Spaces:
Running
Running
vijul.shah
commited on
Commit
·
57d7ed3
1
Parent(s):
c34dd19
init
Browse files
app.py
ADDED
|
@@ -0,0 +1,229 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# takn from: https://huggingface.co/spaces/frgfm/torch-cam/blob/main/app.py
|
| 2 |
+
|
| 3 |
+
# streamlit run app.py
|
| 4 |
+
from io import BytesIO
|
| 5 |
+
import os
|
| 6 |
+
import sys
|
| 7 |
+
import matplotlib.pyplot as plt
|
| 8 |
+
import requests
|
| 9 |
+
import streamlit as st
|
| 10 |
+
import torch
|
| 11 |
+
from PIL import Image
|
| 12 |
+
from torchvision import models
|
| 13 |
+
from torchvision.transforms.functional import normalize, resize, to_pil_image, to_tensor
|
| 14 |
+
from torchvision import transforms
|
| 15 |
+
|
| 16 |
+
from torchcam.methods import CAM
|
| 17 |
+
from torchcam import methods as torchcam_methods
|
| 18 |
+
from torchcam.utils import overlay_mask
|
| 19 |
+
import os.path as osp
|
| 20 |
+
|
| 21 |
+
root_path = osp.abspath(osp.join(__file__, osp.pardir))
|
| 22 |
+
sys.path.append(root_path)
|
| 23 |
+
|
| 24 |
+
from utils import get_model
|
| 25 |
+
from registry_utils import import_registered_modules
|
| 26 |
+
|
| 27 |
+
import_registered_modules()
|
| 28 |
+
# from torchcam.methods._utils import locate_candidate_layer
|
| 29 |
+
|
| 30 |
+
CAM_METHODS = [
|
| 31 |
+
"CAM",
|
| 32 |
+
# "GradCAM",
|
| 33 |
+
# "GradCAMpp",
|
| 34 |
+
# "SmoothGradCAMpp",
|
| 35 |
+
# "ScoreCAM",
|
| 36 |
+
# "SSCAM",
|
| 37 |
+
# "ISCAM",
|
| 38 |
+
# "XGradCAM",
|
| 39 |
+
# "LayerCAM",
|
| 40 |
+
]
|
| 41 |
+
TV_MODELS = [
|
| 42 |
+
"resnet18",
|
| 43 |
+
# "resnet50",
|
| 44 |
+
]
|
| 45 |
+
SR_METHODS = ["GFPGAN", "RealESRGAN", "SRResNet", "CodeFormer", "HAT"]
|
| 46 |
+
UPSCALE = ["2", "3", "4"]
|
| 47 |
+
LABEL_MAP = [
|
| 48 |
+
"left_eye",
|
| 49 |
+
"right_eye",
|
| 50 |
+
]
|
| 51 |
+
|
| 52 |
+
|
| 53 |
+
@torch.no_grad()
|
| 54 |
+
def _load_model(model_configs, device="cpu"):
|
| 55 |
+
model_path = os.path.join(root_path, model_configs["model_path"])
|
| 56 |
+
model_configs.pop("model_path")
|
| 57 |
+
model_dict = torch.load(model_path, map_location=device)
|
| 58 |
+
model = get_model(model_configs=model_configs)
|
| 59 |
+
model.load_state_dict(model_dict)
|
| 60 |
+
model = model.to(device)
|
| 61 |
+
model = model.eval()
|
| 62 |
+
return model
|
| 63 |
+
|
| 64 |
+
|
| 65 |
+
def main():
|
| 66 |
+
# Wide mode
|
| 67 |
+
st.set_page_config(page_title="Pupil Diameter Estimator", layout="wide")
|
| 68 |
+
|
| 69 |
+
# Designing the interface
|
| 70 |
+
st.title("EyeDentify Playground")
|
| 71 |
+
# For newline
|
| 72 |
+
st.write("\n")
|
| 73 |
+
# Set the columns
|
| 74 |
+
cols = st.columns((1, 1))
|
| 75 |
+
# cols = st.columns((1, 1, 1))
|
| 76 |
+
cols[0].header("Input image")
|
| 77 |
+
# cols[1].header("Raw CAM")
|
| 78 |
+
cols[-1].header("Prediction")
|
| 79 |
+
|
| 80 |
+
# Sidebar
|
| 81 |
+
# File selection
|
| 82 |
+
st.sidebar.title("Input selection")
|
| 83 |
+
# Disabling warning
|
| 84 |
+
st.set_option("deprecation.showfileUploaderEncoding", False)
|
| 85 |
+
# Choose your own image
|
| 86 |
+
uploaded_file = st.sidebar.file_uploader(
|
| 87 |
+
"Upload files", type=["png", "jpeg", "jpg"]
|
| 88 |
+
)
|
| 89 |
+
if uploaded_file is not None:
|
| 90 |
+
img = Image.open(BytesIO(uploaded_file.read()), mode="r").convert("RGB")
|
| 91 |
+
|
| 92 |
+
cols[0].image(img, use_column_width=True)
|
| 93 |
+
|
| 94 |
+
# Model selection
|
| 95 |
+
st.sidebar.title("Setup")
|
| 96 |
+
tv_model = st.sidebar.selectbox(
|
| 97 |
+
"Classification model",
|
| 98 |
+
TV_MODELS,
|
| 99 |
+
help="Supported models from Torchvision",
|
| 100 |
+
)
|
| 101 |
+
|
| 102 |
+
# class_choices = [
|
| 103 |
+
# f"{idx + 1} - {class_name}" for idx, class_name in enumerate(LABEL_MAP)
|
| 104 |
+
# ]
|
| 105 |
+
# class_selection = st.sidebar.selectbox(
|
| 106 |
+
# "Class selection", ["Predicted class (argmax)", *class_choices]
|
| 107 |
+
# )
|
| 108 |
+
|
| 109 |
+
img_configs = {"img_size": [32, 64], "means": None, "stds": None}
|
| 110 |
+
# For newline
|
| 111 |
+
st.sidebar.write("\n")
|
| 112 |
+
|
| 113 |
+
if st.sidebar.button("Compute CAM"):
|
| 114 |
+
if uploaded_file is None:
|
| 115 |
+
st.sidebar.error("Please upload an image first")
|
| 116 |
+
|
| 117 |
+
else:
|
| 118 |
+
with st.spinner("Analyzing..."):
|
| 119 |
+
|
| 120 |
+
preprocess_steps = [transforms.ToTensor()]
|
| 121 |
+
|
| 122 |
+
image_size = img_configs["img_size"]
|
| 123 |
+
if image_size is not None:
|
| 124 |
+
preprocess_steps.append(
|
| 125 |
+
transforms.Resize(
|
| 126 |
+
[image_size[0], image_size[-1]],
|
| 127 |
+
interpolation=transforms.InterpolationMode.BICUBIC,
|
| 128 |
+
antialias=True,
|
| 129 |
+
)
|
| 130 |
+
)
|
| 131 |
+
|
| 132 |
+
means = img_configs["means"]
|
| 133 |
+
stds = img_configs["stds"]
|
| 134 |
+
if means is not None and stds is not None:
|
| 135 |
+
preprocess_steps.append(transforms.Normalize(means, stds))
|
| 136 |
+
|
| 137 |
+
preprocess_function = transforms.Compose(preprocess_steps)
|
| 138 |
+
input_img = preprocess_function(img)
|
| 139 |
+
input_img = input_img.unsqueeze(0).to(device="cpu")
|
| 140 |
+
|
| 141 |
+
model_configs = {
|
| 142 |
+
"model_path": root_path
|
| 143 |
+
+ "/pre_trained_models/ResNet18/left_eye.pt",
|
| 144 |
+
"registered_model_name": "ResNet18",
|
| 145 |
+
"num_classes": 1,
|
| 146 |
+
}
|
| 147 |
+
registered_model_name = model_configs["registered_model_name"]
|
| 148 |
+
# default_layer = ""
|
| 149 |
+
if tv_model is not None:
|
| 150 |
+
with st.spinner("Loading model..."):
|
| 151 |
+
model = _load_model(model_configs)
|
| 152 |
+
|
| 153 |
+
if torch.cuda.is_available():
|
| 154 |
+
model = model.cuda()
|
| 155 |
+
|
| 156 |
+
if registered_model_name == "ResNet18":
|
| 157 |
+
target_layer = model.resnet.layer4[-1].conv2
|
| 158 |
+
elif registered_model_name == "ResNet50":
|
| 159 |
+
target_layer = model.resnet.layer4[-1].conv3
|
| 160 |
+
else:
|
| 161 |
+
raise Exception(
|
| 162 |
+
f"No target layer available for selected model: {registered_model_name}"
|
| 163 |
+
)
|
| 164 |
+
|
| 165 |
+
# target_layer = st.sidebar.text_input(
|
| 166 |
+
# "Target layer",
|
| 167 |
+
# default_layer,
|
| 168 |
+
# help='If you want to target several layers, add a "+" separator (e.g. "layer3+layer4")',
|
| 169 |
+
# )
|
| 170 |
+
cam_method = "CAM"
|
| 171 |
+
# cam_method = st.sidebar.selectbox(
|
| 172 |
+
# "CAM method",
|
| 173 |
+
# CAM_METHODS,
|
| 174 |
+
# help="The way your class activation map will be computed",
|
| 175 |
+
# )
|
| 176 |
+
if cam_method is not None:
|
| 177 |
+
# cam_extractor = methods.__dict__[cam_method](
|
| 178 |
+
# model,
|
| 179 |
+
# target_layer=(
|
| 180 |
+
# [s.strip() for s in target_layer.split("+")]
|
| 181 |
+
# if len(target_layer) > 0
|
| 182 |
+
# else None
|
| 183 |
+
# ),
|
| 184 |
+
# )
|
| 185 |
+
cam_extractor = torchcam_methods.__dict__[cam_method](
|
| 186 |
+
model,
|
| 187 |
+
target_layer=target_layer,
|
| 188 |
+
fc_layer=model.resnet.fc,
|
| 189 |
+
input_shape=(3, 32, 64),
|
| 190 |
+
)
|
| 191 |
+
# with torch.no_grad():
|
| 192 |
+
# if input_mask is not None:
|
| 193 |
+
# out = self.model(input_img, input_mask)
|
| 194 |
+
# else:
|
| 195 |
+
# out = self.model(input_img)
|
| 196 |
+
# activation_map = cam_extractor(class_idx=target_class)
|
| 197 |
+
|
| 198 |
+
# Forward the image to the model
|
| 199 |
+
out = model(input_img)
|
| 200 |
+
print("out = ", out)
|
| 201 |
+
|
| 202 |
+
# Select the target class
|
| 203 |
+
# if class_selection == "Predicted class (argmax)":
|
| 204 |
+
# class_idx = out.squeeze(0).argmax().item()
|
| 205 |
+
# else:
|
| 206 |
+
# class_idx = LABEL_MAP.index(class_selection.rpartition(" - ")[-1])
|
| 207 |
+
|
| 208 |
+
# Retrieve the CAM
|
| 209 |
+
# act_maps = cam_extractor(class_idx=target_class)
|
| 210 |
+
act_maps = cam_extractor(0, out)
|
| 211 |
+
# Fuse the CAMs if there are several
|
| 212 |
+
activation_map = (
|
| 213 |
+
act_maps[0]
|
| 214 |
+
if len(act_maps) == 1
|
| 215 |
+
else cam_extractor.fuse_cams(act_maps)
|
| 216 |
+
)
|
| 217 |
+
|
| 218 |
+
# Overlayed CAM
|
| 219 |
+
fig, ax = plt.subplots()
|
| 220 |
+
result = overlay_mask(
|
| 221 |
+
img, to_pil_image(activation_map, mode="F"), alpha=0.5
|
| 222 |
+
)
|
| 223 |
+
ax.imshow(result)
|
| 224 |
+
ax.axis("off")
|
| 225 |
+
cols[-1].pyplot(fig)
|
| 226 |
+
|
| 227 |
+
|
| 228 |
+
if __name__ == "__main__":
|
| 229 |
+
main()
|