File size: 11,250 Bytes
e19aac6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
# Copyright 2024 NVIDIA CORPORATION & AFFILIATES
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#      http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
# SPDX-License-Identifier: Apache-2.0

from PIL import Image
from io import BytesIO
import base64
import numpy as np
import os

import torch
from transformers import StoppingCriteria
from .constants import IMAGE_TOKEN_INDEX

import tempfile
from io import BytesIO


def get_frame_from_vcap(vidcap, num_frames=10, fps=None, frame_count=None):
    import cv2

    if fps == None or frame_count == None:
        # if one of fps or frame_count is None, still recompute
        fps = vidcap.get(cv2.CAP_PROP_FPS)
        frame_count = int(vidcap.get(cv2.CAP_PROP_FRAME_COUNT))
    if fps == 0 or frame_count == 0:
        print("Video file not found. return empty images.")
        return [
            Image.new("RGB", (720, 720)),
        ] * num_frames

    duration = frame_count / fps
    frame_interval = frame_count // num_frames
    if frame_interval == 0 and frame_count <= 1:
        print("frame_interval is equal to 0. return empty image.")
        return [
            Image.new("RGB", (720, 720)),
        ] * num_frames
    # print("duration:", duration, "frames:", frame_count, "intervals:", frame_interval)

    images = []
    count = 0
    success = True
    frame_indices = np.linspace(0, frame_count - 2, num_frames, dtype=int)

    while success:
        # print("frame_count:", frame_count, "count:", count, "num_frames:", num_frames, "frame_interval:", frame_interval)
        if frame_count >= num_frames:
            success, frame = vidcap.read()
            if count in frame_indices:
                img = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
                im_pil = Image.fromarray(img)
                images.append(im_pil)
                if len(images) >= num_frames:
                    return images
            count += 1
        else:
            # Left padding frames if the video is not long enough
            success, frame = vidcap.read()
            if success:
                img = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
                im_pil = Image.fromarray(img)
                images.append(im_pil)
                count += 1
            elif count >= 1:
                width, height = images[-1].size
                images = [Image.new("RGB", (width, height))] * \
                    (num_frames - len(images)) + images
                print("padding frames:", (num_frames - len(images)))
                return images
            else:
                break
    raise ValueError(
        "Did not find enough frames in the video. return empty image.")


def opencv_extract_frames(vpath_or_bytesio, frames=6, fps=None, frame_count=None):
    """
    Extract frames from a video using OpenCV.

    Args:
        vpath_or_bytesio (str or BytesIO): Path to the video file or BytesIO object containing the video.
        frames (int): Number of frames to extract from the video.

    Returns:
        list: List of PIL Images extracted from the video.

    Raises:
        NotImplementedError: If the type of `vpath_or_bytesio` is not supported.
    """
    import cv2

    if isinstance(vpath_or_bytesio, str):
        vidcap = cv2.VideoCapture(vpath_or_bytesio)
        return get_frame_from_vcap(vidcap, frames, fps=fps, frame_count=frame_count)
    elif isinstance(vpath_or_bytesio, (BytesIO,)):
        # assuming mp4
        with tempfile.NamedTemporaryFile(delete=True, suffix=".mp4") as temp_video:
            temp_video.write(vpath_or_bytesio.read())
            temp_video_name = temp_video.name
            vidcap = cv2.VideoCapture(temp_video_name)
            return get_frame_from_vcap(vidcap, frames, fps=fps, frame_count=frame_count)
    else:
        raise NotImplementedError(type(vpath_or_bytesio))


def load_image_from_base64(image):
    return Image.open(BytesIO(base64.b64decode(image)))


def expand2square(pil_img, background_color):
    """
    Expand the given PIL image to a square shape by adding padding.

    Parameters:
    - pil_img: The PIL image to be expanded.
    - background_color: The color of the padding to be added.

    Returns:
    - The expanded PIL image.

    If the image is already square, it is returned as is.
    If the image is wider than it is tall, padding is added to the top and bottom.
    If the image is taller than it is wide, padding is added to the left and right.
    """
    width, height = pil_img.size
    if pil_img.mode == 'L':
        background_color = background_color[0]
    if width == height:
        return pil_img
    elif width > height:
        result = Image.new(pil_img.mode, (width, width), background_color)
        result.paste(pil_img, (0, (width - height) // 2))
        return result
    else:
        result = Image.new(pil_img.mode, (height, height), background_color)
        result.paste(pil_img, ((height - width) // 2, 0))
        return result


def process_image(image_file, data_args, image_folder, pil_preprocess_fn=None):
    processor = data_args.image_processor
    if isinstance(image_file, str):
        if image_folder is not None:
            image = Image.open(os.path.join(
                image_folder, image_file)).convert("RGB")
        else:
            image = Image.open(image_file).convert("RGB")
    else:
        # image is stored in bytearray
        image = image_file.convert("RGB")

    info = None

    if pil_preprocess_fn is not None:
        image = pil_preprocess_fn(image)
        if isinstance(image, tuple):
            image, info = image

    if data_args.image_aspect_ratio == "resize":
        if hasattr(data_args.image_processor, "crop_size"):
            # CLIP vision tower
            crop_size = data_args.image_processor.crop_size
        else:
            # SIGLIP vision tower
            assert hasattr(data_args.image_processor, "size")
            crop_size = data_args.image_processor.size
        image = image.resize((crop_size["height"], crop_size["width"]))
    if data_args.image_aspect_ratio == "pad":

        def expand2square(pil_img, background_color):
            width, height = pil_img.size
            if width == height:
                return pil_img
            elif width > height:
                result = Image.new(
                    pil_img.mode, (width, width), background_color)
                result.paste(pil_img, (0, (width - height) // 2))
                return result
            else:
                result = Image.new(
                    pil_img.mode, (height, height), background_color)
                result.paste(pil_img, ((height - width) // 2, 0))
                return result

        image = expand2square(image, tuple(int(x * 255)
                              for x in processor.image_mean))
        image = processor.preprocess(image, return_tensors="pt")[
            "pixel_values"][0]
    else:
        # Using default behavior of the vision encoder
        # For CLIP, default is central crop
        # For Radio, default is central crop
        # For Siglip, default is resize
        # For InternVIT, default is resize
        image = processor.preprocess(image, return_tensors="pt")[
            "pixel_values"][0]
    if info is not None:
        return image, info
    return image


def process_images(images, image_processor, model_cfg):

    model_cfg.image_processor = image_processor
    new_images = [process_image(image, model_cfg, None) for image in images]

    if all(x.shape == new_images[0].shape for x in new_images):
        new_images = torch.stack(new_images, dim=0)
    return new_images


# Note that newer VILA codebase adds an lstrip option that defaults to False, and the functionality is the same by default
def tokenizer_image_token(
    prompt, tokenizer, image_token_index=IMAGE_TOKEN_INDEX, return_tensors=None
):
    prompt_chunks = [
        tokenizer(chunk).input_ids for chunk in prompt.split("<image>")]

    def insert_separator(X, sep):
        return [ele for sublist in zip(X, [sep] * len(X)) for ele in sublist][:-1]

    input_ids = []
    offset = 0
    if (
        len(prompt_chunks) > 0
        and len(prompt_chunks[0]) > 0
        and prompt_chunks[0][0] == tokenizer.bos_token_id
    ):
        offset = 1
        input_ids.append(prompt_chunks[0][0])

    for x in insert_separator(prompt_chunks, [image_token_index] * (offset + 1)):
        input_ids.extend(x[offset:])

    if return_tensors is not None:
        if return_tensors == "pt":
            return torch.tensor(input_ids, dtype=torch.long)
        raise ValueError(f"Unsupported tensor type: {return_tensors}")
    return input_ids


def is_gemma_tokenizer(tokenizer):
    return "gemma" in tokenizer.__class__.__name__.lower()


def get_model_name_from_path(model_path):
    model_path = model_path.strip("/")
    model_paths = model_path.split("/")
    if model_paths[-1].startswith("checkpoint-"):
        return model_paths[-2] + "_" + model_paths[-1]
    else:
        return model_paths[-1]


class KeywordsStoppingCriteria(StoppingCriteria):
    def __init__(self, keywords, tokenizer, input_ids):
        self.keywords = keywords
        self.keyword_ids = []
        self.max_keyword_len = 0
        for keyword in keywords:
            cur_keyword_ids = tokenizer(keyword).input_ids
            if (
                len(cur_keyword_ids) > 1
                and cur_keyword_ids[0] == tokenizer.bos_token_id
            ):
                cur_keyword_ids = cur_keyword_ids[1:]
            if len(cur_keyword_ids) > self.max_keyword_len:
                self.max_keyword_len = len(cur_keyword_ids)
            self.keyword_ids.append(torch.tensor(cur_keyword_ids))
        self.tokenizer = tokenizer
        self.start_len = input_ids.shape[1]

    def call_for_batch(
        self, output_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs
    ) -> bool:
        offset = min(output_ids.shape[1] -
                     self.start_len, self.max_keyword_len)
        self.keyword_ids = [
            keyword_id.to(output_ids.device) for keyword_id in self.keyword_ids
        ]
        for keyword_id in self.keyword_ids:
            if (output_ids[0, -keyword_id.shape[0]:] == keyword_id).all():
                return True
        outputs = self.tokenizer.batch_decode(
            output_ids[:, -offset:], skip_special_tokens=True
        )[0]
        for keyword in self.keywords:
            if keyword in outputs:
                return True
        return False

    def __call__(
        self, output_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs
    ) -> bool:
        outputs = []
        for i in range(output_ids.shape[0]):
            outputs.append(self.call_for_batch(
                output_ids[i].unsqueeze(0), scores))
        return all(outputs)