using
Pkg
Pkg
.
activate
(
"
../../FastAI.jl/
"
)
Pkg
.
instantiate
(
)
Activating project at `~/Desktop/dev/FastAI.jl`
using
Flux
┌ Info: Precompiling FastAI [5d0beca9-ade8-49ae-ad0b-a3cf890e669f]
└ @ Base loading.jl:1423
┌ Info: Precompiling FastTimeSeries [5337c758-7610-4451-a331-8357b11df7c6]
└ @ Base loading.jl:1423
data
,
blocks
=
load
(
datarecipes
(
)
[
"
ecg5000
"
]
)
;
getobs gets us a sample from the TimeSeriesDataset. It returns a tuple with the input time series and the correspodning label.
input
,
class
=
sample
=
getobs
(
data
,
25
)
(Float32[-0.28834122 -2.2725453 … 1.722784 1.2959242], "1")
Now we create a learning task for time-series classification. This means using the time-series to predict labels. We will use the
TimeSeriesRow block as input and
Label block as the target.
task
=
SupervisedTask
(
blocks
,
(
OneHot
(
)
,
setup
(
TSPreprocessing
,
blocks
[
1
]
,
data
[
1
]
.
table
)
)
)
SupervisedTask(TimeSeriesRow -> Label{SubString{String}})The encodings passed in transform samples into formats suitable as inputs and outputs for a model
Let's check that samples from the created data container conform to the blocks of the learning task:
checkblock
(
task
.
blocks
.
sample
,
sample
)
trueTo get an overview of the learning task created, and as a sanity test, we can use describetask. This shows us what encodings will be applied to which blocks, and how the predicted ŷ values are decoded.
describetask
(
task
)
encoded_sample
=
encodesample
(
task
,
Training
(
)
,
sample
)
(Float32[-0.28937635 -2.2807038 … 1.7289687 1.3005764], Bool[1, 0, 0, 0, 0])
sample
=
getobs
(
data
,
1
)
(Float32[-0.11252183 -2.8272038 … 0.92528623 0.19313742], "1")
showsample
(
task
,
sample
)
showblock
(
blocks
[
1
]
,
sample
[
1
]
)
We will use a StackedLSTM as a backbone model, and a Dense layer at the front for classification.
taskmodel knows how to do this by looking at the datablocks used.
backbone
=
FastTimeSeries
.
Models
.
StackedLSTM
(
1
,
16
,
10
,
2
)
;
We can
tasklossfn to get a loss function suitable for our task.
lossfn
=
tasklossfn
(
task
)
logitcrossentropy (generic function with 1 method)Next we create a pair of training and validation data loaders. They take care of batching and loading the data in parallel in the background.
traindl
,
validdl
=
taskdataloaders
(
data
,
task
,
16
)
;
We will use an
Adam optimzer for this task.
optimizer
=
ADAM
(
0.002
)
ADAM(0.002, (0.9, 0.999), 1.0e-8, IdDict{Any, Any}())We create callbacks to get the accuracy during the training
With the addition of an optimizer and a loss function, we can now create a Learner and start training.
learner
=
Learner
(
model
,
lossfn
;
data
=
(
traindl
,
validdl
)
,
optimizer
=
optimizer
,
callbacks
=
callbacks
)
;
fitonecycle!
(
learner
,
10
,
0.002
)
┌ Info: The GPU function is being called but the GPU is not accessible.
│ Defaulting back to the CPU. (No action is required if you want to run on the CPU).
└ @ Flux /Users/saksham/.julia/packages/Flux/js6mP/src/functor.jl:192
Epoch 1 TrainingPhase(): 100%|██████████████████████████| Time: 0:00:40
┌───────────────┬───────┬─────────┬──────────┐
│ Phase │ Epoch │ Loss │ Accuracy │
├───────────────┼───────┼─────────┼──────────┤
│ TrainingPhase │ 1.0 │ 0.95453 │ 0.65725 │
└───────────────┴───────┴─────────┴──────────┘
Epoch 1 ValidationPhase(): 100%|████████████████████████| Time: 0:00:00
┌─────────────────┬───────┬─────────┬──────────┐
│ Phase │ Epoch │ Loss │ Accuracy │
├─────────────────┼───────┼─────────┼──────────┤
│ ValidationPhase │ 1.0 │ 0.36429 │ 0.9082 │
└─────────────────┴───────┴─────────┴──────────┘
Epoch 2 TrainingPhase(): 100%|██████████████████████████| Time: 0:00:03
┌───────────────┬───────┬─────────┬──────────┐
│ Phase │ Epoch │ Loss │ Accuracy │
├───────────────┼───────┼─────────┼──────────┤
│ TrainingPhase │ 2.0 │ 0.30034 │ 0.9205 │
└───────────────┴───────┴─────────┴──────────┘
Epoch 2 ValidationPhase(): 100%|████████████████████████| Time: 0:00:00
┌─────────────────┬───────┬─────────┬──────────┐
│ Phase │ Epoch │ Loss │ Accuracy │
├─────────────────┼───────┼─────────┼──────────┤
│ ValidationPhase │ 2.0 │ 0.28543 │ 0.91211 │
└─────────────────┴───────┴─────────┴──────────┘
Epoch 3 TrainingPhase(): 100%|██████████████████████████| Time: 0:00:03
┌───────────────┬───────┬────────┬──────────┐
│ Phase │ Epoch │ Loss │ Accuracy │
├───────────────┼───────┼────────┼──────────┤
│ TrainingPhase │ 3.0 │ 0.2677 │ 0.92825 │
└───────────────┴───────┴────────┴──────────┘
Epoch 3 ValidationPhase(): 100%|████████████████████████| Time: 0:00:00
┌─────────────────┬───────┬─────────┬──────────┐
│ Phase │ Epoch │ Loss │ Accuracy │
├─────────────────┼───────┼─────────┼──────────┤
│ ValidationPhase │ 3.0 │ 0.26776 │ 0.91895 │
└─────────────────┴───────┴─────────┴──────────┘
Epoch 4 TrainingPhase(): 100%|██████████████████████████| Time: 0:00:03
┌───────────────┬───────┬─────────┬──────────┐
│ Phase │ Epoch │ Loss │ Accuracy │
├───────────────┼───────┼─────────┼──────────┤
│ TrainingPhase │ 4.0 │ 0.23461 │ 0.9355 │
└───────────────┴───────┴─────────┴──────────┘
Epoch 4 ValidationPhase(): 100%|████████████████████████| Time: 0:00:00
┌─────────────────┬───────┬─────────┬──────────┐
│ Phase │ Epoch │ Loss │ Accuracy │
├─────────────────┼───────┼─────────┼──────────┤
│ ValidationPhase │ 4.0 │ 0.27086 │ 0.92285 │
└─────────────────┴───────┴─────────┴──────────┘
Epoch 5 TrainingPhase(): 100%|██████████████████████████| Time: 0:00:03
┌───────────────┬───────┬─────────┬──────────┐
│ Phase │ Epoch │ Loss │ Accuracy │
├───────────────┼───────┼─────────┼──────────┤
│ TrainingPhase │ 5.0 │ 0.22571 │ 0.9375 │
└───────────────┴───────┴─────────┴──────────┘
Epoch 5 ValidationPhase(): 100%|████████████████████████| Time: 0:00:00
┌─────────────────┬───────┬─────────┬──────────┐
│ Phase │ Epoch │ Loss │ Accuracy │
├─────────────────┼───────┼─────────┼──────────┤
│ ValidationPhase │ 5.0 │ 0.24774 │ 0.93457 │
└─────────────────┴───────┴─────────┴──────────┘
Epoch 6 TrainingPhase(): 100%|██████████████████████████| Time: 0:00:03
┌───────────────┬───────┬─────────┬──────────┐
│ Phase │ Epoch │ Loss │ Accuracy │
├───────────────┼───────┼─────────┼──────────┤
│ TrainingPhase │ 6.0 │ 0.21649 │ 0.9385 │
└───────────────┴───────┴─────────┴──────────┘
Epoch 6 ValidationPhase(): 100%|████████████████████████| Time: 0:00:00
┌─────────────────┬───────┬─────────┬──────────┐
│ Phase │ Epoch │ Loss │ Accuracy │
├─────────────────┼───────┼─────────┼──────────┤
│ ValidationPhase │ 6.0 │ 0.24026 │ 0.93359 │
└─────────────────┴───────┴─────────┴──────────┘
Epoch 7 TrainingPhase(): 100%|██████████████████████████| Time: 0:00:03
┌───────────────┬───────┬─────────┬──────────┐
│ Phase │ Epoch │ Loss │ Accuracy │
├───────────────┼───────┼─────────┼──────────┤
│ TrainingPhase │ 7.0 │ 0.21095 │ 0.93825 │
└───────────────┴───────┴─────────┴──────────┘
Epoch 7 ValidationPhase(): 100%|████████████████████████| Time: 0:00:00
┌─────────────────┬───────┬─────────┬──────────┐
│ Phase │ Epoch │ Loss │ Accuracy │
├─────────────────┼───────┼─────────┼──────────┤
│ ValidationPhase │ 7.0 │ 0.23704 │ 0.93262 │
└─────────────────┴───────┴─────────┴──────────┘
Epoch 8 TrainingPhase(): 100%|██████████████████████████| Time: 0:00:03
┌───────────────┬───────┬─────────┬──────────┐
│ Phase │ Epoch │ Loss │ Accuracy │
├───────────────┼───────┼─────────┼──────────┤
│ TrainingPhase │ 8.0 │ 0.20555 │ 0.93975 │
└───────────────┴───────┴─────────┴──────────┘
Epoch 8 ValidationPhase(): 100%|████████████████████████| Time: 0:00:00
┌─────────────────┬───────┬─────────┬──────────┐
│ Phase │ Epoch │ Loss │ Accuracy │
├─────────────────┼───────┼─────────┼──────────┤
│ ValidationPhase │ 8.0 │ 0.24263 │ 0.93359 │
└─────────────────┴───────┴─────────┴──────────┘
Epoch 9 TrainingPhase(): 100%|██████████████████████████| Time: 0:00:03
┌───────────────┬───────┬─────────┬──────────┐
│ Phase │ Epoch │ Loss │ Accuracy │
├───────────────┼───────┼─────────┼──────────┤
│ TrainingPhase │ 9.0 │ 0.20291 │ 0.94075 │
└───────────────┴───────┴─────────┴──────────┘
Epoch 9 ValidationPhase(): 100%|████████████████████████| Time: 0:00:00
┌─────────────────┬───────┬─────────┬──────────┐
│ Phase │ Epoch │ Loss │ Accuracy │
├─────────────────┼───────┼─────────┼──────────┤
│ ValidationPhase │ 9.0 │ 0.23519 │ 0.93457 │
└─────────────────┴───────┴─────────┴──────────┘
Epoch 10 TrainingPhase(): 100%|█████████████████████████| Time: 0:00:03
┌───────────────┬───────┬─────────┬──────────┐
│ Phase │ Epoch │ Loss │ Accuracy │
├───────────────┼───────┼─────────┼──────────┤
│ TrainingPhase │ 10.0 │ 0.19846 │ 0.942 │
└───────────────┴───────┴─────────┴──────────┘
Epoch 10 ValidationPhase(): 100%|███████████████████████| Time: 0:00:00
┌─────────────────┬───────┬─────────┬──────────┐
│ Phase │ Epoch │ Loss │ Accuracy │
├─────────────────┼───────┼─────────┼──────────┤
│ ValidationPhase │ 10.0 │ 0.23493 │ 0.93457 │
└─────────────────┴───────┴─────────┴──────────┘
We can save the model for later inference using
savetaskmodel:
savetaskmodel
(
"
tsclassification.jld2
"
,
task
,
learner
.
model
;
force
=
true
)