Swap to HF diffusers (#258)

- Swap to HF diffusers (d03edfafdf371af210d3a8c062f61da2683d3e7c)
- camera ready (f1daa6039ead8b008c9f9a424f88cc6ce5d7ae1e)


Co-authored-by: Sanchit Gandhi <sanchit-gandhi@users.noreply.huggingface.co>

README.md

@@ -8,6 +8,7 @@ sdk_version: 3.27.0
 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

app.py

@@ -1,197 +1,139 @@
 import gradio as gr
-import numpy as np
-from audioldm import text_to_audio, build_model
+import torch
+from diffusers import AudioLDMPipeline
 from share_btn import community_icon_html, loading_icon_html, share_js
 
-model_id="haoheliu/AudioLDM-S-Full"
+from transformers import AutoProcessor, ClapModel
 
-audioldm = None
-current_model_name = None
 
-# def predict(input, history=[]):
-#     # tokenize the new input sentence
-#     new_user_input_ids = tokenizer.encode(input + tokenizer.eos_token, return_tensors='pt')
+# make Space compatible with CPU duplicates
+if torch.cuda.is_available():
+    device = "cuda"
+    torch_dtype = torch.float16
+else:
+    device = "cpu"
+    torch_dtype = torch.float32
 
-#     # append the new user input tokens to the chat history
-#     bot_input_ids = torch.cat([torch.LongTensor(history), new_user_input_ids], dim=-1)
+# load the diffusers pipeline
+repo_id = "cvssp/audioldm-m-full"
+pipe = AudioLDMPipeline.from_pretrained(repo_id, torch_dtype=torch_dtype).to(device)
+pipe.unet = torch.compile(pipe.unet)
 
-#     # generate a response 
-#     history = model.generate(bot_input_ids, max_length=1000, pad_token_id=tokenizer.eos_token_id).tolist()
+# CLAP model (only required for automatic scoring)
+clap_model = ClapModel.from_pretrained("sanchit-gandhi/clap-htsat-unfused-m-full").to(device)
+processor = AutoProcessor.from_pretrained("sanchit-gandhi/clap-htsat-unfused-m-full")
 
-#     # 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, model_name="audioldm-m-text-ft"):
-    global audioldm, current_model_name
-    
-    if audioldm is None or model_name != current_model_name:
-        audioldm=build_model(model_name=model_name)
-        current_model_name = model_name
-        
-    # print(text, length, guidance_scale)
-    waveform = text_to_audio(
-        latent_diffusion=audioldm,
-        text=text,
-        seed=random_seed,
-        duration=duration,
+generator = torch.Generator(device)
+
+
+def text2audio(text, negative_prompt, duration, guidance_scale, random_seed, n_candidates):
+    if text is None:
+        raise gr.Error("Please provide a text input.")
+
+    waveforms = pipe(
+        text,
+        audio_length_in_s=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
+        negative_prompt=negative_prompt,
+        num_waveforms_per_prompt=n_candidates if n_candidates else 1,
+        generator=generator.manual_seed(int(random_seed)),
+    )["audios"]
+
+    if waveforms.shape[0] > 1:
+        waveform = score_waveforms(text, waveforms)
+    else:
+        waveform = waveforms[0]
+
+    return gr.make_waveform((16000, waveform), bg_image="bg.png")
+
 
-# 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)
+def score_waveforms(text, waveforms):
+    inputs = processor(text=text, audios=list(waveforms), return_tensors="pt", padding=True)
+    inputs = {key: inputs[key].to(device) for key in inputs}
+    with torch.no_grad():
+        logits_per_text = clap_model(**inputs).logits_per_text  # this is the audio-text similarity score
+        probs = logits_per_text.softmax(dim=-1)  # we can take the softmax to get the label probabilities
+        most_probable = torch.argmax(probs)  # and now select the most likely audio waveform
+    waveform = waveforms[most_probable]
+    return waveform
 
 
 css = """
         a {
-            color: inherit;
-            text-decoration: underline;
-        }
-        .gradio-container {
+            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'] {
+        } .gr-button {
+            color: white; border-color: #000000; background: #000000;
+        } input[type='range'] {
             accent-color: #000000;
-        }
-        .dark input[type='range'] {
+        } .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 {
+        } .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 {
+        } .details:hover {
             text-decoration: underline;
-        }
-        .gr-button {
+        } .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;
+        } .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 {
+        } #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 {
+        } .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 {
+        } .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 {
+        } .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 {
+        } @keyframes spin {
             from {
                 transform: rotate(0deg);
-            }
-            to {
+            } 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 * {
+        } #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 {
+        } #share-btn-container div:nth-child(-n+2){
+            width: auto !important; min-height: 0px !important;
+        } #share-btn-container .wrap {
             display: none !important;
-        }
-        .gr-form{
+        } .gr-form{
             flex: 1 1 50%; border-top-right-radius: 0; border-bottom-right-radius: 0;
-        }
-        #prompt-container{
+        } #prompt-container{
             gap: 0;
-        }
-        #generated_id{
+        } #generated_id{
             min-height: 700px
-        }
-        #setting_id{
-          margin-bottom: 12px;
-          text-align: center;
-          font-weight: 900;
+        } #setting_id{
+          margin-bottom: 12px; text-align: center; font-weight: 900;
         }
 """
 iface = gr.Blocks(css=css)
@@ -202,56 +144,72 @@ with iface:
             <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;
+                  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>
+              </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> <a href="https://huggingface.co/docs/diffusers/main/en/api/pipelines/audioldm">[🧨
+                Diffusers]</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/>
-    """)
+    gr.HTML(
+        """
+        <p>This is the demo for AudioLDM, powered by 🧨 Diffusers. Demo uses the checkpoint <a
+        href="https://huggingface.co/cvssp/audioldm-m-full"> audioldm-m-full </a>. For faster inference without waiting in
+        queue, you may duplicate the space and upgrade to a GPU in the 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")
+            textbox = gr.Textbox(
+                value="A hammer is hitting a wooden surface",
+                max_lines=1,
+                label="Input text",
+                info="Your text is important for the audio quality. Please ensure it is descriptive by using more adjectives.",
+                elem_id="prompt-in",
+            )
+            negative_textbox = gr.Textbox(
+                value="low quality, average quality",
+                max_lines=1,
+                label="Negative prompt",
+                info="Enter a negative prompt not to guide the audio generation. Selecting appropriate negative prompts can improve the audio quality significantly.",
+                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=5, 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, 3, 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")
-              # model_name = gr.Dropdown(
-              #       ["audioldm-m-text-ft", "audioldm-s-text-ft", "audioldm-m-full","audioldm-s-full-v2", "audioldm-s-full", "audioldm-l-full"], value="audioldm-m-full", label="Choose the model to use. audioldm-m-text-ft and audioldm-s-text-ft are recommanded. -s- means small, -m- means medium and -l- means large",
-              #   )
-            ############# 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")]
+                seed = gr.Number(
+                    value=45,
+                    label="Seed",
+                    info="Change this value (any integer number) will lead to a different generation result.",
+                )
+                duration = gr.Slider(2.5, 10, value=5, step=2.5, label="Duration (seconds)")
+                guidance_scale = gr.Slider(
+                    0,
+                    4,
+                    value=2.5,
+                    step=0.5,
+                    label="Guidance scale",
+                    info="Large => better quality and relevancy to text; Small => better diversity",
+                )
+                n_candidates = gr.Slider(
+                    1,
+                    3,
+                    value=3,
+                    step=1,
+                    label="Number waveforms to generate",
+                    info="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",
+                )
+
+            outputs = gr.Video(label="Output", elem_id="output-video")
             btn = gr.Button("Submit").style(full_width=True)
 
         with gr.Group(elem_id="share-btn-container", visible=False):
@@ -259,51 +217,61 @@ with iface:
             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, model_name], outputs=[outputs])
-        btn.click(text2audio, inputs=[
-                  textbox, duration, guidance_scale, seed, n_candidates], outputs=[outputs])
-        
+        btn.click(
+            text2audio,
+            inputs=[textbox, negative_textbox, duration, guidance_scale, seed, n_candidates],
+            outputs=[outputs],
+        )
+
         share_button.click(None, [], [], _js=share_js)
-        gr.HTML('''
+        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>
+                    <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>. Code and demo by 🤗 Hugging Face.</p> <br>
         </div>
-        ''')
-        gr.Examples([
-            ["A hammer is hitting a wooden surface", 5, 2.5, 45, 3, "audioldm-m-full"],
-            ["Peaceful and calming ambient music with singing bowl and other instruments.", 5, 2.5, 45, 3, "audioldm-m-full"],
-            ["A man is speaking in a small room.", 5, 2.5, 45, 3, "audioldm-m-full"],
-            ["A female is speaking followed by footstep sound", 5, 2.5, 45, 3, "audioldm-m-full"],
-            ["Wooden table tapping sound followed by water pouring sound.", 5, 2.5, 45, 3, "audioldm-m-full"],
-        ],
+        """
+        )
+        gr.Examples(
+            [
+                ["A hammer is hitting a wooden surface", "low quality, average quality", 5, 2.5, 45, 3],
+                ["Peaceful and calming ambient music with singing bowl and other instruments.", "low quality, average quality", 5, 2.5, 45, 3],
+                ["A man is speaking in a small room.", "low quality, average quality", 5, 2.5, 45, 3],
+                ["A female is speaking followed by footstep sound", "low quality, average quality", 5, 2.5, 45, 3],
+                ["Wooden table tapping sound followed by water pouring sound.", "low quality, average quality", 5, 2.5, 45, 3],
+            ],
             fn=text2audio,
-            # inputs=[textbox, duration, guidance_scale, seed, n_candidates, model_name],
-            inputs=[textbox, duration, guidance_scale, seed, n_candidates],
+            inputs=[textbox, negative_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>
-                ''')
+        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> <p>4. Using a negative prompt to not guide the diffusion process can improve the
+                audio quality significantly. Try using negative prompts like 'low quality'.</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>
+                    <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(max_size=10).launch(debug=True)
-# iface.launch(debug=True, share=True)

audioldm/__init__.py

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

audioldm/audio/audio_processing.py

@@ -1,100 +0,0 @@
-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

@@ -1,180 +0,0 @@
-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

@@ -1,33 +0,0 @@
-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

@@ -1,170 +0,0 @@
-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",
-        amodel = "HTSAT-tiny",
-        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 = amodel
-        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

@@ -1,25 +0,0 @@
-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

@@ -1,40 +0,0 @@
-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

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

audioldm/clap/open_clip/factory.py

@@ -1,277 +0,0 @@
-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

@@ -1,192 +0,0 @@
-"""
-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)