Update app.py

app.py

@@ -6,14 +6,14 @@ import librosa
 import numpy as np
 import torch
 
-from transformers import SpeechT5Processor, SpeechT5ForTextToSpeech, SpeechT5HifiGan
+# from transformers import SpeechT5Processor, SpeechT5ForTextToSpeech, SpeechT5HifiGan
 from transformers import AutoProcessor, AutoModelForCausalLM
 
 
-checkpoint = "microsoft/speecht5_tts"
-tts_processor = SpeechT5Processor.from_pretrained(checkpoint)
-tts_model = SpeechT5ForTextToSpeech.from_pretrained(checkpoint)
-vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan")
+# checkpoint = "microsoft/speecht5_tts"
+# tts_processor = SpeechT5Processor.from_pretrained(checkpoint)
+# tts_model = SpeechT5ForTextToSpeech.from_pretrained(checkpoint)
+# vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan")
 
 # ic_processor = AutoProcessor.from_pretrained("microsoft/git-base")
 # ic_model = AutoModelForCausalLM.from_pretrained("microsoft/git-base")
@@ -21,40 +21,24 @@ vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan")
 ic_processor = AutoProcessor.from_pretrained("ronniet/git-base-env")
 ic_model = AutoModelForCausalLM.from_pretrained("ronniet/git-base-env")
 
-def tts(text):
-    if len(text.strip()) == 0:
-        return (16000, np.zeros(0).astype(np.int16))
+# def tts(text):
+#     if len(text.strip()) == 0:
+#         return (16000, np.zeros(0).astype(np.int16))
 
-    inputs = tts_processor(text=text, return_tensors="pt")
+#     inputs = tts_processor(text=text, return_tensors="pt")
 
-    # limit input length
-    input_ids = inputs["input_ids"]
-    input_ids = input_ids[..., :tts_model.config.max_text_positions]
+#     # limit input length
+#     input_ids = inputs["input_ids"]
+#     input_ids = input_ids[..., :tts_model.config.max_text_positions]
 
-    # if speaker == "Surprise Me!":
-    #     # load one of the provided speaker embeddings at random
-    #     idx = np.random.randint(len(speaker_embeddings))
-    #     key = list(speaker_embeddings.keys())[idx]
-    #     speaker_embedding = np.load(speaker_embeddings[key])
+#     speaker_embedding = np.load("cmu_us_bdl_arctic-wav-arctic_a0009.npy")
 
-    #     # randomly shuffle the elements
-    #     np.random.shuffle(speaker_embedding)
+#     speaker_embedding = torch.tensor(speaker_embedding).unsqueeze(0)
 
-    #     # randomly flip half the values
-    #     x = (np.random.rand(512) >= 0.5) * 1.0
-    #     x[x == 0] = -1.0
-    #     speaker_embedding *= x
+#     speech = tts_model.generate_speech(input_ids, speaker_embedding, vocoder=vocoder)
 
-        #speaker_embedding = np.random.rand(512).astype(np.float32) * 0.3 - 0.15
-    # else:
-    speaker_embedding = np.load("cmu_us_bdl_arctic-wav-arctic_a0009.npy")
-
-    speaker_embedding = torch.tensor(speaker_embedding).unsqueeze(0)
-
-    speech = tts_model.generate_speech(input_ids, speaker_embedding, vocoder=vocoder)
-
-    speech = (speech.numpy() * 32767).astype(np.int16)
-    return (16000, speech)
+#     speech = (speech.numpy() * 32767).astype(np.int16)
+#     return (16000, speech)
 
 
 # captioner = pipeline(model="microsoft/git-base")
@@ -71,16 +55,16 @@ def predict(image):
     text_ids = ic_model.generate(pixel_values=pixel_values, max_length=50)
     text = ic_processor.batch_decode(text_ids, skip_special_tokens=True)[0]
     
-    audio = tts(text)
+    # audio = tts(text)
     
-    return text, audio
+    return text
 
 # theme = gr.themes.Default(primary_hue="#002A5B")
 
 demo = gr.Interface(
     fn=predict,
     inputs=gr.Image(type="pil",label="Environment"),
-    outputs=[gr.Textbox(label="Caption"), gr.Audio(type="numpy",label="Audio Feedback")],
+    outputs=gr.Textbox(label="Caption"),
     css=".gradio-container {background-color: #002A5B}",
     theme=gr.themes.Soft()
 )