Duplicate from haoheliu/audioldm-text-to-audio-generation

Co-authored-by: haoheliu <haoheliu@users.noreply.huggingface.co>

.gitattributes

@@ -0,0 +1,34 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,4 @@
+*.pyc
+__pycache__
+test.py
+flagged

LICENSE

@@ -0,0 +1,437 @@
+Attribution-NonCommercial-ShareAlike 4.0 International
+
+=======================================================================
+
+Creative Commons Corporation ("Creative Commons") is not a law firm and
+does not provide legal services or legal advice. Distribution of
+Creative Commons public licenses does not create a lawyer-client or
+other relationship. Creative Commons makes its licenses and related
+information available on an "as-is" basis. Creative Commons gives no
+warranties regarding its licenses, any material licensed under their
+terms and conditions, or any related information. Creative Commons
+disclaims all liability for damages resulting from their use to the
+fullest extent possible.
+
+Using Creative Commons Public Licenses
+
+Creative Commons public licenses provide a standard set of terms and
+conditions that creators and other rights holders may use to share
+original works of authorship and other material subject to copyright
+and certain other rights specified in the public license below. The
+following considerations are for informational purposes only, are not
+exhaustive, and do not form part of our licenses.
+
+     Considerations for licensors: Our public licenses are
+     intended for use by those authorized to give the public
+     permission to use material in ways otherwise restricted by
+     copyright and certain other rights. Our licenses are
+     irrevocable. Licensors should read and understand the terms
+     and conditions of the license they choose before applying it.
+     Licensors should also secure all rights necessary before
+     applying our licenses so that the public can reuse the
+     material as expected. Licensors should clearly mark any
+     material not subject to the license. This includes other CC-
+     licensed material, or material used under an exception or
+     limitation to copyright. More considerations for licensors:
+	wiki.creativecommons.org/Considerations_for_licensors
+
+     Considerations for the public: By using one of our public
+     licenses, a licensor grants the public permission to use the
+     licensed material under specified terms and conditions. If
+     the licensor's permission is not necessary for any reason--for
+     example, because of any applicable exception or limitation to
+     copyright--then that use is not regulated by the license. Our
+     licenses grant only permissions under copyright and certain
+     other rights that a licensor has authority to grant. Use of
+     the licensed material may still be restricted for other
+     reasons, including because others have copyright or other
+     rights in the material. A licensor may make special requests,
+     such as asking that all changes be marked or described.
+     Although not required by our licenses, you are encouraged to
+     respect those requests where reasonable. More_considerations
+     for the public: 
+	wiki.creativecommons.org/Considerations_for_licensees
+
+=======================================================================
+
+Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International
+Public License
+
+By exercising the Licensed Rights (defined below), You accept and agree
+to be bound by the terms and conditions of this Creative Commons
+Attribution-NonCommercial-ShareAlike 4.0 International Public License
+("Public License"). To the extent this Public License may be
+interpreted as a contract, You are granted the Licensed Rights in
+consideration of Your acceptance of these terms and conditions, and the
+Licensor grants You such rights in consideration of benefits the
+Licensor receives from making the Licensed Material available under
+these terms and conditions.
+
+
+Section 1 -- Definitions.
+
+  a. Adapted Material means material subject to Copyright and Similar
+     Rights that is derived from or based upon the Licensed Material
+     and in which the Licensed Material is translated, altered,
+     arranged, transformed, or otherwise modified in a manner requiring
+     permission under the Copyright and Similar Rights held by the
+     Licensor. For purposes of this Public License, where the Licensed
+     Material is a musical work, performance, or sound recording,
+     Adapted Material is always produced where the Licensed Material is
+     synched in timed relation with a moving image.
+
+  b. Adapter's License means the license You apply to Your Copyright
+     and Similar Rights in Your contributions to Adapted Material in
+     accordance with the terms and conditions of this Public License.
+
+  c. BY-NC-SA Compatible License means a license listed at
+     creativecommons.org/compatiblelicenses, approved by Creative
+     Commons as essentially the equivalent of this Public License.
+
+  d. Copyright and Similar Rights means copyright and/or similar rights
+     closely related to copyright including, without limitation,
+     performance, broadcast, sound recording, and Sui Generis Database
+     Rights, without regard to how the rights are labeled or
+     categorized. For purposes of this Public License, the rights
+     specified in Section 2(b)(1)-(2) are not Copyright and Similar
+     Rights.
+
+  e. Effective Technological Measures means those measures that, in the
+     absence of proper authority, may not be circumvented under laws
+     fulfilling obligations under Article 11 of the WIPO Copyright
+     Treaty adopted on December 20, 1996, and/or similar international
+     agreements.
+
+  f. Exceptions and Limitations means fair use, fair dealing, and/or
+     any other exception or limitation to Copyright and Similar Rights
+     that applies to Your use of the Licensed Material.
+
+  g. License Elements means the license attributes listed in the name
+     of a Creative Commons Public License. The License Elements of this
+     Public License are Attribution, NonCommercial, and ShareAlike.
+
+  h. Licensed Material means the artistic or literary work, database,
+     or other material to which the Licensor applied this Public
+     License.
+
+  i. Licensed Rights means the rights granted to You subject to the
+     terms and conditions of this Public License, which are limited to
+     all Copyright and Similar Rights that apply to Your use of the
+     Licensed Material and that the Licensor has authority to license.
+
+  j. Licensor means the individual(s) or entity(ies) granting rights
+     under this Public License.
+
+  k. NonCommercial means not primarily intended for or directed towards
+     commercial advantage or monetary compensation. For purposes of
+     this Public License, the exchange of the Licensed Material for
+     other material subject to Copyright and Similar Rights by digital
+     file-sharing or similar means is NonCommercial provided there is
+     no payment of monetary compensation in connection with the
+     exchange.
+
+  l. Share means to provide material to the public by any means or
+     process that requires permission under the Licensed Rights, such
+     as reproduction, public display, public performance, distribution,
+     dissemination, communication, or importation, and to make material
+     available to the public including in ways that members of the
+     public may access the material from a place and at a time
+     individually chosen by them.
+
+  m. Sui Generis Database Rights means rights other than copyright
+     resulting from Directive 96/9/EC of the European Parliament and of
+     the Council of 11 March 1996 on the legal protection of databases,
+     as amended and/or succeeded, as well as other essentially
+     equivalent rights anywhere in the world.
+
+  n. You means the individual or entity exercising the Licensed Rights
+     under this Public License. Your has a corresponding meaning.
+
+
+Section 2 -- Scope.
+
+  a. License grant.
+
+       1. Subject to the terms and conditions of this Public License,
+          the Licensor hereby grants You a worldwide, royalty-free,
+          non-sublicensable, non-exclusive, irrevocable license to
+          exercise the Licensed Rights in the Licensed Material to:
+
+            a. reproduce and Share the Licensed Material, in whole or
+               in part, for NonCommercial purposes only; and
+
+            b. produce, reproduce, and Share Adapted Material for
+               NonCommercial purposes only.
+
+       2. Exceptions and Limitations. For the avoidance of doubt, where
+          Exceptions and Limitations apply to Your use, this Public
+          License does not apply, and You do not need to comply with
+          its terms and conditions.
+
+       3. Term. The term of this Public License is specified in Section
+          6(a).
+
+       4. Media and formats; technical modifications allowed. The
+          Licensor authorizes You to exercise the Licensed Rights in
+          all media and formats whether now known or hereafter created,
+          and to make technical modifications necessary to do so. The
+          Licensor waives and/or agrees not to assert any right or
+          authority to forbid You from making technical modifications
+          necessary to exercise the Licensed Rights, including
+          technical modifications necessary to circumvent Effective
+          Technological Measures. For purposes of this Public License,
+          simply making modifications authorized by this Section 2(a)
+          (4) never produces Adapted Material.
+
+       5. Downstream recipients.
+
+            a. Offer from the Licensor -- Licensed Material. Every
+               recipient of the Licensed Material automatically
+               receives an offer from the Licensor to exercise the
+               Licensed Rights under the terms and conditions of this
+               Public License.
+
+            b. Additional offer from the Licensor -- Adapted Material.
+               Every recipient of Adapted Material from You
+               automatically receives an offer from the Licensor to
+               exercise the Licensed Rights in the Adapted Material
+               under the conditions of the Adapter's License You apply.
+
+            c. No downstream restrictions. You may not offer or impose
+               any additional or different terms or conditions on, or
+               apply any Effective Technological Measures to, the
+               Licensed Material if doing so restricts exercise of the
+               Licensed Rights by any recipient of the Licensed
+               Material.
+
+       6. No endorsement. Nothing in this Public License constitutes or
+          may be construed as permission to assert or imply that You
+          are, or that Your use of the Licensed Material is, connected
+          with, or sponsored, endorsed, or granted official status by,
+          the Licensor or others designated to receive attribution as
+          provided in Section 3(a)(1)(A)(i).
+
+  b. Other rights.
+
+       1. Moral rights, such as the right of integrity, are not
+          licensed under this Public License, nor are publicity,
+          privacy, and/or other similar personality rights; however, to
+          the extent possible, the Licensor waives and/or agrees not to
+          assert any such rights held by the Licensor to the limited
+          extent necessary to allow You to exercise the Licensed
+          Rights, but not otherwise.
+
+       2. Patent and trademark rights are not licensed under this
+          Public License.
+
+       3. To the extent possible, the Licensor waives any right to
+          collect royalties from You for the exercise of the Licensed
+          Rights, whether directly or through a collecting society
+          under any voluntary or waivable statutory or compulsory
+          licensing scheme. In all other cases the Licensor expressly
+          reserves any right to collect such royalties, including when
+          the Licensed Material is used other than for NonCommercial
+          purposes.
+
+
+Section 3 -- License Conditions.
+
+Your exercise of the Licensed Rights is expressly made subject to the
+following conditions.
+
+  a. Attribution.
+
+       1. If You Share the Licensed Material (including in modified
+          form), You must:
+
+            a. retain the following if it is supplied by the Licensor
+               with the Licensed Material:
+
+                 i. identification of the creator(s) of the Licensed
+                    Material and any others designated to receive
+                    attribution, in any reasonable manner requested by
+                    the Licensor (including by pseudonym if
+                    designated);
+
+                ii. a copyright notice;
+
+               iii. a notice that refers to this Public License;
+
+                iv. a notice that refers to the disclaimer of
+                    warranties;
+
+                 v. a URI or hyperlink to the Licensed Material to the
+                    extent reasonably practicable;
+
+            b. indicate if You modified the Licensed Material and
+               retain an indication of any previous modifications; and
+
+            c. indicate the Licensed Material is licensed under this
+               Public License, and include the text of, or the URI or
+               hyperlink to, this Public License.
+
+       2. You may satisfy the conditions in Section 3(a)(1) in any
+          reasonable manner based on the medium, means, and context in
+          which You Share the Licensed Material. For example, it may be
+          reasonable to satisfy the conditions by providing a URI or
+          hyperlink to a resource that includes the required
+          information.
+       3. If requested by the Licensor, You must remove any of the
+          information required by Section 3(a)(1)(A) to the extent
+          reasonably practicable.
+
+  b. ShareAlike.
+
+     In addition to the conditions in Section 3(a), if You Share
+     Adapted Material You produce, the following conditions also apply.
+
+       1. The Adapter's License You apply must be a Creative Commons
+          license with the same License Elements, this version or
+          later, or a BY-NC-SA Compatible License.
+
+       2. You must include the text of, or the URI or hyperlink to, the
+          Adapter's License You apply. You may satisfy this condition
+          in any reasonable manner based on the medium, means, and
+          context in which You Share Adapted Material.
+
+       3. You may not offer or impose any additional or different terms
+          or conditions on, or apply any Effective Technological
+          Measures to, Adapted Material that restrict exercise of the
+          rights granted under the Adapter's License You apply.
+
+
+Section 4 -- Sui Generis Database Rights.
+
+Where the Licensed Rights include Sui Generis Database Rights that
+apply to Your use of the Licensed Material:
+
+  a. for the avoidance of doubt, Section 2(a)(1) grants You the right
+     to extract, reuse, reproduce, and Share all or a substantial
+     portion of the contents of the database for NonCommercial purposes
+     only;
+
+  b. if You include all or a substantial portion of the database
+     contents in a database in which You have Sui Generis Database
+     Rights, then the database in which You have Sui Generis Database
+     Rights (but not its individual contents) is Adapted Material,
+     including for purposes of Section 3(b); and
+
+  c. You must comply with the conditions in Section 3(a) if You Share
+     all or a substantial portion of the contents of the database.
+
+For the avoidance of doubt, this Section 4 supplements and does not
+replace Your obligations under this Public License where the Licensed
+Rights include other Copyright and Similar Rights.
+
+
+Section 5 -- Disclaimer of Warranties and Limitation of Liability.
+
+  a. UNLESS OTHERWISE SEPARATELY UNDERTAKEN BY THE LICENSOR, TO THE
+     EXTENT POSSIBLE, THE LICENSOR OFFERS THE LICENSED MATERIAL AS-IS
+     AND AS-AVAILABLE, AND MAKES NO REPRESENTATIONS OR WARRANTIES OF
+     ANY KIND CONCERNING THE LICENSED MATERIAL, WHETHER EXPRESS,
+     IMPLIED, STATUTORY, OR OTHER. THIS INCLUDES, WITHOUT LIMITATION,
+     WARRANTIES OF TITLE, MERCHANTABILITY, FITNESS FOR A PARTICULAR
+     PURPOSE, NON-INFRINGEMENT, ABSENCE OF LATENT OR OTHER DEFECTS,
+     ACCURACY, OR THE PRESENCE OR ABSENCE OF ERRORS, WHETHER OR NOT
+     KNOWN OR DISCOVERABLE. WHERE DISCLAIMERS OF WARRANTIES ARE NOT
+     ALLOWED IN FULL OR IN PART, THIS DISCLAIMER MAY NOT APPLY TO YOU.
+
+  b. TO THE EXTENT POSSIBLE, IN NO EVENT WILL THE LICENSOR BE LIABLE
+     TO YOU ON ANY LEGAL THEORY (INCLUDING, WITHOUT LIMITATION,
+     NEGLIGENCE) OR OTHERWISE FOR ANY DIRECT, SPECIAL, INDIRECT,
+     INCIDENTAL, CONSEQUENTIAL, PUNITIVE, EXEMPLARY, OR OTHER LOSSES,
+     COSTS, EXPENSES, OR DAMAGES ARISING OUT OF THIS PUBLIC LICENSE OR
+     USE OF THE LICENSED MATERIAL, EVEN IF THE LICENSOR HAS BEEN
+     ADVISED OF THE POSSIBILITY OF SUCH LOSSES, COSTS, EXPENSES, OR
+     DAMAGES. WHERE A LIMITATION OF LIABILITY IS NOT ALLOWED IN FULL OR
+     IN PART, THIS LIMITATION MAY NOT APPLY TO YOU.
+
+  c. The disclaimer of warranties and limitation of liability provided
+     above shall be interpreted in a manner that, to the extent
+     possible, most closely approximates an absolute disclaimer and
+     waiver of all liability.
+
+
+Section 6 -- Term and Termination.
+
+  a. This Public License applies for the term of the Copyright and
+     Similar Rights licensed here. However, if You fail to comply with
+     this Public License, then Your rights under this Public License
+     terminate automatically.
+
+  b. Where Your right to use the Licensed Material has terminated under
+     Section 6(a), it reinstates:
+
+       1. automatically as of the date the violation is cured, provided
+          it is cured within 30 days of Your discovery of the
+          violation; or
+
+       2. upon express reinstatement by the Licensor.
+
+     For the avoidance of doubt, this Section 6(b) does not affect any
+     right the Licensor may have to seek remedies for Your violations
+     of this Public License.
+
+  c. For the avoidance of doubt, the Licensor may also offer the
+     Licensed Material under separate terms or conditions or stop
+     distributing the Licensed Material at any time; however, doing so
+     will not terminate this Public License.
+
+  d. Sections 1, 5, 6, 7, and 8 survive termination of this Public
+     License.
+
+
+Section 7 -- Other Terms and Conditions.
+
+  a. The Licensor shall not be bound by any additional or different
+     terms or conditions communicated by You unless expressly agreed.
+
+  b. Any arrangements, understandings, or agreements regarding the
+     Licensed Material not stated herein are separate from and
+     independent of the terms and conditions of this Public License.
+
+
+Section 8 -- Interpretation.
+
+  a. For the avoidance of doubt, this Public License does not, and
+     shall not be interpreted to, reduce, limit, restrict, or impose
+     conditions on any use of the Licensed Material that could lawfully
+     be made without permission under this Public License.
+
+  b. To the extent possible, if any provision of this Public License is
+     deemed unenforceable, it shall be automatically reformed to the
+     minimum extent necessary to make it enforceable. If the provision
+     cannot be reformed, it shall be severed from this Public License
+     without affecting the enforceability of the remaining terms and
+     conditions.
+
+  c. No term or condition of this Public License will be waived and no
+     failure to comply consented to unless expressly agreed to by the
+     Licensor.
+
+  d. Nothing in this Public License constitutes or may be interpreted
+     as a limitation upon, or waiver of, any privileges and immunities
+     that apply to the Licensor or You, including from the legal
+     processes of any jurisdiction or authority.
+
+=======================================================================
+
+Creative Commons is not a party to its public
+licenses. Notwithstanding, Creative Commons may elect to apply one of
+its public licenses to material it publishes and in those instances
+will be considered the “Licensor.” The text of the Creative Commons
+public licenses is dedicated to the public domain under the CC0 Public
+Domain Dedication. Except for the limited purpose of indicating that
+material is shared under a Creative Commons public license or as
+otherwise permitted by the Creative Commons policies published at
+creativecommons.org/policies, Creative Commons does not authorize the
+use of the trademark "Creative Commons" or any other trademark or logo
+of Creative Commons without its prior written consent including,
+without limitation, in connection with any unauthorized modifications
+to any of its public licenses or any other arrangements,
+understandings, or agreements concerning use of licensed material. For
+the avoidance of doubt, this paragraph does not form part of the
+public licenses.
+
+Creative Commons may be contacted at creativecommons.org.

README.md

@@ -0,0 +1,22 @@
+---
+title: Audioldm Text To Audio Generation
+emoji: 🔊
+colorFrom: indigo
+colorTo: red
+sdk: gradio
+sdk_version: 3.16.2
+app_file: app.py
+pinned: false
+license: bigscience-openrail-m
+duplicated_from: haoheliu/audioldm-text-to-audio-generation
+---
+
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+
+## Reference
+Part of the code from this repo is borrowed from the following repos. We would like to thank the authors of them for their contribution. 
+
+> https://github.com/LAION-AI/CLAP
+> https://github.com/CompVis/stable-diffusion
+> https://github.com/v-iashin/SpecVQGAN 
+> https://github.com/toshas/torch-fidelity

app.py

@@ -0,0 +1,289 @@
+import gradio as gr
+import numpy as np
+from audioldm import text_to_audio, build_model
+from share_btn import community_icon_html, loading_icon_html, share_js
+
+model_id="haoheliu/AudioLDM-S-Full"
+
+audioldm = build_model()
+# audioldm=None
+
+# def predict(input, history=[]):
+#     # tokenize the new input sentence
+#     new_user_input_ids = tokenizer.encode(input + tokenizer.eos_token, return_tensors='pt')
+
+#     # append the new user input tokens to the chat history
+#     bot_input_ids = torch.cat([torch.LongTensor(history), new_user_input_ids], dim=-1)
+
+#     # generate a response 
+#     history = model.generate(bot_input_ids, max_length=1000, pad_token_id=tokenizer.eos_token_id).tolist()
+
+#     # convert the tokens to text, and then split the responses into lines
+#     response = tokenizer.decode(history[0]).split("<|endoftext|>")
+#     response = [(response[i], response[i+1]) for i in range(0, len(response)-1, 2)]  # convert to tuples of list
+#     return response, history
+  
+def text2audio(text, duration, guidance_scale, random_seed, n_candidates):
+    # print(text, length, guidance_scale)
+    waveform = text_to_audio(audioldm, text, random_seed, duration=duration, guidance_scale=guidance_scale, n_candidate_gen_per_text=int(n_candidates)) # [bs, 1, samples]
+    waveform = [gr.make_waveform((16000, wave[0]), bg_image="bg.png") for wave in waveform]
+    # waveform = [(16000, np.random.randn(16000)), (16000, np.random.randn(16000))]
+    if(len(waveform) == 1):
+      waveform = waveform[0]
+    return waveform
+
+# iface = gr.Interface(fn=text2audio, inputs=[
+#         gr.Textbox(value="A man is speaking in a huge room", max_lines=1),
+#         gr.Slider(2.5, 10, value=5, step=2.5),
+#         gr.Slider(0, 5, value=2.5, step=0.5),
+#         gr.Number(value=42)
+#     ], outputs=[gr.Audio(label="Output", type="numpy"), gr.Audio(label="Output", type="numpy")],
+#                 allow_flagging="never"
+#                      )
+# iface.launch(share=True)
+
+
+css = """
+        a {
+            color: inherit;
+            text-decoration: underline;
+        }
+        .gradio-container {
+            font-family: 'IBM Plex Sans', sans-serif;
+        }
+        .gr-button {
+            color: white;
+            border-color: #000000;
+            background: #000000;
+        }
+        input[type='range'] {
+            accent-color: #000000;
+        }
+        .dark input[type='range'] {
+            accent-color: #dfdfdf;
+        }
+        .container {
+            max-width: 730px;
+            margin: auto;
+            padding-top: 1.5rem;
+        }
+        #gallery {
+            min-height: 22rem;
+            margin-bottom: 15px;
+            margin-left: auto;
+            margin-right: auto;
+            border-bottom-right-radius: .5rem !important;
+            border-bottom-left-radius: .5rem !important;
+        }
+        #gallery>div>.h-full {
+            min-height: 20rem;
+        }
+        .details:hover {
+            text-decoration: underline;
+        }
+        .gr-button {
+            white-space: nowrap;
+        }
+        .gr-button:focus {
+            border-color: rgb(147 197 253 / var(--tw-border-opacity));
+            outline: none;
+            box-shadow: var(--tw-ring-offset-shadow), var(--tw-ring-shadow), var(--tw-shadow, 0 0 #0000);
+            --tw-border-opacity: 1;
+            --tw-ring-offset-shadow: var(--tw-ring-inset) 0 0 0 var(--tw-ring-offset-width) var(--tw-ring-offset-color);
+            --tw-ring-shadow: var(--tw-ring-inset) 0 0 0 calc(3px var(--tw-ring-offset-width)) var(--tw-ring-color);
+            --tw-ring-color: rgb(191 219 254 / var(--tw-ring-opacity));
+            --tw-ring-opacity: .5;
+        }
+        #advanced-btn {
+            font-size: .7rem !important;
+            line-height: 19px;
+            margin-top: 12px;
+            margin-bottom: 12px;
+            padding: 2px 8px;
+            border-radius: 14px !important;
+        }
+        #advanced-options {
+            margin-bottom: 20px;
+        }
+        .footer {
+            margin-bottom: 45px;
+            margin-top: 35px;
+            text-align: center;
+            border-bottom: 1px solid #e5e5e5;
+        }
+        .footer>p {
+            font-size: .8rem;
+            display: inline-block;
+            padding: 0 10px;
+            transform: translateY(10px);
+            background: white;
+        }
+        .dark .footer {
+            border-color: #303030;
+        }
+        .dark .footer>p {
+            background: #0b0f19;
+        }
+        .acknowledgments h4{
+            margin: 1.25em 0 .25em 0;
+            font-weight: bold;
+            font-size: 115%;
+        }
+        #container-advanced-btns{
+            display: flex;
+            flex-wrap: wrap;
+            justify-content: space-between;
+            align-items: center;
+        }
+        .animate-spin {
+            animation: spin 1s linear infinite;
+        }
+        @keyframes spin {
+            from {
+                transform: rotate(0deg);
+            }
+            to {
+                transform: rotate(360deg);
+            }
+        }
+        #share-btn-container {
+            display: flex; padding-left: 0.5rem !important; padding-right: 0.5rem !important; background-color: #000000; justify-content: center; align-items: center; border-radius: 9999px !important; width: 13rem;
+            margin-top: 10px;
+            margin-left: auto;
+        }
+        #share-btn {
+            all: initial; color: #ffffff;font-weight: 600; cursor:pointer; font-family: 'IBM Plex Sans', sans-serif; margin-left: 0.5rem !important; padding-top: 0.25rem !important; padding-bottom: 0.25rem !important;right:0;
+        }
+        #share-btn * {
+            all: unset;
+        }
+        #share-btn-container div:nth-child(-n+2){
+            width: auto !important;
+            min-height: 0px !important;
+        }
+        #share-btn-container .wrap {
+            display: none !important;
+        }
+        .gr-form{
+            flex: 1 1 50%; border-top-right-radius: 0; border-bottom-right-radius: 0;
+        }
+        #prompt-container{
+            gap: 0;
+        }
+        #generated_id{
+            min-height: 700px
+        }
+        #setting_id{
+          margin-bottom: 12px;
+          text-align: center;
+          font-weight: 900;
+        }
+"""
+iface = gr.Blocks(css=css)
+
+with iface:
+    gr.HTML(
+        """
+            <div style="text-align: center; max-width: 700px; margin: 0 auto;">
+              <div
+                style="
+                  display: inline-flex;
+                  align-items: center;
+                  gap: 0.8rem;
+                  font-size: 1.75rem;
+                "
+              >
+                <h1 style="font-weight: 900; margin-bottom: 7px; line-height: normal;">
+                  AudioLDM: Text-to-Audio Generation with Latent Diffusion Models
+                </h1>
+              </div>
+              <p style="margin-bottom: 10px; font-size: 94%">
+                <a href="https://arxiv.org/abs/2301.12503">[Paper]</a>  <a href="https://audioldm.github.io/">[Project page]</a>
+              </p>
+            </div>
+        """
+    )
+    gr.HTML("""
+        <h1 style="font-weight: 900; margin-bottom: 7px;">
+        AudioLDM: Text-to-Audio Generation with Latent Diffusion Models
+        </h1>
+        <p>For faster inference without waiting in queue, you may duplicate the space and upgrade to GPU in settings.
+        <br/>
+        <a href="https://huggingface.co/spaces/haoheliu/audioldm-text-to-audio-generation?duplicate=true">
+        <img style="margin-top: 0em; margin-bottom: 0em" src="https://bit.ly/3gLdBN6" alt="Duplicate Space"></a>
+        <p/>
+    """)
+    with gr.Group():
+        with gr.Box():
+            ############# Input
+            textbox = gr.Textbox(value="A hammer is hitting a wooden surface", max_lines=1, label="Input your text here. Your text is important for the audio quality. Please ensure it is descriptive by using more adjectives.", elem_id="prompt-in")
+
+            with gr.Accordion("Click to modify detailed configurations", open=False):
+              seed = gr.Number(value=45, label="Change this value (any integer number) will lead to a different generation result.")
+              duration = gr.Slider(2.5, 10, value=10, step=2.5, label="Duration (seconds)")
+              guidance_scale = gr.Slider(0, 4, value=2.5, step=0.5, label="Guidance scale (Large => better quality and relavancy to text; Small => better diversity)")
+              n_candidates = gr.Slider(1, 5, value=3, step=1, label="Automatic quality control. This number control the number of candidates (e.g., generate three audios and choose the best to show you). A Larger value usually lead to better quality with heavier computation")
+            ############# Output
+            # outputs=gr.Audio(label="Output", type="numpy")
+            outputs=gr.Video(label="Output", elem_id="output-video")
+            
+            # with gr.Group(elem_id="container-advanced-btns"):
+            #   # advanced_button = gr.Button("Advanced options", elem_id="advanced-btn")
+            #   with gr.Group(elem_id="share-btn-container"):
+            #     community_icon = gr.HTML(community_icon_html, visible=False)
+            #     loading_icon = gr.HTML(loading_icon_html, visible=False)
+            #     share_button = gr.Button("Share to community", elem_id="share-btn", visible=False)
+            # outputs=[gr.Audio(label="Output", type="numpy"), gr.Audio(label="Output", type="numpy")]
+            btn = gr.Button("Submit").style(full_width=True)
+
+        with gr.Group(elem_id="share-btn-container", visible=False):
+            community_icon = gr.HTML(community_icon_html)
+            loading_icon = gr.HTML(loading_icon_html)
+            share_button = gr.Button("Share to community", elem_id="share-btn")
+
+        btn.click(text2audio, inputs=[
+                  textbox, duration, guidance_scale, seed, n_candidates], outputs=[outputs])
+
+        share_button.click(None, [], [], _js=share_js)
+        gr.HTML('''
+        <div class="footer" style="text-align: center; max-width: 700px; margin: 0 auto;">
+                    <p>Follow the latest update of AudioLDM on our<a href="https://github.com/haoheliu/AudioLDM" style="text-decoration: underline;" target="_blank"> Github repo</a>
+                    </p>
+                    <br>
+                    <p>Model by <a href="https://twitter.com/LiuHaohe" style="text-decoration: underline;" target="_blank">Haohe Liu</a></p>
+                    <br>
+        </div>
+        ''')
+        gr.Examples([
+            ["A hammer is hitting a wooden surface", 5, 2.5, 45, 3],
+            ["Peaceful and calming ambient music with singing bowl and other instruments.", 5, 2.5, 45, 3],
+            ["A man is speaking in a small room.", 5, 2.5, 45, 3],
+            ["A female is speaking followed by footstep sound", 5, 2.5, 45, 3],
+            ["Wooden table tapping sound followed by water pouring sound.", 5, 2.5, 45, 3],
+        ],
+            fn=text2audio,
+            inputs=[textbox, duration, guidance_scale, seed, n_candidates],
+            outputs=[outputs],
+            cache_examples=True,
+        )
+        gr.HTML('''
+                <div class="acknowledgements">
+                <p>Essential Tricks for Enhancing the Quality of Your Generated Audio</p>
+                <p>1. Try to use more adjectives to describe your sound. For example: "A man is speaking clearly and slowly in a large room" is better than "A man is speaking". This can make sure AudioLDM understands what you want.</p>
+                <p>2. Try to use different random seeds, which can affect the generation quality significantly sometimes.</p>
+                <p>3. It's better to use general terms like 'man' or 'woman' instead of specific names for individuals or abstract objects that humans may not be familiar with, such as 'mummy'.</p>
+                </div>
+                ''')
+        with gr.Accordion("Additional information", open=False):
+            gr.HTML(
+                """
+                <div class="acknowledgments">
+                    <p> We build the model with data from <a href="http://research.google.com/audioset/">AudioSet</a>, <a href="https://freesound.org/">Freesound</a> and <a href="https://sound-effects.bbcrewind.co.uk/">BBC Sound Effect library</a>. We share this demo based on the <a href="https://assets.publishing.service.gov.uk/government/uploads/system/uploads/attachment_data/file/375954/Research.pdf">UK copyright exception</a> of data for academic research. </p>
+                            </div>
+                        """
+            )
+# <p>This demo is strictly for research demo purpose only. For commercial use please <a href="haoheliu@gmail.com">contact us</a>.</p>
+
+iface.queue(concurrency_count=2)
+iface.launch(debug=True)
+# iface.launch(debug=True, share=True)

audioldm/__init__.py

@@ -0,0 +1,3 @@
+from .ldm import LatentDiffusion
+from .utils import seed_everything
+from .pipeline import *

audioldm/audio/audio_processing.py

@@ -0,0 +1,100 @@
+import torch
+import numpy as np
+import librosa.util as librosa_util
+from scipy.signal import get_window
+
+
+def window_sumsquare(
+    window,
+    n_frames,
+    hop_length,
+    win_length,
+    n_fft,
+    dtype=np.float32,
+    norm=None,
+):
+    """
+    # from librosa 0.6
+    Compute the sum-square envelope of a window function at a given hop length.
+
+    This is used to estimate modulation effects induced by windowing
+    observations in short-time fourier transforms.
+
+    Parameters
+    ----------
+    window : string, tuple, number, callable, or list-like
+        Window specification, as in `get_window`
+
+    n_frames : int > 0
+        The number of analysis frames
+
+    hop_length : int > 0
+        The number of samples to advance between frames
+
+    win_length : [optional]
+        The length of the window function.  By default, this matches `n_fft`.
+
+    n_fft : int > 0
+        The length of each analysis frame.
+
+    dtype : np.dtype
+        The data type of the output
+
+    Returns
+    -------
+    wss : np.ndarray, shape=`(n_fft + hop_length * (n_frames - 1))`
+        The sum-squared envelope of the window function
+    """
+    if win_length is None:
+        win_length = n_fft
+
+    n = n_fft + hop_length * (n_frames - 1)
+    x = np.zeros(n, dtype=dtype)
+
+    # Compute the squared window at the desired length
+    win_sq = get_window(window, win_length, fftbins=True)
+    win_sq = librosa_util.normalize(win_sq, norm=norm) ** 2
+    win_sq = librosa_util.pad_center(win_sq, n_fft)
+
+    # Fill the envelope
+    for i in range(n_frames):
+        sample = i * hop_length
+        x[sample : min(n, sample + n_fft)] += win_sq[: max(0, min(n_fft, n - sample))]
+    return x
+
+
+def griffin_lim(magnitudes, stft_fn, n_iters=30):
+    """
+    PARAMS
+    ------
+    magnitudes: spectrogram magnitudes
+    stft_fn: STFT class with transform (STFT) and inverse (ISTFT) methods
+    """
+
+    angles = np.angle(np.exp(2j * np.pi * np.random.rand(*magnitudes.size())))
+    angles = angles.astype(np.float32)
+    angles = torch.autograd.Variable(torch.from_numpy(angles))
+    signal = stft_fn.inverse(magnitudes, angles).squeeze(1)
+
+    for i in range(n_iters):
+        _, angles = stft_fn.transform(signal)
+        signal = stft_fn.inverse(magnitudes, angles).squeeze(1)
+    return signal
+
+
+def dynamic_range_compression(x, normalize_fun=torch.log, C=1, clip_val=1e-5):
+    """
+    PARAMS
+    ------
+    C: compression factor
+    """
+    return normalize_fun(torch.clamp(x, min=clip_val) * C)
+
+
+def dynamic_range_decompression(x, C=1):
+    """
+    PARAMS
+    ------
+    C: compression factor used to compress
+    """
+    return torch.exp(x) / C

audioldm/audio/stft.py

@@ -0,0 +1,180 @@
+import torch
+import torch.nn.functional as F
+import numpy as np
+from scipy.signal import get_window
+from librosa.util import pad_center, tiny
+from librosa.filters import mel as librosa_mel_fn
+
+from audioldm.audio.audio_processing import (
+    dynamic_range_compression,
+    dynamic_range_decompression,
+    window_sumsquare,
+)
+
+
+class STFT(torch.nn.Module):
+    """adapted from Prem Seetharaman's https://github.com/pseeth/pytorch-stft"""
+
+    def __init__(self, filter_length, hop_length, win_length, window="hann"):
+        super(STFT, self).__init__()
+        self.filter_length = filter_length
+        self.hop_length = hop_length
+        self.win_length = win_length
+        self.window = window
+        self.forward_transform = None
+        scale = self.filter_length / self.hop_length
+        fourier_basis = np.fft.fft(np.eye(self.filter_length))
+
+        cutoff = int((self.filter_length / 2 + 1))
+        fourier_basis = np.vstack(
+            [np.real(fourier_basis[:cutoff, :]), np.imag(fourier_basis[:cutoff, :])]
+        )
+
+        forward_basis = torch.FloatTensor(fourier_basis[:, None, :])
+        inverse_basis = torch.FloatTensor(
+            np.linalg.pinv(scale * fourier_basis).T[:, None, :]
+        )
+
+        if window is not None:
+            assert filter_length >= win_length
+            # get window and zero center pad it to filter_length
+            fft_window = get_window(window, win_length, fftbins=True)
+            fft_window = pad_center(fft_window, filter_length)
+            fft_window = torch.from_numpy(fft_window).float()
+
+            # window the bases
+            forward_basis *= fft_window
+            inverse_basis *= fft_window
+
+        self.register_buffer("forward_basis", forward_basis.float())
+        self.register_buffer("inverse_basis", inverse_basis.float())
+
+    def transform(self, input_data):
+        num_batches = input_data.size(0)
+        num_samples = input_data.size(1)
+
+        self.num_samples = num_samples
+
+        # similar to librosa, reflect-pad the input
+        input_data = input_data.view(num_batches, 1, num_samples)
+        input_data = F.pad(
+            input_data.unsqueeze(1),
+            (int(self.filter_length / 2), int(self.filter_length / 2), 0, 0),
+            mode="reflect",
+        )
+        input_data = input_data.squeeze(1)
+
+        forward_transform = F.conv1d(
+            input_data,
+            torch.autograd.Variable(self.forward_basis, requires_grad=False),
+            stride=self.hop_length,
+            padding=0,
+        ).cpu()
+
+        cutoff = int((self.filter_length / 2) + 1)
+        real_part = forward_transform[:, :cutoff, :]
+        imag_part = forward_transform[:, cutoff:, :]
+
+        magnitude = torch.sqrt(real_part**2 + imag_part**2)
+        phase = torch.autograd.Variable(torch.atan2(imag_part.data, real_part.data))
+
+        return magnitude, phase
+
+    def inverse(self, magnitude, phase):
+        recombine_magnitude_phase = torch.cat(
+            [magnitude * torch.cos(phase), magnitude * torch.sin(phase)], dim=1
+        )
+
+        inverse_transform = F.conv_transpose1d(
+            recombine_magnitude_phase,
+            torch.autograd.Variable(self.inverse_basis, requires_grad=False),
+            stride=self.hop_length,
+            padding=0,
+        )
+
+        if self.window is not None:
+            window_sum = window_sumsquare(
+                self.window,
+                magnitude.size(-1),
+                hop_length=self.hop_length,
+                win_length=self.win_length,
+                n_fft=self.filter_length,
+                dtype=np.float32,
+            )
+            # remove modulation effects
+            approx_nonzero_indices = torch.from_numpy(
+                np.where(window_sum > tiny(window_sum))[0]
+            )
+            window_sum = torch.autograd.Variable(
+                torch.from_numpy(window_sum), requires_grad=False
+            )
+            window_sum = window_sum
+            inverse_transform[:, :, approx_nonzero_indices] /= window_sum[
+                approx_nonzero_indices
+            ]
+
+            # scale by hop ratio
+            inverse_transform *= float(self.filter_length) / self.hop_length
+
+        inverse_transform = inverse_transform[:, :, int(self.filter_length / 2) :]
+        inverse_transform = inverse_transform[:, :, : -int(self.filter_length / 2) :]
+
+        return inverse_transform
+
+    def forward(self, input_data):
+        self.magnitude, self.phase = self.transform(input_data)
+        reconstruction = self.inverse(self.magnitude, self.phase)
+        return reconstruction
+
+
+class TacotronSTFT(torch.nn.Module):
+    def __init__(
+        self,
+        filter_length,
+        hop_length,
+        win_length,
+        n_mel_channels,
+        sampling_rate,
+        mel_fmin,
+        mel_fmax,
+    ):
+        super(TacotronSTFT, self).__init__()
+        self.n_mel_channels = n_mel_channels
+        self.sampling_rate = sampling_rate
+        self.stft_fn = STFT(filter_length, hop_length, win_length)
+        mel_basis = librosa_mel_fn(
+            sampling_rate, filter_length, n_mel_channels, mel_fmin, mel_fmax
+        )
+        mel_basis = torch.from_numpy(mel_basis).float()
+        self.register_buffer("mel_basis", mel_basis)
+
+    def spectral_normalize(self, magnitudes, normalize_fun):
+        output = dynamic_range_compression(magnitudes, normalize_fun)
+        return output
+
+    def spectral_de_normalize(self, magnitudes):
+        output = dynamic_range_decompression(magnitudes)
+        return output
+
+    def mel_spectrogram(self, y, normalize_fun=torch.log):
+        """Computes mel-spectrograms from a batch of waves
+        PARAMS
+        ------
+        y: Variable(torch.FloatTensor) with shape (B, T) in range [-1, 1]
+
+        RETURNS
+        -------
+        mel_output: torch.FloatTensor of shape (B, n_mel_channels, T)
+        """
+        assert torch.min(y.data) >= -1, torch.min(y.data)
+        assert torch.max(y.data) <= 1, torch.max(y.data)
+
+        magnitudes, phases = self.stft_fn.transform(y)
+        magnitudes = magnitudes.data
+        mel_output = torch.matmul(self.mel_basis, magnitudes)
+        mel_output = self.spectral_normalize(mel_output, normalize_fun)
+        energy = torch.norm(magnitudes, dim=1)
+
+        log_magnitudes = self.spectral_normalize(magnitudes, normalize_fun)
+
+        return mel_output, log_magnitudes, energy

audioldm/audio/tools.py

@@ -0,0 +1,33 @@
+import torch
+import numpy as np
+
+
+def get_mel_from_wav(audio, _stft):
+    audio = torch.clip(torch.FloatTensor(audio).unsqueeze(0), -1, 1)
+    audio = torch.autograd.Variable(audio, requires_grad=False)
+    melspec, log_magnitudes_stft, energy = _stft.mel_spectrogram(audio)
+    melspec = torch.squeeze(melspec, 0).numpy().astype(np.float32)
+    log_magnitudes_stft = (
+        torch.squeeze(log_magnitudes_stft, 0).numpy().astype(np.float32)
+    )
+    energy = torch.squeeze(energy, 0).numpy().astype(np.float32)
+    return melspec, log_magnitudes_stft, energy
+
+
+# def inv_mel_spec(mel, out_filename, _stft, griffin_iters=60):
+#     mel = torch.stack([mel])
+#     mel_decompress = _stft.spectral_de_normalize(mel)
+#     mel_decompress = mel_decompress.transpose(1, 2).data.cpu()
+#     spec_from_mel_scaling = 1000
+#     spec_from_mel = torch.mm(mel_decompress[0], _stft.mel_basis)
+#     spec_from_mel = spec_from_mel.transpose(0, 1).unsqueeze(0)
+#     spec_from_mel = spec_from_mel * spec_from_mel_scaling
+
+#     audio = griffin_lim(
+#         torch.autograd.Variable(spec_from_mel[:, :, :-1]), _stft._stft_fn, griffin_iters
+#     )
+
+#     audio = audio.squeeze()
+#     audio = audio.cpu().numpy()
+#     audio_path = out_filename
+#     write(audio_path, _stft.sampling_rate, audio)

audioldm/clap/encoders.py

@@ -0,0 +1,169 @@
+import torch
+import torch.nn as nn
+from audioldm.clap.open_clip import create_model
+from audioldm.clap.training.data import get_audio_features
+import torchaudio
+from transformers import RobertaTokenizer
+import torch.nn.functional as F
+
+
+class CLAPAudioEmbeddingClassifierFreev2(nn.Module):
+    def __init__(
+        self,
+        pretrained_path="",
+        key="class",
+        sampling_rate=16000,
+        embed_mode="audio",
+        unconditional_prob=0.1,
+        random_mute=False,
+        max_random_mute_portion=0.5,
+        training_mode=True,
+    ):
+        super().__init__()
+
+        self.key = key
+        self.device = "cpu"
+        self.precision = "fp32"
+        self.amodel = "HTSAT-tiny"  # or 'PANN-14'
+        self.tmodel = "roberta"  # the best text encoder in our training
+        self.enable_fusion = False  # False if you do not want to use the fusion model
+        self.fusion_type = "aff_2d"
+        self.pretrained = pretrained_path
+        self.embed_mode = embed_mode
+        self.embed_mode_orig = embed_mode
+        self.sampling_rate = sampling_rate
+        self.unconditional_prob = unconditional_prob
+        self.random_mute = random_mute
+        self.tokenize = RobertaTokenizer.from_pretrained("roberta-base")
+        self.max_random_mute_portion = max_random_mute_portion
+        self.training_mode = training_mode
+        self.model, self.model_cfg = create_model(
+            self.amodel,
+            self.tmodel,
+            self.pretrained,
+            precision=self.precision,
+            device=self.device,
+            enable_fusion=self.enable_fusion,
+            fusion_type=self.fusion_type,
+        )
+        for p in self.model.parameters():
+            p.requires_grad = False
+
+        self.model.eval()
+
+    def get_unconditional_condition(self, batchsize):
+        self.unconditional_token = self.model.get_text_embedding(
+            self.tokenizer(["", ""])
+        )[0:1]
+        return torch.cat([self.unconditional_token.unsqueeze(0)] * batchsize, dim=0)
+
+    def batch_to_list(self, batch):
+        ret = []
+        for i in range(batch.size(0)):
+            ret.append(batch[i])
+        return ret
+
+    def make_decision(self, probability):
+        if float(torch.rand(1)) < probability:
+            return True
+        else:
+            return False
+
+    def random_uniform(self, start, end):
+        val = torch.rand(1).item()
+        return start + (end - start) * val
+
+    def _random_mute(self, waveform):
+        # waveform: [bs, t-steps]
+        t_steps = waveform.size(-1)
+        for i in range(waveform.size(0)):
+            mute_size = int(
+                self.random_uniform(0, end=int(t_steps * self.max_random_mute_portion))
+            )
+            mute_start = int(self.random_uniform(0, t_steps - mute_size))
+            waveform[i, mute_start : mute_start + mute_size] = 0
+        return waveform
+
+    def cos_similarity(self, waveform, text):
+        # waveform: [bs, t_steps]
+        with torch.no_grad():
+            self.embed_mode = "audio"
+            audio_emb = self(waveform.cuda())
+            self.embed_mode = "text"
+            text_emb = self(text)
+            similarity = F.cosine_similarity(audio_emb, text_emb, dim=2)
+            return similarity.squeeze()
+
+    def forward(self, batch, key=None):
+        # If you want this conditioner to be unconditional, set self.unconditional_prob = 1.0
+        # If you want this conditioner to be fully conditional, set self.unconditional_prob = 0.0
+        if self.model.training == True and not self.training_mode:
+            print(
+                "The pretrained CLAP model should always be in eval mode. Reloading model just in case you change the parameters."
+            )
+            self.model, self.model_cfg = create_model(
+                self.amodel,
+                self.tmodel,
+                self.pretrained,
+                precision=self.precision,
+                device="cuda",
+                enable_fusion=self.enable_fusion,
+                fusion_type=self.fusion_type,
+            )
+            for p in self.model.parameters():
+                p.requires_grad = False
+            self.model.eval()
+
+        # the 'fusion' truncate mode can be changed to 'rand_trunc' if run in unfusion mode
+        if self.embed_mode == "audio":
+            with torch.no_grad():
+                audio_dict_list = []
+                assert (
+                    self.sampling_rate == 16000
+                ), "We only support 16000 sampling rate"
+                if self.random_mute:
+                    batch = self._random_mute(batch)
+                # batch: [bs, 1, t-samples]
+                batch = torchaudio.functional.resample(
+                    batch, orig_freq=self.sampling_rate, new_freq=48000
+                )
+                for waveform in self.batch_to_list(batch):
+                    audio_dict = {}
+                    audio_dict = get_audio_features(
+                        audio_dict,
+                        waveform,
+                        480000,
+                        data_truncating="fusion",
+                        data_filling="repeatpad",
+                        audio_cfg=self.model_cfg["audio_cfg"],
+                    )
+                    audio_dict_list.append(audio_dict)
+                # [bs, 512]
+                embed = self.model.get_audio_embedding(audio_dict_list)
+        elif self.embed_mode == "text":
+            with torch.no_grad():
+                # the 'fusion' truncate mode can be changed to 'rand_trunc' if run in unfusion mode
+                text_data = self.tokenizer(batch)
+                embed = self.model.get_text_embedding(text_data)
+
+        embed = embed.unsqueeze(1)
+        self.unconditional_token = self.model.get_text_embedding(
+            self.tokenizer(["", ""])
+        )[0:1]
+
+        for i in range(embed.size(0)):
+            if self.make_decision(self.unconditional_prob):
+                embed[i] = self.unconditional_token
+
+        # [bs, 1, 512]
+        return embed.detach()
+
+    def tokenizer(self, text):
+        result = self.tokenize(
+            text,
+            padding="max_length",
+            truncation=True,
+            max_length=512,
+            return_tensors="pt",
+        )
+        return {k: v.squeeze(0) for k, v in result.items()}

audioldm/clap/open_clip/__init__.py

@@ -0,0 +1,25 @@
+from .factory import (
+    list_models,
+    create_model,
+    create_model_and_transforms,
+    add_model_config,
+)
+from .loss import ClipLoss, gather_features, LPLoss, lp_gather_features, LPMetrics
+from .model import (
+    CLAP,
+    CLAPTextCfg,
+    CLAPVisionCfg,
+    CLAPAudioCfp,
+    convert_weights_to_fp16,
+    trace_model,
+)
+from .openai import load_openai_model, list_openai_models
+from .pretrained import (
+    list_pretrained,
+    list_pretrained_tag_models,
+    list_pretrained_model_tags,
+    get_pretrained_url,
+    download_pretrained,
+)
+from .tokenizer import SimpleTokenizer, tokenize
+from .transform import image_transform

audioldm/clap/open_clip/bert.py

@@ -0,0 +1,40 @@
+from transformers import BertTokenizer, BertModel
+
+tokenizer = BertTokenizer.from_pretrained("bert-base-uncased")
+model = BertModel.from_pretrained("bert-base-uncased")
+text = "Replace me by any text you'd like."
+
+
+def bert_embeddings(text):
+    # text = "Replace me by any text you'd like."
+    encoded_input = tokenizer(text, return_tensors="pt")
+    output = model(**encoded_input)
+    return output
+
+
+from transformers import RobertaTokenizer, RobertaModel
+
+tokenizer = RobertaTokenizer.from_pretrained("roberta-base")
+model = RobertaModel.from_pretrained("roberta-base")
+text = "Replace me by any text you'd like."
+
+
+def Roberta_embeddings(text):
+    # text = "Replace me by any text you'd like."
+    encoded_input = tokenizer(text, return_tensors="pt")
+    output = model(**encoded_input)
+    return output
+
+
+from transformers import BartTokenizer, BartModel
+
+tokenizer = BartTokenizer.from_pretrained("facebook/bart-base")
+model = BartModel.from_pretrained("facebook/bart-base")
+text = "Replace me by any text you'd like."
+
+
+def bart_embeddings(text):
+    # text = "Replace me by any text you'd like."
+    encoded_input = tokenizer(text, return_tensors="pt")
+    output = model(**encoded_input)
+    return output

audioldm/clap/open_clip/bpe_simple_vocab_16e6.txt.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:924691ac288e54409236115652ad4aa250f48203de50a9e4722a6ecd48d6804a
+size 1356917

audioldm/clap/open_clip/factory.py

@@ -0,0 +1,277 @@
+import json
+import logging
+import os
+import pathlib
+import re
+from copy import deepcopy
+from pathlib import Path
+
+import torch
+
+from .model import CLAP, convert_weights_to_fp16
+from .openai import load_openai_model
+from .pretrained import get_pretrained_url, download_pretrained
+from .transform import image_transform
+
+_MODEL_CONFIG_PATHS = [Path(__file__).parent / f"model_configs/"]
+_MODEL_CONFIGS = {}  # directory (model_name: config) of model architecture configs
+
+
+def _natural_key(string_):
+    return [int(s) if s.isdigit() else s for s in re.split(r"(\d+)", string_.lower())]
+
+
+def _rescan_model_configs():
+    global _MODEL_CONFIGS
+
+    config_ext = (".json",)
+    config_files = []
+    for config_path in _MODEL_CONFIG_PATHS:
+        if config_path.is_file() and config_path.suffix in config_ext:
+            config_files.append(config_path)
+        elif config_path.is_dir():
+            for ext in config_ext:
+                config_files.extend(config_path.glob(f"*{ext}"))
+
+    for cf in config_files:
+        if os.path.basename(cf)[0] == ".":
+            continue  # Ignore hidden files
+
+        with open(cf, "r") as f:
+            model_cfg = json.load(f)
+            if all(a in model_cfg for a in ("embed_dim", "audio_cfg", "text_cfg")):
+                _MODEL_CONFIGS[cf.stem] = model_cfg
+
+    _MODEL_CONFIGS = {
+        k: v
+        for k, v in sorted(_MODEL_CONFIGS.items(), key=lambda x: _natural_key(x[0]))
+    }
+
+
+_rescan_model_configs()  # initial populate of model config registry
+
+
+def load_state_dict(checkpoint_path: str, map_location="cpu", skip_params=True):
+    checkpoint = torch.load(checkpoint_path, map_location=map_location)
+    if isinstance(checkpoint, dict) and "state_dict" in checkpoint:
+        state_dict = checkpoint["state_dict"]
+    else:
+        state_dict = checkpoint
+    if skip_params:
+        if next(iter(state_dict.items()))[0].startswith("module"):
+            state_dict = {k[7:]: v for k, v in state_dict.items()}
+    # for k in state_dict:
+    #     if k.startswith('transformer'):
+    #         v = state_dict.pop(k)
+    #         state_dict['text_branch.' + k[12:]] = v
+    return state_dict
+
+
+def create_model(
+    amodel_name: str,
+    tmodel_name: str,
+    pretrained: str = "",
+    precision: str = "fp32",
+    device: torch.device = torch.device("cpu"),
+    jit: bool = False,
+    force_quick_gelu: bool = False,
+    openai_model_cache_dir: str = os.path.expanduser("~/.cache/clip"),
+    skip_params=True,
+    pretrained_audio: str = "",
+    pretrained_text: str = "",
+    enable_fusion: bool = False,
+    fusion_type: str = "None"
+    # pretrained_image: bool = False,
+):
+    amodel_name = amodel_name.replace(
+        "/", "-"
+    )  # for callers using old naming with / in ViT names
+    pretrained_orig = pretrained
+    pretrained = pretrained.lower()
+    if pretrained == "openai":
+        if amodel_name in _MODEL_CONFIGS:
+            logging.info(f"Loading {amodel_name} model config.")
+            model_cfg = deepcopy(_MODEL_CONFIGS[amodel_name])
+        else:
+            logging.error(
+                f"Model config for {amodel_name} not found; available models {list_models()}."
+            )
+            raise RuntimeError(f"Model config for {amodel_name} not found.")
+
+        logging.info(f"Loading pretrained ViT-B-16 text encoder from OpenAI.")
+        # Hard Code in model name
+        model_cfg["text_cfg"]["model_type"] = tmodel_name
+        model = load_openai_model(
+            "ViT-B-16",
+            model_cfg,
+            device=device,
+            jit=jit,
+            cache_dir=openai_model_cache_dir,
+            enable_fusion=enable_fusion,
+            fusion_type=fusion_type,
+        )
+        # See https://discuss.pytorch.org/t/valueerror-attemting-to-unscale-fp16-gradients/81372
+        if precision == "amp" or precision == "fp32":
+            model = model.float()
+    else:
+        if amodel_name in _MODEL_CONFIGS:
+            logging.info(f"Loading {amodel_name} model config.")
+            model_cfg = deepcopy(_MODEL_CONFIGS[amodel_name])
+        else:
+            logging.error(
+                f"Model config for {amodel_name} not found; available models {list_models()}."
+            )
+            raise RuntimeError(f"Model config for {amodel_name} not found.")
+
+        if force_quick_gelu:
+            # override for use of QuickGELU on non-OpenAI transformer models
+            model_cfg["quick_gelu"] = True
+
+        # if pretrained_image:
+        #     if 'timm_amodel_name' in model_cfg.get('vision_cfg', {}):
+        #         # pretrained weight loading for timm models set via vision_cfg
+        #         model_cfg['vision_cfg']['timm_model_pretrained'] = True
+        #     else:
+        #         assert False, 'pretrained image towers currently only supported for timm models'
+        model_cfg["text_cfg"]["model_type"] = tmodel_name
+        model_cfg["enable_fusion"] = enable_fusion
+        model_cfg["fusion_type"] = fusion_type
+        model = CLAP(**model_cfg)
+
+        if pretrained:
+            checkpoint_path = ""
+            url = get_pretrained_url(amodel_name, pretrained)
+            if url:
+                checkpoint_path = download_pretrained(url, root=openai_model_cache_dir)
+            elif os.path.exists(pretrained_orig):
+                checkpoint_path = pretrained_orig
+            if checkpoint_path:
+                logging.info(
+                    f"Loading pretrained {amodel_name}-{tmodel_name} weights ({pretrained})."
+                )
+                ckpt = load_state_dict(checkpoint_path, skip_params=True)
+                model.load_state_dict(ckpt)
+                param_names = [n for n, p in model.named_parameters()]
+                # for n in param_names:
+                #     print(n, "\t", "Loaded" if n in ckpt else "Unloaded")
+            else:
+                logging.warning(
+                    f"Pretrained weights ({pretrained}) not found for model {amodel_name}."
+                )
+                raise RuntimeError(
+                    f"Pretrained weights ({pretrained}) not found for model {amodel_name}."
+                )
+
+        if pretrained_audio:
+            if amodel_name.startswith("PANN"):
+                if "Cnn14_mAP" in pretrained_audio:  # official checkpoint
+                    audio_ckpt = torch.load(pretrained_audio, map_location="cpu")
+                    audio_ckpt = audio_ckpt["model"]
+                    keys = list(audio_ckpt.keys())
+                    for key in keys:
+                        if (
+                            "spectrogram_extractor" not in key
+                            and "logmel_extractor" not in key
+                        ):
+                            v = audio_ckpt.pop(key)
+                            audio_ckpt["audio_branch." + key] = v
+                elif os.path.basename(pretrained_audio).startswith(
+                    "PANN"
+                ):  # checkpoint trained via HTSAT codebase
+                    audio_ckpt = torch.load(pretrained_audio, map_location="cpu")
+                    audio_ckpt = audio_ckpt["state_dict"]
+                    keys = list(audio_ckpt.keys())
+                    for key in keys:
+                        if key.startswith("sed_model"):
+                            v = audio_ckpt.pop(key)
+                            audio_ckpt["audio_branch." + key[10:]] = v
+                elif os.path.basename(pretrained_audio).startswith(
+                    "finetuned"
+                ):  # checkpoint trained via linear probe codebase
+                    audio_ckpt = torch.load(pretrained_audio, map_location="cpu")
+                else:
+                    raise ValueError("Unknown audio checkpoint")
+            elif amodel_name.startswith("HTSAT"):
+                if "HTSAT_AudioSet_Saved" in pretrained_audio:  # official checkpoint
+                    audio_ckpt = torch.load(pretrained_audio, map_location="cpu")
+                    audio_ckpt = audio_ckpt["state_dict"]
+                    keys = list(audio_ckpt.keys())
+                    for key in keys:
+                        if key.startswith("sed_model") and (
+                            "spectrogram_extractor" not in key
+                            and "logmel_extractor" not in key
+                        ):
+                            v = audio_ckpt.pop(key)
+                            audio_ckpt["audio_branch." + key[10:]] = v
+                elif os.path.basename(pretrained_audio).startswith(
+                    "HTSAT"
+                ):  # checkpoint trained via HTSAT codebase
+                    audio_ckpt = torch.load(pretrained_audio, map_location="cpu")
+                    audio_ckpt = audio_ckpt["state_dict"]
+                    keys = list(audio_ckpt.keys())
+                    for key in keys:
+                        if key.startswith("sed_model"):
+                            v = audio_ckpt.pop(key)
+                            audio_ckpt["audio_branch." + key[10:]] = v
+                elif os.path.basename(pretrained_audio).startswith(
+                    "finetuned"
+                ):  # checkpoint trained via linear probe codebase
+                    audio_ckpt = torch.load(pretrained_audio, map_location="cpu")
+                else:
+                    raise ValueError("Unknown audio checkpoint")
+            else:
+                raise f"this audio encoder pretrained checkpoint is not support"
+
+            model.load_state_dict(audio_ckpt, strict=False)
+            logging.info(
+                f"Loading pretrained {amodel_name} weights ({pretrained_audio})."
+            )
+            param_names = [n for n, p in model.named_parameters()]
+            for n in param_names:
+                print(n, "\t", "Loaded" if n in audio_ckpt else "Unloaded")
+
+        model.to(device=device)
+        if precision == "fp16":
+            assert device.type != "cpu"
+            convert_weights_to_fp16(model)
+
+        if jit:
+            model = torch.jit.script(model)
+
+    return model, model_cfg
+
+
+def create_model_and_transforms(
+    model_name: str,
+    pretrained: str = "",
+    precision: str = "fp32",
+    device: torch.device = torch.device("cpu"),
+    jit: bool = False,
+    force_quick_gelu: bool = False,
+    # pretrained_image: bool = False,
+):
+    model = create_model(
+        model_name,
+        pretrained,
+        precision,
+        device,
+        jit,
+        force_quick_gelu=force_quick_gelu,
+        # pretrained_image=pretrained_image
+    )
+    preprocess_train = image_transform(model.visual.image_size, is_train=True)
+    preprocess_val = image_transform(model.visual.image_size, is_train=False)
+    return model, preprocess_train, preprocess_val
+
+
+def list_models():
+    """enumerate available model architectures based on config files"""
+    return list(_MODEL_CONFIGS.keys())
+
+
+def add_model_config(path):
+    """add model config path or file and update registry"""
+    if not isinstance(path, Path):
+        path = Path(path)
+    _MODEL_CONFIG_PATHS.append(path)
+    _rescan_model_configs()

audioldm/clap/open_clip/feature_fusion.py

@@ -0,0 +1,192 @@
+"""
+Feature Fusion for Varible-Length Data Processing
+AFF/iAFF is referred and modified from https://github.com/YimianDai/open-aff/blob/master/aff_pytorch/aff_net/fusion.py
+According to the paper: Yimian Dai et al, Attentional Feature Fusion, IEEE Winter Conference on Applications of Computer Vision, WACV 2021
+"""
+
+import torch
+import torch.nn as nn
+
+
+class DAF(nn.Module):
+    """
+    直接相加 DirectAddFuse
+    """
+
+    def __init__(self):
+        super(DAF, self).__init__()
+
+    def forward(self, x, residual):
+        return x + residual
+
+
+class iAFF(nn.Module):
+    """
+    多特征融合 iAFF
+    """
+
+    def __init__(self, channels=64, r=4, type="2D"):
+        super(iAFF, self).__init__()
+        inter_channels = int(channels // r)
+
+        if type == "1D":
+            # 本地注意力
+            self.local_att = nn.Sequential(
+                nn.Conv1d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv1d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(channels),
+            )
+
+            # 全局注意力
+            self.global_att = nn.Sequential(
+                nn.AdaptiveAvgPool1d(1),
+                nn.Conv1d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv1d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(channels),
+            )
+
+            # 第二次本地注意力
+            self.local_att2 = nn.Sequential(
+                nn.Conv1d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv1d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(channels),
+            )
+            # 第二次全局注意力
+            self.global_att2 = nn.Sequential(
+                nn.AdaptiveAvgPool1d(1),
+                nn.Conv1d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv1d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(channels),
+            )
+        elif type == "2D":
+            # 本地注意力
+            self.local_att = nn.Sequential(
+                nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(channels),
+            )
+
+            # 全局注意力
+            self.global_att = nn.Sequential(
+                nn.AdaptiveAvgPool2d(1),
+                nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(channels),
+            )
+
+            # 第二次本地注意力
+            self.local_att2 = nn.Sequential(
+                nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(channels),
+            )
+            # 第二次全局注意力
+            self.global_att2 = nn.Sequential(
+                nn.AdaptiveAvgPool2d(1),
+                nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(channels),
+            )
+        else:
+            raise f"the type is not supported"
+
+        self.sigmoid = nn.Sigmoid()
+
+    def forward(self, x, residual):
+        flag = False
+        xa = x + residual
+        if xa.size(0) == 1:
+            xa = torch.cat([xa, xa], dim=0)
+            flag = True
+        xl = self.local_att(xa)
+        xg = self.global_att(xa)
+        xlg = xl + xg
+        wei = self.sigmoid(xlg)
+        xi = x * wei + residual * (1 - wei)
+
+        xl2 = self.local_att2(xi)
+        xg2 = self.global_att(xi)
+        xlg2 = xl2 + xg2
+        wei2 = self.sigmoid(xlg2)
+        xo = x * wei2 + residual * (1 - wei2)
+        if flag:
+            xo = xo[0].unsqueeze(0)
+        return xo
+
+
+class AFF(nn.Module):
+    """
+    多特征融合 AFF
+    """
+
+    def __init__(self, channels=64, r=4, type="2D"):
+        super(AFF, self).__init__()
+        inter_channels = int(channels // r)
+
+        if type == "1D":
+            self.local_att = nn.Sequential(
+                nn.Conv1d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv1d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(channels),
+            )
+            self.global_att = nn.Sequential(
+                nn.AdaptiveAvgPool1d(1),
+                nn.Conv1d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv1d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(channels),
+            )
+        elif type == "2D":
+            self.local_att = nn.Sequential(
+                nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(channels),
+            )
+            self.global_att = nn.Sequential(
+                nn.AdaptiveAvgPool2d(1),
+                nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(channels),
+            )
+        else:
+            raise f"the type is not supported."
+
+        self.sigmoid = nn.Sigmoid()
+
+    def forward(self, x, residual):
+        flag = False
+        xa = x + residual
+        if xa.size(0) == 1:
+            xa = torch.cat([xa, xa], dim=0)
+            flag = True
+        xl = self.local_att(xa)
+        xg = self.global_att(xa)
+        xlg = xl + xg
+        wei = self.sigmoid(xlg)
+        xo = 2 * x * wei + 2 * residual * (1 - wei)
+        if flag:
+            xo = xo[0].unsqueeze(0)
+        return xo

audioldm/clap/open_clip/htsat.py

@@ -0,0 +1,1308 @@
+# Ke Chen
+# knutchen@ucsd.edu
+# HTS-AT: A HIERARCHICAL TOKEN-SEMANTIC AUDIO TRANSFORMER FOR SOUND CLASSIFICATION AND DETECTION
+# Some layers designed on the model
+# below codes are based and referred from https://github.com/microsoft/Swin-Transformer
+# Swin Transformer for Computer Vision: https://arxiv.org/pdf/2103.14030.pdf
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from itertools import repeat
+import collections.abc
+import math
+import warnings
+
+from torch.nn.init import _calculate_fan_in_and_fan_out
+import torch.utils.checkpoint as checkpoint
+
+import random
+
+from torchlibrosa.stft import Spectrogram, LogmelFilterBank
+from torchlibrosa.augmentation import SpecAugmentation
+
+from itertools import repeat
+from .utils import do_mixup, interpolate
+
+from .feature_fusion import iAFF, AFF, DAF
+
+# from PyTorch internals
+def _ntuple(n):
+    def parse(x):
+        if isinstance(x, collections.abc.Iterable):
+            return x
+        return tuple(repeat(x, n))
+
+    return parse
+
+
+to_1tuple = _ntuple(1)
+to_2tuple = _ntuple(2)
+to_3tuple = _ntuple(3)
+to_4tuple = _ntuple(4)
+to_ntuple = _ntuple
+
+
+def drop_path(x, drop_prob: float = 0.0, training: bool = False):
+    """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
+    This is the same as the DropConnect impl I created for EfficientNet, etc networks, however,
+    the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...
+    See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for
+    changing the layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use
+    'survival rate' as the argument.
+    """
+    if drop_prob == 0.0 or not training:
+        return x
+    keep_prob = 1 - drop_prob
+    shape = (x.shape[0],) + (1,) * (
+        x.ndim - 1
+    )  # work with diff dim tensors, not just 2D ConvNets
+    random_tensor = keep_prob + torch.rand(shape, dtype=x.dtype, device=x.device)
+    random_tensor.floor_()  # binarize
+    output = x.div(keep_prob) * random_tensor
+    return output
+
+
+class DropPath(nn.Module):
+    """Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks)."""
+
+    def __init__(self, drop_prob=None):
+        super(DropPath, self).__init__()
+        self.drop_prob = drop_prob
+
+    def forward(self, x):
+        return drop_path(x, self.drop_prob, self.training)
+
+
+class PatchEmbed(nn.Module):
+    """2D Image to Patch Embedding"""
+
+    def __init__(
+        self,
+        img_size=224,
+        patch_size=16,
+        in_chans=3,
+        embed_dim=768,
+        norm_layer=None,
+        flatten=True,
+        patch_stride=16,
+        enable_fusion=False,
+        fusion_type="None",
+    ):
+        super().__init__()
+        img_size = to_2tuple(img_size)
+        patch_size = to_2tuple(patch_size)
+        patch_stride = to_2tuple(patch_stride)
+        self.img_size = img_size
+        self.patch_size = patch_size
+        self.patch_stride = patch_stride
+        self.grid_size = (
+            img_size[0] // patch_stride[0],
+            img_size[1] // patch_stride[1],
+        )
+        self.num_patches = self.grid_size[0] * self.grid_size[1]
+        self.flatten = flatten
+        self.in_chans = in_chans
+        self.embed_dim = embed_dim
+
+        self.enable_fusion = enable_fusion
+        self.fusion_type = fusion_type
+
+        padding = (
+            (patch_size[0] - patch_stride[0]) // 2,
+            (patch_size[1] - patch_stride[1]) // 2,
+        )
+
+        if (self.enable_fusion) and (self.fusion_type == "channel_map"):
+            self.proj = nn.Conv2d(
+                in_chans * 4,
+                embed_dim,
+                kernel_size=patch_size,
+                stride=patch_stride,
+                padding=padding,
+            )
+        else:
+            self.proj = nn.Conv2d(
+                in_chans,
+                embed_dim,
+                kernel_size=patch_size,
+                stride=patch_stride,
+                padding=padding,
+            )
+        self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()
+
+        if (self.enable_fusion) and (
+            self.fusion_type in ["daf_2d", "aff_2d", "iaff_2d"]
+        ):
+            self.mel_conv2d = nn.Conv2d(
+                in_chans,
+                embed_dim,
+                kernel_size=(patch_size[0], patch_size[1] * 3),
+                stride=(patch_stride[0], patch_stride[1] * 3),
+                padding=padding,
+            )
+            if self.fusion_type == "daf_2d":
+                self.fusion_model = DAF()
+            elif self.fusion_type == "aff_2d":
+                self.fusion_model = AFF(channels=embed_dim, type="2D")
+            elif self.fusion_type == "iaff_2d":
+                self.fusion_model = iAFF(channels=embed_dim, type="2D")
+
+    def forward(self, x, longer_idx=None):
+        if (self.enable_fusion) and (
+            self.fusion_type in ["daf_2d", "aff_2d", "iaff_2d"]
+        ):
+            global_x = x[:, 0:1, :, :]
+
+            # global processing
+            B, C, H, W = global_x.shape
+            assert (
+                H == self.img_size[0] and W == self.img_size[1]
+            ), f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
+            global_x = self.proj(global_x)
+            TW = global_x.size(-1)
+            if len(longer_idx) > 0:
+                # local processing
+                local_x = x[longer_idx, 1:, :, :].contiguous()
+                B, C, H, W = local_x.shape
+                local_x = local_x.view(B * C, 1, H, W)
+                local_x = self.mel_conv2d(local_x)
+                local_x = local_x.view(
+                    B, C, local_x.size(1), local_x.size(2), local_x.size(3)
+                )
+                local_x = local_x.permute((0, 2, 3, 1, 4)).contiguous().flatten(3)
+                TB, TC, TH, _ = local_x.size()
+                if local_x.size(-1) < TW:
+                    local_x = torch.cat(
+                        [
+                            local_x,
+                            torch.zeros(
+                                (TB, TC, TH, TW - local_x.size(-1)),
+                                device=global_x.device,
+                            ),
+                        ],
+                        dim=-1,
+                    )
+                else:
+                    local_x = local_x[:, :, :, :TW]
+
+                global_x[longer_idx] = self.fusion_model(global_x[longer_idx], local_x)
+            x = global_x
+        else:
+            B, C, H, W = x.shape
+            assert (
+                H == self.img_size[0] and W == self.img_size[1]
+            ), f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
+            x = self.proj(x)
+
+        if self.flatten:
+            x = x.flatten(2).transpose(1, 2)  # BCHW -> BNC
+        x = self.norm(x)
+        return x
+
+
+class Mlp(nn.Module):
+    """MLP as used in Vision Transformer, MLP-Mixer and related networks"""
+
+    def __init__(
+        self,
+        in_features,
+        hidden_features=None,
+        out_features=None,
+        act_layer=nn.GELU,
+        drop=0.0,
+    ):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+def _no_grad_trunc_normal_(tensor, mean, std, a, b):
+    # Cut & paste from PyTorch official master until it's in a few official releases - RW
+    # Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf
+    def norm_cdf(x):
+        # Computes standard normal cumulative distribution function
+        return (1.0 + math.erf(x / math.sqrt(2.0))) / 2.0
+
+    if (mean < a - 2 * std) or (mean > b + 2 * std):
+        warnings.warn(
+            "mean is more than 2 std from [a, b] in nn.init.trunc_normal_. "
+            "The distribution of values may be incorrect.",
+            stacklevel=2,
+        )
+
+    with torch.no_grad():
+        # Values are generated by using a truncated uniform distribution and
+        # then using the inverse CDF for the normal distribution.
+        # Get upper and lower cdf values
+        l = norm_cdf((a - mean) / std)
+        u = norm_cdf((b - mean) / std)
+
+        # Uniformly fill tensor with values from [l, u], then translate to
+        # [2l-1, 2u-1].
+        tensor.uniform_(2 * l - 1, 2 * u - 1)
+
+        # Use inverse cdf transform for normal distribution to get truncated
+        # standard normal
+        tensor.erfinv_()
+
+        # Transform to proper mean, std
+        tensor.mul_(std * math.sqrt(2.0))
+        tensor.add_(mean)
+
+        # Clamp to ensure it's in the proper range
+        tensor.clamp_(min=a, max=b)
+        return tensor
+
+
+def trunc_normal_(tensor, mean=0.0, std=1.0, a=-2.0, b=2.0):
+    # type: (Tensor, float, float, float, float) -> Tensor
+    r"""Fills the input Tensor with values drawn from a truncated
+    normal distribution. The values are effectively drawn from the
+    normal distribution :math:`\mathcal{N}(\text{mean}, \text{std}^2)`
+    with values outside :math:`[a, b]` redrawn until they are within
+    the bounds. The method used for generating the random values works
+    best when :math:`a \leq \text{mean} \leq b`.
+    Args:
+        tensor: an n-dimensional `torch.Tensor`
+        mean: the mean of the normal distribution
+        std: the standard deviation of the normal distribution
+        a: the minimum cutoff value
+        b: the maximum cutoff value
+    Examples:
+        >>> w = torch.empty(3, 5)
+        >>> nn.init.trunc_normal_(w)
+    """
+    return _no_grad_trunc_normal_(tensor, mean, std, a, b)
+
+
+def variance_scaling_(tensor, scale=1.0, mode="fan_in", distribution="normal"):
+    fan_in, fan_out = _calculate_fan_in_and_fan_out(tensor)
+    if mode == "fan_in":
+        denom = fan_in
+    elif mode == "fan_out":
+        denom = fan_out
+    elif mode == "fan_avg":
+        denom = (fan_in + fan_out) / 2
+
+    variance = scale / denom
+
+    if distribution == "truncated_normal":
+        # constant is stddev of standard normal truncated to (-2, 2)
+        trunc_normal_(tensor, std=math.sqrt(variance) / 0.87962566103423978)
+    elif distribution == "normal":
+        tensor.normal_(std=math.sqrt(variance))
+    elif distribution == "uniform":
+        bound = math.sqrt(3 * variance)
+        tensor.uniform_(-bound, bound)
+    else:
+        raise ValueError(f"invalid distribution {distribution}")
+
+
+def lecun_normal_(tensor):
+    variance_scaling_(tensor, mode="fan_in", distribution="truncated_normal")
+
+
+def window_partition(x, window_size):
+    """
+    Args:
+        x: (B, H, W, C)
+        window_size (int): window size
+    Returns:
+        windows: (num_windows*B, window_size, window_size, C)
+    """
+    B, H, W, C = x.shape
+    x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
+    windows = (
+        x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
+    )
+    return windows
+
+
+def window_reverse(windows, window_size, H, W):
+    """
+    Args:
+        windows: (num_windows*B, window_size, window_size, C)
+        window_size (int): Window size
+        H (int): Height of image
+        W (int): Width of image
+    Returns:
+        x: (B, H, W, C)
+    """
+    B = int(windows.shape[0] / (H * W / window_size / window_size))
+    x = windows.view(
+        B, H // window_size, W // window_size, window_size, window_size, -1
+    )
+    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
+    return x
+
+
+class WindowAttention(nn.Module):
+    r"""Window based multi-head self attention (W-MSA) module with relative position bias.
+    It supports both of shifted and non-shifted window.
+    Args:
+        dim (int): Number of input channels.
+        window_size (tuple[int]): The height and width of the window.
+        num_heads (int): Number of attention heads.
+        qkv_bias (bool, optional):  If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set
+        attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
+        proj_drop (float, optional): Dropout ratio of output. Default: 0.0
+    """
+
+    def __init__(
+        self,
+        dim,
+        window_size,
+        num_heads,
+        qkv_bias=True,
+        qk_scale=None,
+        attn_drop=0.0,
+        proj_drop=0.0,
+    ):
+
+        super().__init__()
+        self.dim = dim
+        self.window_size = window_size  # Wh, Ww
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = qk_scale or head_dim**-0.5
+
+        # define a parameter table of relative position bias
+        self.relative_position_bias_table = nn.Parameter(
+            torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads)
+        )  # 2*Wh-1 * 2*Ww-1, nH
+
+        # get pair-wise relative position index for each token inside the window
+        coords_h = torch.arange(self.window_size[0])
+        coords_w = torch.arange(self.window_size[1])
+        coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Ww
+        coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
+        relative_coords = (
+            coords_flatten[:, :, None] - coords_flatten[:, None, :]
+        )  # 2, Wh*Ww, Wh*Ww
+        relative_coords = relative_coords.permute(
+            1, 2, 0
+        ).contiguous()  # Wh*Ww, Wh*Ww, 2
+        relative_coords[:, :, 0] += self.window_size[0] - 1  # shift to start from 0
+        relative_coords[:, :, 1] += self.window_size[1] - 1
+        relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
+        relative_position_index = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
+        self.register_buffer("relative_position_index", relative_position_index)
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+        trunc_normal_(self.relative_position_bias_table, std=0.02)
+        self.softmax = nn.Softmax(dim=-1)
+
+    def forward(self, x, mask=None):
+        """
+        Args:
+            x: input features with shape of (num_windows*B, N, C)
+            mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
+        """
+        B_, N, C = x.shape
+        qkv = (
+            self.qkv(x)
+            .reshape(B_, N, 3, self.num_heads, C // self.num_heads)
+            .permute(2, 0, 3, 1, 4)
+        )
+        q, k, v = (
+            qkv[0],
+            qkv[1],
+            qkv[2],
+        )  # make torchscript happy (cannot use tensor as tuple)
+
+        q = q * self.scale
+        attn = q @ k.transpose(-2, -1)
+
+        relative_position_bias = self.relative_position_bias_table[
+            self.relative_position_index.view(-1)
+        ].view(
+            self.window_size[0] * self.window_size[1],
+            self.window_size[0] * self.window_size[1],
+            -1,
+        )  # Wh*Ww,Wh*Ww,nH
+        relative_position_bias = relative_position_bias.permute(
+            2, 0, 1
+        ).contiguous()  # nH, Wh*Ww, Wh*Ww
+        attn = attn + relative_position_bias.unsqueeze(0)
+
+        if mask is not None:
+            nW = mask.shape[0]
+            attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(
+                1
+            ).unsqueeze(0)
+            attn = attn.view(-1, self.num_heads, N, N)
+            attn = self.softmax(attn)
+        else:
+            attn = self.softmax(attn)
+
+        attn = self.attn_drop(attn)
+
+        x = (attn @ v).transpose(1, 2).reshape(B_, N, C)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x, attn
+
+    def extra_repr(self):
+        return f"dim={self.dim}, window_size={self.window_size}, num_heads={self.num_heads}"
+
+
+# We use the model based on Swintransformer Block, therefore we can use the swin-transformer pretrained model
+class SwinTransformerBlock(nn.Module):
+    r"""Swin Transformer Block.
+    Args:
+        dim (int): Number of input channels.
+        input_resolution (tuple[int]): Input resulotion.
+        num_heads (int): Number of attention heads.
+        window_size (int): Window size.
+        shift_size (int): Shift size for SW-MSA.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float, optional): Stochastic depth rate. Default: 0.0
+        act_layer (nn.Module, optional): Activation layer. Default: nn.GELU
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+    """
+
+    def __init__(
+        self,
+        dim,
+        input_resolution,
+        num_heads,
+        window_size=7,
+        shift_size=0,
+        mlp_ratio=4.0,
+        qkv_bias=True,
+        qk_scale=None,
+        drop=0.0,
+        attn_drop=0.0,
+        drop_path=0.0,
+        act_layer=nn.GELU,
+        norm_layer=nn.LayerNorm,
+        norm_before_mlp="ln",
+    ):
+        super().__init__()
+        self.dim = dim
+        self.input_resolution = input_resolution
+        self.num_heads = num_heads
+        self.window_size = window_size
+        self.shift_size = shift_size
+        self.mlp_ratio = mlp_ratio
+        self.norm_before_mlp = norm_before_mlp
+        if min(self.input_resolution) <= self.window_size:
+            # if window size is larger than input resolution, we don't partition windows
+            self.shift_size = 0
+            self.window_size = min(self.input_resolution)
+        assert (
+            0 <= self.shift_size < self.window_size
+        ), "shift_size must in 0-window_size"
+
+        self.norm1 = norm_layer(dim)
+        self.attn = WindowAttention(
+            dim,
+            window_size=to_2tuple(self.window_size),
+            num_heads=num_heads,
+            qkv_bias=qkv_bias,
+            qk_scale=qk_scale,
+            attn_drop=attn_drop,
+            proj_drop=drop,
+        )
+
+        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
+        if self.norm_before_mlp == "ln":
+            self.norm2 = nn.LayerNorm(dim)
+        elif self.norm_before_mlp == "bn":
+            self.norm2 = lambda x: nn.BatchNorm1d(dim)(x.transpose(1, 2)).transpose(
+                1, 2
+            )
+        else:
+            raise NotImplementedError
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(
+            in_features=dim,
+            hidden_features=mlp_hidden_dim,
+            act_layer=act_layer,
+            drop=drop,
+        )
+
+        if self.shift_size > 0:
+            # calculate attention mask for SW-MSA
+            H, W = self.input_resolution
+            img_mask = torch.zeros((1, H, W, 1))  # 1 H W 1
+            h_slices = (
+                slice(0, -self.window_size),
+                slice(-self.window_size, -self.shift_size),
+                slice(-self.shift_size, None),
+            )
+            w_slices = (
+                slice(0, -self.window_size),
+                slice(-self.window_size, -self.shift_size),
+                slice(-self.shift_size, None),
+            )
+            cnt = 0
+            for h in h_slices:
+                for w in w_slices:
+                    img_mask[:, h, w, :] = cnt
+                    cnt += 1
+
+            mask_windows = window_partition(
+                img_mask, self.window_size
+            )  # nW, window_size, window_size, 1
+            mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
+            attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
+            attn_mask = attn_mask.masked_fill(
+                attn_mask != 0, float(-100.0)
+            ).masked_fill(attn_mask == 0, float(0.0))
+        else:
+            attn_mask = None
+
+        self.register_buffer("attn_mask", attn_mask)
+
+    def forward(self, x):
+        # pdb.set_trace()
+        H, W = self.input_resolution
+        # print("H: ", H)
+        # print("W: ", W)
+        # pdb.set_trace()
+        B, L, C = x.shape
+        # assert L == H * W, "input feature has wrong size"
+
+        shortcut = x
+        x = self.norm1(x)
+        x = x.view(B, H, W, C)
+
+        # cyclic shift
+        if self.shift_size > 0:
+            shifted_x = torch.roll(
+                x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2)
+            )
+        else:
+            shifted_x = x
+
+        # partition windows
+        x_windows = window_partition(
+            shifted_x, self.window_size
+        )  # nW*B, window_size, window_size, C
+        x_windows = x_windows.view(
+            -1, self.window_size * self.window_size, C
+        )  # nW*B, window_size*window_size, C
+
+        # W-MSA/SW-MSA
+        attn_windows, attn = self.attn(
+            x_windows, mask=self.attn_mask
+        )  # nW*B, window_size*window_size, C
+
+        # merge windows
+        attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
+        shifted_x = window_reverse(attn_windows, self.window_size, H, W)  # B H' W' C
+
+        # reverse cyclic shift
+        if self.shift_size > 0:
+            x = torch.roll(
+                shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2)
+            )
+        else:
+            x = shifted_x
+        x = x.view(B, H * W, C)
+
+        # FFN
+        x = shortcut + self.drop_path(x)
+        x = x + self.drop_path(self.mlp(self.norm2(x)))
+
+        return x, attn
+
+    def extra_repr(self):
+        return (
+            f"dim={self.dim}, input_resolution={self.input_resolution}, num_heads={self.num_heads}, "
+            f"window_size={self.window_size}, shift_size={self.shift_size}, mlp_ratio={self.mlp_ratio}"
+        )
+
+
+class PatchMerging(nn.Module):
+    r"""Patch Merging Layer.
+    Args:
+        input_resolution (tuple[int]): Resolution of input feature.
+        dim (int): Number of input channels.
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+    """
+
+    def __init__(self, input_resolution, dim, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.input_resolution = input_resolution
+        self.dim = dim
+        self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
+        self.norm = norm_layer(4 * dim)
+
+    def forward(self, x):
+        """
+        x: B, H*W, C
+        """
+        H, W = self.input_resolution
+        B, L, C = x.shape
+        assert L == H * W, "input feature has wrong size"
+        assert H % 2 == 0 and W % 2 == 0, f"x size ({H}*{W}) are not even."
+
+        x = x.view(B, H, W, C)
+
+        x0 = x[:, 0::2, 0::2, :]  # B H/2 W/2 C
+        x1 = x[:, 1::2, 0::2, :]  # B H/2 W/2 C
+        x2 = x[:, 0::2, 1::2, :]  # B H/2 W/2 C
+        x3 = x[:, 1::2, 1::2, :]  # B H/2 W/2 C
+        x = torch.cat([x0, x1, x2, x3], -1)  # B H/2 W/2 4*C
+        x = x.view(B, -1, 4 * C)  # B H/2*W/2 4*C
+
+        x = self.norm(x)
+        x = self.reduction(x)
+
+        return x
+
+    def extra_repr(self):
+        return f"input_resolution={self.input_resolution}, dim={self.dim}"
+
+
+class BasicLayer(nn.Module):
+    """A basic Swin Transformer layer for one stage.
+    Args:
+        dim (int): Number of input channels.
+        input_resolution (tuple[int]): Input resolution.
+        depth (int): Number of blocks.
+        num_heads (int): Number of attention heads.
+        window_size (int): Local window size.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
+        norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
+        downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
+    """
+
+    def __init__(
+        self,
+        dim,
+        input_resolution,
+        depth,
+        num_heads,
+        window_size,
+        mlp_ratio=4.0,
+        qkv_bias=True,
+        qk_scale=None,
+        drop=0.0,
+        attn_drop=0.0,
+        drop_path=0.0,
+        norm_layer=nn.LayerNorm,
+        downsample=None,
+        use_checkpoint=False,
+        norm_before_mlp="ln",
+    ):
+
+        super().__init__()
+        self.dim = dim
+        self.input_resolution = input_resolution
+        self.depth = depth
+        self.use_checkpoint = use_checkpoint
+
+        # build blocks
+        self.blocks = nn.ModuleList(
+            [
+                SwinTransformerBlock(
+                    dim=dim,
+                    input_resolution=input_resolution,
+                    num_heads=num_heads,
+                    window_size=window_size,
+                    shift_size=0 if (i % 2 == 0) else window_size // 2,
+                    mlp_ratio=mlp_ratio,
+                    qkv_bias=qkv_bias,
+                    qk_scale=qk_scale,
+                    drop=drop,
+                    attn_drop=attn_drop,
+                    drop_path=drop_path[i]
+                    if isinstance(drop_path, list)
+                    else drop_path,
+                    norm_layer=norm_layer,
+                    norm_before_mlp=norm_before_mlp,
+                )
+                for i in range(depth)
+            ]
+        )
+
+        # patch merging layer
+        if downsample is not None:
+            self.downsample = downsample(
+                input_resolution, dim=dim, norm_layer=norm_layer
+            )
+        else:
+            self.downsample = None
+
+    def forward(self, x):
+        attns = []
+        for blk in self.blocks:
+            if self.use_checkpoint:
+                x = checkpoint.checkpoint(blk, x)
+            else:
+                x, attn = blk(x)
+                if not self.training:
+                    attns.append(attn.unsqueeze(0))
+        if self.downsample is not None:
+            x = self.downsample(x)
+        if not self.training:
+            attn = torch.cat(attns, dim=0)
+            attn = torch.mean(attn, dim=0)
+        return x, attn
+
+    def extra_repr(self):
+        return f"dim={self.dim}, input_resolution={self.input_resolution}, depth={self.depth}"
+
+
+# The Core of HTSAT
+class HTSAT_Swin_Transformer(nn.Module):
+    r"""HTSAT based on the Swin Transformer
+    Args:
+        spec_size (int | tuple(int)): Input Spectrogram size. Default 256
+        patch_size (int | tuple(int)): Patch size. Default: 4
+        path_stride (iot | tuple(int)): Patch Stride for Frequency and Time Axis. Default: 4
+        in_chans (int): Number of input image channels. Default: 1 (mono)
+        num_classes (int): Number of classes for classification head. Default: 527
+        embed_dim (int): Patch embedding dimension. Default: 96
+        depths (tuple(int)): Depth of each HTSAT-Swin Transformer layer.
+        num_heads (tuple(int)): Number of attention heads in different layers.
+        window_size (int): Window size. Default: 8
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4
+        qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float): Override default qk scale of head_dim ** -0.5 if set. Default: None
+        drop_rate (float): Dropout rate. Default: 0
+        attn_drop_rate (float): Attention dropout rate. Default: 0
+        drop_path_rate (float): Stochastic depth rate. Default: 0.1
+        norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm.
+        ape (bool): If True, add absolute position embedding to the patch embedding. Default: False
+        patch_norm (bool): If True, add normalization after patch embedding. Default: True
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False
+        config (module): The configuration Module from config.py
+    """
+
+    def __init__(
+        self,
+        spec_size=256,
+        patch_size=4,
+        patch_stride=(4, 4),
+        in_chans=1,
+        num_classes=527,
+        embed_dim=96,
+        depths=[2, 2, 6, 2],
+        num_heads=[4, 8, 16, 32],
+        window_size=8,
+        mlp_ratio=4.0,
+        qkv_bias=True,
+        qk_scale=None,
+        drop_rate=0.0,
+        attn_drop_rate=0.0,
+        drop_path_rate=0.1,
+        norm_layer=nn.LayerNorm,
+        ape=False,
+        patch_norm=True,
+        use_checkpoint=False,
+        norm_before_mlp="ln",
+        config=None,
+        enable_fusion=False,
+        fusion_type="None",
+        **kwargs,
+    ):
+        super(HTSAT_Swin_Transformer, self).__init__()
+
+        self.config = config
+        self.spec_size = spec_size
+        self.patch_stride = patch_stride
+        self.patch_size = patch_size
+        self.window_size = window_size
+        self.embed_dim = embed_dim
+        self.depths = depths
+        self.ape = ape
+        self.in_chans = in_chans
+        self.num_classes = num_classes
+        self.num_heads = num_heads
+        self.num_layers = len(self.depths)
+        self.num_features = int(self.embed_dim * 2 ** (self.num_layers - 1))
+
+        self.drop_rate = drop_rate
+        self.attn_drop_rate = attn_drop_rate
+        self.drop_path_rate = drop_path_rate
+
+        self.qkv_bias = qkv_bias
+        self.qk_scale = None
+
+        self.patch_norm = patch_norm
+        self.norm_layer = norm_layer if self.patch_norm else None
+        self.norm_before_mlp = norm_before_mlp
+        self.mlp_ratio = mlp_ratio
+
+        self.use_checkpoint = use_checkpoint
+
+        self.enable_fusion = enable_fusion
+        self.fusion_type = fusion_type
+
+        #  process mel-spec ; used only once
+        self.freq_ratio = self.spec_size // self.config.mel_bins
+        window = "hann"
+        center = True
+        pad_mode = "reflect"
+        ref = 1.0
+        amin = 1e-10
+        top_db = None
+        self.interpolate_ratio = 32  # Downsampled ratio
+        # Spectrogram extractor
+        self.spectrogram_extractor = Spectrogram(
+            n_fft=config.window_size,
+            hop_length=config.hop_size,
+            win_length=config.window_size,
+            window=window,
+            center=center,
+            pad_mode=pad_mode,
+            freeze_parameters=True,
+        )
+        # Logmel feature extractor
+        self.logmel_extractor = LogmelFilterBank(
+            sr=config.sample_rate,
+            n_fft=config.window_size,
+            n_mels=config.mel_bins,
+            fmin=config.fmin,
+            fmax=config.fmax,
+            ref=ref,
+            amin=amin,
+            top_db=top_db,
+            freeze_parameters=True,
+        )
+        # Spec augmenter
+        self.spec_augmenter = SpecAugmentation(
+            time_drop_width=64,
+            time_stripes_num=2,
+            freq_drop_width=8,
+            freq_stripes_num=2,
+        )  # 2 2
+        self.bn0 = nn.BatchNorm2d(self.config.mel_bins)
+
+        # split spctrogram into non-overlapping patches
+        self.patch_embed = PatchEmbed(
+            img_size=self.spec_size,
+            patch_size=self.patch_size,
+            in_chans=self.in_chans,
+            embed_dim=self.embed_dim,
+            norm_layer=self.norm_layer,
+            patch_stride=patch_stride,
+            enable_fusion=self.enable_fusion,
+            fusion_type=self.fusion_type,
+        )
+
+        num_patches = self.patch_embed.num_patches
+        patches_resolution = self.patch_embed.grid_size
+        self.patches_resolution = patches_resolution
+
+        # absolute position embedding
+        if self.ape:
+            self.absolute_pos_embed = nn.Parameter(
+                torch.zeros(1, num_patches, self.embed_dim)
+            )
+            trunc_normal_(self.absolute_pos_embed, std=0.02)
+
+        self.pos_drop = nn.Dropout(p=self.drop_rate)
+
+        # stochastic depth
+        dpr = [
+            x.item() for x in torch.linspace(0, self.drop_path_rate, sum(self.depths))
+        ]  # stochastic depth decay rule
+
+        # build layers
+        self.layers = nn.ModuleList()
+        for i_layer in range(self.num_layers):
+            layer = BasicLayer(
+                dim=int(self.embed_dim * 2**i_layer),
+                input_resolution=(
+                    patches_resolution[0] // (2**i_layer),
+                    patches_resolution[1] // (2**i_layer),
+                ),
+                depth=self.depths[i_layer],
+                num_heads=self.num_heads[i_layer],
+                window_size=self.window_size,
+                mlp_ratio=self.mlp_ratio,
+                qkv_bias=self.qkv_bias,
+                qk_scale=self.qk_scale,
+                drop=self.drop_rate,
+                attn_drop=self.attn_drop_rate,
+                drop_path=dpr[
+                    sum(self.depths[:i_layer]) : sum(self.depths[: i_layer + 1])
+                ],
+                norm_layer=self.norm_layer,
+                downsample=PatchMerging if (i_layer < self.num_layers - 1) else None,
+                use_checkpoint=use_checkpoint,
+                norm_before_mlp=self.norm_before_mlp,
+            )
+            self.layers.append(layer)
+
+        self.norm = self.norm_layer(self.num_features)
+        self.avgpool = nn.AdaptiveAvgPool1d(1)
+        self.maxpool = nn.AdaptiveMaxPool1d(1)
+
+        SF = (
+            self.spec_size
+            // (2 ** (len(self.depths) - 1))
+            // self.patch_stride[0]
+            // self.freq_ratio
+        )
+        self.tscam_conv = nn.Conv2d(
+            in_channels=self.num_features,
+            out_channels=self.num_classes,
+            kernel_size=(SF, 3),
+            padding=(0, 1),
+        )
+        self.head = nn.Linear(num_classes, num_classes)
+
+        if (self.enable_fusion) and (
+            self.fusion_type in ["daf_1d", "aff_1d", "iaff_1d"]
+        ):
+            self.mel_conv1d = nn.Sequential(
+                nn.Conv1d(64, 64, kernel_size=5, stride=3, padding=2),
+                nn.BatchNorm1d(64),
+            )
+            if self.fusion_type == "daf_1d":
+                self.fusion_model = DAF()
+            elif self.fusion_type == "aff_1d":
+                self.fusion_model = AFF(channels=64, type="1D")
+            elif self.fusion_type == "iaff_1d":
+                self.fusion_model = iAFF(channels=64, type="1D")
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=0.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+
+    @torch.jit.ignore
+    def no_weight_decay(self):
+        return {"absolute_pos_embed"}
+
+    @torch.jit.ignore
+    def no_weight_decay_keywords(self):
+        return {"relative_position_bias_table"}
+
+    def forward_features(self, x, longer_idx=None):
+        # A deprecated optimization for using a hierarchical output from different blocks
+
+        frames_num = x.shape[2]
+        x = self.patch_embed(x, longer_idx=longer_idx)
+        if self.ape:
+            x = x + self.absolute_pos_embed
+        x = self.pos_drop(x)
+        for i, layer in enumerate(self.layers):
+            x, attn = layer(x)
+        # for x
+        x = self.norm(x)
+        B, N, C = x.shape
+        SF = frames_num // (2 ** (len(self.depths) - 1)) // self.patch_stride[0]
+        ST = frames_num // (2 ** (len(self.depths) - 1)) // self.patch_stride[1]
+        x = x.permute(0, 2, 1).contiguous().reshape(B, C, SF, ST)
+        B, C, F, T = x.shape
+        # group 2D CNN
+        c_freq_bin = F // self.freq_ratio
+        x = x.reshape(B, C, F // c_freq_bin, c_freq_bin, T)
+        x = x.permute(0, 1, 3, 2, 4).contiguous().reshape(B, C, c_freq_bin, -1)
+        # get latent_output
+        fine_grained_latent_output = torch.mean(x, dim=2)
+        fine_grained_latent_output = interpolate(
+            fine_grained_latent_output.permute(0, 2, 1).contiguous(),
+            8 * self.patch_stride[1],
+        )
+
+        latent_output = self.avgpool(torch.flatten(x, 2))
+        latent_output = torch.flatten(latent_output, 1)
+
+        # display the attention map, if needed
+
+        x = self.tscam_conv(x)
+        x = torch.flatten(x, 2)  # B, C, T
+
+        fpx = interpolate(
+            torch.sigmoid(x).permute(0, 2, 1).contiguous(), 8 * self.patch_stride[1]
+        )
+
+        x = self.avgpool(x)
+        x = torch.flatten(x, 1)
+
+        output_dict = {
+            "framewise_output": fpx,  # already sigmoided
+            "clipwise_output": torch.sigmoid(x),
+            "fine_grained_embedding": fine_grained_latent_output,
+            "embedding": latent_output,
+        }
+
+        return output_dict
+
+    def crop_wav(self, x, crop_size, spe_pos=None):
+        time_steps = x.shape[2]
+        tx = torch.zeros(x.shape[0], x.shape[1], crop_size, x.shape[3]).to(x.device)
+        for i in range(len(x)):
+            if spe_pos is None:
+                crop_pos = random.randint(0, time_steps - crop_size - 1)
+            else:
+                crop_pos = spe_pos
+            tx[i][0] = x[i, 0, crop_pos : crop_pos + crop_size, :]
+        return tx
+
+    # Reshape the wavform to a img size, if you want to use the pretrained swin transformer model
+    def reshape_wav2img(self, x):
+        B, C, T, F = x.shape
+        target_T = int(self.spec_size * self.freq_ratio)
+        target_F = self.spec_size // self.freq_ratio
+        assert (
+            T <= target_T and F <= target_F
+        ), "the wav size should less than or equal to the swin input size"
+        # to avoid bicubic zero error
+        if T < target_T:
+            x = nn.functional.interpolate(
+                x, (target_T, x.shape[3]), mode="bicubic", align_corners=True
+            )
+        if F < target_F:
+            x = nn.functional.interpolate(
+                x, (x.shape[2], target_F), mode="bicubic", align_corners=True
+            )
+        x = x.permute(0, 1, 3, 2).contiguous()
+        x = x.reshape(
+            x.shape[0],
+            x.shape[1],
+            x.shape[2],
+            self.freq_ratio,
+            x.shape[3] // self.freq_ratio,
+        )
+        # print(x.shape)
+        x = x.permute(0, 1, 3, 2, 4).contiguous()
+        x = x.reshape(x.shape[0], x.shape[1], x.shape[2] * x.shape[3], x.shape[4])
+        return x
+
+    # Repeat the wavform to a img size, if you want to use the pretrained swin transformer model
+    def repeat_wat2img(self, x, cur_pos):
+        B, C, T, F = x.shape
+        target_T = int(self.spec_size * self.freq_ratio)
+        target_F = self.spec_size // self.freq_ratio
+        assert (
+            T <= target_T and F <= target_F
+        ), "the wav size should less than or equal to the swin input size"
+        # to avoid bicubic zero error
+        if T < target_T:
+            x = nn.functional.interpolate(
+                x, (target_T, x.shape[3]), mode="bicubic", align_corners=True
+            )
+        if F < target_F:
+            x = nn.functional.interpolate(
+                x, (x.shape[2], target_F), mode="bicubic", align_corners=True
+            )
+        x = x.permute(0, 1, 3, 2).contiguous()  # B C F T
+        x = x[:, :, :, cur_pos : cur_pos + self.spec_size]
+        x = x.repeat(repeats=(1, 1, 4, 1))
+        return x
+
+    def forward(
+        self, x: torch.Tensor, mixup_lambda=None, infer_mode=False, device=None
+    ):  # out_feat_keys: List[str] = None):
+
+        if self.enable_fusion and x["longer"].sum() == 0:
+            # if no audio is longer than 10s, then randomly select one audio to be longer
+            x["longer"][torch.randint(0, x["longer"].shape[0], (1,))] = True
+
+        if not self.enable_fusion:
+            x = x["waveform"].to(device=device, non_blocking=True)
+            x = self.spectrogram_extractor(x)  # (batch_size, 1, time_steps, freq_bins)
+            x = self.logmel_extractor(x)  # (batch_size, 1, time_steps, mel_bins)
+            x = x.transpose(1, 3)
+            x = self.bn0(x)
+            x = x.transpose(1, 3)
+            if self.training:
+                x = self.spec_augmenter(x)
+
+            if self.training and mixup_lambda is not None:
+                x = do_mixup(x, mixup_lambda)
+
+            x = self.reshape_wav2img(x)
+            output_dict = self.forward_features(x)
+        else:
+            longer_list = x["longer"].to(device=device, non_blocking=True)
+            x = x["mel_fusion"].to(device=device, non_blocking=True)
+            x = x.transpose(1, 3)
+            x = self.bn0(x)
+            x = x.transpose(1, 3)
+            longer_list_idx = torch.where(longer_list)[0]
+            if self.fusion_type in ["daf_1d", "aff_1d", "iaff_1d"]:
+                new_x = x[:, 0:1, :, :].clone().contiguous()
+                if len(longer_list_idx) > 0:
+                    # local processing
+                    fusion_x_local = x[longer_list_idx, 1:, :, :].clone().contiguous()
+                    FB, FC, FT, FF = fusion_x_local.size()
+                    fusion_x_local = fusion_x_local.view(FB * FC, FT, FF)
+                    fusion_x_local = torch.permute(
+                        fusion_x_local, (0, 2, 1)
+                    ).contiguous()
+                    fusion_x_local = self.mel_conv1d(fusion_x_local)
+                    fusion_x_local = fusion_x_local.view(
+                        FB, FC, FF, fusion_x_local.size(-1)
+                    )
+                    fusion_x_local = (
+                        torch.permute(fusion_x_local, (0, 2, 1, 3))
+                        .contiguous()
+                        .flatten(2)
+                    )
+                    if fusion_x_local.size(-1) < FT:
+                        fusion_x_local = torch.cat(
+                            [
+                                fusion_x_local,
+                                torch.zeros(
+                                    (FB, FF, FT - fusion_x_local.size(-1)),
+                                    device=device,
+                                ),
+                            ],
+                            dim=-1,
+                        )
+                    else:
+                        fusion_x_local = fusion_x_local[:, :, :FT]
+                    # 1D fusion
+                    new_x = new_x.squeeze(1).permute((0, 2, 1)).contiguous()
+                    new_x[longer_list_idx] = self.fusion_model(
+                        new_x[longer_list_idx], fusion_x_local
+                    )
+                    x = new_x.permute((0, 2, 1)).contiguous()[:, None, :, :]
+                else:
+                    x = new_x
+
+            elif self.fusion_type in ["daf_2d", "aff_2d", "iaff_2d", "channel_map"]:
+                x = x  # no change
+
+            if self.training:
+                x = self.spec_augmenter(x)
+            if self.training and mixup_lambda is not None:
+                x = do_mixup(x, mixup_lambda)
+
+            x = self.reshape_wav2img(x)
+            output_dict = self.forward_features(x, longer_idx=longer_list_idx)
+
+        # if infer_mode:
+        #     # in infer mode. we need to handle different length audio input
+        #     frame_num = x.shape[2]
+        #     target_T = int(self.spec_size * self.freq_ratio)
+        #     repeat_ratio = math.floor(target_T / frame_num)
+        #     x = x.repeat(repeats=(1,1,repeat_ratio,1))
+        #     x = self.reshape_wav2img(x)
+        #     output_dict = self.forward_features(x)
+        # else:
+        #     if x.shape[2] > self.freq_ratio * self.spec_size:
+        #         if self.training:
+        #             x = self.crop_wav(x, crop_size=self.freq_ratio * self.spec_size)
+        #             x = self.reshape_wav2img(x)
+        #             output_dict = self.forward_features(x)
+        #         else:
+        #             # Change: Hard code here
+        #             overlap_size = (x.shape[2] - 1) // 4
+        #             output_dicts = []
+        #             crop_size = (x.shape[2] - 1) // 2
+        #             for cur_pos in range(0, x.shape[2] - crop_size - 1, overlap_size):
+        #                 tx = self.crop_wav(x, crop_size = crop_size, spe_pos = cur_pos)
+        #                 tx = self.reshape_wav2img(tx)
+        #                 output_dicts.append(self.forward_features(tx))
+        #             clipwise_output = torch.zeros_like(output_dicts[0]["clipwise_output"]).float().to(x.device)
+        #             framewise_output = torch.zeros_like(output_dicts[0]["framewise_output"]).float().to(x.device)
+        #             for d in output_dicts:
+        #                 clipwise_output += d["clipwise_output"]
+        #                 framewise_output += d["framewise_output"]
+        #             clipwise_output  = clipwise_output / len(output_dicts)
+        #             framewise_output = framewise_output / len(output_dicts)
+        #             output_dict = {
+        #                 'framewise_output': framewise_output,
+        #                 'clipwise_output': clipwise_output
+        #             }
+        #     else: # this part is typically used, and most easy one
+        #         x = self.reshape_wav2img(x)
+        #         output_dict = self.forward_features(x)
+        # x = self.head(x)
+
+        # We process the data in the dataloader part, in that here we only consider the input_T < fixed_T
+
+        return output_dict
+
+
+def create_htsat_model(audio_cfg, enable_fusion=False, fusion_type="None"):
+    try:
+
+        assert audio_cfg.model_name in [
+            "tiny",
+            "base",
+            "large",
+        ], "model name for HTS-AT is wrong!"
+        if audio_cfg.model_name == "tiny":
+            model = HTSAT_Swin_Transformer(
+                spec_size=256,
+                patch_size=4,
+                patch_stride=(4, 4),
+                num_classes=audio_cfg.class_num,
+                embed_dim=96,
+                depths=[2, 2, 6, 2],
+                num_heads=[4, 8, 16, 32],
+                window_size=8,
+                config=audio_cfg,
+                enable_fusion=enable_fusion,
+                fusion_type=fusion_type,
+            )
+        elif audio_cfg.model_name == "base":
+            model = HTSAT_Swin_Transformer(
+                spec_size=256,
+                patch_size=4,
+                patch_stride=(4, 4),
+                num_classes=audio_cfg.class_num,
+                embed_dim=128,
+                depths=[2, 2, 12, 2],
+                num_heads=[4, 8, 16, 32],
+                window_size=8,
+                config=audio_cfg,
+                enable_fusion=enable_fusion,
+                fusion_type=fusion_type,
+            )
+        elif audio_cfg.model_name == "large":
+            model = HTSAT_Swin_Transformer(
+                spec_size=256,
+                patch_size=4,
+                patch_stride=(4, 4),
+                num_classes=audio_cfg.class_num,
+                embed_dim=256,
+                depths=[2, 2, 12, 2],
+                num_heads=[4, 8, 16, 32],
+                window_size=8,
+                config=audio_cfg,
+                enable_fusion=enable_fusion,
+                fusion_type=fusion_type,
+            )
+
+        return model
+    except:
+        raise RuntimeError(
+            f"Import Model for {audio_cfg.model_name} not found, or the audio cfg parameters are not enough."
+        )

audioldm/clap/open_clip/linear_probe.py

@@ -0,0 +1,66 @@
+import numpy as np
+import torch.nn.functional as F
+from torch import nn
+from .model import MLPLayers
+
+
+class LinearProbe(nn.Module):
+    def __init__(self, model, mlp, freeze, in_ch, out_ch, act=None):
+        """
+        Args:
+            model: nn.Module
+            mlp: bool, if True, then use the MLP layer as the linear probe module
+            freeze: bool, if Ture, then freeze all the CLAP model's layers when training the linear probe
+            in_ch: int, the output channel from CLAP model
+            out_ch: int, the output channel from linear probe (class_num)
+            act: torch.nn.functional, the activation function before the loss function
+        """
+        super().__init__()
+        in_ch = 512
+        self.clap_model = model
+        self.clap_model.text_branch = None  # to save memory
+        self.freeze = freeze
+        if mlp:
+            self.lp_layer = MLPLayers(units=[in_ch, in_ch * 2, out_ch])
+        else:
+            self.lp_layer = nn.Linear(in_ch, out_ch)
+
+        if self.freeze:
+            for param in self.clap_model.parameters():
+                param.requires_grad = False
+
+        if act == "None":
+            self.act = None
+        elif act == "relu":
+            self.act = nn.ReLU()
+        elif act == "elu":
+            self.act = nn.ELU()
+        elif act == "prelu":
+            self.act = nn.PReLU(num_parameters=in_ch)
+        elif act == "softmax":
+            self.act = nn.Softmax(dim=-1)
+        elif act == "sigmoid":
+            self.act = nn.Sigmoid()
+
+    def forward(self, x, mix_lambda=None, device=None):
+        """
+        Args:
+            x: waveform, torch.tensor [batch, t_samples] / batch of mel_spec and longer list
+            mix_lambda: torch.tensor [batch], the mixup lambda
+        Returns:
+            class_prob: torch.tensor [batch, class_num]
+
+        """
+        # batchnorm cancel grandient
+        if self.freeze:
+            self.clap_model.eval()
+
+        x = self.clap_model.audio_projection(
+            self.clap_model.audio_branch(x, mixup_lambda=mix_lambda, device=device)[
+                "embedding"
+            ]
+        )
+        out = self.lp_layer(x)
+        if self.act is not None:
+            out = self.act(out)
+        return out

audioldm/clap/open_clip/loss.py

@@ -0,0 +1,398 @@
+from multiprocessing.sharedctypes import Value
+import torch
+import torch.distributed.nn
+from torch import distributed as dist, nn as nn
+from torch.nn import functional as F
+import numpy as np
+from sklearn.metrics import average_precision_score, roc_auc_score, accuracy_score
+
+try:
+    import horovod.torch as hvd
+except ImportError:
+    hvd = None
+
+
+def gather_features(
+    audio_features,
+    text_features,
+    audio_features_mlp=None,
+    text_features_mlp=None,
+    local_loss=False,
+    gather_with_grad=False,
+    rank=0,
+    world_size=1,
+    use_horovod=False,
+    mlp_loss=False,
+):
+    if use_horovod:
+        assert hvd is not None, "Please install horovod"
+        if gather_with_grad:
+            all_audio_features = hvd.allgather(audio_features)
+            all_text_features = hvd.allgather(text_features)
+            if mlp_loss:
+                all_audio_features_mlp = hvd.allgather(audio_features_mlp)
+                all_text_features_mlp = hvd.allgather(text_features_mlp)
+        else:
+            with torch.no_grad():
+                all_audio_features = hvd.allgather(audio_features)
+                all_text_features = hvd.allgather(text_features)
+                if mlp_loss:
+                    all_audio_features_mlp = hvd.allgather(audio_features_mlp)
+                    all_text_features_mlp = hvd.allgather(text_features_mlp)
+            if not local_loss:
+                # ensure grads for local rank when all_* features don't have a gradient
+                gathered_audio_features = list(
+                    all_audio_features.chunk(world_size, dim=0)
+                )
+                gathered_text_features = list(
+                    all_text_features.chunk(world_size, dim=0)
+                )
+                gathered_audio_features[rank] = audio_features
+                gathered_text_features[rank] = text_features
+                all_audio_features = torch.cat(gathered_audio_features, dim=0)
+                all_text_features = torch.cat(gathered_text_features, dim=0)
+                if mlp_loss:
+                    gathered_audio_features_mlp = list(
+                        all_audio_features_mlp.chunk(world_size, dim=0)
+                    )
+                    gathered_text_features_mlp = list(
+                        all_text_features_mlp.chunk(world_size, dim=0)
+                    )
+                    gathered_audio_features_mlp[rank] = audio_features_mlp
+                    gathered_text_features_mlp[rank] = text_features_mlp
+                    all_audio_features_mlp = torch.cat(
+                        gathered_audio_features_mlp, dim=0
+                    )
+                    all_text_features_mlp = torch.cat(gathered_text_features_mlp, dim=0)
+    else:
+        # We gather tensors from all gpus
+        if gather_with_grad:
+            all_audio_features = torch.cat(
+                torch.distributed.nn.all_gather(audio_features), dim=0
+            )
+            all_text_features = torch.cat(
+                torch.distributed.nn.all_gather(text_features), dim=0
+            )
+            if mlp_loss:
+                all_audio_features_mlp = torch.cat(
+                    torch.distributed.nn.all_gather(audio_features_mlp), dim=0
+                )
+                all_text_features_mlp = torch.cat(
+                    torch.distributed.nn.all_gather(text_features_mlp), dim=0
+                )
+        else:
+            gathered_audio_features = [
+                torch.zeros_like(audio_features) for _ in range(world_size)
+            ]
+            gathered_text_features = [
+                torch.zeros_like(text_features) for _ in range(world_size)
+            ]
+            dist.all_gather(gathered_audio_features, audio_features)
+            dist.all_gather(gathered_text_features, text_features)
+            if mlp_loss:
+                gathered_audio_features_mlp = [
+                    torch.zeros_like(audio_features_mlp) for _ in range(world_size)
+                ]
+                gathered_text_features_mlp = [
+                    torch.zeros_like(text_features_mlp) for _ in range(world_size)
+                ]
+                dist.all_gather(gathered_audio_features_mlp, audio_features_mlp)
+                dist.all_gather(gathered_text_features_mlp, text_features_mlp)
+            if not local_loss:
+                # ensure grads for local rank when all_* features don't have a gradient
+                gathered_audio_features[rank] = audio_features
+                gathered_text_features[rank] = text_features
+                if mlp_loss:
+                    gathered_audio_features_mlp[rank] = audio_features_mlp
+                    gathered_text_features_mlp[rank] = text_features_mlp
+
+            all_audio_features = torch.cat(gathered_audio_features, dim=0)
+            all_text_features = torch.cat(gathered_text_features, dim=0)
+            if mlp_loss:
+                all_audio_features_mlp = torch.cat(gathered_audio_features_mlp, dim=0)
+                all_text_features_mlp = torch.cat(gathered_text_features_mlp, dim=0)
+    if mlp_loss:
+        return (
+            all_audio_features,
+            all_text_features,
+            all_audio_features_mlp,
+            all_text_features_mlp,
+        )
+    else:
+        return all_audio_features, all_text_features
+
+
+class ClipLoss(nn.Module):
+    def __init__(
+        self,
+        local_loss=False,
+        gather_with_grad=False,
+        cache_labels=False,
+        rank=0,
+        world_size=1,
+        use_horovod=False,
+        mlp_loss=False,
+        weight_loss_kappa=0,
+    ):
+        super().__init__()
+        self.local_loss = local_loss
+        self.gather_with_grad = gather_with_grad
+        self.cache_labels = cache_labels
+        self.rank = rank
+        self.world_size = world_size
+        self.use_horovod = use_horovod
+        self.mlp_loss = mlp_loss
+        self.weighted_loss = bool(weight_loss_kappa != 0)
+        self.weight_loss_kappa = weight_loss_kappa
+        # cache state
+        self.prev_num_logits = 0
+        self.labels = {}
+
+    def forward(
+        self,
+        audio_features,
+        text_features,
+        logit_scale_a,
+        logit_scale_t=None,
+        audio_features_mlp=None,
+        text_features_mlp=None,
+    ):
+        device = audio_features.device
+        if self.mlp_loss:
+            if self.world_size > 1:
+                (
+                    all_audio_features,
+                    all_text_features,
+                    all_audio_features_mlp,
+                    all_text_features_mlp,
+                ) = gather_features(
+                    audio_features=audio_features,
+                    text_features=text_features,
+                    audio_features_mlp=audio_features_mlp,
+                    text_features_mlp=text_features_mlp,
+                    local_loss=self.local_loss,
+                    gather_with_grad=self.gather_with_grad,
+                    rank=self.rank,
+                    world_size=self.world_size,
+                    use_horovod=self.use_horovod,
+                    mlp_loss=self.mlp_loss,
+                )
+                if self.local_loss:
+                    a_logits_per_audio = (
+                        logit_scale_a * audio_features @ all_text_features_mlp.T
+                    )
+                    a_logits_per_text = (
+                        logit_scale_a * text_features_mlp @ all_audio_features.T
+                    )
+                    t_logits_per_audio = (
+                        logit_scale_t * audio_features_mlp @ all_text_features.T
+                    )
+                    t_logits_per_text = (
+                        logit_scale_t * text_features @ all_audio_features_mlp.T
+                    )
+                else:
+                    a_logits_per_audio = (
+                        logit_scale_a * all_audio_features @ all_text_features_mlp.T
+                    )
+                    a_logits_per_text = a_logits_per_audio.T
+                    t_logits_per_audio = (
+                        logit_scale_t * all_audio_features_mlp @ all_text_features.T
+                    )
+                    t_logits_per_text = t_logits_per_audio.T
+            else:
+                a_logits_per_audio = (
+                    logit_scale_a * audio_features @ text_features_mlp.T
+                )
+                a_logits_per_text = logit_scale_a * text_features_mlp @ audio_features.T
+                t_logits_per_audio = (
+                    logit_scale_t * audio_features_mlp @ text_features.T
+                )
+                t_logits_per_text = logit_scale_t * text_features @ audio_features_mlp.T
+
+            # calculated ground-truth and cache if enabled
+            num_logits = a_logits_per_audio.shape[0]
+            if self.prev_num_logits != num_logits or device not in self.labels:
+                labels = torch.arange(num_logits, device=device, dtype=torch.long)
+                if self.world_size > 1 and self.local_loss:
+                    labels = labels + num_logits * self.rank
+                if self.cache_labels:
+                    self.labels[device] = labels
+                    self.prev_num_logits = num_logits
+            else:
+                labels = self.labels[device]
+
+            if not self.weighted_loss:
+                total_loss = (
+                    F.cross_entropy(a_logits_per_audio, labels)
+                    + F.cross_entropy(a_logits_per_text, labels)
+                    + F.cross_entropy(t_logits_per_audio, labels)
+                    + F.cross_entropy(t_logits_per_text, labels)
+                ) / 4
+            else:
+                audio_weight = (audio_features @ audio_features.T).detach()
+                audio_weight = (
+                    torch.exp(
+                        torch.sum(audio_weight, axis=1)
+                        / (self.weight_loss_kappa * len(audio_weight))
+                    )
+                ).detach()
+                text_weight = (text_features @ text_features.T).detach()
+                text_weight = (
+                    torch.exp(
+                        torch.sum(text_weight, axis=1)
+                        / (self.weight_loss_kappa * len(text_features))
+                    )
+                ).detach()
+                total_loss = (
+                    F.cross_entropy(a_logits_per_audio, labels, weight=audio_weight)
+                    + F.cross_entropy(a_logits_per_text, labels, weight=audio_weight)
+                    + F.cross_entropy(t_logits_per_audio, labels, weight=text_weight)
+                    + F.cross_entropy(t_logits_per_text, labels, weight=text_weight)
+                ) / 4
+        else:
+            if self.world_size > 1:
+                all_audio_features, all_text_features = gather_features(
+                    audio_features=audio_features,
+                    text_features=text_features,
+                    local_loss=self.local_loss,
+                    gather_with_grad=self.gather_with_grad,
+                    rank=self.rank,
+                    world_size=self.world_size,
+                    use_horovod=self.use_horovod,
+                    mlp_loss=self.mlp_loss,
+                )
+
+                if self.local_loss:
+                    logits_per_audio = (
+                        logit_scale_a * audio_features @ all_text_features.T
+                    )
+                    logits_per_text = (
+                        logit_scale_a * text_features @ all_audio_features.T
+                    )
+                else:
+                    logits_per_audio = (
+                        logit_scale_a * all_audio_features @ all_text_features.T
+                    )
+                    logits_per_text = logits_per_audio.T
+            else:
+                logits_per_audio = logit_scale_a * audio_features @ text_features.T
+                logits_per_text = logit_scale_a * text_features @ audio_features.T
+
+            # calculated ground-truth and cache if enabled
+            num_logits = logits_per_audio.shape[0]
+            if self.prev_num_logits != num_logits or device not in self.labels:
+                labels = torch.arange(num_logits, device=device, dtype=torch.long)
+                if self.world_size > 1 and self.local_loss:
+                    labels = labels + num_logits * self.rank
+                if self.cache_labels:
+                    self.labels[device] = labels
+                    self.prev_num_logits = num_logits
+            else:
+                labels = self.labels[device]
+            if not self.weighted_loss:
+                total_loss = (
+                    F.cross_entropy(logits_per_audio, labels)
+                    + F.cross_entropy(logits_per_text, labels)
+                ) / 2
+            else:
+                audio_weight = (all_audio_features @ all_audio_features.T).detach()
+                audio_weight = (
+                    torch.exp(
+                        torch.sum(audio_weight, axis=1)
+                        / (self.weight_loss_kappa * len(all_audio_features))
+                    )
+                ).detach()
+                text_weight = (all_text_features @ all_text_features.T).detach()
+                text_weight = (
+                    torch.exp(
+                        torch.sum(text_weight, axis=1)
+                        / (self.weight_loss_kappa * len(all_text_features))
+                    )
+                ).detach()
+                total_loss = (
+                    F.cross_entropy(logits_per_audio, labels, weight=text_weight)
+                    + F.cross_entropy(logits_per_text, labels, weight=audio_weight)
+                ) / 2
+        return total_loss
+
+
+def lp_gather_features(pred, target, world_size=1, use_horovod=False):
+    if use_horovod:
+        assert hvd is not None, "Please install horovod"
+        with torch.no_grad():
+            all_preds = hvd.allgather(pred)
+            all_targets = hvd.allgath(target)
+    else:
+        gathered_preds = [torch.zeros_like(pred) for _ in range(world_size)]
+        gathered_targets = [torch.zeros_like(target) for _ in range(world_size)]
+
+        dist.all_gather(gathered_preds, pred)
+        dist.all_gather(gathered_targets, target)
+        all_preds = torch.cat(gathered_preds, dim=0)
+        all_targets = torch.cat(gathered_targets, dim=0)
+
+    return all_preds, all_targets
+
+
+def get_map(pred, target):
+    pred = torch.sigmoid(pred).numpy()
+    target = target.numpy()
+    return np.mean(average_precision_score(target, pred, average=None))
+
+
+def get_acc(pred, target):
+    pred = torch.argmax(pred, 1).numpy()
+    target = torch.argmax(target, 1).numpy()
+    return accuracy_score(target, pred)
+
+
+def get_mauc(pred, target):
+    pred = torch.sigmoid(pred).numpy()
+    target = target.numpy()
+    return np.mean(roc_auc_score(target, pred, average=None))
+
+
+class LPMetrics(object):
+    def __init__(self, metric_names=["map", "acc", "mauc"]):
+        self.metrics = []
+        for name in metric_names:
+            self.metrics.append(self.get_metric(name))
+        self.metric_names = metric_names
+
+    def get_metric(self, name):
+        if name == "map":
+            return get_map
+        elif name == "acc":
+            return get_acc
+        elif name == "mauc":
+            return get_mauc
+        else:
+            raise ValueError(f"the metric should be at least one of [map, acc, mauc]")
+
+    def evaluate_mertics(self, pred, target):
+        metric_dict = {}
+        for i in range(len(self.metric_names)):
+            metric_dict[self.metric_names[i]] = self.metrics[i](pred, target)
+        return metric_dict
+
+
+def calc_celoss(pred, target):
+    target = torch.argmax(target, 1).long()
+    return nn.CrossEntropyLoss()(pred, target)
+
+
+class LPLoss(nn.Module):
+    def __init__(self, loss_name):
+        super().__init__()
+        if loss_name == "bce":
+            self.loss_func = nn.BCEWithLogitsLoss()
+        elif loss_name == "ce":
+            self.loss_func = calc_celoss
+        elif loss_name == "mse":
+            self.loss_func = nn.MSELoss()
+        else:
+            raise ValueError(f"the loss func should be at least one of [bce, ce, mse]")
+
+    def forward(self, pred, target):
+        loss = self.loss_func(pred, target)
+        return loss

audioldm/clap/open_clip/model.py

@@ -0,0 +1,936 @@
+""" CLAP Model
+
+Adapted from CLIP: https://github.com/openai/CLIP. Originally MIT License, Copyright (c) 2021 OpenAI.
+Adapted to the Audio Task.
+"""
+
+from collections import OrderedDict
+from dataclasses import dataclass
+from email.mime import audio
+from typing import Tuple, Union, Callable, Optional
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from .timm_model import TimmModel
+import logging
+from .utils import freeze_batch_norm_2d
+
+from .pann_model import create_pann_model
+from .htsat import create_htsat_model
+from transformers import BertModel, RobertaModel, BartModel
+from transformers.tokenization_utils_base import BatchEncoding
+
+
+class MLPLayers(nn.Module):
+    def __init__(self, units=[512, 512, 512], nonlin=nn.ReLU(), dropout=0.1):
+        super(MLPLayers, self).__init__()
+        self.nonlin = nonlin
+        self.dropout = dropout
+
+        sequence = []
+        for u0, u1 in zip(units[:-1], units[1:]):
+            sequence.append(nn.Linear(u0, u1))
+            sequence.append(self.nonlin)
+            sequence.append(nn.Dropout(self.dropout))
+        sequence = sequence[:-2]
+
+        self.sequential = nn.Sequential(*sequence)
+
+    def forward(self, X):
+        X = self.sequential(X)
+        return X
+
+
+class Bottleneck(nn.Module):
+    expansion = 4
+
+    def __init__(self, inplanes, planes, stride=1):
+        super().__init__()
+
+        # all conv layers have stride 1. an avgpool is performed after the second convolution when stride > 1
+        self.conv1 = nn.Conv2d(inplanes, planes, 1, bias=False)
+        self.bn1 = nn.BatchNorm2d(planes)
+
+        self.conv2 = nn.Conv2d(planes, planes, 3, padding=1, bias=False)
+        self.bn2 = nn.BatchNorm2d(planes)
+
+        self.avgpool = nn.AvgPool2d(stride) if stride > 1 else nn.Identity()
+
+        self.conv3 = nn.Conv2d(planes, planes * self.expansion, 1, bias=False)
+        self.bn3 = nn.BatchNorm2d(planes * self.expansion)
+
+        self.relu = nn.ReLU(inplace=True)
+        self.downsample = None
+        self.stride = stride
+
+        if stride > 1 or inplanes != planes * Bottleneck.expansion:
+            # downsampling layer is prepended with an avgpool, and the subsequent convolution has stride 1
+            self.downsample = nn.Sequential(
+                OrderedDict(
+                    [
+                        ("-1", nn.AvgPool2d(stride)),
+                        (
+                            "0",
+                            nn.Conv2d(
+                                inplanes,
+                                planes * self.expansion,
+                                1,
+                                stride=1,
+                                bias=False,
+                            ),
+                        ),
+                        ("1", nn.BatchNorm2d(planes * self.expansion)),
+                    ]
+                )
+            )
+
+    def forward(self, x: torch.Tensor):
+        identity = x
+
+        out = self.relu(self.bn1(self.conv1(x)))
+        out = self.relu(self.bn2(self.conv2(out)))
+        out = self.avgpool(out)
+        out = self.bn3(self.conv3(out))
+
+        if self.downsample is not None:
+            identity = self.downsample(x)
+
+        out += identity
+        out = self.relu(out)
+        return out
+
+
+class AttentionPool2d(nn.Module):
+    def __init__(
+        self, spacial_dim: int, embed_dim: int, num_heads: int, output_dim: int = None
+    ):
+        super().__init__()
+        self.positional_embedding = nn.Parameter(
+            torch.randn(spacial_dim**2 + 1, embed_dim) / embed_dim**0.5
+        )
+        self.k_proj = nn.Linear(embed_dim, embed_dim)
+        self.q_proj = nn.Linear(embed_dim, embed_dim)
+        self.v_proj = nn.Linear(embed_dim, embed_dim)
+        self.c_proj = nn.Linear(embed_dim, output_dim or embed_dim)
+        self.num_heads = num_heads
+
+    def forward(self, x):
+        x = x.reshape(x.shape[0], x.shape[1], x.shape[2] * x.shape[3]).permute(
+            2, 0, 1
+        )  # NCHW -> (HW)NC
+        x = torch.cat([x.mean(dim=0, keepdim=True), x], dim=0)  # (HW+1)NC
+        x = x + self.positional_embedding[:, None, :].to(x.dtype)  # (HW+1)NC
+        x, _ = F.multi_head_attention_forward(
+            query=x,
+            key=x,
+            value=x,
+            embed_dim_to_check=x.shape[-1],
+            num_heads=self.num_heads,
+            q_proj_weight=self.q_proj.weight,
+            k_proj_weight=self.k_proj.weight,
+            v_proj_weight=self.v_proj.weight,
+            in_proj_weight=None,
+            in_proj_bias=torch.cat(
+                [self.q_proj.bias, self.k_proj.bias, self.v_proj.bias]
+            ),
+            bias_k=None,
+            bias_v=None,
+            add_zero_attn=False,
+            dropout_p=0,
+            out_proj_weight=self.c_proj.weight,
+            out_proj_bias=self.c_proj.bias,
+            use_separate_proj_weight=True,
+            training=self.training,
+            need_weights=False,
+        )
+
+        return x[0]
+
+
+class ModifiedResNet(nn.Module):
+    """
+    A ResNet class that is similar to torchvision's but contains the following changes:
+    - There are now 3 "stem" convolutions as opposed to 1, with an average pool instead of a max pool.
+    - Performs anti-aliasing strided convolutions, where an avgpool is prepended to convolutions with stride > 1
+    - The final pooling layer is a QKV attention instead of an average pool
+    """
+
+    def __init__(self, layers, output_dim, heads, image_size=224, width=64):
+        super().__init__()
+        self.output_dim = output_dim
+        self.image_size = image_size
+
+        # the 3-layer stem
+        self.conv1 = nn.Conv2d(
+            3, width // 2, kernel_size=3, stride=2, padding=1, bias=False
+        )
+        self.bn1 = nn.BatchNorm2d(width // 2)
+        self.conv2 = nn.Conv2d(
+            width // 2, width // 2, kernel_size=3, padding=1, bias=False
+        )
+        self.bn2 = nn.BatchNorm2d(width // 2)
+        self.conv3 = nn.Conv2d(width // 2, width, kernel_size=3, padding=1, bias=False)
+        self.bn3 = nn.BatchNorm2d(width)
+        self.avgpool = nn.AvgPool2d(2)
+        self.relu = nn.ReLU(inplace=True)
+
+        # residual layers
+        self._inplanes = width  # this is a *mutable* variable used during construction
+        self.layer1 = self._make_layer(width, layers[0])
+        self.layer2 = self._make_layer(width * 2, layers[1], stride=2)
+        self.layer3 = self._make_layer(width * 4, layers[2], stride=2)
+        self.layer4 = self._make_layer(width * 8, layers[3], stride=2)
+
+        embed_dim = width * 32  # the ResNet feature dimension
+        self.attnpool = AttentionPool2d(image_size // 32, embed_dim, heads, output_dim)
+
+        self.init_parameters()
+
+    def _make_layer(self, planes, blocks, stride=1):
+        layers = [Bottleneck(self._inplanes, planes, stride)]
+
+        self._inplanes = planes * Bottleneck.expansion
+        for _ in range(1, blocks):
+            layers.append(Bottleneck(self._inplanes, planes))
+
+        return nn.Sequential(*layers)
+
+    def init_parameters(self):
+        if self.attnpool is not None:
+            std = self.attnpool.c_proj.in_features**-0.5
+            nn.init.normal_(self.attnpool.q_proj.weight, std=std)
+            nn.init.normal_(self.attnpool.k_proj.weight, std=std)
+            nn.init.normal_(self.attnpool.v_proj.weight, std=std)
+            nn.init.normal_(self.attnpool.c_proj.weight, std=std)
+
+        for resnet_block in [self.layer1, self.layer2, self.layer3, self.layer4]:
+            for name, param in resnet_block.named_parameters():
+                if name.endswith("bn3.weight"):
+                    nn.init.zeros_(param)
+
+    def lock(self, unlocked_groups=0, freeze_bn_stats=False):
+        assert (
+            unlocked_groups == 0
+        ), "partial locking not currently supported for this model"
+        for param in self.parameters():
+            param.requires_grad = False
+        if freeze_bn_stats:
+            freeze_batch_norm_2d(self)
+
+    def stem(self, x):
+        for conv, bn in [
+            (self.conv1, self.bn1),
+            (self.conv2, self.bn2),
+            (self.conv3, self.bn3),
+        ]:
+            x = self.relu(bn(conv(x)))
+        x = self.avgpool(x)
+        return x
+
+    def forward(self, x):
+        x = self.stem(x)
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = self.layer4(x)
+        x = self.attnpool(x)
+
+        return x
+
+
+class LayerNorm(nn.LayerNorm):
+    """Subclass torch's LayerNorm to handle fp16."""
+
+    def forward(self, x: torch.Tensor):
+        orig_type = x.dtype
+        x = F.layer_norm(x, self.normalized_shape, self.weight, self.bias, self.eps)
+        return x.to(orig_type)
+
+
+class QuickGELU(nn.Module):
+    # NOTE This is slower than nn.GELU or nn.SiLU and uses more GPU memory
+    def forward(self, x: torch.Tensor):
+        return x * torch.sigmoid(1.702 * x)
+
+
+class ResidualAttentionBlock(nn.Module):
+    def __init__(self, d_model: int, n_head: int, act_layer: Callable = nn.GELU):
+        super().__init__()
+
+        self.attn = nn.MultiheadAttention(d_model, n_head)
+        self.ln_1 = LayerNorm(d_model)
+        self.mlp = nn.Sequential(
+            OrderedDict(
+                [
+                    ("c_fc", nn.Linear(d_model, d_model * 4)),
+                    ("gelu", act_layer()),
+                    ("c_proj", nn.Linear(d_model * 4, d_model)),
+                ]
+            )
+        )
+        self.ln_2 = LayerNorm(d_model)
+
+    def attention(self, x: torch.Tensor, attn_mask: Optional[torch.Tensor] = None):
+        return self.attn(x, x, x, need_weights=False, attn_mask=attn_mask)[0]
+
+    def forward(self, x: torch.Tensor, attn_mask: Optional[torch.Tensor] = None):
+        x = x + self.attention(self.ln_1(x), attn_mask=attn_mask)
+        x = x + self.mlp(self.ln_2(x))
+        return x
+
+
+class Transformer(nn.Module):
+    def __init__(
+        self, width: int, layers: int, heads: int, act_layer: Callable = nn.GELU
+    ):
+        super().__init__()
+        self.width = width
+        self.layers = layers
+        self.resblocks = nn.ModuleList(
+            [
+                ResidualAttentionBlock(width, heads, act_layer=act_layer)
+                for _ in range(layers)
+            ]
+        )
+
+    def forward(self, x: torch.Tensor, attn_mask: Optional[torch.Tensor] = None):
+        for r in self.resblocks:
+            x = r(x, attn_mask=attn_mask)
+        return x
+
+
+class VisualTransformer(nn.Module):
+    def __init__(
+        self,
+        image_size: int,
+        patch_size: int,
+        width: int,
+        layers: int,
+        heads: int,
+        output_dim: int,
+        act_layer: Callable = nn.GELU,
+    ):
+        super().__init__()
+        self.image_size = image_size
+        self.output_dim = output_dim
+        self.conv1 = nn.Conv2d(
+            in_channels=3,
+            out_channels=width,
+            kernel_size=patch_size,
+            stride=patch_size,
+            bias=False,
+        )
+
+        scale = width**-0.5
+        self.class_embedding = nn.Parameter(scale * torch.randn(width))
+        self.positional_embedding = nn.Parameter(
+            scale * torch.randn((image_size // patch_size) ** 2 + 1, width)
+        )
+        self.ln_pre = LayerNorm(width)
+
+        self.text_branch = Transformer(width, layers, heads, act_layer=act_layer)
+
+        self.ln_post = LayerNorm(width)
+        self.proj = nn.Parameter(scale * torch.randn(width, output_dim))
+
+    def lock(self, unlocked_groups=0, freeze_bn_stats=False):
+        assert (
+            unlocked_groups == 0
+        ), "partial locking not currently supported for this model"
+        for param in self.parameters():
+            param.requires_grad = False
+
+    def forward(self, x: torch.Tensor):
+        x = self.conv1(x)  # shape = [*, width, grid, grid]
+        x = x.reshape(x.shape[0], x.shape[1], -1)  # shape = [*, width, grid ** 2]
+        x = x.permute(0, 2, 1)  # shape = [*, grid ** 2, width]
+        x = torch.cat(
+            [
+                self.class_embedding.to(x.dtype)
+                + torch.zeros(
+                    x.shape[0], 1, x.shape[-1], dtype=x.dtype, device=x.device
+                ),
+                x,
+            ],
+            dim=1,
+        )  # shape = [*, grid ** 2 + 1, width]
+        x = x + self.positional_embedding.to(x.dtype)
+        x = self.ln_pre(x)
+
+        x = x.permute(1, 0, 2)  # NLD -> LND
+        x = self.text_branch(x)
+        x = x.permute(1, 0, 2)  # LND -> NLD
+
+        x = self.ln_post(x[:, 0, :])
+
+        if self.proj is not None:
+            x = x @ self.proj
+
+        return x
+
+
+@dataclass
+class CLAPVisionCfg:
+    layers: Union[Tuple[int, int, int, int], int] = 12
+    width: int = 768
+    patch_size: int = 16
+    image_size: Union[Tuple[int, int], int] = 224
+    timm_model_name: str = (
+        None  # a valid model name overrides layers, width, patch_size
+    )
+    timm_model_pretrained: bool = (
+        False  # use (imagenet) pretrained weights for named model
+    )
+    timm_pool: str = (
+        "avg"  # feature pooling for timm model ('abs_attn', 'rot_attn', 'avg', '')
+    )
+    timm_proj: str = (
+        "linear"  # linear projection for timm model output ('linear', 'mlp', '')
+    )
+
+
+# Audio Config Class
+@dataclass
+class CLAPAudioCfp:
+    model_type: str = "PANN"
+    model_name: str = "Cnn14"
+    sample_rate: int = 48000
+    # Param
+    audio_length: int = 1024
+    window_size: int = 1024
+    hop_size: int = 1024
+    fmin: int = 50
+    fmax: int = 14000
+    class_num: int = 527
+    mel_bins: int = 64
+    clip_samples: int = 480000
+
+
+@dataclass
+class CLAPTextCfg:
+    context_length: int
+    vocab_size: int
+    width: int
+    heads: int
+    layers: int
+    model_type: str
+
+
+class CLAP(nn.Module):
+    def __init__(
+        self,
+        embed_dim: int,
+        audio_cfg: CLAPAudioCfp,
+        text_cfg: CLAPTextCfg,
+        quick_gelu: bool = False,
+        enable_fusion: bool = False,
+        fusion_type: str = "None",
+        joint_embed_shape: int = 512,
+        mlp_act: str = "relu",
+    ):
+        super().__init__()
+        if isinstance(audio_cfg, dict):
+            audio_cfg = CLAPAudioCfp(**audio_cfg)
+        if isinstance(text_cfg, dict):
+            text_cfg = CLAPTextCfg(**text_cfg)
+
+        self.audio_cfg = audio_cfg
+        self.text_cfg = text_cfg
+        self.enable_fusion = enable_fusion
+        self.fusion_type = fusion_type
+        self.joint_embed_shape = joint_embed_shape
+        self.mlp_act = mlp_act
+
+        self.context_length = text_cfg.context_length
+
+        # OpenAI models are pretrained w/ QuickGELU but native nn.GELU is both faster and more
+        # memory efficient in recent PyTorch releases (>= 1.10).
+        # NOTE: timm models always use native GELU regardless of quick_gelu flag.
+        act_layer = QuickGELU if quick_gelu else nn.GELU
+
+        if mlp_act == "relu":
+            mlp_act_layer = nn.ReLU()
+        elif mlp_act == "gelu":
+            mlp_act_layer = nn.GELU()
+        else:
+            raise NotImplementedError
+
+        # audio branch
+        # audio branch parameters
+        if audio_cfg.model_type == "PANN":
+            self.audio_branch = create_pann_model(audio_cfg, enable_fusion, fusion_type)
+        elif audio_cfg.model_type == "HTSAT":
+            self.audio_branch = create_htsat_model(
+                audio_cfg, enable_fusion, fusion_type
+            )
+        else:
+            logging.error(f"Model config for {audio_cfg.model_type} not found")
+            raise RuntimeError(f"Model config for {audio_cfg.model_type} not found.")
+
+        # text branch
+        # text branch parameters
+        if text_cfg.model_type == "transformer":
+            self.text_branch = Transformer(
+                width=text_cfg.width,
+                layers=text_cfg.layers,
+                heads=text_cfg.heads,
+                act_layer=act_layer,
+            )
+            self.vocab_size = text_cfg.vocab_size
+            self.token_embedding = nn.Embedding(text_cfg.vocab_size, text_cfg.width)
+            self.positional_embedding = nn.Parameter(
+                torch.empty(self.context_length, text_cfg.width)
+            )
+            self.ln_final = LayerNorm(text_cfg.width)
+            self.text_transform = MLPLayers(
+                units=[
+                    self.joint_embed_shape,
+                    self.joint_embed_shape,
+                    self.joint_embed_shape,
+                ],
+                dropout=0.1,
+            )
+            self.text_projection = nn.Sequential(
+                nn.Linear(text_cfg.width, self.joint_embed_shape),
+                mlp_act_layer,
+                nn.Linear(self.joint_embed_shape, self.joint_embed_shape),
+            )
+        elif text_cfg.model_type == "bert":
+            self.text_branch = BertModel.from_pretrained("bert-base-uncased")
+            self.text_transform = MLPLayers(
+                units=[
+                    self.joint_embed_shape,
+                    self.joint_embed_shape,
+                    self.joint_embed_shape,
+                ],
+                dropout=0.1,
+            )
+            self.text_projection = nn.Sequential(
+                nn.Linear(768, self.joint_embed_shape),
+                mlp_act_layer,
+                nn.Linear(self.joint_embed_shape, self.joint_embed_shape),
+            )
+        elif text_cfg.model_type == "roberta":
+            self.text_branch = RobertaModel.from_pretrained("roberta-base")
+            self.text_transform = MLPLayers(
+                units=[
+                    self.joint_embed_shape,
+                    self.joint_embed_shape,
+                    self.joint_embed_shape,
+                ],
+                dropout=0.1,
+            )
+            self.text_projection = nn.Sequential(
+                nn.Linear(768, self.joint_embed_shape),
+                mlp_act_layer,
+                nn.Linear(self.joint_embed_shape, self.joint_embed_shape),
+            )
+        elif text_cfg.model_type == "bart":
+            self.text_branch = BartModel.from_pretrained("facebook/bart-base")
+            self.text_transform = MLPLayers(
+                units=[
+                    self.joint_embed_shape,
+                    self.joint_embed_shape,
+                    self.joint_embed_shape,
+                ],
+                dropout=0.1,
+            )
+            self.text_projection = nn.Sequential(
+                nn.Linear(768, self.joint_embed_shape),
+                mlp_act_layer,
+                nn.Linear(self.joint_embed_shape, self.joint_embed_shape),
+            )
+        else:
+            logging.error(f"Model config for {text_cfg.model_type} not found")
+            raise RuntimeError(f"Model config for {text_cfg.model_type} not found.")
+        self.text_branch_type = text_cfg.model_type
+        # text branch parameters
+
+        # audio branch parameters
+        self.audio_transform = MLPLayers(
+            units=[
+                self.joint_embed_shape,
+                self.joint_embed_shape,
+                self.joint_embed_shape,
+            ],
+            dropout=0.1,
+        )
+
+        # below here is text branch parameters
+
+        # ============================================================================================================
+        self.audio_projection = nn.Sequential(
+            nn.Linear(embed_dim, self.joint_embed_shape),
+            mlp_act_layer,
+            nn.Linear(self.joint_embed_shape, self.joint_embed_shape),
+        )
+
+        self.logit_scale_a = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))
+        self.logit_scale_t = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))
+        self.register_buffer("attn_mask", self.build_attention_mask(), persistent=False)
+
+        self.init_text_branch_parameters()
+
+    def init_text_branch_parameters(self):
+        if self.text_branch_type == "transformer":
+            nn.init.normal_(self.token_embedding.weight, std=0.02)
+            nn.init.normal_(self.positional_embedding, std=0.01)
+            proj_std = (self.text_branch.width**-0.5) * (
+                (2 * self.text_branch.layers) ** -0.5
+            )
+            attn_std = self.text_branch.width**-0.5
+            fc_std = (2 * self.text_branch.width) ** -0.5
+            for block in self.text_branch.resblocks:
+                nn.init.normal_(block.attn.in_proj_weight, std=attn_std)
+                nn.init.normal_(block.attn.out_proj.weight, std=proj_std)
+                nn.init.normal_(block.mlp.c_fc.weight, std=fc_std)
+                nn.init.normal_(block.mlp.c_proj.weight, std=proj_std)
+        if self.text_branch_type == "bert" or self.text_branch_type == "roberta":
+            width = self.text_branch.embeddings.word_embeddings.weight.shape[-1]
+        elif self.text_branch_type == "bart":
+            width = self.text_branch.shared.weight.shape[-1]
+        else:
+            width = self.text_branch.width
+        nn.init.constant_(self.logit_scale_a, np.log(1 / 0.07))
+        nn.init.constant_(self.logit_scale_t, np.log(1 / 0.07))
+
+        # deprecated
+        # if hasattr(self.visual, 'init_parameters'):
+        # self.visual.init_parameters()
+
+        # if self.text_projection is not None:
+        #     nn.init.normal_(self.text_projection, std=width**-0.5)
+
+    def build_attention_mask(self):
+        # lazily create causal attention mask, with full attention between the vision tokens
+        # pytorch uses additive attention mask; fill with -inf
+        mask = torch.empty(self.context_length, self.context_length)
+        mask.fill_(float("-inf"))
+        mask.triu_(1)  # zero out the lower diagonal
+        return mask
+
+    def encode_audio(self, audio, device):
+        return self.audio_branch(
+            audio, mixup_lambda=None, device=device
+        )  # mix lambda needs to add
+
+    # def list_of_dict_of_tensor2dict_of_tensor(self, x, device):
+    #     tmp = {}
+    #     for k in x[0].keys():
+    #         tmp[k] = []
+    #         for i in range(len(x)):
+    #             tmp[k].append(x[i][k][:77])
+    #     for k in x[0].keys():
+    #         tmp[k] = torch.tensor(tmp[k]).to(device=device, non_blocking=True)
+    #     return tmp
+
+    def encode_text(self, text, device):
+        if self.text_branch_type == "transformer":
+            text = text.to(device=device, non_blocking=True)
+            x = self.token_embedding(text)  # [batch_size, n_ctx, d_model]
+
+            x = x + self.positional_embedding
+            x = x.permute(1, 0, 2)  # NLD -> LND
+            x = self.text_branch(x, attn_mask=self.attn_mask)
+            x = x.permute(1, 0, 2)  # LND -> NLD
+            x = self.ln_final(x)
+
+            # x.shape = [batch_size, n_ctx, transformer.width]
+            # take features from the eot embedding (eot_token is the highest number in each sequence)
+            x = self.text_projection(x[torch.arange(x.shape[0]), text.argmax(dim=-1)])
+        elif self.text_branch_type == "bert":
+            # text = self.list_of_dict_of_tensor2dict_of_tensor(text, device)
+            # text = BatchEncoding(text)
+            x = self.text_branch(
+                input_ids=text["input_ids"].to(device=device, non_blocking=True),
+                attention_mask=text["attention_mask"].to(
+                    device=device, non_blocking=True
+                ),
+                token_type_ids=text["token_type_ids"].to(
+                    device=device, non_blocking=True
+                ),
+            )["pooler_output"]
+            x = self.text_projection(x)
+        elif self.text_branch_type == "roberta":
+            x = self.text_branch(
+                input_ids=text["input_ids"].to(device=device, non_blocking=True),
+                attention_mask=text["attention_mask"].to(
+                    device=device, non_blocking=True
+                ),
+            )["pooler_output"]
+            x = self.text_projection(x)
+        elif self.text_branch_type == "bart":
+            x = torch.mean(
+                self.text_branch(
+                    input_ids=text["input_ids"].to(device=device, non_blocking=True),
+                    attention_mask=text["attention_mask"].to(
+                        device=device, non_blocking=True
+                    ),
+                )["encoder_last_hidden_state"],
+                axis=1,
+            )
+            x = self.text_projection(x)
+        else:
+            logging.error(f"Model type {self.text_branch_type} not found")
+            raise RuntimeError(f"Model type {self.text_branch_type} not found.")
+        return x
+
+    def forward(self, audio, text, device=None):
+        """Forward audio and text into the CLAP
+
+        Parameters
+        ----------
+        audio: torch.Tensor (batch_size, audio_length)
+            the time-domain audio input / the batch of mel_spec and longer list.
+        text: torch.Tensor () // need to add
+            the text token input
+        """
+        if device is None:
+            if audio is not None:
+                device = audio.device
+            elif text is not None:
+                device = text.device
+        if audio is None and text is None:
+            # a hack to get the logit scale
+            return self.logit_scale_a.exp(), self.logit_scale_t.exp()
+        elif audio is None:
+            return self.encode_text(text, device=device)
+        elif text is None:
+            return self.audio_projection(
+                self.encode_audio(audio, device=device)["embedding"]
+            )
+        audio_features = self.audio_projection(
+            self.encode_audio(audio, device=device)["embedding"]
+        )
+        audio_features = F.normalize(audio_features, dim=-1)
+
+        text_features = self.encode_text(text, device=device)
+        # print("text_features", text_features)
+        # print("text_features.shape", text_features.shape)
+        # print("text_features.type", type(text_features))
+        text_features = F.normalize(text_features, dim=-1)
+
+        audio_features_mlp = self.audio_transform(audio_features)
+        text_features_mlp = self.text_transform(text_features)
+        # Four outputs: audio features (basic & MLP), text features (basic & MLP)
+        return (
+            audio_features,
+            text_features,
+            audio_features_mlp,
+            text_features_mlp,
+            self.logit_scale_a.exp(),
+            self.logit_scale_t.exp(),
+        )
+
+    def get_logit_scale(self):
+        return self.logit_scale_a.exp(), self.logit_scale_t.exp()
+
+    def get_text_embedding(self, data):
+        """Get the text embedding from the model
+
+        Parameters
+        ----------
+        data: torch.Tensor
+            a tensor of text embedding
+
+        Returns
+        ----------
+        text_embed: torch.Tensor
+            a tensor of text_embeds (N, D)
+
+        """
+        device = next(self.parameters()).device
+        for k in data:
+            data[k] = data[k].to(device)
+            if(len(data[k].size()) < 2):
+                data[k] = data[k].unsqueeze(0)
+        text_embeds = self.encode_text(data, device=device)
+        text_embeds = F.normalize(text_embeds, dim=-1)
+
+        return text_embeds
+
+    def get_audio_embedding(self, data):
+        """Get the audio embedding from the model
+
+        Parameters
+        ----------
+        data: a list of dict
+            the audio input dict list from 'get_audio_feature' method
+
+        Returns
+        ----------
+        audio_embed: torch.Tensor
+            a tensor of audio_embeds (N, D)
+
+        """
+        device = next(self.parameters()).device
+        input_dict = {}
+        keys = data[0].keys()
+        for k in keys:
+            input_dict[k] = torch.cat([d[k].unsqueeze(0) for d in data], dim=0).to(
+                device
+            )
+
+        audio_embeds = self.audio_projection(
+            self.encode_audio(input_dict, device=device)["embedding"]
+        )
+        audio_embeds = F.normalize(audio_embeds, dim=-1)
+
+        return audio_embeds
+
+    def audio_infer(self, audio, hopsize=None, device=None):
+        """Forward one audio and produce the audio embedding
+
+        Parameters
+        ----------
+        audio:  (audio_length)
+            the time-domain audio input, notice that it must be only one input
+        hopsize: int
+            the overlap hopsize as the sliding window
+
+        Returns
+        ----------
+        output_dict: {
+            key: [n, (embedding_shape)] if "HTS-AT"
+            or
+            key: [(embedding_shape)] if "PANN"
+        }
+            the list of key values of the audio branch
+
+        """
+
+        assert not self.training, "the inference mode must be run at eval stage"
+        output_dict = {}
+        # PANN
+        if self.audio_cfg.model_type == "PANN":
+            audio_input = audio.unsqueeze(dim=0)
+            output_dict[key] = self.encode_audio(audio_input, device=device)[
+                key
+            ].squeeze(dim=0)
+        elif self.audio_cfg.model_type == "HTSAT":
+            # repeat
+            audio_len = len(audio)
+            k = self.audio_cfg.clip_samples // audio_len
+            if k > 1:
+                audio = audio.repeat(k)
+                audio_len = len(audio)
+
+            if hopsize is None:
+                hopsize = min(hopsize, audio_len)
+
+            if audio_len > self.audio_cfg.clip_samples:
+                audio_input = [
+                    audio[pos : pos + self.audio_cfg.clip_samples].clone()
+                    for pos in range(
+                        0, audio_len - self.audio_cfg.clip_samples, hopsize
+                    )
+                ]
+                audio_input.append(audio[-self.audio_cfg.clip_samples :].clone())
+                audio_input = torch.stack(audio_input)
+                output_dict[key] = self.encode_audio(audio_input, device=device)[key]
+            else:
+                audio_input = audio.unsqueeze(dim=0)
+                output_dict[key] = self.encode_audio(audio_input, device=device)[
+                    key
+                ].squeeze(dim=0)
+
+        return output_dict
+
+
+def convert_weights_to_fp16(model: nn.Module):
+    """Convert applicable model parameters to fp16"""
+
+    def _convert_weights_to_fp16(l):
+        if isinstance(l, (nn.Conv1d, nn.Conv2d, nn.Linear)):
+            l.weight.data = l.weight.data.half()
+            if l.bias is not None:
+                l.bias.data = l.bias.data.half()
+
+        if isinstance(l, nn.MultiheadAttention):
+            for attr in [
+                *[f"{s}_proj_weight" for s in ["in", "q", "k", "v"]],
+                "in_proj_bias",
+                "bias_k",
+                "bias_v",
+            ]:
+                tensor = getattr(l, attr)
+                if tensor is not None:
+                    tensor.data = tensor.data.half()
+
+        for name in ["text_projection", "proj"]:
+            if hasattr(l, name):
+                attr = getattr(l, name)
+                if attr is not None:
+                    attr.data = attr.data.half()
+
+    model.apply(_convert_weights_to_fp16)
+
+
+# Ignore the state dict of the vision part
+def build_model_from_openai_state_dict(
+    state_dict: dict, model_cfg, enable_fusion: bool = False, fusion_type: str = "None"
+):
+
+    embed_dim = model_cfg["embed_dim"]
+    audio_cfg = model_cfg["audio_cfg"]
+    text_cfg = model_cfg["text_cfg"]
+    context_length = state_dict["positional_embedding"].shape[0]
+    vocab_size = state_dict["token_embedding.weight"].shape[0]
+    transformer_width = state_dict["ln_final.weight"].shape[0]
+    transformer_heads = transformer_width // 64
+    transformer_layers = len(
+        set(
+            k.split(".")[2]
+            for k in state_dict
+            if k.startswith(f"transformer.resblocks")
+        )
+    )
+
+    audio_cfg = CLAPAudioCfp(**audio_cfg)
+    text_cfg = CLAPTextCfg(**text_cfg)
+
+    model = CLAP(
+        embed_dim,
+        audio_cfg=audio_cfg,
+        text_cfg=text_cfg,
+        quick_gelu=True,  # OpenAI models were trained with QuickGELU
+        enable_fusion=enable_fusion,
+        fusion_type=fusion_type,
+    )
+    state_dict["logit_scale_a"] = state_dict["logit_scale"]
+    state_dict["logit_scale_t"] = state_dict["logit_scale"]
+    pop_keys = list(state_dict.keys())[::]
+    # pop the visual branch saved weights
+    for key in pop_keys:
+        if key.startswith("visual."):
+            state_dict.pop(key, None)
+
+    for key in ["logit_scale", "input_resolution", "context_length", "vocab_size"]:
+        state_dict.pop(key, None)
+
+    # not use fp16
+    # convert_weights_to_fp16(model)
+    model.load_state_dict(state_dict, strict=False)
+    return model.eval()
+
+
+def trace_model(model, batch_size=256, device=torch.device("cpu")):
+    model.eval()
+    audio_length = model.audio_cfg.audio_length
+    example_audio = torch.ones((batch_size, audio_length), device=device)
+    example_text = torch.zeros(
+        (batch_size, model.context_length), dtype=torch.int, device=device
+    )
+    model = torch.jit.trace_module(
+        model,
+        inputs=dict(
+            forward=(example_audio, example_text),
+            encode_text=(example_text,),
+            encode_image=(example_audio,),
+        ),
+    )
+    model.audio_cfg.audio_length = audio_length  # Question: what does this do?
+    return model

audioldm/clap/open_clip/model_configs/HTSAT-base.json

@@ -0,0 +1,23 @@
+{
+    "embed_dim": 1024,
+    "audio_cfg": {
+        "audio_length": 1024,
+        "clip_samples": 480000,
+        "mel_bins": 64,
+        "sample_rate": 48000,
+        "window_size": 1024,
+        "hop_size": 480,
+        "fmin": 50,
+        "fmax": 14000,
+        "class_num": 527,
+        "model_type": "HTSAT",
+        "model_name": "base"
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
+}

audioldm/clap/open_clip/model_configs/HTSAT-large.json

@@ -0,0 +1,23 @@
+{
+    "embed_dim": 2048,
+    "audio_cfg": {
+        "audio_length": 1024,
+        "clip_samples": 480000,
+        "mel_bins": 64,
+        "sample_rate": 48000,
+        "window_size": 1024,
+        "hop_size": 480,
+        "fmin": 50,
+        "fmax": 14000,
+        "class_num": 527,
+        "model_type": "HTSAT",
+        "model_name": "large"
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
+}

audioldm/clap/open_clip/model_configs/HTSAT-tiny-win-1536.json

@@ -0,0 +1,23 @@
+{
+    "embed_dim": 768,
+    "audio_cfg": {
+        "audio_length": 1024,
+        "clip_samples": 480000,
+        "mel_bins": 64,
+        "sample_rate": 48000,
+        "window_size": 1536,
+        "hop_size": 480,
+        "fmin": 50,
+        "fmax": 14000,
+        "class_num": 527,
+        "model_type": "HTSAT",
+        "model_name": "tiny"
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
+}

audioldm/clap/open_clip/model_configs/HTSAT-tiny.json

@@ -0,0 +1,23 @@
+{
+    "embed_dim": 768,
+    "audio_cfg": {
+        "audio_length": 1024,
+        "clip_samples": 480000,
+        "mel_bins": 64,
+        "sample_rate": 48000,
+        "window_size": 1024,
+        "hop_size": 480,
+        "fmin": 50,
+        "fmax": 14000,
+        "class_num": 527,
+        "model_type": "HTSAT",
+        "model_name": "tiny"
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
+}

audioldm/clap/open_clip/model_configs/PANN-10.json

@@ -0,0 +1,23 @@
+{
+    "embed_dim": 1024,
+    "audio_cfg": {
+        "audio_length": 1024,
+        "clip_samples": 480000,
+        "mel_bins": 64,
+        "sample_rate": 48000,
+        "window_size": 1024,
+        "hop_size": 480,
+        "fmin": 50,
+        "fmax": 14000,
+        "class_num": 527,
+        "model_type": "PANN",
+        "model_name": "Cnn10"
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
+}

audioldm/clap/open_clip/model_configs/PANN-14-fmax-18k.json

@@ -0,0 +1,23 @@
+{
+    "embed_dim": 2048,
+    "audio_cfg": {
+        "audio_length": 1024,
+        "clip_samples": 480000,
+        "mel_bins": 64,
+        "sample_rate": 48000,
+        "window_size": 1024,
+        "hop_size": 480,
+        "fmin": 50,
+        "fmax": 18000,
+        "class_num": 527,
+        "model_type": "PANN",
+        "model_name": "Cnn14"
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
+}

audioldm/clap/open_clip/model_configs/PANN-14-fmax-8k-20s.json

@@ -0,0 +1,23 @@
+{
+    "embed_dim": 2048,
+    "audio_cfg": {
+        "audio_length": 1024,
+        "clip_samples": 960000,
+        "mel_bins": 64,
+        "sample_rate": 48000,
+        "window_size": 1024,
+        "hop_size": 360,
+        "fmin": 50,
+        "fmax": 8000,
+        "class_num": 527,
+        "model_type": "PANN",
+        "model_name": "Cnn14"
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
+}

audioldm/clap/open_clip/model_configs/PANN-14-tiny-transformer.json

@@ -0,0 +1,23 @@
+{
+    "embed_dim": 2048,
+    "audio_cfg": {
+        "audio_length": 1024,
+        "clip_samples": 480000,
+        "mel_bins": 64,
+        "sample_rate": 48000,
+        "window_size": 1024,
+        "hop_size": 480,
+        "fmin": 50,
+        "fmax": 14000,
+        "class_num": 527,
+        "model_type": "PANN",
+        "model_name": "Cnn14"
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 4
+    }
+}

audioldm/clap/open_clip/model_configs/PANN-14-win-1536.json

@@ -0,0 +1,23 @@
+{
+    "embed_dim": 2048,
+    "audio_cfg": {
+        "audio_length": 1024,
+        "clip_samples": 480000,
+        "mel_bins": 64,
+        "sample_rate": 48000,
+        "window_size": 1536,
+        "hop_size": 480,
+        "fmin": 50,
+        "fmax": 14000,
+        "class_num": 527,
+        "model_type": "PANN",
+        "model_name": "Cnn14"
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
+}

audioldm/clap/open_clip/model_configs/PANN-14.json

@@ -0,0 +1,23 @@
+{
+    "embed_dim": 2048,
+    "audio_cfg": {
+        "audio_length": 1024,
+        "clip_samples": 480000,
+        "mel_bins": 64,
+        "sample_rate": 48000,
+        "window_size": 1024,
+        "hop_size": 480,
+        "fmin": 50,
+        "fmax": 14000,
+        "class_num": 527,
+        "model_type": "PANN",
+        "model_name": "Cnn14"
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
+}

audioldm/clap/open_clip/model_configs/PANN-6.json

@@ -0,0 +1,23 @@
+{
+    "embed_dim": 512,
+    "audio_cfg": {
+        "audio_length": 1024,
+        "clip_samples": 480000,
+        "mel_bins": 64,
+        "sample_rate": 48000,
+        "window_size": 1024,
+        "hop_size": 480,
+        "fmin": 50,
+        "fmax": 14000,
+        "class_num": 527,
+        "model_type": "PANN",
+        "model_name": "Cnn6"
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
+}

audioldm/clap/open_clip/model_configs/RN101-quickgelu.json

@@ -0,0 +1,22 @@
+{
+    "embed_dim": 512,
+    "quick_gelu": true,
+    "vision_cfg": {
+        "image_size": 224,
+        "layers": [
+            3,
+            4,
+            23,
+            3
+        ],
+        "width": 64,
+        "patch_size": null
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
+}

audioldm/clap/open_clip/model_configs/RN101.json

@@ -0,0 +1,21 @@
+{
+    "embed_dim": 512,
+    "vision_cfg": {
+        "image_size": 224,
+        "layers": [
+            3,
+            4,
+            23,
+            3
+        ],
+        "width": 64,
+        "patch_size": null
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
+}

audioldm/clap/open_clip/model_configs/RN50-quickgelu.json

@@ -0,0 +1,22 @@
+{
+    "embed_dim": 1024,
+    "quick_gelu": true,
+    "vision_cfg": {
+        "image_size": 224,
+        "layers": [
+            3,
+            4,
+            6,
+            3
+        ],
+        "width": 64,
+        "patch_size": null
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
+}

audioldm/clap/open_clip/model_configs/RN50.json

@@ -0,0 +1,21 @@
+{
+    "embed_dim": 1024,
+    "vision_cfg": {
+        "image_size": 224,
+        "layers": [
+            3,
+            4,
+            6,
+            3
+        ],
+        "width": 64,
+        "patch_size": null
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
+}

audioldm/clap/open_clip/model_configs/RN50x16.json

@@ -0,0 +1,21 @@
+{
+    "embed_dim": 768,
+    "vision_cfg": {
+        "image_size": 384,
+        "layers": [
+            6,
+            8,
+            18,
+            8
+        ],
+        "width": 96,
+        "patch_size": null
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 768,
+        "heads": 12,
+        "layers": 12
+    }
+}

audioldm/clap/open_clip/model_configs/RN50x4.json

@@ -0,0 +1,21 @@
+{
+    "embed_dim": 640,
+    "vision_cfg": {
+        "image_size": 288,
+        "layers": [
+            4,
+            6,
+            10,
+            6
+        ],
+        "width": 80,
+        "patch_size": null
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 640,
+        "heads": 10,
+        "layers": 12
+    }
+}

audioldm/clap/open_clip/model_configs/ViT-B-16.json

@@ -0,0 +1,16 @@
+{
+    "embed_dim": 512,
+    "vision_cfg": {
+        "image_size": 224,
+        "layers": 12,
+        "width": 768,
+        "patch_size": 16
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
+}

audioldm/clap/open_clip/model_configs/ViT-B-32-quickgelu.json

@@ -0,0 +1,17 @@
+{
+    "embed_dim": 512,
+    "quick_gelu": true,
+    "vision_cfg": {
+        "image_size": 224,
+        "layers": 12,
+        "width": 768,
+        "patch_size": 32
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
+}

audioldm/clap/open_clip/model_configs/ViT-B-32.json

@@ -0,0 +1,16 @@
+{
+    "embed_dim": 512,
+    "vision_cfg": {
+        "image_size": 224,
+        "layers": 12,
+        "width": 768,
+        "patch_size": 32
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
+}

audioldm/clap/open_clip/model_configs/ViT-L-14.json

@@ -0,0 +1,16 @@
+{
+    "embed_dim": 768,
+    "vision_cfg": {
+        "image_size": 224,
+        "layers": 24,
+        "width": 1024,
+        "patch_size": 14
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 768,
+        "heads": 12,
+        "layers": 12
+    }
+}

audioldm/clap/open_clip/openai.py

@@ -0,0 +1,156 @@
+""" OpenAI pretrained model functions
+
+Adapted from https://github.com/openai/CLIP. Originally MIT License, Copyright (c) 2021 OpenAI.
+"""
+
+import os
+import warnings
+from typing import Union, List
+
+import torch
+
+from .model import build_model_from_openai_state_dict
+from .pretrained import (
+    get_pretrained_url,
+    list_pretrained_tag_models,
+    download_pretrained,
+)
+
+__all__ = ["list_openai_models", "load_openai_model"]
+
+
+def list_openai_models() -> List[str]:
+    """Returns the names of available CLIP models"""
+    return list_pretrained_tag_models("openai")
+
+
+def load_openai_model(
+    name: str,
+    model_cfg,
+    device: Union[str, torch.device] = "cuda" if torch.cuda.is_available() else "cpu",
+    jit=True,
+    cache_dir=os.path.expanduser("~/.cache/clip"),
+    enable_fusion: bool = False,
+    fusion_type: str = "None",
+):
+    """Load a CLIP model, preserve its text pretrained part, and set in the CLAP model
+
+    Parameters
+    ----------
+    name : str
+        A model name listed by `clip.available_models()`, or the path to a model checkpoint containing the state_dict
+    device : Union[str, torch.device]
+        The device to put the loaded model
+    jit : bool
+        Whether to load the optimized JIT model (default) or more hackable non-JIT model.
+
+    Returns
+    -------
+    model : torch.nn.Module
+        The CLAP model
+    preprocess : Callable[[PIL.Image], torch.Tensor]
+        A torchvision transform that converts a PIL image into a tensor that the returned model can take as its input
+    """
+    if get_pretrained_url(name, "openai"):
+        model_path = download_pretrained(
+            get_pretrained_url(name, "openai"), root=cache_dir
+        )
+    elif os.path.isfile(name):
+        model_path = name
+    else:
+        raise RuntimeError(
+            f"Model {name} not found; available models = {list_openai_models()}"
+        )
+
+    try:
+        # loading JIT archive
+        model = torch.jit.load(model_path, map_location=device if jit else "cpu").eval()
+        state_dict = None
+    except RuntimeError:
+        # loading saved state dict
+        if jit:
+            warnings.warn(
+                f"File {model_path} is not a JIT archive. Loading as a state dict instead"
+            )
+            jit = False
+        state_dict = torch.load(model_path, map_location="cpu")
+
+    if not jit:
+        try:
+            model = build_model_from_openai_state_dict(
+                state_dict or model.state_dict(), model_cfg, enable_fusion, fusion_type
+            ).to(device)
+        except KeyError:
+            sd = {k[7:]: v for k, v in state_dict["state_dict"].items()}
+            model = build_model_from_openai_state_dict(
+                sd, model_cfg, enable_fusion, fusion_type
+            ).to(device)
+
+        if str(device) == "cpu":
+            model.float()
+        return model
+
+    # patch the device names
+    device_holder = torch.jit.trace(
+        lambda: torch.ones([]).to(torch.device(device)), example_inputs=[]
+    )
+    device_node = [
+        n
+        for n in device_holder.graph.findAllNodes("prim::Constant")
+        if "Device" in repr(n)
+    ][-1]
+
+    def patch_device(module):
+        try:
+            graphs = [module.graph] if hasattr(module, "graph") else []
+        except RuntimeError:
+            graphs = []
+
+        if hasattr(module, "forward1"):
+            graphs.append(module.forward1.graph)
+
+        for graph in graphs:
+            for node in graph.findAllNodes("prim::Constant"):
+                if "value" in node.attributeNames() and str(node["value"]).startswith(
+                    "cuda"
+                ):
+                    node.copyAttributes(device_node)
+
+    model.apply(patch_device)
+    patch_device(model.encode_audio)
+    patch_device(model.encode_text)
+
+    # patch dtype to float32 on CPU
+    if str(device) == "cpu":
+        float_holder = torch.jit.trace(
+            lambda: torch.ones([]).float(), example_inputs=[]
+        )
+        float_input = list(float_holder.graph.findNode("aten::to").inputs())[1]
+        float_node = float_input.node()
+
+        def patch_float(module):
+            try:
+                graphs = [module.graph] if hasattr(module, "graph") else []
+            except RuntimeError:
+                graphs = []
+
+            if hasattr(module, "forward1"):
+                graphs.append(module.forward1.graph)
+
+            for graph in graphs:
+                for node in graph.findAllNodes("aten::to"):
+                    inputs = list(node.inputs())
+                    for i in [
+                        1,
+                        2,
+                    ]:  # dtype can be the second or third argument to aten::to()
+                        if inputs[i].node()["value"] == 5:
+                            inputs[i].node().copyAttributes(float_node)
+
+        model.apply(patch_float)
+        patch_float(model.encode_audio)
+        patch_float(model.encode_text)
+        model.float()
+
+    model.audio_branch.audio_length = model.audio_cfg.audio_length
+    return model

audioldm/clap/open_clip/pann_model.py

@@ -0,0 +1,703 @@
+# PANNs: Large-Scale Pretrained Audio Neural Networks for Audio Pattern Recognition
+# Reference from https://github.com/qiuqiangkong/audioset_tagging_cnn
+# Some layers are re-designed for CLAP
+import os
+
+os.environ["NUMBA_CACHE_DIR"] = "/tmp/"
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torchlibrosa.stft import Spectrogram, LogmelFilterBank
+from torchlibrosa.augmentation import SpecAugmentation
+
+from .utils import do_mixup, interpolate, pad_framewise_output
+from .feature_fusion import iAFF, AFF, DAF
+
+
+def init_layer(layer):
+    """Initialize a Linear or Convolutional layer."""
+    nn.init.xavier_uniform_(layer.weight)
+
+    if hasattr(layer, "bias"):
+        if layer.bias is not None:
+            layer.bias.data.fill_(0.0)
+
+def init_bn(bn):
+    """Initialize a Batchnorm layer."""
+    bn.bias.data.fill_(0.0)
+    bn.weight.data.fill_(1.0)
+
+
+class ConvBlock(nn.Module):
+    def __init__(self, in_channels, out_channels):
+
+        super(ConvBlock, self).__init__()
+
+        self.conv1 = nn.Conv2d(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=(3, 3),
+            stride=(1, 1),
+            padding=(1, 1),
+            bias=False,
+        )
+
+        self.conv2 = nn.Conv2d(
+            in_channels=out_channels,
+            out_channels=out_channels,
+            kernel_size=(3, 3),
+            stride=(1, 1),
+            padding=(1, 1),
+            bias=False,
+        )
+
+        self.bn1 = nn.BatchNorm2d(out_channels)
+        self.bn2 = nn.BatchNorm2d(out_channels)
+
+        self.init_weight()
+
+    def init_weight(self):
+        init_layer(self.conv1)
+        init_layer(self.conv2)
+        init_bn(self.bn1)
+        init_bn(self.bn2)
+
+    def forward(self, input, pool_size=(2, 2), pool_type="avg"):
+
+        x = input
+        x = F.relu_(self.bn1(self.conv1(x)))
+        x = F.relu_(self.bn2(self.conv2(x)))
+        if pool_type == "max":
+            x = F.max_pool2d(x, kernel_size=pool_size)
+        elif pool_type == "avg":
+            x = F.avg_pool2d(x, kernel_size=pool_size)
+        elif pool_type == "avg+max":
+            x1 = F.avg_pool2d(x, kernel_size=pool_size)
+            x2 = F.max_pool2d(x, kernel_size=pool_size)
+            x = x1 + x2
+        else:
+            raise Exception("Incorrect argument!")
+
+        return x
+
+
+class ConvBlock5x5(nn.Module):
+    def __init__(self, in_channels, out_channels):
+
+        super(ConvBlock5x5, self).__init__()
+
+        self.conv1 = nn.Conv2d(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=(5, 5),
+            stride=(1, 1),
+            padding=(2, 2),
+            bias=False,
+        )
+
+        self.bn1 = nn.BatchNorm2d(out_channels)
+
+        self.init_weight()
+
+    def init_weight(self):
+        init_layer(self.conv1)
+        init_bn(self.bn1)
+
+    def forward(self, input, pool_size=(2, 2), pool_type="avg"):
+
+        x = input
+        x = F.relu_(self.bn1(self.conv1(x)))
+        if pool_type == "max":
+            x = F.max_pool2d(x, kernel_size=pool_size)
+        elif pool_type == "avg":
+            x = F.avg_pool2d(x, kernel_size=pool_size)
+        elif pool_type == "avg+max":
+            x1 = F.avg_pool2d(x, kernel_size=pool_size)
+            x2 = F.max_pool2d(x, kernel_size=pool_size)
+            x = x1 + x2
+        else:
+            raise Exception("Incorrect argument!")
+
+        return x
+
+
+class AttBlock(nn.Module):
+    def __init__(self, n_in, n_out, activation="linear", temperature=1.0):
+        super(AttBlock, self).__init__()
+
+        self.activation = activation
+        self.temperature = temperature
+        self.att = nn.Conv1d(
+            in_channels=n_in,
+            out_channels=n_out,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=True,
+        )
+        self.cla = nn.Conv1d(
+            in_channels=n_in,
+            out_channels=n_out,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=True,
+        )
+
+        self.bn_att = nn.BatchNorm1d(n_out)
+        self.init_weights()
+
+    def init_weights(self):
+        init_layer(self.att)
+        init_layer(self.cla)
+        init_bn(self.bn_att)
+
+    def forward(self, x):
+        # x: (n_samples, n_in, n_time)
+        norm_att = torch.softmax(torch.clamp(self.att(x), -10, 10), dim=-1)
+        cla = self.nonlinear_transform(self.cla(x))
+        x = torch.sum(norm_att * cla, dim=2)
+        return x, norm_att, cla
+
+    def nonlinear_transform(self, x):
+        if self.activation == "linear":
+            return x
+        elif self.activation == "sigmoid":
+            return torch.sigmoid(x)
+
+
+class Cnn14(nn.Module):
+    def __init__(
+        self,
+        sample_rate,
+        window_size,
+        hop_size,
+        mel_bins,
+        fmin,
+        fmax,
+        classes_num,
+        enable_fusion=False,
+        fusion_type="None",
+    ):
+
+        super(Cnn14, self).__init__()
+
+        window = "hann"
+        center = True
+        pad_mode = "reflect"
+        ref = 1.0
+        amin = 1e-10
+        top_db = None
+
+        self.enable_fusion = enable_fusion
+        self.fusion_type = fusion_type
+
+        # Spectrogram extractor
+        self.spectrogram_extractor = Spectrogram(
+            n_fft=window_size,
+            hop_length=hop_size,
+            win_length=window_size,
+            window=window,
+            center=center,
+            pad_mode=pad_mode,
+            freeze_parameters=True,
+        )
+
+        # Logmel feature extractor
+        self.logmel_extractor = LogmelFilterBank(
+            sr=sample_rate,
+            n_fft=window_size,
+            n_mels=mel_bins,
+            fmin=fmin,
+            fmax=fmax,
+            ref=ref,
+            amin=amin,
+            top_db=top_db,
+            freeze_parameters=True,
+        )
+
+        # Spec augmenter
+        self.spec_augmenter = SpecAugmentation(
+            time_drop_width=64,
+            time_stripes_num=2,
+            freq_drop_width=8,
+            freq_stripes_num=2,
+        )
+
+        self.bn0 = nn.BatchNorm2d(64)
+
+        if (self.enable_fusion) and (self.fusion_type == "channel_map"):
+            self.conv_block1 = ConvBlock(in_channels=4, out_channels=64)
+        else:
+            self.conv_block1 = ConvBlock(in_channels=1, out_channels=64)
+        self.conv_block2 = ConvBlock(in_channels=64, out_channels=128)
+        self.conv_block3 = ConvBlock(in_channels=128, out_channels=256)
+        self.conv_block4 = ConvBlock(in_channels=256, out_channels=512)
+        self.conv_block5 = ConvBlock(in_channels=512, out_channels=1024)
+        self.conv_block6 = ConvBlock(in_channels=1024, out_channels=2048)
+
+        self.fc1 = nn.Linear(2048, 2048, bias=True)
+        self.fc_audioset = nn.Linear(2048, classes_num, bias=True)
+
+        if (self.enable_fusion) and (
+            self.fusion_type in ["daf_1d", "aff_1d", "iaff_1d"]
+        ):
+            self.mel_conv1d = nn.Sequential(
+                nn.Conv1d(64, 64, kernel_size=5, stride=3, padding=2),
+                nn.BatchNorm1d(64),  # No Relu
+            )
+            if self.fusion_type == "daf_1d":
+                self.fusion_model = DAF()
+            elif self.fusion_type == "aff_1d":
+                self.fusion_model = AFF(channels=64, type="1D")
+            elif self.fusion_type == "iaff_1d":
+                self.fusion_model = iAFF(channels=64, type="1D")
+
+        if (self.enable_fusion) and (
+            self.fusion_type in ["daf_2d", "aff_2d", "iaff_2d"]
+        ):
+            self.mel_conv2d = nn.Sequential(
+                nn.Conv2d(1, 64, kernel_size=(5, 5), stride=(6, 2), padding=(2, 2)),
+                nn.BatchNorm2d(64),
+                nn.ReLU(inplace=True),
+            )
+
+            if self.fusion_type == "daf_2d":
+                self.fusion_model = DAF()
+            elif self.fusion_type == "aff_2d":
+                self.fusion_model = AFF(channels=64, type="2D")
+            elif self.fusion_type == "iaff_2d":
+                self.fusion_model = iAFF(channels=64, type="2D")
+        self.init_weight()
+
+    def init_weight(self):
+        init_bn(self.bn0)
+        init_layer(self.fc1)
+        init_layer(self.fc_audioset)
+
+    def forward(self, input, mixup_lambda=None, device=None):
+        """
+        Input: (batch_size, data_length)"""
+
+        if self.enable_fusion and input["longer"].sum() == 0:
+            # if no audio is longer than 10s, then randomly select one audio to be longer
+            input["longer"][torch.randint(0, input["longer"].shape[0], (1,))] = True
+
+        if not self.enable_fusion:
+            x = self.spectrogram_extractor(
+                input["waveform"].to(device=device, non_blocking=True)
+            )  # (batch_size, 1, time_steps, freq_bins)
+            x = self.logmel_extractor(x)  # (batch_size, 1, time_steps, mel_bins)
+
+            x = x.transpose(1, 3)
+            x = self.bn0(x)
+            x = x.transpose(1, 3)
+        else:
+            longer_list = input["longer"].to(device=device, non_blocking=True)
+            x = input["mel_fusion"].to(device=device, non_blocking=True)
+            longer_list_idx = torch.where(longer_list)[0]
+            x = x.transpose(1, 3)
+            x = self.bn0(x)
+            x = x.transpose(1, 3)
+            if self.fusion_type in ["daf_1d", "aff_1d", "iaff_1d"]:
+                new_x = x[:, 0:1, :, :].clone().contiguous()
+                # local processing
+                if len(longer_list_idx) > 0:
+                    fusion_x_local = x[longer_list_idx, 1:, :, :].clone().contiguous()
+                    FB, FC, FT, FF = fusion_x_local.size()
+                    fusion_x_local = fusion_x_local.view(FB * FC, FT, FF)
+                    fusion_x_local = torch.permute(
+                        fusion_x_local, (0, 2, 1)
+                    ).contiguous()
+                    fusion_x_local = self.mel_conv1d(fusion_x_local)
+                    fusion_x_local = fusion_x_local.view(
+                        FB, FC, FF, fusion_x_local.size(-1)
+                    )
+                    fusion_x_local = (
+                        torch.permute(fusion_x_local, (0, 2, 1, 3))
+                        .contiguous()
+                        .flatten(2)
+                    )
+                    if fusion_x_local.size(-1) < FT:
+                        fusion_x_local = torch.cat(
+                            [
+                                fusion_x_local,
+                                torch.zeros(
+                                    (FB, FF, FT - fusion_x_local.size(-1)),
+                                    device=device,
+                                ),
+                            ],
+                            dim=-1,
+                        )
+                    else:
+                        fusion_x_local = fusion_x_local[:, :, :FT]
+                    # 1D fusion
+                    new_x = new_x.squeeze(1).permute((0, 2, 1)).contiguous()
+                    new_x[longer_list_idx] = self.fusion_model(
+                        new_x[longer_list_idx], fusion_x_local
+                    )
+                    x = new_x.permute((0, 2, 1)).contiguous()[:, None, :, :]
+                else:
+                    x = new_x
+            elif self.fusion_type in ["daf_2d", "aff_2d", "iaff_2d", "channel_map"]:
+                x = x  # no change
+
+        if self.training:
+            x = self.spec_augmenter(x)
+        # Mixup on spectrogram
+        if self.training and mixup_lambda is not None:
+            x = do_mixup(x, mixup_lambda)
+        if (self.enable_fusion) and (
+            self.fusion_type in ["daf_2d", "aff_2d", "iaff_2d"]
+        ):
+            global_x = x[:, 0:1, :, :]
+
+            # global processing
+            B, C, H, W = global_x.shape
+            global_x = self.conv_block1(global_x, pool_size=(2, 2), pool_type="avg")
+            if len(longer_list_idx) > 0:
+                local_x = x[longer_list_idx, 1:, :, :].contiguous()
+                TH = global_x.size(-2)
+                # local processing
+                B, C, H, W = local_x.shape
+                local_x = local_x.view(B * C, 1, H, W)
+                local_x = self.mel_conv2d(local_x)
+                local_x = local_x.view(
+                    B, C, local_x.size(1), local_x.size(2), local_x.size(3)
+                )
+                local_x = local_x.permute((0, 2, 1, 3, 4)).contiguous().flatten(2, 3)
+                TB, TC, _, TW = local_x.size()
+                if local_x.size(-2) < TH:
+                    local_x = torch.cat(
+                        [
+                            local_x,
+                            torch.zeros(
+                                (TB, TC, TH - local_x.size(-2), TW),
+                                device=global_x.device,
+                            ),
+                        ],
+                        dim=-2,
+                    )
+                else:
+                    local_x = local_x[:, :, :TH, :]
+
+                global_x[longer_list_idx] = self.fusion_model(
+                    global_x[longer_list_idx], local_x
+                )
+            x = global_x
+        else:
+            x = self.conv_block1(x, pool_size=(2, 2), pool_type="avg")
+
+        x = F.dropout(x, p=0.2, training=self.training)
+        x = self.conv_block2(x, pool_size=(2, 2), pool_type="avg")
+        x = F.dropout(x, p=0.2, training=self.training)
+        x = self.conv_block3(x, pool_size=(2, 2), pool_type="avg")
+        x = F.dropout(x, p=0.2, training=self.training)
+        x = self.conv_block4(x, pool_size=(2, 2), pool_type="avg")
+        x = F.dropout(x, p=0.2, training=self.training)
+        x = self.conv_block5(x, pool_size=(2, 2), pool_type="avg")
+        x = F.dropout(x, p=0.2, training=self.training)
+        x = self.conv_block6(x, pool_size=(1, 1), pool_type="avg")
+        x = F.dropout(x, p=0.2, training=self.training)
+        x = torch.mean(x, dim=3)
+
+        latent_x1 = F.max_pool1d(x, kernel_size=3, stride=1, padding=1)
+        latent_x2 = F.avg_pool1d(x, kernel_size=3, stride=1, padding=1)
+        latent_x = latent_x1 + latent_x2
+        latent_x = latent_x.transpose(1, 2)
+        latent_x = F.relu_(self.fc1(latent_x))
+        latent_output = interpolate(latent_x, 32)
+
+        (x1, _) = torch.max(x, dim=2)
+        x2 = torch.mean(x, dim=2)
+        x = x1 + x2
+        x = F.dropout(x, p=0.5, training=self.training)
+        x = F.relu_(self.fc1(x))
+        embedding = F.dropout(x, p=0.5, training=self.training)
+        clipwise_output = torch.sigmoid(self.fc_audioset(x))
+
+        output_dict = {
+            "clipwise_output": clipwise_output,
+            "embedding": embedding,
+            "fine_grained_embedding": latent_output,
+        }
+        return output_dict
+
+
+class Cnn6(nn.Module):
+    def __init__(
+        self,
+        sample_rate,
+        window_size,
+        hop_size,
+        mel_bins,
+        fmin,
+        fmax,
+        classes_num,
+        enable_fusion=False,
+        fusion_type="None",
+    ):
+
+        super(Cnn6, self).__init__()
+
+        window = "hann"
+        center = True
+        pad_mode = "reflect"
+        ref = 1.0
+        amin = 1e-10
+        top_db = None
+
+        self.enable_fusion = enable_fusion
+        self.fusion_type = fusion_type
+
+        # Spectrogram extractor
+        self.spectrogram_extractor = Spectrogram(
+            n_fft=window_size,
+            hop_length=hop_size,
+            win_length=window_size,
+            window=window,
+            center=center,
+            pad_mode=pad_mode,
+            freeze_parameters=True,
+        )
+
+        # Logmel feature extractor
+        self.logmel_extractor = LogmelFilterBank(
+            sr=sample_rate,
+            n_fft=window_size,
+            n_mels=mel_bins,
+            fmin=fmin,
+            fmax=fmax,
+            ref=ref,
+            amin=amin,
+            top_db=top_db,
+            freeze_parameters=True,
+        )
+
+        # Spec augmenter
+        self.spec_augmenter = SpecAugmentation(
+            time_drop_width=64,
+            time_stripes_num=2,
+            freq_drop_width=8,
+            freq_stripes_num=2,
+        )
+
+        self.bn0 = nn.BatchNorm2d(64)
+
+        self.conv_block1 = ConvBlock5x5(in_channels=1, out_channels=64)
+        self.conv_block2 = ConvBlock5x5(in_channels=64, out_channels=128)
+        self.conv_block3 = ConvBlock5x5(in_channels=128, out_channels=256)
+        self.conv_block4 = ConvBlock5x5(in_channels=256, out_channels=512)
+
+        self.fc1 = nn.Linear(512, 512, bias=True)
+        self.fc_audioset = nn.Linear(512, classes_num, bias=True)
+
+        self.init_weight()
+
+    def init_weight(self):
+        init_bn(self.bn0)
+        init_layer(self.fc1)
+        init_layer(self.fc_audioset)
+
+    def forward(self, input, mixup_lambda=None, device=None):
+        """
+        Input: (batch_size, data_length)"""
+
+        x = self.spectrogram_extractor(input)  # (batch_size, 1, time_steps, freq_bins)
+        x = self.logmel_extractor(x)  # (batch_size, 1, time_steps, mel_bins)
+
+        x = x.transpose(1, 3)
+        x = self.bn0(x)
+        x = x.transpose(1, 3)
+
+        if self.training:
+            x = self.spec_augmenter(x)
+
+        # Mixup on spectrogram
+        if self.training and mixup_lambda is not None:
+            x = do_mixup(x, mixup_lambda)
+
+        x = self.conv_block1(x, pool_size=(2, 2), pool_type="avg")
+        x = F.dropout(x, p=0.2, training=self.training)
+        x = self.conv_block2(x, pool_size=(2, 2), pool_type="avg")
+        x = F.dropout(x, p=0.2, training=self.training)
+        x = self.conv_block3(x, pool_size=(2, 2), pool_type="avg")
+        x = F.dropout(x, p=0.2, training=self.training)
+        x = self.conv_block4(x, pool_size=(2, 2), pool_type="avg")
+        x = F.dropout(x, p=0.2, training=self.training)
+        x = torch.mean(x, dim=3)
+
+        latent_x1 = F.max_pool1d(x, kernel_size=3, stride=1, padding=1)
+        latent_x2 = F.avg_pool1d(x, kernel_size=3, stride=1, padding=1)
+        latent_x = latent_x1 + latent_x2
+        latent_x = latent_x.transpose(1, 2)
+        latent_x = F.relu_(self.fc1(latent_x))
+        latent_output = interpolate(latent_x, 16)
+
+        (x1, _) = torch.max(x, dim=2)
+        x2 = torch.mean(x, dim=2)
+        x = x1 + x2
+        x = F.dropout(x, p=0.5, training=self.training)
+        x = F.relu_(self.fc1(x))
+        embedding = F.dropout(x, p=0.5, training=self.training)
+        clipwise_output = torch.sigmoid(self.fc_audioset(x))
+
+        output_dict = {
+            "clipwise_output": clipwise_output,
+            "embedding": embedding,
+            "fine_grained_embedding": latent_output,
+        }
+
+        return output_dict
+
+
+class Cnn10(nn.Module):
+    def __init__(
+        self,
+        sample_rate,
+        window_size,
+        hop_size,
+        mel_bins,
+        fmin,
+        fmax,
+        classes_num,
+        enable_fusion=False,
+        fusion_type="None",
+    ):
+
+        super(Cnn10, self).__init__()
+
+        window = "hann"
+        center = True
+        pad_mode = "reflect"
+        ref = 1.0
+        amin = 1e-10
+        top_db = None
+
+        self.enable_fusion = enable_fusion
+        self.fusion_type = fusion_type
+
+        # Spectrogram extractor
+        self.spectrogram_extractor = Spectrogram(
+            n_fft=window_size,
+            hop_length=hop_size,
+            win_length=window_size,
+            window=window,
+            center=center,
+            pad_mode=pad_mode,
+            freeze_parameters=True,
+        )
+
+        # Logmel feature extractor
+        self.logmel_extractor = LogmelFilterBank(
+            sr=sample_rate,
+            n_fft=window_size,
+            n_mels=mel_bins,
+            fmin=fmin,
+            fmax=fmax,
+            ref=ref,
+            amin=amin,
+            top_db=top_db,
+            freeze_parameters=True,
+        )
+
+        # Spec augmenter
+        self.spec_augmenter = SpecAugmentation(
+            time_drop_width=64,
+            time_stripes_num=2,
+            freq_drop_width=8,
+            freq_stripes_num=2,
+        )
+
+        self.bn0 = nn.BatchNorm2d(64)
+
+        self.conv_block1 = ConvBlock(in_channels=1, out_channels=64)
+        self.conv_block2 = ConvBlock(in_channels=64, out_channels=128)
+        self.conv_block3 = ConvBlock(in_channels=128, out_channels=256)
+        self.conv_block4 = ConvBlock(in_channels=256, out_channels=512)
+        self.conv_block5 = ConvBlock(in_channels=512, out_channels=1024)
+
+        self.fc1 = nn.Linear(1024, 1024, bias=True)
+        self.fc_audioset = nn.Linear(1024, classes_num, bias=True)
+
+        self.init_weight()
+
+    def init_weight(self):
+        init_bn(self.bn0)
+        init_layer(self.fc1)
+        init_layer(self.fc_audioset)
+
+    def forward(self, input, mixup_lambda=None, device=None):
+        """
+        Input: (batch_size, data_length)"""
+
+        x = self.spectrogram_extractor(input)  # (batch_size, 1, time_steps, freq_bins)
+        x = self.logmel_extractor(x)  # (batch_size, 1, time_steps, mel_bins)
+
+        x = x.transpose(1, 3)
+        x = self.bn0(x)
+        x = x.transpose(1, 3)
+
+        if self.training:
+            x = self.spec_augmenter(x)
+
+        # Mixup on spectrogram
+        if self.training and mixup_lambda is not None:
+            x = do_mixup(x, mixup_lambda)
+
+        x = self.conv_block1(x, pool_size=(2, 2), pool_type="avg")
+        x = F.dropout(x, p=0.2, training=self.training)
+        x = self.conv_block2(x, pool_size=(2, 2), pool_type="avg")
+        x = F.dropout(x, p=0.2, training=self.training)
+        x = self.conv_block3(x, pool_size=(2, 2), pool_type="avg")
+        x = F.dropout(x, p=0.2, training=self.training)
+        x = self.conv_block4(x, pool_size=(2, 2), pool_type="avg")
+        x = F.dropout(x, p=0.2, training=self.training)
+        x = self.conv_block5(x, pool_size=(2, 2), pool_type="avg")
+        x = F.dropout(x, p=0.2, training=self.training)
+        x = torch.mean(x, dim=3)
+
+        latent_x1 = F.max_pool1d(x, kernel_size=3, stride=1, padding=1)
+        latent_x2 = F.avg_pool1d(x, kernel_size=3, stride=1, padding=1)
+        latent_x = latent_x1 + latent_x2
+        latent_x = latent_x.transpose(1, 2)
+        latent_x = F.relu_(self.fc1(latent_x))
+        latent_output = interpolate(latent_x, 32)
+
+        (x1, _) = torch.max(x, dim=2)
+        x2 = torch.mean(x, dim=2)
+        x = x1 + x2
+        x = F.dropout(x, p=0.5, training=self.training)
+        x = F.relu_(self.fc1(x))
+        embedding = F.dropout(x, p=0.5, training=self.training)
+        clipwise_output = torch.sigmoid(self.fc_audioset(x))
+
+        output_dict = {
+            "clipwise_output": clipwise_output,
+            "embedding": embedding,
+            "fine_grained_embedding": latent_output,
+        }
+
+        return output_dict
+
+
+def create_pann_model(audio_cfg, enable_fusion=False, fusion_type="None"):
+    try:
+        ModelProto = eval(audio_cfg.model_name)
+        model = ModelProto(
+            sample_rate=audio_cfg.sample_rate,
+            window_size=audio_cfg.window_size,
+            hop_size=audio_cfg.hop_size,
+            mel_bins=audio_cfg.mel_bins,
+            fmin=audio_cfg.fmin,
+            fmax=audio_cfg.fmax,
+            classes_num=audio_cfg.class_num,
+            enable_fusion=enable_fusion,
+            fusion_type=fusion_type,
+        )
+        return model
+    except:
+        raise RuntimeError(
+            f"Import Model for {audio_cfg.model_name} not found, or the audio cfg parameters are not enough."
+        )

audioldm/clap/open_clip/pretrained.py

@@ -0,0 +1,167 @@
+import hashlib
+import os
+import urllib
+import warnings
+
+from tqdm import tqdm
+
+_RN50 = dict(
+    openai="https://openaipublic.azureedge.net/clip/models/afeb0e10f9e5a86da6080e35cf09123aca3b358a0c3e3b6c78a7b63bc04b6762/RN50.pt",
+    yfcc15m="https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/rn50-quickgelu-yfcc15m-455df137.pt",
+    cc12m="https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/rn50-quickgelu-cc12m-f000538c.pt",
+)
+
+_RN50_quickgelu = dict(
+    openai="https://openaipublic.azureedge.net/clip/models/afeb0e10f9e5a86da6080e35cf09123aca3b358a0c3e3b6c78a7b63bc04b6762/RN50.pt",
+    yfcc15m="https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/rn50-quickgelu-yfcc15m-455df137.pt",
+    cc12m="https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/rn50-quickgelu-cc12m-f000538c.pt",
+)
+
+_RN101 = dict(
+    openai="https://openaipublic.azureedge.net/clip/models/8fa8567bab74a42d41c5915025a8e4538c3bdbe8804a470a72f30b0d94fab599/RN101.pt",
+    yfcc15m="https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/rn101-quickgelu-yfcc15m-3e04b30e.pt",
+)
+
+_RN101_quickgelu = dict(
+    openai="https://openaipublic.azureedge.net/clip/models/8fa8567bab74a42d41c5915025a8e4538c3bdbe8804a470a72f30b0d94fab599/RN101.pt",
+    yfcc15m="https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/rn101-quickgelu-yfcc15m-3e04b30e.pt",
+)
+
+_RN50x4 = dict(
+    openai="https://openaipublic.azureedge.net/clip/models/7e526bd135e493cef0776de27d5f42653e6b4c8bf9e0f653bb11773263205fdd/RN50x4.pt",
+)
+
+_RN50x16 = dict(
+    openai="https://openaipublic.azureedge.net/clip/models/52378b407f34354e150460fe41077663dd5b39c54cd0bfd2b27167a4a06ec9aa/RN50x16.pt",
+)
+
+_RN50x64 = dict(
+    openai="https://openaipublic.azureedge.net/clip/models/be1cfb55d75a9666199fb2206c106743da0f6468c9d327f3e0d0a543a9919d9c/RN50x64.pt",
+)
+
+_VITB32 = dict(
+    openai="https://openaipublic.azureedge.net/clip/models/40d365715913c9da98579312b702a82c18be219cc2a73407c4526f58eba950af/ViT-B-32.pt",
+    laion400m_e31="https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/vit_b_32-quickgelu-laion400m_e31-d867053b.pt",
+    laion400m_e32="https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/vit_b_32-quickgelu-laion400m_e32-46683a32.pt",
+    laion400m_avg="https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/vit_b_32-quickgelu-laion400m_avg-8a00ab3c.pt",
+)
+
+_VITB32_quickgelu = dict(
+    openai="https://openaipublic.azureedge.net/clip/models/40d365715913c9da98579312b702a82c18be219cc2a73407c4526f58eba950af/ViT-B-32.pt",
+    laion400m_e31="https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/vit_b_32-quickgelu-laion400m_e31-d867053b.pt",
+    laion400m_e32="https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/vit_b_32-quickgelu-laion400m_e32-46683a32.pt",
+    laion400m_avg="https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/vit_b_32-quickgelu-laion400m_avg-8a00ab3c.pt",
+)
+
+_VITB16 = dict(
+    openai="https://openaipublic.azureedge.net/clip/models/5806e77cd80f8b59890b7e101eabd078d9fb84e6937f9e85e4ecb61988df416f/ViT-B-16.pt",
+)
+
+_VITL14 = dict(
+    openai="https://openaipublic.azureedge.net/clip/models/b8cca3fd41ae0c99ba7e8951adf17d267cdb84cd88be6f7c2e0eca1737a03836/ViT-L-14.pt",
+)
+
+_PRETRAINED = {
+    "RN50": _RN50,
+    "RN50-quickgelu": _RN50_quickgelu,
+    "RN101": _RN101,
+    "RN101-quickgelu": _RN101_quickgelu,
+    "RN50x4": _RN50x4,
+    "RN50x16": _RN50x16,
+    "ViT-B-32": _VITB32,
+    "ViT-B-32-quickgelu": _VITB32_quickgelu,
+    "ViT-B-16": _VITB16,
+    "ViT-L-14": _VITL14,
+}
+
+
+def list_pretrained(as_str: bool = False):
+    """returns list of pretrained models
+    Returns a tuple (model_name, pretrain_tag) by default or 'name:tag' if as_str == True
+    """
+    return [
+        ":".join([k, t]) if as_str else (k, t)
+        for k in _PRETRAINED.keys()
+        for t in _PRETRAINED[k].keys()
+    ]
+
+
+def list_pretrained_tag_models(tag: str):
+    """return all models having the specified pretrain tag"""
+    models = []
+    for k in _PRETRAINED.keys():
+        if tag in _PRETRAINED[k]:
+            models.append(k)
+    return models
+
+
+def list_pretrained_model_tags(model: str):
+    """return all pretrain tags for the specified model architecture"""
+    tags = []
+    if model in _PRETRAINED:
+        tags.extend(_PRETRAINED[model].keys())
+    return tags
+
+
+def get_pretrained_url(model: str, tag: str):
+    if model not in _PRETRAINED:
+        return ""
+    model_pretrained = _PRETRAINED[model]
+    if tag not in model_pretrained:
+        return ""
+    return model_pretrained[tag]
+
+
+def download_pretrained(url: str, root: str = os.path.expanduser("~/.cache/clip")):
+    os.makedirs(root, exist_ok=True)
+    filename = os.path.basename(url)
+
+    if "openaipublic" in url:
+        expected_sha256 = url.split("/")[-2]
+    else:
+        expected_sha256 = ""
+
+    download_target = os.path.join(root, filename)
+
+    if os.path.exists(download_target) and not os.path.isfile(download_target):
+        raise RuntimeError(f"{download_target} exists and is not a regular file")
+
+    if os.path.isfile(download_target):
+        if expected_sha256:
+            if (
+                hashlib.sha256(open(download_target, "rb").read()).hexdigest()
+                == expected_sha256
+            ):
+                return download_target
+            else:
+                warnings.warn(
+                    f"{download_target} exists, but the SHA256 checksum does not match; re-downloading the file"
+                )
+        else:
+            return download_target
+
+    with urllib.request.urlopen(url) as source, open(download_target, "wb") as output:
+        with tqdm(
+            total=int(source.info().get("Content-Length")),
+            ncols=80,
+            unit="iB",
+            unit_scale=True,
+        ) as loop:
+            while True:
+                buffer = source.read(8192)
+                if not buffer:
+                    break
+
+                output.write(buffer)
+                loop.update(len(buffer))
+
+    if (
+        expected_sha256
+        and hashlib.sha256(open(download_target, "rb").read()).hexdigest()
+        != expected_sha256
+    ):
+        raise RuntimeError(
+            f"Model has been downloaded but the SHA256 checksum does not not match"
+        )
+
+    return download_target

audioldm/clap/open_clip/timm_model.py

@@ -0,0 +1,112 @@
+""" timm model adapter
+
+Wraps timm (https://github.com/rwightman/pytorch-image-models) models for use as a vision tower in CLIP model.
+"""
+from collections import OrderedDict
+
+import torch.nn as nn
+
+try:
+    import timm
+    from timm.models.layers import Mlp, to_2tuple
+    from timm.models.layers.attention_pool2d import RotAttentionPool2d
+    from timm.models.layers.attention_pool2d import (
+        AttentionPool2d as AbsAttentionPool2d,
+    )
+except ImportError as e:
+    timm = None
+
+from .utils import freeze_batch_norm_2d
+
+
+class TimmModel(nn.Module):
+    """timm model adapter
+    # FIXME this adapter is a work in progress, may change in ways that break weight compat
+    """
+
+    def __init__(
+        self,
+        model_name,
+        embed_dim,
+        image_size=224,
+        pool="avg",
+        proj="linear",
+        drop=0.0,
+        pretrained=False,
+    ):
+        super().__init__()
+        if timm is None:
+            raise RuntimeError("Please `pip install timm` to use timm models.")
+
+        self.image_size = to_2tuple(image_size)
+        self.trunk = timm.create_model(model_name, pretrained=pretrained)
+        feat_size = self.trunk.default_cfg.get("pool_size", None)
+        feature_ndim = 1 if not feat_size else 2
+        if pool in ("abs_attn", "rot_attn"):
+            assert feature_ndim == 2
+            # if attn pooling used, remove both classifier and default pool
+            self.trunk.reset_classifier(0, global_pool="")
+        else:
+            # reset global pool if pool config set, otherwise leave as network default
+            reset_kwargs = dict(global_pool=pool) if pool else {}
+            self.trunk.reset_classifier(0, **reset_kwargs)
+        prev_chs = self.trunk.num_features
+
+        head_layers = OrderedDict()
+        if pool == "abs_attn":
+            head_layers["pool"] = AbsAttentionPool2d(
+                prev_chs, feat_size=feat_size, out_features=embed_dim
+            )
+            prev_chs = embed_dim
+        elif pool == "rot_attn":
+            head_layers["pool"] = RotAttentionPool2d(prev_chs, out_features=embed_dim)
+            prev_chs = embed_dim
+        else:
+            assert proj, "projection layer needed if non-attention pooling is used."
+
+        # NOTE attention pool ends with a projection layer, so proj should usually be set to '' if such pooling is used
+        if proj == "linear":
+            head_layers["drop"] = nn.Dropout(drop)
+            head_layers["proj"] = nn.Linear(prev_chs, embed_dim)
+        elif proj == "mlp":
+            head_layers["mlp"] = Mlp(prev_chs, 2 * embed_dim, embed_dim, drop=drop)
+
+        self.head = nn.Sequential(head_layers)
+
+    def lock(self, unlocked_groups=0, freeze_bn_stats=False):
+        """lock modules
+        Args:
+            unlocked_groups (int): leave last n layer groups unlocked (default: 0)
+        """
+        if not unlocked_groups:
+            # lock full model
+            for param in self.trunk.parameters():
+                param.requires_grad = False
+            if freeze_bn_stats:
+                freeze_batch_norm_2d(self.trunk)
+        else:
+            # NOTE: partial freeze requires latest timm (master) branch and is subject to change
+            try:
+                # FIXME import here until API stable and in an official release
+                from timm.models.helpers import group_parameters, group_modules
+            except ImportError:
+                raise RuntimeError(
+                    "Please install latest timm `pip install git+https://github.com/rwightman/pytorch-image-models`"
+                )
+            matcher = self.trunk.group_matcher()
+            gparams = group_parameters(self.trunk, matcher)
+            max_layer_id = max(gparams.keys())
+            max_layer_id = max_layer_id - unlocked_groups
+            for group_idx in range(max_layer_id + 1):
+                group = gparams[group_idx]
+                for param in group:
+                    self.trunk.get_parameter(param).requires_grad = False
+            if freeze_bn_stats:
+                gmodules = group_modules(self.trunk, matcher, reverse=True)
+                gmodules = {k for k, v in gmodules.items() if v <= max_layer_id}
+                freeze_batch_norm_2d(self.trunk, gmodules)
+
+    def forward(self, x):
+        x = self.trunk(x)
+        x = self.head(x)
+        return x

audioldm/clap/open_clip/tokenizer.py

@@ -0,0 +1,197 @@
+""" CLIP tokenizer
+
+Copied from https://github.com/openai/CLIP. Originally MIT License, Copyright (c) 2021 OpenAI.
+"""
+import gzip
+import html
+import os
+from functools import lru_cache
+from typing import Union, List
+
+import ftfy
+import regex as re
+import torch
+
+
+@lru_cache()
+def default_bpe():
+    return os.path.join(
+        os.path.dirname(os.path.abspath(__file__)), "bpe_simple_vocab_16e6.txt.gz"
+    )
+
+
+@lru_cache()
+def bytes_to_unicode():
+    """
+    Returns list of utf-8 byte and a corresponding list of unicode strings.
+    The reversible bpe codes work on unicode strings.
+    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
+    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
+    This is a signficant percentage of your normal, say, 32K bpe vocab.
+    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
+    And avoids mapping to whitespace/control characters the bpe code barfs on.
+    """
+    bs = (
+        list(range(ord("!"), ord("~") + 1))
+        + list(range(ord("¡"), ord("¬") + 1))
+        + list(range(ord("®"), ord("ÿ") + 1))
+    )
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8 + n)
+            n += 1
+    cs = [chr(n) for n in cs]
+    return dict(zip(bs, cs))
+
+
+def get_pairs(word):
+    """Return set of symbol pairs in a word.
+    Word is represented as tuple of symbols (symbols being variable-length strings).
+    """
+    pairs = set()
+    prev_char = word[0]
+    for char in word[1:]:
+        pairs.add((prev_char, char))
+        prev_char = char
+    return pairs
+
+
+def basic_clean(text):
+    text = ftfy.fix_text(text)
+    text = html.unescape(html.unescape(text))
+    return text.strip()
+
+
+def whitespace_clean(text):
+    text = re.sub(r"\s+", " ", text)
+    text = text.strip()
+    return text
+
+
+class SimpleTokenizer(object):
+    def __init__(self, bpe_path: str = default_bpe(), special_tokens=None):
+        self.byte_encoder = bytes_to_unicode()
+        self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}
+        merges = gzip.open(bpe_path).read().decode("utf-8").split("\n")
+        merges = merges[1 : 49152 - 256 - 2 + 1]
+        merges = [tuple(merge.split()) for merge in merges]
+        vocab = list(bytes_to_unicode().values())
+        vocab = vocab + [v + "</w>" for v in vocab]
+        for merge in merges:
+            vocab.append("".join(merge))
+        if not special_tokens:
+            special_tokens = ["<start_of_text>", "<end_of_text>"]
+        else:
+            special_tokens = ["<start_of_text>", "<end_of_text>"] + special_tokens
+        vocab.extend(special_tokens)
+        self.encoder = dict(zip(vocab, range(len(vocab))))
+        self.decoder = {v: k for k, v in self.encoder.items()}
+        self.bpe_ranks = dict(zip(merges, range(len(merges))))
+        self.cache = {t: t for t in special_tokens}
+        special = "|".join(special_tokens)
+        self.pat = re.compile(
+            special + r"""|'s|'t|'re|'ve|'m|'ll|'d|[\p{L}]+|[\p{N}]|[^\s\p{L}\p{N}]+""",
+            re.IGNORECASE,
+        )
+
+        self.vocab_size = len(self.encoder)
+        self.all_special_ids = [self.encoder[t] for t in special_tokens]
+
+    def bpe(self, token):
+        if token in self.cache:
+            return self.cache[token]
+        word = tuple(token[:-1]) + (token[-1] + "</w>",)
+        pairs = get_pairs(word)
+
+        if not pairs:
+            return token + "</w>"
+
+        while True:
+            bigram = min(pairs, key=lambda pair: self.bpe_ranks.get(pair, float("inf")))
+            if bigram not in self.bpe_ranks:
+                break
+            first, second = bigram
+            new_word = []
+            i = 0
+            while i < len(word):
+                try:
+                    j = word.index(first, i)
+                    new_word.extend(word[i:j])
+                    i = j
+                except:
+                    new_word.extend(word[i:])
+                    break
+
+                if word[i] == first and i < len(word) - 1 and word[i + 1] == second:
+                    new_word.append(first + second)
+                    i += 2
+                else:
+                    new_word.append(word[i])
+                    i += 1
+            new_word = tuple(new_word)
+            word = new_word
+            if len(word) == 1:
+                break
+            else:
+                pairs = get_pairs(word)
+        word = " ".join(word)
+        self.cache[token] = word
+        return word
+
+    def encode(self, text):
+        bpe_tokens = []
+        text = whitespace_clean(basic_clean(text)).lower()
+        for token in re.findall(self.pat, text):
+            token = "".join(self.byte_encoder[b] for b in token.encode("utf-8"))
+            bpe_tokens.extend(
+                self.encoder[bpe_token] for bpe_token in self.bpe(token).split(" ")
+            )
+        return bpe_tokens
+
+    def decode(self, tokens):
+        text = "".join([self.decoder[token] for token in tokens])
+        text = (
+            bytearray([self.byte_decoder[c] for c in text])
+            .decode("utf-8", errors="replace")
+            .replace("</w>", " ")
+        )
+        return text
+
+
+_tokenizer = SimpleTokenizer()
+
+
+def tokenize(
+    texts: Union[str, List[str]], context_length: int = 77
+) -> torch.LongTensor:
+    """
+    Returns the tokenized representation of given input string(s)
+
+    Parameters
+    ----------
+    texts : Union[str, List[str]]
+        An input string or a list of input strings to tokenize
+    context_length : int
+        The context length to use; all CLIP models use 77 as the context length
+
+    Returns
+    -------
+    A two-dimensional tensor containing the resulting tokens, shape = [number of input strings, context_length]
+    """
+    if isinstance(texts, str):
+        texts = [texts]
+
+    sot_token = _tokenizer.encoder["<start_of_text>"]
+    eot_token = _tokenizer.encoder["<end_of_text>"]
+    all_tokens = [[sot_token] + _tokenizer.encode(text) + [eot_token] for text in texts]
+    result = torch.zeros(len(all_tokens), context_length, dtype=torch.long)
+
+    for i, tokens in enumerate(all_tokens):
+        if len(tokens) > context_length:
+            tokens = tokens[:context_length]  # Truncate
+        result[i, : len(tokens)] = torch.tensor(tokens)
+
+    return result

audioldm/clap/open_clip/transform.py

@@ -0,0 +1,45 @@
+from torchvision.transforms import (
+    Normalize,
+    Compose,
+    RandomResizedCrop,
+    InterpolationMode,
+    ToTensor,
+    Resize,
+    CenterCrop,
+)
+
+
+def _convert_to_rgb(image):
+    return image.convert("RGB")
+
+
+def image_transform(
+    image_size: int,
+    is_train: bool,
+    mean=(0.48145466, 0.4578275, 0.40821073),
+    std=(0.26862954, 0.26130258, 0.27577711),
+):
+    normalize = Normalize(mean=mean, std=std)
+    if is_train:
+        return Compose(
+            [
+                RandomResizedCrop(
+                    image_size,
+                    scale=(0.9, 1.0),
+                    interpolation=InterpolationMode.BICUBIC,
+                ),
+                _convert_to_rgb,
+                ToTensor(),
+                normalize,
+            ]
+        )
+    else:
+        return Compose(
+            [
+                Resize(image_size, interpolation=InterpolationMode.BICUBIC),
+                CenterCrop(image_size),
+                _convert_to_rgb,
+                ToTensor(),
+                normalize,
+            ]
+        )

audioldm/clap/open_clip/utils.py

@@ -0,0 +1,361 @@
+import numpy as np
+import torch
+from torch import nn as nn
+from torchvision.ops.misc import FrozenBatchNorm2d
+import logging
+# import h5py
+from tqdm import tqdm
+import random
+import json
+import os
+import pathlib
+
+# TODO: (yusong) this not a good place to store those information and does not scale. Need to be fixed later.
+dataset_split = {
+    "audiocaps": ["train", "valid", "test"],
+    "audioset": ["balanced_train", "unbalanced_train", "eval"],
+    "BBCSoundEffects": ["train", "test"],
+    "Clotho": ["train", "test", "valid"],
+    "free_to_use_sounds": ["train", "test"],
+    "paramount_motion": ["train", "test"],
+    "sonniss_game_effects": ["train", "test"],
+    "wesoundeffects": ["train", "test"],
+    "MACS": ["train", "test"],
+    "freesound": ["train", "test"],
+    "FSD50K": ["train", "test", "valid"],
+    "fsd50k_class_label": ["train", "test", "valid"],
+    "esc50": ["train", "test"],
+    "audiostock": ["train", "test"],
+    "freesound_no_overlap_noesc50": ["train", "test"],
+    "epidemic_sound_effects": ["train", "test"],
+    "VGGSound": ["train", "test"],
+    "urbansound8k_class_label": ["train", "test"],
+    "audioset_t5": ["balanced_train", "unbalanced_train", "eval"],
+    "epidemic_sound_effects_t5": ["train", "test"],
+    "WavText5K": ["train", "test"],
+    "esc50_no_overlap": ["train", "test"],
+    "usd8k_no_overlap": ["train", "test"],
+    "fsd50k_200_class_label": ["train", "test", "valid"],
+}
+
+
+def freeze_batch_norm_2d(module, module_match={}, name=""):
+    """
+    Converts all `BatchNorm2d` and `SyncBatchNorm` layers of provided module into `FrozenBatchNorm2d`. If `module` is
+    itself an instance of either `BatchNorm2d` or `SyncBatchNorm`, it is converted into `FrozenBatchNorm2d` and
+    returned. Otherwise, the module is walked recursively and submodules are converted in place.
+
+    Args:
+        module (torch.nn.Module): Any PyTorch module.
+        module_match (dict): Dictionary of full module names to freeze (all if empty)
+        name (str): Full module name (prefix)
+
+    Returns:
+        torch.nn.Module: Resulting module
+
+    Inspired by https://github.com/pytorch/pytorch/blob/a5895f85be0f10212791145bfedc0261d364f103/torch/nn/modules/batchnorm.py#L762
+    """
+    res = module
+    is_match = True
+    if module_match:
+        is_match = name in module_match
+    if is_match and isinstance(
+        module, (nn.modules.batchnorm.BatchNorm2d, nn.modules.batchnorm.SyncBatchNorm)
+    ):
+        res = FrozenBatchNorm2d(module.num_features)
+        res.num_features = module.num_features
+        res.affine = module.affine
+        if module.affine:
+            res.weight.data = module.weight.data.clone().detach()
+            res.bias.data = module.bias.data.clone().detach()
+        res.running_mean.data = module.running_mean.data
+        res.running_var.data = module.running_var.data
+        res.eps = module.eps
+    else:
+        for child_name, child in module.named_children():
+            full_child_name = ".".join([name, child_name]) if name else child_name
+            new_child = freeze_batch_norm_2d(child, module_match, full_child_name)
+            if new_child is not child:
+                res.add_module(child_name, new_child)
+    return res
+
+
+def exist(dataset_name, dataset_type):
+    """
+    Check if dataset exists
+    """
+    if dataset_type in dataset_split[dataset_name]:
+        return True
+    else:
+        return False
+
+
+def get_tar_path_from_dataset_name(
+    dataset_names, dataset_types, islocal, dataset_path, proportion=1, full_dataset=None
+):
+    """
+    Get tar path from dataset name and type
+    """
+    output = []
+    for n in dataset_names:
+        if full_dataset is not None and n in full_dataset:
+            current_dataset_types = dataset_split[n]
+        else:
+            current_dataset_types = dataset_types
+        for s in current_dataset_types:
+            tmp = []
+            if islocal:
+                sizefilepath_ = f"{dataset_path}/{n}/{s}/sizes.json"
+                if not os.path.exists(sizefilepath_):
+                    sizefilepath_ = f"./json_files/{n}/{s}/sizes.json"
+            else:
+                sizefilepath_ = f"./json_files/{n}/{s}/sizes.json"
+            if not os.path.exists(sizefilepath_):
+                continue
+            sizes = json.load(open(sizefilepath_, "r"))
+            for k in sizes.keys():
+                if islocal:
+                    tmp.append(f"{dataset_path}/{n}/{s}/{k}")
+                else:
+                    tmp.append(
+                        f"pipe:aws s3 --cli-connect-timeout 0 cp s3://s-laion-audio/webdataset_tar/{n}/{s}/{k} -"
+                    )
+            if proportion != 1:
+                tmp = random.sample(tmp, int(proportion * len(tmp)))
+            output.append(tmp)
+    return sum(output, [])
+
+
+def get_tar_path_from_txts(txt_path, islocal, proportion=1):
+    """
+    Get tar path from txt path
+    """
+    if isinstance(txt_path, (list, tuple)):
+        return sum(
+            [
+                get_tar_path_from_txts(
+                    txt_path[i], islocal=islocal, proportion=proportion
+                )
+                for i in range(len(txt_path))
+            ],
+            [],
+        )
+    if isinstance(txt_path, str):
+        with open(txt_path) as f:
+            lines = f.readlines()
+        if islocal:
+            lines = [
+                lines[i]
+                .split("\n")[0]
+                .replace("pipe:aws s3 cp s3://s-laion-audio/", "/mnt/audio_clip/")
+                for i in range(len(lines))
+            ]
+        else:
+            lines = [
+                lines[i].split("\n")[0].replace(".tar", ".tar -")
+                for i in range(len(lines))
+            ]
+        if proportion != 1:
+            print("Sampling tars with proportion of {}".format(proportion))
+            lines = random.sample(lines, int(proportion * len(lines)))
+        return lines
+
+
+def get_mix_lambda(mixup_alpha, batch_size):
+    mixup_lambdas = [
+        np.random.beta(mixup_alpha, mixup_alpha, 1)[0] for _ in range(batch_size)
+    ]
+    return np.array(mixup_lambdas).astype(np.float32)
+
+
+def do_mixup(x, mixup_lambda):
+    """
+    Args:
+      x: (batch_size , ...)
+      mixup_lambda: (batch_size,)
+    Returns:
+      out: (batch_size, ...)
+    """
+    out = (
+        x.transpose(0, -1) * mixup_lambda
+        + torch.flip(x, dims=[0]).transpose(0, -1) * (1 - mixup_lambda)
+    ).transpose(0, -1)
+    return out
+
+
+def interpolate(x, ratio):
+    """Interpolate data in time domain. This is used to compensate the
+    resolution reduction in downsampling of a CNN.
+
+    Args:
+      x: (batch_size, time_steps, classes_num)
+      ratio: int, ratio to interpolate
+    Returns:
+      upsampled: (batch_size, time_steps * ratio, classes_num)
+    """
+    (batch_size, time_steps, classes_num) = x.shape
+    upsampled = x[:, :, None, :].repeat(1, 1, ratio, 1)
+    upsampled = upsampled.reshape(batch_size, time_steps * ratio, classes_num)
+    return upsampled
+
+
+def pad_framewise_output(framewise_output, frames_num):
+    """Pad framewise_output to the same length as input frames. The pad value
+    is the same as the value of the last frame.
+    Args:
+      framewise_output: (batch_size, frames_num, classes_num)
+      frames_num: int, number of frames to pad
+    Outputs:
+      output: (batch_size, frames_num, classes_num)
+    """
+    pad = framewise_output[:, -1:, :].repeat(
+        1, frames_num - framewise_output.shape[1], 1
+    )
+    """tensor for padding"""
+
+    output = torch.cat((framewise_output, pad), dim=1)
+    """(batch_size, frames_num, classes_num)"""
+
+
+# def process_ipc(index_path, classes_num, filename):
+#     # load data
+#     logging.info("Load Data...............")
+#     ipc = [[] for _ in range(classes_num)]
+#     with h5py.File(index_path, "r") as f:
+#         for i in tqdm(range(len(f["target"]))):
+#             t_class = np.where(f["target"][i])[0]
+#             for t in t_class:
+#                 ipc[t].append(i)
+#     print(ipc)
+#     np.save(filename, ipc)
+#     logging.info("Load Data Succeed...............")
+
+
+def save_to_dict(s, o_={}):
+    sp = s.split(": ")
+    o_.update({sp[0]: float(sp[1])})
+    return o_
+
+
+def get_data_from_log(txt_path):
+    """
+    Output dictionary from out.txt log file
+    """
+    with open(txt_path) as f:
+        lines = f.readlines()
+    val_data = {}
+    train_data = {}
+    train_losses = []
+    train_losses_epoch = []
+    for i in range(len(lines)):
+        if "| INFO |" in lines[i]:
+            if "Eval Epoch" in lines[i]:
+                if "val_loss" in lines[i]:
+                    # float(regex.sub("", lines[310].split("	")[-1]).replace(" ", ""))
+                    line = lines[i].split("Eval Epoch: ")[-1]
+                    num_epoch = int(line.split("	")[0].split(" ")[0])
+                    d = {
+                        line.split("	")[0]
+                        .split(" ")[1]
+                        .replace(":", ""): float(line.split("	")[0].split(" ")[-1])
+                    }
+                    for i in range(1, len(line.split("	"))):
+                        d = save_to_dict(line.split("	")[i], d)
+                    val_data[num_epoch] = d
+            elif "Train Epoch" in lines[i]:
+                num_epoch = int(lines[i].split("Train Epoch: ")[1][0])
+                loss = float(lines[i].split("Loss: ")[-1].split(" (")[0])
+                train_losses.append(loss)
+                train_losses_epoch.append(num_epoch)
+    for i in range(len(train_losses)):
+        train_data[i] = {
+            "num_epoch": train_losses_epoch[i],
+            "train_loss": train_losses[i],
+        }
+    return train_data, val_data
+
+
+def save_p(obj, filename):
+    import pickle
+
+    try:
+        from deepdiff import DeepDiff
+    except:
+        os.system("pip install deepdiff")
+        from deepdiff import DeepDiff
+    with open(filename, "wb") as file:
+        pickle.dump(obj, file, protocol=pickle.HIGHEST_PROTOCOL)  # highest protocol
+    with open(filename, "rb") as file:
+        z = pickle.load(file)
+    assert (
+        DeepDiff(obj, z, ignore_string_case=True) == {}
+    ), "there is something wrong with the saving process"
+    return
+
+
+def load_p(filename):
+    import pickle
+
+    with open(filename, "rb") as file:
+        z = pickle.load(file)
+    return z
+
+
+def save_json(data, name="data.json"):
+    import json
+
+    with open(name, "w") as fp:
+        json.dump(data, fp)
+    return
+
+
+def load_json(name):
+    import json
+
+    with open(name, "r") as fp:
+        data = json.load(fp)
+    return data
+
+
+from multiprocessing import Process, Manager
+from multiprocessing import Process, Value, Array
+from ctypes import c_wchar
+
+
+def load_class_label(path):
+    # https://stackoverflow.com/questions/48004243/how-to-share-large-read-only-dictionary-list-across-processes-in-multiprocessing
+    # https://stackoverflow.com/questions/45693949/storing-strings-in-a-multiprocessing-sharedctypes-array
+    out = None
+    if path is not None:
+        if pathlib.Path(path).suffix in [".pkl", ".pickle"]:
+            out = load_p(path)
+        elif pathlib.Path(path).suffix in [".json", ".txt"]:
+            out = load_json(path)
+        elif pathlib.Path(path).suffix in [".npy", ".npz"]:
+            out = np.load(path)
+        elif pathlib.Path(path).suffix in [".csv"]:
+            import pandas as pd
+
+            out = pd.read_csv(path)
+    return out
+    # if out is None:
+    #     return None
+    # else:
+    #     key = Array(c_wchar, '\n'.join(list(out.keys())), lock=False)
+    #     val = Array('i', out.values(), lock=False)
+    #     return (key, val)
+
+
+from torch import optim
+
+
+def get_optimizer(params, lr, betas, eps, momentum, optimizer_name):
+    if optimizer_name.lower() == "adamw":
+        optimizer = optim.AdamW(params, lr=lr, betas=betas, eps=eps)
+    elif optimizer_name.lower() == "sgd":
+        optimizer = optim.SGD(params, lr=lr, momentum=momentum)
+    elif optimizer_name.lower() == "adam":
+        optimizer = optim.Adam(params, lr=lr, betas=betas, eps=eps)
+    else:
+        raise ValueError("optimizer name is not correct")
+    return optimizer

audioldm/clap/open_clip/version.py

@@ -0,0 +1 @@
+__version__ = "0.2.1"