Building a FastAI H&M Text Classifier from Scratch (with Kaggle GPU)

Last updated: 26 Apr 2025 · 5 min read

1. Dataset overview

I used H&M scraped product data on Kaggle—9 677 product rows with rich metadata (ID, name, price, material, details text, and a mainCatCode label). Only materials has ~2 % missing values, everything else is squeaky‑clean. The full description is in the dataset card :contentReference[oaicite:0]{index=0}:contentReference[oaicite:1]{index=1}.

Why this dataset? Short, domain‑specific descriptions make it perfect for Lesson‑4 NLP: enough vocabulary to learn but small enough to train on Kaggle’s free GPU.

2. Notebook & environment setup

2.1 Install FastAI & helper libs

!pip install -Uq fastai fastdownload fasttransform scikit-learn

2.2 Load the CSV mounted by Kaggle

from pathlib import Path
import pandas as pd, numpy as np

DATA = Path('/kaggle/input/hm-fashion-product-data')
df   = pd.read_csv(DATA/'hm_catalog.csv')

df.head()

The median description length is 48 tokens (95 th pct ≈ 140 tokens)—well within default seq_len=72.

3. Training a fashion‑aware language model

from fastai.text.all import *

dls_lm = TextDataLoaders.from_df(
    df, text_col='details', is_lm=True,
    valid_pct=0.1, seq_len=72)

learn_lm = language_model_learner(
    dls_lm, AWD_LSTM,
    drop_mult=0.5,
    metrics=[accuracy, Perplexity()])

learn_lm.fine_tune(3, 2e-2)
learn_lm.save_encoder('hm_enc')

After three epochs perplexity fell below 20—good enough to transfer.

4. Fine‑tuning the 322‑class classifier

4.1 Taming rare labels

About 40 categories have < 3 samples. I filtered them out to keep training balanced :contentReference[oaicite:2]{index=2}:contentReference[oaicite:3]{index=3}:

vc = df['mainCatCode'].value_counts()
keep = vc[vc >= 3].index
dsrc = df[df.mainCatCode.isin(keep)].copy()

4.2 Guaranteed stratified split

from sklearn.model_selection import StratifiedShuffleSplit
sss = list(StratifiedShuffleSplit(
    n_splits=1,
    test_size=0.2,
    random_state=42
).split(df_filt,df_filt['mainCatCode']))
train_idx,valid_idx = sss[0]

4.3 Build DataLoaders that share the LM vocab

tok_vocab = dls_lm.vocab                 # token vocab from LM
cat_vocab = dsrc['mainCatCode'].unique().tolist()
mySplitter = IndexSplitter(valid_idx);

dls_clas = TextDataLoaders.from_df(
    dsrc.reset_index(drop=True),
    text_col='details',
    label_col='mainCatCode',
    text_vocab=tok_vocab,
    label_vocab=cat_vocab,
    seq_len=72,
    splitter=mySplitter)

4.4 Fine‑tune in three minutes

learn_clas = text_classifier_learner(
    dls_clas, AWD_LSTM,
    metrics=accuracy).to_fp16()

learn_clas.load_encoder('hm_enc')
learn_clas.fine_tune(3, 2e-3)

Epoch	Train loss	Valid loss	Accuracy
0 (frozen)	4.54	3.15	41.9 %
1	3.22	2.39	51.5 %
2	2.47	1.83	60.0 %
3	2.11	1.73	61.6 %

61 % top‑1 on 322 labels in < 10 min of wall‑clock time—pretty good for a first run!

5. Reading the confusion matrix

I plotted the top‑15 most common categories

Confusion matrix showing dark diagonal and light off‑diagonals — Figure 1 – normalized confusion matrix (darker = better)

The darkest mis‑fires? “ladies_tops_shortsleeve” ↔ “ladies_tops_longsleeve”—no surprise given similar wording.

6. What to try next

Gradual unfreezing & discriminative LR slices (+3‑5 pp).
Label‑smoothing cross‑entropy to soften over‑confidence.
Increase seq_len to 128 to capture full descriptions.
Swap AWD‑LSTM for a pre‑trained MiniLM via Transformers callback (+10‑15 pp).
Write a Gradio demo that autocompletes descriptions & predicts categories.

Key take‑aways

Always share the exact token vocab between language model and classifier to avoid shape mismatches.
Use a custom stratified split when some labels have ≤ 3 samples.
FastAI’s AWD_LSTM + fine_tune makes a solid baseline—you can iterate from there.

Have questions or improvements? Drop them in the comments or ping me @Far__Had.

It's Farhad!