tarekziade's picture
tarekziade HF Staff
Publish MIC usage, slim fixture and reviewed Transformers expectations
cd55ca8 verified
|
Raw History Blame Contribute Delete
4.74 kB

BLIP: image captioning with Transformers

This page is the Transformers contract of the BLIP MIC and also a usage example. doc-builder renders the runnable blocks below as documentation, and its pytest plugin runs them as written against Salesforce/blip-image-captioning-base. 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.

The images are synthetic, so the example runs offline: two distinct colour blocks. A real checkpoint captions them as noise; the slim fixture is trained to name them.

Generate captions for a batch

import numpy as np
import torch
from PIL import Image
from transformers import BlipForConditionalGeneration, BlipProcessor, set_seed
import json  # doc-builder: hide
import os  # doc-builder: hide

model_id = "Salesforce/blip-image-captioning-base"
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 = BlipProcessor.from_pretrained(model_id)
model = BlipForConditionalGeneration.from_pretrained(model_id).to(device)
if variant:  # doc-builder: hide
    model = BlipForConditionalGeneration.from_pretrained(model_id, **variant).to(device)

# Two distinct colour blocks: the fixture can tell them apart, and a real checkpoint sees different inputs.
images = [
    Image.fromarray(np.full((384, 384, 3), (220, 20, 60), dtype=np.uint8)),
    Image.fromarray(np.full((384, 384, 3), (30, 144, 255), dtype=np.uint8)),
]
inputs = processor(images=images, return_tensors="pt").to(device, model.dtype)

generated = model.generate(**inputs, max_new_tokens=8)
print(processor.batch_decode(generated, skip_special_tokens=True))

Conditional captioning

BLIP accepts an optional text prefix. The processor tokenizes it as the decoder prompt.

conditional_inputs = processor(images=images, text=["a picture of", "a picture of"], return_tensors="pt").to(device, model.dtype)
conditional = model.generate(**conditional_inputs, max_new_tokens=8)
print(processor.batch_decode(conditional, skip_special_tokens=True))

One training step

captions = ["a red block", "a blue block"]
targets = processor(text=captions, return_tensors="pt", padding=True).to(device)
labels = targets.input_ids.masked_fill(targets.attention_mask == 0, -100)

with torch.no_grad():  # doc-builder: hide
    # Teacher-forced logits for the reference captions.
    logits = model(pixel_values=inputs.pixel_values, input_ids=targets.input_ids).logits
    logits_probe = logits[:, -1, :8].float()

model.train()
optimizer = torch.optim.SGD(model.parameters(), lr=0.01)
loss = model(pixel_values=inputs.pixel_values, input_ids=targets.input_ids, labels=labels).loss
loss.backward()
before = {name: p.detach().clone() for name, p in model.named_parameters()}  # 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 = (
        "vision_model.encoder.layers.0.self_attn.qkv.weight",
        "text_decoder.bert.encoder.layer.0.crossattention.self.query.weight",
    )
    result = {
        "runtime": {
            "device": model.device.type,
            "dtype": str(model.dtype).removeprefix("torch."),
            "attn_implementation": model.config._attn_implementation,
            "attn_implementations": {n: getattr(model.config, n)._attn_implementation for n in sorted(model.config.sub_configs) if getattr(getattr(model.config, n, None), '_attn_implementation', None)},
        },
        "generated_tokens": generated.tolist(),
        "generated_text": processor.batch_decode(generated, skip_special_tokens=True),
        "conditional_tokens": conditional.tolist(),
        "conditional_text": processor.batch_decode(conditional, skip_special_tokens=True),
        "logits_probe": logits_probe.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, before[n]) for n, p in named.items()),
        },
    }
    print(json.dumps(result))