CLIP text encoder: projected text embeddings with Transformers

This page is the Transformers contract of the CLIP text encoder MIC and also a usage example. doc-builder renders the runnable blocks below as documentation, and its pytest plugin runs them as written against peft-internal-testing/tiny-clip-text-2. 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.

Embed a batch of texts

import torch
from transformers import CLIPTextModelWithProjection, AutoTokenizer, set_seed
import json  # doc-builder: hide
import os  # doc-builder: hide

model_id = "peft-internal-testing/tiny-clip-text-2"
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 and dropout draw random numbers

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = CLIPTextModelWithProjection.from_pretrained(model_id).to(device)
if variant:  # doc-builder: hide
    del model  # doc-builder: hide
    model = CLIPTextModelWithProjection.from_pretrained(model_id, **variant).to(device)  # doc-builder: hide

# Two distinct texts so batch perturbation is meaningful; right padding is the CLIP default.
texts = ["a photo of a cat", "a photo of a red bus on a street"]
inputs = tokenizer(texts, return_tensors="pt", padding=True).to(device)

model.eval()
with torch.no_grad():
    outputs = model(**inputs)

text_embeds = outputs.text_embeds
print(text_embeds.shape)
print(text_embeds[0, :4])

One training step

# A tiny linear probe on the projected text embeddings gives a real supervised loss.
probe = torch.nn.Linear(model.config.projection_dim, 2).to(device, model.dtype)
labels = torch.tensor([0, 1], device=device)

model.train()
optimizer = torch.optim.SGD(list(model.parameters()) + list(probe.parameters()), lr=0.01)
embeds = model(**inputs).text_embeds
loss = torch.nn.functional.cross_entropy(probe(embeds), labels)
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 = ("text_model.encoder.layers.0.self_attn.q_proj.weight", "text_model.encoder.layers.0.mlp.fc1.weight", "text_projection.weight")
    result = {
        "runtime": {
            "device": model.device.type,
            "dtype": str(model.dtype).removeprefix("torch."),
            "attn_implementation": model.config._attn_implementation,
        },
        "text_embeds": text_embeds.float().tolist(),
        "probe_logits": probe(embeds).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))