Spaces:
Running
Running
Category filter (#3)
Browse files- Filter categories (2d04ce888baa749cea99eebb988103adba62a52a)
app.py
CHANGED
|
@@ -12,6 +12,9 @@ ds_results = load_dataset("visionLMsftw/vibe-testing-results", split="train")
|
|
| 12 |
models = get_model_names(ds_results)
|
| 13 |
responses = get_responses(ds_results)
|
| 14 |
|
|
|
|
|
|
|
|
|
|
| 15 |
|
| 16 |
with gr.Blocks(theme=gr.themes.Soft()) as demo:
|
| 17 |
gr.Markdown("# VLMVibeEval")
|
|
@@ -20,27 +23,48 @@ with gr.Blocks(theme=gr.themes.Soft()) as demo:
|
|
| 20 |
|
| 21 |
with gr.Column(visible=True) as model_mode:
|
| 22 |
selected_model = gr.Dropdown(models, label="Choose model")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 23 |
model_output = gr.HTML()
|
| 24 |
current_index = gr.State(value=0)
|
| 25 |
current_html = gr.State(value="")
|
| 26 |
|
| 27 |
-
def load_initial(model):
|
| 28 |
-
html = display_model_responses_html(evaluation_data, responses, model, 0)
|
| 29 |
return html, 5, html
|
| 30 |
-
|
| 31 |
-
def load_more(model, index, html_so_far):
|
| 32 |
-
new_html = display_model_responses_html(evaluation_data, responses, model, index)
|
| 33 |
updated_html = html_so_far + new_html
|
| 34 |
return updated_html, index + 5, updated_html
|
| 35 |
|
| 36 |
-
selected_model.change(
|
| 37 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 38 |
|
| 39 |
-
|
| 40 |
-
|
|
|
|
|
|
|
|
|
|
| 41 |
|
| 42 |
-
|
| 43 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 44 |
|
| 45 |
with gr.Column(visible=False) as example_mode:
|
| 46 |
category = gr.Dropdown(
|
|
@@ -83,14 +107,12 @@ with gr.Blocks(theme=gr.themes.Soft()) as demo:
|
|
| 83 |
align-items: center;
|
| 84 |
justify-content: center;
|
| 85 |
}
|
| 86 |
-
|
| 87 |
#image-modal img {
|
| 88 |
max-width: 90%;
|
| 89 |
max-height: 90%;
|
| 90 |
border-radius: 8px;
|
| 91 |
box-shadow: 0 0 20px rgba(255,255,255,0.3);
|
| 92 |
}
|
| 93 |
-
|
| 94 |
#image-modal .close {
|
| 95 |
position: absolute;
|
| 96 |
top: 20px; right: 30px;
|
|
@@ -100,12 +122,10 @@ with gr.Blocks(theme=gr.themes.Soft()) as demo:
|
|
| 100 |
font-weight: bold;
|
| 101 |
}
|
| 102 |
</style>
|
| 103 |
-
|
| 104 |
<div id="image-modal" onclick="closeModal(event)">
|
| 105 |
<span class="close" onclick="closeModal(event)">×</span>
|
| 106 |
<img id="modal-img" src="" alt="Enlarged Image" />
|
| 107 |
</div>
|
| 108 |
-
|
| 109 |
<script>
|
| 110 |
function openImage(src) {
|
| 111 |
const modal = document.getElementById('image-modal');
|
|
@@ -113,13 +133,11 @@ with gr.Blocks(theme=gr.themes.Soft()) as demo:
|
|
| 113 |
img.src = src;
|
| 114 |
modal.style.display = 'flex';
|
| 115 |
}
|
| 116 |
-
|
| 117 |
function closeModal(event) {
|
| 118 |
if (event.target.id === 'image-modal' || event.target.classList.contains('close')) {
|
| 119 |
document.getElementById('image-modal').style.display = 'none';
|
| 120 |
}
|
| 121 |
}
|
| 122 |
-
|
| 123 |
// Optional: close on ESC key
|
| 124 |
document.addEventListener('keydown', function(e) {
|
| 125 |
if (e.key === "Escape") {
|
|
|
|
| 12 |
models = get_model_names(ds_results)
|
| 13 |
responses = get_responses(ds_results)
|
| 14 |
|
| 15 |
+
default_category = evaluation_data[0]["category"]
|
| 16 |
+
default_example_id = evaluation_data[0]["id"]
|
| 17 |
+
|
| 18 |
|
| 19 |
with gr.Blocks(theme=gr.themes.Soft()) as demo:
|
| 20 |
gr.Markdown("# VLMVibeEval")
|
|
|
|
| 23 |
|
| 24 |
with gr.Column(visible=True) as model_mode:
|
| 25 |
selected_model = gr.Dropdown(models, label="Choose model")
|
| 26 |
+
model_category = gr.Dropdown(
|
| 27 |
+
choices=list(set(ex["category"] for ex in evaluation_data)),
|
| 28 |
+
label="Category",
|
| 29 |
+
value=default_category
|
| 30 |
+
)
|
| 31 |
model_output = gr.HTML()
|
| 32 |
current_index = gr.State(value=0)
|
| 33 |
current_html = gr.State(value="")
|
| 34 |
|
| 35 |
+
def load_initial(model, category):
|
| 36 |
+
html = display_model_responses_html(evaluation_data, responses, model, start_index=0, batch_size=5, category=category)
|
| 37 |
return html, 5, html
|
| 38 |
+
|
| 39 |
+
def load_more(model, index, html_so_far, category):
|
| 40 |
+
new_html = display_model_responses_html(evaluation_data, responses, model, start_index=index, batch_size=5, category=category)
|
| 41 |
updated_html = html_so_far + new_html
|
| 42 |
return updated_html, index + 5, updated_html
|
| 43 |
|
| 44 |
+
selected_model.change(
|
| 45 |
+
load_initial,
|
| 46 |
+
inputs=[selected_model, model_category],
|
| 47 |
+
outputs=[model_output, current_index, current_html]
|
| 48 |
+
)
|
| 49 |
+
model_category.change(
|
| 50 |
+
load_initial,
|
| 51 |
+
inputs=[selected_model, model_category],
|
| 52 |
+
outputs=[model_output, current_index, current_html]
|
| 53 |
+
)
|
| 54 |
|
| 55 |
+
demo.load(
|
| 56 |
+
load_initial,
|
| 57 |
+
inputs=[selected_model, model_category],
|
| 58 |
+
outputs=[model_output, current_index, current_html]
|
| 59 |
+
)
|
| 60 |
|
| 61 |
+
|
| 62 |
+
more_button = gr.Button("Load more")
|
| 63 |
+
more_button.click(
|
| 64 |
+
load_more,
|
| 65 |
+
inputs=[selected_model, current_index, current_html, model_category],
|
| 66 |
+
outputs=[model_output, current_index, current_html]
|
| 67 |
+
)
|
| 68 |
|
| 69 |
with gr.Column(visible=False) as example_mode:
|
| 70 |
category = gr.Dropdown(
|
|
|
|
| 107 |
align-items: center;
|
| 108 |
justify-content: center;
|
| 109 |
}
|
|
|
|
| 110 |
#image-modal img {
|
| 111 |
max-width: 90%;
|
| 112 |
max-height: 90%;
|
| 113 |
border-radius: 8px;
|
| 114 |
box-shadow: 0 0 20px rgba(255,255,255,0.3);
|
| 115 |
}
|
|
|
|
| 116 |
#image-modal .close {
|
| 117 |
position: absolute;
|
| 118 |
top: 20px; right: 30px;
|
|
|
|
| 122 |
font-weight: bold;
|
| 123 |
}
|
| 124 |
</style>
|
|
|
|
| 125 |
<div id="image-modal" onclick="closeModal(event)">
|
| 126 |
<span class="close" onclick="closeModal(event)">×</span>
|
| 127 |
<img id="modal-img" src="" alt="Enlarged Image" />
|
| 128 |
</div>
|
|
|
|
| 129 |
<script>
|
| 130 |
function openImage(src) {
|
| 131 |
const modal = document.getElementById('image-modal');
|
|
|
|
| 133 |
img.src = src;
|
| 134 |
modal.style.display = 'flex';
|
| 135 |
}
|
|
|
|
| 136 |
function closeModal(event) {
|
| 137 |
if (event.target.id === 'image-modal' || event.target.classList.contains('close')) {
|
| 138 |
document.getElementById('image-modal').style.display = 'none';
|
| 139 |
}
|
| 140 |
}
|
|
|
|
| 141 |
// Optional: close on ESC key
|
| 142 |
document.addEventListener('keydown', function(e) {
|
| 143 |
if (e.key === "Escape") {
|
utils.py
CHANGED
|
@@ -12,10 +12,16 @@ def image_to_base64(image):
|
|
| 12 |
def get_examples_by_category(evaluation_data, category):
|
| 13 |
return [ex["id"] for ex in evaluation_data if ex["category"] == category]
|
| 14 |
|
| 15 |
-
def display_model_responses_html(evaluation_data, responses, model, start_index=0, batch_size=5):
|
| 16 |
start_time = time.time()
|
| 17 |
html = ""
|
| 18 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 19 |
image_thumbnail = ex["image_thumbnail"]
|
| 20 |
image_full_url = ex["image_full_url"]
|
| 21 |
img_id = ex["id"]
|
|
@@ -46,6 +52,7 @@ def display_model_responses_html(evaluation_data, responses, model, start_index=
|
|
| 46 |
|
| 47 |
return html
|
| 48 |
|
|
|
|
| 49 |
def display_example_responses_html(evaluation_data, responses, models, example_id):
|
| 50 |
ex = next(e for e in evaluation_data if e["id"] == example_id)
|
| 51 |
image_thumbnail = ex["image_thumbnail"]
|
|
|
|
| 12 |
def get_examples_by_category(evaluation_data, category):
|
| 13 |
return [ex["id"] for ex in evaluation_data if ex["category"] == category]
|
| 14 |
|
| 15 |
+
def display_model_responses_html(evaluation_data, responses, model, start_index=0, batch_size=5, category=None):
|
| 16 |
start_time = time.time()
|
| 17 |
html = ""
|
| 18 |
+
|
| 19 |
+
if category is not None:
|
| 20 |
+
filtered_data = [ex for ex in evaluation_data if ex["category"] == category]
|
| 21 |
+
else:
|
| 22 |
+
filtered_data = evaluation_data
|
| 23 |
+
|
| 24 |
+
for ex in filtered_data[start_index:start_index + batch_size]:
|
| 25 |
image_thumbnail = ex["image_thumbnail"]
|
| 26 |
image_full_url = ex["image_full_url"]
|
| 27 |
img_id = ex["id"]
|
|
|
|
| 52 |
|
| 53 |
return html
|
| 54 |
|
| 55 |
+
|
| 56 |
def display_example_responses_html(evaluation_data, responses, models, example_id):
|
| 57 |
ex = next(e for e in evaluation_data if e["id"] == example_id)
|
| 58 |
image_thumbnail = ex["image_thumbnail"]
|