NextJump / app.py
dylanplummer's picture
update mark thresh
cf90a22
import gradio as gr
import numpy as np
from PIL import Image
import os
import cv2
import math
import spaces
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
from scipy.signal import medfilt, find_peaks
from functools import partial
from passlib.hash import pbkdf2_sha256
from tqdm import tqdm
import pandas as pd
import plotly.express as px
import onnxruntime as ort
import torch
from torchvision import transforms
import torchvision.transforms.functional as F
from huggingface_hub import hf_hub_download
from huggingface_hub import HfApi
plt.style.use('dark_background')
onnx_file = hf_hub_download(repo_id="dylanplummer/ropenet", filename="nextjump.onnx", repo_type="model", token=os.environ['DATASET_SECRET'])
# model_xml = hf_hub_download(repo_id="dylanplummer/ropenet", filename="model.xml", repo_type="model", token=os.environ['DATASET_SECRET'])
# hf_hub_download(repo_id="dylanplummer/ropenet", filename="model.mapping", repo_type="model", token=os.environ['DATASET_SECRET'])
#model_xml = "model_ir/model.xml"
# ie = Core()
# model_ir = ie.read_model(model=model_xml)
# config = {"PERFORMANCE_HINT": "LATENCY"}
# compiled_model_ir = ie.compile_model(model=model_ir, device_name="CPU", config=config)
class SquarePad:
# https://discuss.pytorch.org/t/how-to-resize-and-pad-in-a-torchvision-transforms-compose/71850/9
def __call__(self, image):
w, h = image.size
max_wh = max(w, h)
hp = int((max_wh - w) / 2)
vp = int((max_wh - h) / 2)
padding = (hp, vp, hp, vp)
return F.pad(image, padding, 0, 'constant')
def sigmoid(x):
return 1 / (1 + np.exp(-x))
@spaces.GPU()
def inference(x, count_only_api, api_key,
img_size=288, seq_len=64, stride_length=32, stride_pad=3, batch_size=4,
miss_threshold=0.8, marks_threshold=0.5, median_pred_filter=True, center_crop=True, both_feet=True,
api_call=False,
progress=gr.Progress()):
progress(0, desc="Starting...")
# check if GPU is available
if torch.cuda.is_available():
providers = [("CUDAExecutionProvider", {"device_id": torch.cuda.current_device(),
"user_compute_stream": str(torch.cuda.current_stream().cuda_stream)})]
sess_options = ort.SessionOptions()
ort_sess = ort.InferenceSession(onnx_file, sess_options=sess_options, providers=providers)
else:
ort_sess = ort.InferenceSession(onnx_file)
#api = HfApi(token=os.environ['DATASET_SECRET'])
#out_file = str(uuid.uuid1())
has_access = False
if api_call:
has_access = pbkdf2_sha256.verify(os.environ['DEV_API_TOKEN'], api_key)
if not has_access:
return "Invalid API Key"
cap = cv2.VideoCapture(x)
length = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
period_length_overlaps = np.zeros(length + seq_len)
fps = int(cap.get(cv2.CAP_PROP_FPS))
seconds = length / fps
all_frames = []
frame_i = 1
while cap.isOpened():
ret, frame = cap.read()
if ret is False:
frame = all_frames[-1] # padding will be with last frame
break
frame = cv2.cvtColor(np.uint8(frame), cv2.COLOR_BGR2RGB)
img = Image.fromarray(frame)
all_frames.append(img)
frame_i += 1
cap.release()
length = len(all_frames)
period_lengths = np.zeros(len(all_frames) + seq_len + stride_length)
periodicities = np.zeros(len(all_frames) + seq_len + stride_length)
full_marks = np.zeros(len(all_frames) + seq_len + stride_length)
event_type_logits = np.zeros((len(all_frames) + seq_len + stride_length, 7))
period_length_overlaps = np.zeros(len(all_frames) + seq_len + stride_length)
event_type_logit_overlaps = np.zeros((len(all_frames) + seq_len + stride_length, 7))
for _ in range(seq_len + stride_length): # pad full sequence
all_frames.append(all_frames[-1])
batch_list = []
idx_list = []
for i in tqdm(range(0, length + stride_length - stride_pad, stride_length)):
batch = all_frames[i:i + seq_len]
Xlist = []
print('Preprocessing...')
for img in batch:
transforms_list = []
# if center_crop:
# if width > height:
# transforms_list.append(transforms.Resize((int(width / (height / img_size)), img_size)))
# else:
# transforms_list.append(transforms.Resize((img_size, int(height / (width / img_size)))))
# transforms_list.append(transforms.CenterCrop((img_size, img_size)))
# else:
transforms_list.append(SquarePad())
transforms_list.append(transforms.Resize((img_size, img_size), interpolation=Image.BICUBIC))
transforms_list += [
transforms.ToTensor()]
#transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])]
preprocess = transforms.Compose(transforms_list)
frameTensor = preprocess(img).unsqueeze(0)
Xlist.append(frameTensor)
if len(Xlist) < seq_len:
for _ in range(seq_len - len(Xlist)):
Xlist.append(Xlist[-1])
X = torch.cat(Xlist)
X *= 255
batch_list.append(X.unsqueeze(0))
idx_list.append(i)
print('Running inference...')
if len(batch_list) == batch_size:
batch_X = torch.cat(batch_list)
outputs = ort_sess.run(None, {'video': batch_X.numpy()})
y1pred = outputs[0]
y2pred = outputs[1]
y3pred = outputs[2]
y4pred = outputs[3]
for y1, y2, y3, y4, idx in zip(y1pred, y2pred, y3pred, y4pred, idx_list):
periodLength = y1.squeeze()
periodicity = y2.squeeze()
marks = y3.squeeze()
event_type = y4.squeeze()
period_lengths[idx:idx+seq_len] += periodLength
periodicities[idx:idx+seq_len] += periodicity
full_marks[idx:idx+seq_len] += marks
event_type_logits[idx:idx+seq_len] += event_type
period_length_overlaps[idx:idx+seq_len] += 1
event_type_logit_overlaps[idx:idx+seq_len] += 1
batch_list = []
idx_list = []
progress(i / (length + stride_length - stride_pad), desc="Processing...")
if len(batch_list) != 0: # still some leftover frames
while len(batch_list) != batch_size:
batch_list.append(batch_list[-1])
idx_list.append(idx_list[-1])
batch_X = torch.cat(batch_list)
outputs = ort_sess.run(None, {'video': batch_X.numpy()})
y1pred = outputs[0]
y2pred = outputs[1]
y3pred = outputs[2]
y4pred = outputs[3]
for y1, y2, y3, y4, idx in zip(y1pred, y2pred, y3pred, y4pred, idx_list):
periodLength = y1.squeeze()
periodicity = y2.squeeze()
marks = y3.squeeze()
event_type = y4.squeeze()
period_lengths[idx:idx+seq_len] += periodLength
periodicities[idx:idx+seq_len] += periodicity
full_marks[idx:idx+seq_len] += marks
event_type_logits[idx:idx+seq_len] += event_type
period_length_overlaps[idx:idx+seq_len] += 1
event_type_logit_overlaps[idx:idx+seq_len] += 1
periodLength = np.divide(period_lengths, period_length_overlaps, where=period_length_overlaps!=0)[:length]
periodicity = np.divide(periodicities, period_length_overlaps, where=period_length_overlaps!=0)[:length]
full_marks = np.divide(full_marks, period_length_overlaps, where=period_length_overlaps!=0)[:length]
per_frame_event_type_logits = np.divide(event_type_logits, event_type_logit_overlaps, where=event_type_logit_overlaps!=0)[:length]
event_type_logits = np.mean(per_frame_event_type_logits, axis=0)
# softmax of event type logits
event_type_probs = np.exp(event_type_logits) / np.sum(np.exp(event_type_logits))
per_frame_event_types = np.argmax(per_frame_event_type_logits, axis=1)
if median_pred_filter:
periodicity = medfilt(periodicity, 5)
periodLength = medfilt(periodLength, 5)
periodicity = sigmoid(periodicity)
full_marks = sigmoid(full_marks)
#full_marks_mask = np.int32(full_marks > marks_threshold)
pred_marks_peaks, _ = find_peaks(full_marks, distance=3, height=marks_threshold)
full_marks_mask = np.zeros(len(full_marks))
full_marks_mask[pred_marks_peaks] = 1
periodicity_mask = np.int32(periodicity > miss_threshold)
numofReps = 0
count = []
for i in range(len(periodLength)):
if periodLength[i] < 2 or periodicity_mask[i] == 0:
numofReps += 0
elif full_marks_mask[i]: # high confidence mark detected
if math.modf(numofReps)[0] < 0.2: # probably false positive/late detection
numofReps = float(int(numofReps))
else:
numofReps = float(int(numofReps) + 1.01) # round up
else:
numofReps += max(0, periodicity_mask[i]/(periodLength[i]))
count.append(round(float(numofReps), 2))
count_pred = count[-1]
marks_count_pred = 0
for i in range(len(full_marks) - 1):
# if a jump was counted, and periodicity is high, and the next frame was not counted (to avoid double counting)
if full_marks_mask[i] > 0 and periodicity_mask[i] > 0 and full_marks_mask[i + 1] == 0:
marks_count_pred += 1
if not both_feet:
count_pred = count_pred / 2
marks_count_pred = marks_count_pred / 2
count = np.array(count) / 2
confidence = (np.mean(periodicity[periodicity > miss_threshold]) - miss_threshold) / (1 - miss_threshold)
self_err = abs(count_pred - marks_count_pred)
self_pct_err = self_err / count_pred
total_confidence = confidence * (1 - self_pct_err)
if both_feet:
count_msg = f"## Reps Count (both feet): {count_pred:.1f}, Marks Count (both feet): {marks_count_pred:.1f}, Confidence: {total_confidence:.2f}"
else:
count_msg = f"## Predicted Count (one foot): {count_pred:.1f}, Marks Count (one foot): {marks_count_pred:.1f}, Confidence: {total_confidence:.2f}"
if api_call:
if count_only_api:
return f"{count_pred:.2f} (conf: {total_confidence:.2f})"
else:
return np.array2string(periodLength, formatter={'float_kind':lambda x: "%.2f" % x}).replace('\n', ''), \
np.array2string(periodicity, formatter={'float_kind':lambda x: "%.2f" % x}).replace('\n', ''), \
np.array2string(full_marks, formatter={'float_kind':lambda x: "%.2f" % x}).replace('\n', ''), \
f"reps: {count_pred:.2f}, marks: {marks_count_pred:.1f}, confidence: {total_confidence:.2f}", \
f"single_rope_speed: {event_type_probs[0]:.3f}, double_dutch: {event_type_probs[1]:.3f}, double_unders: {event_type_probs[2]:.3f}, single_bounce: {event_type_probs[3]:.3f}"
jumps_per_second = np.clip(1 / ((periodLength / fps) + 0.01), 0, 10)
jumping_speed = np.copy(jumps_per_second)
misses = periodicity < miss_threshold
jumps_per_second[misses] = 0
frame_type = np.array(['miss' if miss else 'frame' for miss in misses])
frame_type[full_marks > marks_threshold] = 'jump'
per_frame_event_types = np.clip(per_frame_event_types, 0, 6) / 6
df = pd.DataFrame.from_dict({'period length': periodLength,
'jumping speed': jumping_speed,
'jumps per second': jumps_per_second,
'periodicity': periodicity,
'miss': misses,
'frame_type': frame_type,
'event_type': per_frame_event_types,
'jumps': full_marks,
'jumps_size': (full_marks + 0.05) * 10,
'miss_size': np.clip((1 - periodicity) * 0.9 + 0.1, 1, 8),
'seconds': np.linspace(0, seconds, num=len(periodLength))})
event_type_tick_vals = np.linspace(0, 1, num=7)
event_type_colors = ['red', 'orange', 'green', 'blue', 'purple', 'pink', 'black']
fig = px.scatter(data_frame=df,
x='seconds',
y='jumps per second',
#symbol='frame_type',
#symbol_map={'frame': 'circle', 'miss': 'circle-open', 'jump': 'triangle-down'},
color='event_type',
size='jumps_size',
size_max=8,
color_continuous_scale=[(t, c) for t, c in zip(event_type_tick_vals, event_type_colors)],
range_color=(0,1),
title="Jumping speed (jumps-per-second)",
trendline='rolling',
trendline_options=dict(window=16),
trendline_color_override="goldenrod",
trendline_scope='overall',
template="plotly_dark")
fig.update_layout(legend=dict(
orientation="h",
yanchor="bottom",
y=0.98,
xanchor="right",
x=1,
font=dict(
family="Courier",
size=12,
color="black"
),
bgcolor="AliceBlue",
),
paper_bgcolor='rgba(0,0,0,0)',
plot_bgcolor='rgba(0,0,0,0)'
)
# remove white outline from marks
fig.update_traces(marker_line_width = 0)
fig.update_layout(coloraxis_colorbar=dict(
tickvals=event_type_tick_vals,
ticktext=['single<br>rope', 'double<br>dutch', 'double<br>unders', 'single<br>bounces', 'double<br>bounces', 'triple<br>unders', 'other'],
title='event type'
))
hist = px.histogram(df,
x="jumps per second",
template="plotly_dark",
marginal="box",
histnorm='percent',
title="Distribution of jumping speed (jumps-per-second)")
# make a bar plot of the event type distribution
bar = px.bar(x=['single rope', 'double dutch', 'double unders', 'single bounces', 'double bounces', 'triple unders', 'other'],
y=event_type_probs,
template="plotly_dark",
title="Event Type Distribution",
labels={'x': 'event type', 'y': 'probability'},
range_y=[0, 1])
return count_msg, fig, hist, bar
DESCRIPTION = '# NextJump 🦘'
DESCRIPTION += '\n## AI Counting for Competitive Jump Rope'
DESCRIPTION += '\nDemo created by [Dylan Plummer](https://dylan-plummer.github.io/). Check out the [NextJump iOS app](https://apps.apple.com/us/app/nextjump-jump-rope-counter/id6451026115).'
with gr.Blocks(theme='WeixuanYuan/Soft_dark') as demo:
gr.Markdown(DESCRIPTION)
in_video = gr.PlayableVideo(label="Input Video", elem_id='input-video', format='mp4',
width=400, height=400, interactive=True, container=True,
max_length=150)
with gr.Row():
run_button = gr.Button(value="Run", elem_id='run-button', scale=1)
api_dummy_button = gr.Button(value="Run (No Viz)", elem_id='count-only', visible=False, scale=2)
count_only = gr.Checkbox(label="Count Only", visible=False)
api_token = gr.Textbox(label="API Key", elem_id='api-token', visible=False)
with gr.Column(elem_id='output-video-container'):
with gr.Row():
with gr.Column():
out_text = gr.Markdown(label="Predicted Count", elem_id='output-text')
period_length = gr.Textbox(label="Period Length", elem_id='period-length', visible=False)
periodicity = gr.Textbox(label="Periodicity", elem_id='periodicity', visible=False)
#with gr.Column(min_width=480):
#out_video = gr.PlayableVideo(label="Output Video", elem_id='output-video', format='mp4')
with gr.Row():
out_plot = gr.Plot(label="Jumping Speed", elem_id='output-plot')
with gr.Row():
with gr.Column():
out_hist = gr.Plot(label="Speed Histogram", elem_id='output-hist')
with gr.Column():
out_event_type_dist = gr.Plot(label="Event Type Distribution", elem_id='output-event-type-dist')
with gr.Accordion(label="Instructions and more information", open=False):
instructions = "## Instructions:"
instructions += "\n* Upload a video and click 'Run' to get a prediction of the number of jumps (either one foot, or both). This could take a couple minutes!"
instructions += "\n\n## Tips (optional):"
instructions += "\n* Trim the video to start and end of the event"
instructions += "\n* Frame the jumper fully, in the center of the frame"
instructions += "\n* Videos are automatically resized, so higher resolution will not help, but a closer framing of the jumper might help. Try cropping the video differently."
gr.Markdown(instructions)
faq = "## FAQ:"
faq += "\n* **Q:** Does the model recognize misses?\n * **A:** Yes, but if it fails, you can try tuning the miss threshold slider to make it more sensitive."
faq += "\n* **Q:** Does the model recognize double dutch?\n * **A:** Yes, but it is trained on a smaller set of double dutch videos, so it may not work perfectly."
faq += "\n* **Q:** Does the model recognize double unders\n * **A:** Yes, but it is trained on a smaller set of double under videos, so it may not work perfectly. It is also trained to count the rope, not the jumps so you will need to divide the count by 2 to get the traditional double under count."
faq += "\n* **Q:** Does the model count both feet?\n * **A:** Yes, it counts every time the rope goes around no matter the event."
gr.Markdown(faq)
demo_inference = partial(inference, count_only_api=False, api_key=None)
gr.Examples(examples=[
[os.path.join(os.path.dirname(__file__), "files", "dylan.mp4")],
#[os.path.join(os.path.dirname(__file__), "files", "train14.mp4")],
#[os.path.join(os.path.dirname(__file__), "files", "train_17.mp4")],
#[os.path.join(os.path.dirname(__file__), "files", "train13.mp4")],
#[os.path.join(os.path.dirname(__file__), "files", "train_213.mp4")],
#[os.path.join(os.path.dirname(__file__), "files", "train_156.mp4")],
#[os.path.join(os.path.dirname(__file__), "files", "train_202.mp4")],
#[os.path.join(os.path.dirname(__file__), "files", "train_57.mp4")],
#[os.path.join(os.path.dirname(__file__), "files", "train_95.mp4")],
#[os.path.join(os.path.dirname(__file__), "files", "train_253.mp4")],
#[os.path.join(os.path.dirname(__file__), "files", "train_66.mp4")],
#[os.path.join(os.path.dirname(__file__), "files", "train_21.mp4")]
],
inputs=[in_video],
outputs=[out_text, out_plot, out_hist, out_event_type_dist],
fn=demo_inference, cache_examples=os.getenv('SYSTEM') == 'spaces')
run_button.click(demo_inference, [in_video], outputs=[out_text, out_plot, out_hist, out_event_type_dist])
api_inference = partial(inference, api_call=True)
api_dummy_button.click(api_inference, [in_video, count_only, api_token], outputs=[period_length], api_name='inference')
if __name__ == "__main__":
demo.queue(api_open=True, max_size=15).launch(share=False)