tarekziade HF Staff
Publish MIC usage, slim fixture and reviewed Transformers expectations
fef0512 verified | # DPT: depth estimation with Transformers | |
| This page is the Transformers contract of the DPT 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 `Intel/dpt-large`. The MIC runner runs the same blocks against the slim fixture and the 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. | |
| ## Predict a depth map | |
| ```py runnable:depth | |
| import torch | |
| from transformers import AutoImageProcessor, AutoModelForDepthEstimation, set_seed | |
| import json # doc-builder: hide | |
| import os # doc-builder: hide | |
| import numpy as np # doc-builder: hide | |
| model_id = "Intel/dpt-large" | |
| 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 | |
| processor = AutoImageProcessor.from_pretrained(model_id) | |
| model = AutoModelForDepthEstimation.from_pretrained(model_id).to(device) | |
| if variant: # doc-builder: hide | |
| del model # doc-builder: hide | |
| model = AutoModelForDepthEstimation.from_pretrained(model_id, **variant).to(device) # doc-builder: hide | |
| model.eval() | |
| # two small synthetic images: a gradient and a checkerboard, so the batch is meaningful | |
| h, w = 64, 64 | |
| gradient = torch.linspace(0, 1, h).view(h, 1, 1).expand(h, w, 3) | |
| checker = (torch.arange(h).view(h, 1) // 8 + torch.arange(w).view(1, w) // 8) % 2 | |
| checker = checker.float().unsqueeze(-1).expand(h, w, 3) | |
| # uint8: the processor rescales by 1/255, which would flatten floats in [0, 1]. | |
| images = [(gradient * 255).byte().permute(2, 0, 1).numpy(), (checker * 255).byte().permute(2, 0, 1).numpy()] | |
| inputs = processor(images=images, return_tensors="pt") | |
| inputs = inputs.to(device, dtype=model.dtype) | |
| with torch.no_grad(): | |
| outputs = model(**inputs) | |
| predicted_depth = outputs.predicted_depth | |
| post_processed = processor.post_process_depth_estimation( | |
| outputs, target_sizes=[(h, w), (h, w)] | |
| ) | |
| print(predicted_depth.shape) | |
| ``` | |
| ## One training step | |
| ```py runnable:depth:2 | |
| # depth targets: smooth ramps matching the synthetic images | |
| labels = torch.stack([ | |
| torch.linspace(0.1, 1.0, h).view(h, 1).expand(h, w), | |
| torch.linspace(1.0, 0.1, h).view(h, 1).expand(h, w), | |
| ]).to(device=device, dtype=model.dtype) | |
| model.train() | |
| optimizer = torch.optim.SGD([p for p in model.parameters() if p.requires_grad], lr=0.01) | |
| # DPTForDepthEstimation implements no training loss: regress the predicted depth | |
| # (at the pixel_values resolution) onto the resized targets with L1. | |
| predicted = model(**inputs).predicted_depth | |
| target = torch.nn.functional.interpolate(labels[:, None], size=predicted.shape[-2:], mode="bilinear")[:, 0] | |
| loss = torch.nn.functional.l1_loss(predicted.float(), target.float()) | |
| 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 = ( | |
| "dpt.encoder.layer.0.attention.attention.query.weight", | |
| "dpt.encoder.layer.0.intermediate.dense.weight", | |
| "head.head.0.weight", | |
| ) | |
| result = { | |
| "runtime": { | |
| "device": model.device.type, | |
| "dtype": str(model.dtype).removeprefix("torch."), | |
| "attn_implementation": model.config._attn_implementation, | |
| }, | |
| "predicted_depth_probe": predicted_depth.flatten()[:4].float().tolist(), | |
| "post_processed_depth_probe": post_processed[0]["predicted_depth"].flatten()[:4].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)) | |
| ``` | |