cached demo

.gitattributes

@@ -29,3 +29,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zstandard filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+*.faiss filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,12 +1,13 @@
 ---
 title: Fashion Classification
-emoji: 📊
-colorFrom: pink
-colorTo: green
+emoji: 👁
+colorFrom: gray
+colorTo: pink
 sdk: streamlit
 sdk_version: 1.10.0
-app_file: app.py
+app_file: data_analysis_app.py
 pinned: false
+license: apache-2.0
 ---
 
 Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

data_analysis_app.py

@@ -0,0 +1,73 @@
+from datasets import load_dataset
+import streamlit as st
+
+from data_utils import get_embedding
+
+from bokeh.plotting import figure,show
+from bokeh.io import push_notebook, output_notebook
+# output_notebook()
+from bokeh.palettes import d3
+
+from bokeh.models import ColumnDataSource, Grid, LinearAxis, Plot, Scatter
+from bokeh.transform import factor_cmap, factor_mark
+import base64
+from io import BytesIO
+
+label_columns=["gender","subCategory","masterCategory"]
+
+model_interest=['facebook/deit-tiny-patch16-224', # very small model 5M param model
+                'microsoft/beit-base-patch16-224', # big model
+                "facebook/dino-vits8",
+                "facebook/levit-128S"]
+
+def convert_base64(img):
+    buffered = BytesIO()
+    img.save(buffered, format="JPEG")
+    img_str = base64.b64encode(buffered.getvalue()).decode("utf-8")
+    return "data:image/jpeg;base64,"+img_str
+                    
+@st.experimental_singleton
+def cache_embedding(model_name):
+    dataset=load_dataset("ceyda/fashion-products-small", split="train")
+    dataset=dataset.shuffle(seed=100) #pick a random seed
+    viz_dat=dataset.train_test_split(0.1,shuffle=False) #일부를 visualization위해서 뽑시단
+    viz_dat=viz_dat["test"]
+    embedding = get_embedding(model_name,viz_dat)
+    embedding["image"]=embedding["image"].apply(convert_base64)
+    labels = {label:viz_dat.unique(label) for label in label_columns}
+    return embedding,labels
+    
+@st.experimental_singleton
+def cache_graph(model_name,color_column):
+    embedding,labels=cache_embedding(model_name)
+    
+    color_palette = (d3['Category20'][20]+d3['Category20b'][20]+d3['Category20c'][20])[:len(labels[color_column])]
+    source = ColumnDataSource(data=embedding)
+    # colors = factor_cmap('gender', palette=["purple","navy","green","blue","pink"], factors=embedding["gender"].unique()) 
+    
+    
+    TOOLS="hover,crosshair,pan,wheel_zoom,zoom_in,zoom_out,box_zoom,reset,tap,save,box_select,lasso_select,"
+    TOOLTIPS = """
+        <div>
+            <div>
+                <img
+                    src="@image" height="42" alt="@image" width="42"
+                    style="float: left; margin: 0px 15px 15px 0px;"
+                    border="2"
+                ></img>
+            </div>
+    """
+    p = figure(tools=TOOLS,tooltips=TOOLTIPS)
+    
+    p.scatter(x="x", y="y", source=source,
+              # marker=factor_mark('gender', ['circle', 'circle_cross', 'circle_dot','circle_x','circle_y'], labels["gender"]),
+              color=factor_cmap(color_column, color_palette, labels[color_column])
+             )
+ 
+    return p
+
+
+model_name=st.sidebar.selectbox("Model",model_interest)
+color_column=st.selectbox("Color by",label_columns)
+p=cache_graph(model_name,color_column)
+st.bokeh_chart(p, use_container_width=False)

data_utils.py

@@ -0,0 +1,48 @@
+from datasets import load_dataset
+from PIL import Image
+import os
+import pandas as pd
+from transformers import AutoFeatureExtractor,AutoModel
+from faiss.contrib.inspect_tools import get_flat_data 
+import pymde
+import numpy as np
+
+def get_embedding(model_name,viz_dat):
+
+    index_file=f"./indexes/{model_name.split('/')[1]}.faiss"
+    
+    if os.path.exists(index_file):
+        viz_dat.load_faiss_index('embeddings', index_file)
+    else:
+        feature_extractor = AutoFeatureExtractor.from_pretrained(model_name)
+        model = AutoModel.from_pretrained(model_name)
+        # model.to("cuda:0")
+        def embed(x):
+            images=x["image"]
+            inputs = feature_extractor(images=images, return_tensors="pt")
+            # inputs.to("cuda:0")
+            outputs = model(**inputs,output_hidden_states= True)
+            final_emb=outputs.pooler_output.detach().cpu().numpy() # this line depends on the model you are using
+            x["embeddings"]=final_emb
+            return x
+        # Add embeddings to dataset
+        viz_dat = viz_dat.map(embed,batched=True,batch_size=20)
+        viz_dat.add_faiss_index(column='embeddings')
+        viz_dat.save_faiss_index('embeddings',index_file)
+        
+    embedding_file=f"./indexes/{model_name.split('/')[1]}.npy"   
+    if os.path.exists(embedding_file):
+        embedding = np.load(embedding_file) # load
+    else:    
+        index=viz_dat.get_index("embeddings").faiss_index
+        embeddings=get_flat_data(index)
+        embedding=pymde.preserve_neighbors(embeddings, verbose=True).embed().numpy()
+        np.save(embedding_file, embedding) # save
+    
+    embedding=pd.DataFrame(embedding,columns=["x","y"])
+    embedding["image"]=viz_dat["image"]
+    embedding["gender"]=viz_dat["gender"]
+    embedding["masterCategory"]=viz_dat["masterCategory"]
+    embedding["subCategory"]=viz_dat["subCategory"]   
+    
+    return embedding

indexes/beit-base-patch16-224.faiss

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eef8b38b055d02f2dbfd6ac5c2cbba0cb670ef225285ba375e4baf63034ef589
+size 13117485

indexes/beit-base-patch16-224.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b42f5b1e9a149b412a120a63999ea7a5d89438aef5651e200acbdcb226e75418
+size 34288

indexes/deit-tiny-patch16-224.faiss

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5df0a6edbaf0b746bffc2ff740d783a4fa6be961dedece7ec069a37e1688341a
+size 3279405

indexes/deit-tiny-patch16-224.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a5bb0e5186ea3125d548838806af53cffc3fbb53ed752e06a7af9b2487c997df
+size 34288

indexes/dino-vits8.faiss

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:81965f420ca2c66ad55ad0e3300bbce2d5fd366b2c9e5e3ae48ecc88e4a00d97
+size 6558765

indexes/dino-vits8.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1499f7501e455d088c5621512fd94a819fd26940271782cbf047db9e49359df0
+size 34288

indexes/levit-128S.faiss

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d1d9fec9e282812e613e312c036f8998b7177f6d19f787b61a5cd9810916ba8e
+size 6558765

indexes/levit-128S.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8757326676da484d4aa542eacc54f74fb8c8c743697fde4d3dd0dd44e2eeb1f9
+size 34288

requirements.txt

@@ -0,0 +1,6 @@
+# didn't pin versions
+datasets
+transformers
+pymde
+bokeh==2.4.1
+faiss-cpu