import streamlit as st from src.data import StoreDataLoader from src.model import Model_Load import matplotlib.pyplot as plt import seaborn as sns import plotly.graph_objects as go from sklearn.metrics import mean_absolute_error,mean_squared_error import numpy as np import pandas as pd from src.prediction import test_prediction,val_prediction,create_week_date_featues import plotly.express as px #----------------hide menubar and footer---------------------------------- hide_streamlit_style = """ """ st.markdown(hide_streamlit_style, unsafe_allow_html=True) #------------------------------------------------------------- ## Load model object model_obj=Model_Load() #-------------------------------------------------------------- @st.cache_data def convert_df(df): return df.to_csv(index=False).encode('utf-8') #----------------------------------------------------------------- ## Title of Page st.markdown("""

Product Demand Forecasting Dashboard by AI Planet

""", unsafe_allow_html=True) #--------------------------------------------------------------------------------------------------------------------- # select the model(Sidebar) with st.sidebar: st.markdown("""
""", unsafe_allow_html=True) option=st.selectbox("Select Model",['TFT','Prophet']) #------------------------------------------------------------------------------------------------------------ # TFT if option=='TFT': #-------------------------------------------------------------------------------------------------------- ## TFT data path and load path='data/train.csv' obj=StoreDataLoader(path) train_dataset,test_dataset,training,validation,earliest_time=obj.tft_data() print(f"TRAINING ::START DATE ::{train_dataset['date'].min()} :: END DATE ::{train_dataset['date'].max()}") print(f"TESTING ::START DATE ::{test_dataset['date'].min()} :: END DATE ::{test_dataset['date'].max()}") list_store=train_dataset['store'].unique() list_items=train_dataset['item'].unique() #--------------------------------------------------------------------------------------------------------- try: # load the pre trained tft model model=model_obj.store_model_load(option) with st.sidebar: # st.success('Model Loaded successfully', icon="✅") # select the store id store=st.selectbox("Select Store ID",list_store) # select the item id item=st.selectbox("Select Product ID",list_items) #-------------------------------------------------------------------------------------------------------------- ## prediction on testing data testing_results=test_prediction(model,train_dataset=train_dataset,test_dataset=test_dataset ,earliest_time=earliest_time,store_id=store,item_id=item) # find kpi rmse=np.around(np.sqrt(mean_squared_error(testing_results['Lead_1'],testing_results['prediction'])),2) mae=np.around(mean_absolute_error(testing_results['Lead_1'],testing_results['prediction']),2) print(f"TEST DATA = Item ID : {item} :: MAE : {mae} :: RMSE : {rmse}") #--------------------------------------tft future prediction------------------------------------------- final_data=pd.concat([train_dataset,test_dataset]) consumer_data=final_data.loc[(final_data['store']==store) & (final_data['item']==item)] consumer_data.fillna(0,inplace=True) date_list=[] demand_prediction=[] for i in range(30): # select last 150 records as an enocer + decoder data encoder_data = consumer_data[lambda x: x.days_from_start > x.days_from_start.max() - 150] last_data = consumer_data[lambda x: x.days_from_start == x.days_from_start.max()] # prediction date and time date_list.append(encoder_data.tail(1).iloc[-1,:]['date']) # prediction for the last 30 records test_prediction = model.predict(encoder_data, mode="prediction", trainer_kwargs=dict(accelerator="cpu"), return_x=True) # create the next day record decoder_data = pd.concat( [last_data.assign(date=lambda x: x.date + pd.offsets.DateOffset(i)) for i in range(1, 2)], ignore_index=True, ) # find the hours_from_start & days_from_start decoder_data["hours_from_start"] = (decoder_data["date"] - earliest_time).dt.seconds / 60 / 60 + (decoder_data["date"] - earliest_time).dt.days * 24 decoder_data['hours_from_start'] = decoder_data['hours_from_start'].astype('int') decoder_data["hours_from_start"] += encoder_data["hours_from_start"].max() + 1 - decoder_data["hours_from_start"].min() # add time index consistent with "data" decoder_data["days_from_start"] = (decoder_data["date"] - earliest_time).apply(lambda x:x.days) # adding the datetime features decoder_data=create_week_date_featues(decoder_data,'date') # last timestep predicted record as assume next day actual demand(for more day forecasting) decoder_data['sales']=float(test_prediction.output[0][-1]) # append this prediction into the list demand_prediction.append(float(test_prediction.output[0][-1])) # update prediction time idx decoder_data['time_idx']=int(test_prediction.x['decoder_time_idx'][0][-1]) # add the next day record into the original data consumer_data=pd.concat([consumer_data,decoder_data]) # fina lag and update consumer_data['lag_1']=consumer_data['sales'].shift(1) consumer_data['lag_5']=consumer_data['sales'].shift(5) # reset the index consumer_data=consumer_data.reset_index(drop=True) # forecast values for the next 30 days/timesteps d2=pd.DataFrame({"date":date_list,"prediction":demand_prediction})[['date','prediction']] # update the store and item ids d2['store']=store d2['item']=item #----------------------------TFT and Prophet model KPI---------------------------------------- with st.sidebar: st.markdown(f"""
Models Evalution
""",unsafe_allow_html=True) st.dataframe(pd.DataFrame({"KPI":['RMSE','MAE'],"TFT":[7.73,6.17],"Prophet":[7.32,6.01]}).set_index('KPI'),width=300) # d2=pd.DataFrame({"KPI":['RMSE','MAE','RMSE','MAE'],"model":['TFT','TFT','Prophet','Prophet'],"Score":[7.73,6.17,7.32,6.01]}) # fig = px.bar(d2, x="KPI", y="Score", # color='model', barmode='group', # height=200,width=300,text_auto=True,) # st.plotly_chart(fig) #------------------------------------Prophet model KPI--------------------------------------------------------- st.markdown(f"""
KPI :: {item}
""",unsafe_allow_html=True) st.dataframe(pd.DataFrame({"KPI":['RMSE','MAE'],"TFT":[rmse,mae]}).set_index('KPI'),width=300) #-------------------------------------------------------------------------------------------------------------- # tabs tab1,tab2=st.tabs(['📈Forecast Plot','🗃Forecast Table']) #tab3-'🗃Actual Table' #------------------------------------------------Tab-1----------------------------------------------------------- tab1.markdown("""

📈 Forecast Plot

""", unsafe_allow_html=True) # change dtype on prediction column testing_results['prediction']=testing_results['prediction'].apply(lambda x:round(x)) testing_results['date']=testing_results['date'].dt.date d2['prediction']=d2['prediction'].apply(lambda x:round(x)) d2['date']=d2['date'].dt.date # training_data=train_dataset.loc[(train_dataset['store']==store)&(train_dataset['item']==item)][['date','Lead_1']].iloc[-60:,:] #---------------------------------------------forecast plot--------------------------------------------- fig = go.Figure([ # go.Scatter(x=training_data['date'],y=training_data['Lead_1'],name='Train Observed',line=dict(color='rgba(50, 205, 50, 0.7)')), #go.Scatter(x=y_train_pred['ds'],y=y_train_pred['yhat'],name='Prophet Pred.(10 Item)',line=dict(color='blue', dash='dot')), go.Scatter(x=testing_results['date'], y=testing_results['Lead_1'],name='Observed',line=dict(color='rgba(218, 112, 214, 0.5)')), go.Scatter(x=testing_results['date'],y=testing_results['prediction'],name='Historical Forecast',line=dict(color='#9400D3', dash='dash')), go.Scatter(x=d2['date'],y=d2['prediction'],name='Future Forecast',line=dict(color='Dark Orange', dash='dot'))]) fig.update_layout( xaxis_title='Date', yaxis_title='Order Demand', margin=dict(l=0, r=0, t=50, b=0), xaxis=dict(title_font=dict(size=20)), yaxis=dict(title_font=dict(size=20))) fig.update_layout(width=700,height=400) tab1.plotly_chart(fig) #----------------------------------------------Tab-2------------------------------------------------------------ tab2.markdown("""

📃 Forecast Table

""", unsafe_allow_html=True) final_r=pd.concat([d2[['date','store','item','prediction']],testing_results[['date','store','item','prediction']]]).sort_values('date').drop_duplicates().reset_index(drop=True) csv = convert_df(final_r) tab2.dataframe(final_r,width=500) tab2.download_button( "Download", csv, "file.csv", "text/csv", key='download-csv' ) #--------------------------------Tab-3---------------------------------------------- # tab3.markdown(""" #
#

# 📈 # # Actual Dataset # # # # #

#
# """, unsafe_allow_html=True) # train_a=train_dataset.loc[(train_dataset['store']==store) & (train_dataset['item']==item)][['date','store','item','sales']] # test_a=test_dataset.loc[(test_dataset['store']==store) & (test_dataset['item']==item)][['date','store','item','sales']] # actual_final_data=pd.concat([train_a,test_a]) # actual_final_data['date']=actual_final_data['date'].dt.date # tab3.dataframe(actual_final_data,width=500) except: st.sidebar.error('Model Not Loaded successfully!',icon="🚨") #+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ #+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ elif option=='Prophet': print("prophet") #---------------------------------------------------Data---------------------------------------------------- # Prophet data path='data/train.csv' obj=StoreDataLoader(path) fb_train_data,fb_test_data,item_dummay,store_dummay=obj.fb_data() # st.write(fb_train_data.columns) # st.write(fb_test_data.columns) # print(fb_test_data.columns) print(f"TRAINING ::START DATE ::{fb_train_data['ds'].min()} :: END DATE ::{fb_train_data['ds'].max()}") print(f"TESTING ::START DATE ::{fb_test_data['ds'].min()} :: END DATE ::{fb_test_data['ds'].max()}") train_new=fb_train_data.drop('y',axis=1) test_new=fb_test_data.drop('y',axis=1) #----------------------------------------------model Load---------------------------------------------------- try: fb_model=model_obj.store_model_load(option) # with st.sidebar: # st.success('Model Loaded successfully', icon="✅") #-------------------------------------select store & item --------------------------------------------------- list_items=item_dummay.columns list_store=store_dummay.columns with st.sidebar: store=st.selectbox("Select Store",list_store) item=st.selectbox("Select Product",list_items) #------------------------------------------prediction--------------------------------------------------------------- test_prediction=fb_model.predict(test_new.loc[test_new[item]==1]) train_prediction=fb_model.predict(train_new.loc[train_new[item]==1]) y_true_test=fb_test_data.loc[fb_test_data[item]==1] y_true_train=fb_train_data.loc[fb_train_data[item]==1] y_train_pred=train_prediction[['ds','yhat']].iloc[-60:,:] y_train_true=y_true_train[['ds','y']].iloc[-60:,:] y_test_pred=test_prediction[['ds','yhat']] y_test_true=y_true_test[['ds','y']] #----------------------------------------KPI--------------------------------------------------------------- rmse=np.sqrt(mean_squared_error(y_test_true['y'],y_test_pred['yhat'])) mae=mean_absolute_error(y_test_true['y'],y_test_pred['yhat']) #---------------------------------future prediction--------------------------------------- fb_final=pd.concat([fb_train_data,fb_test_data]) # extract the data for selected store and item fb_consumer=fb_final.loc[(fb_final[store]==1) & (fb_final[item]==1)] # list of dates and prediction date_list=[] prediction_list=[] # predicting the next 30 days product demand for i in range(30): # select only date record next_prediction=fb_consumer.tail(1).drop('y',axis=1) # drop target of last 01/01/2015 00:00:00 # predict next timestep demand prediction=fb_model.predict(next_prediction) # pass other feature value to the model # append date and predicted demand date_list.append(prediction['ds'][0]) ## append the datetime of prediction prediction_list.append(prediction['yhat'][0]) ## append the next timestep prediction #--------------------------next timestep data simulate------------------------------------------------------------- last_data = fb_consumer[lambda x: x.ds == x.ds.max()] # last date present in data # next timestep decoder_data = pd.concat( [last_data.assign(ds=lambda x: x.ds + pd.offsets.DateOffset(i)) for i in range(1, 2)], ignore_index=True, ) # update next timestep datetime covariates decoder_data=create_week_date_featues(decoder_data,'ds') # update last day demand prediction to the here as an actual demand value(using for more future timestep prediction) decoder_data['sales']=prediction['yhat'][0] # assume next timestep prediction as actual # update this next record into the original data fb_consumer=pd.concat([fb_consumer,decoder_data]) # append that next timestep data to into main data # find shift of power usage and update into the datset fb_consumer['lag_1']=fb_consumer['sales'].shift(1) fb_consumer['lag_5']=fb_consumer['sales'].shift(5) fb_consumer=fb_consumer.reset_index(drop=True) # reset_index future_prediction=pd.DataFrame({"ds":date_list,"yhat":prediction_list}) future_prediction['store']=store future_prediction['item']=item with st.sidebar: st.markdown(f"""
Models Evalution
""",unsafe_allow_html=True) st.dataframe(pd.DataFrame({"KPI":['RMSE','MAE'],"TFT":[7.73,6.17],"Prophet":[7.32,6.01]}).set_index('KPI'),width=300) st.markdown(f"""
KPI :: {item}
""",unsafe_allow_html=True) st.dataframe(pd.DataFrame({"KPI":['RMSE','MAE'],"Prophet":[rmse,mae]}).set_index('KPI'),width=300) #---------------------------------------Tabs----------------------------------------------------------------------- tab1,tab2=st.tabs(['📈Forecast Plot','🗃Forecast Table']) #tab3- '🗃Actual Table' #-------------------------------------------Tab-1=Forecast plot--------------------------------------------------- tab1.markdown("""

📈 Forecast Plot

""", unsafe_allow_html=True) ## round fig. y_train_true['y']=y_train_true['y'].astype('int') y_train_pred['yhat']=y_train_pred['yhat'].astype('int') y_test_true['y']=y_test_true['y'].astype('int') y_test_pred['yhat']=y_test_pred['yhat'].astype('int') future_prediction['yhat']=future_prediction['yhat'].astype('int') y_train_true['ds']=y_train_true['ds'].dt.date y_train_pred['ds']=y_train_pred['ds'].dt.date y_test_true['ds']=y_test_true['ds'].dt.date y_test_pred['ds']=y_test_pred['ds'].dt.date future_prediction['ds']=future_prediction['ds'].dt.date #-----------------------------plot--------------------------------------------------------------------------------------------- fig = go.Figure([ # go.Scatter(x=y_train_true['ds'],y=y_train_true['y'],name='Train Observed',line=dict(color='rgba(50, 205, 50, 0.7)' )), # go.Scatter(x=y_train_pred['ds'],y=y_train_pred['yhat'],name='Prophet Pred.(10 Item)',line=dict(color='#32CD32', dash='dot')), go.Scatter(x=y_test_true['ds'], y=y_test_true['y'],name='Observed',line=dict(color='rgba(218, 112, 214, 0.5)')), go.Scatter(x=y_test_pred['ds'],y=y_test_pred['yhat'],name='Historical Forecast',line=dict(color='#9400D3', dash='dash')), go.Scatter(x=future_prediction['ds'],y=future_prediction['yhat'],name='Future Forecast',line=dict(color='Dark Orange', dash='dot'))]) fig.update_layout( xaxis_title='Date', yaxis_title='Order Demand', margin=dict(l=0, r=0, t=50, b=0), xaxis=dict(title_font=dict(size=20)), yaxis=dict(title_font=dict(size=20))) fig.update_layout(width=700,height=400) tab1.plotly_chart(fig) #----------------------------------------Tab-2------------------------------------------------------------ results=y_test_pred.reset_index() results['store']='store_1' results['item']=item tab2.markdown("""

📃 Forecast Table

""", unsafe_allow_html=True) final_r=pd.concat([future_prediction[['ds','store','item','yhat']],results[['ds','store','item','yhat']]]).sort_values('ds').drop_duplicates().reset_index(drop=True) csv = convert_df(final_r) tab2.dataframe(final_r,width=500) tab2.download_button( "Download", csv, "file.csv", "text/csv", key='download-csv' ) #------------------------------------------Tab-3-------------------------------------------------- # train_a=fb_train_data.loc[fb_train_data[item]==1][['ds','sales']] # # train_a['store']=1 # # train_a['item']=item # test_a=fb_test_data.loc[fb_test_data[item]==1][['ds','sales']] # # test_a['store']=1 # # test_a['item']=item.split('_')[-1] # actual_final_data=pd.concat([train_a,test_a]) # actual_final_data['store']=1 # actual_final_data['item']=item.split('_')[-1] # actual_final_data['ds']=actual_final_data['ds'].dt.date # actual_final_data.rename({"ds":'date'},inplace=True) # tab3.dataframe(actual_final_data[['date','store','item','sales']],width=500) except: st.sidebar.error('Model Not Loaded successfully!',icon="🚨")