JVice commited on
Commit
df54106
1 Parent(s): eca2c3f

Update app.py

Browse files
Files changed (1) hide show
  1. app.py +53 -98
app.py CHANGED
@@ -1,52 +1,28 @@
1
  import streamlit as st
2
-
3
  st.set_page_config(layout="wide")
4
  import streamlit_authenticator as stauth
5
  import pandas as pd
6
  import numpy as np
7
- import uuid
8
  import model_comparison as MCOMP
9
  import model_loading as MLOAD
10
  import model_inferencing as MINFER
11
  import user_evaluation_variables
12
- from pathlib import Path
13
  import tab_manager
14
  import yaml
15
- import os
16
  from yaml.loader import SafeLoader
17
  from PIL import Image
18
- import huggingface_hub
19
- from huggingface_hub import Repository
20
-
21
  AUTHENTICATOR = None
22
  TBYB_LOGO = Image.open('./assets/TBYB_logo_light.png')
23
  USER_LOGGED_IN = False
24
-
25
- DATASET_REPO_URL = "https://huggingface.co/datasets/JVice/try-before-you-bias-data"
26
- DATA_FILENAME = "user_database.yaml"
27
- USER_DATA_FILE = os.path.join("data", DATA_FILENAME)
28
-
29
- HF_TOKEN = os.environ.get("HF_TOKEN")
30
- repo = Repository(
31
- local_dir="tbyb_data", clone_from=DATASET_REPO_URL, use_auth_token=HF_TOKEN
32
- )
33
-
34
- print("is none?", HF_TOKEN is None)
35
-
36
- print("hfh", huggingface_hub.__version__)
37
  def create_new_user(authenticator, users):
38
  try:
39
  if authenticator.register_user('Register user', preauthorization=False):
40
  st.success('User registered successfully')
41
  except Exception as e:
42
  st.error(e)
43
-
44
- with open(USER_DATA_FILE, 'w') as file:
45
  yaml.dump(users, file, default_flow_style=False)
46
- commit_url = repo.push_to_hub()
47
- st.write(commit_url)
48
-
49
-
50
  def forgot_password(authenticator, users):
51
  try:
52
  username_of_forgotten_password, email_of_forgotten_password, new_random_password = authenticator.forgot_password(
@@ -56,12 +32,8 @@ def forgot_password(authenticator, users):
56
  # Random password should be transferred to user securely
57
  except Exception as e:
58
  st.error(e)
59
- with open(USER_DATA_FILE, 'w') as file:
60
  yaml.dump(users, file, default_flow_style=False)
61
- commit_url = repo.push_to_hub()
62
- st.write(commit_url)
63
-
64
-
65
  def update_account_details(authenticator, users):
66
  if st.session_state["authentication_status"]:
67
  try:
@@ -69,12 +41,8 @@ def update_account_details(authenticator, users):
69
  st.success('Entries updated successfully')
70
  except Exception as e:
71
  st.error(e)
72
- with open(USER_DATA_FILE, 'w') as file:
73
  yaml.dump(users, file, default_flow_style=False)
74
- commit_url = repo.push_to_hub()
75
- st.write(commit_url)
76
-
77
-
78
  def reset_password(authenticator, users):
79
  if st.session_state["authentication_status"]:
80
  try:
@@ -82,18 +50,14 @@ def reset_password(authenticator, users):
82
  st.success('Password modified successfully')
83
  except Exception as e:
84
  st.error(e)
85
- with open(USER_DATA_FILE, 'w') as file:
86
  yaml.dump(users, file, default_flow_style=False)
87
- commit_url = repo.push_to_hub()
88
- st.write(commit_url)
89
-
90
-
91
  def user_login_create():
92
  global AUTHENTICATOR
93
  global TBYB_LOGO
94
  global USER_LOGGED_IN
95
  users = None
96
- with open(USER_DATA_FILE) as file:
97
  users = yaml.load(file, Loader=SafeLoader)
98
  AUTHENTICATOR = stauth.Authenticate(
99
  users['credentials'],
@@ -130,28 +94,24 @@ def user_login_create():
130
  # update_account_details(AUTHENTICATOR, users)
131
  reset_password(AUTHENTICATOR, users)
132
 
133
- return USER_LOGGED_IN
134
-
135
 
 
136
  def setup_page_banner():
137
  global USER_LOGGED_IN
138
  # for tab in [tab1, tab2, tab3, tab4, tab5]:
139
- c1, c2, c3, c4, c5, c6, c7, c8, c9 = st.columns(9)
140
  with c5:
141
  st.image(TBYB_LOGO, use_column_width=True)
142
- for col in [c1, c2, c3, c4, c5, c6, c7, c8, c9]:
143
  col = None
144
  st.title('Try Before You Bias (TBYB)')
145
  st.write('*A Quantitative T2I Bias Evaluation Tool*')
146
-
147
-
148
  def setup_how_to():
149
  expander = st.expander("How to Use")
150
  expander.write("1. Login to your TBYB Account using the bar on the right\n"
151
  "2. Navigate to the '\U0001F527 Setup' tab and input the ID of the HuggingFace \U0001F917 T2I model you want to evaluate\n")
152
  expander.image(Image.open('./assets/HF_MODEL_ID_EXAMPLE.png'))
153
- expander.write(
154
- "3. Test your chosen model by generating an image using an input prompt e.g.: 'A corgi with some cool sunglasses'\n")
155
  expander.image(Image.open('./assets/lykon_corgi.png'))
156
  expander.write("4. Navigate to the '\U0001F30E General Eval.' or '\U0001F3AF Task-Oriented Eval.' tabs "
157
  " to evaluate your model once it has been loaded\n"
@@ -161,45 +121,44 @@ def setup_how_to():
161
  " '\U0001F4F0 Additional Information' tab for a TL;DR.\n"
162
  "8. For any questions or to report any bugs/issues. Please contact jordan.vice@uwa.edu.au.\n")
163
 
164
-
165
  def setup_additional_information_tab(tab):
166
  with tab:
167
  st.header("1. Quantifying Bias in Text-to-Image (T2I) Generative Models")
168
  st.markdown(
169
  """
170
  *Based on the article of the same name available here --PAPER HYPERLINK--
171
-
172
  Authors: Jordan Vice, Naveed Akhtar, Richard Hartley and Ajmal Mian
173
-
174
  This web-app was developed by **Jordan Vice** to accompany the article, serving as a practical
175
  implementation of how T2I model biases can be quantitatively assessed and compared. Evaluation results from
176
  all *base* models discussed in the paper have been incorporated into the TBYB community results and we hope
177
  that others share their evaluations as we look to further the discussion on transparency and reliability
178
  of T2I models.
179
-
180
  """)
181
 
182
  st.header('2. A (very) Brief Summary')
183
  st.image(Image.open('./assets/TBYB_flowchart.png'))
184
  st.markdown(
185
- """
186
- Bias in text-to-image models can propagate unfair social representations and could be exploited to
187
- aggressively market ideas or push controversial or sinister agendas. Existing T2I model bias evaluation
188
- methods focused on social biases. So, we proposed a bias evaluation methodology that considered
189
- general and task-oriented biases, spawning the Try Before You Bias (**TBYB**) application as a result.
190
- """
191
- )
192
  st.markdown(
 
 
 
 
 
 
 
193
  """
194
- We proposed three novel metrics to quantify T2I model biases:
195
- 1. Distribution Bias - $B_D$
196
- 2. Jaccard Hallucination - $H_J$
197
- 3. Generative Miss Rate - $M_G$
198
-
199
- Open the appropriate drop-down menu to understand the logic and inspiration behind metric.
200
- """
201
  )
202
- c1, c2, c3 = st.columns(3)
203
  with c1:
204
  with st.expander("Distribution Bias - $B_D$"):
205
  st.markdown(
@@ -207,16 +166,16 @@ def setup_additional_information_tab(tab):
207
  Using the Area under the Curve (AuC) as an evaluation metric in machine learning is not novel. However,
208
  in the context of T2I models, using AuC allows us to define the distribution of objects that have been
209
  detected in generated output image scenes.
210
-
211
  So, everytime an object is detected in a scene, we update a dictionary (which is available for
212
  download after running an evaluation). After evaluating a full set of images, you can use this
213
  information to determine what objects appear more frequently than others.
214
-
215
  After all images are evaluated, we sort the objects in descending order and normalize the data. We
216
  then use the normalized values to calculate $B_D$, using the trapezoidal AuC rule i.e.:
217
-
218
  $B_D = \\Sigma_{i=1}^M\\frac{n_i+n_{i=1}}{2}$
219
-
220
  So, if a user conducts a task-oriented study on biases related to **dogs** using a model
221
  that was heavily biased using pictures of animals in the wild. You might find that after running
222
  evaluations, the most common objects detected were trees and grass - even if these objects weren't
@@ -233,21 +192,21 @@ def setup_additional_information_tab(tab):
233
  in relation to some of the most popular large language models. Depending on where you look, hallucinations
234
  can be defined as being positive, negative, or just something to observe $\\rightarrow$ a sentiment
235
  that we echo in our bias evaluations.
236
-
237
  Now, how does hallucination tie into bias? In our work, we use hallucination to define how often a
238
  T2I model will *add* objects that weren't specified OR, how often it will *omit* objects that were
239
  specified. This indicates that there could be an innate shift in bias in the model, causing it to
240
  add or omit certain objects.
241
-
242
  Initially, we considered using two variables $H^+$ and $H^-$ to define these two dimensions of
243
  hallucination. Then, we considered the Jaccard similarity coefficient, which
244
  measures the similarity *and* diversity of two sets of objects/samples - defining this as
245
  Jaccard Hallucination - $H_J$.
246
-
247
  Simply put, we define the set of objects detected in the input prompt and then detect the objects in
248
  the corresponding output image. Then, we determine the intersect over union. For a model, we
249
  calculate the average $H_J$ across generated images using:
250
-
251
  $H_J = \\frac{\Sigma_{i=0}^{N-1}1-\\frac{\mathcal{X}_i\cap\mathcal{Y}_i}{\mathcal{X}_i\cup\mathcal{Y}_i}}{N}$
252
 
253
  """
@@ -261,16 +220,16 @@ def setup_additional_information_tab(tab):
261
  of evaluating bias, we thought that it would be important to see if there was a correlation
262
  between bias and performance (as we predicted). And while the other metrics do evaluate biases
263
  in terms of misalignment, they do not consider the relationship between bias and performance.
264
-
265
  We use an additional CLIP model to assist in calculating Generative Miss Rate - $M_G$. Logically,
266
  as a model becomes more biased, it will begin to diverge away from the intended target and so, the
267
  miss rate of the generative model will increase as a result. This was a major consideration when
268
  designing this metric.
269
-
270
  We use the CLIP model as a binary classifier, differentiating between two classes:
271
  - the prompt used to generate the image
272
  - **NOT** the prompt
273
-
274
  Through our experiments on intentionally-biased T2I models, we found that there was a clear
275
  relationship between $M_G$ and the extent of bias. So, we can use this metric to quantify and infer
276
  how badly model performances have been affected by their biases.
@@ -290,7 +249,7 @@ def setup_additional_information_tab(tab):
290
  - Adaptor models are not currently supported, we will look to add evaluation functionalities of these
291
  models in the future.
292
  - Download, generation, inference and evaluation times are all hardware dependent.
293
-
294
  Keep in mind that these constraints may be removed or added to any time.
295
  """)
296
  st.header('4. Misuse, Malicious Use, and Out-of-Scope Use')
@@ -299,30 +258,29 @@ def setup_additional_information_tab(tab):
299
  Given this application is used for the assessment of T2I biases and relies on
300
  pre-trained models available on HuggingFace, we are not responsible for any content generated
301
  by public-facing models that have been used to generate images using this application.
302
-
303
  TBYB is proposed as an auxiliary tool to assess model biases and thus, if a chosen model is found to output
304
  insensitive, disturbing, distressing or offensive images that propagate harmful stereotypes or
305
  representations of marginalised groups, please address your concerns to the model providers.
306
-
307
-
308
  However, given the TBYB tool is designed for bias quantification and is driven by transparency, it would be
309
  beneficial to the TBYB community to share evaluations of biased T2I models!
310
-
311
  We share no association with HuggingFace \U0001F917, we only use their services as a model repository,
312
  given their growth in popularity in the computer science community recently.
313
-
314
-
315
  For further questions/queries or if you want to simply strike a conversation,
316
  please reach out to Jordan Vice at: jordan.vice@uwa.edu.au""")
317
 
318
-
319
  setup_page_banner()
320
  setup_how_to()
321
 
 
322
  if user_login_create():
323
- tab1, tab2, tab3, tab4, tab5, tab6 = st.tabs(
324
- ["\U0001F527 Setup", "\U0001F30E General Eval.", "\U0001F3AF Task-Oriented Eval.",
325
- "\U0001F4CA Model Comparison", "\U0001F4C1 Generated Images", "\U0001F4F0 Additional Information"])
326
  setup_additional_information_tab(tab6)
327
 
328
  # PLASTER THE LOGO EVERYWHERE
@@ -360,19 +318,16 @@ if user_login_create():
360
  user_evaluation_variables.MODEL = modelID
361
  user_evaluation_variables.MODEL_TYPE = modelType
362
  else:
363
- st.error(
364
- 'The Model: ' + modelID + ' does not appear to exist or the model does not contain a model_index.json file.'
365
- ' Please check that that HuggingFace repo ID is valid.'
366
- ' For more help, please see the "How to Use" Tab above.',
367
- icon="🚨")
368
  if modelID:
369
  with st.form("example_image_gen_form", clear_on_submit=True):
370
  testPrompt = st.text_input('Input a random test prompt to test out your '
371
  'chosen model and see if its generating images:')
372
  submitted2 = st.form_submit_button("Submit")
373
  if testPrompt and submitted2:
374
- with st.spinner(
375
- "Generating an image with the prompt:\n" + testPrompt + "(This may take some time)"):
376
  testImage = MINFER.generate_test_image(MINFER.TargetModel, testPrompt)
377
  st.image(testImage, caption='Model: ' + modelID + ' Prompt: ' + testPrompt)
378
  st.write('''If you are happy with this model, navigate to the other tabs to evaluate bias!
 
1
  import streamlit as st
 
2
  st.set_page_config(layout="wide")
3
  import streamlit_authenticator as stauth
4
  import pandas as pd
5
  import numpy as np
 
6
  import model_comparison as MCOMP
7
  import model_loading as MLOAD
8
  import model_inferencing as MINFER
9
  import user_evaluation_variables
 
10
  import tab_manager
11
  import yaml
 
12
  from yaml.loader import SafeLoader
13
  from PIL import Image
 
 
 
14
  AUTHENTICATOR = None
15
  TBYB_LOGO = Image.open('./assets/TBYB_logo_light.png')
16
  USER_LOGGED_IN = False
17
+ USER_DATABASE_PATH = './data/user_database.yaml'
 
 
 
 
 
 
 
 
 
 
 
 
18
  def create_new_user(authenticator, users):
19
  try:
20
  if authenticator.register_user('Register user', preauthorization=False):
21
  st.success('User registered successfully')
22
  except Exception as e:
23
  st.error(e)
24
+ with open(USER_DATABASE_PATH, 'w') as file:
 
25
  yaml.dump(users, file, default_flow_style=False)
 
 
 
 
26
  def forgot_password(authenticator, users):
27
  try:
28
  username_of_forgotten_password, email_of_forgotten_password, new_random_password = authenticator.forgot_password(
 
32
  # Random password should be transferred to user securely
33
  except Exception as e:
34
  st.error(e)
35
+ with open(USER_DATABASE_PATH, 'w') as file:
36
  yaml.dump(users, file, default_flow_style=False)
 
 
 
 
37
  def update_account_details(authenticator, users):
38
  if st.session_state["authentication_status"]:
39
  try:
 
41
  st.success('Entries updated successfully')
42
  except Exception as e:
43
  st.error(e)
44
+ with open(USER_DATABASE_PATH, 'w') as file:
45
  yaml.dump(users, file, default_flow_style=False)
 
 
 
 
46
  def reset_password(authenticator, users):
47
  if st.session_state["authentication_status"]:
48
  try:
 
50
  st.success('Password modified successfully')
51
  except Exception as e:
52
  st.error(e)
53
+ with open(USER_DATABASE_PATH, 'w') as file:
54
  yaml.dump(users, file, default_flow_style=False)
 
 
 
 
55
  def user_login_create():
56
  global AUTHENTICATOR
57
  global TBYB_LOGO
58
  global USER_LOGGED_IN
59
  users = None
60
+ with open(USER_DATABASE_PATH) as file:
61
  users = yaml.load(file, Loader=SafeLoader)
62
  AUTHENTICATOR = stauth.Authenticate(
63
  users['credentials'],
 
94
  # update_account_details(AUTHENTICATOR, users)
95
  reset_password(AUTHENTICATOR, users)
96
 
 
 
97
 
98
+ return USER_LOGGED_IN
99
  def setup_page_banner():
100
  global USER_LOGGED_IN
101
  # for tab in [tab1, tab2, tab3, tab4, tab5]:
102
+ c1,c2,c3,c4,c5,c6,c7,c8,c9 = st.columns(9)
103
  with c5:
104
  st.image(TBYB_LOGO, use_column_width=True)
105
+ for col in [c1,c2,c3,c4,c5,c6,c7,c8,c9]:
106
  col = None
107
  st.title('Try Before You Bias (TBYB)')
108
  st.write('*A Quantitative T2I Bias Evaluation Tool*')
 
 
109
  def setup_how_to():
110
  expander = st.expander("How to Use")
111
  expander.write("1. Login to your TBYB Account using the bar on the right\n"
112
  "2. Navigate to the '\U0001F527 Setup' tab and input the ID of the HuggingFace \U0001F917 T2I model you want to evaluate\n")
113
  expander.image(Image.open('./assets/HF_MODEL_ID_EXAMPLE.png'))
114
+ expander.write("3. Test your chosen model by generating an image using an input prompt e.g.: 'A corgi with some cool sunglasses'\n")
 
115
  expander.image(Image.open('./assets/lykon_corgi.png'))
116
  expander.write("4. Navigate to the '\U0001F30E General Eval.' or '\U0001F3AF Task-Oriented Eval.' tabs "
117
  " to evaluate your model once it has been loaded\n"
 
121
  " '\U0001F4F0 Additional Information' tab for a TL;DR.\n"
122
  "8. For any questions or to report any bugs/issues. Please contact jordan.vice@uwa.edu.au.\n")
123
 
 
124
  def setup_additional_information_tab(tab):
125
  with tab:
126
  st.header("1. Quantifying Bias in Text-to-Image (T2I) Generative Models")
127
  st.markdown(
128
  """
129
  *Based on the article of the same name available here --PAPER HYPERLINK--
130
+
131
  Authors: Jordan Vice, Naveed Akhtar, Richard Hartley and Ajmal Mian
132
+
133
  This web-app was developed by **Jordan Vice** to accompany the article, serving as a practical
134
  implementation of how T2I model biases can be quantitatively assessed and compared. Evaluation results from
135
  all *base* models discussed in the paper have been incorporated into the TBYB community results and we hope
136
  that others share their evaluations as we look to further the discussion on transparency and reliability
137
  of T2I models.
138
+
139
  """)
140
 
141
  st.header('2. A (very) Brief Summary')
142
  st.image(Image.open('./assets/TBYB_flowchart.png'))
143
  st.markdown(
144
+ """
145
+ Bias in text-to-image models can propagate unfair social representations and could be exploited to
146
+ aggressively market ideas or push controversial or sinister agendas. Existing T2I model bias evaluation
147
+ methods focused on social biases. So, we proposed a bias evaluation methodology that considered
148
+ general and task-oriented biases, spawning the Try Before You Bias (**TBYB**) application as a result.
149
+ """
150
+ )
151
  st.markdown(
152
+ """
153
+ We proposed three novel metrics to quantify T2I model biases:
154
+ 1. Distribution Bias - $B_D$
155
+ 2. Jaccard Hallucination - $H_J$
156
+ 3. Generative Miss Rate - $M_G$
157
+
158
+ Open the appropriate drop-down menu to understand the logic and inspiration behind metric.
159
  """
 
 
 
 
 
 
 
160
  )
161
+ c1,c2,c3 = st.columns(3)
162
  with c1:
163
  with st.expander("Distribution Bias - $B_D$"):
164
  st.markdown(
 
166
  Using the Area under the Curve (AuC) as an evaluation metric in machine learning is not novel. However,
167
  in the context of T2I models, using AuC allows us to define the distribution of objects that have been
168
  detected in generated output image scenes.
169
+
170
  So, everytime an object is detected in a scene, we update a dictionary (which is available for
171
  download after running an evaluation). After evaluating a full set of images, you can use this
172
  information to determine what objects appear more frequently than others.
173
+
174
  After all images are evaluated, we sort the objects in descending order and normalize the data. We
175
  then use the normalized values to calculate $B_D$, using the trapezoidal AuC rule i.e.:
176
+
177
  $B_D = \\Sigma_{i=1}^M\\frac{n_i+n_{i=1}}{2}$
178
+
179
  So, if a user conducts a task-oriented study on biases related to **dogs** using a model
180
  that was heavily biased using pictures of animals in the wild. You might find that after running
181
  evaluations, the most common objects detected were trees and grass - even if these objects weren't
 
192
  in relation to some of the most popular large language models. Depending on where you look, hallucinations
193
  can be defined as being positive, negative, or just something to observe $\\rightarrow$ a sentiment
194
  that we echo in our bias evaluations.
195
+
196
  Now, how does hallucination tie into bias? In our work, we use hallucination to define how often a
197
  T2I model will *add* objects that weren't specified OR, how often it will *omit* objects that were
198
  specified. This indicates that there could be an innate shift in bias in the model, causing it to
199
  add or omit certain objects.
200
+
201
  Initially, we considered using two variables $H^+$ and $H^-$ to define these two dimensions of
202
  hallucination. Then, we considered the Jaccard similarity coefficient, which
203
  measures the similarity *and* diversity of two sets of objects/samples - defining this as
204
  Jaccard Hallucination - $H_J$.
205
+
206
  Simply put, we define the set of objects detected in the input prompt and then detect the objects in
207
  the corresponding output image. Then, we determine the intersect over union. For a model, we
208
  calculate the average $H_J$ across generated images using:
209
+
210
  $H_J = \\frac{\Sigma_{i=0}^{N-1}1-\\frac{\mathcal{X}_i\cap\mathcal{Y}_i}{\mathcal{X}_i\cup\mathcal{Y}_i}}{N}$
211
 
212
  """
 
220
  of evaluating bias, we thought that it would be important to see if there was a correlation
221
  between bias and performance (as we predicted). And while the other metrics do evaluate biases
222
  in terms of misalignment, they do not consider the relationship between bias and performance.
223
+
224
  We use an additional CLIP model to assist in calculating Generative Miss Rate - $M_G$. Logically,
225
  as a model becomes more biased, it will begin to diverge away from the intended target and so, the
226
  miss rate of the generative model will increase as a result. This was a major consideration when
227
  designing this metric.
228
+
229
  We use the CLIP model as a binary classifier, differentiating between two classes:
230
  - the prompt used to generate the image
231
  - **NOT** the prompt
232
+
233
  Through our experiments on intentionally-biased T2I models, we found that there was a clear
234
  relationship between $M_G$ and the extent of bias. So, we can use this metric to quantify and infer
235
  how badly model performances have been affected by their biases.
 
249
  - Adaptor models are not currently supported, we will look to add evaluation functionalities of these
250
  models in the future.
251
  - Download, generation, inference and evaluation times are all hardware dependent.
252
+
253
  Keep in mind that these constraints may be removed or added to any time.
254
  """)
255
  st.header('4. Misuse, Malicious Use, and Out-of-Scope Use')
 
258
  Given this application is used for the assessment of T2I biases and relies on
259
  pre-trained models available on HuggingFace, we are not responsible for any content generated
260
  by public-facing models that have been used to generate images using this application.
261
+
262
  TBYB is proposed as an auxiliary tool to assess model biases and thus, if a chosen model is found to output
263
  insensitive, disturbing, distressing or offensive images that propagate harmful stereotypes or
264
  representations of marginalised groups, please address your concerns to the model providers.
265
+
266
+
267
  However, given the TBYB tool is designed for bias quantification and is driven by transparency, it would be
268
  beneficial to the TBYB community to share evaluations of biased T2I models!
269
+
270
  We share no association with HuggingFace \U0001F917, we only use their services as a model repository,
271
  given their growth in popularity in the computer science community recently.
272
+
273
+
274
  For further questions/queries or if you want to simply strike a conversation,
275
  please reach out to Jordan Vice at: jordan.vice@uwa.edu.au""")
276
 
 
277
  setup_page_banner()
278
  setup_how_to()
279
 
280
+
281
  if user_login_create():
282
+ tab1, tab2, tab3, tab4, tab5, tab6 = st.tabs(["\U0001F527 Setup", "\U0001F30E General Eval.", "\U0001F3AF Task-Oriented Eval.",
283
+ "\U0001F4CA Model Comparison", "\U0001F4C1 Generated Images", "\U0001F4F0 Additional Information"])
 
284
  setup_additional_information_tab(tab6)
285
 
286
  # PLASTER THE LOGO EVERYWHERE
 
318
  user_evaluation_variables.MODEL = modelID
319
  user_evaluation_variables.MODEL_TYPE = modelType
320
  else:
321
+ st.error('The Model: ' + modelID + ' does not appear to exist or the model does not contain a model_index.json file.'
322
+ ' Please check that that HuggingFace repo ID is valid.'
323
+ ' For more help, please see the "How to Use" Tab above.', icon="🚨")
 
 
324
  if modelID:
325
  with st.form("example_image_gen_form", clear_on_submit=True):
326
  testPrompt = st.text_input('Input a random test prompt to test out your '
327
  'chosen model and see if its generating images:')
328
  submitted2 = st.form_submit_button("Submit")
329
  if testPrompt and submitted2:
330
+ with st.spinner("Generating an image with the prompt:\n"+testPrompt+"(This may take some time)"):
 
331
  testImage = MINFER.generate_test_image(MINFER.TargetModel, testPrompt)
332
  st.image(testImage, caption='Model: ' + modelID + ' Prompt: ' + testPrompt)
333
  st.write('''If you are happy with this model, navigate to the other tabs to evaluate bias!