Add no repeat n gram (#1)

- Add no repeat n gram (7e700b0cef3a572a312ce3980e8447a67cad1f2b)
- Update app.py (18ea58bc4b6f3f7fe183b5f240cfb97023d12ef9)
- add step=1 (ab3c9117453f46a47a1515b4a51ba347c6578101)
- Add num beams (4da5c39c3ff75666d61e0f5c6ca54e1f29dff943)
- fix small nit (2c122c8a5af180a698588e126f6559bfd531ba2a)
- fix small errors (c3dc4f20f1ef386d870f884ec8ba145a95afdb87)
- Update app.py (390933bb5e1f27425db2d20f5edcbcb886f1004a)


Co-authored-by: Younes Belkada <[email protected]>

app.py

@@ -14,9 +14,9 @@ hf_writer = gr.HuggingFaceDatasetSaver(HF_API_TOKEN, "huggingface/bloom_internal
 
 
 examples = [
-    ['A "whatpu" is a small, furry animal native to Tanzania. An example of a sentence that uses the word whatpu is: We were traveling in Africa and we saw these very cute whatpus. To do a "farduddle" means to jump up and down really fast. An example of a sentence that uses the word farduddle is:', 8, 0.1, 0, 0.9, False],
-    ['def quicksort(l):', 8, 0.1, 0, 0.9, False],
-    ['Q: ¿Cómo te llamas? A: What is your name? Q: ¿Qué edad tienes? A: How old are you? Q: ¿Dónde vives?', 8, 0.1, 0, 0.9, False]
+    ['A "whatpu" is a small, furry animal native to Tanzania. An example of a sentence that uses the word whatpu is: We were traveling in Africa and we saw these very cute whatpus. To do a "farduddle" means to jump up and down really fast. An example of a sentence that uses the word farduddle is:', 8, 0, 0, 0.1, 0, 0.9, False],
+    ['def quicksort(l):', 8, 0, 0, 0.1, 0, 0.9, False],
+    ['Q: ¿Cómo te llamas? A: What is your name? Q: ¿Qué edad tienes? A: How old are you? Q: ¿Dónde vives?', 8, 0, 0, 0.1, 0, 0.9, False]
 ]
 
 def safe_text(text):
@@ -35,11 +35,18 @@ def query(payload):
     print(response)
     return json.loads(response.content.decode("utf-8"))
     
-def inference(input_sentence, max_length, temperature,top_k, top_p, greedy_decoding, seed=42):
+def inference(input_sentence, max_length, no_repeat_ngram_size, num_beams, temperature,top_k, top_p, greedy_decoding, seed=42):
     top_k = None if top_k == 0 else top_k
+    do_sample = False if num_beams > 0 else not greedy_decoding
+    num_beams = None if (greedy_decoding or num_beams == 0) else num_beams
+    no_repeat_ngram_size = None if num_beams is None else no_repeat_ngram_size
+    top_p = None if num_beams else top_p
+
+    early_stopping = None if num_beams is None else num_beams > 0
+
     payload = {"inputs": input_sentence,
                "parameters": {"max_new_tokens": max_length, "top_k": top_k, "top_p": top_p, "temperature": temperature,
-                              "do_sample": not greedy_decoding, "seed": seed}}
+                              "do_sample": do_sample, "seed": seed, "early_stopping":early_stopping, "no_repeat_ngram_size":no_repeat_ngram_size, "num_beams":num_beams}}
     data = query(
         payload
     )
@@ -51,7 +58,9 @@ gr.Interface(
     inference, 
     [
         gr.inputs.Textbox(label="Input"),
-        gr.inputs.Slider(1, 64, default=8, label="Tokens to generate"),
+        gr.inputs.Slider(1, 64, default=8, step=1, label="Tokens to generate"),
+        gr.inputs.Slider(1, 10, default=2, step=1, label="No repeat N gram"),
+        gr.inputs.Slider(0, 10, default=5, step=1, label="Num beams"),
         gr.inputs.Slider(0.0, 1.0, default=0.1, step=0.05, label="Temperature"),
         gr.inputs.Slider(0, 64, default=0, step=1, label="Top K"),
         gr.inputs.Slider(0.0, 10, default=0.9, step=0.05, label="Top P"),