tarekziade HF Staff
Publish MIC usage, slim fixture and reviewed Transformers expectations
0ce0675 verified |
Download integrations/transformers/usage.md from huggingface-mic/poolformer_s12-mic: direct link, hf CLI and curl.
- Browser
- Download file 4.08 kB
-
https://huggingface.co/huggingface-mic/poolformer_s12-mic/resolve/main/integrations/transformers/usage.md
- Command line
-
hf download hf://huggingface-mic/poolformer_s12-mic/integrations/transformers/usage.md
-
curl -L -o usage.md https://huggingface.co/huggingface-mic/poolformer_s12-mic/resolve/main/integrations/transformers/usage.md
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))