def main(args):
global g_output_dir
g_output_dir = pj( ut.get_repo_dir(), "out/hough" )
img_dir = pj( ut.get_repo_dir(), "data/hough" )
img_fn = args.image_file
img = cv2.imread( pj(img_dir, img_fn) )
edged = ut.detect_edges(img)
edged_fn = img_fn[:-4] + "_edged.png"
cv2.imwrite( pj(g_output_dir, edged_fn), edged )
hough_lines(np.copy(edged), np.copy(img), img_fn)
hough_lines_p(np.copy(edged), np.copy(img), img_fn)
bw = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
hough_circles(bw, np.copy(img), img_fn)
elif self.tuning == TuningParam.lamb:
assert self.use_ppc, 'need to use ppc to tune lambda'
*_, ppc_ind = self.get_ppc_off(frame,
p_int) # ppc_ind: (B, F, T, J)
ppc_ind_throw = torch.gather(ppc_ind, 2, tof).squeeze() # B, F, J
ppc_ind_throw = torch.gather(ppc_ind_throw, 1,
ball_field_ind).squeeze() # B, J
return ppc_ind_throw
if __name__ == '__main__':
# args
parser = argparse.ArgumentParser()
parser.add_argument('-d',
'--data_dir',
default=os.path.join(get_repo_dir(), 'data'),
help='path to directory with data files')
parser.add_argument('-j',
'--num_workers',
default=4,
help='number of data loader workers',
type=int)
parser.add_argument('-b',
'--batch_size',
default=4,
help='batch size to use',
type=int)
parser.add_argument('-e',
'--epochs',
default=5,
help='number of training epochs',
self.input_noise_std)
out = self.net(net_input)
total_loss = self.loss(out, image)
total_loss.backward()
if iter_num == self.iters[self.iter_idx]:
self.iter_idx += 1
self.save(iter_num, torch_to_np(out))
if self.iter_idx == len(self.iters):
self.iter_idx = 0
def run_all(self):
for image, target in tqdm(self.dataloader):
self.run(image.to(self.device))
data = CIFAR10(root=os.path.join(get_repo_dir(), 'data'),
train=True,
download=True,
transform=identity_transform)
save_dir = os.path.join(get_repo_dir(), 'data/transforms')
os.makedirs(save_dir, exist_ok=True)
#first = True
#hashes = set()
#if first:
# for im, t in data:
# hashes.add(image_hash(im))
# assert len(data) == len(hashes), 'colliding hash'
# pickle.dump(hashes, open(os.path.join(save_dir, 'hashes.pkl'), 'wb'))
#else:
# hashes_old = pickle.load(open(os.path.join(save_dir, 'hashes.pkl'), 'rb'))
def play_eppa_cpu(track_df,
game_id,
play_id,
viz_df=False,
save_np=False,
stats_df=False,
viz_true_proj=False,
save_all_dfs=False,
out_dir_path=join(get_repo_dir(), 'output/{}'),
simulatePass=None):
play_df = track_df[(track_df.playId == play_id)
& (track_df.gameId == game_id)].sort_values(
by='frameId')
events = set(play_df.event.unique())
if 'pass_forward' not in events:
raise ValueError('No Pass Forward in play')
if 'fumble' in events:
raise ValueError('Fumble in play')
ball_snap_frame = play_df.loc[
(play_df.nflId == 0) & (play_df.event == 'ball_snap')].frameId.iloc[0]
pass_forward_frame = play_df.loc[(play_df.nflId == 0) & (
(play_df.event == 'pass_forward')
| (play_df.event == 'pass_shovel'))].frameId.sort_values().iloc[0]
play_df['frames_since_snap'] = play_df.frameId - ball_snap_frame
pass_arriveds = play_df.loc[(play_df.nflId == 0) &
(play_df.event == 'pass_arrived'),
'frameId'].to_list()
true_pass_found = len(pass_arriveds) > 0
if true_pass_found:
true_pass_completed = ('pass_outcome_caught'
in events) or ('pass_outcome_touchdown'
in events)
pass_arrived_frame = pass_arriveds[0]
true_T_frames = pass_arrived_frame - pass_forward_frame
true_x, true_y = play_df.loc[(play_df.nflId == 0) &
(play_df.event == 'pass_arrived'),
['x', 'y']].iloc[0].to_numpy(dtype=dt)
print(
f"True Pass: t: {pass_forward_frame} x:{true_x} y:{true_y} T:{true_T_frames/10}"
)
true_T_idx = np.rint(true_T_frames).astype(int) - 1
true_x_idx = np.clip((true_x).astype(int), 0,
len(xx[0]) - 1) # this is a bug
true_y_idx = np.clip((true_y).astype(int) + 1, 0,
len(xx) - 1) # this is a bug
true_f_idx = np.ravel_multi_index((true_y_idx, true_x_idx), xx.shape)
(true_x_idxd,
true_y_idxd), true_T_idxd = field_locs[true_f_idx], T[true_T_idx]
print(
f"True Pass idxs: t: {pass_forward_frame} x:{true_x_idx} y:{true_y_idx} f:{true_f_idx} T:{true_T_idx}"
)
print(
f"True Pass idxd: t: {pass_forward_frame} x:{true_x_idxd} y:{true_y_idxd} T:{true_T_idxd}"
)
else:
print(f"No True Pass Found: No pass_arrived event \n")
if simulatePass != None:
sim_frame, true_x, true_y, true_T = simulatePass
print(f"Simming Pass: t: {sim_frame} x:{true_x} y:{true_y} T:{true_T}")
true_T_idx = np.rint(true_T * 10).astype(int) - 1
true_x_idx = np.clip(dt(true_x).astype(int), 0,
len(xx[0]) - 1) # this is a bug
true_y_idx = np.clip(dt(true_y).astype(int) + 1, 0,
len(xx) - 1) # this is a bug
true_f_idx = np.ravel_multi_index((true_y_idx, true_x_idx), xx.shape)
(true_x_idxd,
true_y_idxd), true_T_idxd = field_locs[true_f_idx], T[true_T_idx]
print(
f"Sim Pass idxs: t: {pass_forward_frame} x:{true_x_idx} y:{true_y_idx} f:{true_f_idx} T:{true_T_idx}"
)
print(
f"Sim Pass idxd: t: {pass_forward_frame} x:{true_x_idxd} y:{true_y_idxd} T:{true_T_idxd}"
)
# per play epa model
def getEPAModel():
epvals = np.array([7, -7, 3, -3, 2, -2, 0])
joined_df = pbp_joined[(pbp_joined.playId == play_id)
& (pbp_joined.gameId == game_id)]
epa_df = pd.DataFrame({'play_endpoint_x': x})
test = {}
for feat in epa_model.feature_names:
test[feat] = [joined_df.iloc[0][feat]]
epa_df[feat] = joined_df.iloc[0][feat]
first_df = pd.DataFrame(test)
dtest = treelite_runtime.Batch.from_npy2d(
first_df[cols_when_model_builds_ep].values)
ypred = epa_predictor.predict(dtest)
ep = np.sum(ypred * epvals, axis=1)
epa_df["before_play_ep"] = ep[0]
###### EP FOR A INCOMPLETE #######
epa_df_incomp = epa_df.copy(deep=True)
old_down = joined_df.iloc[0]['down_x']
epa_df_incomp['isFirstDown'] = 0
for d in range(1, 6):
epa_df_incomp['down%d' % d] = 1 if (d == old_down + 1) else 0
# offense is other team inverted if turnover on downs
epa_df_incomp['yardline_100'] = np.where(
epa_df_incomp.down5 == 1, 100 - epa_df_incomp['yardline_100'],
epa_df_incomp['yardline_100'])
epa_df_incomp['ydstogo'] = np.where(epa_df_incomp.down5 == 1, 10,
epa_df_incomp['ydstogo'])
epa_df_incomp['down1'] = np.where(epa_df_incomp.down5 == 1, 1,
epa_df_incomp['down1'])
dtest = treelite_runtime.Batch.from_npy2d(
epa_df_incomp[cols_when_model_builds_ep].values)
ypred = epa_predictor.predict(dtest)
ep = np.sum(ypred * epvals, axis=1)
epa_df['xep_inc'] = ep
epa_df['xepa_inc'] = np.where(
epa_df_incomp.down5 == 1,
-epa_df['xep_inc'] - epa_df_incomp['before_play_ep'],
epa_df['xep_inc'] - epa_df_incomp['before_play_ep'])
###### EP FOR A CATCH #######
epa_df['los'] = 110 - epa_df['yardline_100']
epa_df['first_down_line'] = epa_df['los'] + epa_df["ydstogo"]
epa_df['expected_end_up_line'] = epa_df['play_endpoint_x']
epa_df['isFirstDown'] = 0
epa_df['isFirstDown'] = np.where(
epa_df['expected_end_up_line'] >= epa_df['first_down_line'], 1, 0)
epa_df['yardline_100'] = np.round(110 - epa_df['expected_end_up_line'])
epa_df['yardline_100'] = np.clip(epa_df['yardline_100'], 0, 100)
epa_df['ydstogo'] = epa_df['first_down_line'] - epa_df[
'expected_end_up_line']
epa_df['ydstogo'] = np.where(epa_df['isFirstDown'] == 1, 10,
epa_df['ydstogo'])
old_down = joined_df.iloc[0]['down_x']
for d in range(1, 6):
epa_df['down%d' % d] = 1 if (d == old_down + 1) else 0
epa_df['down1'] = np.where(epa_df['isFirstDown'] == 1, 1,
epa_df['down1'])
epa_df['down2'] = np.where(epa_df['isFirstDown'] == 1, 0,
epa_df['down2'])
epa_df['down3'] = np.where(epa_df['isFirstDown'] == 1, 0,
epa_df['down3'])
epa_df['down4'] = np.where(epa_df['isFirstDown'] == 1, 0,
epa_df['down4'])
epa_df['down5'] = np.where(
epa_df['isFirstDown'] == 1, 0,
epa_df['down5']) # down 5 means turnover on downs
# offense is other team inverted if turnover on downs
epa_df['yardline_100'] = np.where(epa_df.down5 == 1,
100 - epa_df['yardline_100'],
epa_df['yardline_100'])
epa_df['ydstogo'] = np.where(epa_df.down5 == 1, 10, epa_df['ydstogo'])
epa_df['down1'] = np.where(epa_df.down5 == 1, 1, epa_df['down1'])
### UPDATE EPA VARIABLES ###
dtest = treelite_runtime.Batch.from_npy2d(
epa_df[cols_when_model_builds_ep].values)
ypred = epa_predictor.predict(dtest)
ep = np.sum(ypred * epvals, axis=1)
epa_df['xep'] = ep # ep after play
# SCORE SAFETIES
epa_df['xep'] = np.where(epa_df['play_endpoint_x'] <= 10, -2,
epa_df['xep'])
# SCORE Tds
epa_df['xep'] = np.where(epa_df['play_endpoint_x'] >= 110, 7,
epa_df['xep'])
epa_df['xepa'] = np.where(
epa_df.down5 == 1, -epa_df['xep'] - epa_df['before_play_ep'],
epa_df['xep'] - epa_df['before_play_ep']) # if turnover
only_vals = epa_df[[
"play_endpoint_x", "xep", "xepa", "xep_inc", "xepa_inc"
]].rename(columns={
'xep': 'xep_comp',
'xepa': 'xepa_comp'
}) # THIS CONTAINS THE EPA VALUES BASED ON PLAY ENDPOINT
# breakpoint()
return only_vals
epa_df = getEPAModel()
# input: tracking data for a single frame
# output: frame_eppa (F, T)... (writes intermediate F,T to disk)
def frame_eppa(frame_id):
nonlocal play_df
frame_df = play_df.loc[play_df.frameId == frame_id].reset_index()
ball_start = frame_df.loc[frame_df.position == 'QB',
['x', 'y']].iloc[0].to_numpy(dtype=dt)
t = frame_df.frames_since_snap.iloc[0]
frame_df = frame_df.loc[(frame_df.nflId != 0) & (
frame_df.position != 'QB')] # remove ball and qb from df
reach_vecs = (field_locs - ball_start).astype(dt) # (F, 2)
frame_df = frame_df.drop_duplicates(
subset='nflId').sort_values('nflId').reset_index()
# project motion by reaction time
frame_df['v_x_r'] = frame_df.a_x * params.reax_t + frame_df.v_x
frame_df['v_y_r'] = frame_df.a_y * params.reax_t + frame_df.v_y
frame_df['v_r_mag'] = np.linalg.norm(np.array(
[frame_df.v_x_r, frame_df.v_y_r], dtype=dt),
axis=0)
frame_df['v_r_theta'] = np.arctan2(frame_df.v_y_r, frame_df.v_x_r)
frame_df[
'x_r'] = frame_df.x + frame_df.v_x * params.reax_t + 0.5 * frame_df.a_x * params.reax_t**2
frame_df[
'y_r'] = frame_df.y + frame_df.v_y * params.reax_t + 0.5 * frame_df.a_y * params.reax_t**2
player_teams = frame_df['team_pos'].to_numpy() # J,
player_team_names = frame_df['teamAbbr'].to_numpy() # J,
player_ids = frame_df['nflId'].to_numpy()
player_names = frame_df['displayName'].to_numpy()
reaction_player_locs = frame_df[['x_r',
'y_r']].to_numpy(dtype=dt) # (J, 2)
reaction_player_vels = frame_df[['v_x_r',
'v_y_r']].to_numpy(dtype=dt) # (J,2)
# intercept vector between each player and field location
int_d_vec = field_locs[:, None, :] - reaction_player_locs # F, J, 2
int_d_mag = np.linalg.norm(int_d_vec, axis=2) # F, J
int_d_theta = np.arctan2(int_d_vec[..., 1], int_d_vec[..., 0])
# projecting player velocity on d_vec to get initial speed along d_vec
int_s0 = np.clip(
np.sum(int_d_vec * reaction_player_vels, axis=2) / int_d_mag,
-params.s_max, params.s_max) # F, J,
# calculate time to int to ball loc on d_vec based on phys model
t_lt_smax = (params.s_max - int_s0) / params.a_max # F, J,
d_lt_smax = t_lt_smax * ((int_s0 + params.s_max) / 2) # F, J,
# if accelerating would overshoot, then t = -v0/a + sqrt(v0^2/a^2 + 2x/a) (from kinematics)
t_lt_smax = np.where(
d_lt_smax > int_d_mag, -int_s0 / params.a_max +
np.sqrt((int_s0 / params.a_max)**2 + 2 * int_d_mag / params.a_max),
t_lt_smax)
d_lt_smax = np.clip(d_lt_smax, 0, int_d_mag)
d_at_smax = int_d_mag - d_lt_smax # F, J,
t_at_smax = d_at_smax / params.s_max # F, J,
t_tot = t_lt_smax + t_at_smax + params.reax_t # F, J,
# int success if T-t_tot = dT < 0. Put through sigmoid to add temporal uncertainty around
int_dT = T[None, :, None] - t_tot[:, None, :] # F, T, J
p_int = (1 / (1. + np.exp(
-np.pi / np.sqrt(3.0) / params.tti_sigma * int_dT, dtype=dt))
) # F, T, J
p_int_def = 1 - np.prod(
(1 - p_int[:, :, (player_teams == 'DEF')]), axis=-1) # F, T
p_int_off = 1 - np.prod(
(1 - p_int[:, :, (player_teams == 'OFF')]), axis=-1) # F, T
p_int_off_adj = p_int_off * (1 - p_int_def)
# adjusting indiv p_int to renormalize sum
p_int_ind = p_int.copy()
p_int_ind[:, :, player_teams == 'DEF'] *= (
p_int_def / np.sum(p_int_ind, where=(player_teams == 'DEF')))[...,
None]
p_int_ind[:, :, player_teams == 'OFF'] *= (
p_int_off_adj /
np.sum(p_int_ind, where=(player_teams == 'OFF')))[..., None]
# projected locations at T (F, T, J)
d_proj = np.select([
T[None, :, None] <= params.reax_t, T[None, :, None] <=
(t_lt_smax + params.reax_t)[:, None, :], True
], [
0, (int_s0[:, None, :] *
(T[None, :, None] - params.reax_t)) + 0.5 * params.a_max *
(T[None, :, None] - params.reax_t)**2,
d_lt_smax[:, None, :] + params.s_max *
(T[None, :, None] - t_lt_smax[:, None, :] - params.reax_t)
])
d_proj = np.minimum(
d_proj,
int_d_mag[:,
None, :]) # no player is going to purposefully overshoot
s_proj = np.select([
T[None, :, None] <= params.reax_t, T[None, :, None] <=
(t_lt_smax + params.reax_t)[:, None, :], T[None, :, None] <=
(t_lt_smax + t_at_smax + params.reax_t)[:, None, :], True
], [
int_s0[:, None, :], int_s0[:, None, :] + params.a_max *
(T[None, :, None] - params.reax_t), params.s_max, params.s_max
])
x_proj = reaction_player_locs[None, None, :, 0] + d_proj * np.cos(
int_d_theta[:, None, :]) # (F, T, J)
y_proj = reaction_player_locs[None, None, :, 1] + d_proj * np.sin(
int_d_theta[:, None, :]) # (F, T, J)
v_x_proj = s_proj * np.cos(int_d_theta[:, None, :]) # (F, T, J)
v_y_proj = s_proj * np.sin(int_d_theta[:, None, :]) # (F, T, J)
# input: qb_loc (2,), t=frames_after_snap (int)
# output: (P(L,T)|t) int probability of each pass (F, T)
def get_hist_trans_prob():
""" P(L|t) """
ball_start_idx = np.rint(ball_start).astype(int)
# mask for zeroing out parts of the field that are too far to be thrown to per the L_given_t model
L_t_mask = np.zeros_like(xx, dtype=dt) # (Y, X)
L_t_mask[max(ball_start_idx[1] +
y_min, 0):min(ball_start_idx[1] + y_max,
len(y) - 1),
max(ball_start_idx[0] +
x_min, 0):min(ball_start_idx[0] + x_max,
len(x) - 1)] = 1.
L_t_mask = L_t_mask.flatten() # (F,)
# # we clip reach vecs to be used to index into L_given_t.
# # eg if qb is far right, then the left field will be clipped to y=-39 and later zeroed out
# reach_vecs_int = np.rint(reach_vecs).astype(int)
# clipped_reach_vecs = np.stack((np.clip(reach_vecs_int[:,0], x_min, x_max),
# np.clip(-reach_vecs_int[:,1], y_min, y_max))) # (2, F)
# t_i = max(t-t_min, 0)
# L_given_t = L_given_ts[t_i, clipped_reach_vecs[1]-y_min, clipped_reach_vecs[0]-x_min] * L_t_mask # (F,) ; index with y and then x
L_given_t = L_t_mask # changed L_given_t to uniform after discussion
L_given_t /= L_given_t.sum(
) # renormalize since part of L|t may have been off field
""" P(T|L) """
# we find T|L for sufficiently close spots (1 < L <= 60)
reach_dist_int = np.rint(np.linalg.norm(
reach_vecs, axis=1)).astype(int) # (F,)
reach_dist_in_bounds_idx = (reach_dist_int > 1) & (reach_dist_int
<= 60)
reach_dist_in_bounds = reach_dist_int[reach_dist_in_bounds_idx]
# yards 1-60 map to indices 0-59 so subtract 1 to get index
# spots with dist 0 shouldn't wrap around to index -1 so clip
T_given_L_subset = T_given_Ls[np.clip(
reach_dist_in_bounds - 1, 0, 59
)] # (F~, T) ; F~ is subset of F that is in [1, 60] yds from ball
T_given_L = np.zeros((len(field_locs), len(T)), dtype=dt) # (F, T)
# fill in the subset of values computed above
T_given_L[reach_dist_in_bounds_idx] = T_given_L_subset
L_T_given_t = L_given_t[:, None] * T_given_L # (F, T)
L_T_given_t /= L_T_given_t.sum(
) # normalize all passes after some have been chopped off
return L_T_given_t
def get_ppc():
# use p_int as memoized values for integration
# trajectory integration
g = 10.72468 # y/s/s
dx = reach_vecs[:, 0] # F
dy = reach_vecs[:, 1] # F
vx = dx[:, None] / T[None, :] # F, T
vy = dy[:, None] / T[None, :] # F, T
vz_0 = (T * g) / 2 # T
# note that idx (i, j, k) into below arrays is invalid when j < k
traj_ts = np.tile(T, (len(field_locs), len(T), 1)) # (F, T, T)
traj_locs_x_idx = np.rint(
np.clip((ball_start[0] + vx[:, :, None] * T), 0,
len(x) - 1)).astype(int) # F, T, T
traj_locs_y_idx = np.rint(
np.clip((ball_start[1] + vy[:, :, None] * T), 0,
len(y) - 1)).astype(int) # F, T, T
traj_locs_z = 2.0 + vz_0[
None, :,
None] * traj_ts - 0.5 * g * traj_ts * traj_ts # F, T, T
path_idxs = np.ravel_multi_index(
np.stack((traj_locs_y_idx, traj_locs_x_idx)).reshape(2, -1),
xx.shape) # (F*T*T,)
traj_t_idxs = np.rint(10 * traj_ts - 1).flatten().astype(
int) # (F, T, T)
# TODO
ppi_ind_dt = p_int_ind[path_idxs, traj_t_idxs] # F*T*T, J
ppi_ind_dt = ppi_ind_dt.reshape(
(*traj_locs_x_idx.shape,
len(reaction_player_locs))) # F, T, T, J
# ppi_off_dt = p_int_off[path_idxs, traj_t_idxs] # F*T*T,
# ppi_off_dt = ppi_off_dt.reshape(traj_locs_x_idx.shape) # F, T, T
# ppi_def_dt = p_int_def[path_idxs, traj_t_idxs] # F*T*T,
# ppi_def_dt = ppi_def_dt.reshape(traj_locs_x_idx.shape) # F, T, T
# ppi_off_adj_dt = p_int_off_adj[path_idxs, traj_t_idxs] # F*T*T,
# ppi_off_adj_dt = ppi_off_adj_dt.reshape(traj_locs_x_idx.shape) # F, T, T
# account for ball height on traj and normalize each locations int probability
lambda_z = np.where(
(traj_locs_z < params.z_max) & (traj_locs_z > params.z_min), 1,
0) # F, T, T # maybe change this to a normal distribution
ppi_ind_dt = ppi_ind_dt * lambda_z[:, :, :, None] # F, T, T
# ppi_off_dt = ppi_off_dt * lambda_z[:, :, :, None] # F, T, T
# ppi_off_adj_dt = ppi_off_adj_dt * lambda_z[:, :, :, None] # F, T, T
# ppi_def_dt = ppi_def_dt * lambda_z[:, :, :, None] # F, T, T
# P(ball is still in play at t)
ball_in_play_traj = np.cumprod(1 - ppi_ind_dt, axis=-1)
ball_in_play_traj = np.roll(ball_in_play_traj, 1,
axis=-1) # F, T, T
ball_in_play_traj[:, :, 0] = 1
# ppi_def_adj_dt = ball_in_play_traj * ppi_def_dt # F, T, T
# ppi_off_adj_dt = ball_in_play_traj * ppi_off_adj_dt # F, T, T
ppi_ind_adj_dt = ball_in_play_traj[...,
None] * ppi_ind_dt # F, T, T, J
# ppi_def_traj = np.cumsum(ppi_def_adj_dt, axis=-1)
# ppi_def_traj = np.einsum('ijj->ij', ppi_def_traj) # (F, T)
# ppi_off_traj = np.cumsum(ppi_off_adj_dt, axis=-1)
# ppi_off_traj = np.einsum('ijj->ij', ppi_off_traj) # (F, T)
ppi_ind_traj = np.cumsum(ppi_ind_adj_dt, axis=-2)
ppi_ind_traj = np.einsum('ijjk->ijk', ppi_ind_traj) # (F, T, J)
ppi_off_traj = np.sum(ppi_ind_traj,
where=(player_teams == 'OFF'),
axis=-1) # (F, T)
ppi_def_traj = np.sum(ppi_ind_traj,
where=(player_teams == 'DEF'),
axis=-1) # (F, T)
ppi_rem_traj = 1 - ppi_def_traj - ppi_off_traj # (F, T)
return ppi_off_traj, ppi_def_traj, ppi_ind_traj
def get_xyac():
x_proj_def = x_proj[:, :, player_teams == "DEF"] # F, T, J
v_proj_def = s_proj[:, :, player_teams == "DEF"]
y_proj_def = y_proj[:, :, player_teams == "DEF"]
x_proj_relv = x_proj_def - field_locs[:, None, None,
0] # F, T, J #GET RELATIVE COORDS
y_proj_relv = y_proj_def - field_locs[:, None, None,
1] # F, T, J #GET RELATIVE COORDS
distances_to_ball = np.stack((x_proj_relv, y_proj_relv),
axis=-1) # F, T, J, 2
distance_mags = np.linalg.norm(distances_to_ball,
axis=-1) # F, T, J
sorted_indices = np.argsort(distance_mags, axis=2) # F, T, J
distance_mags = np.take_along_axis(distance_mags,
sorted_indices,
axis=2) # F, T, J
x_proj_sorted = np.take_along_axis(x_proj_relv,
sorted_indices,
axis=2)
y_proj_sorted = np.take_along_axis(y_proj_relv,
sorted_indices,
axis=2)
v_proj_sorted = np.take_along_axis(v_proj_def,
sorted_indices,
axis=2)
just_top_5_distances = distance_mags[:, :, 0:5] # F, T, 5
just_top_5_x_proj = x_proj_sorted[:, :, 0:5] # F, T, 5
just_top_5_y_proj = y_proj_sorted[:, :, 0:5] # F, T, 5
just_top_5_v_proj = v_proj_sorted[:, :, 0:5] # F, T, 5
just_top_5_distances = np.reshape(
just_top_5_distances,
(-1, just_top_5_distances.shape[2])) # (F*T, 5)
just_top_5_x_proj = np.reshape(just_top_5_x_proj,
just_top_5_distances.shape)
just_top_5_y_proj = np.reshape(just_top_5_y_proj,
just_top_5_distances.shape)
just_top_5_v_proj = np.reshape(just_top_5_v_proj,
just_top_5_distances.shape)
endpoints = np.repeat(field_locs, repeats=len(T), axis=0) # FxT, 2
# assert((field_locs[:, None, :]+np.zeros_like(T[None, None, :])).reshape(end))
times = np.repeat(T[None, :], repeats=len(field_locs), axis=0)
times_shaped = times.reshape(
(times.shape[0] * times.shape[1])) # FxT, 1
value_array = np.array(
[-2.5, 2.5, 7.5, 12.5, 17.5, 22.5, 27.5, 30])
field_df = pd.DataFrame({
'pass_endpoint_x':
endpoints[:, 0],
'pass_endpoint_y':
endpoints[:, 1],
'frame_thrown':
t,
'time_of_flight':
times_shaped,
'1-closest-defender-distance':
just_top_5_distances[:, 0],
'2-closest-defender-distance':
just_top_5_distances[:, 1],
'3-closest-defender-distance':
just_top_5_distances[:, 2],
'4-closest-defender-distance':
just_top_5_distances[:, 3],
'5-closest-defender-distance':
just_top_5_distances[:, 4],
'1-closest-defender-x':
just_top_5_x_proj[:, 0],
'2-closest-defender-x':
just_top_5_x_proj[:, 1],
'3-closest-defender-x':
just_top_5_x_proj[:, 2],
'4-closest-defender-x':
just_top_5_x_proj[:, 3],
'5-closest-defender-x':
just_top_5_x_proj[:, 4],
'1-closest-defender-y':
just_top_5_y_proj[:, 0],
'2-closest-defender-y':
just_top_5_y_proj[:, 1],
'3-closest-defender-y':
just_top_5_y_proj[:, 2],
'4-closest-defender-y':
just_top_5_y_proj[:, 3],
'5-closest-defender-y':
just_top_5_y_proj[:, 4],
'1-closest-defender-speed':
just_top_5_v_proj[:, 0],
'2-closest-defender-speed':
just_top_5_v_proj[:, 1],
'3-closest-defender-speed':
just_top_5_v_proj[:, 2],
'4-closest-defender-speed':
just_top_5_v_proj[:, 3],
'5-closest-defender-speed':
just_top_5_v_proj[:, 4],
"y":
endpoints[:, 1]
})
# CALCULTE XYAC
dtest = treelite_runtime.Batch.from_npy2d(
field_df[cols_when_model_builds].values)
ypred = xyac_predictor.predict(dtest)
y_vals = np.sum(ypred * value_array, axis=1)
field_df['xyac'] = y_vals
field_df['play_endpoint_x'] = np.round(field_df['xyac'] +
field_df['pass_endpoint_x'])
field_df['play_endpoint_x'] = field_df['play_endpoint_x'] + .5
field_df['play_endpoint_x'] = np.clip(field_df['play_endpoint_x'],
0.5, 119.5)
return field_df[[
'pass_endpoint_x', 'pass_endpoint_y', 'time_of_flight', 'xyac',
'play_endpoint_x'
]]
nonlocal epa_df
# completion model
ppc_off, ppc_def, ppc_ind = get_ppc() # (F, T), (F, T), (F, T, J)
# value model
epa_xyac_df = get_xyac().merge(epa_df,
how='left',
on='play_endpoint_x')
xyac = epa_xyac_df.xyac.to_numpy().reshape(ppc_off.shape)
# end_x = epa_xyac_df.play_endpoint_x.to_numpy().reshape(ppc_off.shape)
xepa_comp = epa_xyac_df.xepa_comp.to_numpy().reshape(
ppc_off.shape) # F, T
xepa_inc = epa_xyac_df.xepa_inc.iloc[0]
xepa_diff = xepa_comp - xepa_inc
# assert(h_trans_prob.shape == ppc_off.shape)
# breakpoint()
# transition model
h_trans_prob = get_hist_trans_prob() # (F, T)
trans = h_trans_prob * np.power(ppc_off, params.alpha) # F, T
trans /= trans.sum()
# output metrics (eppa1 uses ppc_off as catch prob, eppa2 uses catch_prob)
ind_eppa1_wo_value = ppc_ind * trans[..., None] # F, T, J
ind_eppa1 = ind_eppa1_wo_value * xepa_diff[..., None] # F, T, J
# ind_eppa2_wo_value = catch_prob_ind * trans[..., None] # F, T, J
# ind_eppa2 = ind_eppa2_wo_value * xepa_diff[..., None] # F, T, J
eppa1_pass_val = (ppc_off * xepa_comp) + (1 - ppc_off) * xepa_inc
eppa1 = eppa1_pass_val * trans # F, T
# eppa2_pass_val = (catch_prob_off*xepa_comp)+(1-catch_prob_off)*xepa_inc
# eppa2 = eppa2_pass_val*trans # F, T
eppa1_wo_value = ppc_off * trans
# eppa2_wo_value = catch_prob_off*trans
# maxEppa1Val = eppa1_pass_val[np.unravel_index(np.argmax(eppa1, axis=None), eppa1.shape)]
# maxEppa2Val = eppa2_pass_val[np.unravel_index(np.argmax(eppa2, axis=None), eppa2.shape)]
field_df = pd.DataFrame()
player_stats_df = pd.DataFrame()
passes_df = pd.DataFrame()
proj_df = pd.DataFrame()
if viz_true_proj and true_pass_found:
proj_df['nflId'] = frame_df['nflId']
proj_df['reax_x'] = frame_df['x_r']
proj_df['reax_y'] = frame_df['y_r']
proj_df['reax_v_x'] = frame_df['v_x_r']
proj_df['reax_v_y'] = frame_df['v_y_r']
proj_df['d_vec_x'] = int_d_vec[true_f_idx, :, 0]
proj_df['d_vec_y'] = int_d_vec[true_f_idx, :, 1]
proj_df['d_mag'] = int_d_mag[true_f_idx]
proj_df['int_s0'] = int_s0[true_f_idx]
proj_df['t_lt_smax'] = t_lt_smax[true_f_idx]
proj_df['d_lt_smax'] = d_lt_smax[true_f_idx]
proj_df['t_at_smax'] = t_at_smax[true_f_idx]
proj_df['d_at_smax'] = d_at_smax[true_f_idx]
proj_df['t_tot'] = t_tot[true_f_idx]
proj_df['int_dT'] = int_dT[true_f_idx, true_T_idx]
proj_df['p_int'] = p_int[true_f_idx, true_T_idx]
proj_df['p_int_adj'] = p_int_adj[true_f_idx, true_T_idx]
proj_df['d_proj'] = d_proj[true_f_idx, true_T_idx] # J
proj_df['s_proj'] = s_proj[true_f_idx, true_T_idx] # J
proj_df['proj_x'] = x_proj[true_f_idx, true_T_idx] # J
proj_df['proj_y'] = y_proj[true_f_idx, true_T_idx] # J
proj_df['proj_v_x'] = v_x_proj[true_f_idx, true_T_idx] # J
proj_df['proj_v_y'] = v_y_proj[true_f_idx, true_T_idx] # J
proj_df['ppc_ind'] = ppc_ind[true_f_idx, true_T_idx]
# proj_df['catch_prob'] = catch_prob_ind[true_f_idx, true_T_idx] # only valid for off players
proj_df['frameId'] = frame_id
if save_np:
week = games_df.loc[games_df.gameId ==
game_id].week.to_numpy()[0].item()
dir = out_dir_path.format(f'{week}/{game_id}/{play_id}')
Path(dir).mkdir(parents=True, exist_ok=True)
# np.savez_compressed(f'{dir}/{frame_id}', players_ind_info=ind_info,
# ppc_ind=ppc_ind, h_trans=h_trans_prob, xepa=xepa, eppa_ind=eppa_ind)
# np.savez_compressed(f'{dir}/{frame_id}', ind_info=ind_info, eppa_ind=eppa_ppc_ind)
if stats_df:
# ind_eppa1 = ind_eppa1.sum(axis=(0, 1)) # J
# ind_eppa1_wo_value = ind_eppa1_wo_value.sum(axis=(0, 1)) # J
player_stats_df = pd.DataFrame(
{
'gameId': game_id,
'playId': play_id,
'frameId': frame_id,
'frame_after_snap': t,
'nflId': player_ids,
'displayName': player_names,
'team': player_team_names,
'team_pos': player_teams,
'ind_eppa1': ind_eppa1.sum(axis=(0, 1)),
'ind_eppa1_wo_value': ind_eppa1_wo_value.sum(axis=(0, 1)),
},
# 'ind_eppa2': ind_eppa2.sum(axis=(0, 1)),
# 'ind_eppa2_wo_value': ind_eppa2_wo_value.sum(axis=(0, 1))},
index=player_ids)
off_team_name = play_df.loc[play_df.team_pos ==
'OFF'].teamAbbr.iloc[0]
def_team_name = play_df.loc[play_df.team_pos ==
'DEF'].teamAbbr.iloc[0]
# fill out passes df
row = {
'gameId': game_id,
'playId': play_id,
'frameId': frame_id,
'frames_after_snap': t,
'off_team': off_team_name,
'def_team': def_team_name,
'eppa1_tot': eppa1.sum(), # 'eppa2_tot': eppa2.sum(),
'xepa_inc': xepa_inc
}
# for name, arr in {'eppa1': eppa1, 'eppa2': eppa2, 'ppc_off': ppc_off, 'catch_prob': catch_prob_off, 'eppa1_pass_val': eppa1_pass_val,
# 'eppa2_pass_val': eppa2_pass_val, 'eppa1_wo_value': eppa1_wo_value, 'eppa2_wo_value': eppa2_wo_value}.items():
for name, arr in {
'eppa1': eppa1,
'ppc_off': ppc_off,
'eppa1_pass_val': eppa1_pass_val,
'eppa1_wo_value': eppa1_wo_value
}.items():
f, T_idx = np.unravel_index(arr.argmax(), arr.shape)
end_x, end_y = field_locs[f]
row[f"max_{name}_x"] = end_x
row[f"max_{name}_y"] = end_y
row[f"max_{name}_T"] = T[T_idx]
# row[f"max_{name}_catch_prob"] = catch_prob_off[f, T_idx]
row[f"max_{name}_ppc_off"] = ppc_off[f, T_idx]
row[f"max_{name}_pint_off"] = p_int_off[f, T_idx]
row[f"max_{name}_ppc_def"] = ppc_def[f, T_idx]
row[f"max_{name}_pint_def"] = p_int_def[f, T_idx]
row[f"max_{name}_xyac"] = xyac[f, T_idx]
row[f"max_{name}_xepa_comp"] = xepa_comp[f, T_idx]
row[f"max_{name}_trans"] = trans[f, T_idx]
row[f"max_{name}_trans_tof"] = h_trans_prob[f, T_idx]
row[f"max_{name}_eppa1"] = eppa1[f, T_idx]
# row[f"max_{name}_eppa2"] = eppa2[f, T_idx]
row[f"max_{name}_eppa1_xval"] = eppa1_pass_val[f, T_idx]
# row[f"max_{name}_eppa2_xval"] = eppa2_pass_val[f, T_idx]
row[f"max_{name}_eppa1_wo_value"] = eppa1_wo_value[f, T_idx]
# row[f"max_{name}_eppa2_wo_value"] = eppa2_wo_value[f, T_idx]
if (true_pass_found and frame_id == pass_forward_frame) or (
simulatePass and frame_id == sim_frame):
row[f"true_x"] = true_x_idxd
row[f"true_y"] = true_y_idxd
row[f"true_T"] = true_T_idxd
# row[f"true_catch_prob"] = catch_prob_off[true_f_idx, true_T_idx]
row[f"true_ppc_off"] = ppc_off[true_f_idx, true_T_idx]
row[f"true_pint_off"] = p_int_off[true_f_idx, true_T_idx]
row[f"true_ppc_def"] = ppc_def[true_f_idx, true_T_idx]
row[f"true_pint_def"] = p_int_def[true_f_idx, true_T_idx]
row[f"true_xyac"] = xyac[true_f_idx, true_T_idx]
row[f"true_xepa_comp"] = xepa_comp[true_f_idx, true_T_idx]
row[f"true_xepa"] = xepa_comp[
true_f_idx,
true_T_idx] if true_pass_completed else xepa_inc
row[f"true_trans"] = trans[true_f_idx, true_T_idx]
row[f"true_trans_tof"] = h_trans_prob[true_f_idx, true_T_idx]
row[f"true_eppa1"] = eppa1[true_f_idx, true_T_idx]
# row[f"true_eppa2"] = eppa2[true_f_idx, true_T_idx]
row[f"true_eppa1_xval"] = eppa1_pass_val[true_f_idx,
true_T_idx]
# row[f"true_eppa2_xval"] = eppa2_pass_val[true_f_idx, true_T_idx]
row[f"true_eppa1_wo_value"] = eppa1_wo_value[true_f_idx,
true_T_idx]
# row[f"true_eppa2_wo_value"] = eppa2_wo_value[true_f_idx, true_T_idx]
passes_df = pd.DataFrame(row, index=[t])
if viz_df:
field_df = pd.DataFrame({
'frameId':
frame_id,
'ball_end_x':
field_locs[:, 0],
'ball_end_y':
field_locs[:, 1],
'eppa1m':
eppa1.max(axis=1),
'eppa1s':
eppa1.sum(axis=1),
'eppa1_val':
eppa1_pass_val[np.arange(eppa1_pass_val.shape[0]),
eppa1.argmax(axis=1)],
# 'eppa2': eppa2.max(axis=1),
# 'p_int_off_only': p_int_off_only.max(axis=1),
'p_int_off':
p_int_off.max(axis=1),
# 'p_int_adj_off': np.max(p_int_adj, axis=(1, 2), where=(player_teams=='OFF')),
'ppc_off':
ppc_off.max(axis=1),
'ppc_def':
ppc_def.max(axis=1),
'eppa1_wo_value':
eppa1_wo_value.max(axis=1),
# 'eppa2_wo_value': eppa2_wo_value.max(axis=1),
'eppa1_wo_trans':
eppa1_pass_val.max(axis=1),
# 'eppa2_wo_trans': eppa2_pass_val.max(axis=1),
'xyac':
xyac.max(axis=1),
'xepa_comp':
xepa_comp.max(axis=1) - xepa_inc,
'xepa_inc':
xepa_inc,
'trans':
trans.sum(axis=1),
})
field_df.loc[field_df.ball_end_x < ball_start[0] -
10, :] = np.nan # remove backward passes
# breakpoint()
return field_df, passes_df, player_stats_df, proj_df
field_dfs = pd.DataFrame()
player_stats_df = pd.DataFrame()
passes_df = pd.DataFrame()
proj_df = pd.DataFrame()
if pass_forward_frame < (min_t_frame + ball_snap_frame):
return play_df, field_dfs, passes_df, player_stats_df
if not simulatePass:
for fid in tqdm(
range(
min_t_frame + ball_snap_frame,
min(pass_forward_frame, max_t_frame + ball_snap_frame) +
1)):
# for fid in tqdm(range(pass_forward_frame, pass_forward_frame+1)):
field_df, passes, player_stats, projs = frame_eppa(fid)
field_dfs = field_dfs.append(field_df, ignore_index=True)
passes_df = passes_df.append(passes, ignore_index=True)
player_stats_df = player_stats_df.append(player_stats,
ignore_index=True)
proj_df = proj_df.append(projs, ignore_index=True)
else:
for fid in tqdm(range(sim_frame, sim_frame + 1)):
# for fid in tqdm(range(pass_forward_frame, pass_forward_frame+1)):
field_df, passes, player_stats, projs = frame_eppa(fid)
field_dfs = field_dfs.append(field_df, ignore_index=True)
passes_df = passes_df.append(passes, ignore_index=True)
player_stats_df = player_stats_df.append(player_stats,
ignore_index=True)
proj_df = proj_df.append(projs, ignore_index=True)
if true_pass_found & viz_true_proj:
play_df = pd.merge(play_df,
proj_df,
on=['frameId', 'nflId'],
how='left')
if save_all_dfs:
week = games_df.loc[games_df.gameId ==
game_id].week.to_numpy()[0].item()
dir = out_dir_path.format(f'{week}/{game_id}/{play_id}')
Path(dir).mkdir(parents=True, exist_ok=True)
proj_df.to_pickle(f"{dir}/true_pass_player_proj.pkl", )
passes_df.to_pickle(f"{dir}/passes.pkl", )
player_stats_df.to_pickle(f"{dir}/player_stats.pkl", )
field_dfs.to_pickle(f"{dir}/field_viz.pkl", )
return play_df, field_dfs, passes_df, player_stats_df
# params={'parallel_comp': 32}, verbose=True) #.so for ubuntu, .dylib for mac
# bst = joblib.load(join(get_repo_dir(), "analysis/models/in/epa_model_rishav_no_time.model"))
# xgb.plot_importance(bst)
# scores = bst.get_score(importance_type='gain')
# print(scores.keys())
# cols_when_model_builds = bst.feature_names
# print("ADDED NEW SHIT")
# model = treelite.Model.from_xgboost(bst)
# toolchain = 'gcc'
# model.export_lib(toolchain=toolchain, libpath=join(get_repo_dir(), 'analysis/models/in/epa_no_time_mymodel.so'),
# params={'parallel_comp': 32}, verbose=True) #.so for ubuntu, .dylib for mac
# file loading and prep
path_shared = join(get_repo_dir(), 'data/{}')
dt = np.float64
games_df = pd.read_csv(path_shared.format('games.csv'))
plays_df = pd.read_csv(path_shared.format('plays.csv'))
players_df = pd.read_csv(path_shared.format('players.csv'))
pbp_2018 = pd.read_csv(path_shared.format('play_by_play_2018.csv'),
low_memory=False)
pbp_joined = pd.merge(plays_df,
pbp_2018,
how="left",
left_on=["gameId", "playId"],
right_on=["old_game_id", "play_id"])
pbp_joined["retractable"] = np.where(
import os
import sys
import shutil
from utils import get_repo_dir
from tqdm import tqdm
trans_dir = os.path.join(get_repo_dir(), 'data/transforms')
to_delete = []
for hash_dir in tqdm([f for f in os.scandir(trans_dir) if f.is_dir()]):
if len(list(os.scandir(hash_dir))) != 200:
to_delete.append(hash_dir)
if to_delete:
if (len(sys.argv) > 1 and sys.argv[1] == 'DELETE' and input(
f'Are you sure you want to delete {len(to_delete)} directories? (y/N): '
).lower() in ['y', 'yes']):
for hash_dir in to_delete:
print(f'Deleting directory {hash_dir.name}')
shutil.rmtree(hash_dir.path)
else:
print(f'{len(to_delete)} incomplete directories found.')
if input('Print? (y/N): ').lower() in ['y', 'yes']:
for hash_dir in to_delete:
print(hash_dir.name)
else:
print('No incomplete directories found.')
import pandas as pd
from models.play_eppa_gpu import *
from pathlib import Path
import traceback
import os
from tqdm import tqdm
from utils import get_repo_dir
WEEK_START = 1
WEEK_END = 2
out_dir_path = os.path.join(get_repo_dir(), 'output/{}') # for cloud runs
path_shared = os.path.join(get_repo_dir(), 'data/{}')
# main loop
for week in range(WEEK_START, WEEK_END):
track_df = pd.read_csv(
path_shared.format(f'week{week}_norm.csv',
usecols=[
'nflId', 'displayName', 'position', 'team_pos',
'x', 'y', 'v_x', 'v_y', 'v_mag', 'v_theta',
'a_x', 'a_y', 'a_mag', 'a_theta'
]))
plays = sorted(
list(
set(
map(
lambda x: (x[0].item(), x[1].item()),
track_df.groupby(['gameId', 'playId'],
as_index=False).first()[[
# -*- coding: utf-8 -*-
import os
import numpy as np
import pandas as pd
pd.options.mode.chained_assignment = None
import torch
from models.params import params
from models.consts import *
from models.frame_eppa import frame_eppa, set_device
from utils import get_repo_dir
from tqdm import tqdm
out_dir = os.path.join(get_repo_dir(), 'def_opt_output')
os.makedirs(out_dir, exist_ok=True)
out_dir_path = os.path.join(out_dir, '{}') # for cloud runs
path_shared = os.path.join(get_repo_dir(), 'data/{}')
plan_horizon = 0.8 # how far ahead to project (in seconds) while optimizing path
plan_res = 0.1 # how long (in seconds) to execute plan. less than plan_horizon
frame_dt = 0.1 # how far apart (in seconds) the frames are
plan_res_linspace = np.linspace(frame_dt, plan_res,
np.round(plan_res /
frame_dt).astype(int))[:, None]
# 10*dt is the number of frames to increment by
res_incr = int(10 * plan_res)
horizon_incr = int(10 * plan_horizon)
reduce_eppa_mode = 'sum' # choose from {'sum', 'max', 'softmax'}
softmax_temp = 1000 # used for 'softmax' reduce_eppa_mode
week_game_plays = [