Merge pull request #2 from sdsc-ordes/fix-classifier-with-image-preprocessing

README.md

@@ -5,6 +5,7 @@ colorFrom: yellow
 colorTo: red
 sdk: streamlit
 sdk_version: 1.39.0
+python_version: 3.10
 app_file: call_models/entry_and_hotdog.py
 pinned: false
 license: apache-2.0

call_models/entry_and_hotdog.py

@@ -25,6 +25,8 @@ from transformers import AutoModelForImageClassification
 
 # setup for the ML model on huggingface (our wrapper)
 os.environ["PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION"] = "python"
+#classifier_revision = '0f9c15e2db4d64e7f622ade518854b488d8d35e6'
+classifier_revision = 'main' # default/latest version
 # and the dataset of observations (hf dataset in our space)
 dataset_id = "Saving-Willy/Happywhale-kaggle"
 data_files = "data/train-00000-of-00001.parquet"
@@ -106,6 +108,7 @@ def push_observation(tab_log=None):
 if __name__ == "__main__":
 
     g_logger.info("App started.")
+    g_logger.warning(f"[D] Streamlit version: {st.__version__}. Python version: {os.sys.version}")
 
     #g_logger.debug("debug message")
     #g_logger.info("info message")
@@ -220,7 +223,8 @@ if __name__ == "__main__":
     if tab_inference.button("Identify with cetacean classifier"):
         #pipe = pipeline("image-classification", model="Saving-Willy/cetacean-classifier", trust_remote_code=True)
         cetacean_classifier = AutoModelForImageClassification.from_pretrained("Saving-Willy/cetacean-classifier", 
-                                                                            revision='0f9c15e2db4d64e7f622ade518854b488d8d35e6', trust_remote_code=True)
+                                                                            revision=classifier_revision,
+                                                                            trust_remote_code=True)
         
         if st.session_state.image is None:
             # TODO: cleaner design to disable the button until data input done?

call_models/input_handling.py

@@ -6,6 +6,8 @@ import hashlib
 import logging
 
 import streamlit as st
+import cv2
+import numpy as np
 
 m_logger = logging.getLogger(__name__)
 # we can set the log level locally for funcs in this module
@@ -135,7 +137,12 @@ def setup_input(viewcontainer: st.delta_generator.DeltaGenerator=None, _allowed_
 
     if uploaded_filename is not None:
         # Display the uploaded image
-        image = Image.open(uploaded_filename)
+        #image = Image.open(uploaded_filename)
+        # load image using cv2 format, so it is compatible with the ML models
+        file_bytes = np.asarray(bytearray(uploaded_filename.read()), dtype=np.uint8)
+        image = cv2.imdecode(file_bytes, 1)
+
+
         viewcontainer.image(image, caption='Uploaded Image.', use_column_width=True)
         # store the image in the session state
         st.session_state.image = image

call_models/requirements.txt

@@ -1,17 +1,30 @@
-transformers
-streamlit
-huggingface_hub
-torch
+numpy==1.24
+pandas==2.2.3
 
-pandas
-numpy
 
-datasets
-
-# for nice map tiles
+# frontend
+streamlit==1.39.0
+## for nice map tiles
 folium
 streamlit_folium
+folium==0.18.0
+streamlit_folium==0.23.1
+
+# backend 
+datasets==3.0.2
+
+
+# running ML models
+
+## to use ML models hosted on HF 
+huggingface-hub==0.26.1
+transformers==4.46.0
+#torch
+## +minimal reqs for locally running model - with PIL load
+pytorch-lightning<=1.5.10
+timm==0.5.4
+pillow==10.4.0
 
-# for ceatatean
-pytorch_lightning 
-timm
+## also need opencv for handling images in the later model with preprocessing
+opencv-python-headless==4.5.5.64
+albumentations==1.1.0