Spaces:
Build error
Build error
natexcvi
commited on
Commit
•
0edd049
1
Parent(s):
0e92099
Add fecnet
Browse files- .gitignore +5 -1
- Dockerfile +1 -1
- app.py +5 -17
- auth.py +19 -0
- dev-requirements.txt +5 -0
- model/facial_expression_embedding.ipynb +808 -0
- model/fecnet.py +72 -0
- model/fecnet_test.py +20 -0
- model.py → model/model.py +0 -0
- model/openface_model.py +27 -0
- model/openface_model_test.py +14 -0
- requirements.txt +6 -1
- routers/fecnet_router.py +33 -0
- routers/openface_router.py +33 -0
.gitignore
CHANGED
@@ -2,4 +2,8 @@
|
|
2 |
.env
|
3 |
__pycache__
|
4 |
.vscode
|
5 |
-
.cache
|
|
|
|
|
|
|
|
|
|
2 |
.env
|
3 |
__pycache__
|
4 |
.vscode
|
5 |
+
.cache
|
6 |
+
.DS_Store
|
7 |
+
model/images
|
8 |
+
model/data
|
9 |
+
model/training_dataset
|
Dockerfile
CHANGED
@@ -3,7 +3,7 @@ FROM python:3.7
|
|
3 |
WORKDIR /code
|
4 |
|
5 |
COPY ./requirements.txt /code/requirements.txt
|
6 |
-
RUN apt-get update && apt-get install ffmpeg libsm6 libxext6 -y
|
7 |
RUN pip install --no-cache-dir --upgrade -r /code/requirements.txt
|
8 |
|
9 |
RUN useradd -m -u 1000 user
|
|
|
3 |
WORKDIR /code
|
4 |
|
5 |
COPY ./requirements.txt /code/requirements.txt
|
6 |
+
RUN apt-get update && apt-get install ffmpeg libsm6 libxext6 docker-ce docker-ce-cli containerd.io docker-buildx-plugin docker-compose-plugin -y
|
7 |
RUN pip install --no-cache-dir --upgrade -r /code/requirements.txt
|
8 |
|
9 |
RUN useradd -m -u 1000 user
|
app.py
CHANGED
@@ -1,20 +1,18 @@
|
|
1 |
-
import hmac
|
2 |
import os
|
3 |
-
from typing import Union
|
4 |
|
5 |
import numpy as np
|
6 |
-
from fastapi import Depends, FastAPI, File,
|
7 |
-
from fastapi.security import APIKeyQuery
|
8 |
|
9 |
-
from
|
|
|
|
|
10 |
from schema import EmbeddingResponse, SimilarityResponse
|
11 |
|
12 |
app = FastAPI(
|
13 |
title="Facial Expression Embedding Service",
|
14 |
)
|
15 |
|
16 |
-
|
17 |
-
client_token: str = os.getenv("CLIENT_TOKEN", "")
|
18 |
|
19 |
model = Model(
|
20 |
os.getenv("MODEL_REPO_ID", ""),
|
@@ -23,16 +21,6 @@ model = Model(
|
|
23 |
)
|
24 |
|
25 |
|
26 |
-
async def validate_token(
|
27 |
-
token: Union[str, None] = Depends(api_key),
|
28 |
-
):
|
29 |
-
if token is None:
|
30 |
-
raise HTTPException(status.HTTP_401_UNAUTHORIZED, "No token provided")
|
31 |
-
if not hmac.compare_digest(token, client_token):
|
32 |
-
raise HTTPException(status.HTTP_401_UNAUTHORIZED, "Invalid token")
|
33 |
-
return token
|
34 |
-
|
35 |
-
|
36 |
@app.post(
|
37 |
"/embed",
|
38 |
status_code=status.HTTP_200_OK,
|
|
|
|
|
1 |
import os
|
|
|
2 |
|
3 |
import numpy as np
|
4 |
+
from fastapi import Depends, FastAPI, File, Response, UploadFile, status
|
|
|
5 |
|
6 |
+
from auth import validate_token
|
7 |
+
from model.model import Model
|
8 |
+
from routers.fecnet_router import router as fecnet_router
|
9 |
from schema import EmbeddingResponse, SimilarityResponse
|
10 |
|
11 |
app = FastAPI(
|
12 |
title="Facial Expression Embedding Service",
|
13 |
)
|
14 |
|
15 |
+
app.include_router(fecnet_router)
|
|
|
16 |
|
17 |
model = Model(
|
18 |
os.getenv("MODEL_REPO_ID", ""),
|
|
|
21 |
)
|
22 |
|
23 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
24 |
@app.post(
|
25 |
"/embed",
|
26 |
status_code=status.HTTP_200_OK,
|
auth.py
ADDED
@@ -0,0 +1,19 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import hmac
|
2 |
+
import os
|
3 |
+
from typing import Union
|
4 |
+
|
5 |
+
from fastapi import Depends, HTTPException, status
|
6 |
+
from fastapi.security import APIKeyQuery
|
7 |
+
|
8 |
+
api_key = APIKeyQuery(name="token", auto_error=False)
|
9 |
+
client_token: str = os.getenv("CLIENT_TOKEN", "")
|
10 |
+
|
11 |
+
|
12 |
+
async def validate_token(
|
13 |
+
token: Union[str, None] = Depends(api_key),
|
14 |
+
):
|
15 |
+
if token is None:
|
16 |
+
raise HTTPException(status.HTTP_401_UNAUTHORIZED, "No token provided")
|
17 |
+
if not hmac.compare_digest(token, client_token):
|
18 |
+
raise HTTPException(status.HTTP_401_UNAUTHORIZED, "Invalid token")
|
19 |
+
return token
|
dev-requirements.txt
ADDED
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
ipykernel
|
2 |
+
plotly
|
3 |
+
retrying
|
4 |
+
swifter
|
5 |
+
nbformat
|
model/facial_expression_embedding.ipynb
ADDED
@@ -0,0 +1,808 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"cells": [
|
3 |
+
{
|
4 |
+
"attachments": {},
|
5 |
+
"cell_type": "markdown",
|
6 |
+
"metadata": {},
|
7 |
+
"source": [
|
8 |
+
"# Facial Expression Embedding"
|
9 |
+
]
|
10 |
+
},
|
11 |
+
{
|
12 |
+
"cell_type": "code",
|
13 |
+
"execution_count": 1,
|
14 |
+
"metadata": {},
|
15 |
+
"outputs": [
|
16 |
+
{
|
17 |
+
"name": "stderr",
|
18 |
+
"output_type": "stream",
|
19 |
+
"text": [
|
20 |
+
"2023-04-19 12:14:54.399884: I tensorflow/core/platform/cpu_feature_guard.cc:193] This TensorFlow binary is optimized with oneAPI Deep Neural Network Library (oneDNN) to use the following CPU instructions in performance-critical operations: AVX2 AVX512F AVX512_VNNI FMA\n",
|
21 |
+
"To enable them in other operations, rebuild TensorFlow with the appropriate compiler flags.\n"
|
22 |
+
]
|
23 |
+
}
|
24 |
+
],
|
25 |
+
"source": [
|
26 |
+
"import numpy as np\n",
|
27 |
+
"import os\n",
|
28 |
+
"import random\n",
|
29 |
+
"from typing import *\n",
|
30 |
+
"import tensorflow as tf\n",
|
31 |
+
"from pathlib import Path\n",
|
32 |
+
"from tensorflow.keras import applications\n",
|
33 |
+
"from tensorflow.keras import layers\n",
|
34 |
+
"from tensorflow.keras import losses\n",
|
35 |
+
"from tensorflow.keras import optimizers\n",
|
36 |
+
"from tensorflow.keras import metrics\n",
|
37 |
+
"from tensorflow.keras import Model\n",
|
38 |
+
"from tensorflow.keras.applications import resnet\n",
|
39 |
+
"import pandas as pd\n",
|
40 |
+
"import mediapipe as mp\n",
|
41 |
+
"import plotly.express as px\n",
|
42 |
+
"from plotly.subplots import make_subplots\n",
|
43 |
+
"import plotly.graph_objects as go\n",
|
44 |
+
"import requests\n",
|
45 |
+
"from tqdm import tqdm\n",
|
46 |
+
"import base64\n",
|
47 |
+
"from concurrent.futures import ThreadPoolExecutor, as_completed\n",
|
48 |
+
"from retrying import retry\n",
|
49 |
+
"import swifter\n"
|
50 |
+
]
|
51 |
+
},
|
52 |
+
{
|
53 |
+
"attachments": {},
|
54 |
+
"cell_type": "markdown",
|
55 |
+
"metadata": {},
|
56 |
+
"source": [
|
57 |
+
"## Dataset Loading"
|
58 |
+
]
|
59 |
+
},
|
60 |
+
{
|
61 |
+
"cell_type": "code",
|
62 |
+
"execution_count": 9,
|
63 |
+
"metadata": {},
|
64 |
+
"outputs": [],
|
65 |
+
"source": [
|
66 |
+
"IMAGE_DIR = \"images\"\n",
|
67 |
+
"TRAINING_DATASET = \"training_dataset\""
|
68 |
+
]
|
69 |
+
},
|
70 |
+
{
|
71 |
+
"cell_type": "code",
|
72 |
+
"execution_count": 3,
|
73 |
+
"metadata": {},
|
74 |
+
"outputs": [],
|
75 |
+
"source": [
|
76 |
+
"@retry(stop_max_attempt_number=3)\n",
|
77 |
+
"def image_downloader(url: str):\n",
|
78 |
+
" get_name = lambda url: base64.urlsafe_b64encode(url.encode()).decode()\n",
|
79 |
+
" Path(IMAGE_DIR).mkdir(exist_ok=True)\n",
|
80 |
+
" filename = get_name(url)\n",
|
81 |
+
" if os.path.exists(os.path.join(IMAGE_DIR, filename)):\n",
|
82 |
+
" return filename\n",
|
83 |
+
" res = requests.get(url, timeout=10)\n",
|
84 |
+
" if not res.ok:\n",
|
85 |
+
" return None\n",
|
86 |
+
" with open(os.path.join(IMAGE_DIR, filename), \"wb\") as f:\n",
|
87 |
+
" f.write(res.content)\n",
|
88 |
+
" return filename\n"
|
89 |
+
]
|
90 |
+
},
|
91 |
+
{
|
92 |
+
"cell_type": "code",
|
93 |
+
"execution_count": 4,
|
94 |
+
"metadata": {},
|
95 |
+
"outputs": [],
|
96 |
+
"source": [
|
97 |
+
"def get_column_names():\n",
|
98 |
+
" names = []\n",
|
99 |
+
" for i in range(1, 4):\n",
|
100 |
+
" names += [\n",
|
101 |
+
" f\"img{i}_url\",\n",
|
102 |
+
" f\"img{i}_tl_col\",\n",
|
103 |
+
" f\"img{i}_br_col\",\n",
|
104 |
+
" f\"img{i}_tl_row\",\n",
|
105 |
+
" f\"img{i}_br_row\",\n",
|
106 |
+
" ]\n",
|
107 |
+
" names += [\"triplet_type\"]\n",
|
108 |
+
" for i in range(6):\n",
|
109 |
+
" names += [f\"annotator{i+1}_id\", f\"annotation{i+1}\"]\n",
|
110 |
+
" return names\n",
|
111 |
+
"\n",
|
112 |
+
"\n",
|
113 |
+
"def get_local_storage_column_names():\n",
|
114 |
+
" names = []\n",
|
115 |
+
" for i in range(1, 4):\n",
|
116 |
+
" names += [\n",
|
117 |
+
" f\"img{i}_id\",\n",
|
118 |
+
" f\"img{i}_tl_col\",\n",
|
119 |
+
" f\"img{i}_br_col\",\n",
|
120 |
+
" f\"img{i}_tl_row\",\n",
|
121 |
+
" f\"img{i}_br_row\",\n",
|
122 |
+
" ]\n",
|
123 |
+
" names += [\"triplet_type\"]\n",
|
124 |
+
" names += [\"annotator1_id\"]\n",
|
125 |
+
" names += [\"annotation\"]\n",
|
126 |
+
" return names\n",
|
127 |
+
"\n",
|
128 |
+
"\n",
|
129 |
+
"def get_label(annotations: pd.Series):\n",
|
130 |
+
" def mode(x):\n",
|
131 |
+
" s = pd.Series(x)\n",
|
132 |
+
" if s.value_counts(normalize=True).max() < 0.5:\n",
|
133 |
+
" return np.nan\n",
|
134 |
+
" return s.mode().at[0]\n",
|
135 |
+
"\n",
|
136 |
+
" return annotations.swifter.apply(mode)\n",
|
137 |
+
"\n",
|
138 |
+
"\n",
|
139 |
+
"def fecnet_dataset_loader(dataset_csv: str):\n",
|
140 |
+
" if isinstance(dataset_csv, bytes):\n",
|
141 |
+
" dataset_csv = dataset_csv.decode()\n",
|
142 |
+
" df = pd.read_csv(\n",
|
143 |
+
" dataset_csv, header=None, names=get_column_names(), nrows=10000\n",
|
144 |
+
" ) # TODO: remove nrows\n",
|
145 |
+
"\n",
|
146 |
+
" # download images\n",
|
147 |
+
" df[\"img1_url\"] = df[\"img1_url\"].swifter.apply(image_downloader)\n",
|
148 |
+
" df[\"img2_url\"] = df[\"img2_url\"].swifter.apply(image_downloader)\n",
|
149 |
+
" df[\"img3_url\"] = df[\"img3_url\"].swifter.apply(image_downloader)\n",
|
150 |
+
" df.dropna(subset=[\"img1_url\", \"img2_url\", \"img3_url\"], inplace=True)\n",
|
151 |
+
"\n",
|
152 |
+
" # determine label\n",
|
153 |
+
" df[\"label\"] = get_label(\n",
|
154 |
+
" pd.Series(df[[f\"annotation{i}\" for i in range(1, 7)]].values.tolist())\n",
|
155 |
+
" )\n",
|
156 |
+
" df.dropna(subset=[\"label\"], inplace=True)\n",
|
157 |
+
"\n",
|
158 |
+
" samples = {\n",
|
159 |
+
" \"img1\": [],\n",
|
160 |
+
" \"img1_box\": [],\n",
|
161 |
+
" \"img2\": [],\n",
|
162 |
+
" \"img2_box\": [],\n",
|
163 |
+
" \"img3\": [],\n",
|
164 |
+
" \"img3_box\": [],\n",
|
165 |
+
" }\n",
|
166 |
+
"\n",
|
167 |
+
" for _, row in df.iterrows():\n",
|
168 |
+
" img1_idx, img2_idx, img3_idx = 1, 2, 3\n",
|
169 |
+
" if row.label == 1:\n",
|
170 |
+
" img1_idx, img3_idx = img3_idx, img1_idx\n",
|
171 |
+
" elif row.label == 2:\n",
|
172 |
+
" img2_idx, img3_idx = img3_idx, img2_idx\n",
|
173 |
+
" bounding_boxes = (\n",
|
174 |
+
" (\n",
|
175 |
+
" (row[f\"img{img1_idx}_tl_col\"], row[f\"img{img1_idx}_tl_row\"]),\n",
|
176 |
+
" (row[f\"img{img1_idx}_br_col\"], row[f\"img{img1_idx}_br_row\"]),\n",
|
177 |
+
" ),\n",
|
178 |
+
" (\n",
|
179 |
+
" (row[f\"img{img2_idx}_tl_col\"], row[f\"img{img2_idx}_tl_row\"]),\n",
|
180 |
+
" (row[f\"img{img2_idx}_br_col\"], row[f\"img{img2_idx}_br_row\"]),\n",
|
181 |
+
" ),\n",
|
182 |
+
" (\n",
|
183 |
+
" (row[f\"img{img3_idx}_tl_col\"], row[f\"img{img3_idx}_tl_row\"]),\n",
|
184 |
+
" (row[f\"img{img3_idx}_br_col\"], row[f\"img{img3_idx}_br_row\"]),\n",
|
185 |
+
" ),\n",
|
186 |
+
" )\n",
|
187 |
+
" samples[\"img1\"].append(row[f\"img{img1_idx}_url\"])\n",
|
188 |
+
" samples[\"img1_box\"].append(bounding_boxes[0])\n",
|
189 |
+
" samples[\"img2\"].append(row[f\"img{img2_idx}_url\"])\n",
|
190 |
+
" samples[\"img2_box\"].append(bounding_boxes[1])\n",
|
191 |
+
" samples[\"img3\"].append(row[f\"img{img3_idx}_url\"])\n",
|
192 |
+
" samples[\"img3_box\"].append(bounding_boxes[2])\n",
|
193 |
+
" return samples\n"
|
194 |
+
]
|
195 |
+
},
|
196 |
+
{
|
197 |
+
"cell_type": "code",
|
198 |
+
"execution_count": 5,
|
199 |
+
"metadata": {},
|
200 |
+
"outputs": [],
|
201 |
+
"source": [
|
202 |
+
"def extract_landmarks(image):\n",
|
203 |
+
" with mp.solutions.face_mesh.FaceMesh(\n",
|
204 |
+
" static_image_mode=True,\n",
|
205 |
+
" max_num_faces=1,\n",
|
206 |
+
" refine_landmarks=True,\n",
|
207 |
+
" min_detection_confidence=0.5,\n",
|
208 |
+
" ) as face_mesh:\n",
|
209 |
+
" results = face_mesh.process(image.numpy())\n",
|
210 |
+
" if results.multi_face_landmarks:\n",
|
211 |
+
" landmarks = results.multi_face_landmarks[0]\n",
|
212 |
+
" landmarks = np.array(\n",
|
213 |
+
" [[lm.x, lm.y, lm.z] for lm in landmarks.landmark], dtype=np.float32\n",
|
214 |
+
" )\n",
|
215 |
+
" landmarks = landmarks.flatten()\n",
|
216 |
+
" else:\n",
|
217 |
+
" landmarks = np.zeros(478 * 3, dtype=np.float32)\n",
|
218 |
+
" return landmarks\n",
|
219 |
+
"\n",
|
220 |
+
"\n",
|
221 |
+
"def preprocess_image(filename: str, tl: Tuple[float, float], br: Tuple[float, float]):\n",
|
222 |
+
" image_string = tf.io.read_file(tf.strings.join([IMAGE_DIR, \"/\", filename]))\n",
|
223 |
+
" image = tf.image.decode_jpeg(image_string, channels=3)\n",
|
224 |
+
" image = tf.image.convert_image_dtype(image, tf.uint8)\n",
|
225 |
+
"\n",
|
226 |
+
" # crop image\n",
|
227 |
+
" tl = tf.cast(tf.multiply(tl, tf.cast(tf.shape(image)[:2][::-1], tf.float32)), tf.int32)\n",
|
228 |
+
" br = tf.cast(tf.multiply(br, tf.cast(tf.shape(image)[:2][::-1], tf.float32)), tf.int32)\n",
|
229 |
+
" image = tf.image.crop_to_bounding_box(\n",
|
230 |
+
" image, tl[1], tl[0], br[1] - tl[1], br[0] - tl[0]\n",
|
231 |
+
" )\n",
|
232 |
+
"\n",
|
233 |
+
" # extract landmarks using facemesh\n",
|
234 |
+
" return tf.py_function(extract_landmarks, [image], tf.float32)\n",
|
235 |
+
" # return image\n",
|
236 |
+
"\n",
|
237 |
+
"\n",
|
238 |
+
"ImgType = Tuple[str, Tuple[float, float], Tuple[float, float]]\n",
|
239 |
+
"\n",
|
240 |
+
"\n",
|
241 |
+
"def preprocess_triplets(triplet: dict):\n",
|
242 |
+
" anchor: ImgType = (triplet[\"img1\"], triplet[\"img1_box\"][0], triplet[\"img1_box\"][1])\n",
|
243 |
+
" positive: ImgType = (\n",
|
244 |
+
" triplet[\"img2\"],\n",
|
245 |
+
" triplet[\"img2_box\"][0],\n",
|
246 |
+
" triplet[\"img2_box\"][1],\n",
|
247 |
+
" )\n",
|
248 |
+
" negative: ImgType = (\n",
|
249 |
+
" triplet[\"img3\"],\n",
|
250 |
+
" triplet[\"img3_box\"][0],\n",
|
251 |
+
" triplet[\"img3_box\"][1],\n",
|
252 |
+
" )\n",
|
253 |
+
" return (\n",
|
254 |
+
" preprocess_image(*anchor),\n",
|
255 |
+
" preprocess_image(*positive),\n",
|
256 |
+
" preprocess_image(*negative),\n",
|
257 |
+
" )\n"
|
258 |
+
]
|
259 |
+
},
|
260 |
+
{
|
261 |
+
"cell_type": "code",
|
262 |
+
"execution_count": 6,
|
263 |
+
"metadata": {},
|
264 |
+
"outputs": [
|
265 |
+
{
|
266 |
+
"data": {
|
267 |
+
"application/vnd.jupyter.widget-view+json": {
|
268 |
+
"model_id": "dc545e1717044afd9ddc0b32fecd72c3",
|
269 |
+
"version_major": 2,
|
270 |
+
"version_minor": 0
|
271 |
+
},
|
272 |
+
"text/plain": [
|
273 |
+
"Pandas Apply: 0%| | 0/10000 [00:00<?, ?it/s]"
|
274 |
+
]
|
275 |
+
},
|
276 |
+
"metadata": {},
|
277 |
+
"output_type": "display_data"
|
278 |
+
},
|
279 |
+
{
|
280 |
+
"data": {
|
281 |
+
"application/vnd.jupyter.widget-view+json": {
|
282 |
+
"model_id": "0e653babd8104a72a0e4425167b44005",
|
283 |
+
"version_major": 2,
|
284 |
+
"version_minor": 0
|
285 |
+
},
|
286 |
+
"text/plain": [
|
287 |
+
"Pandas Apply: 0%| | 0/10000 [00:00<?, ?it/s]"
|
288 |
+
]
|
289 |
+
},
|
290 |
+
"metadata": {},
|
291 |
+
"output_type": "display_data"
|
292 |
+
},
|
293 |
+
{
|
294 |
+
"data": {
|
295 |
+
"application/vnd.jupyter.widget-view+json": {
|
296 |
+
"model_id": "5fe7ddb8d18c41c599a69b9ca87cd8cd",
|
297 |
+
"version_major": 2,
|
298 |
+
"version_minor": 0
|
299 |
+
},
|
300 |
+
"text/plain": [
|
301 |
+
"Pandas Apply: 0%| | 0/10000 [00:00<?, ?it/s]"
|
302 |
+
]
|
303 |
+
},
|
304 |
+
"metadata": {},
|
305 |
+
"output_type": "display_data"
|
306 |
+
},
|
307 |
+
{
|
308 |
+
"data": {
|
309 |
+
"application/vnd.jupyter.widget-view+json": {
|
310 |
+
"model_id": "38e2dc2d2517465f896d4fe4ca05da75",
|
311 |
+
"version_major": 2,
|
312 |
+
"version_minor": 0
|
313 |
+
},
|
314 |
+
"text/plain": [
|
315 |
+
"Pandas Apply: 0%| | 0/8143 [00:00<?, ?it/s]"
|
316 |
+
]
|
317 |
+
},
|
318 |
+
"metadata": {},
|
319 |
+
"output_type": "display_data"
|
320 |
+
},
|
321 |
+
{
|
322 |
+
"name": "stderr",
|
323 |
+
"output_type": "stream",
|
324 |
+
"text": [
|
325 |
+
"2023-04-19 13:04:21.919175: I tensorflow/core/platform/cpu_feature_guard.cc:193] This TensorFlow binary is optimized with oneAPI Deep Neural Network Library (oneDNN) to use the following CPU instructions in performance-critical operations: AVX2 AVX512F AVX512_VNNI FMA\n",
|
326 |
+
"To enable them in other operations, rebuild TensorFlow with the appropriate compiler flags.\n"
|
327 |
+
]
|
328 |
+
}
|
329 |
+
],
|
330 |
+
"source": [
|
331 |
+
"if not TRAINING_DATASET:\n",
|
332 |
+
" df = fecnet_dataset_loader(\"data/faceexp-comparison-data-train-public.csv\")\n",
|
333 |
+
" image_count = len(df)\n",
|
334 |
+
"\n",
|
335 |
+
" dataset = tf.data.Dataset.from_tensor_slices(\n",
|
336 |
+
" df\n",
|
337 |
+
" )\n",
|
338 |
+
"\n",
|
339 |
+
" dataset = dataset.shuffle(buffer_size=1024)\n",
|
340 |
+
" dataset = dataset.map(preprocess_triplets)\n",
|
341 |
+
"else:\n",
|
342 |
+
" dataset = tf.data.Dataset.load(TRAINING_DATASET)\n",
|
343 |
+
"\n",
|
344 |
+
"train_dataset = dataset.take(round(image_count * 0.8))\n",
|
345 |
+
"val_dataset = dataset.skip(round(image_count * 0.8))\n",
|
346 |
+
"\n",
|
347 |
+
"train_dataset = train_dataset.batch(32, drop_remainder=False)\n",
|
348 |
+
"train_dataset = train_dataset.prefetch(8)\n",
|
349 |
+
"\n",
|
350 |
+
"val_dataset = val_dataset.batch(32, drop_remainder=False)\n",
|
351 |
+
"val_dataset = val_dataset.prefetch(8)"
|
352 |
+
]
|
353 |
+
},
|
354 |
+
{
|
355 |
+
"attachments": {},
|
356 |
+
"cell_type": "markdown",
|
357 |
+
"metadata": {},
|
358 |
+
"source": [
|
359 |
+
"## Data Visualisation"
|
360 |
+
]
|
361 |
+
},
|
362 |
+
{
|
363 |
+
"cell_type": "code",
|
364 |
+
"execution_count": 10,
|
365 |
+
"metadata": {},
|
366 |
+
"outputs": [],
|
367 |
+
"source": [
|
368 |
+
"def visualise_face_mesh(landmarks):\n",
|
369 |
+
" landmarks = landmarks.reshape(-1, 3)\n",
|
370 |
+
" fig = go.Figure(\n",
|
371 |
+
" go.Mesh3d(\n",
|
372 |
+
" x=landmarks[:, 0],\n",
|
373 |
+
" y=landmarks[:, 1],\n",
|
374 |
+
" z=landmarks[:, 2],\n",
|
375 |
+
" color=\"lightpink\",\n",
|
376 |
+
" opacity=0.50,\n",
|
377 |
+
" contour=dict( \n",
|
378 |
+
" color=\"grey\",\n",
|
379 |
+
" width=1,\n",
|
380 |
+
" ),\n",
|
381 |
+
" )\n",
|
382 |
+
" )\n",
|
383 |
+
" fig.show()\n",
|
384 |
+
"\n",
|
385 |
+
"def visualise_face_mesh_triplets(anchor, positive, negative):\n",
|
386 |
+
" fig = make_subplots(\n",
|
387 |
+
" rows=anchor.shape[0],\n",
|
388 |
+
" cols=3,\n",
|
389 |
+
" specs=[[{\"type\": \"surface\"}, {\"type\": \"surface\"}, {\"type\": \"surface\"}]],\n",
|
390 |
+
" )\n",
|
391 |
+
" for i in range(anchor.shape[0]):\n",
|
392 |
+
" fig.add_trace(\n",
|
393 |
+
" go.Mesh3d(\n",
|
394 |
+
" x=anchor[i, :, 0],\n",
|
395 |
+
" y=anchor[i, :, 1],\n",
|
396 |
+
" z=anchor[i, :, 2],\n",
|
397 |
+
" color=\"lightpink\",\n",
|
398 |
+
" opacity=0.50,\n",
|
399 |
+
" ),\n",
|
400 |
+
" row=i + 1,\n",
|
401 |
+
" col=1,\n",
|
402 |
+
" )\n",
|
403 |
+
" fig.add_trace(\n",
|
404 |
+
" go.Mesh3d(\n",
|
405 |
+
" x=positive[i, :, 0],\n",
|
406 |
+
" y=positive[i, :, 1],\n",
|
407 |
+
" z=positive[i, :, 2],\n",
|
408 |
+
" color=\"lightpink\",\n",
|
409 |
+
" opacity=0.50,\n",
|
410 |
+
" ),\n",
|
411 |
+
" row=i + 1,\n",
|
412 |
+
" col=2,\n",
|
413 |
+
" )\n",
|
414 |
+
" fig.add_trace(\n",
|
415 |
+
" go.Mesh3d(\n",
|
416 |
+
" x=negative[i, :, 0],\n",
|
417 |
+
" y=negative[i, :, 1],\n",
|
418 |
+
" z=negative[i, :, 2],\n",
|
419 |
+
" color=\"lightpink\",\n",
|
420 |
+
" opacity=0.50,\n",
|
421 |
+
" ),\n",
|
422 |
+
" row=i + 1,\n",
|
423 |
+
" col=3,\n",
|
424 |
+
" )\n",
|
425 |
+
" fig.show()\n",
|
426 |
+
"\n",
|
427 |
+
"def visualise_image(image):\n",
|
428 |
+
" fig = px.imshow(image)\n",
|
429 |
+
" fig.show()\n",
|
430 |
+
"\n",
|
431 |
+
"\n",
|
432 |
+
"def visualise_triplet(anchor, positive, negative):\n",
|
433 |
+
" pass\n"
|
434 |
+
]
|
435 |
+
},
|
436 |
+
{
|
437 |
+
"cell_type": "code",
|
438 |
+
"execution_count": 160,
|
439 |
+
"metadata": {},
|
440 |
+
"outputs": [],
|
441 |
+
"source": [
|
442 |
+
"# ds_samples = list(dataset.as_numpy_iterator())"
|
443 |
+
]
|
444 |
+
},
|
445 |
+
{
|
446 |
+
"cell_type": "code",
|
447 |
+
"execution_count": 161,
|
448 |
+
"metadata": {},
|
449 |
+
"outputs": [],
|
450 |
+
"source": [
|
451 |
+
"# visualise_face_mesh(ds_samples[30][0])"
|
452 |
+
]
|
453 |
+
},
|
454 |
+
{
|
455 |
+
"attachments": {},
|
456 |
+
"cell_type": "markdown",
|
457 |
+
"metadata": {},
|
458 |
+
"source": [
|
459 |
+
"## Model Definition"
|
460 |
+
]
|
461 |
+
},
|
462 |
+
{
|
463 |
+
"cell_type": "code",
|
464 |
+
"execution_count": 11,
|
465 |
+
"metadata": {},
|
466 |
+
"outputs": [],
|
467 |
+
"source": [
|
468 |
+
"def create_embedding_model():\n",
|
469 |
+
" input_layer = layers.Input(shape=(478, 3))\n",
|
470 |
+
" flatten = layers.Flatten()(input_layer)\n",
|
471 |
+
" dense1 = layers.Dense(512, activation=\"relu\")(flatten)\n",
|
472 |
+
" dense1 = layers.BatchNormalization()(dense1)\n",
|
473 |
+
" dense2 = layers.Dense(256, activation=\"relu\")(dense1)\n",
|
474 |
+
" dense2 = layers.BatchNormalization()(dense2)\n",
|
475 |
+
" output = layers.Dense(16)(dense2)\n",
|
476 |
+
"\n",
|
477 |
+
" embedding = Model(input_layer, output, name=\"Embedding\")\n",
|
478 |
+
" return embedding"
|
479 |
+
]
|
480 |
+
},
|
481 |
+
{
|
482 |
+
"cell_type": "code",
|
483 |
+
"execution_count": 12,
|
484 |
+
"metadata": {},
|
485 |
+
"outputs": [
|
486 |
+
{
|
487 |
+
"name": "stdout",
|
488 |
+
"output_type": "stream",
|
489 |
+
"text": [
|
490 |
+
"Model: \"Embedding\"\n",
|
491 |
+
"_________________________________________________________________\n",
|
492 |
+
" Layer (type) Output Shape Param # \n",
|
493 |
+
"=================================================================\n",
|
494 |
+
" input_1 (InputLayer) [(None, 478, 3)] 0 \n",
|
495 |
+
" \n",
|
496 |
+
" flatten (Flatten) (None, 1434) 0 \n",
|
497 |
+
" \n",
|
498 |
+
" dense (Dense) (None, 512) 734720 \n",
|
499 |
+
" \n",
|
500 |
+
" batch_normalization (BatchN (None, 512) 2048 \n",
|
501 |
+
" ormalization) \n",
|
502 |
+
" \n",
|
503 |
+
" dense_1 (Dense) (None, 256) 131328 \n",
|
504 |
+
" \n",
|
505 |
+
" batch_normalization_1 (Batc (None, 256) 1024 \n",
|
506 |
+
" hNormalization) \n",
|
507 |
+
" \n",
|
508 |
+
" dense_2 (Dense) (None, 16) 4112 \n",
|
509 |
+
" \n",
|
510 |
+
"=================================================================\n",
|
511 |
+
"Total params: 873,232\n",
|
512 |
+
"Trainable params: 871,696\n",
|
513 |
+
"Non-trainable params: 1,536\n",
|
514 |
+
"_________________________________________________________________\n"
|
515 |
+
]
|
516 |
+
}
|
517 |
+
],
|
518 |
+
"source": [
|
519 |
+
"embedding = create_embedding_model()\n",
|
520 |
+
"embedding.summary()"
|
521 |
+
]
|
522 |
+
},
|
523 |
+
{
|
524 |
+
"cell_type": "code",
|
525 |
+
"execution_count": 13,
|
526 |
+
"metadata": {},
|
527 |
+
"outputs": [],
|
528 |
+
"source": [
|
529 |
+
"class DistanceLayer(layers.Layer):\n",
|
530 |
+
" def __init__(self, **kwargs):\n",
|
531 |
+
" super().__init__(**kwargs)\n",
|
532 |
+
"\n",
|
533 |
+
" def call(self, anchor, positive, negative):\n",
|
534 |
+
" ap_distance = tf.reduce_sum(tf.square(anchor - positive), -1)\n",
|
535 |
+
" an_distance = tf.reduce_sum(tf.square(anchor - negative), -1)\n",
|
536 |
+
" return (ap_distance, an_distance)\n"
|
537 |
+
]
|
538 |
+
},
|
539 |
+
{
|
540 |
+
"cell_type": "code",
|
541 |
+
"execution_count": 14,
|
542 |
+
"metadata": {},
|
543 |
+
"outputs": [],
|
544 |
+
"source": [
|
545 |
+
"anchor_input = layers.Input(name=\"anchor\", shape=(478,) + (3,))\n",
|
546 |
+
"positive_input = layers.Input(name=\"positive\", shape=(478,) + (3,))\n",
|
547 |
+
"negative_input = layers.Input(name=\"negative\", shape=(478,) + (3,))\n",
|
548 |
+
"\n",
|
549 |
+
"distances = DistanceLayer()(\n",
|
550 |
+
" embedding(anchor_input),\n",
|
551 |
+
" embedding(positive_input),\n",
|
552 |
+
" embedding(negative_input),\n",
|
553 |
+
")\n",
|
554 |
+
"\n",
|
555 |
+
"siamese_network = Model(\n",
|
556 |
+
" inputs=[anchor_input, positive_input, negative_input], outputs=distances\n",
|
557 |
+
")"
|
558 |
+
]
|
559 |
+
},
|
560 |
+
{
|
561 |
+
"cell_type": "code",
|
562 |
+
"execution_count": 15,
|
563 |
+
"metadata": {},
|
564 |
+
"outputs": [],
|
565 |
+
"source": [
|
566 |
+
"class SiameseModel(Model):\n",
|
567 |
+
" \"\"\"The Siamese Network model with a custom training and testing loops.\n",
|
568 |
+
"\n",
|
569 |
+
" Computes the triplet loss using the three embeddings produced by the\n",
|
570 |
+
" Siamese Network.\n",
|
571 |
+
"\n",
|
572 |
+
" The triplet loss is defined as:\n",
|
573 |
+
" L(A, P, N) = max(‖f(A) - f(P)‖² - ‖f(A) - f(N)‖² + margin, 0)\n",
|
574 |
+
" \"\"\"\n",
|
575 |
+
"\n",
|
576 |
+
" def __init__(self, siamese_network, margin=0.2):\n",
|
577 |
+
" super().__init__()\n",
|
578 |
+
" self.siamese_network = siamese_network\n",
|
579 |
+
" self.margin = margin\n",
|
580 |
+
" self.loss_tracker = metrics.Mean(name=\"loss\")\n",
|
581 |
+
" self.acc_tracker = metrics.Mean(name=\"accuracy\")\n",
|
582 |
+
"\n",
|
583 |
+
" def call(self, inputs):\n",
|
584 |
+
" return self.siamese_network(inputs)\n",
|
585 |
+
"\n",
|
586 |
+
" def train_step(self, data):\n",
|
587 |
+
" # GradientTape is a context manager that records every operation that\n",
|
588 |
+
" # you do inside. We are using it here to compute the loss so we can get\n",
|
589 |
+
" # the gradients and apply them using the optimizer specified in\n",
|
590 |
+
" # `compile()`.\n",
|
591 |
+
" with tf.GradientTape() as tape:\n",
|
592 |
+
" loss = self._compute_loss(data)\n",
|
593 |
+
"\n",
|
594 |
+
" # Storing the gradients of the loss function with respect to the\n",
|
595 |
+
" # weights/parameters.\n",
|
596 |
+
" gradients = tape.gradient(loss, self.siamese_network.trainable_weights)\n",
|
597 |
+
"\n",
|
598 |
+
" # Applying the gradients on the model using the specified optimizer\n",
|
599 |
+
" self.optimizer.apply_gradients(\n",
|
600 |
+
" zip(gradients, self.siamese_network.trainable_weights)\n",
|
601 |
+
" )\n",
|
602 |
+
"\n",
|
603 |
+
" # Let's update and return the training loss metric.\n",
|
604 |
+
" self.loss_tracker.update_state(loss)\n",
|
605 |
+
" return {\"loss\": self.loss_tracker.result()}\n",
|
606 |
+
"\n",
|
607 |
+
" def test_step(self, data):\n",
|
608 |
+
" loss = self._compute_loss(data)\n",
|
609 |
+
"\n",
|
610 |
+
" # Let's update and return the loss metric.\n",
|
611 |
+
" self.loss_tracker.update_state(loss)\n",
|
612 |
+
" self.acc_tracker.update_state(\n",
|
613 |
+
" self._compute_accuracy(data)\n",
|
614 |
+
" )\n",
|
615 |
+
" return {\"loss\": self.loss_tracker.result(), \"accuracy\": self.acc_tracker.result()}\n",
|
616 |
+
"\n",
|
617 |
+
" def _compute_loss(self, data):\n",
|
618 |
+
" # The output of the network is a tuple containing the distances\n",
|
619 |
+
" # between the anchor and the positive example, and the anchor and\n",
|
620 |
+
" # the negative example.\n",
|
621 |
+
" ap_distance, an_distance = self.siamese_network(data)\n",
|
622 |
+
"\n",
|
623 |
+
" # Computing the Triplet Loss by subtracting both distances and\n",
|
624 |
+
" # making sure we don't get a negative value.\n",
|
625 |
+
" loss = ap_distance - an_distance\n",
|
626 |
+
" loss = tf.maximum(loss + self.margin, 0.0)\n",
|
627 |
+
" return loss\n",
|
628 |
+
" \n",
|
629 |
+
" def _compute_accuracy(self, data):\n",
|
630 |
+
" ap_distance, an_distance = self.siamese_network(data)\n",
|
631 |
+
" return tf.cast(ap_distance < an_distance, tf.float32)\n",
|
632 |
+
"\n",
|
633 |
+
" @property\n",
|
634 |
+
" def metrics(self):\n",
|
635 |
+
" # We need to list our metrics here so the `reset_states()` can be\n",
|
636 |
+
" # called automatically.\n",
|
637 |
+
" return [self.loss_tracker, self.acc_tracker]"
|
638 |
+
]
|
639 |
+
},
|
640 |
+
{
|
641 |
+
"attachments": {},
|
642 |
+
"cell_type": "markdown",
|
643 |
+
"metadata": {},
|
644 |
+
"source": [
|
645 |
+
"## Model Fitting"
|
646 |
+
]
|
647 |
+
},
|
648 |
+
{
|
649 |
+
"cell_type": "code",
|
650 |
+
"execution_count": 16,
|
651 |
+
"metadata": {},
|
652 |
+
"outputs": [
|
653 |
+
{
|
654 |
+
"name": "stdout",
|
655 |
+
"output_type": "stream",
|
656 |
+
"text": [
|
657 |
+
"Epoch 1/40\n",
|
658 |
+
"1/1 [==============================] - 1010s 1010s/step - loss: 0.6925 - val_loss: 0.4313 - val_accuracy: 0.5041\n",
|
659 |
+
"Epoch 2/40\n",
|
660 |
+
"1/1 [==============================] - ETA: 0s - loss: 0.3325"
|
661 |
+
]
|
662 |
+
},
|
663 |
+
{
|
664 |
+
"ename": "KeyboardInterrupt",
|
665 |
+
"evalue": "",
|
666 |
+
"output_type": "error",
|
667 |
+
"traceback": [
|
668 |
+
"\u001b[0;31m---------------------------------------------------------------------------\u001b[0m",
|
669 |
+
"\u001b[0;31mKeyboardInterrupt\u001b[0m Traceback (most recent call last)",
|
670 |
+
"\u001b[0;32m/var/folders/9f/9272h8p951zfstxdwzn0cj2h0000gn/T/ipykernel_61590/1150799400.py\u001b[0m in \u001b[0;36m<module>\u001b[0;34m\u001b[0m\n\u001b[1;32m 1\u001b[0m \u001b[0msiamese_model\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mSiameseModel\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0msiamese_network\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 2\u001b[0m \u001b[0msiamese_model\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mcompile\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0moptimizer\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0moptimizers\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mAdam\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;36m0.0005\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mweighted_metrics\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0;34m\"accuracy\"\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m----> 3\u001b[0;31m \u001b[0msiamese_model\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mfit\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mtrain_dataset\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mepochs\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0;36m40\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mvalidation_data\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0mval_dataset\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m",
|
671 |
+
"\u001b[0;32m~/Git/FacialExpressionSyncService/.venv/lib/python3.7/site-packages/keras/utils/traceback_utils.py\u001b[0m in \u001b[0;36merror_handler\u001b[0;34m(*args, **kwargs)\u001b[0m\n\u001b[1;32m 63\u001b[0m \u001b[0mfiltered_tb\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0;32mNone\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 64\u001b[0m \u001b[0;32mtry\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m---> 65\u001b[0;31m \u001b[0;32mreturn\u001b[0m \u001b[0mfn\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m*\u001b[0m\u001b[0margs\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m**\u001b[0m\u001b[0mkwargs\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m 66\u001b[0m \u001b[0;32mexcept\u001b[0m \u001b[0mException\u001b[0m \u001b[0;32mas\u001b[0m \u001b[0me\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 67\u001b[0m \u001b[0mfiltered_tb\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0m_process_traceback_frames\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0me\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0m__traceback__\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
|
672 |
+
"\u001b[0;32m~/Git/FacialExpressionSyncService/.venv/lib/python3.7/site-packages/keras/engine/training.py\u001b[0m in \u001b[0;36mfit\u001b[0;34m(self, x, y, batch_size, epochs, verbose, callbacks, validation_split, validation_data, shuffle, class_weight, sample_weight, initial_epoch, steps_per_epoch, validation_steps, validation_batch_size, validation_freq, max_queue_size, workers, use_multiprocessing)\u001b[0m\n\u001b[1;32m 1703\u001b[0m \u001b[0muse_multiprocessing\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0muse_multiprocessing\u001b[0m\u001b[0;34m,\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 1704\u001b[0m \u001b[0mreturn_dict\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0;32mTrue\u001b[0m\u001b[0;34m,\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m-> 1705\u001b[0;31m \u001b[0m_use_cached_eval_dataset\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0;32mTrue\u001b[0m\u001b[0;34m,\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m 1706\u001b[0m )\n\u001b[1;32m 1707\u001b[0m val_logs = {\n",
|
673 |
+
"\u001b[0;32m~/Git/FacialExpressionSyncService/.venv/lib/python3.7/site-packages/keras/utils/traceback_utils.py\u001b[0m in \u001b[0;36merror_handler\u001b[0;34m(*args, **kwargs)\u001b[0m\n\u001b[1;32m 63\u001b[0m \u001b[0mfiltered_tb\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0;32mNone\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 64\u001b[0m \u001b[0;32mtry\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m---> 65\u001b[0;31m \u001b[0;32mreturn\u001b[0m \u001b[0mfn\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m*\u001b[0m\u001b[0margs\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m**\u001b[0m\u001b[0mkwargs\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m 66\u001b[0m \u001b[0;32mexcept\u001b[0m \u001b[0mException\u001b[0m \u001b[0;32mas\u001b[0m \u001b[0me\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 67\u001b[0m \u001b[0mfiltered_tb\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0m_process_traceback_frames\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0me\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0m__traceback__\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
|
674 |
+
"\u001b[0;32m~/Git/FacialExpressionSyncService/.venv/lib/python3.7/site-packages/keras/engine/training.py\u001b[0m in \u001b[0;36mevaluate\u001b[0;34m(self, x, y, batch_size, verbose, sample_weight, steps, callbacks, max_queue_size, workers, use_multiprocessing, return_dict, **kwargs)\u001b[0m\n\u001b[1;32m 2038\u001b[0m ):\n\u001b[1;32m 2039\u001b[0m \u001b[0mcallbacks\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mon_test_batch_begin\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mstep\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m-> 2040\u001b[0;31m \u001b[0mtmp_logs\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mtest_function\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0miterator\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m 2041\u001b[0m \u001b[0;32mif\u001b[0m \u001b[0mdata_handler\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mshould_sync\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 2042\u001b[0m \u001b[0mcontext\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0masync_wait\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
|
675 |
+
"\u001b[0;32m~/Git/FacialExpressionSyncService/.venv/lib/python3.7/site-packages/tensorflow/python/util/traceback_utils.py\u001b[0m in \u001b[0;36merror_handler\u001b[0;34m(*args, **kwargs)\u001b[0m\n\u001b[1;32m 148\u001b[0m \u001b[0mfiltered_tb\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0;32mNone\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 149\u001b[0m \u001b[0;32mtry\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m--> 150\u001b[0;31m \u001b[0;32mreturn\u001b[0m \u001b[0mfn\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m*\u001b[0m\u001b[0margs\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m**\u001b[0m\u001b[0mkwargs\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m 151\u001b[0m \u001b[0;32mexcept\u001b[0m \u001b[0mException\u001b[0m \u001b[0;32mas\u001b[0m \u001b[0me\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 152\u001b[0m \u001b[0mfiltered_tb\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0m_process_traceback_frames\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0me\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0m__traceback__\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
|
676 |
+
"\u001b[0;32m~/Git/FacialExpressionSyncService/.venv/lib/python3.7/site-packages/tensorflow/python/eager/polymorphic_function/polymorphic_function.py\u001b[0m in \u001b[0;36m__call__\u001b[0;34m(self, *args, **kwds)\u001b[0m\n\u001b[1;32m 878\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 879\u001b[0m \u001b[0;32mwith\u001b[0m \u001b[0mOptionalXlaContext\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0m_jit_compile\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m--> 880\u001b[0;31m \u001b[0mresult\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0m_call\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m*\u001b[0m\u001b[0margs\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m**\u001b[0m\u001b[0mkwds\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m 881\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 882\u001b[0m \u001b[0mnew_tracing_count\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mexperimental_get_tracing_count\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
|
677 |
+
"\u001b[0;32m~/Git/FacialExpressionSyncService/.venv/lib/python3.7/site-packages/tensorflow/python/eager/polymorphic_function/polymorphic_function.py\u001b[0m in \u001b[0;36m_call\u001b[0;34m(self, *args, **kwds)\u001b[0m\n\u001b[1;32m 917\u001b[0m \u001b[0;31m# In this case we have not created variables on the first call. So we can\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 918\u001b[0m \u001b[0;31m# run the first trace but we should fail if variables are created.\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m--> 919\u001b[0;31m \u001b[0mresults\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0m_variable_creation_fn\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m*\u001b[0m\u001b[0margs\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m**\u001b[0m\u001b[0mkwds\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m 920\u001b[0m \u001b[0;32mif\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0m_created_variables\u001b[0m \u001b[0;32mand\u001b[0m \u001b[0;32mnot\u001b[0m \u001b[0mALLOW_DYNAMIC_VARIABLE_CREATION\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 921\u001b[0m raise ValueError(\"Creating variables on a non-first call to a function\"\n",
|
678 |
+
"\u001b[0;32m~/Git/FacialExpressionSyncService/.venv/lib/python3.7/site-packages/tensorflow/python/eager/polymorphic_function/tracing_compiler.py\u001b[0m in \u001b[0;36m__call__\u001b[0;34m(self, *args, **kwargs)\u001b[0m\n\u001b[1;32m 133\u001b[0m filtered_flat_args) = self._maybe_define_function(args, kwargs)\n\u001b[1;32m 134\u001b[0m return concrete_function._call_flat(\n\u001b[0;32m--> 135\u001b[0;31m filtered_flat_args, captured_inputs=concrete_function.captured_inputs) # pylint: disable=protected-access\n\u001b[0m\u001b[1;32m 136\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 137\u001b[0m \u001b[0;34m@\u001b[0m\u001b[0mproperty\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
|
679 |
+
"\u001b[0;32m~/Git/FacialExpressionSyncService/.venv/lib/python3.7/site-packages/tensorflow/python/eager/polymorphic_function/monomorphic_function.py\u001b[0m in \u001b[0;36m_call_flat\u001b[0;34m(self, args, captured_inputs, cancellation_manager)\u001b[0m\n\u001b[1;32m 1744\u001b[0m \u001b[0;31m# No tape is watching; skip to running the function.\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 1745\u001b[0m return self._build_call_outputs(self._inference_function.call(\n\u001b[0;32m-> 1746\u001b[0;31m ctx, args, cancellation_manager=cancellation_manager))\n\u001b[0m\u001b[1;32m 1747\u001b[0m forward_backward = self._select_forward_and_backward_functions(\n\u001b[1;32m 1748\u001b[0m \u001b[0margs\u001b[0m\u001b[0;34m,\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
|
680 |
+
"\u001b[0;32m~/Git/FacialExpressionSyncService/.venv/lib/python3.7/site-packages/tensorflow/python/eager/polymorphic_function/monomorphic_function.py\u001b[0m in \u001b[0;36mcall\u001b[0;34m(self, ctx, args, cancellation_manager)\u001b[0m\n\u001b[1;32m 381\u001b[0m \u001b[0minputs\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0margs\u001b[0m\u001b[0;34m,\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 382\u001b[0m \u001b[0mattrs\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0mattrs\u001b[0m\u001b[0;34m,\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m--> 383\u001b[0;31m ctx=ctx)\n\u001b[0m\u001b[1;32m 384\u001b[0m \u001b[0;32melse\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 385\u001b[0m outputs = execute.execute_with_cancellation(\n",
|
681 |
+
"\u001b[0;32m~/Git/FacialExpressionSyncService/.venv/lib/python3.7/site-packages/tensorflow/python/eager/execute.py\u001b[0m in \u001b[0;36mquick_execute\u001b[0;34m(op_name, num_outputs, inputs, attrs, ctx, name)\u001b[0m\n\u001b[1;32m 51\u001b[0m \u001b[0mctx\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mensure_initialized\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 52\u001b[0m tensors = pywrap_tfe.TFE_Py_Execute(ctx._handle, device_name, op_name,\n\u001b[0;32m---> 53\u001b[0;31m inputs, attrs, num_outputs)\n\u001b[0m\u001b[1;32m 54\u001b[0m \u001b[0;32mexcept\u001b[0m \u001b[0mcore\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0m_NotOkStatusException\u001b[0m \u001b[0;32mas\u001b[0m \u001b[0me\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 55\u001b[0m \u001b[0;32mif\u001b[0m \u001b[0mname\u001b[0m \u001b[0;32mis\u001b[0m \u001b[0;32mnot\u001b[0m \u001b[0;32mNone\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
|
682 |
+
"\u001b[0;31mKeyboardInterrupt\u001b[0m: "
|
683 |
+
]
|
684 |
+
}
|
685 |
+
],
|
686 |
+
"source": [
|
687 |
+
"siamese_model = SiameseModel(siamese_network)\n",
|
688 |
+
"siamese_model.compile(optimizer=optimizers.Adam(0.0005), weighted_metrics=[\"accuracy\"])\n",
|
689 |
+
"siamese_model.fit(train_dataset, epochs=40, validation_data=val_dataset)"
|
690 |
+
]
|
691 |
+
},
|
692 |
+
{
|
693 |
+
"attachments": {},
|
694 |
+
"cell_type": "markdown",
|
695 |
+
"metadata": {},
|
696 |
+
"source": [
|
697 |
+
"## Evaluate Model"
|
698 |
+
]
|
699 |
+
},
|
700 |
+
{
|
701 |
+
"cell_type": "code",
|
702 |
+
"execution_count": null,
|
703 |
+
"metadata": {},
|
704 |
+
"outputs": [],
|
705 |
+
"source": [
|
706 |
+
"test_df = fecnet_dataset_loader(\"data/faceexp-comparison-data-test-public.csv\")\n",
|
707 |
+
"\n",
|
708 |
+
"test_dataset = tf.data.Dataset.from_tensor_slices(\n",
|
709 |
+
" test_df\n",
|
710 |
+
")\n",
|
711 |
+
"\n",
|
712 |
+
"test_dataset = test_dataset.map(preprocess_triplets)"
|
713 |
+
]
|
714 |
+
},
|
715 |
+
{
|
716 |
+
"cell_type": "code",
|
717 |
+
"execution_count": null,
|
718 |
+
"metadata": {},
|
719 |
+
"outputs": [],
|
720 |
+
"source": [
|
721 |
+
"siamese_model.evaluate(test_dataset)"
|
722 |
+
]
|
723 |
+
},
|
724 |
+
{
|
725 |
+
"cell_type": "code",
|
726 |
+
"execution_count": 168,
|
727 |
+
"metadata": {},
|
728 |
+
"outputs": [
|
729 |
+
{
|
730 |
+
"name": "stdout",
|
731 |
+
"output_type": "stream",
|
732 |
+
"text": [
|
733 |
+
"Positive similarity: 0.96082336\n",
|
734 |
+
"Negative similarity 0.85784876\n",
|
735 |
+
"Positive-Negative similarity 0.8386482\n"
|
736 |
+
]
|
737 |
+
}
|
738 |
+
],
|
739 |
+
"source": [
|
740 |
+
"sample = next(iter(train_dataset))\n",
|
741 |
+
"# visualise_face_mesh_triplet(*sample)\n",
|
742 |
+
"\n",
|
743 |
+
"anchor, positive, negative = sample\n",
|
744 |
+
"anchor_embedding, positive_embedding, negative_embedding = (\n",
|
745 |
+
" embedding(tf.reshape(anchor, (-1, 478,3))),\n",
|
746 |
+
" embedding(tf.reshape(positive, (-1, 478,3))),\n",
|
747 |
+
" embedding(tf.reshape(negative, (-1, 478,3))),\n",
|
748 |
+
")\n",
|
749 |
+
"cosine_similarity = metrics.CosineSimilarity()\n",
|
750 |
+
"\n",
|
751 |
+
"positive_similarity = cosine_similarity(anchor_embedding, positive_embedding)\n",
|
752 |
+
"print(\"Positive similarity:\", positive_similarity.numpy())\n",
|
753 |
+
"\n",
|
754 |
+
"negative_similarity = cosine_similarity(anchor_embedding, negative_embedding)\n",
|
755 |
+
"print(\"Negative similarity\", negative_similarity.numpy())\n",
|
756 |
+
"\n",
|
757 |
+
"positive_negative_similarity = cosine_similarity(positive_embedding, negative_embedding)\n",
|
758 |
+
"print(\"Positive-Negative similarity\", positive_negative_similarity.numpy())"
|
759 |
+
]
|
760 |
+
},
|
761 |
+
{
|
762 |
+
"cell_type": "code",
|
763 |
+
"execution_count": 8,
|
764 |
+
"metadata": {},
|
765 |
+
"outputs": [
|
766 |
+
{
|
767 |
+
"name": "stderr",
|
768 |
+
"output_type": "stream",
|
769 |
+
"text": [
|
770 |
+
"INFO: Created TensorFlow Lite XNNPACK delegate for CPU.\n"
|
771 |
+
]
|
772 |
+
}
|
773 |
+
],
|
774 |
+
"source": [
|
775 |
+
"# tf.data.Dataset.save(dataset, \"training_dataset\")"
|
776 |
+
]
|
777 |
+
},
|
778 |
+
{
|
779 |
+
"cell_type": "code",
|
780 |
+
"execution_count": null,
|
781 |
+
"metadata": {},
|
782 |
+
"outputs": [],
|
783 |
+
"source": []
|
784 |
+
}
|
785 |
+
],
|
786 |
+
"metadata": {
|
787 |
+
"kernelspec": {
|
788 |
+
"display_name": ".venv",
|
789 |
+
"language": "python",
|
790 |
+
"name": "python3"
|
791 |
+
},
|
792 |
+
"language_info": {
|
793 |
+
"codemirror_mode": {
|
794 |
+
"name": "ipython",
|
795 |
+
"version": 3
|
796 |
+
},
|
797 |
+
"file_extension": ".py",
|
798 |
+
"mimetype": "text/x-python",
|
799 |
+
"name": "python",
|
800 |
+
"nbconvert_exporter": "python",
|
801 |
+
"pygments_lexer": "ipython3",
|
802 |
+
"version": "3.7.9"
|
803 |
+
},
|
804 |
+
"orig_nbformat": 4
|
805 |
+
},
|
806 |
+
"nbformat": 4,
|
807 |
+
"nbformat_minor": 2
|
808 |
+
}
|
model/fecnet.py
ADDED
@@ -0,0 +1,72 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import sys
|
3 |
+
from importlib import import_module, invalidate_caches
|
4 |
+
from importlib.util import module_from_spec, spec_from_file_location
|
5 |
+
from tempfile import TemporaryDirectory
|
6 |
+
|
7 |
+
import cv2
|
8 |
+
import numpy as np
|
9 |
+
import plotly.express as px
|
10 |
+
import requests
|
11 |
+
import torch
|
12 |
+
from git import Repo
|
13 |
+
from huggingface_hub import hf_hub_download
|
14 |
+
|
15 |
+
|
16 |
+
class FECNetModel:
|
17 |
+
def __init__(self, hf_token: str) -> None:
|
18 |
+
self.hf_token = hf_token
|
19 |
+
repo_dir = TemporaryDirectory()
|
20 |
+
Repo.clone_from(
|
21 |
+
"https://github.com/AmirSh15/FECNet.git",
|
22 |
+
repo_dir.name,
|
23 |
+
)
|
24 |
+
invalidate_caches()
|
25 |
+
sys.path.append(repo_dir.name)
|
26 |
+
fecnet_module_path = os.path.join(repo_dir.name, "models", "FECNet.py")
|
27 |
+
with open(fecnet_module_path, "r") as f:
|
28 |
+
content = f.read()
|
29 |
+
content = content.replace(
|
30 |
+
"cuda",
|
31 |
+
"cpu",
|
32 |
+
)
|
33 |
+
with open(fecnet_module_path, "w") as f:
|
34 |
+
f.write(content)
|
35 |
+
spec = spec_from_file_location("FECNet", fecnet_module_path)
|
36 |
+
fecnet_module = module_from_spec(spec) # type: ignore
|
37 |
+
spec.loader.exec_module(fecnet_module) # type: ignore
|
38 |
+
|
39 |
+
self.model = self.__load_model(
|
40 |
+
self.__download_weights(repo_dir.name), fecnet_module.FECNet
|
41 |
+
)
|
42 |
+
|
43 |
+
def __download_weights(self, model_dir: str) -> str:
|
44 |
+
model_path = hf_hub_download(
|
45 |
+
"natexcvi/pretrained-fecnet",
|
46 |
+
"fecnet.pt",
|
47 |
+
token=self.hf_token,
|
48 |
+
)
|
49 |
+
return model_path
|
50 |
+
|
51 |
+
def __load_model(self, model_path: str, model_class):
|
52 |
+
model = model_class(pretrained=False)
|
53 |
+
model_weights = torch.load(model_path, map_location=torch.device("cpu"))
|
54 |
+
model.load_state_dict(model_weights)
|
55 |
+
model.eval()
|
56 |
+
return model.double()
|
57 |
+
|
58 |
+
def predict(self, image: np.ndarray):
|
59 |
+
pred = self.model.forward(image)
|
60 |
+
return pred
|
61 |
+
|
62 |
+
def distance(a, b):
|
63 |
+
return np.linalg.norm(a - b)
|
64 |
+
|
65 |
+
def embed_image(self, image) -> np.ndarray:
|
66 |
+
image = cv2.imdecode(image, cv2.IMREAD_COLOR)
|
67 |
+
image = cv2.resize(image, (224, 224))
|
68 |
+
image = np.transpose(image, (2, 0, 1))
|
69 |
+
image = np.expand_dims(image, axis=0)
|
70 |
+
image = torch.from_numpy(image).double()
|
71 |
+
pred = self.predict(image)
|
72 |
+
return pred.detach().numpy()
|
model/fecnet_test.py
ADDED
@@ -0,0 +1,20 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
|
3 |
+
import numpy as np
|
4 |
+
import pytest
|
5 |
+
from dotenv import load_dotenv
|
6 |
+
from fecnet import FECNetModel
|
7 |
+
|
8 |
+
load_dotenv()
|
9 |
+
|
10 |
+
|
11 |
+
@pytest.fixture()
|
12 |
+
def model():
|
13 |
+
return FECNetModel(hf_token=os.getenv("HF_TOKEN"))
|
14 |
+
|
15 |
+
|
16 |
+
def test_embed(model: FECNetModel):
|
17 |
+
image = open("testdata/face_pic.jpeg", "rb").read()
|
18 |
+
image_arr = np.asarray(bytearray(image), dtype=np.uint8)
|
19 |
+
rep = model.embed_image(image_arr)
|
20 |
+
assert rep.shape == (1, 16)
|
model.py → model/model.py
RENAMED
File without changes
|
model/openface_model.py
ADDED
@@ -0,0 +1,27 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import cv2
|
2 |
+
import numpy as np
|
3 |
+
import openface
|
4 |
+
import docker
|
5 |
+
|
6 |
+
|
7 |
+
class OpenFaceModel:
|
8 |
+
def __init__(self) -> None:
|
9 |
+
self.client = docker.from_env()
|
10 |
+
self.client.images.pull("bamos/openface")
|
11 |
+
|
12 |
+
def preprocess(self, image: bytes) -> np.ndarray:
|
13 |
+
raise NotImplemented
|
14 |
+
|
15 |
+
def embed(self, aligned_face):
|
16 |
+
container = self.client.containers.run(
|
17 |
+
"bamos/openface",
|
18 |
+
"python /root/openface/demos/classifier.py infer /root/openface/models/openface/celeb-classifier.nn4.small2.v1.pkl -",
|
19 |
+
detach=True,
|
20 |
+
stdin_open=True,
|
21 |
+
tty=True,
|
22 |
+
)
|
23 |
+
raise NotImplemented
|
24 |
+
return rep
|
25 |
+
|
26 |
+
def similarity(self, rep1, rep2):
|
27 |
+
return np.linalg.norm(rep1 - rep2, ord=2)
|
model/openface_model_test.py
ADDED
@@ -0,0 +1,14 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import pytest
|
2 |
+
from openface_model import OpenFaceModel
|
3 |
+
|
4 |
+
|
5 |
+
@pytest.fixture()
|
6 |
+
def model():
|
7 |
+
return OpenFaceModel()
|
8 |
+
|
9 |
+
@pytest.mark.skip(reason="Not implemented")
|
10 |
+
def test_embed(model):
|
11 |
+
image = open("../testdata/face_pic.jpeg", "rb").read()
|
12 |
+
aligned_face = model.preprocess(image)
|
13 |
+
rep = model.embed(aligned_face)
|
14 |
+
assert rep.shape == (128,)
|
requirements.txt
CHANGED
@@ -9,4 +9,9 @@ mediapipe
|
|
9 |
pandas
|
10 |
pytest
|
11 |
python-dotenv
|
12 |
-
bcrypt
|
|
|
|
|
|
|
|
|
|
|
|
9 |
pandas
|
10 |
pytest
|
11 |
python-dotenv
|
12 |
+
bcrypt
|
13 |
+
openface
|
14 |
+
dlib
|
15 |
+
docker
|
16 |
+
torch
|
17 |
+
gitpython
|
routers/fecnet_router.py
ADDED
@@ -0,0 +1,33 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from fastapi import APIRouter, Depends, File, UploadFile, status
|
2 |
+
|
3 |
+
from auth import validate_token
|
4 |
+
from schema import EmbeddingResponse, SimilarityResponse
|
5 |
+
|
6 |
+
router = APIRouter(
|
7 |
+
prefix="/fecnet",
|
8 |
+
tags=["fecnet"],
|
9 |
+
dependencies=[Depends(validate_token)],
|
10 |
+
)
|
11 |
+
|
12 |
+
|
13 |
+
@router.get(
|
14 |
+
"/embed",
|
15 |
+
status_code=status.HTTP_200_OK,
|
16 |
+
response_model=EmbeddingResponse,
|
17 |
+
)
|
18 |
+
async def calculate_embedding(
|
19 |
+
image: UploadFile = File(...),
|
20 |
+
):
|
21 |
+
return {"message": "Hello World"}
|
22 |
+
|
23 |
+
|
24 |
+
@router.get(
|
25 |
+
"/similarity",
|
26 |
+
status_code=status.HTTP_200_OK,
|
27 |
+
response_model=SimilarityResponse,
|
28 |
+
)
|
29 |
+
async def calculate_similarity_score(
|
30 |
+
image1: UploadFile = File(...),
|
31 |
+
image2: UploadFile = File(...),
|
32 |
+
):
|
33 |
+
return {"message": "Hello World"}
|
routers/openface_router.py
ADDED
@@ -0,0 +1,33 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from fastapi import APIRouter, Depends, File, UploadFile, status
|
2 |
+
|
3 |
+
from auth import validate_token
|
4 |
+
from schema import EmbeddingResponse, SimilarityResponse
|
5 |
+
|
6 |
+
router = APIRouter(
|
7 |
+
prefix="/openface",
|
8 |
+
tags=["openface"],
|
9 |
+
dependencies=[Depends(validate_token)],
|
10 |
+
)
|
11 |
+
|
12 |
+
|
13 |
+
@router.get(
|
14 |
+
"/embed",
|
15 |
+
status_code=status.HTTP_200_OK,
|
16 |
+
response_model=EmbeddingResponse,
|
17 |
+
)
|
18 |
+
async def calculate_embedding(
|
19 |
+
image: UploadFile = File(...),
|
20 |
+
):
|
21 |
+
return {"message": "Hello World"}
|
22 |
+
|
23 |
+
|
24 |
+
@router.get(
|
25 |
+
"/similarity",
|
26 |
+
status_code=status.HTTP_200_OK,
|
27 |
+
response_model=SimilarityResponse,
|
28 |
+
)
|
29 |
+
async def calculate_similarity_score(
|
30 |
+
image1: UploadFile = File(...),
|
31 |
+
image2: UploadFile = File(...),
|
32 |
+
):
|
33 |
+
return {"message": "Hello World"}
|