norm2 = np.sum(train_tracklets[:, :, 12:], axis=(1, 2), keepdims=True)
# train_tracklets[:,:,:12] /= norm1
# train_tracklets[:,:,12:] /= norm2
# test_tracks, infosets = dt.load_whole_named_dataset('6_tracklets_large_calib_deconvoluted_test')
# test_tracklets = test_tracks.reshape((-1, 17, 24, 1))
# test_labels = np.repeat(infosets[:, 0], 6)
pid_model = cmod.VeryComplexConvTrackletPID()
pid_model.compile(
optimizer=tf.train.AdamOptimizer(learning_rate=0.001),
loss='binary_crossentropy',
metrics=['accuracy',
cmet.PionEfficiencyAtElectronEfficiency(0.9)],
)
pid_model.fit(
train_tracklets,
train_labels,
batch_size=512,
epochs=100,
validation_split=0.2,
)
# validation_data=(test_tracklets, test_labels),)
units = pid_model.get_weights()[0]
u1 = units[:, :, 0, 0]
u2 = units[:, :, 0, 1]
train_labels = train_infosets[:, 0]
test_tracks, test_infosets = dt.load_whole_named_dataset('6_tracklets_large_test')
test_tracks = np.expand_dims(test_tracks, axis=-1)
test_labels = test_infosets[:, 0]
for i in range(100):
test_tracks /= np.sum(test_tracks, axis=(2,3), keepdims=True)
train_tracks /= np.sum(train_tracks, axis=(2,3), keepdims=True)
tracklet_pid_model = cmod.PartialTrackletConvPID()
track_pid_model = cmod.TrackletModelMultiplexer(tracklet_pid_model)
track_pid_model.compile(optimizer=tf.train.AdamOptimizer(learning_rate=0.0001),
loss='binary_crossentropy',
metrics=['accuracy', cmet.PionEfficiencyAtElectronEfficiency(0.9)],
)
history = track_pid_model.fit(train_tracks,
train_labels,
batch_size=512,
epochs=1,
validation_data=(test_tracks, test_labels),)
conv_unit = track_pid_model.get_weights()[0][:,:,0,0]
# plt.imshow(conv_unit)
# plt.show()
new_train_tracks = dt.project_conv_unit_out_of_dataset(train_tracks, conv_unit, [1,1], 'valid').reshape((-1, 6, 17, 24))
new_test_tracks = dt.project_conv_unit_out_of_dataset(test_tracks, conv_unit, [1,1], 'valid').reshape((-1, 6, 17, 24))