20 lines
4.3 KiB
JSON
20 lines
4.3 KiB
JSON
{
|
|
"summary": "The code installs libraries, loads the CLIP model, and processes a dataset before performing zero-shot classification and calculating top-1/top-5 accuracy on ImageNet dataset.",
|
|
"details": [
|
|
{
|
|
"comment": "This code installs required libraries, loads OpenAI's CLIP model, and retrieves Imagenet classes and templates. It also imports the ImageNetV2Dataset from a specific repository.",
|
|
"location": "\"/media/root/Toshiba XG3/works/CLIP/docs/src/notebooks/Prompt_Engineering_for_ImageNet.py\":0-35",
|
|
"content": "#! pip install ftfy regex tqdm\n#! pip install git+https://github.com/openai/CLIP.git\nimport numpy as np\nimport torch\nimport clip\nfrom tqdm.notebook import tqdm\nfrom pkg_resources import packaging\nprint(\"Torch version:\", torch.__version__)\nclip.available_models()\nmodel, preprocess = clip.load(\"ViT-B/32\")\ninput_resolution = model.visual.input_resolution\ncontext_length = model.context_length\nvocab_size = model.vocab_size\nprint(\"Model parameters:\", f\"{np.sum([int(np.prod(p.shape)) for p in model.parameters()]):,}\")\nprint(\"Input resolution:\", input_resolution)\nprint(\"Context length:\", context_length)\nprint(\"Vocab size:\", vocab_size)\nimport json\nimagenet_data = json.loads(open(\"imagenet_data.json\",\"r\").read())\nimagenet_classes = imagenet_data['imagenet_classes']\nimagenet_templates = imagenet_data['imagenet_templates']\nprint(f\"{len(imagenet_classes)} classes, {len(imagenet_templates)} templates\")\n# execute:\n# ! pip install git+https://github.com/modestyachts/ImageNetV2_pytorch\nfrom imagenetv2_pytorch import ImageNetV2Dataset"
|
|
},
|
|
{
|
|
"comment": "Code snippet performs zero-shot classification for the ImageNet dataset. It generates embeddings for given class names using text templates, averages them and stores in zeroshot_weights. The accuracy function calculates accuracy based on the output and target values.",
|
|
"location": "\"/media/root/Toshiba XG3/works/CLIP/docs/src/notebooks/Prompt_Engineering_for_ImageNet.py\":37-58",
|
|
"content": "images = ImageNetV2Dataset(transform=preprocess)\nloader = torch.utils.data.DataLoader(images, batch_size=32, num_workers=2)\ndef zeroshot_classifier(classnames, templates):\n with torch.no_grad():\n zeroshot_weights = []\n for classname in tqdm(classnames):\n texts = [template.format(classname) for template in templates] #format with class\n texts = clip.tokenize(texts).cuda() #tokenize\n class_embeddings = model.encode_text(texts) #embed with text encoder\n class_embeddings /= class_embeddings.norm(dim=-1, keepdim=True)\n class_embedding = class_embeddings.mean(dim=0)\n class_embedding /= class_embedding.norm()\n zeroshot_weights.append(class_embedding)\n zeroshot_weights = torch.stack(zeroshot_weights, dim=1).cuda()\n return zeroshot_weights\nzeroshot_weights = zeroshot_classifier(imagenet_classes, imagenet_templates)\ndef accuracy(output, target, topk=(1,)):\n pred = output.topk(max(topk), 1, True, True)[1].t()"
|
|
},
|
|
{
|
|
"comment": "The code calculates the top-1 and top-5 accuracy of a model's predictions on ImageNet dataset. It computes the accuracy by comparing predicted probabilities with ground truth labels, averages them over all images in the batch, and prints the results.",
|
|
"location": "\"/media/root/Toshiba XG3/works/CLIP/docs/src/notebooks/Prompt_Engineering_for_ImageNet.py\":59-83",
|
|
"content": " correct = pred.eq(target.view(1, -1).expand_as(pred))\n return [float(correct[:k].reshape(-1).float().sum(0, keepdim=True).cpu().numpy()) for k in topk]\nwith torch.no_grad():\n top1, top5, n = 0., 0., 0.\n for i, (images, target) in enumerate(tqdm(loader)):\n images = images.cuda()\n target = target.cuda()\n # predict\n image_features = model.encode_image(images)\n image_features /= image_features.norm(dim=-1, keepdim=True)\n logits = 100. * image_features @ zeroshot_weights\n # measure accuracy\n acc1, acc5 = accuracy(logits, target, topk=(1, 5))\n top1 += acc1\n top5 += acc5\n n += images.size(0)\ntop1 = (top1 / n) * 100\ntop5 = (top5 / n) * 100 \nprint(f\"Top-1 accuracy: {top1:.2f}\")\nprint(f\"Top-5 accuracy: {top5:.2f}\")"
|
|
}
|
|
]
|
|
} |