adds model selection from wandb

CnLgGm025_SpeUni1024_DI_EATL.nemo

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:94d7a7443c60dddb1e7ddfdcf5f2e85a2a857b2bc31bf20b4aedf8979b866183
-size 566712320

app.py

@@ -1,16 +1,51 @@
+import wandb
 import gradio as gr
+from datetime import datetime, timedelta
 import nemo.collections.asr as nemo_asr
 
-model = nemo_asr.models.ASRModel.restore_from("CnLgGm025_SpeUni1024_DI_EATL.nemo")
-model.eval()
+wandb_api = wandb.Api()
+
+all_artifacts_versions = [
+    version
+    for version in [
+        collection.versions()
+        for collection in wandb_api.artifact_type(
+            type_name="model", project="nemo-experiments"
+        ).collections()
+    ]
+]
+
+latest_artifacts = [
+    artifact
+    for artifact_versions in all_artifacts_versions
+    for artifact in artifact_versions
+    if (
+        datetime.fromisoformat(artifact.created_at)
+        > datetime.now() - timedelta(days=180)  # last 180 days
+        and artifact.state != "DELETED"
+    )
+]
+
+models = {artifact.name: None for artifact in latest_artifacts}
+
+
+def lazy_load_models(models_names):
+    for model_name in models_names:
+        model = models[model_name]
+        if not model:
+            models[model_name] = nemo_asr.models.ASRModel.restore_from(
+                list(filter(lambda x: x.name == model_name, latest_artifacts))[0].file()
+            )
+            models[model_name].eval()
 
 
 def transcribe(audio_mic, audio_file, models_names):
+    lazy_load_models(models_names)
     # transcribe audio_mic and audio_file separately
     # because transcribe() fails is path is empty
     transcription_mic = "\n".join(
         [
-            f"{model_name} => {model.transcribe([audio_mic])[0]}"
+            f"{model_name} => {models[model_name].transcribe([audio_mic])[0]}"
             for model_name in models_names
         ]
         if audio_mic
@@ -18,7 +53,7 @@ def transcribe(audio_mic, audio_file, models_names):
     )
     transcription_file = "\n".join(
         [
-            f"{model_name} => {model.transcribe([audio_file])[0]}"
+            f"{model_name} => {models[model_name].transcribe([audio_file])[0]}"
             for model_name in models_names
         ]
         if audio_file
@@ -27,12 +62,24 @@ def transcribe(audio_mic, audio_file, models_names):
     return transcription_mic, transcription_file
 
 
-selection = ["a", "b", "c"]
+model_selection = list(models.keys())
 
 demo = gr.Blocks()
 
 with demo:
-    models_names = gr.CheckboxGroup(selection)
+    gr.Markdown(
+        """
+        # ﷽
+        These are the latest* Tarteel models.
+
+        Please note that the models are lazy loaded.
+        This means that the first time you use a model,
+        it might take some time to be downloaded and loaded for inference.
+
+        *: last 180 days since the space was launched.
+        To update the list, restart the space.
+        """
+    )
     with gr.Row():
         audio_mic = gr.Audio(source="microphone", type="filepath", label="Microphone")
         audio_file = gr.Audio(source="upload", type="filepath", label="File")
@@ -41,6 +88,8 @@ with demo:
         output_mic = gr.TextArea(label="Microphone Transcription")
         output_file = gr.TextArea(label="Audio Transcription")
 
+    models_names = gr.CheckboxGroup(model_selection, label="Select Models to Use")
+
     b1 = gr.Button("Transcribe")
 
     b1.click(