Upload 13 files

GPT_eval_multi.py

@@ -0,0 +1,144 @@
+import os 
+import torch
+import numpy as np
+from torch.utils.tensorboard import SummaryWriter
+import json
+# import clip
+from CLIP import clip
+
+import options.option_transformer as option_trans
+import models.vqvae as vqvae
+import utils.utils_model as utils_model
+import eval_trans_per as eval_trans
+from dataset import dataset_TM_eval
+import models.t2m_trans as trans
+from options.get_eval_option import get_opt
+from models.evaluator_wrapper import EvaluatorModelWrapper
+import warnings
+from tqdm import trange
+warnings.filterwarnings('ignore')
+
+##### ---- Exp dirs ---- #####
+os.chdir('/root/autodl-tmp/SATO')
+args = option_trans.get_args_parser()
+torch.manual_seed(args.seed)
+
+args.out_dir = os.path.join(args.out_dir, f'{args.exp_name}')
+os.makedirs(args.out_dir, exist_ok = True)
+
+##### ---- Logger ---- #####
+logger = utils_model.get_logger(args.out_dir)
+writer = SummaryWriter(args.out_dir)
+logger.info(json.dumps(vars(args), indent=4, sort_keys=True))
+
+from utils.word_vectorizer import WordVectorizer
+w_vectorizer = WordVectorizer('./glove', 'our_vab')
+val_loader = dataset_TM_eval.DATALoader(args.dataname, True, 32, w_vectorizer)
+
+dataset_opt_path = 'checkpoints/kit/Comp_v6_KLD005/opt.txt' if args.dataname == 'kit' else 'checkpoints/t2m/Comp_v6_KLD005/opt.txt'
+
+wrapper_opt = get_opt(dataset_opt_path, torch.device('cuda'))
+eval_wrapper = EvaluatorModelWrapper(wrapper_opt)
+
+##### ---- Network ---- #####
+
+## load clip model and datasets
+
+clip_model, clip_preprocess = clip.load(args.clip_path, device=torch.device('cuda'), jit=False)  # Must set jit=False for training
+clip.model.convert_weights(clip_model)  # Actually this line is unnecessary since clip by default already on float16
+clip_model.eval()
+for p in clip_model.parameters():
+    p.requires_grad = False
+
+net = vqvae.HumanVQVAE(args, ## use args to define different parameters in different quantizers
+                       args.nb_code,
+                       args.code_dim,
+                       args.output_emb_width,
+                       args.down_t,
+                       args.stride_t,
+                       args.width,
+                       args.depth,
+                       args.dilation_growth_rate)
+
+
+trans_encoder = trans.Text2Motion_Transformer(num_vq=args.nb_code, 
+                                embed_dim=args.embed_dim_gpt, 
+                                clip_dim=args.clip_dim, 
+                                block_size=args.block_size, 
+                                num_layers=args.num_layers, 
+                                n_head=args.n_head_gpt, 
+                                drop_out_rate=args.drop_out_rate, 
+                                fc_rate=args.ff_rate)
+
+
+print ('loading checkpoint from {}'.format(args.resume_pth))
+ckpt = torch.load(args.resume_pth, map_location='cpu')
+net.load_state_dict(ckpt['net'], strict=True)
+net.eval()
+net.cuda()
+
+if args.resume_trans is not None:
+    print ('loading transformer checkpoint from {}'.format(args.resume_trans))
+    ckpt = torch.load(args.resume_trans, map_location='cpu')
+    trans_encoder.load_state_dict(ckpt['trans'], strict=True)
+trans_encoder.train()
+trans_encoder.cuda()
+print('checkpoints loading successfully')
+
+fid = []
+fid_per=[]
+div = []
+top1 = []
+top2 = []
+top3 = []
+matching = []
+multi = []
+repeat_time = 20
+fid_word_perb=[]
+        
+for i in range(repeat_time):
+    print('repeat_time: ',i)
+    best_fid,best_fid_word_perb,best_fid_per, best_iter, best_div, best_top1, best_top2, best_top3, best_matching, best_multi, writer, logger = eval_trans.evaluation_transformer_test(args.out_dir, val_loader, net, trans_encoder, logger, writer, 0, best_fid=1000,best_fid_word_perb=1000,best_fid_perturbation=1000, best_iter=0, best_div=100, best_top1=0, best_top2=0, best_top3=0, best_matching=100, best_multi=0, clip_model=clip_model, eval_wrapper=eval_wrapper, draw=False, savegif=False, save=False, savenpy=(i==0))
+    fid.append(best_fid)
+    fid_word_perb.append(best_fid_word_perb)
+    fid_per.append(best_fid_per)
+    div.append(best_div)
+    top1.append(best_top1)
+    top2.append(best_top2)
+    top3.append(best_top3)
+    matching.append(best_matching)
+    multi.append(best_multi)
+
+    # print('fid: ', sum(fid)/(i+1))
+    # print('fid_per',sum(fid_per)/(i+1))
+    # print('div: ', sum(div)/(i+1))
+    # print('top1: ', sum(top1)/(i+1))
+    # print('top2: ', sum(top2)/(i+1))
+    # print('top3: ', sum(top3)/(i+1))
+    # print('matching: ', sum(matching)/(i+1))
+    # print('multi: ', sum(multi)/(i+1))
+
+
+print('final result:')
+print('fid: ', sum(fid)/repeat_time)
+print('fid_word_perb',sum(fid_word_perb)/repeat_time)
+print('fid_per',sum(fid_per)/repeat_time)
+print('div: ', sum(div)/repeat_time)
+print('top1: ', sum(top1)/repeat_time)
+print('top2: ', sum(top2)/repeat_time)
+print('top3: ', sum(top3)/repeat_time)
+print('matching: ', sum(matching)/repeat_time)
+# print('multi: ', sum(multi)/repeat_time)
+
+fid = np.array(fid)
+fid_word_perb=np.array(fid_word_perb)
+fid_per=np.array(fid_per)
+div = np.array(div)
+top1 = np.array(top1)
+top2 = np.array(top2)
+top3 = np.array(top3)
+matching = np.array(matching)
+# multi = np.array(multi)
+# msg_final = f"FID. {np.mean(fid):.3f}, FID_PERB.{np.mean(fid_per):.3f}conf. {np.std(fid)*1.96/np.sqrt(repeat_time):.3f}, Diversity. {np.mean(div):.3f}, conf. {np.std(div)*1.96/np.sqrt(repeat_time):.3f}, TOP1. {np.mean(top1):.3f}, conf. {np.std(top1)*1.96/np.sqrt(repeat_time):.3f}, TOP2. {np.mean(top2):.3f}, conf. {np.std(top2)*1.96/np.sqrt(repeat_time):.3f}, TOP3. {np.mean(top3):.3f}, conf. {np.std(top3)*1.96/np.sqrt(repeat_time):.3f}, Matching. {np.mean(matching):.3f}, conf. {np.std(matching)*1.96/np.sqrt(repeat_time):.3f}, Multi. {np.mean(multi):.3f}, conf. {np.std(multi)*1.96/np.sqrt(repeat_time):.3f}"
+msg_final = f"FID. {np.mean(fid):.3f}, {np.std(fid)*1.96/np.sqrt(repeat_time):.3f}, FID_word_perb.{np.mean(fid_word_perb):.3f}, {np.std(fid_word_perb)*1.96/np.sqrt(repeat_time):.3f},FID_PERB.{np.mean(fid_per):.3f}, conf. {np.std(fid)*1.96/np.sqrt(repeat_time):.3f}, Diversity. {np.mean(div):.3f}, conf. {np.std(div)*1.96/np.sqrt(repeat_time):.3f}, TOP1. {np.mean(top1):.3f}, conf. {np.std(top1)*1.96/np.sqrt(repeat_time):.3f}, TOP2. {np.mean(top2):.3f}, conf. {np.std(top2)*1.96/np.sqrt(repeat_time):.3f}, TOP3. {np.mean(top3):.3f}, conf. {np.std(top3)*1.96/np.sqrt(repeat_time):.3f}, Matching. {np.mean(matching):.3f}, conf. {np.std(matching)*1.96/np.sqrt(repeat_time):.3f}, conf. {np.std(multi)*1.96/np.sqrt(repeat_time):.3f}"
+logger.info(msg_final)

LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright 2023 tencent
+
+   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.

README.md

@@ -1,13 +1,227 @@
----
-title: SATO
-emoji: 🏃
-colorFrom: green
-colorTo: purple
-sdk: gradio
-sdk_version: 4.28.3
-app_file: app.py
-pinned: false
-license: mit
----
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# SATO: Stable Text-to-Motion Framework
+
+[Wenshuo chen*](https://github.com/shurdy123), [Hongru Xiao*](https://github.com/Hongru0306), [Erhang Zhang*](https://github.com/zhangerhang), [Lijie Hu](https://sites.google.com/view/lijiehu/homepage), [Lei Wang](https://leiwangr.github.io/), [Mengyuan Liu](), [Chen Chen](https://www.crcv.ucf.edu/chenchen/)
+
+[![Website shields.io](https://img.shields.io/website?url=http%3A//poco.is.tue.mpg.de)](https://sato-team.github.io/Stable-Text-to-Motion-Framework/) [![YouTube Badge](https://img.shields.io/badge/YouTube-Watch-red?style=flat-square&logo=youtube)]()  [![arXiv](https://img.shields.io/badge/arXiv-2308.12965-00ff00.svg)]()  
+## Existing Challenges
+A fundamental challenge inherent in text-to-motion tasks stems from the variability of textual inputs. Even when conveying similar or the same meanings and intentions, texts can exhibit considerable variations in vocabulary and structure due to individual user preferences or linguistic nuances. Despite the considerable advancements made in these models, we find a notable weakness: all of them demonstrate instability in prediction when encountering minor textual perturbations, such as synonym substitutions. In the following demonstration, we showcase the instability of predictions generated by the previous method when presented with different user inputs conveying identical semantic meaning.
+<!-- <div style="display:flex;">
+    <img src="assets/run_lola.gif" width="45%" style="margin-right: 1%;">
+    <img src="assets/yt_solo.gif" width="45%">
+</div> -->
+
+<p align="center">
+<table align="center">
+  <tr>
+    <th colspan="4">Original text: A man kicks something or someone with his left leg.</th>
+  </tr>
+  <tr>
+    <th align="center"><u><a href="https://github.com/Mael-zys/T2M-GPT"><nobr>T2M-GPT</nobr> </a></u></th>
+    <th align="center"><u><a href="https://guytevet.github.io/mdm-page/"><nobr>MDM</nobr> </a></u></th>
+    <th align="center"><u><a href="https://github.com/EricGuo5513/momask-codes"><nobr>MoMask</nobr> </a></u></th>
+  </tr>
+
+  <tr>
+    <td width="250" align="center"><img src="images/example/kick/gpt.gif" width="150px" height="150px" alt="gif"></td>
+    <td width="250" align="center"><img src="images/example/kick/mdm.gif" width="150px" height="150px" alt="gif"></td>
+    <td width="250" align="center"><img src="images/example/kick/momask.gif" width="150px" height="150px" alt="gif"></td>
+  </tr>
+
+  <tr>
+    <th colspan="4" >Perturbed text: A human boots something or someone with his left leg.</th>
+  </tr>
+  <tr>
+    <th align="center"><u><a href="https://github.com/Mael-zys/T2M-GPT"><nobr>T2M-GPT</nobr> </a></u></th>
+    <th align="center"><u><a href="https://guytevet.github.io/mdm-page/"><nobr>MDM</nobr> </a></u></th>
+    <th align="center"><u><a href="https://github.com/EricGuo5513/momask-codes"><nobr>MoMask</nobr> </a></u></th>
+  </tr>
+
+  <tr>
+    <td width="250" align="center"><img src="images/example/boot/gpt.gif" width="150px" height="150px" alt="gif"></td>
+    <td width="250" align="center"><img src="images/example/boot/mdm.gif" width="150px" height="150px" alt="gif"></td>
+    <td width="250" align="center"><img src="images/example/boot/momask.gif" width="150px" height="150px" alt="gif"></td>
+  </tr>
+</table>
+</p>
+
+## Motivation
+![motivation](images/motivation.png)
+The model's inconsistent outputs are accompanied by unstable attention patterns. We further elucidate the aforementioned experimental findings: When perturbed text is inputted, the model exhibits unstable attention, often neglecting critical text elements necessary for accurate motion prediction. This instability further complicates the encoding of text into consistent embeddings, leading to a cascade of consecutive temporal motion generation errors.
+
+
+## Visualization
+<p align="center">
+<table align="center">
+  <tr>
+    <th colspan="4">Original text: person is walking normally in a circle.</th>
+  </tr>
+  <tr>
+    <th align="center"><u><a href="https://github.com/Mael-zys/T2M-GPT"><nobr>T2M-GPT</nobr> </a></u></th>
+    <th align="center"><u><a href="https://guytevet.github.io/mdm-page/"><nobr>MDM</nobr> </a></u></th>
+    <th align="center"><u><a href="https://github.com/EricGuo5513/momask-codes"><nobr>MoMask</nobr> </a></u></th>
+    <th align="center"><nobr>SATO</nobr> </th>
+  </tr>
+
+  <tr>
+    <td width="250" align="center"><img src="images/visualization/circle/gpt.gif" width="150px" height="150px" alt="gif"></td>
+    <td width="250" align="center"><img src="images/visualization/circle/mdm.gif" width="150px" height="150px" alt="gif"></td>
+    <td width="250" align="center"><img src="images/visualization/circle/momask.gif" width="150px" height="150px" alt="gif"></td>
+    <td width="250" align="center"><img src="images/visualization/circle/sato.gif" width="150px" height="150px" alt="gif"></td>
+  </tr>
+
+  <tr>
+    <th colspan="4" >Perturbed text: <span style="color: red;">human</span> is walking <span style="color: red;">usually</span> in a <span style="color: red;">loop.</th>
+  </tr>
+  <tr>
+    <th align="center"><u><a href="https://github.com/Mael-zys/T2M-GPT"><nobr>T2M-GPT</nobr> </a></u></th>
+    <th align="center"><u><a href="https://guytevet.github.io/mdm-page/"><nobr>MDM</nobr> </a></u></th>
+    <th align="center"><u><a href="https://github.com/EricGuo5513/momask-codes"><nobr>MoMask</nobr> </a></u></th>
+    <th align="center"><nobr>SATO</nobr> </th>
+  </tr>
+
+  <tr>
+    <td width="250" align="center"><img src="images/visualization/loop/gpt.gif" width="150px" height="150px" alt="gif"></td>
+    <td width="250" align="center"><img src="images/visualization/loop/mdm.gif" width="150px" height="150px" alt="gif"></td>
+    <td width="250" align="center"><img src="images/visualization/loop/momask.gif" width="150px" height="150px" alt="gif"></td>
+    <td width="250" align="center"><img src="images/visualization/loop/sato.gif" width="150px" height="150px" alt="gif"></td>
+  </tr>
+</table>
+</p>
+<center>
+  <h3>
+    <p style="color: blue;">Explanation: T2M-GPT, MDM, and MoMask all don't walk in a loop.</p>
+  </h3>
+
+<p align="center">
+<table align="center">
+  <tr>
+    <th colspan="4">Original text: a person uses his right arm to help himself to stand up.</th>
+  </tr>
+  <tr>
+    <th align="center"><u><a href="https://github.com/Mael-zys/T2M-GPT"><nobr>T2M-GPT</nobr> </a></u></th>
+    <th align="center"><u><a href="https://guytevet.github.io/mdm-page/"><nobr>MDM</nobr> </a></u></th>
+    <th align="center"><u><a href="https://github.com/EricGuo5513/momask-codes"><nobr>MoMask</nobr> </a></u></th>
+    <th align="center"><nobr>SATO</nobr> </th>
+  </tr>
+
+  <tr>
+    <td width="250" align="center"><img src="images/visualization/use/gpt.gif" width="150px" height="150px" alt="gif"></td>
+    <td width="250" align="center"><img src="images/visualization/use/mdm.gif" width="150px" height="150px" alt="gif"></td>
+    <td width="250" align="center"><img src="images/visualization/use/momask.gif" width="150px" height="150px" alt="gif"></td>
+    <td width="250" align="center"><img src="images/visualization/use/sato.gif" width="150px" height="150px" alt="gif"></td>
+  </tr>
+
+  <tr>
+    <th colspan="4" >Perturbed text: A human <span style="color: red;">utilizes</span> his right arm to help himself to stand up.</th>
+  </tr>
+  <tr>
+    <th align="center"><u><a href="https://github.com/Mael-zys/T2M-GPT"><nobr>T2M-GPT</nobr> </a></u></th>
+    <th align="center"><u><a href="https://guytevet.github.io/mdm-page/"><nobr>MDM</nobr> </a></u></th>
+    <th align="center"><u><a href="https://github.com/EricGuo5513/momask-codes"><nobr>MoMask</nobr> </a></u></th>
+    <th align="center"><nobr>SATO</nobr> </th>
+  </tr>
+
+  <tr>
+    <td width="250" align="center"><img src="images/visualization/utilize/gpt.gif" width="150px" height="150px" alt="gif"></td>
+    <td width="250" align="center"><img src="images/visualization/utilize/mdm.gif" width="150px" height="150px" alt="gif"></td>
+    <td width="250" align="center"><img src="images/visualization/utilize/momask.gif" width="150px" height="150px" alt="gif"></td>
+    <td width="250" align="center"><img src="images/visualization/utilize/sato.gif" width="150px" height="150px" alt="gif"></td>
+  </tr>
+</table>
+</p>
+<center>
+  <h3>
+    <p style="color: blue;">Explanation: T2M-GPT, MDM, and MoMask all lack the action of transitioning from squatting to standing up, resulting in a catastrophic error.</p>
+  </h3>
+
+    
+## How to Use the Code
+
+* [1. Setup and Installation](#setup)
+
+* [2.Dependencies](#Dependencies)
+
+* [3. Quick Start](#quickstart)
+
+* [4. Datasets](#datasets)
+
+* [4. Train](#train)
+
+* [5. Evaluation](#eval)
+
+* [6. Acknowledgments](#acknowledgements)
+
+    
+## Setup and Installation <a name="setup"></a>
+
+Clone the repository: 
+
+```shell
+git clone https://github.com/sato-team/Stable-Text-to-motion-Framework.git
+```
+
+Create fresh conda environment and install all the dependencies:
+
+```
+conda env create -f environment.yml
+conda activate SATO
+```
+
+The code was tested on Python 3.8 and PyTorch 1.8.1.
+
+## Dependencies<a name="Dependencies"></a>
+
+```shell
+bash dataset/prepare/download_extractor.sh
+bash dataset/prepare/download_glove.sh
+```
+
+## **Quick Start**<a name="quickstart"></a>
+
+A quick reference guide for using our code is provided in quickstart.ipynb.
+
+## Datasets<a name="datasets"></a>
+
+We are using two 3D human motion-language dataset: HumanML3D and KIT-ML. For both datasets, you could find the details as well as download [link](https://github.com/EricGuo5513/HumanML3D).
+We perturbed the input texts based on the two datasets mentioned. You can access the perturbed text dataset through the following [link](https://drive.google.com/file/d/1XLvu2jfG1YKyujdANhYHV_NfFTyOJPvP/view?usp=sharing).
+Take HumanML3D for an example, the dataset structure should look like this:  
+```
+./dataset/HumanML3D/
+├── new_joint_vecs/
+├── texts/ # You need to replace the 'texts' folder in the original dataset with the 'texts' folder from our dataset.
+├── Mean.npy 
+├── Std.npy 
+├── train.txt
+├── val.txt
+├── test.txt
+├── train_val.txt
+└── all.txt
+```
+### **Train**<a name="train"></a>
+
+We will release the training code soon.
+
+### **Evaluation**<a name="eval"></a>
+
+You can download the pretrained models in this [link](https://drive.google.com/drive/folders/1rs8QPJ3UPzLW4H3vWAAX9hJn4ln7m_ts?usp=sharing). 
+
+```shell
+python eval_t2m.py --resume-pth pretrained/net_best_fid.pth --clip_path pretrained/clip_best_fid.pth
+```
+
+## Acknowledgements<a name="acknowledgements"></a>
+
+We appreciate helps from :
+
+- Open Source Code：[T2M-GPT](https://github.com/Mael-zys/T2M-GPT), [MoMask ](https://github.com/EricGuo5513/momask-codes), [MDM](https://guytevet.github.io/mdm-page/), etc.
+- [Hongru Xiao](https://github.com/Hongru0306), [Erhang Zhang](https://github.com/zhangerhang), [Lijie Hu](https://sites.google.com/view/lijiehu/homepage), [Lei Wang](https://leiwangr.github.io/), [Mengyuan Liu](), [Chen Chen](https://www.crcv.ucf.edu/chenchen/) for discussions and guidance throughout the project, which has been instrumental to our work.
+- [Zhen Zhao](https://github.com/Zanebla) for project website.
+- If you find our work helpful, we would appreciate it if you could give our project a star!
+## Citing<a name="citing"></a>
+
+If you find this code useful for your research, please consider citing the following paper:
+
+```bibtex
+
+```
+

VQ_eval.py

@@ -0,0 +1,95 @@
+import os
+import json
+
+import torch
+from torch.utils.tensorboard import SummaryWriter
+import numpy as np
+import models.vqvae as vqvae
+import options.option_vq as option_vq
+import utils.utils_model as utils_model
+from dataset import dataset_TM_eval
+import utils.eval_trans as eval_trans
+from options.get_eval_option import get_opt
+from models.evaluator_wrapper import EvaluatorModelWrapper
+import warnings
+warnings.filterwarnings('ignore')
+import numpy as np
+##### ---- Exp dirs ---- #####
+args = option_vq.get_args_parser()
+torch.manual_seed(args.seed)
+
+args.out_dir = os.path.join(args.out_dir, f'{args.exp_name}')
+os.makedirs(args.out_dir, exist_ok = True)
+
+##### ---- Logger ---- #####
+logger = utils_model.get_logger(args.out_dir)
+writer = SummaryWriter(args.out_dir)
+logger.info(json.dumps(vars(args), indent=4, sort_keys=True))
+
+
+from utils.word_vectorizer import WordVectorizer
+w_vectorizer = WordVectorizer('./glove', 'our_vab')
+
+
+dataset_opt_path = 'checkpoints/kit/Comp_v6_KLD005/opt.txt' if args.dataname == 'kit' else 'checkpoints/t2m/Comp_v6_KLD005/opt.txt'
+
+wrapper_opt = get_opt(dataset_opt_path, torch.device('cuda'))
+eval_wrapper = EvaluatorModelWrapper(wrapper_opt)
+
+
+##### ---- Dataloader ---- #####
+args.nb_joints = 21 if args.dataname == 'kit' else 22
+
+val_loader = dataset_TM_eval.DATALoader(args.dataname, True, 32, w_vectorizer, unit_length=2**args.down_t)
+
+##### ---- Network ---- #####
+net = vqvae.HumanVQVAE(args, ## use args to define different parameters in different quantizers
+                       args.nb_code,
+                       args.code_dim,
+                       args.output_emb_width,
+                       args.down_t,
+                       args.stride_t,
+                       args.width,
+                       args.depth,
+                       args.dilation_growth_rate,
+                       args.vq_act,
+                       args.vq_norm)
+
+if args.resume_pth : 
+    logger.info('loading checkpoint from {}'.format(args.resume_pth))
+    ckpt = torch.load(args.resume_pth, map_location='cpu')
+    net.load_state_dict(ckpt['net'], strict=True)
+net.train()
+net.cuda()
+
+fid = []
+div = []
+top1 = []
+top2 = []
+top3 = []
+matching = []
+repeat_time = 20
+for i in range(repeat_time):
+    best_fid, best_iter, best_div, best_top1, best_top2, best_top3, best_matching, writer, logger = eval_trans.evaluation_vqvae(args.out_dir, val_loader, net, logger, writer, 0, best_fid=1000, best_iter=0, best_div=100, best_top1=0, best_top2=0, best_top3=0, best_matching=100, eval_wrapper=eval_wrapper, draw=False, save=False, savenpy=(i==0))
+    fid.append(best_fid)
+    div.append(best_div)
+    top1.append(best_top1)
+    top2.append(best_top2)
+    top3.append(best_top3)
+    matching.append(best_matching)
+print('final result:')
+print('fid: ', sum(fid)/repeat_time)
+print('div: ', sum(div)/repeat_time)
+print('top1: ', sum(top1)/repeat_time)
+print('top2: ', sum(top2)/repeat_time)
+print('top3: ', sum(top3)/repeat_time)
+print('matching: ', sum(matching)/repeat_time)
+
+fid = np.array(fid)
+div = np.array(div)
+top1 = np.array(top1)
+top2 = np.array(top2)
+top3 = np.array(top3)
+matching = np.array(matching)
+msg_final = f"FID. {np.mean(fid):.3f}, conf. {np.std(fid)*1.96/np.sqrt(repeat_time):.3f}, Diversity. {np.mean(div):.3f}, conf. {np.std(div)*1.96/np.sqrt(repeat_time):.3f}, TOP1. {np.mean(top1):.3f}, conf. {np.std(top1)*1.96/np.sqrt(repeat_time):.3f}, TOP2. {np.mean(top2):.3f}, conf. {np.std(top2)*1.96/np.sqrt(repeat_time):.3f}, TOP3. {np.mean(top3):.3f}, conf. {np.std(top3)*1.96/np.sqrt(repeat_time):.3f}, Matching. {np.mean(matching):.3f}, conf. {np.std(matching)*1.96/np.sqrt(repeat_time):.3f}"
+logger.info(msg_final)

attack.py

@@ -0,0 +1,182 @@
+import torch
+import numpy as np
+# device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+
+class PGDAttacker():
+    def __init__(self, radius, steps, step_size, random_start, norm_type, ascending=True):
+        self.radius = radius  # attack radius
+        self.steps = steps # how many step to conduct pgd
+        self.step_size = step_size  # coefficient of PGD
+        self.random_start = random_start
+        self.norm_type = norm_type # which norm of your noise
+        self.ascending = ascending # perform gradient ascending, i.e, to maximum the loss
+
+    def output(self, x, model, tokens_lens, text_token):
+        
+        x = x + model.positional_embedding.type(model.dtype)
+
+        x = x.permute(1, 0, 2)  # NLD -> LND
+        x, weight = model.transformer(x)
+        x = x.permute(1, 0, 2)  # LND -> NLD
+        x = model.ln_final(x).type(model.dtype)
+        x = x[torch.arange(x.shape[0]), text_token.argmax(dim=-1)] @ model.text_projection
+
+        attention_weights_all = []
+        for i in range(len(tokens_lens)):
+            attention_weights = weight[-1][i][min(76, tokens_lens[i])][:1+min(75, max(tokens_lens))][1:][:-1]
+            attention_weights_all.append(attention_weights)
+        attention_weights = torch.stack(attention_weights_all, dim=0)
+
+        return x, attention_weights
+
+    def perturb(self, device, m_tokens_len, bs, criterion, x, y,a_indices,encoder, tokens_lens=None, model=None, text_token=None):
+        if self.steps==0 or self.radius==0:
+            return x.clone()
+
+        adv_x = x.clone()
+
+        if self.random_start:
+            if self.norm_type == 'l-infty':
+                adv_x += 2 * (torch.rand_like(x) - 0.5) * self.radius
+            else:
+                adv_x += 2 * (torch.rand_like(x) - 0.5) * self.radius / self.steps
+            self._clip_(adv_x, x)
+
+        ''' temporarily shutdown autograd of model to improve pgd efficiency '''
+        # adv_x, attention_weights = self.output(adv_x, model, tokens_lens, text_token)
+
+        # model.eval()
+        encoder.eval()
+        for pp in encoder.parameters():
+            pp.requires_grad = False
+
+        for step in range(self.steps):
+            adv_x_o = adv_x.clone()
+            adv_x.requires_grad_()
+            _y = encoder(a_indices,adv_x)
+            loss = criterion(y.to(device), _y, m_tokens_len, bs)
+            grad = torch.autograd.grad(loss, [adv_x])[0]
+
+            with torch.no_grad():
+                if not self.ascending: grad.mul_(-1)
+
+                if self.norm_type == 'l-infty':
+                    adv_x.add_(torch.sign(grad), alpha=self.step_size)
+                else:
+                    if self.norm_type == 'l2':
+                        grad_norm = (grad.reshape(grad.shape[0],-1)**2).sum(dim=1).sqrt()
+                    elif self.norm_type == 'l1':
+                        grad_norm = grad.reshape(grad.shape[0],-1).abs().sum(dim=1)
+                    grad_norm = grad_norm.reshape( -1, *( [1] * (len(x.shape)-1) ) )
+                    scaled_grad = grad / (grad_norm + 1e-10)
+                    adv_x.add_(scaled_grad, alpha=self.step_size)
+
+                self._clip_(adv_x, adv_x_o)
+
+        ''' reopen autograd of model after pgd '''
+        # decoder.trian()
+        for pp in encoder.parameters():
+            pp.requires_grad = True
+
+        return adv_x  # , attention_weights
+    
+    def perturb_random(self, criterion, x, data, decoder,y,target_model,encoder=None):
+        if self.steps==0 or self.radius==0:
+            return x.clone()
+        adv_x = x.clone()
+        if self.norm_type == 'l-infty':
+            adv_x += 2 * (torch.rand_like(x) - 0.5) * self.radius
+        else:
+            adv_x += 2 * (torch.rand_like(x) - 0.5) * self.radius / self.steps
+        self._clip_(adv_x, x)
+        return adv_x.data
+    
+    def perturb_iat(self, criterion, x, data, decoder,y,target_model,encoder=None):
+        if self.steps==0 or self.radius==0:
+            return x.clone()
+        
+        B = x.shape[0]
+        L = x.shape[1]
+        H = x.shape[2]
+        nb_num = 8
+        
+        alpha = torch.rand(B,L,nb_num,1).to(device)
+        
+        A_1 = x.unsqueeze(2).expand(B,L,nb_num,H)
+        A_2 = x.unsqueeze(1).expand(B,L,L,H)
+        rand_idx = []
+        for i in range(L):
+            rand_idx.append(np.random.choice(L,nb_num,replace=False))
+        rand_idx = np.array(rand_idx)
+        rand_idx = torch.tensor(rand_idx).long().reshape(1,L,1,nb_num).expand(B,L,H,nb_num).to(device)
+        # A_2 = A_2[:,np.arange(0,L), rand_idx,:]
+        A_2 = torch.gather(A_2.reshape(B,L,H,L),-1,rand_idx).reshape(B,L,nb_num, H)
+        A_e = A_1 - A_2
+        # A_e
+        # adv_x = (A_e * alpha).sum(dim=-1) + x.clone()
+        
+        adv_x = x.clone()
+
+        if self.random_start:
+            if self.norm_type == 'l-infty':
+                adv_x += 2 * (torch.rand_like(x) - 0.5) * self.radius
+            else:
+                adv_x += 2 * (torch.rand_like(x) - 0.5) * self.radius / self.steps
+        self._clip_(adv_x, x)
+
+        # assert adv_x.shape[0] == 8
+
+        ''' temporarily shutdown autograd of model to improve pgd efficiency '''
+        # model.eval()
+        decoder.eval()
+        for pp in decoder.parameters():
+            pp.requires_grad = False
+            
+        adv_x = x.clone()
+        
+        alpha.requires_grad_()
+
+        for step in range(self.steps):
+            alpha.requires_grad_()
+            dot_Ae_alpha = (A_e * alpha).sum(dim=-2)
+            # print("dot_Ae_alpha:", dot_Ae_alpha.shape)
+            
+            adv_x.add_(torch.sign(dot_Ae_alpha), alpha=self.step_size)
+            
+            self._clip_(adv_x, x)
+            
+            if encoder is None:
+                adv_x_input = adv_x.squeeze(-1)
+            else:
+                adv_x_input = adv_x
+            
+            _y = target_model(adv_x_input, data,decoder,encoder)
+            loss = criterion(y.to(device), _y)
+            grad = torch.autograd.grad(loss, [alpha],retain_graph=True)[0]
+            # with torch.no_grad():
+            with torch.no_grad():
+                if not self.ascending: grad.mul_(-1)
+                assert self.norm_type == 'l-infty'
+                alpha = alpha.detach()+ grad * 0.01
+
+        ''' reopen autograd of model after pgd '''
+        # decoder.trian()
+        for pp in decoder.parameters():
+            pp.requires_grad = True
+
+        return adv_x.data
+
+    def _clip_(self, adv_x, x):
+        adv_x -= x
+        if self.norm_type == 'l-infty':
+            adv_x.clamp_(-self.radius, self.radius)
+        else:
+            if self.norm_type == 'l2':
+                norm = (adv_x.reshape(adv_x.shape[0],-1)**2).sum(dim=1).sqrt()
+            elif self.norm_type == 'l1':
+                norm = adv_x.reshape(adv_x.shape[0],-1).abs().sum(dim=1)
+            norm = norm.reshape( -1, *( [1] * (len(x.shape)-1) ) )
+            adv_x /= (norm + 1e-10)
+            adv_x *= norm.clamp(max=self.radius)
+        adv_x += x
+        adv_x.clamp_(0, 1)

environment.yml

@@ -0,0 +1,121 @@
+name: SATO
+channels:
+  - pytorch
+  - defaults
+dependencies:
+  - _libgcc_mutex=0.1=main
+  - _openmp_mutex=4.5=1_gnu
+  - blas=1.0=mkl
+  - bzip2=1.0.8=h7b6447c_0
+  - ca-certificates=2021.7.5=h06a4308_1
+  - certifi=2021.5.30=py38h06a4308_0
+  - cudatoolkit=10.1.243=h6bb024c_0
+  - ffmpeg=4.3=hf484d3e_0
+  - freetype=2.10.4=h5ab3b9f_0
+  - gmp=6.2.1=h2531618_2
+  - gnutls=3.6.15=he1e5248_0
+  - intel-openmp=2021.3.0=h06a4308_3350
+  - jpeg=9b=h024ee3a_2
+  - lame=3.100=h7b6447c_0
+  - lcms2=2.12=h3be6417_0
+  - ld_impl_linux-64=2.35.1=h7274673_9
+  - libffi=3.3=he6710b0_2
+  - libgcc-ng=9.3.0=h5101ec6_17
+  - libgomp=9.3.0=h5101ec6_17
+  - libiconv=1.15=h63c8f33_5
+  - libidn2=2.3.2=h7f8727e_0
+  - libpng=1.6.37=hbc83047_0
+  - libstdcxx-ng=9.3.0=hd4cf53a_17
+  - libtasn1=4.16.0=h27cfd23_0
+  - libtiff=4.2.0=h85742a9_0
+  - libunistring=0.9.10=h27cfd23_0
+  - libuv=1.40.0=h7b6447c_0
+  - libwebp-base=1.2.0=h27cfd23_0
+  - lz4-c=1.9.3=h295c915_1
+  - mkl=2021.3.0=h06a4308_520
+  - mkl-service=2.4.0=py38h7f8727e_0
+  - mkl_fft=1.3.0=py38h42c9631_2
+  - mkl_random=1.2.2=py38h51133e4_0
+  - ncurses=6.2=he6710b0_1
+  - nettle=3.7.3=hbbd107a_1
+  - ninja=1.10.2=hff7bd54_1
+  - numpy=1.20.3=py38hf144106_0
+  - numpy-base=1.20.3=py38h74d4b33_0
+  - olefile=0.46=py_0
+  - openh264=2.1.0=hd408876_0
+  - openjpeg=2.3.0=h05c96fa_1
+  - openssl=1.1.1k=h27cfd23_0
+  - pillow=8.3.1=py38h2c7a002_0
+  - pip=21.0.1=py38h06a4308_0
+  - python=3.8.11=h12debd9_0_cpython
+  - pytorch=1.8.1=py3.8_cuda10.1_cudnn7.6.3_0
+  - readline=8.1=h27cfd23_0
+  - setuptools=52.0.0=py38h06a4308_0
+  - six=1.16.0=pyhd3eb1b0_0
+  - sqlite=3.36.0=hc218d9a_0
+  - tk=8.6.10=hbc83047_0
+  - torchaudio=0.8.1=py38
+  - torchvision=0.9.1=py38_cu101
+  - typing_extensions=3.10.0.0=pyh06a4308_0
+  - wheel=0.37.0=pyhd3eb1b0_0
+  - xz=5.2.5=h7b6447c_0
+  - zlib=1.2.11=h7b6447c_3
+  - zstd=1.4.9=haebb681_0
+  - pip:
+    - absl-py==0.13.0
+    - backcall==0.2.0
+    - cachetools==4.2.2
+    - charset-normalizer==2.0.4
+    - chumpy==0.70
+    - cycler==0.10.0
+    - decorator==5.0.9
+    - google-auth==1.35.0
+    - google-auth-oauthlib==0.4.5
+    - grpcio==1.39.0
+    - idna==3.2
+    - imageio==2.9.0
+    - ipdb==0.13.9
+    - ipython==7.26.0
+    - ipython-genutils==0.2.0
+    - jedi==0.18.0
+    - joblib==1.0.1
+    - kiwisolver==1.3.1
+    - markdown==3.3.4
+    - matplotlib==3.4.3
+    - matplotlib-inline==0.1.2
+    - oauthlib==3.1.1
+    - pandas==1.3.2
+    - parso==0.8.2
+    - pexpect==4.8.0
+    - pickleshare==0.7.5
+    - prompt-toolkit==3.0.20
+    - protobuf==3.17.3
+    - ptyprocess==0.7.0
+    - pyasn1==0.4.8
+    - pyasn1-modules==0.2.8
+    - pygments==2.10.0
+    - pyparsing==2.4.7
+    - python-dateutil==2.8.2
+    - pytz==2021.1
+    - pyyaml==5.4.1
+    - requests==2.26.0
+    - requests-oauthlib==1.3.0
+    - rsa==4.7.2
+    - scikit-learn==0.24.2
+    - scipy==1.7.1
+    - sklearn==0.0
+    - smplx==0.1.28
+    - tensorboard==2.6.0
+    - tensorboard-data-server==0.6.1
+    - tensorboard-plugin-wit==1.8.0
+    - threadpoolctl==2.2.0
+    - toml==0.10.2
+    - tqdm==4.62.2
+    - traitlets==5.0.5
+    - urllib3==1.26.6
+    - wcwidth==0.2.5
+    - werkzeug==2.0.1
+    - git+https://github.com/openai/CLIP.git
+    - git+https://github.com/nghorbani/human_body_prior
+    - gdown
+    - moviepy

eval_trans_per.py

@@ -0,0 +1,653 @@
+import os
+
+# import clip
+from CLIP.clip import clip
+import numpy as np
+import torch
+from scipy import linalg
+from tqdm import tqdm
+import visualization.plot_3d_global as plot_3d
+from utils.motion_process import recover_from_ric
+from tqdm import trange
+
+def tensorborad_add_video_xyz(writer, xyz, nb_iter, tag, nb_vis=4, title_batch=None, outname=None):
+    xyz = xyz[:1]
+    bs, seq = xyz.shape[:2]
+    xyz = xyz.reshape(bs, seq, -1, 3)
+    plot_xyz = plot_3d.draw_to_batch(xyz.cpu().numpy(),title_batch, outname)
+    plot_xyz =np.transpose(plot_xyz, (0, 1, 4, 2, 3)) 
+    writer.add_video(tag, plot_xyz, nb_iter, fps = 20)
+
+@torch.no_grad()        
+def evaluation_vqvae(out_dir, val_loader, net, logger, writer, nb_iter, best_fid, best_iter, best_div, best_top1, best_top2, best_top3, best_matching, eval_wrapper, draw = True, save = True, savegif=False, savenpy=False) : 
+    net.eval()
+    nb_sample = 0
+    
+    draw_org = []
+    draw_pred = []
+    draw_text = []
+
+
+    motion_annotation_list = []
+    motion_pred_list = []
+
+    R_precision_real = 0
+    R_precision = 0
+
+    nb_sample = 0
+    matching_score_real = 0
+    matching_score_pred = 0
+    for batch in val_loader:
+        word_embeddings, pos_one_hots, caption, sent_len, motion, m_length, token, name = batch
+
+        motion = motion.cuda()
+        et, em = eval_wrapper.get_co_embeddings(word_embeddings, pos_one_hots, sent_len, motion, m_length)
+        bs, seq = motion.shape[0], motion.shape[1]
+
+        num_joints = 21 if motion.shape[-1] == 251 else 22
+        
+        pred_pose_eval = torch.zeros((bs, seq, motion.shape[-1])).cuda()
+
+        for i in range(bs):
+            pose = val_loader.dataset.inv_transform(motion[i:i+1, :m_length[i], :].detach().cpu().numpy())
+            pose_xyz = recover_from_ric(torch.from_numpy(pose).float().cuda(), num_joints)
+
+
+            pred_pose, loss_commit, perplexity = net(motion[i:i+1, :m_length[i]])
+            pred_denorm = val_loader.dataset.inv_transform(pred_pose.detach().cpu().numpy())
+            pred_xyz = recover_from_ric(torch.from_numpy(pred_denorm).float().cuda(), num_joints)
+            
+            if savenpy:
+                np.save(os.path.join(out_dir, name[i]+'_gt.npy'), pose_xyz[:, :m_length[i]].cpu().numpy())
+                np.save(os.path.join(out_dir, name[i]+'_pred.npy'), pred_xyz.detach().cpu().numpy())
+
+            pred_pose_eval[i:i+1,:m_length[i],:] = pred_pose
+
+            if i < min(4, bs):
+                draw_org.append(pose_xyz)
+                draw_pred.append(pred_xyz)
+                draw_text.append(caption[i])
+
+        et_pred, em_pred = eval_wrapper.get_co_embeddings(word_embeddings, pos_one_hots, sent_len, pred_pose_eval, m_length)
+
+        motion_pred_list.append(em_pred)
+        motion_annotation_list.append(em)
+            
+        temp_R, temp_match = calculate_R_precision(et.cpu().numpy(), em.cpu().numpy(), top_k=3, sum_all=True)
+        R_precision_real += temp_R
+        matching_score_real += temp_match
+        temp_R, temp_match = calculate_R_precision(et_pred.cpu().numpy(), em_pred.cpu().numpy(), top_k=3, sum_all=True)
+        R_precision += temp_R
+        matching_score_pred += temp_match
+
+        nb_sample += bs
+
+    motion_annotation_np = torch.cat(motion_annotation_list, dim=0).cpu().numpy()
+    motion_pred_np = torch.cat(motion_pred_list, dim=0).cpu().numpy()
+    gt_mu, gt_cov  = calculate_activation_statistics(motion_annotation_np)
+    mu, cov= calculate_activation_statistics(motion_pred_np)
+
+    diversity_real = calculate_diversity(motion_annotation_np, 300 if nb_sample > 300 else 100)
+    diversity = calculate_diversity(motion_pred_np, 300 if nb_sample > 300 else 100)
+
+    R_precision_real = R_precision_real / nb_sample
+    R_precision = R_precision / nb_sample
+
+    matching_score_real = matching_score_real / nb_sample
+    matching_score_pred = matching_score_pred / nb_sample
+
+    fid = calculate_frechet_distance(gt_mu, gt_cov, mu, cov)
+
+    msg = f"--> \t Eva. Iter {nb_iter} :, FID. {fid:.4f}, Diversity Real. {diversity_real:.4f}, Diversity. {diversity:.4f}, R_precision_real. {R_precision_real}, R_precision. {R_precision}, matching_score_real. {matching_score_real}, matching_score_pred. {matching_score_pred}"
+    logger.info(msg)
+    
+    if draw:
+        writer.add_scalar('./Test/FID', fid, nb_iter)
+        writer.add_scalar('./Test/Diversity', diversity, nb_iter)
+        writer.add_scalar('./Test/top1', R_precision[0], nb_iter)
+        writer.add_scalar('./Test/top2', R_precision[1], nb_iter)
+        writer.add_scalar('./Test/top3', R_precision[2], nb_iter)
+        writer.add_scalar('./Test/matching_score', matching_score_pred, nb_iter)
+
+    
+        if nb_iter % 5000 == 0 : 
+            for ii in range(4):
+                tensorborad_add_video_xyz(writer, draw_org[ii], nb_iter, tag='./Vis/org_eval'+str(ii), nb_vis=1, title_batch=[draw_text[ii]], outname=[os.path.join(out_dir, 'gt'+str(ii)+'.gif')] if savegif else None)
+            
+        if nb_iter % 5000 == 0 : 
+            for ii in range(4):
+                tensorborad_add_video_xyz(writer, draw_pred[ii], nb_iter, tag='./Vis/pred_eval'+str(ii), nb_vis=1, title_batch=[draw_text[ii]], outname=[os.path.join(out_dir, 'pred'+str(ii)+'.gif')] if savegif else None)   
+
+    
+    if fid < best_fid : 
+        print(fid,best_fid)
+
+        msg = f"--> --> \t FID Improved from {best_fid:.5f} to {fid:.5f} !!!"
+        logger.info(msg)
+        best_fid, best_iter = fid, nb_iter
+        if save:
+            torch.save({'net' : net.state_dict()}, os.path.join(out_dir, 'net_best_fid.pth'))
+    
+    if abs(diversity_real - diversity) < abs(diversity_real - best_div) : 
+        msg = f"--> --> \t Diversity Improved from {best_div:.5f} to {diversity:.5f} !!!"
+        logger.info(msg)
+        best_div = diversity
+        if save:
+            torch.save({'net' : net.state_dict()}, os.path.join(out_dir, 'net_best_div.pth'))
+
+    if R_precision[0] > best_top1 : 
+        msg = f"--> --> \t Top1 Improved from {best_top1:.4f} to {R_precision[0]:.4f} !!!"
+        logger.info(msg)
+        best_top1 = R_precision[0]
+        if save:
+            torch.save({'net' : net.state_dict()}, os.path.join(out_dir, 'net_best_top1.pth'))
+
+    if R_precision[1] > best_top2 : 
+        msg = f"--> --> \t Top2 Improved from {best_top2:.4f} to {R_precision[1]:.4f} !!!"
+        logger.info(msg)
+        best_top2 = R_precision[1]
+    
+    if R_precision[2] > best_top3 : 
+        msg = f"--> --> \t Top3 Improved from {best_top3:.4f} to {R_precision[2]:.4f} !!!"
+        logger.info(msg)
+        best_top3 = R_precision[2]
+    
+    if matching_score_pred < best_matching : 
+        msg = f"--> --> \t matching_score Improved from {best_matching:.5f} to {matching_score_pred:.5f} !!!"
+        logger.info(msg)
+        best_matching = matching_score_pred
+        if save:
+            torch.save({'net' : net.state_dict()}, os.path.join(out_dir, 'net_best_matching.pth'))
+
+    if save:
+        torch.save({'net' : net.state_dict()}, os.path.join(out_dir, 'net_last.pth'))
+
+    net.train()
+    return best_fid, best_iter, best_div, best_top1, best_top2, best_top3, best_matching, writer, logger
+
+
+@torch.no_grad()        
+def evaluation_transformer(out_dir, val_loader, net, trans, logger, writer, nb_iter, best_fid, best_fid_syn,best_fid_perturbation,best_iter, best_div, best_top1, best_top2, best_top3, best_matching, clip_model, eval_wrapper, draw = True, save = True, savegif=False,PGD=None,crit=None) : 
+
+    trans.eval()
+    #这里是不是应该clip也eval()
+    nb_sample = 0
+    
+    draw_org = []
+    draw_pred = []
+    draw_text = []
+    draw_text_pred = []
+
+    motion_annotation_list = []
+    motion_pred_list = []
+    motion_pred_per_list = []
+    R_precision_real = 0
+    R_precision = 0
+    matching_score_real = 0
+    matching_score_pred = 0
+
+    nb_sample = 0
+    for i in range(1):
+        for batch in tqdm(val_loader):
+            word_embeddings, pos_one_hots, clip_text, clip_text_perb, sent_len, pose, m_length, token, name = batch
+
+            bs, seq = pose.shape[:2]
+            num_joints = 21 if pose.shape[-1] == 251 else 22
+            
+            text = clip.tokenize(clip_text, truncate=True).cuda()
+            text_perb = clip.tokenize(clip_text_perb, truncate=True).cuda()
+
+
+            feat_clip_text = clip_model.encode_text(text)[0].float()
+            feat_clip_text_per = clip_model.encode_text(text_perb)[0].float()
+
+            
+            pred_pose_eval = torch.zeros((bs, seq, pose.shape[-1])).cuda()
+            pred_pose_eval_per = torch.zeros((bs, seq, pose.shape[-1])).cuda()
+            pred_len = torch.ones(bs).long()
+            pred_len_per = torch.ones(bs).long()
+
+            for k in range(bs):
+                try:
+                    index_motion = trans.sample(feat_clip_text[k:k+1], False)
+                    index_motion_per = trans.sample(feat_clip_text_per[k:k+1], False)
+                except:
+                    # print('---------------------')
+                    index_motion = torch.ones(1,1).cuda().long()
+                    index_motion_per = torch.ones(1,1).cuda().long()
+
+                pred_pose = net.forward_decoder(index_motion)
+                pred_pose_per = net.forward_decoder(index_motion_per)
+                
+                cur_len = pred_pose.shape[1]
+                cur_len_per = pred_pose_per.shape[1]
+
+                pred_len[k] = min(cur_len, seq)
+                pred_len_per[k] = min(cur_len_per, seq)
+                pred_pose_eval[k:k+1, :cur_len] = pred_pose[:, :seq]
+                pred_pose_eval_per[k:k+1, :cur_len_per] = pred_pose_per[:, :seq]
+
+                if draw:
+                    pred_denorm = val_loader.dataset.inv_transform(pred_pose.detach().cpu().numpy())
+                    pred_xyz = recover_from_ric(torch.from_numpy(pred_denorm).float().cuda(), num_joints)
+
+                    if i == 0 and k < 4:
+                        draw_pred.append(pred_xyz)
+                        draw_text_pred.append(clip_text[k])
+
+            et_pred, em_pred = eval_wrapper.get_co_embeddings(word_embeddings, pos_one_hots, sent_len, pred_pose_eval, pred_len)
+            et_pred_per, em_pred_per = eval_wrapper.get_co_embeddings(word_embeddings, pos_one_hots, sent_len, pred_pose_eval_per, pred_len_per)
+            
+            if i == 0:
+                pose = pose.cuda().float()
+                
+                et, em = eval_wrapper.get_co_embeddings(word_embeddings, pos_one_hots, sent_len, pose, m_length)
+                motion_annotation_list.append(em)
+                motion_pred_list.append(em_pred)
+                motion_pred_per_list.append(em_pred_per)
+
+                if draw:
+                    pose = val_loader.dataset.inv_transform(pose.detach().cpu().numpy())
+                    pose_xyz = recover_from_ric(torch.from_numpy(pose).float().cuda(), num_joints)
+
+
+                    for j in range(min(4, bs)):
+                        draw_org.append(pose_xyz[j][:m_length[j]].unsqueeze(0))
+                        draw_text.append(clip_text[j])
+
+                temp_R, temp_match = calculate_R_precision(et.cpu().numpy(), em.cpu().numpy(), top_k=3, sum_all=True)
+                R_precision_real += temp_R
+                matching_score_real += temp_match
+                temp_R, temp_match = calculate_R_precision(et_pred.cpu().numpy(), em_pred.cpu().numpy(), top_k=3, sum_all=True)
+                R_precision += temp_R
+                matching_score_pred += temp_match
+
+                nb_sample += bs
+
+    motion_annotation_np = torch.cat(motion_annotation_list, dim=0).cpu().numpy()
+    motion_pred_np = torch.cat(motion_pred_list, dim=0).cpu().numpy()
+    motion_pred_per_np = torch.cat(motion_pred_per_list, dim=0).cpu().numpy()
+    gt_mu, gt_cov  = calculate_activation_statistics(motion_annotation_np)
+    mu, cov= calculate_activation_statistics(motion_pred_np)
+    mu_per, cov_per= calculate_activation_statistics(motion_pred_per_np)
+
+    diversity_real = calculate_diversity(motion_annotation_np, 300 if nb_sample > 300 else 100)
+    diversity = calculate_diversity(motion_pred_np, 300 if nb_sample > 300 else 100)
+
+    R_precision_real = R_precision_real / nb_sample
+    R_precision = R_precision / nb_sample
+
+    matching_score_real = matching_score_real / nb_sample
+    matching_score_pred = matching_score_pred / nb_sample
+
+
+    fid = calculate_frechet_distance(gt_mu, gt_cov, mu, cov)
+    fid_syn = calculate_frechet_distance(gt_mu,gt_cov,mu_per,cov_per)
+    fid_perturbation = calculate_frechet_distance(mu_per, cov_per, mu, cov)
+
+    msg = f"--> \t Eva. Iter {nb_iter} :, FID. {fid:.4f},FID_syn{fid_syn:.5f},FID_perturbation_and_origin.{fid_perturbation:.5f} Diversity Real. {diversity_real:.4f}, Diversity. {diversity:.4f}, R_precision_real. {R_precision_real}, R_precision. {R_precision}, matching_score_real. {matching_score_real}, matching_score_pred. {matching_score_pred}"
+    logger.info(msg)
+    
+    
+    if draw:
+        writer.add_scalar('./Test/FID', fid, nb_iter)
+        writer.add_scalar('./Test/FID_perturbation_and_origin', fid_perturbation, nb_iter)
+        writer.add_scalar('./Test/FID_syn', fid_syn, nb_iter)
+        writer.add_scalar('./Test/Diversity', diversity, nb_iter)
+        writer.add_scalar('./Test/top1', R_precision[0], nb_iter)
+        writer.add_scalar('./Test/top2', R_precision[1], nb_iter)
+        writer.add_scalar('./Test/top3', R_precision[2], nb_iter)
+        writer.add_scalar('./Test/matching_score', matching_score_pred, nb_iter)
+
+    
+        # if nb_iter % 10000 == 0 : 
+        #     for ii in range(4):
+        #         tensorborad_add_video_xyz(writer, draw_org[ii], nb_iter, tag='./Vis/org_eval'+str(ii), nb_vis=1, title_batch=[draw_text[ii]], outname=[os.path.join(out_dir, 'gt'+str(ii)+'.gif')] if savegif else None)
+            
+        # if nb_iter % 10000 == 0 : 
+        #     for ii in range(4):
+        #         tensorborad_add_video_xyz(writer, draw_pred[ii], nb_iter, tag='./Vis/pred_eval'+str(ii), nb_vis=1, title_batch=[draw_text_pred[ii]], outname=[os.path.join(out_dir, 'pred'+str(ii)+'.gif')] if savegif else None)
+
+    if isinstance(best_fid, tuple):
+        best_fid=best_fid[0]
+    if isinstance(best_fid_perturbation, tuple):
+        best_fid_perturbation=best_fid_perturbation[0]
+    if fid < best_fid : 
+        msg = f"--> --> \t FID Improved from {best_fid:.5f} to {fid:.5f} !!!"
+        logger.info(msg)
+        best_fid, best_iter = fid, nb_iter
+        if save:
+            state_dict = clip_model.state_dict()
+            torch.save(state_dict, os.path.join(out_dir, 'clip_best.pth'))
+            torch.save({'trans' : trans.state_dict()}, os.path.join(out_dir, 'net_best_fid.pth'))
+            msg = f"--> --> \t Current FID is {fid:.5f} !!!"
+            logger.info(msg)
+    if fid_syn < best_fid_syn:
+        msg = f"--> --> \t FID_syn {best_fid_syn:.5f} to {fid_syn:.5f} !!!"
+        logger.info(msg)
+        best_fid_syn = fid_syn
+
+    if fid_perturbation < best_fid_perturbation : 
+        msg = f"--> --> \t FID_perturbation_and_origin {best_fid_perturbation:.5f} to {fid_perturbation:.5f} !!!"
+        logger.info(msg)
+        best_fid_perturbation = fid_perturbation
+       
+    if matching_score_pred < best_matching : 
+        msg = f"--> --> \t matching_score Improved from {best_matching:.5f} to {matching_score_pred:.5f} !!!"
+        logger.info(msg)
+        best_matching = matching_score_pred
+
+    if abs(diversity_real - diversity) < abs(diversity_real - best_div) : 
+        msg = f"--> --> \t Diversity Improved from {best_div:.5f} to {diversity:.5f} !!!"
+        logger.info(msg)
+        best_div = diversity
+
+    if R_precision[0] > best_top1 : 
+        msg = f"--> --> \t Top1 Improved from {best_top1:.4f} to {R_precision[0]:.4f} !!!"
+        logger.info(msg)
+        best_top1 = R_precision[0]
+
+    if R_precision[1] > best_top2 : 
+        msg = f"--> --> \t Top2 Improved from {best_top2:.4f} to {R_precision[1]:.4f} !!!"
+        logger.info(msg)
+        best_top2 = R_precision[1]
+    
+    if R_precision[2] > best_top3 : 
+        msg = f"--> --> \t Top3 Improved from {best_top3:.4f} to {R_precision[2]:.4f} !!!"
+        logger.info(msg)
+        best_top3 = R_precision[2]
+
+    if save:
+        state_dict = clip_model.state_dict()
+        torch.save(state_dict, os.path.join(out_dir, 'clip_last.pth'))
+        torch.save({'trans' : trans.state_dict()}, os.path.join(out_dir, 'net_last.pth'))
+
+    trans.train()
+    return best_fid, best_fid_syn, best_fid_perturbation, best_iter, best_div, best_top1, best_top2, best_top3, best_matching, writer, logger
+
+
+@torch.no_grad()        
+def evaluation_transformer_test(out_dir, val_loader, net, trans, logger, writer, nb_iter, best_fid,best_fid_word_perb,best_fid_perturbation, best_iter, best_div, best_top1, best_top2, best_top3, best_matching, best_multi, clip_model, eval_wrapper, draw = True, save = True, savegif=False, savenpy=False) : 
+
+    trans.eval()
+    nb_sample = 0
+    
+    draw_org = []
+    draw_pred = []
+    draw_text = []
+    draw_text_pred = []
+    draw_name = []
+
+    motion_annotation_list = []
+    motion_pred_list = []
+    motion_pred_per_list = []
+
+    motion_multimodality = []
+    R_precision_real = 0
+    R_precision = 0
+    matching_score_real = 0
+    matching_score_pred = 0
+
+    nb_sample = 0
+    
+    for batch in tqdm(val_loader, desc="Validation Progress"):
+
+        word_embeddings, pos_one_hots, clip_text, clip_text_perb, sent_len, pose, m_length, token, name = batch
+        bs, seq = pose.shape[:2]
+        num_joints = 21 if pose.shape[-1] == 251 else 22
+        
+        text = clip.tokenize(clip_text, truncate=True).cuda()
+        text_perb = clip.tokenize(clip_text_perb, truncate=True).cuda()
+        feat_clip_text = clip_model.encode_text(text)[0].float()
+        feat_clip_text_per = clip_model.encode_text(text_perb)[0].float()
+        
+        
+        motion_multimodality_batch = []
+        for i in range(1):
+            pred_pose_eval = torch.zeros((bs, seq, pose.shape[-1])).cuda()
+            pred_pose_eval_per = torch.zeros((bs, seq, pose.shape[-1])).cuda()
+            
+            pred_len = torch.ones(bs).long()
+            pred_len_per = torch.ones(bs).long()
+            
+            for k in range(bs):
+                try:
+                    index_motion = trans.sample(feat_clip_text[k:k+1], True)
+                    index_motion_per = trans.sample(feat_clip_text_per[k:k+1], True)
+                except:
+                    index_motion = torch.ones(1,1).cuda().long()
+                    index_motion_per = torch.ones(1,1).cuda().long()
+
+                pred_pose = net.forward_decoder(index_motion)
+                pred_pose_per = net.forward_decoder(index_motion_per)
+                cur_len = pred_pose.shape[1]
+                cur_len_per = pred_pose_per.shape[1]
+                
+                pred_len[k] = min(cur_len, seq)
+                pred_len_per[k] = min(cur_len_per, seq)
+                
+                pred_pose_eval[k:k+1, :cur_len] = pred_pose[:, :seq]
+                pred_pose_eval_per[k:k+1, :cur_len_per] = pred_pose_per[:, :seq]
+
+                if i == 0 and (draw or savenpy):
+                    pred_denorm = val_loader.dataset.inv_transform(pred_pose.detach().cpu().numpy())
+                    pred_xyz = recover_from_ric(torch.from_numpy(pred_denorm).float().cuda(), num_joints)
+
+                    if savenpy:
+                        np.save(os.path.join(out_dir, name[k]+'_pred.npy'), pred_xyz.detach().cpu().numpy())
+
+                    if draw:
+                        if i == 0:
+                            draw_pred.append(pred_xyz)
+                            draw_text_pred.append(clip_text[k])
+                            draw_name.append(name[k])
+
+            et_pred, em_pred = eval_wrapper.get_co_embeddings(word_embeddings, pos_one_hots, sent_len, pred_pose_eval, pred_len)
+            et_pred_per, em_pred_per = eval_wrapper.get_co_embeddings(word_embeddings, pos_one_hots, sent_len, pred_pose_eval_per, pred_len_per)
+
+            # motion_multimodality_batch.append(em_pred.reshape(bs, 1, -1))
+            
+            if i == 0:
+                pose = pose.cuda().float()
+                
+                et, em = eval_wrapper.get_co_embeddings(word_embeddings, pos_one_hots, sent_len, pose, m_length)
+                motion_annotation_list.append(em)
+                motion_pred_list.append(em_pred)
+                motion_pred_per_list.append(em_pred_per)
+
+                if draw or savenpy:
+                    pose = val_loader.dataset.inv_transform(pose.detach().cpu().numpy())
+                    pose_xyz = recover_from_ric(torch.from_numpy(pose).float().cuda(), num_joints)
+
+                    if savenpy:
+                        for j in range(bs):
+                            np.save(os.path.join(out_dir, name[j]+'_gt.npy'), pose_xyz[j][:m_length[j]].unsqueeze(0).cpu().numpy())
+
+                    if draw:
+                        for j in range(bs):
+                            draw_org.append(pose_xyz[j][:m_length[j]].unsqueeze(0))
+                            draw_text.append(clip_text[j])
+
+                temp_R, temp_match = calculate_R_precision(et.cpu().numpy(), em.cpu().numpy(), top_k=3, sum_all=True)
+                R_precision_real += temp_R
+                matching_score_real += temp_match
+                temp_R, temp_match = calculate_R_precision(et_pred.cpu().numpy(), em_pred.cpu().numpy(), top_k=3, sum_all=True)
+                R_precision += temp_R
+                matching_score_pred += temp_match
+
+                nb_sample += bs
+
+        # motion_multimodality.append(torch.cat(motion_multimodality_batch, dim=1))
+
+    motion_annotation_np = torch.cat(motion_annotation_list, dim=0).cpu().numpy()
+    motion_pred_np = torch.cat(motion_pred_list, dim=0).cpu().numpy()
+    motion_pred_per_np = torch.cat(motion_pred_per_list, dim=0).cpu().numpy()
+    gt_mu, gt_cov  = calculate_activation_statistics(motion_annotation_np)
+    mu, cov= calculate_activation_statistics(motion_pred_np)                # mu cov使用的是motion_perb_np
+    mu_per, cov_per= calculate_activation_statistics(motion_pred_per_np)
+    gt_mu[np.isnan(gt_mu) | np.isinf(gt_mu)] = 0.0
+    gt_cov[np.isnan(gt_cov) | np.isinf(gt_cov)] = 0.0
+    mu[np.isnan(mu) | np.isinf(mu)] = 0.0
+    cov[np.isnan(cov) | np.isinf(cov)] = 0.0
+    mu_per[np.isnan(mu_per) | np.isinf(mu_per)] = 0.0
+    cov_per[np.isnan(cov_per) | np.isinf(cov_per)] = 0.0
+    diversity_real = calculate_diversity(motion_annotation_np, 300 if nb_sample > 300 else 100)
+    diversity = calculate_diversity(motion_pred_np, 300 if nb_sample > 300 else 100)
+
+    R_precision_real = R_precision_real / nb_sample
+    R_precision = R_precision / nb_sample
+
+    matching_score_real = matching_score_real / nb_sample
+    matching_score_pred = matching_score_pred / nb_sample
+
+    multimodality = 0
+    # motion_multimodality = torch.cat(motion_multimodality, dim=0).cpu().numpy()
+    # multimodality = calculate_multimodality(motion_multimodality, 10)
+    try:
+        fid = calculate_frechet_distance(gt_mu, gt_cov, mu, cov)
+        fid_perturbation = calculate_frechet_distance(mu_per, cov_per, mu, cov)
+        fid_word_perb = calculate_frechet_distance(gt_mu,gt_cov,mu_per,cov_per)
+    except:
+        print('数据有问题！！')
+    msg = f"--> \t Eva. Iter {nb_iter} :, FID. {fid:.4f}, FID_syn. {fid_word_perb:.5f}, FID_Perturbation. {fid_perturbation:.4f}, Diversity Real. {diversity_real:.4f}, Diversity. {diversity:.4f}, R_precision_real. {R_precision_real}, R_precision. {R_precision}, matching_score_real. {matching_score_real}, matching_score_pred. {matching_score_pred}, multimodality. {multimodality:.4f}"
+    logger.info(msg)
+    
+    
+    if draw:
+        for ii in range(len(draw_org)):
+            tensorborad_add_video_xyz(writer, draw_org[ii], nb_iter, tag='./Vis/'+draw_name[ii]+'_org', nb_vis=1, title_batch=[draw_text[ii]], outname=[os.path.join(out_dir, draw_name[ii]+'_skel_gt.gif')] if savegif else None)
+        
+            tensorborad_add_video_xyz(writer, draw_pred[ii], nb_iter, tag='./Vis/'+draw_name[ii]+'_pred', nb_vis=1, title_batch=[draw_text_pred[ii]], outname=[os.path.join(out_dir, draw_name[ii]+'_skel_pred.gif')] if savegif else None)
+
+    trans.train()
+    return fid,fid_word_perb,fid_perturbation, best_iter, diversity, R_precision[0], R_precision[1], R_precision[2], matching_score_pred, multimodality, writer, logger
+
+# (X - X_train)*(X - X_train) = -2X*X_train + X*X + X_train*X_train
+def euclidean_distance_matrix(matrix1, matrix2):
+    """
+        Params:
+        -- matrix1: N1 x D
+        -- matrix2: N2 x D
+        Returns:
+        -- dist: N1 x N2
+        dist[i, j] == distance(matrix1[i], matrix2[j])
+    """
+    assert matrix1.shape[1] == matrix2.shape[1]
+    d1 = -2 * np.dot(matrix1, matrix2.T)    # shape (num_test, num_train)
+    d2 = np.sum(np.square(matrix1), axis=1, keepdims=True)    # shape (num_test, 1)
+    d3 = np.sum(np.square(matrix2), axis=1)     # shape (num_train, )
+    dists = np.sqrt(d1 + d2 + d3)  # broadcasting
+    return dists
+
+
+
+def calculate_top_k(mat, top_k):
+    size = mat.shape[0]
+    gt_mat = np.expand_dims(np.arange(size), 1).repeat(size, 1)
+    bool_mat = (mat == gt_mat)
+    correct_vec = False
+    top_k_list = []
+    for i in range(top_k):
+#         print(correct_vec, bool_mat[:, i])
+        correct_vec = (correct_vec | bool_mat[:, i])
+        # print(correct_vec)
+        top_k_list.append(correct_vec[:, None])
+    top_k_mat = np.concatenate(top_k_list, axis=1)
+    return top_k_mat
+
+
+def calculate_R_precision(embedding1, embedding2, top_k, sum_all=False):
+    dist_mat = euclidean_distance_matrix(embedding1, embedding2)
+    matching_score = dist_mat.trace()
+    argmax = np.argsort(dist_mat, axis=1)
+    top_k_mat = calculate_top_k(argmax, top_k)
+    if sum_all:
+        return top_k_mat.sum(axis=0), matching_score
+    else:
+        return top_k_mat, matching_score
+
+def calculate_multimodality(activation, multimodality_times):
+    assert len(activation.shape) == 3
+    assert activation.shape[1] > multimodality_times
+    num_per_sent = activation.shape[1]
+
+    first_dices = np.random.choice(num_per_sent, multimodality_times, replace=False)
+    second_dices = np.random.choice(num_per_sent, multimodality_times, replace=False)
+    dist = linalg.norm(activation[:, first_dices] - activation[:, second_dices], axis=2)
+    return dist.mean()
+
+
+def calculate_diversity(activation, diversity_times):
+    assert len(activation.shape) == 2
+    assert activation.shape[0] > diversity_times
+    num_samples = activation.shape[0]
+
+    first_indices = np.random.choice(num_samples, diversity_times, replace=False)
+    second_indices = np.random.choice(num_samples, diversity_times, replace=False)
+    dist = linalg.norm(activation[first_indices] - activation[second_indices], axis=1)
+    return dist.mean()
+
+
+
+def calculate_frechet_distance(mu1, sigma1, mu2, sigma2, eps=1e-6):
+
+    mu1 = np.atleast_1d(mu1)
+    mu2 = np.atleast_1d(mu2)
+
+    sigma1 = np.atleast_2d(sigma1)
+    sigma2 = np.atleast_2d(sigma2)
+
+    assert mu1.shape == mu2.shape, \
+        'Training and test mean vectors have different lengths'
+    assert sigma1.shape == sigma2.shape, \
+        'Training and test covariances have different dimensions'
+
+    diff = mu1 - mu2
+
+    # Product might be almost singular
+    covmean, _ = linalg.sqrtm(sigma1.dot(sigma2), disp=False)
+    if not np.isfinite(covmean).all():
+        msg = ('fid calculation produces singular product; '
+               'adding %s to diagonal of cov estimates') % eps
+        print(msg)
+        offset = np.eye(sigma1.shape[0]) * eps
+        covmean = linalg.sqrtm((sigma1 + offset).dot(sigma2 + offset))
+
+    # Numerical error might give slight imaginary component
+    if np.iscomplexobj(covmean):
+        if not np.allclose(np.diagonal(covmean).imag, 0, atol=1e-3):
+            m = np.max(np.abs(covmean.imag))
+            raise ValueError('Imaginary component {}'.format(m))
+        covmean = covmean.real
+
+    tr_covmean = np.trace(covmean)
+
+    return (diff.dot(diff) + np.trace(sigma1)
+            + np.trace(sigma2) - 2 * tr_covmean)
+
+
+
+def calculate_activation_statistics(activations):
+
+    mu = np.mean(activations, axis=0)
+    cov = np.cov(activations, rowvar=False)
+    return mu, cov
+
+
+def calculate_frechet_feature_distance(feature_list1, feature_list2):
+    feature_list1 = np.stack(feature_list1)
+    feature_list2 = np.stack(feature_list2)
+
+    # normalize the scale
+    mean = np.mean(feature_list1, axis=0)
+    std = np.std(feature_list1, axis=0) + 1e-10
+    feature_list1 = (feature_list1 - mean) / std
+    feature_list2 = (feature_list2 - mean) / std
+
+    dist = calculate_frechet_distance(
+        mu1=np.mean(feature_list1, axis=0), 
+        sigma1=np.cov(feature_list1, rowvar=False),
+        mu2=np.mean(feature_list2, axis=0), 
+        sigma2=np.cov(feature_list2, rowvar=False),
+    )
+    return dist

losses.py

@@ -0,0 +1,102 @@
+from CLIP.clip import clip
+from CLIP.clip import model
+import torch
+
+def topk_overlap_loss(gt, pred, K=2, metric='l1'):
+    idx = torch.argsort(gt, descending=True)
+    # print(idx)
+    idx = idx[:K]
+    pred_TopK_1 = pred.gather(-1,idx)
+    gt_Topk_1 = gt.gather(-1,idx)
+
+    idx_pred = torch.argsort(pred, descending=True)
+    idx_pred = idx_pred[:K]
+    try:
+        gt_TopK_2 = gt.gather(-1, idx_pred)
+    except Exception as e:
+        print(e)
+        print(gt.shape)
+        print(idx_pred.shape)
+    pred_TopK_2 = pred.gather(-1, idx_pred)
+
+    gt_Topk_1_normed = torch.nn.functional.softmax(gt_Topk_1, dim=-1)
+    pred_TopK_1_normed = torch.nn.functional.softmax(pred_TopK_1, dim=-1)
+    gt_TopK_2_normed = torch.nn.functional.softmax(gt_TopK_2, dim=-1)
+    pred_TopK_2_normed = torch.nn.functional.softmax(pred_TopK_2, dim=-1)
+
+    def kl(a,b):
+        return torch.nn.functional.kl_div(a.log(), b, reduction="batchmean")
+
+    def jsd(a,b):
+        loss = kl(a,b) + kl(b,a)
+        loss /= 2
+        return loss
+
+
+    if metric == 'l1':
+        loss = torch.abs((pred_TopK_1 - gt_Topk_1)) + torch.abs(gt_TopK_2 - pred_TopK_2)
+        loss = loss/(2*K)
+    elif metric == "l2":
+        loss = torch.norm(pred_TopK_1 - gt_Topk_1, p=2) + torch.norm(gt_TopK_2 - pred_TopK_2, p=2)
+        loss = loss/(2*K)
+    elif metric == "kl-full":
+        loss = kl(gt,pred)
+    elif metric == "jsd-full":
+        loss = jsd(gt,pred)
+    elif metric == "kl-topk":
+        loss = kl(gt_Topk_1_normed,pred_TopK_1_normed) + kl(gt_TopK_2_normed,pred_TopK_2_normed)
+        loss /=2
+    elif metric == "jsd-topk":
+        loss = jsd(gt_Topk_1_normed, pred_TopK_1_normed) + jsd(gt_TopK_2_normed, pred_TopK_2_normed)
+        loss /= 2
+    return loss
+
+def topk_overlap_loss_batch(gt,pred,K=2,metric='l1'):
+    idx = torch.argsort(gt,dim=1,descending=True)
+    # print(idx)
+    idx = idx[:,:K]
+    pred_TopK_1 = pred.gather(1,idx)
+    gt_Topk_1 = gt.gather(1,idx)
+
+    idx_pred = torch.argsort(pred,dim=1,descending=True)
+    idx_pred = idx_pred[:,:K]
+    try:
+        gt_TopK_2 = gt.gather(1, idx_pred)
+    except Exception as e:
+        print(e)
+        print(gt.shape)
+        print(idx_pred.shape)
+    pred_TopK_2 = pred.gather(1, idx_pred)
+
+    gt_Topk_1_normed = torch.nn.functional.softmax(gt_Topk_1,dim=-1)
+    pred_TopK_1_normed = torch.nn.functional.softmax(pred_TopK_1,dim=-1)
+    gt_TopK_2_normed = torch.nn.functional.softmax(gt_TopK_2,dim=-1)
+    pred_TopK_2_normed = torch.nn.functional.softmax(pred_TopK_2,dim=-1)
+
+    def kl(a,b):
+        return torch.nn.functional.kl_div(a.log(), b, reduction="batchmean")
+
+    def jsd(a,b):
+        loss = kl(a,b) + kl(b,a)
+        loss /= 2
+        return loss
+
+
+    if metric == 'l1':
+        loss = torch.abs((pred_TopK_1 - gt_Topk_1)) + torch.abs(gt_TopK_2 - pred_TopK_2)
+        loss = loss/(2*K)
+    elif metric == "l2":
+        loss = torch.norm(pred_TopK_1 - gt_Topk_1, p=2) + torch.norm(gt_TopK_2 - pred_TopK_2, p=2)
+        loss = loss/(2*K)
+    elif metric == "kl-full":
+        loss = kl(gt,pred)
+    elif metric == "jsd-full":
+        loss = jsd(gt,pred)
+    elif metric == "kl-topk":
+        loss = kl(gt_Topk_1_normed,pred_TopK_1_normed) + kl(gt_TopK_2_normed,pred_TopK_2_normed)
+        loss /=2
+    elif metric == "jsd-topk":
+        loss = jsd(gt_Topk_1_normed, pred_TopK_1_normed) + jsd(gt_TopK_2_normed, pred_TopK_2_normed)
+        loss /= 2
+    return loss
+

metrics.py

@@ -0,0 +1,192 @@
+import torch
+import numpy as np
+import json
+import os
+import shutil
+from copy import deepcopy
+import torch
+import torch.nn as nn
+
+from sklearn.utils import shuffle
+# from tqdm import tqdm
+import time
+
+def tvd(predictions, targets): #accepts two numpy arrays of dimension: (num. instances, )
+    return (0.5 * np.abs(predictions - targets)).sum()
+
+def batch_tvd(predictions, targets,reduce=True): #accepts two Torch tensors... " "
+    if reduce == False:
+        return (0.5 * torch.abs(predictions - targets))
+    else:
+        return (0.5 * torch.abs(predictions - targets)).sum()
+def get_sorting_index_with_noise_from_lengths(lengths, noise_frac):
+    if noise_frac > 0:
+        noisy_lengths = [x + np.random.randint(np.floor(-x * noise_frac), np.ceil(x * noise_frac)) for x in lengths]
+    else:
+        noisy_lengths = lengths
+    return np.argsort(noisy_lengths)
+
+
+device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+
+
+def kld(a1, a2):
+    # (B, *, A), #(B, *, A)
+    a1 = torch.clamp(a1, 0, 1)
+    a2 = torch.clamp(a2, 0, 1)
+    log_a1 = torch.log(a1 + 1e-10)
+    log_a2 = torch.log(a2 + 1e-10)
+
+    kld = a1 * (log_a1 - log_a2)
+    kld = kld.sum(-1)
+
+    return kld
+
+
+def jsd(p, q):
+    m = 0.5 * (p + q)
+    jsd = 0.5 * (kld(p, m) + kld(q, m))  # for each instance in the batch
+
+    return jsd.unsqueeze(-1)  # jsd.squeeze(1).sum()
+
+
+def tvd(predictions, targets): #accepts two numpy arrays of dimension: (num. instances, )
+    return (0.5 * np.abs(predictions - targets)).sum()
+
+def batch_tvd(predictions, targets): #accepts two Torch tensors... " "
+    return (0.5 * torch.abs(predictions - targets)).sum()
+
+
+
+def batch_jaccard_similarity(gt, pred):
+    intersection = torch.min(gt, pred).sum(dim=1)  
+    union = torch.max(gt, pred).sum(dim=1)  
+    similarity = intersection / union 
+    return similarity
+
+def jaccard_similarity(gt, pred, top_k=2):
+    
+    gt_top_k = torch.topk(gt, top_k, dim=1).values
+    pred_top_k = torch.topk(pred, top_k, dim=1).values
+    
+    
+    jaccard_sim = batch_jaccard_similarity(gt_top_k, pred_top_k)
+    
+    
+    mean_similarity = jaccard_sim.mean()
+    
+    return mean_similarity
+
+
+
+def intersection_of_two_tensor(t1, t2):
+    combined = torch.cat((t1, t2))
+    uniques, counts = combined.unique(return_counts=True)
+    intersection = uniques[counts > 1]
+    return intersection
+
+def topK_overlap_true_loss(a,b,K=2):
+    t1 = torch.argsort(a, descending=True)
+    t2 = torch.argsort(b, descending=True)
+    t1 = t1.detach().cpu().numpy()
+    t2 = t2.detach().cpu().numpy()
+    N = t1.shape[0]
+    loss = []
+    for i in range(N):
+        inset = np.intersect1d(t1[i,:K],t2[i,:K])
+        overlap = len(inset)/K
+        # print(overlap)
+        loss.append(overlap)
+    return np.mean(loss)
+
+
+class AverageMeter():
+    def __init__(self):
+        self.cnt = 0
+        self.sum = 0
+        self.mean = 0
+
+    def update(self, val, cnt):
+        self.cnt += cnt
+        self.sum += val * cnt
+        self.mean = self.sum / self.cnt
+
+    def average(self):
+        return self.mean
+
+    def total(self):
+        return self.sum
+
+
+
+def topk_overlap_loss(gt,pred,K=2,metric='l1'):
+    idx = torch.argsort(gt,dim=1,descending=True)
+    # print(idx)
+    idx = idx[:,:K]
+    pred_TopK_1 = pred.gather(1,idx)
+    gt_Topk_1 = gt.gather(1,idx)
+
+    idx_pred = torch.argsort(pred,dim=1,descending=True)
+    idx_pred = idx_pred[:,:K]
+    try:
+        gt_TopK_2 = gt.gather(1, idx_pred)
+    except Exception as e:
+        print(e)
+        print(gt.shape)
+        print(idx_pred.shape)
+    pred_TopK_2 = pred.gather(1, idx_pred)
+
+    gt_Topk_1_normed = torch.nn.functional.softmax(gt_Topk_1,dim=-1)
+    pred_TopK_1_normed = torch.nn.functional.softmax(pred_TopK_1,dim=-1)
+    gt_TopK_2_normed = torch.nn.functional.softmax(gt_TopK_2,dim=-1)
+    pred_TopK_2_normed = torch.nn.functional.softmax(pred_TopK_2,dim=-1)
+
+    def kl(a,b):
+        return torch.nn.functional.kl_div(a.log(), b, reduction="batchmean")
+
+    def jsd(a,b):
+        loss = kl(a,b) + kl(b,a)
+        loss /= 2
+        return loss
+
+
+    if metric == 'l1':
+        loss = torch.abs((pred_TopK_1 - gt_Topk_1)) + torch.abs(gt_TopK_2 - pred_TopK_2)
+        loss = loss/(2*K)
+    elif metric == "l2":
+        loss = torch.norm(pred_TopK_1 - gt_Topk_1, p=2) + torch.norm(gt_TopK_2 - pred_TopK_2, p=2)
+        loss = loss/(2*K)
+    elif metric == "kl-full":
+        loss = kl(gt,pred)
+    elif metric == "jsd-full":
+        loss = jsd(gt,pred)
+    elif metric == "kl-topk":
+        loss = kl(gt_Topk_1_normed,pred_TopK_1_normed) + kl(gt_TopK_2_normed,pred_TopK_2_normed)
+        loss /=2
+    elif metric == "jsd-topk":
+        loss = jsd(gt_Topk_1_normed, pred_TopK_1_normed) + jsd(gt_TopK_2_normed, pred_TopK_2_normed)
+        loss /= 2
+    return loss
+
+if __name__ == '__main__':
+
+    from torch.autograd import gradcheck
+    import torch
+    import torch.nn as nn
+
+    # intersection_of_two_tensor(t1[i], t2[i])
+
+    t1 = torch.tensor(
+        np.array([[100, 2, 3, 4],
+                  [2, 1, 3, 7]],),requires_grad=True, dtype=torch.double
+    )
+    print(t1.shape)
+    t2 = torch.tensor(
+        np.array([[1, 2, 3, 4],
+                  [2, 4, 6, 7]]),requires_grad=True, dtype=torch.double
+    )
+    print(t2.shape)
+
+
+
+    print(topK_overlap_true_loss(torch.argsort(t1,descending=True),torch.argsort(t2,descending=True),K=2))

render_final.py

@@ -0,0 +1,198 @@
+from models.rotation2xyz import Rotation2xyz
+import numpy as np
+from trimesh import Trimesh
+import os
+os.environ['PYOPENGL_PLATFORM'] = "osmesa"
+
+import torch
+from visualize.simplify_loc2rot import joints2smpl
+import pyrender
+import matplotlib.pyplot as plt
+
+import io
+import imageio
+from shapely import geometry
+import trimesh
+from pyrender.constants import RenderFlags
+import math
+# import ffmpeg
+from PIL import Image
+
+class WeakPerspectiveCamera(pyrender.Camera):
+    def __init__(self,
+                 scale,
+                 translation,
+                 znear=pyrender.camera.DEFAULT_Z_NEAR,
+                 zfar=None,
+                 name=None):
+        super(WeakPerspectiveCamera, self).__init__(
+            znear=znear,
+            zfar=zfar,
+            name=name,
+        )
+        self.scale = scale
+        self.translation = translation
+
+    def get_projection_matrix(self, width=None, height=None):
+        P = np.eye(4)
+        P[0, 0] = self.scale[0]
+        P[1, 1] = self.scale[1]
+        P[0, 3] = self.translation[0] * self.scale[0]
+        P[1, 3] = -self.translation[1] * self.scale[1]
+        P[2, 2] = -1
+        return P
+
+def render(motions, outdir='test_vis', device_id=0, name=None, pred=True):
+    frames, njoints, nfeats = motions.shape
+    MINS = motions.min(axis=0).min(axis=0)
+    MAXS = motions.max(axis=0).max(axis=0)
+
+    height_offset = MINS[1]
+    motions[:, :, 1] -= height_offset
+    trajec = motions[:, 0, [0, 2]]
+
+    j2s = joints2smpl(num_frames=frames, device_id=0, cuda=True)
+    rot2xyz = Rotation2xyz(device=torch.device("cuda:0"))
+    faces = rot2xyz.smpl_model.faces
+
+    if (not os.path.exists(outdir + name+'_pred.pt') and pred) or (not os.path.exists(outdir + name+'_gt.pt') and not pred): 
+        print(f'Running SMPLify, it may take a few minutes.')
+        motion_tensor, opt_dict = j2s.joint2smpl(motions)  # [nframes, njoints, 3]
+
+        vertices = rot2xyz(torch.tensor(motion_tensor).clone(), mask=None,
+                                        pose_rep='rot6d', translation=True, glob=True,
+                                        jointstype='vertices',
+                                        vertstrans=True)
+
+        if pred:
+            torch.save(vertices, outdir + name+'_pred.pt')
+        else:
+            torch.save(vertices, outdir + name+'_gt.pt')
+    else:
+        if pred:
+            vertices = torch.load(outdir + name+'_pred.pt')
+        else:
+            vertices = torch.load(outdir + name+'_gt.pt')
+    frames = vertices.shape[3] # shape: 1, nb_frames, 3, nb_joints
+    print (vertices.shape)
+    MINS = torch.min(torch.min(vertices[0], axis=0)[0], axis=1)[0]
+    MAXS = torch.max(torch.max(vertices[0], axis=0)[0], axis=1)[0]
+    # vertices[:,:,1,:] -= MINS[1] + 1e-5
+
+
+    out_list = []
+    
+    minx = MINS[0] - 0.5
+    maxx = MAXS[0] + 0.5
+    minz = MINS[2] - 0.5 
+    maxz = MAXS[2] + 0.5
+    polygon = geometry.Polygon([[minx, minz], [minx, maxz], [maxx, maxz], [maxx, minz]])
+    polygon_mesh = trimesh.creation.extrude_polygon(polygon, 1e-5)
+
+    vid = []
+    for i in range(frames):
+        if i % 10 == 0:
+            print(i)
+
+        mesh = Trimesh(vertices=vertices[0, :, :, i].squeeze().tolist(), faces=faces)
+
+        base_color = (0.11, 0.53, 0.8, 0.5)
+        ## OPAQUE rendering without alpha
+        ## BLEND rendering consider alpha 
+        material = pyrender.MetallicRoughnessMaterial(
+            metallicFactor=0.7,
+            alphaMode='OPAQUE',
+            baseColorFactor=base_color
+        )
+
+
+        mesh = pyrender.Mesh.from_trimesh(mesh, material=material)
+
+        polygon_mesh.visual.face_colors = [0, 0, 0, 0.21]
+        polygon_render = pyrender.Mesh.from_trimesh(polygon_mesh, smooth=False)
+
+        bg_color = [1, 1, 1, 0.8]
+        scene = pyrender.Scene(bg_color=bg_color, ambient_light=(0.4, 0.4, 0.4))
+        
+        sx, sy, tx, ty = [0.75, 0.75, 0, 0.10]
+
+        camera = pyrender.PerspectiveCamera(yfov=(np.pi / 3.0))
+
+        light = pyrender.DirectionalLight(color=[1,1,1], intensity=300)
+
+        scene.add(mesh)
+
+        c = np.pi / 2
+
+        scene.add(polygon_render, pose=np.array([[ 1, 0, 0, 0],
+
+        [ 0, np.cos(c), -np.sin(c), MINS[1].cpu().numpy()],
+
+        [ 0, np.sin(c), np.cos(c), 0],
+
+        [ 0, 0, 0, 1]]))
+
+        light_pose = np.eye(4)
+        light_pose[:3, 3] = [0, -1, 1]
+        scene.add(light, pose=light_pose.copy())
+
+        light_pose[:3, 3] = [0, 1, 1]
+        scene.add(light, pose=light_pose.copy())
+
+        light_pose[:3, 3] = [1, 1, 2]
+        scene.add(light, pose=light_pose.copy())
+
+
+        c = -np.pi / 6
+
+        scene.add(camera, pose=[[ 1, 0, 0, (minx+maxx).cpu().numpy()/2],
+
+                                [ 0, np.cos(c), -np.sin(c), 1.5],
+
+                                [ 0, np.sin(c), np.cos(c), max(4, minz.cpu().numpy()+(1.5-MINS[1].cpu().numpy())*2, (maxx-minx).cpu().numpy())],
+
+                                [ 0, 0, 0, 1]
+                                ])
+        
+        # render scene
+        r = pyrender.OffscreenRenderer(960, 960)
+
+        color, _ = r.render(scene, flags=RenderFlags.RGBA)
+        # Image.fromarray(color).save(outdir+name+'_'+str(i)+'.png')
+
+        vid.append(color)
+
+        r.delete()
+
+    out = np.stack(vid, axis=0)
+    if pred:
+        imageio.mimsave(outdir + name+'_pred.gif', out, fps=20)
+    else:
+        imageio.mimsave(outdir + name+'_gt.gif', out, fps=20)
+    
+
+
+
+
+if __name__ == "__main__":
+    import argparse
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--filedir", type=str, default='/CV/xhr/xhr_project/Paper/text2Pose/t2m/T2M-GPT-main/visualization/pose_np', help='motion npy file dir')
+    parser.add_argument('--motion-list', default=None, nargs="1", type=str, help="motion name list")
+    args = parser.parse_args()
+
+    filename_list = args.motion_list
+    filedir = args.filedir
+    
+    for filename in filename_list:
+        motions = np.load(filedir + filename+'.npy')
+        print('pred', motions.shape, filename)
+        render(motions[0], outdir=filedir, device_id=0, name=filename, pred=True)
+
+        # motions = np.load(filedir + filename+'_pred.npy')
+        # print('pred', motions.shape, filename)
+        # render(motions[0], outdir=filedir, device_id=0, name=filename, pred=True)
+
+        # motions = np.load(filedir + filename+'_gt.npy')
+        # print('gt', motions.shape, filename)
+        # render(motions[0], outdir=filedir, device_id=0, name=filename, pred=False)

requirements.txt

@@ -0,0 +1,57 @@
+absl-py==0.13.0
+backcall==0.2.0
+cachetools==4.2.2
+charset-normalizer==2.0.4
+chumpy==0.70
+cycler==0.10.0
+decorator==5.0.9
+google-auth==1.35.0
+google-auth-oauthlib==0.4.5
+grpcio==1.39.0
+idna==3.2
+imageio==2.9.0
+ipdb==0.13.9
+ipython==7.26.0
+ipython-genutils==0.2.0
+jedi==0.18.0
+joblib==1.0.1
+kiwisolver==1.3.1
+markdown==3.3.4
+matplotlib==3.4.3
+matplotlib-inline==0.1.2
+oauthlib==3.1.1
+pandas==1.3.2
+parso==0.8.2
+pexpect==4.8.0
+pickleshare==0.7.5
+prompt-toolkit==3.0.20
+protobuf==3.17.3
+ptyprocess==0.7.0
+pyasn1==0.4.8
+pyasn1-modules==0.2.8
+pygments==2.10.0
+pyparsing==2.4.7
+python-dateutil==2.8.2
+pytz==2021.1
+pyyaml==5.4.1
+requests==2.26.0
+requests-oauthlib==1.3.0
+rsa==4.7.2
+scikit-learn==0.24.2
+scipy==1.7.1
+sklearn==0.0
+smplx==0.1.28
+tensorboard==2.6.0
+tensorboard-data-server==0.6.1
+tensorboard-plugin-wit==1.8.0
+threadpoolctl==2.2.0
+toml==0.10.2
+tqdm==4.62.2
+traitlets==5.0.5
+urllib3==1.26.6
+wcwidth==0.2.5
+werkzeug==2.0.1
+# git+https://github.com/openai/CLIP.git
+# git+https://github.com/nghorbani/human_body_prior
+gdown
+moviepy

train_vq.py

@@ -0,0 +1,175 @@
+import os
+import json
+
+import torch
+import torch.optim as optim
+from torch.utils.tensorboard import SummaryWriter
+
+import models.vqvae as vqvae
+import utils.losses as losses 
+import options.option_vq as option_vq
+import utils.utils_model as utils_model
+from dataset import dataset_VQ, dataset_TM_eval
+import utils.eval_trans as eval_trans
+from options.get_eval_option import get_opt
+from models.evaluator_wrapper import EvaluatorModelWrapper
+import warnings
+warnings.filterwarnings('ignore')
+from utils.word_vectorizer import WordVectorizer
+from multiprocessing import freeze_support
+def update_lr_warm_up(optimizer, nb_iter, warm_up_iter, lr):
+
+    current_lr = lr * (nb_iter + 1) / (warm_up_iter + 1)
+    for param_group in optimizer.param_groups:
+        param_group["lr"] = current_lr
+
+    return optimizer, current_lr
+
+##### ---- Exp dirs ---- #####
+args = option_vq.get_args_parser()
+torch.manual_seed(args.seed)
+
+args.out_dir = os.path.join(args.out_dir, f'{args.exp_name}')
+os.makedirs(args.out_dir, exist_ok = True)
+
+##### ---- Logger ---- #####
+logger = utils_model.get_logger(args.out_dir)
+writer = SummaryWriter(args.out_dir)
+logger.info(json.dumps(vars(args), indent=4, sort_keys=True))
+
+
+
+w_vectorizer = WordVectorizer('./glove', 'our_vab')
+# w_vectorizer = WordVectorizer('D:\project\\faithfulpose\T2M-GPT-main\glove', 'our_vab')
+
+
+if args.dataname == 'kit' : 
+    dataset_opt_path = 'checkpoints/kit/Comp_v6_KLD005/opt.txt'
+    # dataset_opt_path = 'D:\project\\faithfulpose\T2M-GPT-main\checkpoints\kit\Comp_v6_KLD005\opt.txt' 
+    args.nb_joints = 21
+    
+else :
+    dataset_opt_path = 'checkpoints/t2m/Comp_v6_KLD005/opt.txt'
+    args.nb_joints = 22
+
+logger.info(f'Training on {args.dataname}, motions are with {args.nb_joints} joints')
+
+wrapper_opt = get_opt(dataset_opt_path, torch.device('cuda'))
+eval_wrapper = EvaluatorModelWrapper(wrapper_opt)
+
+
+##### ---- Dataloader ---- #####
+train_loader = dataset_VQ.DATALoader(args.dataname,
+                                        args.batch_size,
+                                        window_size=args.window_size,
+                                        unit_length=2**args.down_t)
+
+train_loader_iter = dataset_VQ.cycle(train_loader)
+
+val_loader = dataset_TM_eval.DATALoader(args.dataname, False,
+                                        32,
+                                        w_vectorizer,
+                                        unit_length=2**args.down_t)
+
+##### ---- Network ---- #####
+net = vqvae.HumanVQVAE(args, ## use args to define different parameters in different quantizers
+                       args.nb_code,
+                       args.code_dim,
+                       args.output_emb_width,
+                       args.down_t,
+                       args.stride_t,
+                       args.width,
+                       args.depth,
+                       args.dilation_growth_rate,
+                       args.vq_act,
+                       args.vq_norm)
+
+
+if args.resume_pth : 
+    logger.info('loading checkpoint from {}'.format(args.resume_pth))
+    ckpt = torch.load(args.resume_pth, map_location='cpu')
+    net.load_state_dict(ckpt['net'], strict=True)
+net.train()
+net.cuda()
+
+##### ---- Optimizer & Scheduler ---- #####
+optimizer = optim.AdamW(net.parameters(), lr=args.lr, betas=(0.9, 0.99), weight_decay=args.weight_decay)
+scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, milestones=args.lr_scheduler, gamma=args.gamma)
+  
+
+Loss = losses.ReConsLoss(args.recons_loss, args.nb_joints)
+
+##### ------ warm-up ------- #####
+avg_recons, avg_perplexity, avg_commit = 0., 0., 0.
+
+
+for nb_iter in range(1, args.warm_up_iter):
+    
+    optimizer, current_lr = update_lr_warm_up(optimizer, nb_iter, args.warm_up_iter, args.lr)
+    
+    gt_motion = next(train_loader_iter)
+    gt_motion = gt_motion.cuda().float() # (bs, 64, dim)
+
+    pred_motion, loss_commit, perplexity = net(gt_motion)
+    loss_motion = Loss(pred_motion, gt_motion)
+    loss_vel = Loss.forward_vel(pred_motion, gt_motion)
+    
+    loss = loss_motion + args.commit * loss_commit + args.loss_vel * loss_vel
+    
+    optimizer.zero_grad()
+    loss.backward()
+    optimizer.step()
+
+    avg_recons += loss_motion.item()
+    avg_perplexity += perplexity.item()
+    avg_commit += loss_commit.item()
+    
+    if nb_iter % args.print_iter ==  0 :
+        avg_recons /= args.print_iter
+        avg_perplexity /= args.print_iter
+        avg_commit /= args.print_iter
+        
+        logger.info(f"Warmup. Iter {nb_iter} :  lr {current_lr:.5f} \t Commit. {avg_commit:.5f} \t PPL. {avg_perplexity:.2f} \t Recons.  {avg_recons:.5f}")
+        
+        avg_recons, avg_perplexity, avg_commit = 0., 0., 0.
+
+    # ##### ---- Training ---- #####
+avg_recons, avg_perplexity, avg_commit = 0., 0., 0.
+best_fid, best_iter, best_div, best_top1, best_top2, best_top3, best_matching, writer, logger = eval_trans.evaluation_vqvae(args.out_dir, val_loader, net, logger, writer, 0, best_fid=1000, best_iter=0, best_div=100, best_top1=0, best_top2=0, best_top3=0, best_matching=100, eval_wrapper=eval_wrapper)
+
+for nb_iter in range(1, args.total_iter + 1):
+
+    gt_motion = next(train_loader_iter)
+    gt_motion = gt_motion.cuda().float() # bs, nb_joints, joints_dim, seq_len
+    
+    pred_motion, loss_commit, perplexity = net(gt_motion)
+    loss_motion = Loss(pred_motion, gt_motion)
+    loss_vel = Loss.forward_vel(pred_motion, gt_motion)
+    
+    loss = loss_motion + args.commit * loss_commit + args.loss_vel * loss_vel
+    
+    optimizer.zero_grad()
+    loss.backward()
+    optimizer.step()
+    scheduler.step()
+    
+    avg_recons += loss_motion.item()
+    avg_perplexity += perplexity.item()
+    avg_commit += loss_commit.item()
+    
+    if nb_iter % args.print_iter ==  0 :
+        avg_recons /= args.print_iter
+        avg_perplexity /= args.print_iter
+        avg_commit /= args.print_iter
+        
+        writer.add_scalar('./Train/L1', avg_recons, nb_iter)
+        writer.add_scalar('./Train/PPL', avg_perplexity, nb_iter)
+        writer.add_scalar('./Train/Commit', avg_commit, nb_iter)
+        
+        logger.info(f"Train. Iter {nb_iter} : \t Commit. {avg_commit:.5f} \t PPL. {avg_perplexity:.2f} \t Recons.  {avg_recons:.5f}")
+        
+        avg_recons, avg_perplexity, avg_commit = 0., 0., 0.,
+
+    if nb_iter % args.eval_iter==0 :
+        best_fid, best_iter, best_div, best_top1, best_top2, best_top3, best_matching, writer, logger = eval_trans.evaluation_vqvae(args.out_dir, val_loader, net, logger, writer, nb_iter, best_fid, best_iter, best_div, best_top1, best_top2, best_top3, best_matching, eval_wrapper=eval_wrapper)
+