{"id":18631014,"url":"https://github.com/aimagelab/speaksee","last_synced_at":"2025-08-04T07:10:17.481Z","repository":{"id":57469760,"uuid":"132164301","full_name":"aimagelab/speaksee","owner":"aimagelab","description":"PyTorch library for Visual-Semantic tasks","archived":false,"fork":false,"pushed_at":"2022-11-16T04:17:04.000Z","size":71985,"stargazers_count":29,"open_issues_count":2,"forks_count":7,"subscribers_count":8,"default_branch":"master","last_synced_at":"2025-07-28T12:42:36.580Z","etag":null,"topics":["caption-generation","pytorch","visual-semantic"],"latest_commit_sha":null,"homepage":"http://aimagelab.ing.unimore.it","language":"Python","has_issues":true,"has_wiki":null,"has_pages":null,"mirror_url":null,"source_name":null,"license":"bsd-3-clause","status":null,"scm":"git","pull_requests_enabled":true,"icon_url":"https://github.com/aimagelab.png","metadata":{"files":{"readme":"README.md","changelog":null,"contributing":null,"funding":null,"license":"LICENSE","code_of_conduct":null,"threat_model":null,"audit":null,"citation":null,"codeowners":null,"security":null,"support":null}},"created_at":"2018-05-04T16:31:43.000Z","updated_at":"2025-06-27T20:14:05.000Z","dependencies_parsed_at":"2023-01-22T06:45:18.150Z","dependency_job_id":null,"html_url":"https://github.com/aimagelab/speaksee","commit_stats":null,"previous_names":[],"tags_count":0,"template":false,"template_full_name":null,"purl":"pkg:github/aimagelab/speaksee","repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/aimagelab%2Fspeaksee","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/aimagelab%2Fspeaksee/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/aimagelab%2Fspeaksee/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/aimagelab%2Fspeaksee/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/aimagelab","download_url":"https://codeload.github.com/aimagelab/speaksee/tar.gz/refs/heads/master","sbom_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/aimagelab%2Fspeaksee/sbom","host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":268660139,"owners_count":24286028,"icon_url":"https://github.com/github.png","version":null,"created_at":"2022-05-30T11:31:42.601Z","updated_at":"2022-07-04T15:15:14.044Z","status":"online","status_checked_at":"2025-08-04T02:00:09.867Z","response_time":79,"last_error":null,"robots_txt_status":"success","robots_txt_updated_at":"2025-07-24T06:49:26.215Z","robots_txt_url":"https://github.com/robots.txt","online":true,"can_crawl_api":true,"host_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub","repositories_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories","repository_names_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repository_names","owners_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners"}},"keywords":["caption-generation","pytorch","visual-semantic"],"created_at":"2024-11-07T05:05:35.365Z","updated_at":"2025-08-04T07:10:17.448Z","avatar_url":"https://github.com/aimagelab.png","language":"Python","funding_links":[],"categories":[],"sub_categories":[],"readme":"![Speaksee Logo](_static/logo.png)\n\nSpeaksee is a Python package that provides utilities for working with Visual-Semantic data, developed at AImageLab.\n\n## Installation\nTo have a working installation, make sure you have Python 3.5+. You can then install speaksee via pip: \n```\npip install speaksee\n```\n\n### From source\nYou can also install speaksee from source with:\n\n```\ngit clone https://github.com/aimagelab/speaksee\ncd speaksee\npip install -e .\n```\n\nand obtain fresh upgrades without reinstalling it, simply running:\n\n```\ngit pull\n```\n\n## Example(s)\n\n### Pre-processing visual data\n``` python\nfrom speaksee.data import ImageField, TextField\nfrom speaksee.data.pipeline import EncodeCNN\nfrom speaksee.data.dataset import COCO\nfrom torchvision.models import resnet101\nfrom torchvision.transforms import Compose, Normalize\nfrom torch import nn\nimport torch\nfrom tqdm import tqdm\n\ndevice = torch.device('cuda')\n\n# Preprocess with some fancy cnn and transformation\ncnn = resnet101(pretrained=True).to(device)\ncnn.avgpool.forward = lambda x : x.mean(-1).mean(-1)\ncnn.fc = nn.Sequential()\n\ntransforms = Compose([\n    ToTensor(),\n    Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])\n])\n\nprepro_pipeline = EncodeCNN(cnn, transforms)\nimage_field = ImageField(preprocessing=prepro_pipeline, precomp_path='/nas/houston/lorenzo/fc2k_coco.hdf5')\n```\n\n### Pre-processing textual data\n``` python\n# Pipeline for text\ntext_field = TextField(eos_token='\u003ceos\u003e', lower=True, tokenize='spacy', remove_punctuation=True)\n```\n\n### Calling a dataset\n``` python\n# Create the dataset\ndataset = COCO(image_field, text_field, '/tmp/coco/images/',\n               '/nas/houston/lorenzo/vse/data/coco/annotations',\n               '/nas/houston/lorenzo/vse/data/coco/annotations')\ntrain_dataset, val_dataset, test_dataset = dataset.splits\n#image_field.precomp(dataset)  # do this once, or to refresh cache (we might change this in the near future)\ntext_field.build_vocab(train_dataset, val_dataset, min_freq=5)\n```\n\n### Training a model\n``` python\nfrom speaksee.models import FC\nmodel = FC(len(text_field.vocab), 2048, 512, 512, dropout_prob_lm=0).to(device)\n\nfrom speaksee.data import DataLoader\ndataloader_train = DataLoader(train_dataset, batch_size=16, shuffle=True)\ndataloader_val = DataLoader(val_dataset, batch_size=16)\n\nfrom torch.optim import Adam\nfrom torch.optim.lr_scheduler import StepLR\nfrom torch.nn import NLLLoss\noptim = Adam(model.parameters(), lr=5e-4)\nscheduler = StepLR(optim, step_size=3, gamma=.8)\nloss_fn = NLLLoss(ignore_index=text_field.vocab.stoi['\u003cpad\u003e'])\n\nfor e in range(50):\n    # Training\n    model.train()\n    running_loss = .0\n    with tqdm(desc='Epoch %d - train' % e, unit='it', total=len(dataloader_train)) as pbar:\n        for it, (images, captions )in enumerate(dataloader_train):\n            images, captions = images.to(device), captions.to(device)\n            out = model(images, captions)\n            optim.zero_grad()\n            loss = loss_fn(out.view(-1, len(text_field.vocab)), captions.view(-1))\n            loss.backward()\n            optim.step()\n\n            running_loss += loss.item()\n            pbar.set_postfix(loss=running_loss / (it+1))\n            pbar.update()\n\n    if e % 3 == 0 and model.ss_prob \u003c .25:\n        model.ss_prob += .05\n\n    # Validation\n    model.eval()\n    running_loss = .0\n    with tqdm(desc='Epoch %d - val' % e, unit='it', total=len(dataloader_val)) as pbar:\n        for it, (images, captions )in enumerate(dataloader_val):\n            images, captions = images.to(device), captions.to(device)\n            out = model(images, captions)\n            loss = loss_fn(out.view(-1, len(text_field.vocab)), captions.view(-1))\n\n            running_loss += loss.item()\n            pbar.set_postfix(loss=running_loss / (it+1))\n            pbar.update()\n\n    # Serialize model\n    torch.save({\n        'epoch': e,\n        'val_loss': running_loss / len(iter(dataloader_val)),\n        'state_dict': model.state_dict(),\n        'optimizer': optim.state_dict(),\n    }, '/nas/houston/lorenzo/fc_epoch_%03d.pth' % e)\n```\n\n### Evaluating a model\n``` python\nfrom speaksee.evaluation import Cider\nfrom speaksee.evaluation import PTBTokenizer\ndict_dataset_val = val_dataset.image_dictionary({'image': image_field, 'text': RawField()})\ndict_dataloader_val = DataLoader(dict_dataset_val, batch_size=16)\ngen = {}\ngts = {}\nwith tqdm(desc='Validation', unit='it', total=len(dict_dataloader_val)) as pbar:\n    for it, (images, caps_gt) in enumerate(iter(dict_dataloader_val)):\n        images = images.to(device)\n        with torch.no_grad():\n            out = model.beam_search(images, 50, text_field.vocab.stoi['\u003ceos\u003e'], 2, out_size=1)\n        caps_gen = text_field.decode(out)\n        for i, (gts_i, gen_i) in enumerate(zip(caps_gt, caps_gen)):\n            gen['%d_%d' % (it, i)] = [gen_i, ]\n            gts['%d_%d' % (it, i)] = gts_i\n        pbar.update()\n\ngts = PTBTokenizer.tokenize(gts)\ngen = PTBTokenizer.tokenize(gen)\nval_cider, _ = Cider().compute_score(gts, gen)\nprint(\"CIDEr is %f\" % val_cider)\n```\n\n### Model zoo\n| Model        | CIDEr | Download   |\n|--------------|-------|------------|\n| FC-2k (beam) | 93.8 | [Download](http://aimagelab.ing.unimore.it/speaksee/model_zoo/fc_epoch_029.pth)        |\n| Bottomup Topdown with sentinel | 117.8 | [Download](http://aimagelab.ing.unimore.it/speaksee/model_zoo/bottomup_topdown_sentinel_relu_epoch_063.pth)      \n\n\n### The team\nSpeaksee is currently maintained by [Lorenzo Baraldi](http://www.lorenzobaraldi.com),\n [Marcella Cornia](http://imagelab.ing.unimore.it/imagelab/person.asp?idpersona=90) and [Matteo Stefanini](https://github.com/MatteoStefanini)\n","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Faimagelab%2Fspeaksee","html_url":"https://awesome.ecosyste.ms/projects/github.com%2Faimagelab%2Fspeaksee","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Faimagelab%2Fspeaksee/lists"}