Spaces:
Paused
Paused
backup
Browse files
app.py
CHANGED
@@ -1,4 +1,3 @@
|
|
1 |
-
|
2 |
import csv
|
3 |
import json
|
4 |
import os
|
@@ -20,28 +19,30 @@ import torchvision
|
|
20 |
from huggingface_hub import HfApi, login, snapshot_download
|
21 |
from PIL import Image
|
22 |
|
23 |
-
|
24 |
-
|
25 |
-
login(token=session_token)
|
26 |
|
27 |
|
28 |
csv.field_size_limit(sys.maxsize)
|
29 |
|
30 |
np.random.seed(int(time.time()))
|
31 |
|
32 |
-
with open(
|
33 |
-
knn_results
|
34 |
|
35 |
with open("imagenet-labels.json") as f:
|
36 |
wnid_to_label = json.load(f)
|
37 |
|
38 |
-
with open(
|
39 |
id_to_labels = json.load(f)
|
40 |
|
|
|
41 |
|
42 |
-
bad_items = open(
|
43 |
-
bad_items = [x.split(
|
44 |
-
bad_items = [int(x) for x in bad_items if x !=
|
|
|
|
|
45 |
|
46 |
# download and extract folders
|
47 |
|
@@ -54,7 +55,9 @@ gdown.cached_download(
|
|
54 |
|
55 |
# EXTRACT if needed
|
56 |
|
57 |
-
if not os.path.exists("./imagenet_traning_samples") or not os.path.exists(
|
|
|
|
|
58 |
torchvision.datasets.utils.extract_archive(
|
59 |
from_path="data.zip",
|
60 |
to_path="./",
|
@@ -64,11 +67,12 @@ if not os.path.exists("./imagenet_traning_samples") or not os.path.exists("./knn
|
|
64 |
imagenet_hard = datasets.load_dataset("taesiri/imagenet-hard", split="validation")
|
65 |
|
66 |
|
67 |
-
def update_snapshot():
|
68 |
output_dir = snapshot_download(
|
69 |
-
|
|
|
|
|
70 |
)
|
71 |
-
total_size = len(imagenet_hard)
|
72 |
files = glob(f"{output_dir}/*.json")
|
73 |
|
74 |
df = pd.DataFrame()
|
@@ -83,42 +87,35 @@ def update_snapshot():
|
|
83 |
rows.append(tdf)
|
84 |
|
85 |
df = pd.DataFrame(rows, columns=columns)
|
|
|
86 |
|
87 |
-
return df
|
88 |
-
|
89 |
|
90 |
-
# df = update_snapshot()
|
91 |
|
92 |
-
|
|
|
|
|
93 |
|
94 |
-
|
95 |
-
|
96 |
-
|
97 |
-
all_ids = bad_items
|
98 |
|
99 |
-
|
100 |
-
|
101 |
-
|
|
|
|
|
|
|
|
|
102 |
|
103 |
-
weights = [id_counts[id] for id in all_ids]
|
104 |
-
inverse_weights = [1 / (count + 1) for count in weights]
|
105 |
-
normalized_weights = [w / sum(inverse_weights) for w in inverse_weights]
|
106 |
-
|
107 |
-
sampled_ids = np.random.choice(all_ids, size=sample_size, replace=False, p=normalized_weights)
|
108 |
-
return sampled_ids
|
109 |
-
|
110 |
-
|
111 |
-
def generate_dataset():
|
112 |
-
df, total_size = update_snapshot()
|
113 |
-
random_indices = sample_ids(df, total_size, NUMBER_OF_IMAGES)
|
114 |
random_images = [imagenet_hard[int(i)]["image"] for i in random_indices]
|
115 |
random_gt_ids = [imagenet_hard[int(i)]["label"] for i in random_indices]
|
116 |
-
random_gt_labels = [imagenet_hard[int(x)]["english_label"]
|
117 |
|
118 |
data = []
|
119 |
for i, image in enumerate(random_images):
|
120 |
data.append(
|
121 |
-
{
|
122 |
"id": random_indices[i],
|
123 |
"image": image,
|
124 |
"correct_label": random_gt_labels[i],
|
@@ -128,83 +125,86 @@ def generate_dataset():
|
|
128 |
return data
|
129 |
|
130 |
|
131 |
-
|
132 |
def string_to_image(text):
|
133 |
-
text = text.replace(
|
134 |
# Create a blank white square image
|
135 |
img = np.ones((220, 75, 3))
|
136 |
|
137 |
-
# Create a figure and axis object
|
138 |
fig, ax = plt.subplots(figsize=(6, 2.25))
|
139 |
-
|
140 |
-
# Plot the blank white image
|
141 |
ax.imshow(img, extent=[0, 1, 0, 1])
|
142 |
-
|
143 |
-
# Set the text in the center
|
144 |
-
ax.text(0.5, 0.75, text, fontsize=18, ha='center', va='center')
|
145 |
-
|
146 |
-
# Remove the axis labels and ticks
|
147 |
ax.set_xticks([])
|
148 |
ax.set_yticks([])
|
149 |
ax.set_xticklabels([])
|
150 |
ax.set_yticklabels([])
|
151 |
-
|
152 |
-
# Remove the axis spines
|
153 |
for spine in ax.spines.values():
|
154 |
spine.set_visible(False)
|
155 |
|
156 |
-
# Return the figure
|
157 |
return fig
|
158 |
|
159 |
|
|
|
|
|
|
|
|
|
160 |
|
161 |
-
|
162 |
-
|
163 |
-
with open('./trainingset_filenames.json', 'r') as f:
|
164 |
-
trainingset_filenames = json.load(f)
|
165 |
-
|
166 |
-
nns = knn_results[qid][:15]
|
167 |
-
labels = [wnid_to_label[trainingset_filenames[f"{x}"]] for x in nns]
|
168 |
-
label_counts = {x: labels.count(x) for x in set(labels)}
|
169 |
-
# sort by count
|
170 |
-
label_counts = {k: v for k, v in sorted(label_counts.items(), key=lambda item: item[1], reverse=True)}
|
171 |
-
# percetage
|
172 |
-
label_counts = {k: v/len(labels) for k, v in label_counts.items()}
|
173 |
-
return label_counts
|
174 |
-
|
175 |
-
|
176 |
-
from glob import glob
|
177 |
|
178 |
-
all_samples = glob('./imagenet_traning_samples/*.JPEG')
|
179 |
-
qid_to_sample = {int(x.split('/')[-1].split('.')[0].split('_')[0]): x for x in all_samples}
|
180 |
|
181 |
def get_training_samples(qid):
|
182 |
-
labels_id = imagenet_hard[int(qid)][
|
183 |
samples = [qid_to_sample[x] for x in labels_id]
|
184 |
return samples
|
185 |
|
186 |
|
187 |
-
knn_cache_path = "knn_cache_for_imagenet_hard"
|
188 |
-
imagenet_training_samples_path = "imagenet_traning_samples"
|
189 |
-
|
190 |
def load_sample(data, current_index):
|
191 |
image_id = data[current_index]["id"]
|
192 |
qimage = data[current_index]["image"]
|
193 |
-
|
194 |
labels = data[current_index]["correct_label"]
|
195 |
return qimage, labels
|
196 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
197 |
|
198 |
|
199 |
def update_app(decision, data, current_index, history, username):
|
|
|
200 |
if current_index == -1:
|
201 |
-
|
202 |
-
|
203 |
-
if current_index
|
204 |
time_stamp = int(time.time())
|
205 |
|
206 |
image_id = data[current_index]["id"]
|
207 |
-
# convert to percentage
|
208 |
dicision_dict = {
|
209 |
"id": int(image_id),
|
210 |
"user_id": username,
|
@@ -228,23 +228,56 @@ def update_app(decision, data, current_index, history, username):
|
|
228 |
|
229 |
os.remove(temp_filename)
|
230 |
|
231 |
-
|
232 |
-
|
|
|
233 |
|
234 |
-
current_index
|
235 |
-
|
236 |
-
image_id = data[current_index]["id"]
|
237 |
-
training_samples_image = get_training_samples(image_id)
|
238 |
-
training_samples_image = [Image.open(x).convert('RGB') for x in training_samples_image]
|
239 |
|
240 |
-
|
241 |
-
|
242 |
-
|
|
|
|
|
|
|
|
|
|
|
243 |
|
244 |
-
|
|
|
|
|
|
|
|
|
245 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
246 |
|
247 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
248 |
#query_image{
|
249 |
height: auto !important;
|
250 |
}
|
@@ -256,51 +289,87 @@ newcss = '''
|
|
256 |
#sample_gallery {
|
257 |
height: auto !important;
|
258 |
}
|
259 |
-
|
260 |
|
261 |
with gr.Blocks(css=newcss) as demo:
|
262 |
data_gr = gr.State({})
|
263 |
current_index = gr.State(-1)
|
264 |
history = gr.State({})
|
265 |
-
|
266 |
gr.Markdown("# Cleaning ImageNet-Hard!")
|
267 |
|
268 |
random_str = "".join(
|
269 |
random.choice(string.ascii_lowercase + string.digits) for _ in range(5)
|
270 |
)
|
271 |
|
272 |
-
|
|
|
|
|
273 |
|
274 |
with gr.Column():
|
275 |
with gr.Row():
|
276 |
accept_btn = gr.Button(value="Accept")
|
277 |
myabe_btn = gr.Button(value="Not Sure!")
|
278 |
reject_btn = gr.Button(value="Reject")
|
279 |
-
with gr.Row():
|
280 |
query_image = gr.Image(type="pil", label="Query", elem_id="query_image")
|
281 |
with gr.Column():
|
282 |
-
label_plot = gr.Plot(
|
283 |
-
|
284 |
-
|
285 |
-
|
286 |
-
|
287 |
-
|
288 |
|
289 |
accept_btn.click(
|
290 |
update_app,
|
291 |
inputs=[accept_btn, data_gr, current_index, history, username],
|
292 |
-
outputs=[
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
293 |
)
|
294 |
myabe_btn.click(
|
295 |
update_app,
|
296 |
inputs=[myabe_btn, data_gr, current_index, history, username],
|
297 |
-
outputs=[
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
298 |
)
|
299 |
|
300 |
reject_btn.click(
|
301 |
update_app,
|
302 |
inputs=[reject_btn, data_gr, current_index, history, username],
|
303 |
-
outputs=[
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
304 |
)
|
305 |
|
306 |
demo.launch()
|
|
|
|
|
1 |
import csv
|
2 |
import json
|
3 |
import os
|
|
|
19 |
from huggingface_hub import HfApi, login, snapshot_download
|
20 |
from PIL import Image
|
21 |
|
22 |
+
# session_token = os.environ.get("SessionToken")
|
23 |
+
# login(token=session_token)
|
|
|
24 |
|
25 |
|
26 |
csv.field_size_limit(sys.maxsize)
|
27 |
|
28 |
np.random.seed(int(time.time()))
|
29 |
|
30 |
+
with open("./imagenet_hard_nearest_indices.pkl", "rb") as f:
|
31 |
+
knn_results = pickle.load(f)
|
32 |
|
33 |
with open("imagenet-labels.json") as f:
|
34 |
wnid_to_label = json.load(f)
|
35 |
|
36 |
+
with open("id_to_label.json", "r") as f:
|
37 |
id_to_labels = json.load(f)
|
38 |
|
39 |
+
imagenet_training_samples_path = "imagenet_traning_samples"
|
40 |
|
41 |
+
bad_items = open("./ex2.txt", "r").read().split("\n")
|
42 |
+
bad_items = [x.split(".")[0] for x in bad_items]
|
43 |
+
bad_items = [int(x) for x in bad_items if x != ""]
|
44 |
+
|
45 |
+
NUMBER_OF_IMAGES = 100 # len(bad_items)
|
46 |
|
47 |
# download and extract folders
|
48 |
|
|
|
55 |
|
56 |
# EXTRACT if needed
|
57 |
|
58 |
+
if not os.path.exists("./imagenet_traning_samples") or not os.path.exists(
|
59 |
+
"./knn_cache_for_imagenet_hard"
|
60 |
+
):
|
61 |
torchvision.datasets.utils.extract_archive(
|
62 |
from_path="data.zip",
|
63 |
to_path="./",
|
|
|
67 |
imagenet_hard = datasets.load_dataset("taesiri/imagenet-hard", split="validation")
|
68 |
|
69 |
|
70 |
+
def update_snapshot(username):
|
71 |
output_dir = snapshot_download(
|
72 |
+
repo_id="taesiri/imagenet_hard_review_data",
|
73 |
+
allow_patterns="*.json",
|
74 |
+
repo_type="dataset",
|
75 |
)
|
|
|
76 |
files = glob(f"{output_dir}/*.json")
|
77 |
|
78 |
df = pd.DataFrame()
|
|
|
87 |
rows.append(tdf)
|
88 |
|
89 |
df = pd.DataFrame(rows, columns=columns)
|
90 |
+
df = df[df["user_id"] == username]
|
91 |
|
92 |
+
return df
|
|
|
93 |
|
|
|
94 |
|
95 |
+
def generate_dataset(username):
|
96 |
+
global NUMBER_OF_IMAGES
|
97 |
+
df = update_snapshot(username)
|
98 |
|
99 |
+
all_images = set(bad_items)
|
100 |
+
answered = set(df.id)
|
101 |
+
remaining = list(all_images - answered)
|
|
|
102 |
|
103 |
+
if len(remaining) < NUMBER_OF_IMAGES and len(remaining) > 0:
|
104 |
+
NUMBER_OF_IMAGES = len(remaining)
|
105 |
+
random_indices = list(remaining)
|
106 |
+
elif len(remaining) == 0:
|
107 |
+
return []
|
108 |
+
else:
|
109 |
+
random_indices = np.random.choice(remaining, NUMBER_OF_IMAGES, replace=False)
|
110 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
111 |
random_images = [imagenet_hard[int(i)]["image"] for i in random_indices]
|
112 |
random_gt_ids = [imagenet_hard[int(i)]["label"] for i in random_indices]
|
113 |
+
random_gt_labels = [imagenet_hard[int(x)]["english_label"] for x in random_indices]
|
114 |
|
115 |
data = []
|
116 |
for i, image in enumerate(random_images):
|
117 |
data.append(
|
118 |
+
{
|
119 |
"id": random_indices[i],
|
120 |
"image": image,
|
121 |
"correct_label": random_gt_labels[i],
|
|
|
125 |
return data
|
126 |
|
127 |
|
|
|
128 |
def string_to_image(text):
|
129 |
+
text = text.replace("_", " ").lower().replace(", ", "\n")
|
130 |
# Create a blank white square image
|
131 |
img = np.ones((220, 75, 3))
|
132 |
|
|
|
133 |
fig, ax = plt.subplots(figsize=(6, 2.25))
|
|
|
|
|
134 |
ax.imshow(img, extent=[0, 1, 0, 1])
|
135 |
+
ax.text(0.5, 0.75, text, fontsize=18, ha="center", va="center")
|
|
|
|
|
|
|
|
|
136 |
ax.set_xticks([])
|
137 |
ax.set_yticks([])
|
138 |
ax.set_xticklabels([])
|
139 |
ax.set_yticklabels([])
|
|
|
|
|
140 |
for spine in ax.spines.values():
|
141 |
spine.set_visible(False)
|
142 |
|
|
|
143 |
return fig
|
144 |
|
145 |
|
146 |
+
all_samples = glob("./imagenet_traning_samples/*.JPEG")
|
147 |
+
qid_to_sample = {
|
148 |
+
int(x.split("/")[-1].split(".")[0].split("_")[0]): x for x in all_samples
|
149 |
+
}
|
150 |
|
151 |
+
# user-e3z5b
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
152 |
|
|
|
|
|
153 |
|
154 |
def get_training_samples(qid):
|
155 |
+
labels_id = imagenet_hard[int(qid)]["label"]
|
156 |
samples = [qid_to_sample[x] for x in labels_id]
|
157 |
return samples
|
158 |
|
159 |
|
|
|
|
|
|
|
160 |
def load_sample(data, current_index):
|
161 |
image_id = data[current_index]["id"]
|
162 |
qimage = data[current_index]["image"]
|
163 |
+
|
164 |
labels = data[current_index]["correct_label"]
|
165 |
return qimage, labels
|
166 |
+
|
167 |
+
|
168 |
+
def preprocessing(data, current_index, history, username):
|
169 |
+
data = generate_dataset(username)
|
170 |
+
|
171 |
+
if len(data) == 0:
|
172 |
+
fake_plot = string_to_image("No more images to review")
|
173 |
+
empty_image = Image.new("RGB", (224, 224))
|
174 |
+
return (
|
175 |
+
empty_image,
|
176 |
+
fake_plot,
|
177 |
+
current_index,
|
178 |
+
history,
|
179 |
+
data,
|
180 |
+
None,
|
181 |
+
)
|
182 |
+
|
183 |
+
current_index = 0
|
184 |
+
qimage, labels = load_sample(data, current_index)
|
185 |
+
image_id = data[current_index]["id"]
|
186 |
+
training_samples_image = get_training_samples(image_id)
|
187 |
+
training_samples_image = [
|
188 |
+
Image.open(x).convert("RGB") for x in training_samples_image
|
189 |
+
]
|
190 |
+
|
191 |
+
# labels is a list of labels, conver it to a string
|
192 |
+
labels = ", ".join(labels)
|
193 |
+
label_plot = string_to_image(labels)
|
194 |
+
|
195 |
+
return qimage, label_plot, current_index, history, data, training_samples_image
|
196 |
|
197 |
|
198 |
def update_app(decision, data, current_index, history, username):
|
199 |
+
global NUMBER_OF_IMAGES
|
200 |
if current_index == -1:
|
201 |
+
return
|
202 |
+
|
203 |
+
if current_index == NUMBER_OF_IMAGES - 1:
|
204 |
time_stamp = int(time.time())
|
205 |
|
206 |
image_id = data[current_index]["id"]
|
207 |
+
# convert to percentage
|
208 |
dicision_dict = {
|
209 |
"id": int(image_id),
|
210 |
"user_id": username,
|
|
|
228 |
|
229 |
os.remove(temp_filename)
|
230 |
|
231 |
+
fake_plot = string_to_image("Thank you for your time!")
|
232 |
+
empty_image = Image.new("RGB", (224, 224))
|
233 |
+
return empty_image, fake_plot, current_index, history, data, None
|
234 |
|
235 |
+
if current_index >= 0 and current_index < NUMBER_OF_IMAGES - 1:
|
236 |
+
time_stamp = int(time.time())
|
|
|
|
|
|
|
237 |
|
238 |
+
image_id = data[current_index]["id"]
|
239 |
+
# convert to percentage
|
240 |
+
dicision_dict = {
|
241 |
+
"id": int(image_id),
|
242 |
+
"user_id": username,
|
243 |
+
"time": time_stamp,
|
244 |
+
"decision": decision,
|
245 |
+
}
|
246 |
|
247 |
+
# upload the decision to the server
|
248 |
+
temp_filename = f"results_{username}_{time_stamp}.json"
|
249 |
+
# convert decision_dict to json and save it on the disk
|
250 |
+
with open(temp_filename, "w") as f:
|
251 |
+
json.dump(dicision_dict, f)
|
252 |
|
253 |
+
api = HfApi()
|
254 |
+
api.upload_file(
|
255 |
+
path_or_fileobj=temp_filename,
|
256 |
+
path_in_repo=temp_filename,
|
257 |
+
repo_id="taesiri/imagenet_hard_review_data",
|
258 |
+
repo_type="dataset",
|
259 |
+
)
|
260 |
+
|
261 |
+
os.remove(temp_filename)
|
262 |
+
|
263 |
+
# Load the Next Image
|
264 |
|
265 |
+
current_index += 1
|
266 |
+
qimage, labels = load_sample(data, current_index)
|
267 |
+
image_id = data[current_index]["id"]
|
268 |
+
training_samples_image = get_training_samples(image_id)
|
269 |
+
training_samples_image = [
|
270 |
+
Image.open(x).convert("RGB") for x in training_samples_image
|
271 |
+
]
|
272 |
+
|
273 |
+
# labels is a list of labels, conver it to a string
|
274 |
+
labels = ", ".join(labels)
|
275 |
+
label_plot = string_to_image(labels)
|
276 |
+
|
277 |
+
return qimage, label_plot, current_index, history, data, training_samples_image
|
278 |
+
|
279 |
+
|
280 |
+
newcss = """
|
281 |
#query_image{
|
282 |
height: auto !important;
|
283 |
}
|
|
|
289 |
#sample_gallery {
|
290 |
height: auto !important;
|
291 |
}
|
292 |
+
"""
|
293 |
|
294 |
with gr.Blocks(css=newcss) as demo:
|
295 |
data_gr = gr.State({})
|
296 |
current_index = gr.State(-1)
|
297 |
history = gr.State({})
|
298 |
+
|
299 |
gr.Markdown("# Cleaning ImageNet-Hard!")
|
300 |
|
301 |
random_str = "".join(
|
302 |
random.choice(string.ascii_lowercase + string.digits) for _ in range(5)
|
303 |
)
|
304 |
|
305 |
+
with gr.Row():
|
306 |
+
username = gr.Textbox(label="Username", value=f"user-{random_str}")
|
307 |
+
prepare_btn = gr.Button(value="Load Samples")
|
308 |
|
309 |
with gr.Column():
|
310 |
with gr.Row():
|
311 |
accept_btn = gr.Button(value="Accept")
|
312 |
myabe_btn = gr.Button(value="Not Sure!")
|
313 |
reject_btn = gr.Button(value="Reject")
|
314 |
+
with gr.Row():
|
315 |
query_image = gr.Image(type="pil", label="Query", elem_id="query_image")
|
316 |
with gr.Column():
|
317 |
+
label_plot = gr.Plot(
|
318 |
+
label="Is this a correct label for this image?", type="fig"
|
319 |
+
)
|
320 |
+
training_samples = gr.Gallery(
|
321 |
+
type="pil", label="Training samples", elem_id="sample_gallery"
|
322 |
+
)
|
323 |
|
324 |
accept_btn.click(
|
325 |
update_app,
|
326 |
inputs=[accept_btn, data_gr, current_index, history, username],
|
327 |
+
outputs=[
|
328 |
+
query_image,
|
329 |
+
label_plot,
|
330 |
+
current_index,
|
331 |
+
history,
|
332 |
+
data_gr,
|
333 |
+
training_samples,
|
334 |
+
],
|
335 |
)
|
336 |
myabe_btn.click(
|
337 |
update_app,
|
338 |
inputs=[myabe_btn, data_gr, current_index, history, username],
|
339 |
+
outputs=[
|
340 |
+
query_image,
|
341 |
+
label_plot,
|
342 |
+
current_index,
|
343 |
+
history,
|
344 |
+
data_gr,
|
345 |
+
training_samples,
|
346 |
+
],
|
347 |
)
|
348 |
|
349 |
reject_btn.click(
|
350 |
update_app,
|
351 |
inputs=[reject_btn, data_gr, current_index, history, username],
|
352 |
+
outputs=[
|
353 |
+
query_image,
|
354 |
+
label_plot,
|
355 |
+
current_index,
|
356 |
+
history,
|
357 |
+
data_gr,
|
358 |
+
training_samples,
|
359 |
+
],
|
360 |
+
)
|
361 |
+
|
362 |
+
prepare_btn.click(
|
363 |
+
preprocessing,
|
364 |
+
inputs=[data_gr, current_index, history, username],
|
365 |
+
outputs=[
|
366 |
+
query_image,
|
367 |
+
label_plot,
|
368 |
+
current_index,
|
369 |
+
history,
|
370 |
+
data_gr,
|
371 |
+
training_samples,
|
372 |
+
],
|
373 |
)
|
374 |
|
375 |
demo.launch()
|