Hugging Face's logo

Spaces:
Marqo
/
Ecommerce-Embedding-Benchmarks
Running

App Files Community
Ecommerce-Embedding-Benchmarks
File size: 9,730 Bytes 
8504de8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
 
import gradio as gr
import pandas as pd

# ---------------      100k-Marqo-Ecommerce-Easy     -------------------

# Embedding Models for 100k-Marqo-Ecommerce-Easy
embedding_models = [
    '<a href="https://docs.aws.amazon.com/bedrock/latest/userguide/titan-multiemb-models.html">Amazon-Titan-MultiModal</a>', 
    '<a href="https://huggingface.co/jinaai/jina-clip-v1">Jina-V1-CLIP</a>', 
    '<a href="https://cohere.com/blog/introducing-embed-v3">Cohere-embedding-v3</a>', 
    '<a href="https://cloud.google.com/vertex-ai">GCP-Vertex</a>', 
    '<a href="https://huggingface.co/timm/ViT-SO400M-14-SigLIP">ViT-SO400M-14-SigLip</a>', 
    '<a href="https://huggingface.co/timm/ViT-B-16-SigLIP">ViT-B-16-SigLip</a>', 
    '<a href="https://huggingface.co/timm/ViT-L-16-SigLIP-384">ViT-L-16-SigLip</a>', 
    '<a href="https://huggingface.co/Marqo/marqo-ecommerce-embeddings-B">Marqo-Ecommerce-B</a>', 
    '<a href="https://huggingface.co/Marqo/marqo-ecommerce-embeddings-L">Marqo-Ecommerce-L</a>'
]

# GS-Text2Image-1m
gs_text2_image_1m_mAP = [0.694, 0.480, 0.358, 0.740, 0.792, 0.701, 0.754, 0.842, 0.879]
gs_text2_image_1m_Recall = [0.868, 0.638, 0.515, 0.910, 0.935, 0.87, 0.907, 0.961, 0.971]
gs_text2_image_1m_MRR = [0.693, 0.480, 0.358, 0.740, 0.792, 0.701, 0.754, 0.842, 0.879]
gs_text2_image_1m_ndcg = [0.733, 0.511, 0.389, 0.779, 0.825, 0.739, 0.789, 0.871, 0.901]

gs_text2_image_1m_data = {
    "Embedding Model": embedding_models,
    "mAP": gs_text2_image_1m_mAP,
    "R@10": gs_text2_image_1m_Recall,
    "MRR": gs_text2_image_1m_MRR,
    "nDCG@10": gs_text2_image_1m_ndcg
}
gs_text2_image_1m_df = pd.DataFrame(gs_text2_image_1m_data)
gs_text2_image_1m_df = pd.DataFrame(gs_text2_image_1m_df).sort_values(by="mAP", ascending=False).reset_index(drop=True)

# GS-Category2Image-1m
gs_category2_image_1m_mAP = [0.308, 0.175, 0.136, 0.417, 0.423, 0.347, 0.392, 0.479, 0.515]
gs_category2_image_1m_Precision = [0.231, 0.122, 0.110, 0.298, 0.302, 0.252, 0.281, 0.336, 0.358]
gs_category2_image_1m_MRR = [0.558, 0.369, 0.315, 0.636, 0.644, 0.594, 0.627, 0.744, 0.764]
gs_category2_image_1m_ndcg = [0.377, 0.229, 0.178, 0.481, 0.487, 0.414, 0.458, 0.558, 0.590]

gs_category2_image_1m_data = {
    "Embedding Model": embedding_models,
    "mAP": gs_category2_image_1m_mAP,
    "P@10": gs_category2_image_1m_Precision,
    "MRR": gs_category2_image_1m_MRR,
    "nDCG@10": gs_category2_image_1m_ndcg
}
gs_category2_image_1m_df = pd.DataFrame(gs_category2_image_1m_data)
gs_category2_image_1m_df = pd.DataFrame(gs_category2_image_1m_df).sort_values(by="mAP", ascending=False).reset_index(drop=True)

# AP-Text2Image-3m
ap_text2_image_3m_mAP = [0.762, 0.530, 0.433, 0.808, 0.860, 0.797, 0.842, 0.897, 0.928]
ap_text2_image_3m_Recall = [0.889, 0.699, 0.597, 0.933, 0.954, 0.917, 0.940, 0.967, 0.978]
ap_text2_image_3m_MRR = [0.763, 0.530, 0.433, 0.808, 0.860, 0.797, 0.842, 0.897, 0.928]
ap_text2_image_3m_ndcg = [0.791, 0.565, 0.465, 0.837, 0.882, 0.825, 0.865, 0.914, 0.940]

ap_text2_image_3m_data = {
    "Embedding Model": embedding_models,
    "mAP": ap_text2_image_3m_mAP,
    "R@10": ap_text2_image_3m_Recall,
    "MRR": ap_text2_image_3m_MRR,
    "nDCG@10": ap_text2_image_3m_ndcg
}
ap_text2_image_3m_df = pd.DataFrame(ap_text2_image_3m_data)
ap_text2_image_3m_df = pd.DataFrame(ap_text2_image_3m_df).sort_values(by="mAP", ascending=False).reset_index(drop=True)


# ---------------       Marqo-Ecommerce-Hard      -------------------

# Embedding Models for Marqo-Ecommerce-Hard
hard_embedding_models = [
    '<a href="https://docs.aws.amazon.com/bedrock/latest/userguide/titan-multiemb-models.html">Amazon-Titan-MultiModal</a>', 
    '<a href="https://huggingface.co/jinaai/jina-clip-v1">Jina-V1-CLIP</a>', 
    '<a href="https://huggingface.co/timm/ViT-SO400M-14-SigLIP">ViT-SO400M-14-SigLip</a>', 
    '<a href="https://huggingface.co/timm/ViT-B-16-SigLIP">ViT-B-16-SigLip</a>', 
    '<a href="https://huggingface.co/timm/ViT-L-16-SigLIP-384">ViT-L-16-SigLip</a>', 
    '<a href="https://huggingface.co/Marqo/marqo-ecommerce-embeddings-B">Marqo-Ecommerce-B</a>', 
    '<a href="https://huggingface.co/Marqo/marqo-ecommerce-embeddings-L">Marqo-Ecommerce-L</a>'
]

# GS-Text2Image-1m
hard_gs_text2_image_1m_mAP = [0.475, 0.285, 0.573, 0.476, 0.540, 0.623, 0.682]
hard_gs_text2_image_1m_Recall = [0.648, 0.402, 0.763, 0.660, 0.722, 0.832, 0.878]
hard_gs_text2_image_1m_MRR = [0.475, 0.285, 0.574, 0.477, 0.540, 0.624, 0.683]
hard_gs_text2_image_1m_ndcg = [0.509, 0.306, 0.613, 0.513, 0.577, 0.668, 0.726]

hard_gs_text2_image_1m_data = {
    "Embedding Model": hard_embedding_models,
    "mAP": hard_gs_text2_image_1m_mAP,
    "R@10": hard_gs_text2_image_1m_Recall,
    "MRR": hard_gs_text2_image_1m_MRR,
    "nDCG@10": hard_gs_text2_image_1m_ndcg
}
hard_gs_text2_image_1m_df = pd.DataFrame(hard_gs_text2_image_1m_data)
hard_gs_text2_image_1m_df = pd.DataFrame(hard_gs_text2_image_1m_df).sort_values(by="mAP", ascending=False).reset_index(drop=True)

# GS-Category2Image-1m
hard_gs_category2_image_1m_mAP = [0.246, 0.123, 0.352, 0.277, 0.324, 0.423, 0.463]
hard_gs_category2_image_1m_Precision = [0.429, 0.275, 0.516, 0.458, 0.497, 0.629, 0.652]
hard_gs_category2_image_1m_MRR = [0.642, 0.504, 0.707, 0.660, 0.687, 0.810, 0.822]
hard_gs_category2_image_1m_ndcg = [0.446, 0.294, 0.529, 0.473, 0.509, 0.644, 0.666]

hard_gs_category2_image_1m_data = {
    "Embedding Model": hard_embedding_models,
    "mAP": hard_gs_category2_image_1m_mAP,
    "P@10": hard_gs_category2_image_1m_Precision,
    "MRR": hard_gs_category2_image_1m_MRR,
    "nDCG@10": hard_gs_category2_image_1m_ndcg
}
hard_gs_category2_image_1m_df = pd.DataFrame(hard_gs_category2_image_1m_data)
hard_gs_category2_image_1m_df = pd.DataFrame(hard_gs_category2_image_1m_df).sort_values(by="mAP", ascending=False).reset_index(drop=True)

# AP-Text2Image-3m
hard_ap_text2_image_3m_mAP = [0.456, 0.265, 0.560, 0.480, 0.544, 0.592, 0.658]
hard_ap_text2_image_3m_Recall = [0.627, 0.378, 0.742, 0.650, 0.715, 0.795, 0.854]
hard_ap_text2_image_3m_MRR = [0.457, 0.266, 0.564, 0.484, 0.548, 0.597, 0.663]
hard_ap_text2_image_3m_ndcg = [0.491, 0.285, 0.599, 0.515, 0.580, 0.637, 0.703]

hard_ap_text2_image_3m_data = {
    "Embedding Model": hard_embedding_models,
    "mAP": hard_ap_text2_image_3m_mAP,
    "R@10": hard_ap_text2_image_3m_Recall,
    "MRR": hard_ap_text2_image_3m_MRR,
    "nDCG@10": hard_ap_text2_image_3m_ndcg
}
hard_ap_text2_image_3m_df = pd.DataFrame(hard_ap_text2_image_3m_data)
hard_ap_text2_image_3m_df = pd.DataFrame(hard_ap_text2_image_3m_df).sort_values(by="mAP", ascending=False).reset_index(drop=True)


# Gradio interface with sortable Dataframe
with gr.Blocks(css="""
    .gradio-container {
        display: flex;
        justify-content: center;
        align-items: center;
        min-height: 100vh;
        flex-direction: column;
    }
""") as demo:
    gr.Markdown("# Ecommerce Embedding Model Benchmarks")
    
    gr.Markdown("This Space contains benchmark results conducted as part of the release of our ecommerce embedding models: [**`Marqo-Ecommerce-L`**](https://huggingface.co/Marqo/marqo-ecommerce-embeddings-L) and [**`Marqo-Ecommerce-B`**](https://huggingface.co/Marqo/marqo-ecommerce-embeddings-B). ")
    gr.Markdown('The benchmarks are separated into \'Marqo-Ecommerce-Hard\' and \'100k-Marqo-Ecommerce-Easy\'. The "easy" dataset is about 10-30 times smaller, and  designed to accommodate rate-limited models, specifically Cohere-Embeddings-v3 and GCP-Vertex. The "hard"  dataset represents the true challenge, since it contains four million ecommerce product listings, which pushes these models to their limits in a real-world, ecommerce scenario. ' )
    gr.Markdown('Within both these scenarios, the models were benchmarked against three different tasks:')
    gr.Markdown('- **Google Shopping Text-to-Image**')
    gr.Markdown('- **Google Shopping Category-to-Image**')
    gr.Markdown('- **Amazon Products Text-to-Image**')
    gr.Markdown('As part of this launch, we also released two evaluation datasets: [`Marqo/google-shopping-general-eval`](https://huggingface.co/datasets/Marqo/google-shopping-general-eval) and [`Marqo/amazon-products-eval`](https://huggingface.co/datasets/Marqo/amazon-products-eval).')
    gr.Markdown('For more information on these models, benchmark results, and how you can run these evaluations yourself, visit our [blog post](https://www.marqo.ai/blog).')

    # Hard
    gr.Markdown('## Marqo-Ecommerce-Hard')
    gr.Markdown('### Google Shopping Text to Image 1m')
    gr.Dataframe(value=hard_gs_text2_image_1m_df, headers="keys", interactive=True, datatype=["html", "number", "number", "number", "number"])

    gr.Markdown('### Google Shopping Category to Image 1m')
    gr.Dataframe(value=hard_gs_category2_image_1m_df, headers="keys", interactive=True, datatype=["html", "number", "number", "number", "number"])

    gr.Markdown('###  Amazon Products Text to Image 3m')
    gr.Dataframe(value=hard_ap_text2_image_3m_df, headers="keys", interactive=True, datatype=["html", "number", "number", "number", "number"])
    
    # Easy
    gr.Markdown('## 100k-Marqo-Ecommerce-Easy')
    gr.Markdown('### Google Shopping Text to Image')
    gr.Dataframe(value=gs_text2_image_1m_df, headers="keys", interactive=True, datatype=["html", "number", "number", "number", "number"])

    gr.Markdown('### Google Shopping Category to Image')
    gr.Dataframe(value=gs_category2_image_1m_df, headers="keys", interactive=True, datatype=["html", "number", "number", "number", "number"])

    gr.Markdown('### Amazon Products Text to Image')
    gr.Dataframe(value=ap_text2_image_3m_df, headers="keys", interactive=True, datatype=["html", "number", "number", "number", "number"])

demo.launch()