Add missing files

.gitignore

@@ -0,0 +1,3 @@
+*mic_env/*
+**__pycache__**
+*.pyc

requirements.txt

@@ -0,0 +1,8 @@
+plotly==5.1.0
+streamlit==0.84.1
+git+https://github.com/huggingface/transformers.git
+torchvision==0.10.0
+mtranslate==1.8
+black==21.7b0
+flax==0.3.4
+sentencepiece==0.1.96

session.py

@@ -0,0 +1,89 @@
+#
+# Code for managing session state, which is needed for multi-input forms
+# See https://github.com/streamlit/streamlit/issues/1557
+#
+# This code is taken from
+# https://gist.github.com/okld/0aba4869ba6fdc8d49132e6974e2e662
+#
+from streamlit.hashing import _CodeHasher
+from streamlit.report_thread import get_report_ctx
+from streamlit.server.server import Server
+
+
+class _SessionState:
+    def __init__(self, session, hash_funcs):
+        """Initialize SessionState instance."""
+        self.__dict__["_state"] = {
+            "data": {},
+            "hash": None,
+            "hasher": _CodeHasher(hash_funcs),
+            "is_rerun": False,
+            "session": session,
+        }
+
+    def __call__(self, **kwargs):
+        """Initialize state data once."""
+        for item, value in kwargs.items():
+            if item not in self._state["data"]:
+                self._state["data"][item] = value
+
+    def __getitem__(self, item):
+        """Return a saved state value, None if item is undefined."""
+        return self._state["data"].get(item, None)
+
+    def __getattr__(self, item):
+        """Return a saved state value, None if item is undefined."""
+        return self._state["data"].get(item, None)
+
+    def __setitem__(self, item, value):
+        """Set state value."""
+        self._state["data"][item] = value
+
+    def __setattr__(self, item, value):
+        """Set state value."""
+        self._state["data"][item] = value
+
+    def clear(self):
+        """Clear session state and request a rerun."""
+        self._state["data"].clear()
+        self._state["session"].request_rerun()
+
+    def sync(self):
+        """
+        Rerun the app with all state values up to date from the beginning to
+        fix rollbacks.
+        """
+        data_to_bytes = self._state["hasher"].to_bytes(self._state["data"], None)
+
+        # Ensure to rerun only once to avoid infinite loops
+        # caused by a constantly changing state value at each run.
+        #
+        # Example: state.value += 1
+        if self._state["is_rerun"]:
+            self._state["is_rerun"] = False
+
+        elif self._state["hash"] is not None:
+            if self._state["hash"] != data_to_bytes:
+                self._state["is_rerun"] = True
+                self._state["session"].request_rerun()
+
+        self._state["hash"] = data_to_bytes
+
+
+def _get_session():
+    session_id = get_report_ctx().session_id
+    session_info = Server.get_current()._get_session_info(session_id)
+
+    if session_info is None:
+        raise RuntimeError("Couldn't get your Streamlit Session object.")
+
+    return session_info.session
+
+
+def _get_state(hash_funcs=None):
+    session = _get_session()
+
+    if not hasattr(session, "_custom_session_state"):
+        session._custom_session_state = _SessionState(session, hash_funcs)
+
+    return session._custom_session_state

utils.py

@@ -3,7 +3,6 @@ import torch
 import numpy as np
 from torchvision.transforms import CenterCrop, ConvertImageDtype, Normalize, Resize
 from torchvision.transforms.functional import InterpolationMode
-from transformers import MBart50TokenizerFast
 from PIL import Image
 
 
@@ -28,24 +27,8 @@ class Transform(torch.nn.Module):
 
 transform = Transform(224)
 
-
 def get_transformed_image(image):
     if image.shape[-1] == 3 and isinstance(image, np.ndarray):
         image = image.transpose(2, 0, 1)
         image = torch.tensor(image)
-    return transform(image).unsqueeze(0).permute(0, 2, 3, 1).numpy()
-
-tokenizer = MBart50TokenizerFast.from_pretrained("facebook/mbart-large-50")
-
-language_mapping = {
-    "english": "en_XX",
-    "german": "de_DE",
-    "french": "fr_XX",
-    "spanish": "es_XX"
-}
-
-def generate_sequence(model, pixel_values, lang_code):
-    lang_code = language_mapping[lang_code]
-    output_ids = model.generate(input_ids=pixel_values, decoder_start_token_id=tokenizer.lang_code_to_id[lang_code], max_length=64, num_beams=4)
-    output_sequence = tokenizer.batch_decode(output_ids[0], skip_special_tokens=True, max_length=64)
-    return output_sequence
+    return transform(image).unsqueeze(0).permute(0, 2, 3, 1).numpy()