-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy path1_🤖_Recommender.py
205 lines (189 loc) · 9.05 KB
/
1_🤖_Recommender.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
import streamlit as st
import tensorflow as tf
import random
import numpy as np
import matplotlib as mpl
import matplotlib.pyplot as plt
import pandas as pd
from sklearn.linear_model import LogisticRegression
from streamlit_extras.stylable_container import stylable_container
pd.options.mode.chained_assignment = None
N_IMGS = 6
# init session state
if 'feedback_df' not in st.session_state:
st.session_state['feedback_df'] = pd.DataFrame({
"price_delta": [0],
"label_delta": [0],
"color_delta": [0],
"closest": [0],
})
if "indices" not in st.session_state:
st.session_state["indices"] = range(N_IMGS)
st.title("Fashion Recommender 🤖")
new_setting_button = st.sidebar.button("New random setting", key="new_setting")
if new_setting_button:
st.session_state["indices"] = [i + N_IMGS for i in st.session_state["indices"]]
np.random.seed(st.session_state["indices"][0])
random.seed(st.session_state["indices"][0])
colors = ["#5933d6", "#a80390", "#000000"]
cmaps = {}
for color in colors:
cmaps[color] = mpl.colors.LinearSegmentedColormap.from_list(
"custom", ["#ffffff", color]
)
itemtype = ["t-shirt", "trouser", "sneaker"]
itemtype_num_orig = [0, 1, 7]
itemtype_num = [1, 2, 2.5]
def get_price(itemtype_num, color):
noise = random.gauss(0, 0.5)
if color == "#000000":
return (itemtype_num + 1) * 10
elif color == "#a80390":
return (itemtype_num + 1) * 20 + noise
else:
return (itemtype_num + 1) * 22 + noise
def number_to_itemtype(input_num):
for i, num in enumerate(itemtype_num):
if input_num == num:
return itemtype[i]
def number_to_itemtype_num(input_num):
for i, num in enumerate(itemtype_num_orig):
if input_num == num:
return itemtype_num[i]
(x_train, y_train), (_, _) = tf.keras.datasets.fashion_mnist.load_data()
x_train = x_train[(y_train == 0) | (y_train == 1) | (y_train == 7)]
y_train = y_train[(y_train == 0) | (y_train == 1) | (y_train == 7)]
y_train = np.array([number_to_itemtype_num(y) for y in y_train])
picture_df = pd.DataFrame({
"index": st.session_state["indices"],
"label": y_train[st.session_state["indices"]],
})
picture_df["color"] = [random.choice(colors) for _ in range(N_IMGS)]
picture_df["price"] = [
np.round(get_price(label, color) , 2)
for label, color in zip(picture_df["label"], picture_df["color"])
]
picture_df["label_str"] = [number_to_itemtype(num) for num in picture_df["label"]]
picture_df["price_delta"] = np.abs(picture_df["price"] - picture_df["price"].iloc[0])
picture_df["label_delta"] = np.abs(picture_df["label"] - picture_df["label"].iloc[0])
picture_df["color_delta"] = (picture_df["color"] != picture_df["color"].iloc[0]).astype(int)
def display_image(index, color_map="viridis"):
import matplotlib.pyplot as plt
plt.imshow(x_train[index] / 255, cmap=color_map)
plt.axis('off')
st.pyplot(plt)
ref_cols = st.columns([3, 3, 3])
with ref_cols[1]:
st.write("Currently selected item:")
display_image(picture_df["index"].iloc[0], color_map=cmaps[picture_df["color"].iloc[0]])
st.write(f"Item: **{picture_df['label_str'].iloc[0]}**\n\nPrice: {picture_df['price'].iloc[0]}€")
st.write("Now the recommender should choose the next best item for you from the following options:")
expander_recommend = st.sidebar.expander("Tune recommendations", expanded=False)
with expander_recommend:
should_recommend = st.checkbox("Recommend?")
price_weight = st.number_input("Price weight", min_value=-100_000.0, max_value=100_000.0, value=0.0, step=0.1)
label_weight = st.number_input("Item weight", min_value=-100_000.0, max_value=100_000.0, value=0.0, step=0.1)
color_weight = st.number_input("Color weight", min_value=-100_000.0, max_value=100_000.0, value=0.0, step=0.1)
if should_recommend:
distances = (
price_weight * picture_df["price_delta"] +
label_weight * picture_df["label_delta"] +
color_weight * picture_df["color_delta"]
)
picture_df["distance"] = distances
picture_df = picture_df.sort_values("distance")
extender_feedback = st.sidebar.expander("Collect Feedback", expanded=False)
with extender_feedback:
record_feedback = st.checkbox("Record feedback?")
if record_feedback:
st.info("You can provide feedback on the recommendations by clicking 👍 or 👎. This feedback is recorded and can be used for training a customized ML model.")
picture_df = picture_df.iloc[1:].copy()
picture_df = picture_df.reset_index(drop=True)
picture_df["distance_feedback"] = -1
img_cols = st.columns(N_IMGS-1)
for i, col in enumerate(img_cols):
with col:
display_image(
picture_df["index"].iloc[i],
color_map=cmaps[picture_df["color"].iloc[i]]
)
st.write(f"Item: **{picture_df['label_str'].iloc[i]}**")
st.write(f"Price: {picture_df['price'].iloc[i]}€")
if should_recommend:
st.write(f"Distance: {picture_df['distance'].iloc[i]:.2f}")
if record_feedback:
# green button for closest, red for not closest
further_cols = st.columns([1, 1])
with further_cols[0]:
with stylable_container(
"green",
css_styles="""
button {
background-color: #4CAF50;
color: black;
}""",
):
if st.button("👍", key=i, use_container_width=True):
picture_df["distance_feedback"].loc[i] = 0
update_df = picture_df.copy()
update_df["closest"] = 0
update_df = update_df[update_df["distance_feedback"] == 0]
st.session_state['feedback_df'] = pd.concat(
[st.session_state['feedback_df'], update_df[["price_delta", "label_delta", "color_delta", "closest"]]],
)
with further_cols[1]:
with stylable_container(
"red",
css_styles="""
button {
background-color: #a86c71;
color: black;
}""",
):
if st.button("👎", key=i+100, use_container_width=True):
picture_df["distance_feedback"].loc[i] = 1
update_df = picture_df.copy()
update_df["closest"] = 1
update_df = update_df[update_df["distance_feedback"] == 1]
st.session_state['feedback_df'] = pd.concat(
[st.session_state['feedback_df'], update_df[["price_delta", "label_delta", "color_delta", "closest"]]],
)
with st.expander("Show feedback", expanded=False):
display_feedback = st.session_state['feedback_df'].copy()
# rename the columns for better readability all _delta should be turned into difference
display_feedback.columns = [col.replace("_delta", " difference").replace("label", "item").title() for col in display_feedback.columns]
display_feedback["Best Choice"] = display_feedback["Closest"].replace({0: True, 1: False})
st.dataframe(display_feedback[["Price Difference", "Item Difference", "Color Difference", "Best Choice"]])
if st.button("Reset feedback", use_container_width=False):
st.session_state['feedback_df'] = pd.DataFrame({
"price_delta": [0],
"label_delta": [0],
"color_delta": [0],
"closest": [0],
})
with st.expander("ML recommendation for the weights", expanded=False):
fit_df = st.session_state['feedback_df'].copy()
# balance the dataset such that 50/50 are closest and not closest
n_closest = fit_df["closest"].sum()
n_not_closest = fit_df.shape[0] - n_closest
if n_closest > n_not_closest:
fit_df = pd.concat([
fit_df[fit_df["closest"] == 1].sample(n=n_not_closest, random_state=st.session_state["indices"][0]),
fit_df[fit_df["closest"] == 0]
])
elif n_closest < n_not_closest:
fit_df = pd.concat([
fit_df[fit_df["closest"] == 1],
fit_df[fit_df["closest"] == 0].sample(n=n_closest, random_state=st.session_state["indices"][0])
])
if fit_df.shape[0] > 1:
st.info("This is a simple logistic regression model that predicts which is the best next choice according to the user's preference. The input features are the differences in price, item, and color between the images. The more good examples you provide the better the model gets.")
X = fit_df[["price_delta", "label_delta", "color_delta"]]
y = fit_df["closest"]
clf = LogisticRegression(random_state=st.session_state["indices"][0]).fit(X, y)
for i, coef in enumerate(clf.coef_[0]):
if np.any(np.abs(clf.coef_[0]) < 0.01):
coef *= 10_000
st.write(f'{X.columns[i].replace("_delta", "").replace("label", "item").title()} weight: {coef:.2f}')
else:
st.warning("Please provide more feedback to train the model.")