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 probably encounter in apply if you happen to ever do pc imaginative and prescient in knowledgeable context. A “few” samples can imply wherever from a number of hundred to some tens of hundreds of photographs. As a sensible instance, we’ll give attention to classifying photographs as canines or cats, in a dataset containing 4,000 photos of cats and canines (2,000 cats, 2,000 canines). We’ll use 2,000 photos for coaching – 1,000 for validation, and 1,000 for testing.

In Chapter 5 of the Deep Studying with R ebook we assessment three methods for tackling this drawback. The primary of those is coaching a small mannequin from scratch on what little information you could have (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 last accuracy of 97%). On this publish we’ll cowl solely the second and third methods.

The relevance of deep studying for small-data issues

You’ll generally hear that deep studying solely works when plenty of information is on the market. That is legitimate partly: one basic attribute of deep studying is that it will 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 plenty of coaching examples can be found. That is very true for issues the place the enter samples are very high-dimensional, like photographs.

However what constitutes plenty of samples is relative – relative to the dimensions and depth of the community you’re making an attempt to coach, for starters. It isn’t doable to coach a convnet to resolve a fancy drawback with only a few tens of samples, however a number of hundred can doubtlessly suffice if the mannequin is small and properly regularized and the duty is straightforward. As a result of convnets be taught 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 information, 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 possibly can take, say, an image-classification or speech-to-text mannequin skilled on a large-scale dataset and reuse it on a considerably completely different drawback with solely minor modifications. Particularly, within the case of pc imaginative and prescient, many pretrained fashions (often skilled 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 possibly can obtain the unique dataset from https://www.kaggle.com/c/dogs-vs-cats/information (you’ll must create a Kaggle account if you happen to don’t have already got one – don’t fear, the method is painless).

The photographs are medium-resolution coloration JPEGs. Listed here are some examples:

Unsurprisingly, the dogs-versus-cats Kaggle competitors in 2013 was received by entrants who used convnets. The very best entries achieved as much as 95% accuracy. Beneath you’ll find yourself with a 97% accuracy, despite the fact that you’ll prepare your fashions on lower than 10% of the information that was out there to the opponents.

This dataset accommodates 25,000 photographs of canines 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 check 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, "prepare")
dir.create(train_dir)
validation_dir <- file.path(base_dir, "validation")
dir.create(validation_dir)
test_dir <- file.path(base_dir, "check")
dir.create(test_dir)

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

train_dogs_dir <- file.path(train_dir, "canines")
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, "canines")
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, "canines")
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 typical and extremely efficient method 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 skilled on a big dataset, usually on a large-scale image-classification process. If this unique dataset is giant sufficient and normal sufficient, then the spatial hierarchy of options realized 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, despite the fact that these new issues could contain fully completely different courses than these of the unique process. For example, you may prepare a community on ImageNet (the place courses are largely animals and on a regular basis objects) after which repurpose this skilled community for one thing as distant as figuring out furnishings gadgets in photographs. Such portability of realized 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 contemplate a big convnet skilled on the ImageNet dataset (1.4 million labeled photographs and 1,000 completely different courses). ImageNet accommodates many animal courses, together with completely different species of cats and canines, and you’ll thus anticipate to carry out properly on the dogs-versus-cats classification drawback.

You’ll use the VGG16 structure, developed by Karen Simonyan and Andrew Zisserman in 2014; it’s a easy and extensively used convnet structure for ImageNet. Though it’s an older mannequin, removed from the present state-of-the-art and considerably heavier than many different current fashions, I selected it as a result of its structure is much like what you’re already aware of and is straightforward to know with out introducing any new ideas. This can be your first encounter with one in all these cutesy mannequin names – VGG, ResNet, Inception, Inception-ResNet, Xception, and so forth; you’ll get used to them, as a result of they’ll come up ceaselessly if you happen to hold 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 realized by a earlier community to extract attention-grabbing options from new samples. These options are then run via a brand new classifier, which is skilled from scratch.

As you noticed beforehand, convnets used for picture classification comprise two elements: they begin with a sequence of pooling and convolution layers, and so they finish with a densely related 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 skilled 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 related classifier as properly? On the whole, doing so needs to be averted. The reason being that the representations realized by the convolutional base are more likely 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 more likely to be helpful whatever the computer-vision drawback at hand. However the representations realized by the classifier will essentially be particular to the set of courses on which the mannequin was skilled – they’ll solely include details about the presence chance of this or that class in the whole image. Moreover, representations present in densely related layers now not include any details about the place objects are positioned within the enter picture: these layers eliminate the notion of house, whereas the item location remains to be described by convolutional function maps. For issues the place object location issues, densely related 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 (reminiscent of visible edges, colours, and textures), whereas layers which can be increased up extract more-abstract ideas (reminiscent of “cat ear” or “canine eye”). So in case your new dataset differs lots from the dataset on which the unique mannequin was skilled, you might be higher off utilizing solely the primary few layers of the mannequin to do function extraction, slightly than utilizing the whole convolutional base.

On this case, as a result of the ImageNet class set accommodates a number of canine and cat courses, it’s more likely to be helpful to reuse the knowledge contained within the densely related layers of the unique mannequin. However we’ll select to not, with a purpose to cowl the extra normal case the place the category set of the brand new drawback doesn’t overlap the category set of the unique mannequin.

Let’s put this in apply by utilizing the convolutional base of the VGG16 community, skilled on ImageNet, to extract attention-grabbing options from cat and canine photographs, after which prepare a dogs-versus-cats classifier on high of those options.

The VGG16 mannequin, amongst others, comes prepackaged with Keras. Right here’s the checklist 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 related classifier on high of the community. By default, this densely related classifier corresponds to the 1,000 courses from ImageNet. Since you intend to make use of your personal densely related classifier (with solely two courses: 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 only optionally available: if you happen to 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 straightforward convnets you’re already aware of:

Layer (kind)                     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 related classifier.

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

• Working 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 related classifier much like these you noticed partly 1 of this ebook. This resolution 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 cause, this method received’t will let you use information augmentation.

• Extending the mannequin you could have (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 cause, this method is much costlier than the primary.

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

As a result of fashions behave identical to layers, you possibly can add a mannequin (like conv_base) to a sequential mannequin identical to you’d add a layer.

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

That is what the mannequin appears like now:

Layer (kind)                     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 possibly can see, the convolutional base of VGG16 has 14,714,688 parameters, which could be very giant. The classifier you’re including on high has 2 million parameters.

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

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 shall be skilled. That’s a complete of 4 weight tensors: two per layer (the principle weight matrix and the bias vector). Word that to ensure that these modifications to take impact, you need to first compile the mannequin. When you ever modify weight trainability after compilation, it’s best to then recompile the mannequin, or these modifications shall be ignored.

Utilizing information augmentation

Overfitting is brought on by having too few samples to be taught from, rendering you unable to coach a mannequin that may generalize to new information. Given infinite information, your mannequin could be uncovered to each doable side of the information distribution at hand: you’d by no means overfit. Information augmentation takes the method of producing extra coaching information from current coaching samples, by augmenting the samples by way of a variety of random transformations that yield believable-looking photographs. 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 facets of the information and generalize higher.

In Keras, this may be executed by configuring a variety 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 photos.
• width_shift and height_shift are ranges (as a fraction of complete width or top) inside which to randomly translate photos vertically or horizontally.
• shear_range is for randomly making use of shearing transformations.
• zoom_range is for randomly zooming inside photos.
• horizontal_flip is for randomly flipping half the pictures horizontally – related when there are not any assumptions of horizontal asymmetry (for instance, real-world photos).
• fill_mode is the technique used for filling in newly created pixels, which might seem after a rotation or a width/top shift.

Now we will prepare 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,              # Information generator
  target_size = c(150, 150),  # Resizes all photographs 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 possibly can see, you attain a validation accuracy of about 90%.

Superb-tuning

One other extensively used method for mannequin reuse, complementary to function extraction, is fine-tuning
Superb-tuning consists of unfreezing a number of 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 totally related 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, with a purpose 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 prepare a randomly initialized classifier on high. For a similar cause, it’s solely doable to fine-tune the highest layers of the convolutional base as soon as the classifier on high has already been skilled. If the classifier isn’t already skilled, then the error sign propagating via the community throughout coaching shall be too giant, and the representations beforehand realized by the layers being fine-tuned shall 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 prepare 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 appears like:

Layer (kind)                     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 all the layers from block3_conv1 and on. Why not fine-tune the whole convolutional base? You possibly can. However it’s essential to contemplate 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 drawback. There could be fast-decreasing returns in fine-tuning decrease layers.
• The extra parameters you’re coaching, the extra you’re prone to overfitting. The convolutional base has 15 million parameters, so it might be dangerous to aim to coach it in your small dataset.

Thus, on this state of affairs, it’s a superb technique to fine-tune solely a few 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 possibly can start fine-tuning the community. You’ll do that with the RMSProp optimizer, utilizing a really low studying price. The explanation for utilizing a low studying price is that you simply need 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 giant 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). It’s possible you’ll marvel, how may accuracy keep steady or enhance if the loss isn’t reducing? 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 enhancing even when this isn’t mirrored within the common loss.

Now you can lastly consider this mannequin on the check 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 check accuracy of 96.5%. Within the unique Kaggle competitors round this dataset, this may have been one of many high outcomes. However utilizing trendy deep-learning methods, you managed to succeed in this end result utilizing solely a small fraction of the coaching information out there (about 10%). There’s a enormous distinction between with the ability to prepare on 20,000 samples in comparison with 2,000 samples!

Take-aways: utilizing convnets with small datasets

Right here’s what it’s best to take away from the workouts prior to now two sections:

• Convnets are the very best kind of machine-learning fashions for computer-vision duties. It’s doable to coach one from scratch even on a really small dataset, with first rate outcomes.
• On a small dataset, overfitting would be the foremost problem. Information augmentation is a robust option to struggle overfitting once you’re working with picture information.
• It’s straightforward to reuse an current convnet on a brand new dataset by way of function extraction. It is a worthwhile method for working with small picture datasets.
• As a complement to function extraction, you need to use fine-tuning, which adapts to a brand new drawback a few of the representations beforehand realized by an current mannequin. This pushes efficiency a bit additional.

Now you could have a strong set of instruments for coping with image-classification issues – specifically with small datasets.