Spaces:
Runtime error
Runtime error
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,280 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import pickle
|
2 |
+
import pandas as pd
|
3 |
+
import tensorflow as tf
|
4 |
+
from tensorflow.keras.models import load_model
|
5 |
+
from collections import Counter
|
6 |
+
|
7 |
+
# Creating a numpy array of shape (8, 16, 1)
|
8 |
+
import cv2
|
9 |
+
import numpy as np
|
10 |
+
import matplotlib.pyplot as plt
|
11 |
+
import gradio as gr
|
12 |
+
|
13 |
+
flow_field = np.ones((128,256), dtype = np.uint8)
|
14 |
+
|
15 |
+
# Changing the left input side
|
16 |
+
flow_field[:,0] = 3
|
17 |
+
# Changing the right output side
|
18 |
+
flow_field[:,-1] = 4
|
19 |
+
# Changing the top layer
|
20 |
+
flow_field[0,:] = 2
|
21 |
+
# Changing the bottom layer
|
22 |
+
flow_field[-1,:] = 2
|
23 |
+
|
24 |
+
def nvs_loss(y_pred, rho=10, nu=0.0001): #arbitary rho and nu(Later use values of air)
|
25 |
+
u,v,p = tf.split(y_pred, 3, axis=3)
|
26 |
+
|
27 |
+
#First order derivative
|
28 |
+
du_dx, du_dy = tf.image.image_gradients(u) # tf.image.image_gradients returns a tuple containing two tensors: u-grad along the x dir and u-grad along the y dir
|
29 |
+
dv_dx, dv_dy = tf.image.image_gradients(v)
|
30 |
+
dp_dx, dp_dy = tf.image.image_gradients(p)
|
31 |
+
|
32 |
+
#Second order derivatives
|
33 |
+
du_dx2, du_dydx = tf.image.image_gradients(du_dx) # du_dydx will be unused
|
34 |
+
du_dxdy, du_dy2 = tf.image.image_gradients(du_dy) # du_dxdy will be unused
|
35 |
+
|
36 |
+
dv_dx2, dv_dydx = tf.image.image_gradients(dv_dx)
|
37 |
+
dv_dxdy, dv_dy2 = tf.image.image_gradients(dv_dy)
|
38 |
+
|
39 |
+
#Momentum equation
|
40 |
+
er1_tensor = tf.math.multiply(u, du_dx) + tf.math.multiply(v, du_dy) + 1.0*dp_dx/rho - nu*(du_dx2 + du_dy2)
|
41 |
+
er2_tensor = tf.math.multiply(u, dv_dx) + tf.math.multiply(v, dv_dy) + 1.0*dp_dy/rho - nu*(dv_dx2 + dv_dy2)
|
42 |
+
|
43 |
+
# # #Continuity equation
|
44 |
+
er3_tensor = du_dx + dv_dy
|
45 |
+
|
46 |
+
er1 = tf.reduce_mean(er1_tensor)
|
47 |
+
er2 = tf.reduce_mean(er2_tensor)
|
48 |
+
er3 = tf.reduce_mean(er3_tensor)
|
49 |
+
|
50 |
+
return er1*er1 + er2*er2 + er3*er3
|
51 |
+
|
52 |
+
# Initiating the Loss Function-
|
53 |
+
def custom_loss(y_true, y_pred):
|
54 |
+
nv_loss = nvs_loss(y_pred)
|
55 |
+
mse_loss = tf.reduce_mean(tf.square(y_true-y_pred)) # Try mse loss function here
|
56 |
+
return mse_loss + nv_loss
|
57 |
+
|
58 |
+
import torch
|
59 |
+
import matplotlib
|
60 |
+
def colorize(value, vmin=None, vmax=None, cmap='gray_r', invalid_val=-99, invalid_mask=None, background_color=(128, 128, 128, 255), gamma_corrected=False, value_transform=None):
|
61 |
+
"""Converts a depth map to a color image.
|
62 |
+
|
63 |
+
Args:
|
64 |
+
value (torch.Tensor, numpy.ndarry): Input depth map. Shape: (H, W) or (1, H, W) or (1, 1, H, W). All singular dimensions are squeezed
|
65 |
+
vmin (float, optional): vmin-valued entries are mapped to start color of cmap. If None, value.min() is used. Defaults to None.
|
66 |
+
vmax (float, optional): vmax-valued entries are mapped to end color of cmap. If None, value.max() is used. Defaults to None.
|
67 |
+
cmap (str, optional): matplotlib colormap to use. Defaults to 'magma_r'.
|
68 |
+
invalid_val (int, optional): Specifies value of invalid pixels that should be colored as 'background_color'. Defaults to -99.
|
69 |
+
invalid_mask (numpy.ndarray, optional): Boolean mask for invalid regions. Defaults to None.
|
70 |
+
background_color (tuple[int], optional): 4-tuple RGB color to give to invalid pixels. Defaults to (128, 128, 128, 255).
|
71 |
+
gamma_corrected (bool, optional): Apply gamma correction to colored image. Defaults to False.
|
72 |
+
value_transform (Callable, optional): Apply transform function to valid pixels before coloring. Defaults to None.
|
73 |
+
|
74 |
+
Returns:
|
75 |
+
numpy.ndarray, dtype - uint8: Colored depth map. Shape: (H, W, 4)
|
76 |
+
"""
|
77 |
+
if isinstance(value, torch.Tensor):
|
78 |
+
value = value.detach().cpu().numpy()
|
79 |
+
|
80 |
+
value = value.squeeze()
|
81 |
+
if invalid_mask is None:
|
82 |
+
invalid_mask = value == invalid_val
|
83 |
+
mask = np.logical_not(invalid_mask)
|
84 |
+
|
85 |
+
# normalize
|
86 |
+
# vmin = np.percentile(value[mask],2) if vmin is None else vmin
|
87 |
+
# vmax = np.percentile(value[mask],85) if vmax is None else vmax
|
88 |
+
vmin = np.min(value[mask]) if vmin is None else vmin
|
89 |
+
vmax = np.max(value[mask]) if vmax is None else vmax
|
90 |
+
if vmin != vmax:
|
91 |
+
value = (value - vmin) / (vmax - vmin) # vmin..vmax
|
92 |
+
else:
|
93 |
+
# Avoid 0-division
|
94 |
+
value = value * 0.
|
95 |
+
|
96 |
+
# squeeze last dim if it exists
|
97 |
+
# grey out the invalid values
|
98 |
+
|
99 |
+
value[invalid_mask] = np.nan
|
100 |
+
cmapper = matplotlib.cm.get_cmap(cmap)
|
101 |
+
if value_transform:
|
102 |
+
value = value_transform(value)
|
103 |
+
# value = value / value.max()
|
104 |
+
value = cmapper(value, bytes=True) # (nxmx4)
|
105 |
+
|
106 |
+
# img = value[:, :, :]
|
107 |
+
img = value[...]
|
108 |
+
img[invalid_mask] = background_color
|
109 |
+
|
110 |
+
# return img.transpose((2, 0, 1))
|
111 |
+
if gamma_corrected:
|
112 |
+
# gamma correction
|
113 |
+
img = img / 255
|
114 |
+
img = np.power(img, 2.2)
|
115 |
+
img = img * 255
|
116 |
+
img = img.astype(np.uint8)
|
117 |
+
return img
|
118 |
+
|
119 |
+
def img_preprocess(image, h, w):
|
120 |
+
# Convert the drawn image to grayscale
|
121 |
+
img_gray = cv2.cvtColor(image, cv2.COLOR_RGB2GRAY)
|
122 |
+
|
123 |
+
# Threshold the grayscale image to create a binary image
|
124 |
+
_, binary_img = cv2.threshold(img_gray, 1, 255, cv2.THRESH_BINARY)
|
125 |
+
|
126 |
+
# Perform flood fill starting from a point inside the shape. Fill the inside with pixel value 0
|
127 |
+
seed_point = (int(h/2), int(w/2))
|
128 |
+
retval, flooded_image, mask, rect = cv2.floodFill(binary_img, None, seed_point, 0)
|
129 |
+
flooded_image = (flooded_image/255).astype(np.uint8)
|
130 |
+
return flooded_image
|
131 |
+
|
132 |
+
def patch_stiching(flooded_image, h, w, x0, y0): # ((x0, y0) = center of channel, (w1, h1) = height and width of patch)
|
133 |
+
flow_field_updated = np.copy(flow_field)
|
134 |
+
print('flow field updated - ', flow_field_updated[:,-1])
|
135 |
+
flow_field_updated[int(x0-w/2):int(x0+w/2),int(y0-h/2):int(y0+h/2)] = flooded_image
|
136 |
+
|
137 |
+
|
138 |
+
# flow_field_updated is the main thing that we will use to make our predictions on -
|
139 |
+
test_img = np.expand_dims(flow_field_updated, axis = 0)
|
140 |
+
test_img = np.expand_dims(test_img, axis = 3) # Shape of test_img = (1, 128, 256)
|
141 |
+
return test_img
|
142 |
+
|
143 |
+
# Define grid points
|
144 |
+
x_points = np.linspace(0, 255, 256)
|
145 |
+
y_points = np.linspace(0, 127, 128)
|
146 |
+
X, Y = np.meshgrid(x_points, y_points)
|
147 |
+
|
148 |
+
def return_quiver_plot(u, v):
|
149 |
+
velocity = np.sqrt(u**2 + v**2)
|
150 |
+
ax = plt.subplot()
|
151 |
+
ax.imshow(velocity, origin = 'lower', extent = (0,256, 0,128), cmap = 'gray')
|
152 |
+
q = ax.quiver(X[5::8,5::8], Y[5::8,5::8], u[5::8,5::8], u[5::8,5::8], pivot = 'middle', color = 'red')
|
153 |
+
# ax.quiverkey(q, X=0.9, Y=1.05, U=2,
|
154 |
+
# label='m/s', labelpos='E')
|
155 |
+
# plt.title("Velocity distribution")
|
156 |
+
# plt.show()
|
157 |
+
return q
|
158 |
+
|
159 |
+
def squeeze_function(img):
|
160 |
+
img = np.squeeze(img, axis = 0)
|
161 |
+
img = np.squeeze(img, axis = 2)
|
162 |
+
return img
|
163 |
+
|
164 |
+
# Taking a shape from the user on sketchpad and placing it inside the fluid flow -
|
165 |
+
|
166 |
+
h, w = 48, 48 # patch_size in which the obstacle will be drawn
|
167 |
+
x0, y0 = 64, 128 # (x0, y0) = center of channel
|
168 |
+
|
169 |
+
def fill_shape_with_pixels(img): #img is taken by gradio as uint8
|
170 |
+
if img is None:
|
171 |
+
return np.zeros((h, w), dtype=np.uint8) # "No input sketch"
|
172 |
+
# Calling the the flooded image function to fill inside the obstacle
|
173 |
+
flooded_image = img_preprocess(img, h, w)
|
174 |
+
# Performing patch statching to put the obstacle at the required center position
|
175 |
+
test_img = patch_stiching(flooded_image, h, w, x0, y0)
|
176 |
+
|
177 |
+
# Loading and Compiling the Model
|
178 |
+
model_path = "/content/drive/MyDrive/Pinns_Loss_file.h5"
|
179 |
+
model = load_model(model_path, compile = False)
|
180 |
+
model.compile(loss=custom_loss, optimizer=tf.keras.optimizers.AdamW(learning_rate = 0.0001), metrics=['mae', 'cosine_proximity'])
|
181 |
+
|
182 |
+
# Making Model prediction from input sketch shape
|
183 |
+
prediction = model.predict(test_img) # (prediction.shape = (1, 128, 256, 3))
|
184 |
+
u_pred, v_pred, p_pred = np.split(prediction, 3, axis=3) # shape of u_pred, v_pred, p_pred = (1, 128, 256, 1)
|
185 |
+
|
186 |
+
# Making test_img in shape required by zero_pixel_location
|
187 |
+
req_img = squeeze_function(test_img)
|
188 |
+
|
189 |
+
# Storing the location of 0 pixel values
|
190 |
+
#req_img = req_img.astype(int)
|
191 |
+
zero_pixel_locations = np.argwhere(req_img == 0)
|
192 |
+
|
193 |
+
# Reducing the dimensions-
|
194 |
+
u_profile = u_pred[0][:,:,0] # shape of u profile to compatible shape (H, W) = (128, 256)
|
195 |
+
v_profile = v_pred[0][:,:,0]
|
196 |
+
p_profile = p_pred[0][:,:,0]
|
197 |
+
p_profile[p_profile>1.6] = 1.6
|
198 |
+
|
199 |
+
# Creating a copy of the above profiles-
|
200 |
+
u_profile_dash = np.copy(u_profile)
|
201 |
+
v_profile_dash = np.copy(v_profile)
|
202 |
+
|
203 |
+
# Creating a copy of the above profiles-
|
204 |
+
u_profile_dash_1 = np.copy(u_profile)
|
205 |
+
v_profile_dash_1 = np.copy(v_profile)
|
206 |
+
|
207 |
+
|
208 |
+
# Hollowing the obstacle out from the u and v plots. Origin of imae is lop left and origin of plot is top right
|
209 |
+
for y, x in zero_pixel_locations:
|
210 |
+
u_profile_dash[128 - y, x] = 0
|
211 |
+
v_profile_dash[128 - y, x] = 0
|
212 |
+
# will be used for image
|
213 |
+
u_profile_dash_1[y, x] = 0
|
214 |
+
v_profile_dash_1[y, x] = 0
|
215 |
+
|
216 |
+
|
217 |
+
# Quiver Plot
|
218 |
+
quiver_plot = plt.figure(figsize = (14,6), edgecolor = "gray")
|
219 |
+
velocity = np.sqrt(u_profile_dash_1**2 + v_profile_dash_1**2)
|
220 |
+
ax = plt.subplot()
|
221 |
+
ax.imshow(velocity, cmap = 'gray', extent = (0,256, 0,128))
|
222 |
+
q = ax.quiver(X[5::7,5::7], Y[5::7,5::7], u_profile_dash[5::7,5::7], v_profile_dash[5::7,5::7], pivot = 'middle', color = 'red')
|
223 |
+
ax.quiverkey(q, X=0.9, Y=1.07, U=2,
|
224 |
+
label='m/s', labelpos='E')
|
225 |
+
plt.title("Velocity distribution", fontsize = 11)
|
226 |
+
plt.xlabel("Length of Channel", fontsize = 11)
|
227 |
+
plt.ylabel("Height of Channel", fontsize = 11)
|
228 |
+
|
229 |
+
# StreamLine Plot
|
230 |
+
streamline_plot = plt.figure(figsize = (14,6), edgecolor = "gray")
|
231 |
+
plt.streamplot(X, Y, u_profile_dash, v_profile_dash, density = 3.5)
|
232 |
+
plt.axis('scaled')
|
233 |
+
plt.title("Streamline Plot", fontsize = 11)
|
234 |
+
plt.xlabel("Length of Channel", fontsize = 11)
|
235 |
+
plt.ylabel("Height of Channel", fontsize = 11)
|
236 |
+
|
237 |
+
# Colorize taken from ZoeDepth Model
|
238 |
+
u_colored = colorize(u_profile, cmap = 'jet')
|
239 |
+
#cbar_u = plt.colorbar(u_profile,fraction=0.025, pad=0.05)
|
240 |
+
v_colored = colorize(v_profile, cmap = 'jet')
|
241 |
+
#cbar_v = plt.colorbar(v_colored,fraction=0.025, pad=0.05)
|
242 |
+
p_colored = colorize(p_profile, cmap = 'jet')
|
243 |
+
#cbar_p = plt.colorbar(p_colored,fraction=0.025, pad=0.05)
|
244 |
+
|
245 |
+
|
246 |
+
return colorize(req_img, cmap = 'jet'), quiver_plot, streamline_plot, u_colored, v_colored, p_colored
|
247 |
+
|
248 |
+
# Importing gr.Blocks()
|
249 |
+
|
250 |
+
with gr.Blocks(theme="Taithrah/Minimal") as demo:
|
251 |
+
gr.Markdown(
|
252 |
+
"""
|
253 |
+
# Physics Constrained DNN for Predicting Mean Turbulent Flows
|
254 |
+
The App solves 2-D incompressible steady state NS equations for any given 2-D closed geometry. Geometry needs to be drawn around the center of the patch.\n
|
255 |
+
It predicts the streamlines,horizontal & vertical velocity profiles and the pressure profiles using a hybrid loss function.\n
|
256 |
+
""")
|
257 |
+
with gr.Row():
|
258 |
+
with gr.Column():
|
259 |
+
input_sketch = gr.Image(label = "Draw any Obstacle contour around the patch center",
|
260 |
+
tool="sketch", source="canvas", shape=(h, w), brush_radius = 3)
|
261 |
+
Process_button = gr.Button("Process Flow Parameters")
|
262 |
+
|
263 |
+
with gr.Column():
|
264 |
+
filled_channel = gr.Image(label = "Drawn object inside a Channel of dimensions 128*256", container = True)
|
265 |
+
|
266 |
+
with gr.Row():
|
267 |
+
quiver_plot = gr.Plot(label = "Velocity Distribution Around The Obstacle", scale = 2)
|
268 |
+
|
269 |
+
with gr.Row():
|
270 |
+
streamline_plot = gr.Plot(label = "Stream Lines Around The Obstacle", scale = 2)
|
271 |
+
|
272 |
+
with gr.Row():
|
273 |
+
u_image = gr.Image(label = "Horizontal Velocity")
|
274 |
+
v_image = gr.Image(label = "Vertical Velocity")
|
275 |
+
p_image = gr.Image(label = "Pressure")
|
276 |
+
|
277 |
+
|
278 |
+
Process_button.click(fn=fill_shape_with_pixels, inputs=input_sketch, outputs=[filled_channel, quiver_plot, streamline_plot, u_image, v_image, p_image])
|
279 |
+
|
280 |
+
demo.launch(debug=True, server_name = "0.0.0.0", share = True, inline = False)
|