a few of the highest layers of a frozen mannequin base used for function extraction, and collectively coaching each the newly added a part of the mannequin (on this case, the absolutely linked classifier) and these high layers. That is referred to as fine-tuning as a result of it barely adjusts the extra summary
representations of the mannequin being reused, as a way to make them extra related for the issue at hand.

I said earlier that it’s essential to freeze the convolution base of VGG16 so as to have the ability to practice a randomly initialized classifier on high. For a similar purpose, it’s solely potential to fine-tune the highest layers of the convolutional base as soon as the classifier on high has already been educated. If the classifier isn’t already educated, then the error sign propagating via the community throughout coaching can be too massive, and the representations beforehand discovered by the layers being fine-tuned can be destroyed. Thus the steps for fine-tuning a community are as follows:

• Add your customized community on high of an already-trained base community.
• Freeze the bottom community.
• Practice the half you added.
• Unfreeze some layers within the base community.
• Collectively practice each these layers and the half you added.

You already accomplished the primary three steps when doing function extraction. Let’s proceed with step 4: you’ll unfreeze your conv_base after which freeze particular person layers inside it.

As a reminder, that is what your convolutional base seems like:

Layer (sort)                     Output Form          Param #  
================================================================
input_1 (InputLayer)             (None, 150, 150, 3)   0        
________________________________________________________________
block1_conv1 (Convolution2D)     (None, 150, 150, 64)  1792     
________________________________________________________________
block1_conv2 (Convolution2D)     (None, 150, 150, 64)  36928    
________________________________________________________________
block1_pool (MaxPooling2D)       (None, 75, 75, 64)    0        
________________________________________________________________
block2_conv1 (Convolution2D)     (None, 75, 75, 128)   73856    
________________________________________________________________
block2_conv2 (Convolution2D)     (None, 75, 75, 128)   147584   
________________________________________________________________
block2_pool (MaxPooling2D)       (None, 37, 37, 128)   0        
________________________________________________________________
block3_conv1 (Convolution2D)     (None, 37, 37, 256)   295168   
________________________________________________________________
block3_conv2 (Convolution2D)     (None, 37, 37, 256)   590080   
________________________________________________________________
block3_conv3 (Convolution2D)     (None, 37, 37, 256)   590080   
________________________________________________________________
block3_pool (MaxPooling2D)       (None, 18, 18, 256)   0        
________________________________________________________________
block4_conv1 (Convolution2D)     (None, 18, 18, 512)   1180160  
________________________________________________________________
block4_conv2 (Convolution2D)     (None, 18, 18, 512)   2359808  
________________________________________________________________
block4_conv3 (Convolution2D)     (None, 18, 18, 512)   2359808  
________________________________________________________________
block4_pool (MaxPooling2D)       (None, 9, 9, 512)     0        
________________________________________________________________
block5_conv1 (Convolution2D)     (None, 9, 9, 512)     2359808  
________________________________________________________________
block5_conv2 (Convolution2D)     (None, 9, 9, 512)     2359808  
________________________________________________________________
block5_conv3 (Convolution2D)     (None, 9, 9, 512)     2359808  
________________________________________________________________
block5_pool (MaxPooling2D)       (None, 4, 4, 512)     0        
================================================================
Complete params: 14714688

You’ll fine-tune the entire layers from block3_conv1 and on. Why not fine-tune all the convolutional base? You may. However you have to take into account the next:

• Earlier layers within the convolutional base encode more-generic, reusable options, whereas layers increased up encode more-specialized options. It’s extra helpful to fine-tune the extra specialised options, as a result of these are those that have to be repurposed in your new downside. There can be fast-decreasing returns in fine-tuning decrease layers.
• The extra parameters you’re coaching, the extra you’re susceptible to overfitting. The convolutional base has 15 million parameters, so it will be dangerous to aim to coach it in your small dataset.

Thus, on this state of affairs, it’s an excellent technique to fine-tune solely a number of the layers within the convolutional base. Let’s set this up, ranging from the place you left off within the earlier instance.

unfreeze_weights(conv_base, from = "block3_conv1")

Now you’ll be able to start fine-tuning the community. You’ll do that with the RMSProp optimizer, utilizing a really low studying charge. The explanation for utilizing a low studying charge is that you simply wish to restrict the magnitude of the modifications you make to the representations of the three layers you’re fine-tuning. Updates which can be too massive could hurt these representations.

mannequin %>% compile(
  loss = "binary_crossentropy",
  optimizer = optimizer_rmsprop(lr = 1e-5),
  metrics = c("accuracy")
)

historical past <- mannequin %>% fit_generator(
  train_generator,
  steps_per_epoch = 100,
  epochs = 100,
  validation_data = validation_generator,
  validation_steps = 50
)

Let’s plot our outcomes:

You’re seeing a pleasant 6% absolute enchancment in accuracy, from about 90% to above 96%.

Word that the loss curve doesn’t present any actual enchancment (in reality, it’s deteriorating). Chances are you’ll surprise, how might accuracy keep steady or enhance if the loss isn’t lowering? The reply is straightforward: what you show is a mean of pointwise loss values; however what issues for accuracy is the distribution of the loss values, not their common, as a result of accuracy is the results of a binary thresholding of the category chance predicted by the mannequin. The mannequin should still be bettering even when this isn’t mirrored within the common loss.

Now you can lastly consider this mannequin on the take a look at information:

test_generator <- flow_images_from_directory(
  test_dir,
  test_datagen,
  target_size = c(150, 150),
  batch_size = 20,
  class_mode = "binary"
)

mannequin %>% evaluate_generator(test_generator, steps = 50)

$loss
[1] 0.2158171

$acc
[1] 0.965

Right here you get a take a look at accuracy of 96.5%. Within the unique Kaggle competitors round this dataset, this is able to have been one of many high outcomes. However utilizing trendy deep-learning strategies, you managed to achieve this consequence utilizing solely a small fraction of the coaching information out there (about 10%). There’s a enormous distinction between with the ability to practice on 20,000 samples in comparison with 2,000 samples!

Take-aways: utilizing convnets with small datasets

Right here’s what it is best to take away from the workout routines up to now two sections:

• Convnets are the very best sort of machine-learning fashions for computer-vision duties. It’s potential to coach one from scratch even on a really small dataset, with respectable outcomes.
• On a small dataset, overfitting would be the foremost difficulty. Knowledge augmentation is a strong approach to battle overfitting while you’re working with picture information.
• It’s simple to reuse an present convnet on a brand new dataset by way of function extraction. This can be a beneficial method for working with small picture datasets.
• As a complement to function extraction, you should utilize fine-tuning, which adapts to a brand new downside a number of the representations beforehand discovered by an present mannequin. This pushes efficiency a bit additional.

Now you’ve got a strong set of instruments for coping with image-classification issues – particularly with small datasets.