Add some models

app.py

@@ -12,23 +12,59 @@ TRANSLATION_NL_TO_EN = "translation_en_to_nl"
 
 GENERATOR_LIST = [
     {
-        "model_name": "yhavinga/longt5-local-eff-large-nl8-voc8k-ddwn-512beta-512l-nedd-256ccmatrix-en-nl",
-        "desc": "longT5 large nl8 256cc/512beta/512l en->nl",
+        "model_name": "Helsinki-NLP/opus-mt-en-nl",
+        "desc": "Opus MT en->nl",
         "task": TRANSLATION_NL_TO_EN,
-        "split_sentences": False,
+        "split_sentences": True,
+    },
+    {
+        "model_name": "yhavinga/t5-small-24L-ccmatrix-multi",
+        "desc": "T5 small nl24 ccmatrix en->nl",
+        "task": TRANSLATION_NL_TO_EN,
+        "split_sentences": True,
     },
     {
-        "model_name": "yhavinga/longt5-local-eff-large-nl8-voc8k-ddwn-512beta-512-nedd-en-nl",
-        "desc": "longT5 large nl8 512beta/512l en->nl",
+        "model_name": "yhavinga/longt5-local-eff-large-nl8-voc8k-ddwn-512beta-512l-nedd-256ccmatrix-en-nl",
+        "desc": "longT5 large nl8 256cc/512beta/512l en->nl",
         "task": TRANSLATION_NL_TO_EN,
         "split_sentences": False,
     },
     {
-        "model_name": "yhavinga/t5-small-24L-ccmatrix-multi",
-        "desc": "T5 small nl24 ccmatrix en->nl",
+        "model_name": "yhavinga/byt5-small-ccmatrix-en-nl",
+        "desc": "ByT5 small ccmatrix en->nl",
         "task": TRANSLATION_NL_TO_EN,
         "split_sentences": True,
     },
+    # {
+    #     "model_name": "yhavinga/t5-eff-large-8l-nedd-en-nl",
+    #     "desc": "T5 eff large nl8 en->nl",
+    #     "task": TRANSLATION_NL_TO_EN,
+    #     "split_sentences": True,
+    # },
+    # {
+    #     "model_name": "yhavinga/t5-base-36L-ccmatrix-multi",
+    #     "desc": "T5 base nl36 ccmatrix en->nl",
+    #     "task": TRANSLATION_NL_TO_EN,
+    #     "split_sentences": True,
+    # },
+    # {
+    #     "model_name": "yhavinga/longt5-local-eff-large-nl8-voc8k-ddwn-512beta-512-nedd-en-nl",
+    #     "desc": "longT5 large nl8 512beta/512l en->nl",
+    #     "task": TRANSLATION_NL_TO_EN,
+    #     "split_sentences": False,
+    # },
+    # {
+    #     "model_name": "yhavinga/t5-base-36L-nedd-x-en-nl-300",
+    #     "desc": "T5 base 36L nedd en->nl 300",
+    #     "task": TRANSLATION_NL_TO_EN,
+    #     "split_sentences": True,
+    # },
+    # {
+    #     "model_name": "yhavinga/long-t5-local-small-ccmatrix-en-nl",
+    #     "desc": "longT5 small ccmatrix en->nl",
+    #     "task": TRANSLATION_NL_TO_EN,
+    #     "split_sentences": True,
+    # },
 ]
 
 
@@ -64,7 +100,7 @@ It was a quite young girl, unknown to me, with a hood over her head, and with la
 
 “My father is very ill,” she said without a word of introduction. “The nurse is frightened. Could you come in and help?”"""
     st.session_state["text"] = st.text_area(
-        "Enter text", st.session_state.prompt_box, height=300
+        "Enter text", st.session_state.prompt_box, height=250
     )
     num_beams = st.sidebar.number_input("Num beams", min_value=1, max_value=10, value=1)
     num_beam_groups = st.sidebar.number_input(
@@ -83,6 +119,7 @@ and the [Huggingface text generation interface doc](https://huggingface.co/trans
         "num_beams": num_beams,
         "num_beam_groups": num_beam_groups,
         "length_penalty": length_penalty,
+        "early_stopping": True,
     }
 
     if st.button("Run"):

generator.py

@@ -30,9 +30,19 @@ def load_model(model_name):
     if tokenizer.pad_token is None:
         print("Adding pad_token to the tokenizer")
         tokenizer.pad_token = tokenizer.eos_token
-    model = AutoModelForSeq2SeqLM.from_pretrained(
-        model_name, from_flax=True, use_auth_token=get_access_token()
-    )
+    try:
+        model = AutoModelForSeq2SeqLM.from_pretrained(
+            model_name, use_auth_token=get_access_token()
+        )
+    except EnvironmentError:
+        try:
+            model = AutoModelForSeq2SeqLM.from_pretrained(
+                model_name, from_flax=True, use_auth_token=get_access_token()
+            )
+        except EnvironmentError:
+            model = AutoModelForSeq2SeqLM.from_pretrained(
+                model_name, from_tf=True, use_auth_token=get_access_token()
+            )
     if device != -1:
         model.to(f"cuda:{device}")
     return tokenizer, model
@@ -66,24 +76,13 @@ class Generator:
             for key in self.gen_kwargs:
                 if key in self.model.config.__dict__:
                     self.gen_kwargs[key] = self.model.config.__dict__[key]
-                    print(
-                        "Setting",
-                        key,
-                        "to",
-                        self.gen_kwargs[key],
-                        "for model",
-                        self.model_name,
-                    )
             try:
                 if self.task in self.model.config.task_specific_params:
                     task_specific_params = self.model.config.task_specific_params[
                         self.task
                     ]
-                    self.prefix = (
-                        task_specific_params["prefix"]
-                        if "prefix" in task_specific_params
-                        else ""
-                    )
+                    if "prefix" in task_specific_params:
+                        self.prefix = task_specific_params["prefix"]
                     for key in self.gen_kwargs:
                         if key in task_specific_params:
                             self.gen_kwargs[key] = task_specific_params[key]
@@ -95,7 +94,7 @@ class Generator:
         text = re.sub(r"\n{2,}", "\n", text)
         generate_kwargs = {**self.gen_kwargs, **generate_kwargs}
 
-        # if there are newlines in the text, and the model needs line-splitting, split the text
+        # if there are newlines in the text, and the model needs line-splitting, split the text and recurse
         if re.search(r"\n", text) and self.split_sentences:
             lines = text.splitlines()
             translated = [self.generate(line, **generate_kwargs)[0] for line in lines]

requirements.txt

@@ -5,9 +5,10 @@ protobuf<3.20
 streamlit>=1.4.0,<=1.10.0
 torch
 transformers>=4.13.0
-mtranslate
+langdetect
 psutil
 jax[cuda]==0.3.16
 chex>=0.1.4
 ##jaxlib==0.1.67
 flax>=0.5.3
+sentencepiece