Low-level ops

Intro

The fastai library simplifies training fast and accurate neural nets using modern best practices. See the fastai website to get started. The library is based on research into deep learning best practices undertaken at fast.ai, and includes “out of the box” support for vision, text, tabular, and collab (collaborative filtering) models.

Dataset

Get dataset and build cnn_learner.

URLs_MNIST_SAMPLE()
tfms = aug_transforms(do_flip = FALSE)
path = 'mnist_sample'
bs = 20
data = ImageDataLoaders_from_folder(path, batch_tfms = tfms, size = 26, bs = bs)
learn = cnn_learner(data, xresnet50_deep(), metrics = accuracy)

Modify channels to 1:

init = learn$model[0][0][0][['in_channels']]
print(init)
# 3
learn$model[0][0][0][['in_channels']] %f% 1L
print(learn$model[0][0][0][['in_channels']])
# 1

Here, one can observe a special assignment %f%. It helps for safe modification of layer parameters.

How to see and modify other parameters of the layer? First see names:

names(learn$model[0][0][0])

Cnn layer parameters:

 [1] "add_module"                "apply"                     "bfloat16"                 
 [4] "bias"                      "buffers"                   "children"                 
 [7] "cpu"                       "cuda"                      "dilation"                 
[10] "double"                    "dump_patches"              "eval"                     
[13] "extra_repr"                "float"                     "forward"                  
[16] "groups"                    "half"                      "has_children"             
[19] "in_channels"               "kernel_size"               "load_state_dict"          
[22] "modules"                   "named_buffers"             "named_children"           
[25] "named_modules"             "named_parameters"          "out_channels"             
[28] "output_padding"            "padding"                   "padding_mode"             
[31] "parameters"                "register_backward_hook"    "register_buffer"          
[34] "register_forward_hook"     "register_forward_pre_hook" "register_parameter"       
[37] "requires_grad_"            "reset_parameters"          "share_memory"             
[40] "state_dict"                "stride"                    "T_destination"            
[43] "to"                        "train"                     "training"                 
[46] "transposed"                "type"                      "weight"                   
[49] "zero_grad"   

Kernel size from (3, 3) to 9.

print(learn$model[0][0][0])
# Conv2d(1, 32, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False))
learn$model[0][0][0][['kernel_size']] %f%  reticulate::tuple(list(9L,9L))
print(learn$model[0][0][0])
# Conv2d(1, 32, kernel_size=(9, 9), stride=(2, 2), padding=(1, 1), bias=False)

Inplace tensor ops

In addition, one could replace values inside tensors with the same assignment.

For single in-place value modification:

x = tensor(c(1,2), c(3,4))
# tensor([[1., 2.],
#         [3., 4.]])
print(x[0][0])
# tensor(1.)

# Now change it to 99.
x[0][0] %f% 99
print(x[0][0])
# tensor(99.)

print(x)
# tensor([[99.,  2.],
#         [ 3.,  4.]])

Modify 2 or more values:

print(x[0])
# tensor([99.,  2.])
# change to 55, 55
x[0] %f% c(55,55)
# tensor([55., 55.])

Slicing

How to slice tensors? For less confusion, the slicing is the same as in python. narrow function requires a tensor and its dimensions. Let’s see an example:

a = tensor(array(1:100, c(3,3,3,3)))
a$shape
# torch.Size([3, 3, 3, 3])

We can extract and play with tensor dimensions:

a %>% narrow('[:,:,:,:]')

tensor([[[[ 1, 28, 55],
          [10, 37, 64],
          [19, 46, 73]],

         [[ 4, 31, 58],
          [13, 40, 67],
          [22, 49, 76]],

         [[ 7, 34, 61],
          [16, 43, 70],
          [25, 52, 79]]],


        [[[ 2, 29, 56],
          [11, 38, 65],
          [20, 47, 74]],

         [[ 5, 32, 59],
          [14, 41, 68],
          [23, 50, 77]],

         [[ 8, 35, 62],
          [17, 44, 71],
          [26, 53, 80]]],


        [[[ 3, 30, 57],
          [12, 39, 66],
          [21, 48, 75]],

         [[ 6, 33, 60],
          [15, 42, 69],
          [24, 51, 78]],

         [[ 9, 36, 63],
          [18, 45, 72],
          [27, 54, 81]]]], dtype=torch.int32)

a %>% narrow('[0,:,:,:]')

tensor([[[[ 1, 28, 55],
          [10, 37, 64],
          [19, 46, 73]],

         [[ 4, 31, 58],
          [13, 40, 67],
          [22, 49, 76]],

         [[ 7, 34, 61],
          [16, 43, 70],
          [25, 52, 79]]]], dtype=torch.int32)

a %>% narrow("[:,0,:,:]")

tensor([[[ 1, 28, 55],
         [10, 37, 64],
         [19, 46, 73]],

        [[ 2, 29, 56],
         [11, 38, 65],
         [20, 47, 74]],

        [[ 3, 30, 57],
         [12, 39, 66],
         [21, 48, 75]]], dtype=torch.int32)

a %>% narrow('[:,0,0,:]')

tensor([[ 1, 28, 55],
        [ 2, 29, 56],
        [ 3, 30, 57]], dtype=torch.int32)

a %>% narrow("[1,1,1,:]")

tensor([14, 41, 68], dtype=torch.int32)

NN module

How to build a model with fastai Module class? Simple!

Prepare data:

library(magrittr)
library(fastai)
library(zeallot)

if(!file.exists('mnist.pkl.gz')) {
  download.file('http://deeplearning.net/data/mnist/mnist.pkl.gz','mnist.pkl.gz')
  R.utils::gunzip("mnist.pkl.gz", remove=FALSE)
}
  
c(c(x_train, y_train), c(x_valid, y_valid), res) %<-%
  reticulate::py_load_object('mnist.pkl', encoding = 'latin-1')

x_train = x_train[1:500,1:784]
x_valid = x_valid[1:500,1:784]

y_train = as.integer(y_train)[1:500]
y_valid = as.integer(y_valid)[1:500]

Plot example:

example = array_reshape(x_train[1,], c(28,28))

example %>% show_image(cmap = 'gray') %>% plot()

TensorDataset = torch()$utils$data$TensorDataset

bs = 32
train_ds = TensorDataset(tensor(x_train), tensor(y_train))
valid_ds = TensorDataset(tensor(x_valid), tensor(y_valid))
train_dl = TfmdDL(train_ds, bs = bs, shuffle = TRUE)
valid_dl = TfmdDL(valid_ds, bs = 2 * bs)
dls = Data_Loaders(train_dl, valid_dl)

one = one_batch(dls)
x = one[[1]]
y = one[[2]]
x$shape; y$shape

nn = nn()
Functional = torch()$nn$functional

model = nn_module(function(self) {
  
  self$lin1 = nn$Linear(784L, 50L, bias=TRUE)
  self$lin2 = nn$Linear(50L, 10L, bias=TRUE)
  
  forward = function(y) {
    x = self$lin1(y)
    x = Functional$relu(x)
    self$lin2(x)
  }
})

learn = Learner(dls, model, loss_func=nn$CrossEntropyLoss(), metrics=accuracy)

learn %>% summary()

learn %>% fit_one_cycle(1, 1e-2)