varifocal/code/sync_defocus_from_rgbd_fast/train/train-lvf-c.py [146:257]: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - fwConstant = 3 fnumConstant = 128 fw = np.full((L), fwConstant, dtype=int) # filter width fnum = np.append(np.full((L-1), fnumConstant, dtype=int), OUTPUT_DIM*(INTERLEAVE_RATE**2)) # output channels at each layer layers_params = {} layers = {} prev_layers = {} for i in range(0, L): if i==0: # first layer in_dim, out_dim = INPUT_DIM*(INTERLEAVE_RATE**2), fnum[i] elif i==L-1: # last layer in_dim, out_dim = fnum[i-1], OUTPUT_DIM*(INTERLEAVE_RATE**2) else: in_dim, out_dim = fnum[i-1], fnum[i] layers_params[i] = {'weights':init_weights([fw[i],fw[i],in_dim, out_dim], 'xavier',xavier_params=(in_dim, out_dim),r=weightVarScale), 'bias':tf.Variable(tf.truncated_normal([out_dim],stddev=bias_stddev))} # build layers:: print("input data:", x_in.shape) if RENORMALIZE_INPUT: x_in = (x_in - 0.5*RENORM_SCALE) for i in range(0, L): if i==0: prev_layers[i] = x_in elif (i<3) or (i%2==0): prev_layers[i] = layers[i-1] else: prev_layers[i] = layers[i-1] + layers[i-3] print('(skip connection: %d, %d)'%(i-1, i-3)) if i==L-1: # last layer layers[i] = tf.nn.tanh(tf.layers.batch_normalization(tf.nn.bias_add(tf.nn.conv2d(prev_layers[i],layers_params[i]['weights'],strides=[1,1,1,1], padding='SAME', data_format='NCHW'), layers_params[i]['bias'], data_format='NCHW'),axis=1)) else: layers[i] = activationFunc(tf.layers.batch_normalization(tf.nn.bias_add(tf.nn.conv2d(prev_layers[i],layers_params[i]['weights'],strides=[1,1,1,1], padding='SAME', data_format='NCHW'), layers_params[i]['bias'], data_format='NCHW'),axis=1)) print("layer %d:" % i, layers[i].shape) # renormalize to desired scale x_out = tf.add(0.5, tf.scalar_mul(0.5, layers[L-1]), name='x_out') print("output tensor:", x_out.shape) return deinterleave(INTERLEAVE_RATE, x_out) # construct model model_op = model(X) # target, prediction labels, predictions = Y, model_op # shared rmse_intensity = tf.losses.mean_squared_error(labels=labels, predictions=predictions) psnr_intensity = 20*log10(RENORM_SCALE) - 10*log10(rmse_intensity) ssim_intensity = tf.reduce_mean(tf.image.ssim(tf.transpose(labels, [0,2,3,1]), tf.transpose(predictions, [0,2,3,1]), max_val = 1.0)) if USE_LOG_L2_BOTH: labels_dx, labels_dy = calImageGradients(labels) preds_dx, preds_dy = calImageGradients(predictions) rmse_grad_x, rmse_grad_y = tf.losses.mean_squared_error(labels=labels_dx, predictions=preds_dx), tf.losses.mean_squared_error(labels=labels_dy, predictions=preds_dy) psnr_grad_x, psnr_grad_y = -20*log10(RENORM_SCALE) + 10*log10(rmse_grad_x), -20*log10(RENORM_SCALE) + 10*log10(rmse_grad_y) loss = -psnr_intensity + 0.5*(psnr_grad_x + psnr_grad_y) elif USE_LOG_L1_BOTH: log_diff_intensity = log10(tf.reduce_mean(tf.abs(labels-predictions))) labels_dx, labels_dy = calImageGradients(labels) preds_dx, preds_dy = calImageGradients(predictions) log_diff_grad_x = log10(tf.reduce_mean(tf.abs(labels_dx-preds_dx))) log_diff_grad_y = log10(tf.reduce_mean(tf.abs(labels_dy-preds_dy))) loss = log_diff_intensity + 0.5*(log_diff_grad_x + log_diff_grad_y) # training optimizer optimizer = tf.train.AdamOptimizer(learning_rate=LEARNING_RATE, beta1=0.9, beta2=0.999, epsilon=1e-8).minimize(loss=loss) grads_and_vars = tf.train.AdamOptimizer(learning_rate=LEARNING_RATE, beta1=0.9, beta2=0.999, epsilon=1e-8).compute_gradients(loss, tf.trainable_variables()) # set up saver saver = tf.train.Saver(max_to_keep=5) # initialization init = tf.global_variables_initializer() # create a summary to monitor cost tensor training_summary = tf.summary.scalar("training loss", loss, family=TF_SUMMARY_FLAG) validation_summary = tf.summary.scalar("validation loss", loss, family=TF_SUMMARY_FLAG) # start training with tf.Session() as sess: # run the initialization sess.run(init) # op to write logs to Tensorboard summary_writer = tf.summary.FileWriter(LOGS_DIR, graph=tf.get_default_graph()) # restore model if trained if RESTORE_TRAINED_MODEL or MODE!="TRAIN": ckpt = tf.train.get_checkpoint_state(RESTORE_LOGS_DIR) if ckpt and ckpt.model_checkpoint_path: saver.restore(sess, tf.train.latest_checkpoint(RESTORE_LOGS_DIR)) print("model restored from %s" % RESTORE_LOGS_DIR) else: input("ERROR: NO RESTORED MODEL...") min_ell_eval_mean = [] #------------------------------------------------------------------------------- # training if MODE == "TRAIN": print("training mode") for e in range(0, NUM_EPOCHS): print("--training epoch:", e) - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - varifocal/code/sync_focalstack_from_rgbd/train/train-lfs.py [146:257]: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - fwConstant = 3 fnumConstant = 128 fw = np.full((L), fwConstant, dtype=int) # filter width fnum = np.append(np.full((L-1), fnumConstant, dtype=int), OUTPUT_DIM*(INTERLEAVE_RATE**2)) # output channels at each layer layers_params = {} layers = {} prev_layers = {} for i in range(0, L): if i==0: # first layer in_dim, out_dim = INPUT_DIM*(INTERLEAVE_RATE**2), fnum[i] elif i==L-1: # last layer in_dim, out_dim = fnum[i-1], OUTPUT_DIM*(INTERLEAVE_RATE**2) else: in_dim, out_dim = fnum[i-1], fnum[i] layers_params[i] = {'weights':init_weights([fw[i],fw[i],in_dim, out_dim], 'xavier',xavier_params=(in_dim, out_dim),r=weightVarScale), 'bias':tf.Variable(tf.truncated_normal([out_dim],stddev=bias_stddev))} # build layers:: print("input data:", x_in.shape) if RENORMALIZE_INPUT: x_in = (x_in - 0.5*RENORM_SCALE) for i in range(0, L): if i==0: prev_layers[i] = x_in elif (i<3) or (i%2==0): prev_layers[i] = layers[i-1] else: prev_layers[i] = layers[i-1] + layers[i-3] print('(skip connection: %d, %d)'%(i-1, i-3)) if i==L-1: # last layer layers[i] = tf.nn.tanh(tf.layers.batch_normalization(tf.nn.bias_add(tf.nn.conv2d(prev_layers[i],layers_params[i]['weights'],strides=[1,1,1,1], padding='SAME', data_format='NCHW'), layers_params[i]['bias'], data_format='NCHW'),axis=1)) else: layers[i] = activationFunc(tf.layers.batch_normalization(tf.nn.bias_add(tf.nn.conv2d(prev_layers[i],layers_params[i]['weights'],strides=[1,1,1,1], padding='SAME', data_format='NCHW'), layers_params[i]['bias'], data_format='NCHW'),axis=1)) print("layer %d:" % i, layers[i].shape) # renormalize to desired scale x_out = tf.add(0.5, tf.scalar_mul(0.5, layers[L-1]), name='x_out') print("output tensor:", x_out.shape) return deinterleave(INTERLEAVE_RATE, x_out) # construct model model_op = model(X) # target, prediction labels, predictions = Y, model_op # shared rmse_intensity = tf.losses.mean_squared_error(labels=labels, predictions=predictions) psnr_intensity = 20*log10(RENORM_SCALE) - 10*log10(rmse_intensity) ssim_intensity = tf.reduce_mean(tf.image.ssim(tf.transpose(labels, [0,2,3,1]), tf.transpose(predictions, [0,2,3,1]), max_val = 1.0)) if USE_LOG_L2_BOTH: labels_dx, labels_dy = calImageGradients(labels) preds_dx, preds_dy = calImageGradients(predictions) rmse_grad_x, rmse_grad_y = tf.losses.mean_squared_error(labels=labels_dx, predictions=preds_dx), tf.losses.mean_squared_error(labels=labels_dy, predictions=preds_dy) psnr_grad_x, psnr_grad_y = -20*log10(RENORM_SCALE) + 10*log10(rmse_grad_x), -20*log10(RENORM_SCALE) + 10*log10(rmse_grad_y) loss = -psnr_intensity + 0.5*(psnr_grad_x + psnr_grad_y) elif USE_LOG_L1_BOTH: log_diff_intensity = log10(tf.reduce_mean(tf.abs(labels-predictions))) labels_dx, labels_dy = calImageGradients(labels) preds_dx, preds_dy = calImageGradients(predictions) log_diff_grad_x = log10(tf.reduce_mean(tf.abs(labels_dx-preds_dx))) log_diff_grad_y = log10(tf.reduce_mean(tf.abs(labels_dy-preds_dy))) loss = log_diff_intensity + 0.5*(log_diff_grad_x + log_diff_grad_y) # training optimizer optimizer = tf.train.AdamOptimizer(learning_rate=LEARNING_RATE, beta1=0.9, beta2=0.999, epsilon=1e-8).minimize(loss=loss) grads_and_vars = tf.train.AdamOptimizer(learning_rate=LEARNING_RATE, beta1=0.9, beta2=0.999, epsilon=1e-8).compute_gradients(loss, tf.trainable_variables()) # set up saver saver = tf.train.Saver(max_to_keep=5) # initialization init = tf.global_variables_initializer() # create a summary to monitor cost tensor training_summary = tf.summary.scalar("training loss", loss, family=TF_SUMMARY_FLAG) validation_summary = tf.summary.scalar("validation loss", loss, family=TF_SUMMARY_FLAG) # start training with tf.Session() as sess: # run the initialization sess.run(init) # op to write logs to Tensorboard summary_writer = tf.summary.FileWriter(LOGS_DIR, graph=tf.get_default_graph()) # restore model if trained if RESTORE_TRAINED_MODEL or MODE!="TRAIN": ckpt = tf.train.get_checkpoint_state(RESTORE_LOGS_DIR) if ckpt and ckpt.model_checkpoint_path: saver.restore(sess, tf.train.latest_checkpoint(RESTORE_LOGS_DIR)) print("model restored from %s" % RESTORE_LOGS_DIR) else: input("ERROR: NO RESTORED MODEL...") min_ell_eval_mean = [] #------------------------------------------------------------------------------- # training if MODE == "TRAIN": print("training mode") for e in range(0, NUM_EPOCHS): print("--training epoch:", e) - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -