t

app.py

@@ -6,6 +6,12 @@ import gradio as gr
 import json
 
 # # from transformers import AutoModelForCausalLM, AutoTokenizer
+def get():
+    pass
+def get():
+    pass;
+
+
 
 # stage, commit, push
 
@@ -28,6 +34,8 @@ temperature = gr.inputs.Slider(
     minimum=0, maximum=1.5, default=0.8, label="temperature")
 top_p = gr.inputs.Slider(minimum=0, maximum=1.0,
                          default=0.9, label="top_p")
+top_k = gr.inputs.Slider(minimum=0, maximum=100,
+                         default=40, label="top_p")
 
 # gradio checkbutton
 
@@ -93,12 +101,19 @@ def get_generated_text(generated_text):
 
 
 
-def f(context, temperature, top_p, max_length, model_idx, SPACE_VERIFICATION_KEY):
+def f(context, temperature, top_p, top_k, max_length, model_idx, SPACE_VERIFICATION_KEY):
     try:
 
         if os.environ['SPACE_VERIFICATION_KEY'] != SPACE_VERIFICATION_KEY:
             return "invalid SPACE_VERIFICATION_KEY; see project secrets to view key"
 
+        try:
+            set_seed(randint(1, 256))
+        except Exception as e:
+            return "Exception while setting seed: " + str(e)
+
+        top_k = (top_k==0 and None) or top_k
+        # TODO write a function to generate the payload, it's becoming repetitive
         # maybe try "0" instead or 1, or "1"
         # use GPT-J-6B
         if model_idx == 0:
@@ -112,7 +127,7 @@ def f(context, temperature, top_p, max_length, model_idx, SPACE_VERIFICATION_KEY
                     # context becomes the previous generated context
                     # NOTE I've set return_full_text to false, see how this plays out
                     # change max_length from max_length>250 and 250 or max_length to 250
-                    payload = {"inputs": context, "parameters": {"return_full_text":False, "max_new_tokens": 250, "temperature": temperature, "top_p": top_p}}
+                    payload = {"inputs": context, "parameters": {"return_full_text":False, "max_new_tokens": 250, "temperature": temperature, "top_p": top_p, "top_k": top_k}}
                     response = requests.request("POST", API_URL, data=json.dumps(payload), headers=headers)
                     context = json.loads(response.content.decode("utf-8"))#[0]['generated_text']
                     # context = get_generated_text(generated_context)
@@ -158,10 +173,10 @@ def f(context, temperature, top_p, max_length, model_idx, SPACE_VERIFICATION_KEY
         elif model_idx == 1:
             # use GPT-2
             #
-            try:
-                set_seed(randint(1, 2**31))
-            except Exception as e:
-                return "Exception while setting seed: " + str(e)
+            # try:
+            #     set_seed(randint(1, 2**31))
+            # except Exception as e:
+            #     return "Exception while setting seed: " + str(e)
             # return sequences specifies how many to return
             
             # for some reson indexing with 'generated-text' doesn't work
@@ -178,7 +193,7 @@ def f(context, temperature, top_p, max_length, model_idx, SPACE_VERIFICATION_KEY
                 # however in the source that's what's used
                 # NOTE I think max_new_tokens is working now and punctuation characters count too
                 # NOTE set max_length to max_length to allow input text of any size
-                generated_text = generator(context, max_length=896, max_new_tokens=max_length, top_p=top_p, temperature=temperature, num_return_sequences=1)
+                generated_text = generator(context, max_length=896, max_new_tokens=max_length, top_p=top_p, top_k=top_k, temperature=temperature, num_return_sequences=1)
             except Exception as e:
                 return "Exception while generating text: " + str(e)
             # [0][0]['generated_text']
@@ -196,12 +211,13 @@ def f(context, temperature, top_p, max_length, model_idx, SPACE_VERIFICATION_KEY
             # TODO use fallback gpt-2 inference api for this as well 
             # TODO or just make it an option in the menu "GPT-2 inference"
         elif model_idx == 2: 
+            
             url = "https://api-inference.huggingface.co/models/distilgpt2"
             generated_text = ""#context #""
             # NOTE adding repetition penalty parameter
             # NOTE maybe leave tha parameter and just write a function to remove repetitions
             while len(generated_text) < max_length:
-                payload = {"inputs": context, "parameters": {"repetition_penalty":20.0,"return_full_text":False, "max_new_tokens": 250, "temperature": temperature, "top_p": top_p}}
+                payload = {"inputs": context, "parameters": {"repetition_penalty":20.0,"return_full_text":False, "max_new_tokens": 250, "temperature": temperature, "top_p": top_p, "top_k": top_k}}
                 response = requests.request("POST", url, data=json.dumps(payload), headers=headers)
                 context = json.loads(response.content.decode("utf-8"))
                 context = get_generated_text(context).strip()
@@ -219,7 +235,7 @@ def f(context, temperature, top_p, max_length, model_idx, SPACE_VERIFICATION_KEY
 
             generated_text = ""#context #""
             while len(generated_text) < max_length:
-                payload = {"inputs": context, "parameters": {"return_full_text":False, "max_new_tokens": 250, "temperature": temperature, "top_p": top_p}}
+                payload = {"inputs": context, "parameters": {"return_full_text":False, "max_new_tokens": 250, "temperature": temperature, "top_p": top_p, "top_k": top_k}}
                 response = requests.request("POST", url, data=json.dumps(payload), headers=headers)
                 context = json.loads(response.content.decode("utf-8"))
                 context = get_generated_text(context).strip()
@@ -231,7 +247,7 @@ def f(context, temperature, top_p, max_length, model_idx, SPACE_VERIFICATION_KEY
             generated_text = ""#context #""
             # NOTE we're actually using max_new_tokens and min_new_tokens
             while len(generated_text) < max_length:
-                payload = {"inputs": context, "parameters": {"return_full_text":False, "max_new_tokens": 250, "temperature": temperature, "top_p": top_p}}
+                payload = {"inputs": context, "parameters": {"return_full_text":False, "max_new_tokens": 250, "temperature": temperature, "top_p": top_p, "top_k": top_k}}
                 response = requests.request("POST", url, data=json.dumps(payload), headers=headers)
                 context = json.loads(response.content.decode("utf-8"))
                 context = get_generated_text(context).strip()
@@ -247,6 +263,7 @@ iface = gr.Interface(f, [
     "text",
     temperature,
     top_p,
+    top_k, 
     gr.inputs.Slider(
         minimum=20, maximum=512, default=30, label="max length"),
     gr.inputs.Dropdown(["GPT-J-6B", "GPT2", "DistilGPT2", "GPT-Large", "GPT-Neo-2.7B"], type="index", label="model", default="GPT2"),