How to use learn.predict with my model? (I get a type error)

I have built an MLP with a standard Learner (also tried a TabularLearner but did not help with this problem). It trains properly and I can save the resulting model. However any attempt to get inferences from it fails with:
TypeError: list indices must be integers or slices, not list

Full stack trace after the code. This is the full training script, with the inference line at the end. I have a separate inference script, but tried adding the predict line here in case I messed up the model setup for inference somehow:

import pandas as pd
from fastai.tabular.all import *
import torch.nn as nn
import torch.nn.functional as F

source = '/Users/pierznewton-john/repos/gauquelin-ai/app/data/prepared.csv'
df = pd.read_csv(source)

# Continuous independent vars come after col 39
continuous_cols = df.columns[39:].tolist()  

# get targets
target_cols = df.columns[:39].tolist() 

# Split the dataset into training and validation sets
splits = RandomSplitter(0.1, 42)(range_of(df))

# Define preprocessing steps
procs = [Categorify, FillMissing, Normalize]

df[continuous_cols] = df[continuous_cols].astype(float)
for col in target_cols:
    df[col] = df[col].astype(float)

# Create the TabularPandas object
to = TabularPandas(df, procs=procs, cont_names=continuous_cols, y_names=target_cols, splits=splits, y_block=MultiCategoryBlock())

# Create the DataLoaders
dls = to.dataloaders(bs=64)  # Adjust batch size as needed

class SimpleMLP(nn.Module):
    def __init__(self, input_size, layer_sizes=[100, 50], output_size=39):
        super(SimpleMLP, self).__init__()
        self.layers = nn.ModuleList()
        last_size = input_size
        for size in layer_sizes:
            self.layers.append(nn.Linear(last_size, size))
            last_size = size
        self.output_layer = nn.Linear(last_size, output_size)

    def forward(self, x_cat, x_cont):
        # x_cat is empty and can be ignored
        x = x_cont
        for layer in self.layers:
            x = F.relu(layer(x))
        x = self.output_layer(x)
        return x
def multi_label_accuracy(preds, targets, thresh=0.5):
    "Compute accuracy for multi-label classification"
    preds = preds.sigmoid()  # Apply sigmoid to get probabilities
    preds = (preds > thresh).float()  # Convert to binary predictions
    correct = (preds == targets).float()  # Compare predictions with targets
    return correct.mean()  # Return mean accuracy

input_size = len(dls.cont_names)  # Assuming only continuous features are used as input
output_size = 39 
model = SimpleMLP(input_size=265, layer_sizes=[100, 50], output_size=39)

learn = Learner(dls, model, loss_func=nn.BCEWithLogitsLoss(), metrics=[multi_label_accuracy])
val_loss, val_accuracy = learn.validate()
print(f"Validation loss: {val_loss}")
print(f"Validation accuracy: {val_accuracy}")
learn.model_dir = "trained"'my_model')

# This fails...
row, clas, probs = learn.predict(df.iloc[0])

Full error:

Traceback (most recent call last):
  File "/Users/pedroponting/repos/gauquelin-ai/app/training/", line 70, in <module>
    row, clas, probs = learn.predict(df.iloc[0])
  File "/Users/pedroponting/.pyenv/versions/3.9.8/lib/python3.9/site-packages/fastai/", line 320, in predict
    dl = self.dls.test_dl([item], rm_type_tfms=rm_type_tfms, num_workers=0)
  File "/Users/pedroponting/.pyenv/versions/3.9.8/lib/python3.9/site-packages/fastai/tabular/", line 54, in test_dl
    if process: to.process()
  File "/Users/pedroponting/.pyenv/versions/3.9.8/lib/python3.9/site-packages/fastai/tabular/", line 180, in process
    def process(self): self.procs(self)
  File "/Users/pedroponting/.pyenv/versions/3.9.8/lib/python3.9/site-packages/fastcore/", line 208, in __call__
    def __call__(self, o): return compose_tfms(o, tfms=self.fs, split_idx=self.split_idx)
  File "/Users/pedroponting/.pyenv/versions/3.9.8/lib/python3.9/site-packages/fastcore/", line 158, in compose_tfms
    x = f(x, **kwargs)
  File "/Users/pedroponting/.pyenv/versions/3.9.8/lib/python3.9/site-packages/fastcore/", line 81, in __call__
    def __call__(self, x, **kwargs): return self._call('encodes', x, **kwargs)
  File "/Users/pedroponting/.pyenv/versions/3.9.8/lib/python3.9/site-packages/fastcore/", line 107, in _call
  File "/Users/pedroponting/.pyenv/versions/3.9.8/lib/python3.9/site-packages/fastcore/", line 91, in _call
    return self._do_call(getattr(self, fn), x, **kwargs)
  File "/Users/pedroponting/.pyenv/versions/3.9.8/lib/python3.9/site-packages/fastcore/", line 97, in _do_call
    return retain_type(f(x, **kwargs), x, ret)
  File "/Users/pedroponting/.pyenv/versions/3.9.8/lib/python3.9/site-packages/fastcore/", line 120, in __call__
    return f(*args, **kwargs)
  File "/Users/pedroponting/.pyenv/versions/3.9.8/lib/python3.9/site-packages/fastai/tabular/", line 308, in encodes
    missing = pd.isnull(to.conts)
  File "/Users/pedroponting/.pyenv/versions/3.9.8/lib/python3.9/site-packages/fastai/tabular/", line 208, in f
    def f(o): return o[list(getattr(o,nm+'_names'))]
  File "/Users/pedroponting/.pyenv/versions/3.9.8/lib/python3.9/site-packages/fastcore/", line 88, in __getitem__
    def __getitem__(self, k): return self.items[list(k) if isinstance(k,CollBase) else k]
TypeError: list indices must be integers or slices, not list

I haven’t personally experienced this but I did come across this forum post which may help.

Could you share your dataset that goes with this code (or provide some fake-data that has a similar structure) so that folks can troubleshoot by training the model and trying to run inference?

Sorry for not responding sooner. I did see that post too, but the solution involves using get_preds instead of predict. Using that post I was able to get predictions from data in a data loader, which is definitely a good start. I would like to use predict though as it is more convenient for single items and gives a more readily useable output.