MIC CLIP-integration (files_sha256 a7bce4f31eb6)

README.md

@@ -0,0 +1,26 @@
+---
+tags:
+  - mic
+  - test-artifact
+base_model:
+  - openai/clip-vit-base-patch32
+---
+
+# MIC: CLIP zero-shot image classification
+
+Model integration contract (MIC) for one usage path, across the checkpoints that share it. This repository holds the contract code and the manifest; each framework keeps its own expected results and pins this repository by revision. The model repositories are unchanged. It is a test artifact, not a model: exclude it from model search.
+
+Usage path: CLIPProcessor + CLIPModel, image-caption matching, batched.
+
+| File | What it is |
+| --- | --- |
+| [`integration.json`](integration.json) | Manifest: checkpoints, which fixture covers each, frameworks, entry points, variants, unsupported targets, comparison rules |
+| [`integrations/transformers/usage.md`](integrations/transformers/usage.md) | Transformers contract, written as doc-builder runnable markdown: the same blocks are documentation and test |
+
+| Checkpoint | Slim fixture (every framework PR) |
+| --- | --- |
+| [`openai/clip-vit-base-patch32`](https://huggingface.co/openai/clip-vit-base-patch32) | [`hf-internal-testing/proto-clip-vit-base-patch32-slim`](https://huggingface.co/hf-internal-testing/proto-clip-vit-base-patch32-slim) |
+
+Checked outputs: `image_to_text`, `text_to_image`, `logits_per_image`, `text_kwargs_input_ids`, `image_processor_config_keys`, `training.loss`, `training.gradients`, `training.updated`.
+
+A fixture pass and a real-checkpoint pass have different scopes; either establishes agreement on fixed inputs and a recorded runtime, not correctness on arbitrary inputs.

integration.json

@@ -0,0 +1,90 @@
+{
+  "mic": "0.3",
+  "name": "CLIP zero-shot image classification",
+  "usage": "CLIPProcessor + CLIPModel, image-caption matching, batched",
+  "architecture": "CLIPModel",
+  "checkpoints": {
+    "openai/clip-vit-base-patch32": {
+      "fixture": "hf-internal-testing/proto-clip-vit-base-patch32-slim"
+    }
+  },
+  "fixtures": {
+    "hf-internal-testing/proto-clip-vit-base-patch32-slim": {
+      "built_from": "openai/clip-vit-base-patch32",
+      "covers": [
+        "openai/clip-vit-base-patch32"
+      ],
+      "build": {
+        "seed": 17
+      }
+    }
+  },
+  "integrations": {
+    "transformers": {
+      "entry_point": "integrations/transformers/usage.md",
+      "format": "runnable-markdown",
+      "requires": [
+        "transformers>=5",
+        "torch",
+        "Pillow"
+      ],
+      "targets": [
+        "fixtures",
+        "checkpoints"
+      ],
+      "variants": {
+        "default": {},
+        "eager": {
+          "attn_implementation": "eager"
+        },
+        "bfloat16": {
+          "dtype": "bfloat16"
+        }
+      },
+      "training": true,
+      "resources": {
+        "fixture": {
+          "memory_gb": 1,
+          "download_mb": 20
+        },
+        "openai/clip-vit-base-patch32": {
+          "memory_gb": 2,
+          "download_mb": 605
+        }
+      }
+    }
+  },
+  "outputs": {
+    "runtime": {
+      "compare": "exact"
+    },
+    "image_to_text": {
+      "compare": "exact"
+    },
+    "text_to_image": {
+      "compare": "exact"
+    },
+    "logits_per_image": {
+      "compare": "allclose",
+      "atol": 0.0001,
+      "rtol": 0.0001
+    },
+    "text_kwargs_input_ids": {
+      "compare": "exact"
+    },
+    "image_processor_config_keys": {
+      "compare": "exact"
+    },
+    "training.loss": {
+      "compare": "allclose",
+      "atol": 0.0001,
+      "rtol": 0.0001
+    },
+    "training.gradients": {
+      "compare": "all_true"
+    },
+    "training.updated": {
+      "compare": "true"
+    }
+  }
+}

integrations/transformers/usage.md

@@ -0,0 +1,83 @@
+# CLIP: zero-shot image classification with Transformers
+
+This page is the Transformers contract of the CLIP MIC and also a usage example. [doc-builder](https://github.com/huggingface/doc-builder/blob/main/docs/runnable-code-blocks.md) renders the `runnable` blocks below as documentation, and its pytest plugin runs them as written against `openai/clip-vit-base-patch32`. The MIC runner runs the same blocks against each slim fixture and each checkpoint in [`integration.json`](../../integration.json). Lines marked `# doc-builder: hide` only matter there: they swap in the pinned checkpoint or variant and print the result record.
+
+## Match images to captions
+
+```py runnable:clip
+import torch
+from PIL import Image
+from transformers import CLIPModel, CLIPProcessor, set_seed
+import json  # doc-builder: hide
+import os  # doc-builder: hide
+
+model_id = "openai/clip-vit-base-patch32"
+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 = CLIPProcessor.from_pretrained(model_id)
+model = CLIPModel.from_pretrained(model_id).to(device)
+if variant:  # doc-builder: hide
+    model = CLIPModel.from_pretrained(model_id, **variant).to(device)
+
+colors = {"red": (255, 0, 0), "green": (0, 255, 0), "blue": (0, 0, 255), "yellow": (255, 255, 0)}
+images = [Image.new("RGB", (256, 256), rgb) for rgb in colors.values()]
+captions = [f"a photo of a {name} square" for name in colors]
+inputs = processor(text=captions, images=images, return_tensors="pt", padding=True).to(device, model.dtype)  # floating-point inputs follow the model's dtype
+
+with torch.no_grad():
+    logits_per_image = model(**inputs).logits_per_image
+best = logits_per_image.softmax(dim=1).argmax(dim=1)
+for name, index in zip(colors, best.tolist()):
+    print(f"{name} image -> {captions[index]!r}")
+```
+
+## Pass options to one modality
+
+A processor call can route options to one of its inputs, here padding to the text only. Integrations also read the preprocessing configuration without building a processor.
+
+```py runnable:clip:3
+from transformers.models.auto.image_processing_auto import get_image_processor_config
+
+text_only = processor(text=captions, images=images, return_tensors="pt", text_kwargs={"padding": True})
+print(text_only["input_ids"].shape)
+
+image_config = get_image_processor_config(model_id)
+print(sorted(image_config))
+```
+
+## One training step
+
+```py runnable:clip:3
+model.train()
+optimizer = torch.optim.SGD(model.parameters(), lr=0.01)
+loss = model(**inputs, return_loss=True).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 = ("visual_projection.weight", "text_projection.weight")
+    result = {
+        "runtime": {
+            "device": model.device.type,
+            "dtype": str(model.dtype).removeprefix("torch."),
+            "attn_implementation": model.config._attn_implementation,
+        },
+        "image_to_text": best.tolist(),
+        "text_to_image": logits_per_image.argmax(dim=0).tolist(),
+        "logits_per_image": logits_per_image.float().tolist(),
+        "text_kwargs_input_ids": text_only["input_ids"].tolist(),
+        "image_processor_config_keys": sorted(image_config),
+        "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))
+```