-
Notifications
You must be signed in to change notification settings - Fork 71
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
* added README for LLM preprocess demo * Fix pip problem in gpu ci * Add aes module in LLM demo * param parser for the LLM preprocessing main * rename the LLM preprocessing demo path * video processing demo with ocr See merge request: !840
- Loading branch information
Showing
12 changed files
with
926 additions
and
2 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,18 @@ | ||
| It's a demo for LLM video/image generating training preprocessing. | ||
|
|
||
| Based on BMF, it's flexible to build and integrate algorithms into whole pipeline of preprocessing. | ||
|
|
||
| Two part of them are included: | ||
| 1. Clip processing | ||
| The input video will be split according to scene change, and subtitles in the video will be detected and cropped by OCR module, and the video quality will be assessed by BMF provided aesthetic module. | ||
| After that, the finalized video clips will be encoded as output. | ||
| ## Prerequisites | ||
| Please pip install all the dependencies in `requirement.txt` | ||
| ## Run | ||
| ``` | ||
| python main.py --input_file <your test video> | ||
| ``` | ||
| There will be output info and clips to be stored in `clip_output` of current path. | ||
|
|
||
| 2. Caption | ||
| Please reference the README in "bmf/bmf/demo/fast_caption_module" |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,153 @@ | ||
| #!/usr/bin/env python | ||
| # -*- coding: utf-8 -*- | ||
|
|
||
| from module_utils import SyncModule | ||
| import os | ||
| import time | ||
| import json | ||
| import pdb | ||
| import os.path as osp | ||
| import numpy as np | ||
|
|
||
| os.environ["OMP_NUM_THREADS"] = "8" | ||
| import onnxruntime as ort | ||
| import torch | ||
| import logging | ||
| import cv2 | ||
|
|
||
|
|
||
| def get_logger(): | ||
| return logging.getLogger("main") | ||
|
|
||
|
|
||
| LOGGER = get_logger() | ||
|
|
||
|
|
||
| def flex_resize_aesv2(img, desired_size=[448, 672], pad_color=[0, 0, 0]): | ||
| old_h, old_w = img.shape[:2] # old_size is in (height, width) format | ||
| if desired_size[0] >= desired_size[1]: | ||
| if old_h < old_w: # rotate the honrizontal video | ||
| img = np.rot90(img, k=1, axes=(1, 0)) | ||
| else: | ||
| if old_h > old_w: # rotate the vertical video | ||
| img = np.rot90(img, k=1, axes=(1, 0)) | ||
| old_h, old_w = img.shape[:2] | ||
|
|
||
| if old_w / old_h > (desired_size[1] / desired_size[0]): | ||
| ratio = desired_size[0] / old_h | ||
| else: | ||
| ratio = desired_size[1] / old_w | ||
| img = cv2.resize(img, None, fx=ratio, fy=ratio) | ||
| h, w, _ = img.shape | ||
| h_crop = (h - desired_size[0]) // 2 | ||
| w_crop = (w - desired_size[1]) // 2 | ||
| img = img[h_crop:h_crop + desired_size[0], | ||
| w_crop:w_crop + desired_size[1], :] | ||
| return img | ||
|
|
||
|
|
||
| class Aesmod: | ||
|
|
||
| def __init__(self, model_path, model_version, output_path): | ||
| self._frm_idx = 0 | ||
| self._frm_scores = [] | ||
| self._model_version = model_version | ||
| self._output_path = output_path | ||
|
|
||
| # model_dir = osp.join(osp.abspath(osp.dirname("__file__")), "models") | ||
| # aesmod_ort_model_path = osp.realpath( | ||
| # osp.join(model_dir, "aes_transonnx_update3.onnx") | ||
| # ) | ||
| self.use_gpu = False | ||
| aesmod_ort_model_path = model_path | ||
| print(aesmod_ort_model_path) | ||
| LOGGER.info("loading aesthetic ort inference session") | ||
| self.ort_session = ort.InferenceSession(aesmod_ort_model_path) | ||
|
|
||
| self.resize_reso = [672, 448] | ||
|
|
||
| def preprocess(self, frame): | ||
| frame = flex_resize_aesv2(frame) | ||
| # print('using flex_resize_aesv2', frame.shape) | ||
| frame = (frame.astype(np.float32) / 255.0 - | ||
| np.array([0.485, 0.456, 0.406], dtype="float32")) / (np.array( | ||
| [0.229, 0.224, 0.225], dtype="float32")) | ||
| frame = np.transpose(frame, (2, 0, 1)) | ||
| frame = np.expand_dims(frame, 0) | ||
| return frame | ||
|
|
||
| @staticmethod | ||
| def tensor_to_list(tensor): | ||
| if tensor.requires_grad: | ||
| return tensor.detach().cpu().flatten().tolist() | ||
| else: | ||
| return tensor.cpu().flatten().tolist() | ||
|
|
||
| @staticmethod | ||
| def score_pred_mapping(raw_scores, raw_min=2.60, raw_max=7.42): | ||
| pred_score = np.clip( | ||
| np.sum([x * (i + 1) for i, x in enumerate(raw_scores)]), raw_min, | ||
| raw_max) | ||
| pred_score = np.sqrt( | ||
| (pred_score - raw_min) / (raw_max - raw_min)) * 100 | ||
| return float(np.clip(pred_score, 0, 100.0)) | ||
|
|
||
| def process(self, frames): | ||
| frames = [ | ||
| frame | ||
| if frame.flags["C_CONTIGUOUS"] else np.ascontiguousarray(frame) | ||
| for frame in frames | ||
| ] | ||
| frame = self.preprocess(frames[0]) | ||
| print("after preprocess shape", frame.shape) | ||
| if not frame.flags["C_CONTIGUOUS"]: | ||
| frame = np.ascontiguousarray(frame, dtype=np.float32) | ||
|
|
||
| t1 = time.time() | ||
| if self.use_gpu: | ||
| with torch.no_grad(): | ||
| input_batch = torch.from_numpy(frame).contiguous().cuda() | ||
| preds, _ = self.trt_model(input_batch) | ||
| raw_score = self.tensor_to_list(preds) | ||
| else: | ||
|
|
||
| raw_score = self.ort_session.run(None, {"input": frame}) | ||
| raw_score = raw_score[0][0] | ||
| score = self.score_pred_mapping(raw_score) | ||
| self._frm_scores.append(score) | ||
| self._frm_idx += 1 | ||
| t2 = time.time() | ||
| LOGGER.info(f"[Aesmod] inference time: {(t2 - t1) * 1000:0.1f} ms") | ||
| return frames[0] | ||
|
|
||
| def clean(self): | ||
| nr_score = round(np.mean(self._frm_scores), 2) | ||
| results = { | ||
| "aesthetic": nr_score, | ||
| "aesthetic_version": self._model_version | ||
| } | ||
| LOGGER.info(f"overall prediction {json.dumps(results)}") | ||
| with open(self._output_path, "w") as outfile: | ||
| json.dump(results, outfile, indent=4, ensure_ascii=False) | ||
|
|
||
|
|
||
| class aesmod_module(SyncModule): | ||
|
|
||
| def __init__(self, node=None, option=None): | ||
| output_path = option.get("result_path", 0) | ||
| params = option.get("params", {}) | ||
| model_version = params.get("model_version", "v1.0") | ||
| model_path = params.get("model_path", | ||
| "../../models/aes_transonnx_update3.onnx") | ||
| self._nrp = Aesmod(model_path, model_version, output_path) | ||
| SyncModule.__init__(self, | ||
| node, | ||
| nb_in=1, | ||
| in_fmt="rgb24", | ||
| out_fmt="rgb24") | ||
|
|
||
| def core_process(self, frames): | ||
| return self._nrp.process(frames) | ||
|
|
||
| def clean(self): | ||
| self._nrp.clean() |
Oops, something went wrong.