Machine Learning in Elixir: Error in Chapter 9, identifying cats and dogs again

Community PragProg Customers
,
1

@seanmor5

the code provided I am running into an error when training the first model that uses transfer learning with ONNX model mobilenetv2-7. I am a bit stuck here, not really sure what may be wrong.

warning: Axon.Optimizers.adam/1 is deprecated. Use Polaris.Optimizers.adam/1 instead
  machine-learning-in-elixir-src/StopReinventingTheWheel.livemd#cell:s4yqqx5n4j6oyja2b2c57iqes4fvcbny:2


12:22:35.339 [warning] found unexpected key in the initial parameters map: "mobilenetv20_output_pred_fwd"
Epoch: 0, Batch: 700, accuracy: 0.7747430 loss: 0.4431267
Batch: 6, accuracy: 0.9151786 loss: 0.1636844

** (ArgumentError) argument at position 1 is not compatible with compiled function template.

{
 <<<<< Expected <<<<<
 #Nx.Tensor<
   f32[32][channels: 3][height: 160][width: 160]
 >
 ==========
 #Nx.Tensor<
   f32[26][channels: 3][height: 160][width: 160]
 >
 >>>>> Argument >>>>>
 , 
 <<<<< Expected <<<<<
 #Nx.Tensor<
   s64[32][1]
 >
 ==========
 #Nx.Tensor<
   s64[26][1]
 >
 >>>>> Argument >>>>>
 }

    (nx 0.6.2) lib/nx/defn.ex:323: anonymous fn/7 in Nx.Defn.compile_flatten/5
    (elixir 1.15.2) lib/enum.ex:1819: Enum."-map_reduce/3-lists^mapfoldl/2-0-"/3
    (nx 0.6.2) lib/nx/lazy_container.ex:61: Nx.LazyContainer.Tuple.traverse/3
    (nx 0.6.2) lib/nx/defn.ex:320: Nx.Defn.compile_flatten/5
    (nx 0.6.2) lib/nx/defn.ex:312: anonymous fn/4 in Nx.Defn.compile/3
    (stdlib 5.0.2) timer.erl:270: :timer.tc/2
    (axon 0.6.0) lib/axon/loop.ex:1805: anonymous fn/4 in Axon.Loop.run_epoch/5
    /data/machine-learning-in-elixir-src/StopReinventingTheWheel.livemd#cell:s4yqqx5n4j6oyja2b2c57iqes4fvcbny:10: (file)
2

@seanmor5

The length of the val_paths has to be divisible by the batch number so that the tensors come out the right shape.
You can just add this line

val_paths = Enum.take(val_paths, 224)

after setting the val_paths. I also removed the Enum.take line at the end because the length of the train_paths is 24000 and that is also divisible by a batch size of 32.
so that cell would look like this now:

{test_paths, train_paths} =
  Path.wildcard("train/*.jpg")
  |> Enum.shuffle()
  |> Enum.split(1000)

{test_paths, val_paths} = test_paths |> Enum.split(750)
val_paths = Enum.take(val_paths, 224)

batch_size = 32
target_height = 160
target_width = 160

train_pipeline =
  CatsAndDogs.pipeline_with_augmentations(
    train_paths,
    batch_size,
    target_height,
    target_width
  )

val_pipeline =
  CatsAndDogs.pipeline(
    val_paths,
    batch_size,
    target_height,
    target_width
  )

test_pipeline =
  CatsAndDogs.pipeline(
    test_paths,
    batch_size,
    target_height,
    target_width
  )

# Enum.take(train_pipeline, 1)
1 Like
3

Thanks so much Josh! That worked like a charm. I will keep that in mind next time. I did not dare to take a guess what could be going on given it was the exact book example.