Intercept call. We want to train the conv2d but keep the radio model in eval mode.

Windows to the full feature map Args: windows: local window features (num_windows*B, window_size, window_size, C) window_size

Apply patch splitting upscale operation. Args: x: Input tensor of shape (B, N, C) where N = num_prefix_tokens + H*W u

Forward function for feature maps before classifying each pixel with ``self.cls_seg`` fc. Args: inputs (list[Tensor]): Li

Get filler tokens interpolated to the required spatial size. Args: H: Target height in patches W: Target width in pat

Uses e^`sim` as a weighted vote for the corresponding label in `labels`. Returns the label that received the most vote weight.

Calculate k-NN Top-1 classification accuracy. Args: * train_split_embeddings: training split embeddings. * train_split_labels: training s

Utility function to pad a tensor a return a validity mask.

Transform inputs for decoder. Args: inputs (list[Tensor]): List of multi-level img features. Returns: Tensor:

Computes the accuracy over the k top predictions for the specified values of k

Pad to 'same' shape outputs.

Pad to 'same' shape outputs.

Returns canonicalized `text` (lowercase and punctuation removed). From: https://github.com/google-research/big_vision/blob/53f18caf27a9419231bbf0

E-RADIO employs windowed attention, which may be sensitive to the choice of this parameter, especially in cases of uneven partitionin

Context Manager. Puts the patch generator into video mode, with the specified number of temporal frames. In video mode, the

Evaluation routine.

Forward pass through decoder only. Args: global_tokens: Global tokens from encoder (B, max_tokens, C) global_token_ma

x_list contains a list of tensors to nest together and run

Build zero-shot classifier weights. Args: model: CLIP model instance tokenizer: CLIP tokenizer instance device: Devi