tarekziade's picture
tarekziade HF Staff
Publish MIC usage, slim fixture and reviewed Transformers expectations
0ce0675 verified
|
Raw History Blame Contribute Delete
4.08 kB

PoolFormer: image classification with Transformers

This page is the Transformers contract of the PoolFormer MIC and also a usage example. doc-builder renders the runnable blocks below as documentation, and its pytest plugin runs them as written against sail/poolformer_s12. The MIC runner runs the same blocks against each slim fixture and each checkpoint in integration.json. Lines marked # doc-builder: hide only matter there: they swap in the pinned checkpoint or variant and print the result record.

Classify two images

import torch
from transformers import AutoImageProcessor, AutoModelForImageClassification, set_seed
import json  # doc-builder: hide
import os  # doc-builder: hide
import numpy as np  # doc-builder: hide

model_id = "sail/poolformer_s12"
model_id = os.environ.get("MIC_MODEL", model_id)  # doc-builder: hide
variant = json.loads(os.environ.get("MIC_MODEL_KWARGS", "{}"))  # doc-builder: hide
device = "cuda" if torch.cuda.is_available() else "cpu"
set_seed(0)  # seeds torch, NumPy, and Python: loading, dropout, and data augmentation draw random numbers

processor = AutoImageProcessor.from_pretrained(model_id)
model = AutoModelForImageClassification.from_pretrained(model_id).to(device)
if variant:  # doc-builder: hide
    del model  # doc-builder: hide
    model = AutoModelForImageClassification.from_pretrained(model_id, **variant).to(device)  # doc-builder: hide

def textured_array(height, width, colour, seed):
    # The base colour plus a gradient and seeded noise: every patch differs, so patch order, crops and resizing all change the pixels.
    rng = np.random.default_rng(seed)
    y, x = np.mgrid[0:height, 0:width]
    ramp = (48 * (x / width + y / height) - 48)[..., None]
    pixels = np.asarray(colour, dtype=np.float64) + ramp + rng.normal(0, 24, (height, width, 3))
    return np.clip(pixels, 0, 255).astype(np.uint8)

# Two textured synthetic images that also differ in colour: global pooling averages random noise away.
images = [textured_array(224, 224, (200, 30, 30), seed=0), textured_array(224, 224, (30, 30, 200), seed=1)]
inputs = processor(images=images, return_tensors="pt").to(device, dtype=model.dtype)

with torch.no_grad():
    logits = model(**inputs).logits
predicted = logits.argmax(dim=-1)
print([model.config.id2label[i] for i in predicted.tolist()])

One training step

# Train on labels the model does not predict: the overfit fixture predicts its own outputs,
# and the near-zero loss leaves bfloat16 weights unchanged.
labels = (predicted + 1) % model.config.num_labels

model.train()
optimizer = torch.optim.SGD(model.parameters(), lr=0.01)
loss = model(**inputs, labels=labels).loss
loss.backward()
before = {n: p.detach().cpu().clone() for n, p in model.named_parameters() if p.requires_grad}  # doc-builder: hide
optimizer.step()
print(f"loss: {loss.item():.4f}")

if os.environ.get("MIC_RUNNER"):  # doc-builder: hide
    named = dict(model.named_parameters())
    checked = (
        next(n for n in named if n.endswith("classifier.weight")),
        next(n for n in named if n.endswith("poolformer.encoder.block.0.0.output.conv1.weight")),
    )
    result = {
        "runtime": {
            "device": model.device.type,
            "dtype": str(model.dtype).removeprefix("torch."),
            "attn_implementation": model.config._attn_implementation,
        },
        "predicted_classes": predicted.tolist(),
        "logits_probe": logits[:, torch.linspace(0, logits.shape[-1] - 1, 16).long()].float().tolist(),
        "training": {
            "loss": loss.item(),
            "gradients": {n: bool(named[n].grad is not None and torch.count_nonzero(named[n].grad) > 0 and torch.isfinite(named[n].grad).all()) for n in checked},
            "updated": any(not torch.equal(p.detach().cpu(), before[n]) for n, p in model.named_parameters() if n in before),
        },
    }
    print(json.dumps(result))