Adding EleutherAI's own API

duplex.py

@@ -48,12 +48,12 @@ tts_model_name = "facebook/tts_transformer-es-css10"
 speak_es = gr.Interface.load(f"huggingface/{tts_model_name}", api_key=HF_AUTH_TOKEN)
 transcribe_es = lambda input_file: asr_es(input_file, chunk_length_s=5, stride_length_s=1)["text"]
 def generate_es(text, **kwargs):
-    # max_length=100, top_k=100, top_p=50, temperature=0.95, do_sample=True, do_clean=True
+    # text="Promtp", max_length=100, top_k=100, top_p=50, temperature=0.95, do_sample=True, do_clean=True
     api_uri = "https://hf.space/embed/bertin-project/bertin-gpt-j-6B/+/api/predict/"
-    response = requests.post(api_uri, data=json.dumps({"data": [text, 100, 100, 50, 0.95, True, True]}))
+    response = requests.post(api_uri, data=json.dumps({"data": [text, kwargs["max_length"], 100, 50, 0.95, True, True]}))
     if response.ok:
         if DEBUG:
-            print(response.json())
+            print("Spanish response >", response.json())
         return response.json()["data"][0]
     else:
         return ""
@@ -71,17 +71,22 @@ asr_en = pipeline(
 tts_model_name = "facebook/fastspeech2-en-ljspeech"
 speak_en = gr.Interface.load(f"huggingface/{tts_model_name}", api_key=HF_AUTH_TOKEN)
 transcribe_en = lambda input_file: asr_en(input_file, chunk_length_s=5, stride_length_s=1)["text"]
-generate_iface = gr.Interface.load("huggingface/EleutherAI/gpt-j-6B", api_key=HF_AUTH_TOKEN)
+# generate_iface = gr.Interface.load("huggingface/EleutherAI/gpt-j-6B", api_key=HF_AUTH_TOKEN)
 
 empty_audio = 'empty.flac'
 sf.write(empty_audio, [], 16000)
 deuncase = gr.Interface.load("huggingface/pere/DeUnCaser", api_key=HF_AUTH_TOKEN)
 
 def generate_en(text, **kwargs):
-    response = generate_iface(text)
-    if DEBUG:
-        print(response)
-    return response or ""
+    api_uri = "https://api.eleuther.ai/completion"
+    #--data-raw '{"context":"Promtp","top_p":0.9,"temp":0.8,"response_length":128,"remove_input":true}'
+    response = requests.post(api_uri, data=json.dumps({"context": text, "top_p": 0.9, "temp": 0.8, "response_length": kwargs["max_length"], "remove_input": True}))
+    if response.ok:
+        if DEBUG:
+            print("English response >", response.json())
+        return response.json()[0]["generated_text"].lstrip()
+    else:
+        return ""
 
 
 def select_lang(lang):
@@ -135,7 +140,7 @@ def chat_with_gpt(lang, agent, user, context, audio_in, history):
     # agent = AGENT
     # user = USER
     generation_kwargs = {
-        "max_length": 25,
+        "max_length": 50,
         # "top_k": top_k,
         # "top_p": top_p,
         # "temperature": temperature,
@@ -159,14 +164,18 @@ def chat_with_gpt(lang, agent, user, context, audio_in, history):
             break
     context += history_context
     for _ in range(5):
-        response = generate(f"{context}\n\n{user}: {message}.\n", **generation_kwargs)
+        response = generate(f"{context}\n\n{user}: {message}.\n", context_length=context_length, **generation_kwargs)
         if DEBUG:
             print("\n-----" + response + "-----\n")
-        response = response.split("\n")[-1]
-        if agent in response and response.split(agent)[-1]:
-            response = response.split(agent)[-1]
-        if user in response and response.split(user)[-1]:
-            response = response.split(user)[-1]
+        # response = response.split("\n")[-1]
+        # if agent in response and response.split(agent)[-1]:
+        #     response = response.split(agent)[-1]
+        # if user in response and response.split(user)[-1]:
+        #     response = response.split(user)[-1]
+        # Take the first response
+        response = [
+            r for r in response.split(f"{AGENT}:") if r.strip()
+        ][0].split(USER)[0].replace(f"{AGENT}:", "\n").strip()
         if response and response[0] in string.punctuation:
             response = response[1:].strip()
         if response.strip().startswith(f"{user}: {message}"):