def run_training(hyper_param, model):
global batch_size, max_feature_size
'''
Train RSVP for a number of steps.
Args:
hyper_param: three elements, layer & feat & model
model:
Returns:
'''
# initialize the summary to write
csv_writer_acc, csv_writer_auc = autorun_util.csv_writer(model, hyper_param['feat'])
# Get the sets of images and labels for training, validation, and
# test on RSVP.
set_batch_size(125)
data_sets = rsvp_input_data.read_data_sets(EEG_DATA_MAT,
FLAGS.fake_data,
reshape_t=False)
# Tell TensorFlow that the model will be built into the default Graph.
with tf.Graph().as_default():
# Generate placeholders for the images and labels.
# Build a Graph that computes predictions from the inference model.
# Generate placeholders for the images and labels.
images_placeholder, labels_placeholder, keep_prob = \
placeholder_inputs(batch_size)
define_learning_rate()
# Build a Graph that computes predictions from the inference model.
results = autorun_deconv_lasso.select_running_cnn(images_placeholder,
None,
keep_prob, None,
None, None, None, None,
mode=0,
layer=hyper_param['layer'],
feat=hyper_param['feat'],
cnn_id=model
)
logits = results
# Add to the Graph the Ops for loss calculation.
loss = rsvp_quick_cnn_model.loss(logits, labels_placeholder)
# Add to the Graph the Ops that calculate and apply gradients.
train_op = rsvp_quick_cnn_model.training(loss, lr, global_step)
# Add the Op to compare the logits to the labels during evaluation.
eval_correct = rsvp_quick_cnn_model.evaluation(logits, labels_placeholder)
# Build the summary operation based on the TF collection of Summaries.
summary_op = tf.merge_all_summaries()
# Create a saver for writing training checkpoints.
# Create a session for running Ops on the Graph.
sess = tf.Session()
# Run the Op to initialize the variables.
# Instantiate a SummaryWriter to output summaries and the Graph.
summary_writer = tf.train.SummaryWriter(FLAGS.train_dir,
graph=sess.graph)
saver = tf.train.Saver()
init = tf.initialize_all_variables()
sess.run(init)
checkpoint = tf.train.get_checkpoint_state(FLAGS.train_dir)
if checkpoint and checkpoint.model_checkpoint_path:
saver.restore(sess, checkpoint.model_checkpoint_path)
print("Successfully loaded:", checkpoint.model_checkpoint_path)
else:
print("Could not find old network weights")
# saver.restore(sess, "/home/e/Downloads/models/-2649")
# print("Model restored.")
# And then after everything is built, start the training loop.
for step in range(0,FLAGS.max_steps):
start_time = time.time()
# Fill a feed dictionary with the actual set of images and labels
# for this particular training step.
if mode == autorun_deconv_lasso.TEST:
feed_dict = fill_feed_dict(data_sets.train,
0.5, #0.0625, # was .5 works good
images_placeholder,
labels_placeholder,
keep_prob)
# Run one step of the model. The return values are the activations
# from the `train_op` (which is discarded) and the `loss` Op. To
# inspect the values of your Ops or variables, you may include them
# in the list passed to sess.run() and the value tensors will be
# returned in the tuple from the call.
_, loss_value = sess.run([train_op, loss],
feed_dict=feed_dict)
duration = time.time() - start_time
# Write the summaries and print an overview fairly often.
if step % check_step == 0:
# Print status to stdout.
print('Step %d: loss = %.4f (%.3f sec)' % (step, loss_value, duration))
# Update the events file.
summary_str = sess.run(summary_op, feed_dict=feed_dict)
summary_writer.add_summary(summary_str, step)
# Save a checkpoint and evaluate the model periodically.
if (step + 1) % check_step == 0 or (step + 1) == FLAGS.max_steps:
saver.save(sess, FLAGS.train_dir, global_step=step)
# Evaluate against the training set.
print('Training Data Eval:')
do_eval(sess,
eval_correct,
logits,
images_placeholder,
labels_placeholder,
keep_prob,
data_sets.train)
# Evaluate against the validation set.
print('Validation Data Eval:')
#do_eval(sess,
# eval_correct,
# logits,
# images_placeholder,
# labels_placeholder,
# keep_prob,
# filter_num, image_num,
# data_sets.validation)
# Evaluate against the test set.
print('Test Data Eval:')
do_eval(sess,
eval_correct,
logits,
images_placeholder,
labels_placeholder,
keep_prob,
data_sets.test)
################################### REBUILD MODEL #####################################
temp = set(tf.all_variables())
set_batch_size(1)
images_placeholder2, labels_placeholder2, max_features_pl2, keep_prob2, filter_num2, image_num2, max_act_pl2, max_ind_pl2 = \
placeholder_inputs2(batch_size)
feed_dict2 = fill_feed_dict2(data_sets.train,
max_features_pl2, None,
1.0,
images_placeholder2,
labels_placeholder2,
keep_prob2,
filter_num2, image_num2, max_act_pl2, max_ind_pl2)
# Build a Graph that computes predictions from the inference model.
returnTensors = autorun_deconv_lasso.select_running_cnn(images_placeholder2,
max_features_pl2,
keep_prob2, 1,
filter_num2, image_num2,
max_act_pl2, max_ind_pl2,
mode=autorun_deconv_lasso.FIND_MAX_ACTIVATION_GET_SWITCHES,
layer=hyper_param['layer'],
feat=hyper_param['feat'],
cnn_id=model
)
# writer = tf.train.SummaryWriter("/home/e/deconvgraph/deconv_logs", sess.graph)
sess.run(tf.initialize_variables(set(tf.all_variables()) - temp))
################################### REBUILD MODEL END #####################################
updates = autorun_deconv_lasso.get_update_threshold()
clear_vars = autorun_deconv_lasso.get_clear_variables()
# Find maximum activation
num_layers = hyper_param['layer']
top_nth = []
feat_num = []
max_ind_list = []
max_image_ind_list = []
max_activation_list = []
max_labels_list = []
#select the nth highest feature
for n in range(1000):
for step in range(0, 19000 , 20):
print('test pass' + str(step))
feed_dict2 = fill_feed_dict2(data_sets.train,
max_features_pl2, None,
1.0,
images_placeholder2,
labels_placeholder2,
keep_prob2,
filter_num2, image_num2, max_act_pl2, max_ind_pl2,
0, step, 0.0, 0, # placeholders
batch_offset=step) # batch offset
returnTensorVals = sess.run(returnTensors, feed_dict=feed_dict2)
_ = returnTensorVals
_ = returnTensorVals
# unpack results
max_ind = returnTensorVals[-3 * num_layers:-2*num_layers]
max_image_ind = returnTensorVals[-2*num_layers:-num_layers]
max_activation = returnTensorVals[-num_layers:]
# Convert to python lists
if n==0:
for l in range(0, num_layers):
list_len = len(max_ind[l].tolist())
top_nth.append([n]*list_len)
feat_num.append(range(list_len))
max_ind_list.append(max_ind[l].tolist())
max_image_ind_list.append(max_image_ind[l].tolist())
max_activation_list.append(max_activation[l].tolist())
else:
for l in range(0, num_layers):
list_len = len(max_ind[l].tolist())
top_nth[l].extend([n]*list_len)
feat_num[l].extend(range(list_len))
max_ind_list[l].extend(max_ind[l].tolist())
max_image_ind_list[l].extend(max_image_ind[l].tolist())
max_activation_list[l].extend(max_activation[l].tolist())
# update threshold
returnTensorsTmp = list(returnTensors)
returnTensorsTmp.extend(updates)
# update max threshold
sess.run(returnTensorsTmp, feed_dict=feed_dict2)
# restart at batch zero
feed_dict2 = fill_feed_dict2(data_sets.train,
max_features_pl2, None,
1.0,
images_placeholder2,
labels_placeholder2,
keep_prob2,
filter_num2, image_num2, max_act_pl2, max_ind_pl2,
0, 0, 0.0, 0, # placeholders
batch_offset=0) # batch offset
# clear data
returnTensorsTmp = list(returnTensors)
returnTensorsTmp.extend(clear_vars)
# update max threshold
sess.run(returnTensorsTmp, feed_dict=feed_dict2)
for l in range(0, num_layers):
max_labels_np_tmp = data_sets.train.get_labels(max_image_ind_list[l])
max_labels_list.append(max_labels_np_tmp.tolist())
# find the top 9 activations from all features in top layer
cur_layer = 1
max_activation_info = zip(max_labels_list[cur_layer], max_activation_list[cur_layer], max_image_ind_list[cur_layer], max_ind_list[cur_layer],
feat_num[cur_layer], top_nth[cur_layer])
sorted_activations_neg = sorted(max_activation_info, key=lambda x: (x[0], -x[1]))
sorted_activations_pos = sorted(max_activation_info, key=lambda x: (-x[0], -x[1]))
print(Counter(max_labels_list[cur_layer]))
# Reconstruct
################################### REBUILD MODEL #####################################
temp = set(tf.all_variables())
set_batch_size(1)
images_placeholder2, labels_placeholder2, max_features_pl2, keep_prob2, filter_num2, image_num2, max_act_pl2, max_ind_pl2 = \
placeholder_inputs2(batch_size)
feed_dict2 = fill_feed_dict2(data_sets.train,
max_features_pl2, None,
1.0,
images_placeholder2,
labels_placeholder2,
keep_prob2,
filter_num2, image_num2, max_act_pl2, max_ind_pl2)
# Build a Graph that computes predictions from the inference model.
returnTensors = autorun_deconv_lasso.select_running_cnn(images_placeholder2,
max_features_pl2,
keep_prob2, 1,
filter_num2, image_num2,
max_act_pl2, max_ind_pl2,
mode=autorun_deconv_lasso.DECONV_LASSO,
layer=hyper_param['layer'],
feat=hyper_param['feat'],
cnn_id=model
)
# writer = tf.train.SummaryWriter("/home/e/deconvgraph/deconv_logs", sess.graph)
sess.run(tf.initialize_variables(set(tf.all_variables()) - temp))
################################### REBUILD MODEL END #####################################
# top 9 positive/negative label reconstructions
for i in range(0,2):
input_reconstructions = []
input_images = []
batch_nums = []
max_acts = []
max_indicies = []
max_filters = []
for top_nth in range(0, 10000):
if i == 0:
_, max_act_val, batch_num, max_ind_val, filter_num_val, _ = sorted_activations_neg[top_nth]
else:
_, max_act_val, batch_num, max_ind_val, filter_num_val, _ = sorted_activations_pos[top_nth]
all_features_mask = np.zeros(
shape=max_feature_size,
dtype=np.float32)
#ind_max_unraveled = np.unravel_index(max_ind_val , (1,16,16,1))
#ind_max_unraveled2 = np.unravel_index(filter_num_val , (1,16,16,hyper_param['feat'][1]))
#sum_ind = tuple(map(lambda a, b: a + b, ind_max_unraveled, ind_max_unraveled2))
#all_features_mask
all_features_mask[0, :, :, filter_num_val] = 1.0
all_features = np.zeros(
shape=max_feature_size,
dtype=np.float32)
feed_dict2 = fill_feed_dict2(data_sets.train,
max_features_pl2, all_features,
1.0,
images_placeholder2,
labels_placeholder2,
keep_prob2,
filter_num2, image_num2, max_act_pl2, max_ind_pl2,
0, batch_num, 0.0, 0, # placeholders
batch_offset=batch_num) # batch offset
returnTensorVals = sess.run(returnTensors, feed_dict=feed_dict2)
all_features = returnTensorVals[hyper_param['layer'] * 2 - 1] * all_features_mask
feed_dict2 = fill_feed_dict2(data_sets.train,
max_features_pl2, all_features,
1.0,
images_placeholder2,
labels_placeholder2,
keep_prob2,
filter_num2, image_num2, max_act_pl2, max_ind_pl2,
0, batch_num, 0.0, 0, # placeholders
batch_offset=batch_num) # batch offset
returnTensorVals = sess.run(returnTensors, feed_dict=feed_dict2)
image = feed_dict2[images_placeholder2].copy()
max_acts.append(max_act_val)
max_indicies.append(max_ind_val)
max_filters.append(filter_num_val)
batch_nums.append(batch_num)
input_images.append(image) # unpack
input_reconstructions.append(returnTensorVals[hyper_param['layer'] - 1])
top_nth_reconstructions_np = np.array([input_reconstructions])
top_nth_images_np = np.array([input_images])
save_location = roi_property.SAVE_DIR+'deconv/'
if i == 0:
print('Writing neg10.mat')
spio.savemat(save_location + 'neg10.mat',
dict(recon = top_nth_reconstructions_np, images = top_nth_images_np,
batch_nums=batch_nums, max_acts=max_acts,max_indicies=max_indicies,
max_filters=max_filters))
else:
print('Writing pos10.mat')
spio.savemat(save_location + 'pos10.mat',
dict(recon=top_nth_reconstructions_np, images=top_nth_images_np,
batch_nums=batch_nums, max_acts=max_acts, max_indicies=max_indicies,
max_filters=max_filters))
return
def run_training(hyper_param, model, name_idx, sub_idx):
global batch_size, max_feature_size
'''
Train RSVP for a number of steps.
Args:
hyper_param: three elements, layer & feat & model
model:
Returns:
'''
# initialize the summary to write
csv_writer_acc, csv_writer_auc = autorun_util.csv_writer(model, hyper_param['feat'])
# Get the sets of images and labels for training, validation, and
# test on RSVP.
set_batch_size(roi_property.BATCH_SIZE)
eeg_data = autorun_util.str_name(name_idx, sub_idx)
eeg_data_dir = roi_property.FILE_DIR + \
'rsvp_data/mat_sub/' + eeg_data
eeg_data_mat = eeg_data_dir + '.mat'
data_sets = rsvp_input_data.read_all_data(eeg_data_mat)
# Tell TensorFlow that the model will be built into the default Graph.
with tf.Graph().as_default():
# Generate placeholders for the images and labels.
images_placeholder, labels_placeholder, keep_prob = placeholder_inputs(
FLAGS.batch_size, data_sets.feature_shape[3])
# Build a Graph that computes predictions from the inference model.
logits = autorun_infer.select_running_cnn(images_placeholder,
keep_prob,
layer=hyper_param['layer'],
feat=hyper_param['feat'],
cnn_id=model)
# Add to the Graph the Ops for loss calculation.
loss = rsvp_quick_cnn_model.loss(logits, labels_placeholder)
# Add to the Graph the Ops that calculate and apply gradients.
train_op = rsvp_quick_cnn_model.training(loss, FLAGS.learning_rate)
# Add the Op to compare the logits to the labels during evaluation.
eval_correct = rsvp_quick_cnn_model.evaluation(logits, labels_placeholder)
# Build the summary operation based on the TF collection of Summaries.
summary_op = tf.merge_all_summaries()
# Create a saver for writing training checkpoints.
saver = tf.train.Saver()
# Create a session for running Ops on the Graph.
sess = tf.Session()
# Run the Op to initialize the variables.
init = tf.initialize_all_variables()
sess.run(init)
checkpoint = tf.train.get_checkpoint_state(FLAGS.train_dir)
if checkpoint and checkpoint.model_checkpoint_path:
saver.restore(sess, checkpoint.model_checkpoint_path)
print("Successfully loaded:", checkpoint.model_checkpoint_path)
else:
print("Could not find old network weights")
################################### FIND SWITCHES START #####################################
temp = set(tf.all_variables())
images_placeholder2, labels_placeholder2, max_features_pl2, keep_prob2, filter_num2, image_num2, max_act_pl2, max_ind_pl2 = \
placeholder_inputs2(batch_size)
feed_dict2 = fill_feed_dict2(data_sets.train,
max_features_pl2, None,
1.0,
images_placeholder2,
labels_placeholder2,
keep_prob2,
filter_num2, image_num2, max_act_pl2, max_ind_pl2)
# Build a Graph that computes predictions from the inference model.
returnTensors = autorun_deconv_lasso.select_running_cnn(images_placeholder2,
max_features_pl2,
keep_prob2, hyper_param['layer'] - 1,
filter_num2, image_num2,
max_act_pl2, max_ind_pl2,
mode=autorun_deconv_lasso.FIND_MAX_ACTIVATION_GET_SWITCHES,
layer=hyper_param['layer'],
feat=hyper_param['feat'],
cnn_id=model
)
# writer = tf.train.SummaryWriter("/home/e/deconvgraph/deconv_logs", sess.graph)
sess.run(tf.initialize_variables(set(tf.all_variables()) - temp))
################################### FIND SWITCHES END #####################################
updates = autorun_deconv_lasso.get_update_threshold()
clear_vars = autorun_deconv_lasso.get_clear_variables()
# Find maximum activation
num_layers = hyper_param['layer']
top_nth = []
feat_num = []
max_ind_list = []
max_image_ind_list = []
max_activation_list = []
max_labels_list = []
#select the nth highest feature
for n in range(1000):
for step in range(0, 167):
print('test pass' + str(step))
feed_dict2 = fill_feed_dict2(data_sets.train,
max_features_pl2, None,
1.0,
images_placeholder2,
labels_placeholder2,
keep_prob2,
filter_num2, image_num2, max_act_pl2, max_ind_pl2,
0, step, 0.0, 0, # placeholders
batch_offset=step) # batch offset
return_tensor_vals = sess.run(returnTensors, feed_dict=feed_dict2)
# unpack results
max_ind = return_tensor_vals[-3 * num_layers:-2*num_layers]
max_image_ind = return_tensor_vals[-2*num_layers:-num_layers]
max_activation = return_tensor_vals[-num_layers:]
# Convert to python lists
if n==0:
for l in range(0, num_layers):
list_len = len(max_ind[l].tolist())
top_nth.append([n]*list_len)
feat_num.append(range(list_len))
max_ind_list.append(max_ind[l].tolist())
max_image_ind_list.append(max_image_ind[l].tolist())
max_activation_list.append(max_activation[l].tolist())
else:
for l in range(0, num_layers):
list_len = len(max_ind[l].tolist())
top_nth[l].extend([n]*list_len)
feat_num[l].extend(range(list_len))
max_ind_list[l].extend(max_ind[l].tolist())
max_image_ind_list[l].extend(max_image_ind[l].tolist())
max_activation_list[l].extend(max_activation[l].tolist())
# update threshold
returnTensorsTmp = list(returnTensors)
returnTensorsTmp.extend(updates)
# update max threshold
sess.run(returnTensorsTmp, feed_dict=feed_dict2)
# restart at batch zero
feed_dict2 = fill_feed_dict2(data_sets.train,
max_features_pl2, None,
1.0,
images_placeholder2,
labels_placeholder2,
keep_prob2,
filter_num2, image_num2, max_act_pl2, max_ind_pl2,
0, 0, 0.0, 0, # placeholders
batch_offset=0) # batch offset
# clear data
returnTensorsTmp = list(returnTensors)
returnTensorsTmp.extend(clear_vars)
# update max threshold
sess.run(returnTensorsTmp, feed_dict=feed_dict2)
for l in range(0, num_layers):
max_labels_np_tmp = data_sets.train.get_labels(max_image_ind_list[l])
max_labels_list.append(max_labels_np_tmp.tolist())
# find the top 9 activations from all features in top layer
cur_layer = 1
max_activation_info = zip(max_labels_list[cur_layer], max_activation_list[cur_layer], max_image_ind_list[cur_layer], max_ind_list[cur_layer],
feat_num[cur_layer], top_nth[cur_layer])
sorted_activations_neg = sorted(max_activation_info, key=lambda x: (x[0], -x[1]))
sorted_activations_pos = sorted(max_activation_info, key=lambda x: (-x[0], -x[1]))
print(Counter(max_labels_list[cur_layer]))
################################### Reconstruct MODEL #####################################
temp = set(tf.all_variables())
set_batch_size(1)
images_placeholder2, labels_placeholder2, max_features_pl2, keep_prob2, filter_num2, image_num2, max_act_pl2, max_ind_pl2 = \
placeholder_inputs2(batch_size)
feed_dict2 = fill_feed_dict2(data_sets.train,
max_features_pl2, None,
1.0,
images_placeholder2,
labels_placeholder2,
keep_prob2,
filter_num2, image_num2, max_act_pl2, max_ind_pl2)
# Build a Graph that computes predictions from the inference model.
returnTensors = autorun_deconv_lasso.select_running_cnn(images_placeholder2,
max_features_pl2,
keep_prob2, hyper_param['layer'] - 1,
filter_num2, image_num2,
max_act_pl2, max_ind_pl2,
mode=autorun_deconv_lasso.DECONV_LASSO,
layer=hyper_param['layer'],
feat=hyper_param['feat'],
cnn_id=model
)
# writer = tf.train.SummaryWriter("/home/e/deconvgraph/deconv_logs", sess.graph)
sess.run(tf.initialize_variables(set(tf.all_variables()) - temp))
################################### Reconstruct MODEL END #####################################
# top 9 positive/negative label reconstructions
for i in range(0, 2):
input_reconstructions = []
input_images = []
batch_nums = []
max_acts = []
max_indicies = []
max_filters = []
for top_nth in range(0, 32000):
if i == 0:
_, max_act_val, batch_num, max_ind_val, filter_num_val, _ = sorted_activations_neg[top_nth]
else:
_, max_act_val, batch_num, max_ind_val, filter_num_val, _ = sorted_activations_pos[top_nth]
all_features_mask = np.zeros(
shape=max_feature_size,
dtype=np.float32)
#ind_max_unraveled = np.unravel_index(max_ind_val , (1,16,16,1))
#ind_max_unraveled2 = np.unravel_index(filter_num_val , (1,16,16,hyper_param['feat'][1]))
#sum_ind = tuple(map(lambda a, b: a + b, ind_max_unraveled, ind_max_unraveled2))
#all_features_mask
all_features_mask[0, :, :, filter_num_val] = 1.0
all_features = np.zeros(
shape=max_feature_size,
dtype=np.float32)
feed_dict2 = fill_feed_dict2(data_sets.train,
max_features_pl2, all_features,
1.0,
images_placeholder2,
labels_placeholder2,
keep_prob2,
filter_num2, image_num2, max_act_pl2, max_ind_pl2,
0, batch_num, 0.0, 0, # placeholders
batch_offset=batch_num) # batch offset
return_tensor_vals = sess.run(returnTensors, feed_dict=feed_dict2)
all_features = return_tensor_vals[hyper_param['layer'] * 2 - 1] * all_features_mask
feed_dict2 = fill_feed_dict2(data_sets.train,
max_features_pl2, all_features,
1.0,
images_placeholder2,
labels_placeholder2,
keep_prob2,
filter_num2, image_num2, max_act_pl2, max_ind_pl2,
0, batch_num, 0.0, 0, # placeholders
batch_offset=batch_num) # batch offset
return_tensor_vals = sess.run(returnTensors, feed_dict=feed_dict2)
image = feed_dict2[images_placeholder2].copy()
max_acts.append(max_act_val)
max_indicies.append(max_ind_val)
max_filters.append(filter_num_val)
batch_nums.append(batch_num)
input_images.append(image) # unpack
input_reconstructions.append(return_tensor_vals[hyper_param['layer'] - 1])
top_nth_reconstructions_np = np.array([input_reconstructions])
top_nth_images_np = np.array([input_images])
save_location = roi_property.SAVE_DIR+'deconv/'
if i == 0:
print('Writing neg10.mat')
spio.savemat(save_location + 'neg10.mat',
dict(recon = top_nth_reconstructions_np, images = top_nth_images_np,
batch_nums=batch_nums, max_acts=max_acts,max_indicies=max_indicies,
max_filters=max_filters))
else:
print('Writing pos10.mat')
spio.savemat(save_location + 'pos10.mat',
dict(recon=top_nth_reconstructions_np, images=top_nth_images_np,
batch_nums=batch_nums, max_acts=max_acts, max_indicies=max_indicies,
max_filters=max_filters))
return
def run_training(hyper_param, model):
global batch_size, max_feature_size
'''
Train RSVP for a number of steps.
Args:
hyper_param: three elements, layer & feat & model
model:
Returns:
'''
# initialize the summary to write
csv_writer_acc, csv_writer_auc = autorun_util.csv_writer(
model, hyper_param['feat'])
# Get the sets of images and labels for training, validation, and
# test on RSVP.
set_batch_size(125)
data_sets = rsvp_input_data.read_data_sets(EEG_DATA_MAT,
FLAGS.fake_data,
reshape_t=False)
# Tell TensorFlow that the model will be built into the default Graph.
with tf.Graph().as_default():
# Generate placeholders for the images and labels.
# Build a Graph that computes predictions from the inference model.
# Generate placeholders for the images and labels.
images_placeholder, labels_placeholder, keep_prob = \
placeholder_inputs(batch_size)
define_learning_rate()
# Build a Graph that computes predictions from the inference model.
results = autorun_deconv_lasso.select_running_cnn(
images_placeholder,
None,
keep_prob,
None,
None,
None,
None,
None,
mode=0,
layer=hyper_param['layer'],
feat=hyper_param['feat'],
cnn_id=model)
logits = results
# Add to the Graph the Ops for loss calculation.
loss = rsvp_quick_cnn_model.loss(logits, labels_placeholder)
# Add to the Graph the Ops that calculate and apply gradients.
train_op = rsvp_quick_cnn_model.training(loss, lr, global_step)
# Add the Op to compare the logits to the labels during evaluation.
eval_correct = rsvp_quick_cnn_model.evaluation(logits,
labels_placeholder)
# Build the summary operation based on the TF collection of Summaries.
summary_op = tf.merge_all_summaries()
# Create a saver for writing training checkpoints.
# Create a session for running Ops on the Graph.
sess = tf.Session()
# Run the Op to initialize the variables.
# Instantiate a SummaryWriter to output summaries and the Graph.
summary_writer = tf.train.SummaryWriter(FLAGS.train_dir,
graph=sess.graph)
saver = tf.train.Saver()
init = tf.initialize_all_variables()
sess.run(init)
checkpoint = tf.train.get_checkpoint_state(FLAGS.train_dir)
if checkpoint and checkpoint.model_checkpoint_path:
saver.restore(sess, checkpoint.model_checkpoint_path)
print("Successfully loaded:", checkpoint.model_checkpoint_path)
else:
print("Could not find old network weights")
# saver.restore(sess, "/home/e/Downloads/models/-2649")
# print("Model restored.")
# And then after everything is built, start the training loop.
for step in range(0, FLAGS.max_steps):
start_time = time.time()
# Fill a feed dictionary with the actual set of images and labels
# for this particular training step.
if mode == autorun_deconv_lasso.TEST:
feed_dict = fill_feed_dict(
data_sets.train,
0.5, #0.0625, # was .5 works good
images_placeholder,
labels_placeholder,
keep_prob)
# Run one step of the model. The return values are the activations
# from the `train_op` (which is discarded) and the `loss` Op. To
# inspect the values of your Ops or variables, you may include them
# in the list passed to sess.run() and the value tensors will be
# returned in the tuple from the call.
_, loss_value = sess.run([train_op, loss], feed_dict=feed_dict)
duration = time.time() - start_time
# Write the summaries and print an overview fairly often.
if step % check_step == 0:
# Print status to stdout.
print('Step %d: loss = %.4f (%.3f sec)' %
(step, loss_value, duration))
# Update the events file.
summary_str = sess.run(summary_op, feed_dict=feed_dict)
summary_writer.add_summary(summary_str, step)
# Save a checkpoint and evaluate the model periodically.
if (step + 1) % check_step == 0 or (step +
1) == FLAGS.max_steps:
saver.save(sess, FLAGS.train_dir, global_step=step)
# Evaluate against the training set.
print('Training Data Eval:')
do_eval(sess, eval_correct, logits, images_placeholder,
labels_placeholder, keep_prob, data_sets.train)
# Evaluate against the validation set.
print('Validation Data Eval:')
#do_eval(sess,
# eval_correct,
# logits,
# images_placeholder,
# labels_placeholder,
# keep_prob,
# filter_num, image_num,
# data_sets.validation)
# Evaluate against the test set.
print('Test Data Eval:')
do_eval(sess, eval_correct, logits, images_placeholder,
labels_placeholder, keep_prob, data_sets.test)
################################### REBUILD MODEL #####################################
temp = set(tf.all_variables())
set_batch_size(1)
images_placeholder2, labels_placeholder2, max_features_pl2, keep_prob2, filter_num2, image_num2, max_act_pl2, max_ind_pl2 = \
placeholder_inputs2(batch_size)
feed_dict2 = fill_feed_dict2(data_sets.train, max_features_pl2, None,
1.0, images_placeholder2,
labels_placeholder2, keep_prob2,
filter_num2, image_num2, max_act_pl2,
max_ind_pl2)
# Build a Graph that computes predictions from the inference model.
returnTensors = autorun_deconv_lasso.select_running_cnn(
images_placeholder2,
max_features_pl2,
keep_prob2,
1,
filter_num2,
image_num2,
max_act_pl2,
max_ind_pl2,
mode=autorun_deconv_lasso.FIND_MAX_ACTIVATION_GET_SWITCHES,
layer=hyper_param['layer'],
feat=hyper_param['feat'],
cnn_id=model)
# writer = tf.train.SummaryWriter("/home/e/deconvgraph/deconv_logs", sess.graph)
sess.run(tf.initialize_variables(set(tf.all_variables()) - temp))
################################### REBUILD MODEL END #####################################
updates = autorun_deconv_lasso.get_update_threshold()
clear_vars = autorun_deconv_lasso.get_clear_variables()
# Find maximum activation
num_layers = hyper_param['layer']
top_nth = []
feat_num = []
max_ind_list = []
max_image_ind_list = []
max_activation_list = []
max_labels_list = []
#select the nth highest feature
for n in range(1000):
for step in range(0, 19000, 20):
print('test pass' + str(step))
feed_dict2 = fill_feed_dict2(
data_sets.train,
max_features_pl2,
None,
1.0,
images_placeholder2,
labels_placeholder2,
keep_prob2,
filter_num2,
image_num2,
max_act_pl2,
max_ind_pl2,
0,
step,
0.0,
0, # placeholders
batch_offset=step) # batch offset
returnTensorVals = sess.run(returnTensors,
feed_dict=feed_dict2)
_ = returnTensorVals
_ = returnTensorVals
# unpack results
max_ind = returnTensorVals[-3 * num_layers:-2 * num_layers]
max_image_ind = returnTensorVals[-2 * num_layers:-num_layers]
max_activation = returnTensorVals[-num_layers:]
# Convert to python lists
if n == 0:
for l in range(0, num_layers):
list_len = len(max_ind[l].tolist())
top_nth.append([n] * list_len)
feat_num.append(range(list_len))
max_ind_list.append(max_ind[l].tolist())
max_image_ind_list.append(max_image_ind[l].tolist())
max_activation_list.append(max_activation[l].tolist())
else:
for l in range(0, num_layers):
list_len = len(max_ind[l].tolist())
top_nth[l].extend([n] * list_len)
feat_num[l].extend(range(list_len))
max_ind_list[l].extend(max_ind[l].tolist())
max_image_ind_list[l].extend(max_image_ind[l].tolist())
max_activation_list[l].extend(max_activation[l].tolist())
# update threshold
returnTensorsTmp = list(returnTensors)
returnTensorsTmp.extend(updates)
# update max threshold
sess.run(returnTensorsTmp, feed_dict=feed_dict2)
# restart at batch zero
feed_dict2 = fill_feed_dict2(
data_sets.train,
max_features_pl2,
None,
1.0,
images_placeholder2,
labels_placeholder2,
keep_prob2,
filter_num2,
image_num2,
max_act_pl2,
max_ind_pl2,
0,
0,
0.0,
0, # placeholders
batch_offset=0) # batch offset
# clear data
returnTensorsTmp = list(returnTensors)
returnTensorsTmp.extend(clear_vars)
# update max threshold
sess.run(returnTensorsTmp, feed_dict=feed_dict2)
for l in range(0, num_layers):
max_labels_np_tmp = data_sets.train.get_labels(
max_image_ind_list[l])
max_labels_list.append(max_labels_np_tmp.tolist())
# find the top 9 activations from all features in top layer
cur_layer = 1
max_activation_info = zip(max_labels_list[cur_layer],
max_activation_list[cur_layer],
max_image_ind_list[cur_layer],
max_ind_list[cur_layer], feat_num[cur_layer],
top_nth[cur_layer])
sorted_activations_neg = sorted(max_activation_info,
key=lambda x: (x[0], -x[1]))
sorted_activations_pos = sorted(max_activation_info,
key=lambda x: (-x[0], -x[1]))
print(Counter(max_labels_list[cur_layer]))
# Reconstruct
################################### REBUILD MODEL #####################################
temp = set(tf.all_variables())
set_batch_size(1)
images_placeholder2, labels_placeholder2, max_features_pl2, keep_prob2, filter_num2, image_num2, max_act_pl2, max_ind_pl2 = \
placeholder_inputs2(batch_size)
feed_dict2 = fill_feed_dict2(data_sets.train, max_features_pl2, None,
1.0, images_placeholder2,
labels_placeholder2, keep_prob2,
filter_num2, image_num2, max_act_pl2,
max_ind_pl2)
# Build a Graph that computes predictions from the inference model.
returnTensors = autorun_deconv_lasso.select_running_cnn(
images_placeholder2,
max_features_pl2,
keep_prob2,
1,
filter_num2,
image_num2,
max_act_pl2,
max_ind_pl2,
mode=autorun_deconv_lasso.DECONV_LASSO,
layer=hyper_param['layer'],
feat=hyper_param['feat'],
cnn_id=model)
# writer = tf.train.SummaryWriter("/home/e/deconvgraph/deconv_logs", sess.graph)
sess.run(tf.initialize_variables(set(tf.all_variables()) - temp))
################################### REBUILD MODEL END #####################################
# top 9 positive/negative label reconstructions
for i in range(0, 2):
input_reconstructions = []
input_images = []
batch_nums = []
max_acts = []
max_indicies = []
max_filters = []
for top_nth in range(0, 10000):
if i == 0:
_, max_act_val, batch_num, max_ind_val, filter_num_val, _ = sorted_activations_neg[
top_nth]
else:
_, max_act_val, batch_num, max_ind_val, filter_num_val, _ = sorted_activations_pos[
top_nth]
all_features_mask = np.zeros(shape=max_feature_size,
dtype=np.float32)
#ind_max_unraveled = np.unravel_index(max_ind_val , (1,16,16,1))
#ind_max_unraveled2 = np.unravel_index(filter_num_val , (1,16,16,hyper_param['feat'][1]))
#sum_ind = tuple(map(lambda a, b: a + b, ind_max_unraveled, ind_max_unraveled2))
#all_features_mask
all_features_mask[0, :, :, filter_num_val] = 1.0
all_features = np.zeros(shape=max_feature_size,
dtype=np.float32)
feed_dict2 = fill_feed_dict2(
data_sets.train,
max_features_pl2,
all_features,
1.0,
images_placeholder2,
labels_placeholder2,
keep_prob2,
filter_num2,
image_num2,
max_act_pl2,
max_ind_pl2,
0,
batch_num,
0.0,
0, # placeholders
batch_offset=batch_num) # batch offset
returnTensorVals = sess.run(returnTensors,
feed_dict=feed_dict2)
all_features = returnTensorVals[hyper_param['layer'] * 2 -
1] * all_features_mask
feed_dict2 = fill_feed_dict2(
data_sets.train,
max_features_pl2,
all_features,
1.0,
images_placeholder2,
labels_placeholder2,
keep_prob2,
filter_num2,
image_num2,
max_act_pl2,
max_ind_pl2,
0,
batch_num,
0.0,
0, # placeholders
batch_offset=batch_num) # batch offset
returnTensorVals = sess.run(returnTensors,
feed_dict=feed_dict2)
image = feed_dict2[images_placeholder2].copy()
max_acts.append(max_act_val)
max_indicies.append(max_ind_val)
max_filters.append(filter_num_val)
batch_nums.append(batch_num)
input_images.append(image) # unpack
input_reconstructions.append(
returnTensorVals[hyper_param['layer'] - 1])
top_nth_reconstructions_np = np.array([input_reconstructions])
top_nth_images_np = np.array([input_images])
save_location = roi_property.SAVE_DIR + 'deconv/'
if i == 0:
print('Writing neg10.mat')
spio.savemat(
save_location + 'neg10.mat',
dict(recon=top_nth_reconstructions_np,
images=top_nth_images_np,
batch_nums=batch_nums,
max_acts=max_acts,
max_indicies=max_indicies,
max_filters=max_filters))
else:
print('Writing pos10.mat')
spio.savemat(
save_location + 'pos10.mat',
dict(recon=top_nth_reconstructions_np,
images=top_nth_images_np,
batch_nums=batch_nums,
max_acts=max_acts,
max_indicies=max_indicies,
max_filters=max_filters))
return
def run_training(hyper_param, model, name_idx, sub_idx):
global batch_size, max_feature_size
'''
Train RSVP for a number of steps.
Args:
hyper_param: three elements, layer & feat & model
model:
Returns:
'''
# initialize the summary to write
csv_writer_acc, csv_writer_auc = autorun_util.csv_writer(model, hyper_param['feat'])
# Get the sets of images and labels for training, validation, and
# test on RSVP.
set_batch_size(roi_property.BATCH_SIZE)
eeg_data = autorun_util.str_name(name_idx, sub_idx)
eeg_data_dir = roi_property.FILE_DIR + \
'rsvp_data/mat_sub/' + eeg_data
eeg_data_mat = eeg_data_dir + '.mat'
data_sets = rsvp_input_data.read_all_data(eeg_data_mat)
# Tell TensorFlow that the model will be built into the default Graph.
with tf.Graph().as_default():
# Generate placeholders for the images and labels.
images_placeholder, labels_placeholder, keep_prob = placeholder_inputs(
FLAGS.batch_size, data_sets.feature_shape[3])
# Build a Graph that computes predictions from the inference model.
logits = autorun_infer.select_running_cnn(images_placeholder,
keep_prob,
layer=hyper_param['layer'],
feat=hyper_param['feat'],
cnn_id=model)
# Add to the Graph the Ops for loss calculation.
loss = rsvp_quick_cnn_model.loss(logits, labels_placeholder)
# Add to the Graph the Ops that calculate and apply gradients.
train_op = rsvp_quick_cnn_model.training(loss, FLAGS.learning_rate)
# Add the Op to compare the logits to the labels during evaluation.
eval_correct = rsvp_quick_cnn_model.evaluation(logits, labels_placeholder)
# Build the summary operation based on the TF collection of Summaries.
# summary_op = tf.merge_all_summaries
summary_op = tf.summary.merge_all()
# Create a saver for writing training checkpoints.
saver = tf.train.Saver()
# Create a session for running Ops on the Graph.
sess = tf.Session()
# Run the Op to initialize the variables.
init = tf.initialize_all_variables()
sess.run(init)
checkpoint = tf.train.get_checkpoint_state(FLAGS.train_dir)
if checkpoint and checkpoint.model_checkpoint_path:
saver.restore(sess, checkpoint.model_checkpoint_path)
print("Successfully loaded:", checkpoint.model_checkpoint_path)
else:
print("Could not find old network weights")
################################### FIND SWITCHES START #####################################
temp = set(tf.all_variables())
images_placeholder2, labels_placeholder2, max_features_pl2, keep_prob2, filter_num2, image_num2, max_act_pl2, max_ind_pl2 = \
placeholder_inputs2(batch_size)
feed_dict2 = fill_feed_dict2(data_sets.train,
max_features_pl2, None,
1.0,
images_placeholder2,
labels_placeholder2,
keep_prob2,
filter_num2, image_num2, max_act_pl2, max_ind_pl2)
# Build a Graph that computes predictions from the inference model.
returnTensors = autorun_deconv_lasso.select_running_cnn(images_placeholder2,
max_features_pl2,
keep_prob2, hyper_param['layer'] - 1,
filter_num2, image_num2,
max_act_pl2, max_ind_pl2,
mode=autorun_deconv_lasso.FIND_MAX_ACTIVATION_GET_SWITCHES,
layer=hyper_param['layer'],
feat=hyper_param['feat'],
cnn_id=model
)
# writer = tf.train.SummaryWriter("/home/e/deconvgraph/deconv_logs", sess.graph)
sess.run(tf.initialize_variables(set(tf.all_variables()) - temp))
# sess.run(tf.global_variables_initializer(tf.global_variables()-temp))
################################### FIND SWITCHES END #####################################
updates = autorun_deconv_lasso.get_update_threshold()
clear_vars = autorun_deconv_lasso.get_clear_variables()
# Find maximum activation
num_layers = hyper_param['layer']
top_nth = []
feat_num = []
max_ind_list = []
max_image_ind_list = []
max_activation_list = []
max_labels_list = []
#select the nth highest feature
for n in range(1000):
for step in range(0, 2):
print('test pass' + str(step))
feed_dict2 = fill_feed_dict2(data_sets.train,
max_features_pl2, None,
1.0,
images_placeholder2,
labels_placeholder2,
keep_prob2,
filter_num2, image_num2, max_act_pl2, max_ind_pl2,
0, step, 0.0, 0, # placeholders
batch_offset=step) # batch offset
return_tensor_vals = sess.run(returnTensors, feed_dict=feed_dict2)
# unpack results
max_ind = return_tensor_vals[-3 * num_layers:-2*num_layers]
max_image_ind = return_tensor_vals[-2*num_layers:-num_layers]
max_activation = return_tensor_vals[-num_layers:]
# Convert to python lists
if n==0:
for l in range(0, num_layers):
list_len = len(max_ind[l].tolist())
top_nth.append([n]*list_len)
feat_num.append(range(list_len))
max_ind_list.append(max_ind[l].tolist())
max_image_ind_list.append(max_image_ind[l].tolist())
max_activation_list.append(max_activation[l].tolist())
else:
for l in range(0, num_layers):
list_len = len(max_ind[l].tolist())
top_nth[l].extend([n]*list_len)
feat_num[l].extend(range(list_len))
max_ind_list[l].extend(max_ind[l].tolist())
max_image_ind_list[l].extend(max_image_ind[l].tolist())
max_activation_list[l].extend(max_activation[l].tolist())
# update threshold
returnTensorsTmp = list(returnTensors)
returnTensorsTmp.extend(updates)
# update max threshold
sess.run(returnTensorsTmp, feed_dict=feed_dict2)
# restart at batch zero
feed_dict2 = fill_feed_dict2(data_sets.train,
max_features_pl2, None,
1.0,
images_placeholder2,
labels_placeholder2,
keep_prob2,
filter_num2, image_num2, max_act_pl2, max_ind_pl2,
0, 0, 0.0, 0, # placeholders
batch_offset=0) # batch offset
# clear data
returnTensorsTmp = list(returnTensors)
returnTensorsTmp.extend(clear_vars)
# update max threshold
sess.run(returnTensorsTmp, feed_dict=feed_dict2)
for l in range(0, num_layers):
max_labels_np_tmp = data_sets.train.get_labels(max_image_ind_list[l])
max_labels_list.append(max_labels_np_tmp.tolist())
# find the top 9 activations from all features in top layer
cur_layer = 1
max_activation_info = zip(max_labels_list[cur_layer], max_activation_list[cur_layer], max_image_ind_list[cur_layer], max_ind_list[cur_layer],
feat_num[cur_layer], top_nth[cur_layer])
sorted_activations_neg = sorted(max_activation_info, key=lambda x: (x[0], -x[1]))
sorted_activations_pos = sorted(max_activation_info, key=lambda x: (-x[0], -x[1]))
print(Counter(max_labels_list[cur_layer]))
################################### Reconstruct MODEL #####################################
temp = set(tf.all_variables())
set_batch_size(1)
images_placeholder2, labels_placeholder2, max_features_pl2, keep_prob2, filter_num2, image_num2, max_act_pl2, max_ind_pl2 = \
placeholder_inputs2(batch_size)
feed_dict2 = fill_feed_dict2(data_sets.train,
max_features_pl2, None,
1.0,
images_placeholder2,
labels_placeholder2,
keep_prob2,
filter_num2, image_num2, max_act_pl2, max_ind_pl2)
# Build a Graph that computes predictions from the inference model.
returnTensors = autorun_deconv_lasso.select_running_cnn(images_placeholder2,
max_features_pl2,
keep_prob2, hyper_param['layer'] - 1,
filter_num2, image_num2,
max_act_pl2, max_ind_pl2,
mode=autorun_deconv_lasso.DECONV_LASSO,
layer=hyper_param['layer'],
feat=hyper_param['feat'],
cnn_id=model
)
# writer = tf.train.SummaryWriter("/home/e/deconvgraph/deconv_logs", sess.graph)
sess.run(tf.initialize_variables(set(tf.all_variables()) - temp))
################################### Reconstruct MODEL END #####################################
# top 9 positive/negative label reconstructions
for i in range(0, 2):
input_reconstructions = []
input_images = []
batch_nums = []
max_acts = []
max_indicies = []
max_filters = []
for top_nth in range(0, 32000):
if i == 0:
_, max_act_val, batch_num, max_ind_val, filter_num_val, _ = sorted_activations_neg[top_nth]
else:
_, max_act_val, batch_num, max_ind_val, filter_num_val, _ = sorted_activations_pos[top_nth]
all_features_mask = np.zeros(
shape=max_feature_size,
dtype=np.float32)
#ind_max_unraveled = np.unravel_index(max_ind_val , (1,16,16,1))
#ind_max_unraveled2 = np.unravel_index(filter_num_val , (1,16,16,hyper_param['feat'][1]))
#sum_ind = tuple(map(lambda a, b: a + b, ind_max_unraveled, ind_max_unraveled2))
#all_features_mask
all_features_mask[0, :, :, filter_num_val] = 1.0
all_features = np.zeros(
shape=max_feature_size,
dtype=np.float32)
feed_dict2 = fill_feed_dict2(data_sets.train,
max_features_pl2, all_features,
1.0,
images_placeholder2,
labels_placeholder2,
keep_prob2,
filter_num2, image_num2, max_act_pl2, max_ind_pl2,
0, batch_num, 0.0, 0, # placeholders
batch_offset=batch_num) # batch offset
return_tensor_vals = sess.run(returnTensors, feed_dict=feed_dict2)
all_features = return_tensor_vals[hyper_param['layer'] * 2 - 1] * all_features_mask
feed_dict2 = fill_feed_dict2(data_sets.train,
max_features_pl2, all_features,
1.0,
images_placeholder2,
labels_placeholder2,
keep_prob2,
filter_num2, image_num2, max_act_pl2, max_ind_pl2,
0, batch_num, 0.0, 0, # placeholders
batch_offset=batch_num) # batch offset
return_tensor_vals = sess.run(returnTensors, feed_dict=feed_dict2)
image = feed_dict2[images_placeholder2].copy()
max_acts.append(max_act_val)
max_indicies.append(max_ind_val)
max_filters.append(filter_num_val)
batch_nums.append(batch_num)
input_images.append(image) # unpack
input_reconstructions.append(return_tensor_vals[hyper_param['layer'] - 1])
top_nth_reconstructions_np = np.array([input_reconstructions])
top_nth_images_np = np.array([input_images])
save_location = roi_property.SAVE_DIR+'deconv/'
if i == 0:
print('Writing neg10.mat')
spio.savemat(save_location + 'neg10.mat',
dict(recon = top_nth_reconstructions_np, images = top_nth_images_np,
batch_nums=batch_nums, max_acts=max_acts,max_indicies=max_indicies,
max_filters=max_filters))
else:
print('Writing pos10.mat')
spio.savemat(save_location + 'pos10.mat',
dict(recon=top_nth_reconstructions_np, images=top_nth_images_np,
batch_nums=batch_nums, max_acts=max_acts, max_indicies=max_indicies,
max_filters=max_filters))
return
