@huggingface/transformers 3.2.3 → 3.3.0

package/src/models.js

@@ -270,8 +270,11 @@ async function getSession(pretrained_model_name_or_path, fileName, options) {
     } else if (session_options.externalData !== undefined) {
         externalDataPromises = session_options.externalData.map(async (ext) => {
             // if the external data is a string, fetch the file and replace the string with its content
+            // @ts-expect-error TS2339
             if (typeof ext.data === "string") {
+                // @ts-expect-error TS2339
                 const ext_buffer = await getModelFile(pretrained_model_name_or_path, ext.data, true, options);
+                // @ts-expect-error TS2698
                 return { ...ext, data: ext_buffer };
             }
             return ext;
@@ -529,14 +532,23 @@ async function encoderForward(self, model_inputs) {
         encoderFeeds.inputs_embeds = await self.encode_text({ input_ids: model_inputs.input_ids });
     }
     if (session.inputNames.includes('token_type_ids') && !encoderFeeds.token_type_ids) {
+        if (!encoderFeeds.input_ids) {
+            throw new Error('Both `input_ids` and `token_type_ids` are missing in the model inputs.');
+        }
         // Assign default `token_type_ids` (all zeroes) to the `encoderFeeds` if the model expects it,
         // but they weren't created by the tokenizer.
-        encoderFeeds.token_type_ids = new Tensor(
-            'int64',
-            new BigInt64Array(encoderFeeds.input_ids.data.length),
-            encoderFeeds.input_ids.dims
-        )
+        encoderFeeds.token_type_ids = zeros_like(encoderFeeds.input_ids);
     }
+    if (session.inputNames.includes('pixel_mask') && !encoderFeeds.pixel_mask) {
+        if (!encoderFeeds.pixel_values) {
+            throw new Error('Both `pixel_values` and `pixel_mask` are missing in the model inputs.');
+        }
+        // Assign default `pixel_mask` (all ones) to the `encoderFeeds` if the model expects it,
+        // but they weren't created by the processor.
+        const dims = encoderFeeds.pixel_values.dims;
+        encoderFeeds.pixel_mask = ones([dims[0], dims[2], dims[3]]);
+    }
+    
     return await sessionRun(session, encoderFeeds);
 }
 
@@ -1519,6 +1531,7 @@ export class PreTrainedModel extends Callable {
                 if (this.config.model_type === 'musicgen') {
                     // Custom logic (TODO: move to Musicgen class)
                     decoder_input_ids = Array.from({
+                        // @ts-expect-error TS2339
                         length: batch_size * this.config.decoder.num_codebooks
                     }, () => [decoder_start_token_id]);
 
@@ -1848,11 +1861,13 @@ export class PreTrainedModel extends Callable {
     async encode_image({ pixel_values }) {
         // image_inputs === { pixel_values }
         const features = (await sessionRun(this.sessions['vision_encoder'], { pixel_values })).image_features;
+        // @ts-expect-error TS2339
         if (!this.config.num_image_tokens) {
             console.warn(
                 'The number of image tokens was not set in the model configuration. ' +
                 `Setting it to the number of features detected by the vision encoder (${features.dims[1]}).`
             )
+            // @ts-expect-error TS2339
             this.config.num_image_tokens = features.dims[1];
         }
         return features;
@@ -3280,6 +3295,7 @@ export class WhisperForConditionalGeneration extends WhisperPreTrainedModel {
 
         if (generation_config.return_token_timestamps) {
             outputs["token_timestamps"] = this._extract_token_timestamps(
+                // @ts-expect-error TS2345
                 outputs,
                 generation_config.alignment_heads,
                 generation_config.num_frames,
@@ -3315,6 +3331,7 @@ export class WhisperForConditionalGeneration extends WhisperPreTrainedModel {
             );
         }
 
+        // @ts-expect-error TS2339
         let median_filter_width = this.config.median_filter_width;
         if (median_filter_width === undefined) {
             console.warn("Model config has no `median_filter_width`, using default value of 7.")
@@ -3325,6 +3342,7 @@ export class WhisperForConditionalGeneration extends WhisperPreTrainedModel {
         const batch = generate_outputs.cross_attentions;
         // Create a list with `decoder_layers` elements, each a tensor of shape
         // (batch size, attention_heads, output length, input length).
+        // @ts-expect-error TS2339
         const cross_attentions = Array.from({ length: this.config.decoder_layers },
             // Concatenate the cross attentions for each layer across sequence length dimension.
             (_, i) => cat(batch.map(x => x[i]), 2)
@@ -3468,6 +3486,7 @@ export class LlavaForConditionalGeneration extends LlavaPreTrainedModel {
         attention_mask,
     }) {
 
+        // @ts-expect-error TS2339
         const image_token_index = this.config.image_token_index;
 
         const idsList = input_ids.tolist();
@@ -4453,6 +4472,7 @@ export class Qwen2VLForConditionalGeneration extends Qwen2VLPreTrainedModel {
                 const image_nums = vision_tokens.filter(x => x == image_token_id).length;
                 const video_nums = vision_tokens.filter(x => x == video_token_id).length;
 
+                /** @type {number[][]} */
                 let llm_pos_ids_list = [];
                 let st = 0;
                 let remain_images = image_nums;
@@ -4522,6 +4542,7 @@ export class Qwen2VLForConditionalGeneration extends Qwen2VLPreTrainedModel {
                 // NOTE: Each item in llm_pos_ids_list is an array of shape (3, text_len),
                 // meaning to perform concatenation along dim=1, we can do the following:
                 const num_items = llm_pos_ids_list.reduce((acc, x) => acc + x.length, 0);
+                /** @type {number[]} */
                 const llm_positions = new Array(num_items);
                 let index = 0;
                 for (let x = 0; x < 3; ++x) {
@@ -4562,9 +4583,10 @@ export class Qwen2VLForConditionalGeneration extends Qwen2VLPreTrainedModel {
                     { length: 3 * data.length },
                     (_, i) => data[i % data.length]
                 );
+                /** @type {bigint[]} */
                 const mrope_position_deltas = Array.from(
                     { length: dims[0] },
-                    (_, i) => max(data.subarray(dims[1] * i, dims[1] * (i + 1)))[0] + 1 + dims[1]
+                    (_, i) => max(data.subarray(dims[1] * i, dims[1] * (i + 1)))[0] + 1n + BigInt(dims[1])
                 );
 
                 return [
@@ -5135,7 +5157,7 @@ export class DPTModel extends DPTPreTrainedModel { }
  * 
  * **Example:** Depth estimation w/ `Xenova/dpt-hybrid-midas`.
  * ```javascript
- * import { DPTForDepthEstimation, AutoProcessor, RawImage, interpolate, max } from '@huggingface/transformers';
+ * import { DPTForDepthEstimation, AutoProcessor, RawImage, interpolate_4d } from '@huggingface/transformers';
  * 
  * // Load model and processor
  * const model_id = 'Xenova/dpt-hybrid-midas';
@@ -5144,7 +5166,7 @@ export class DPTModel extends DPTPreTrainedModel { }
  * 
  * // Load image from URL
  * const url = 'http://images.cocodataset.org/val2017/000000039769.jpg';
- * const image = await RawImage.fromURL(url);
+ * const image = await RawImage.read(url);
  * 
  * // Prepare image for the model
  * const inputs = await processor(image);
@@ -5153,10 +5175,15 @@ export class DPTModel extends DPTPreTrainedModel { }
  * const { predicted_depth } = await model(inputs);
  * 
  * // Interpolate to original size
- * const prediction = interpolate(predicted_depth, image.size.reverse(), 'bilinear', false);
+ * const prediction = (await interpolate_4d(predicted_depth.unsqueeze(1), {
+     * size: image.size.reverse(),
+     * mode: 'bilinear',
+ * })).squeeze(1);
  * 
  * // Visualize the prediction
- * const formatted = prediction.mul_(255 / max(prediction.data)[0]).to('uint8');
+ * const min = prediction.min().item();
+ * const max = prediction.max().item();
+ * const formatted = prediction.sub_(min).div_(max - min).mul_(255).to('uint8');
  * const depth = RawImage.fromTensor(formatted);
  * // RawImage {
  * //   data: Uint8Array(307200) [ 85, 85, 84, ... ],
@@ -5206,11 +5233,7 @@ export class GLPNPreTrainedModel extends PreTrainedModel { }
 export class GLPNModel extends GLPNPreTrainedModel { }
 
 /**
- * GLPN Model transformer with a lightweight depth estimation head on top e.g. for KITTI, NYUv2.
- * 
- * **Example:** Depth estimation w/ `Xenova/glpn-kitti`.
- * ```javascript
- * import { GLPNForDepthEstimation, AutoProcessor, RawImage, interpolate, max } from '@huggingface/transformers';
+ * import { GLPNForDepthEstimation, AutoProcessor, RawImage, interpolate_4d } from '@huggingface/transformers';
  * 
  * // Load model and processor
  * const model_id = 'Xenova/glpn-kitti';
@@ -5219,7 +5242,7 @@ export class GLPNModel extends GLPNPreTrainedModel { }
  * 
  * // Load image from URL
  * const url = 'http://images.cocodataset.org/val2017/000000039769.jpg';
- * const image = await RawImage.fromURL(url);
+ * const image = await RawImage.read(url);
  * 
  * // Prepare image for the model
  * const inputs = await processor(image);
@@ -5228,13 +5251,18 @@ export class GLPNModel extends GLPNPreTrainedModel { }
  * const { predicted_depth } = await model(inputs);
  * 
  * // Interpolate to original size
- * const prediction = interpolate(predicted_depth, image.size.reverse(), 'bilinear', false);
+ * const prediction = (await interpolate_4d(predicted_depth.unsqueeze(1), {
+     * size: image.size.reverse(),
+     * mode: 'bilinear',
+ * })).squeeze(1);
  * 
  * // Visualize the prediction
- * const formatted = prediction.mul_(255 / max(prediction.data)[0]).to('uint8');
+ * const min = prediction.min().item();
+ * const max = prediction.max().item();
+ * const formatted = prediction.sub_(min).div_(max - min).mul_(255).to('uint8');
  * const depth = RawImage.fromTensor(formatted);
  * // RawImage {
- * //   data: Uint8Array(307200) [ 207, 169, 154, ... ],
+ * //   data: Uint8Array(307200) [ 85, 85, 84, ... ],
  * //   width: 640,
  * //   height: 480,
  * //   channels: 1
@@ -5409,6 +5437,8 @@ export class Dinov2WithRegistersForImageClassification extends Dinov2WithRegiste
     }
 }
 //////////////////////////////////////////////////
+export class GroundingDinoPreTrainedModel extends PreTrainedModel { }
+export class GroundingDinoForObjectDetection extends GroundingDinoPreTrainedModel { }
 
 //////////////////////////////////////////////////
 export class YolosPreTrainedModel extends PreTrainedModel { }
@@ -6107,6 +6137,9 @@ export class WavLMForAudioFrameClassification extends WavLMPreTrainedModel {
     }
 }
 
+export class StyleTextToSpeech2PreTrainedModel extends PreTrainedModel { }
+export class StyleTextToSpeech2Model extends StyleTextToSpeech2PreTrainedModel { }
+
 //////////////////////////////////////////////////
 // SpeechT5 models
 /**
@@ -6201,10 +6234,12 @@ export class SpeechT5ForTextToSpeech extends SpeechT5PreTrainedModel {
 
         const { encoder_outputs, encoder_attention_mask } = await encoderForward(this, model_inputs);
 
+        // @ts-expect-error TS2339
         const r = encoder_outputs.dims[1] / this.config.reduction_factor;
         const maxlen = Math.floor(r * maxlenratio);
         const minlen = Math.floor(r * minlenratio);
 
+        // @ts-expect-error TS2339
         const num_mel_bins = this.config.num_mel_bins;
 
         let spectrogramParts = [];
@@ -6569,11 +6604,13 @@ export class MusicgenForConditionalGeneration extends PreTrainedModel { // NOTE:
      */
     _apply_and_filter_by_delay_pattern_mask(outputs) {
         const [bs_x_codebooks, seqLength] = outputs.dims;
+        // @ts-expect-error TS2339
         const num_codebooks = this.config.decoder.num_codebooks;
         const upperBound = (seqLength - num_codebooks);
 
         let newDataSize = 0;
         for (let i = 0; i < outputs.size; ++i) {
+            // @ts-expect-error TS2339
             if (outputs.data[i] === this.config.decoder.pad_token_id) {
                 continue;
             }
@@ -6603,7 +6640,9 @@ export class MusicgenForConditionalGeneration extends PreTrainedModel { // NOTE:
         let clonedInputIds = structuredClone(input_ids);
         for (let i = 0; i < clonedInputIds.length; ++i) {
             for (let j = 0; j < clonedInputIds[i].length; ++j) {
+                // @ts-expect-error TS2339
                 if ((i % this.config.decoder.num_codebooks) >= j) {
+                    // @ts-expect-error TS2339
                     clonedInputIds[i][j] = BigInt(this.config.decoder.pad_token_id);
                 }
             }
@@ -6760,6 +6799,9 @@ export class MultiModalityCausalLM extends MultiModalityPreTrainedModel {
         'past_key_values',
     ];
 
+    /**
+     * @param {ConstructorParameters<typeof MultiModalityPreTrainedModel>} args
+     */
     constructor(...args) {
         super(...args);
 
@@ -7061,6 +7103,8 @@ const MODEL_MAPPING_NAMES_ENCODER_ONLY = new Map([
 
     ['maskformer', ['MaskFormerModel', MaskFormerModel]],
     ['mgp-str', ['MgpstrForSceneTextRecognition', MgpstrForSceneTextRecognition]],
+
+    ['style_text_to_speech_2', ['StyleTextToSpeech2Model', StyleTextToSpeech2Model]],
 ]);
 
 const MODEL_MAPPING_NAMES_ENCODER_DECODER = new Map([
@@ -7305,6 +7349,7 @@ const MODEL_FOR_OBJECT_DETECTION_MAPPING_NAMES = new Map([
 const MODEL_FOR_ZERO_SHOT_OBJECT_DETECTION_MAPPING_NAMES = new Map([
     ['owlvit', ['OwlViTForObjectDetection', OwlViTForObjectDetection]],
     ['owlv2', ['Owlv2ForObjectDetection', Owlv2ForObjectDetection]],
+    ['grounding-dino', ['GroundingDinoForObjectDetection', GroundingDinoForObjectDetection]],
 ]);
 
 const MODEL_FOR_IMAGE_SEGMENTATION_MAPPING_NAMES = new Map([
@@ -7728,10 +7773,17 @@ export class SequenceClassifierOutput extends ModelOutput {
     /**
      * @param {Object} output The output of the model.
      * @param {Tensor} output.logits classification (or regression if config.num_labels==1) scores (before SoftMax).
+     * @param {Record<string, Tensor>} [output.attentions] Object of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length, sequence_length)`.
+     * Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
      */
-    constructor({ logits }) {
+    constructor({ logits, ...attentions }) {
         super();
         this.logits = logits;
+        const attentions_list = Object.values(attentions);
+        if (attentions_list.length > 0) {
+            // Only set attentions if they are not empty
+            this.attentions = attentions_list;
+        }
     }
 }
 

package/src/ops/registry.js

@@ -36,6 +36,16 @@ export class TensorOpRegistry {
         // executionProviders: ['webgpu'],
     };
 
+    static get nearest_interpolate_4d() {
+        if (!this._nearest_interpolate_4d) {
+            this._nearest_interpolate_4d = wrap(
+                [8, 10, 18, 0, 58, 129, 1, 10, 41, 10, 1, 120, 10, 0, 10, 0, 10, 1, 115, 18, 1, 121, 34, 6, 82, 101, 115, 105, 122, 101, 42, 18, 10, 4, 109, 111, 100, 101, 34, 7, 110, 101, 97, 114, 101, 115, 116, 160, 1, 3, 18, 1, 114, 90, 31, 10, 1, 120, 18, 26, 10, 24, 8, 1, 18, 20, 10, 3, 18, 1, 98, 10, 3, 18, 1, 99, 10, 3, 18, 1, 104, 10, 3, 18, 1, 119, 90, 15, 10, 1, 115, 18, 10, 10, 8, 8, 7, 18, 4, 10, 2, 8, 4, 98, 31, 10, 1, 121, 18, 26, 10, 24, 8, 1, 18, 20, 10, 3, 18, 1, 98, 10, 3, 18, 1, 99, 10, 3, 18, 1, 104, 10, 3, 18, 1, 119, 66, 2, 16, 21],
+                this.session_options,
+                'y',
+            );
+        }
+        return this._nearest_interpolate_4d;
+    }
     static get bilinear_interpolate_4d() {
         if (!this._bilinear_interpolate_4d) {
             this._bilinear_interpolate_4d = wrap(

package/src/pipelines.js

@@ -64,12 +64,13 @@ import {
     round,
 } from './utils/maths.js';
 import {
-    read_audio
+    read_audio,
+    RawAudio
 } from './utils/audio.js';
 import {
     Tensor,
     mean_pooling,
-    interpolate,
+    interpolate_4d,
     quantize_embeddings,
     topk,
 } from './utils/tensor.js';
@@ -294,6 +295,7 @@ export class TextClassificationPipeline extends (/** @type {new (options: TextPi
 
         // TODO: Use softmax tensor function
         const function_to_apply =
+            // @ts-expect-error TS2339
             this.model.config.problem_type === 'multi_label_classification'
                 ? batch => batch.sigmoid()
                 : batch => new Tensor(
@@ -302,6 +304,7 @@ export class TextClassificationPipeline extends (/** @type {new (options: TextPi
                     batch.dims,
                 ); // single_label_classification (default)
 
+        // @ts-expect-error TS2339
         const id2label = this.model.config.id2label;
 
         const toReturn = [];
@@ -404,6 +407,7 @@ export class TokenClassificationPipeline extends (/** @type {new (options: TextP
         const outputs = await this.model(model_inputs)
 
         const logits = outputs.logits;
+        // @ts-expect-error TS2339
         const id2label = this.model.config.id2label;
 
         const toReturn = [];
@@ -743,11 +747,14 @@ export class Text2TextGenerationPipeline extends (/** @type {new (options: TextP
 
 
         // Add global prefix, if present
+        // @ts-expect-error TS2339
         if (this.model.config.prefix) {
+            // @ts-expect-error TS2339
             texts = texts.map(x => this.model.config.prefix + x)
         }
 
         // Handle task specific params:
+        // @ts-expect-error TS2339
         const task_specific_params = this.model.config.task_specific_params
         if (task_specific_params && task_specific_params[this.task]) {
             // Add prefixes, if present
@@ -1486,6 +1493,7 @@ export class AudioClassificationPipeline extends (/** @type {new (options: Audio
         const sampling_rate = this.processor.feature_extractor.config.sampling_rate;
         const preparedAudios = await prepareAudios(audio, sampling_rate);
 
+        // @ts-expect-error TS2339
         const id2label = this.model.config.id2label;
 
         const toReturn = [];
@@ -1796,6 +1804,7 @@ export class AutomaticSpeechRecognitionPipeline extends (/** @type {new (options
             audio = [/** @type {AudioInput} */ (audio)];
         }
 
+        // @ts-expect-error TS2339
         const time_precision = this.processor.feature_extractor.config.chunk_length / this.model.config.max_source_positions;
         const hop_length = this.processor.feature_extractor.config.hop_length;
 
@@ -1861,7 +1870,9 @@ export class AutomaticSpeechRecognitionPipeline extends (/** @type {new (options
 
                 // TODO: Right now we only get top beam
                 if (return_timestamps === 'word') {
+                    // @ts-expect-error TS2339
                     chunk.tokens = data.sequences.tolist()[0];
+                    // @ts-expect-error TS2339
                     chunk.token_timestamps = data.token_timestamps.tolist()[0].map(
                         (/** @type {number} */ x) => round(x, 2)
                     );
@@ -1906,7 +1917,7 @@ export class AutomaticSpeechRecognitionPipeline extends (/** @type {new (options
             const max_new_tokens = Math.floor(aud.length / sampling_rate) * 6;
             const outputs = await this.model.generate({ max_new_tokens, ...kwargs, ...inputs });
 
-            const text = this.processor.batch_decode(outputs, { skip_special_tokens: true })[0];
+            const text = this.processor.batch_decode(/** @type {Tensor} */(outputs), { skip_special_tokens: true })[0];
             toReturn.push({ text });
         }
         return single ? toReturn[0] : toReturn;
@@ -2055,6 +2066,7 @@ export class ImageClassificationPipeline extends (/** @type {new (options: Image
         const { pixel_values } = await this.processor(preparedImages);
         const output = await this.model({ pixel_values });
 
+        // @ts-expect-error TS2339
         const id2label = this.model.config.id2label;
 
         /** @type {ImageClassificationOutput[]} */
@@ -2169,6 +2181,7 @@ export class ImageSegmentationPipeline extends (/** @type {new (options: ImagePi
             }
         }
 
+        // @ts-expect-error TS2339
         const id2label = this.model.config.id2label;
 
         /** @type {ImageSegmentationPipelineOutput[]} */
@@ -2395,6 +2408,7 @@ export class ObjectDetectionPipeline extends (/** @type {new (options: ImagePipe
         const processed = this.processor.image_processor.post_process_object_detection(output, threshold, imageSizes);
 
         // Add labels
+        // @ts-expect-error TS2339
         const id2label = this.model.config.id2label;
 
         // Format output
@@ -2539,13 +2553,35 @@ export class ZeroShotObjectDetectionPipeline extends (/** @type {new (options: T
             // Run model with both text and pixel inputs
             const output = await this.model({ ...text_inputs, pixel_values });
 
-            // @ts-ignore
-            const processed = this.processor.image_processor.post_process_object_detection(output, threshold, imageSize, true)[0];
-            let result = processed.boxes.map((box, i) => ({
-                score: processed.scores[i],
-                label: candidate_labels[processed.classes[i]],
-                box: get_bounding_box(box, !percentage),
-            })).sort((a, b) => b.score - a.score);
+            let result;
+            if('post_process_grounded_object_detection' in this.processor) {
+                // @ts-ignore
+                const processed = this.processor.post_process_grounded_object_detection(
+                    output,
+                    text_inputs.input_ids,
+                    {
+                        // TODO: support separate threshold values
+                        box_threshold: threshold,
+                        text_threshold: threshold,
+                        target_sizes: imageSize,
+                    },
+                )[0];
+                result = processed.boxes.map((box, i) => ({
+                    score: processed.scores[i],
+                    label: processed.labels[i],
+                    box: get_bounding_box(box, !percentage),
+                }))
+            } else {
+                // @ts-ignore
+                const processed = this.processor.image_processor.post_process_object_detection(output, threshold, imageSize, true)[0];
+                result = processed.boxes.map((box, i) => ({
+                    score: processed.scores[i],
+                    label: candidate_labels[processed.classes[i]],
+                    box: get_bounding_box(box, !percentage),
+                }))
+            }
+            result.sort((a, b) => b.score - a.score);
+
             if (top_k !== null) {
                 result = result.slice(0, top_k);
             }
@@ -2614,6 +2650,7 @@ export class DocumentQuestionAnsweringPipeline extends (/** @type {new (options:
         // Run model
         const output = await this.model.generate({
             inputs: pixel_values,
+            // @ts-expect-error TS2339
             max_length: this.model.config.decoder.max_position_embeddings,
             decoder_input_ids,
             ...generate_kwargs,
@@ -2664,7 +2701,7 @@ export class DocumentQuestionAnsweringPipeline extends (/** @type {new (options:
  * const synthesizer = await pipeline('text-to-speech', 'Xenova/speecht5_tts', { quantized: false });
  * const speaker_embeddings = 'https://huggingface.co/datasets/Xenova/transformers.js-docs/resolve/main/speaker_embeddings.bin';
  * const out = await synthesizer('Hello, my dog is cute', { speaker_embeddings });
- * // {
+ * // RawAudio {
  * //   audio: Float32Array(26112) [-0.00005657337896991521, 0.00020583874720614403, ...],
  * //   sampling_rate: 16000
  * // }
@@ -2684,7 +2721,7 @@ export class DocumentQuestionAnsweringPipeline extends (/** @type {new (options:
  * ```javascript
  * const synthesizer = await pipeline('text-to-speech', 'Xenova/mms-tts-fra');
  * const out = await synthesizer('Bonjour');
- * // {
+ * // RawAudio {
  * //   audio: Float32Array(23808) [-0.00037693005288019776, 0.0003325853613205254, ...],
  * //   sampling_rate: 16000
  * // }
@@ -2729,11 +2766,12 @@ export class TextToAudioPipeline extends (/** @type {new (options: TextToAudioPi
         // Generate waveform
         const { waveform } = await this.model(inputs);
 
+        // @ts-expect-error TS2339
         const sampling_rate = this.model.config.sampling_rate;
-        return {
-            audio: waveform.data,
+        return new RawAudio(
+            waveform.data,
             sampling_rate,
-        }
+        )
     }
 
     async _call_text_to_spectrogram(text_inputs, { speaker_embeddings }) {
@@ -2773,10 +2811,10 @@ export class TextToAudioPipeline extends (/** @type {new (options: TextToAudioPi
         const { waveform } = await this.model.generate_speech(input_ids, speaker_embeddings, { vocoder: this.vocoder });
 
         const sampling_rate = this.processor.feature_extractor.config.sampling_rate;
-        return {
-            audio: waveform.data,
+        return new RawAudio(
+            waveform.data,
             sampling_rate,
-        }
+        )
     }
 }
 
@@ -2886,11 +2924,23 @@ export class DepthEstimationPipeline extends (/** @type {new (options: ImagePipe
 
         const toReturn = [];
         for (let i = 0; i < preparedImages.length; ++i) {
-            const prediction = interpolate(predicted_depth[i], preparedImages[i].size.reverse(), 'bilinear', false);
-            const formatted = prediction.mul_(255 / max(prediction.data)[0]).to('uint8');
+            const batch = predicted_depth[i];
+            const [height, width] = batch.dims.slice(-2);
+            const [new_width, new_height] = preparedImages[i].size;
+
+            // Interpolate to original size
+            const prediction = (await interpolate_4d(batch.view(1, 1, height, width), {
+                size: [new_height, new_width],
+                mode: 'bilinear',
+            })).view(new_height, new_width);
+
+            const minval = /** @type {number} */(prediction.min().item());
+            const maxval = /** @type {number} */(prediction.max().item());
+            const formatted = prediction.sub(minval).div_(maxval - minval).mul_(255).to('uint8').unsqueeze(0);
+            const depth = RawImage.fromTensor(formatted);
             toReturn.push({
-                predicted_depth: predicted_depth[i],
-                depth: RawImage.fromTensor(formatted),
+                predicted_depth: prediction,
+                depth,
             });
         }
 
@@ -3368,4 +3418,4 @@ async function loadItems(mapping, model, pretrainedOptions) {
     }
 
     return result;
-}
+}

package/src/tokenizers.js

@@ -47,10 +47,8 @@ import {
 import { Template } from '@huggingface/jinja';
 
 import {
-    WHISPER_LANGUAGE_MAPPING,
-    whisper_language_to_code,
+    WHISPER_LANGUAGE_MAPPING
 } from './models/whisper/common_whisper.js';
-import { GITHUB_ISSUE_URL } from './utils/constants.js';
 
 /**
  * @typedef {Object} TokenizerProperties Additional tokenizer-specific properties.
@@ -535,7 +533,7 @@ class Unigram extends TokenizerModel {
      * Create a new Unigram tokenizer model.
      * @param {Object} config The configuration object for the Unigram model.
      * @param {number} config.unk_id The ID of the unknown token
-     * @param {any[][]} config.vocab A 2D array representing a mapping of tokens to scores.
+     * @param {[string, number][]} config.vocab A 2D array representing a mapping of tokens to scores.
      * @param {Object} moreConfig Additional configuration object for the Unigram model.
      */
     constructor(config, moreConfig) {
@@ -543,11 +541,10 @@ class Unigram extends TokenizerModel {
 
         const vocabSize = config.vocab.length;
         this.vocab = new Array(vocabSize);
+        /** @type {number[]} */
         this.scores = new Array(vocabSize);
         for (let i = 0; i < vocabSize; ++i) {
-            const piece = config.vocab[i];
-            this.vocab[i] = piece[0];
-            this.scores[i] = piece[1];
+            [this.vocab[i], this.scores[i]] = config.vocab[i];
         }
 
         this.unk_token_id = config.unk_id;