If distributed is initialized print only on rank 0.

Load and sample frames as input into the model.

Saves frames if only raw video is available.

Load model from pretrained checkpoint.

If distributed is initialized, print the rank.

Args: clip (torch.tensor): Video clip to be cropped. Size is (T, C, H, W)

Load image frames from a directory, with an optional filter function.

Encode inputs using the Qwen3VL embedder. Args: inputs: Dict containing 'text', 'image', 'video', 'instruction' etc.

Compute similarity matrix between query and passage representations.

Gather tensors across distributed processes.

Compute Average Precision (AP)@K for multiple true labels. AP is the average of precision@i for i where the i-th item is relevant.

Generate embeddings using MMEBEmbeddingModel's encode_input method. Supports DDP distributed gathering.

Evaluates predictions against true labels for a list of test cases. Handles multi-label by ensuring `label` is a list of true positiv

Truncate filename and append MD5 hash to ensure path compatibility

Returns: { "qid1": {"docA": 2, "docB": 1}, "qid2": {"docC": 3}, ... }

Compute Normalized Discounted Cumulative Gain (NDCG@K) with optional graded relevance and exponential gain. Args: predic

Compute Precision@K for multiple true labels. Precision@K = (Number of relevant items in top K) / K

Print table in terminal, split into multiple lines if too wide

Compute Recall@K for multiple true labels. Recall@K = (Number of relevant items in top K) / (Total number of relevant items)

Save as TSV file

Helper function to get common elements for calculations.

Pool the last non-padded token.

Build mapping: cand_id(str) -> candidate_sample(dict)

Center cropping implementation from ADM. https://github.com/openai/guided-diffusion/blob/8fb3ad9197f16bbc40620447b2742e13458d2831/guided_diff

Center cropping implementation from ADM. https://github.com/openai/guided-diffusion/blob/8fb3ad9197f16bbc40620447b2742e13458d2831/guided_diff

Collect all evaluation results

Compute summary results

Download specified subset and split of a single dataset

Export NextQA dataset in a format loadable by datasets.load_dataset Args: output_dir: Save directory dataset_name: Datas

Export MSRVTT text-to-video retrieval dataset in a format loadable by datasets.load_dataset Args: output_dir: Output directory

Export video classification dataset in a format loadable by datasets.load_dataset Args: output_dir: Save directory datas

Export ViDoRe text-to-document image retrieval dataset for HuggingFace datasets.load_dataset Args: output_dir: Output directory

Compute F1-score@K. F1@K = 2 * (Precision@K * Recall@K) / (Precision@K + Recall@K)

Forward pass for contrastive learning / evaluation.

Use ffprobe to get total frame count.

Args: clip (torch.tensor): Video clip to be normalized. Size is (T, C, H, W) Returns: flipped clip (torch.tensor): Size is (T

Compute Hit@K (or Hit Rate@K). Returns 1.0 if at least one true label is found in the top K predictions, 0.0 otherwise.

Load and concatenate multiple subsets from a Hugging Face dataset (e.g. MVBench)