Merge remote-tracking branch 'upstream/master'

Sync with master.

research/gan/README.md

@@ -23,7 +23,7 @@ Maintainers of TFGAN:
 
 1. [CIFAR10](#cifar10)
 
-1. [Image compression (coming soon)](#compression)
+1. [Image compression](#compression)
 
 ## MNIST
 <a id='mnist'></a>

research/gan/cifar/launch_jobs.sh

@@ -89,7 +89,7 @@ if [[ "$gan_type" == "unconditional" ]]; then
     --dataset_dir=${DATASET_DIR} \
     --eval_real_images=false \
     --conditional_eval=false \
-    --max_number_of_evaluation=1
+    --max_number_of_evaluations=1
   Banner "Finished unconditional evaluation. See ${UNCONDITIONAL_EVAL_DIR} for output images."
 fi
 
@@ -116,6 +116,6 @@ if [[ "$gan_type" == "conditional" ]]; then
     --dataset_dir=${DATASET_DIR} \
     --eval_real_images=false \
     --conditional_eval=true \
-    --max_number_of_evaluation=1
+    --max_number_of_evaluations=1
   Banner "Finished conditional evaluation. See ${CONDITIONAL_EVAL_DIR} for output images."
 fi

research/gan/image_compression/launch_jobs.sh

@@ -80,5 +80,5 @@ python "${git_repo}/research/gan/image_compression/eval.py" \
   --checkpoint_dir=${MODEL_TRAIN_DIR} \
   --eval_dir=${MODEL_EVAL_DIR} \
   --dataset_dir=${DATASET_DIR} \
-  --max_number_of_evaluation=1
+  --max_number_of_evaluations=1
 Banner "Finished evaluation. See ${MODEL_EVAL_DIR} for output images."

research/gan/mnist/launch_jobs.sh

@@ -102,7 +102,7 @@ if [[ "$gan_type" == "unconditional" ]]; then
     --dataset_dir=${DATASET_DIR} \
     --eval_real_images=false \
     --classifier_filename=${FROZEN_GRAPH} \
-    --max_number_of_evaluation=1
+    --max_number_of_evaluations=1
   Banner "Finished unconditional evaluation. See ${UNCONDITIONAL_EVAL_DIR} for output images."
 fi
 
@@ -127,7 +127,7 @@ if [[ "$gan_type" == "conditional" ]]; then
     --checkpoint_dir=${CONDITIONAL_TRAIN_DIR} \
     --eval_dir=${CONDITIONAL_EVAL_DIR} \
     --classifier_filename=${FROZEN_GRAPH} \
-    --max_number_of_evaluation=1
+    --max_number_of_evaluations=1
   Banner "Finished conditional evaluation. See ${CONDITIONAL_EVAL_DIR} for output images."
 fi
 
@@ -152,6 +152,6 @@ if [[ "$gan_type" == "infogan" ]]; then
     --checkpoint_dir=${INFOGAN_TRAIN_DIR} \
     --eval_dir=${INFOGAN_EVAL_DIR} \
     --classifier_filename=${FROZEN_GRAPH} \
-    --max_number_of_evaluation=1
+    --max_number_of_evaluations=1
   Banner "Finished InfoGAN evaluation. See ${INFOGAN_EVAL_DIR} for output images."
 fi

research/lfads/README.md

@@ -7,7 +7,7 @@ This code implements the model from the paper "[LFADS - Latent Factor Analysis v
 
 The code is written in Python 2.7.6. You will also need:
 
-* **TensorFlow** version 1.2.1 ([install](https://www.tensorflow.org/install/)) -
+* **TensorFlow** version 1.5 ([install](https://www.tensorflow.org/install/)) -
 * **NumPy, SciPy, Matplotlib** ([install SciPy stack](https://www.scipy.org/install.html), contains all of them)
 * **h5py** ([install](https://pypi.python.org/pypi/h5py))
 
@@ -98,7 +98,18 @@ $ python run_lfads.py --kind=train \
 --output_filename_stem="" \
 --ic_prior_var_max=0.1 \
 --prior_ar_atau=10.0 \
---do_train_io_only=false
+--do_train_io_only=false \
+--do_train_encoder_only=false
+
+# Run LFADS on chaotic rnn data with no input pulses (g = 1.5) with Gaussian noise
+$ python run_lfads.py --kind=train \
+--data_dir=/tmp/rnn_synth_data_v1.0/ \
+--data_filename_stem=gaussian_chaotic_rnn_no_inputs \
+--lfads_save_dir=/tmp/lfads_chaotic_rnn_inputs_g2p5 \
+--co_dim=1 \
+--factors_dim=20 \
+--output_dist=gaussian
+
 
 # Run LFADS on chaotic rnn data with input pulses (g = 2.5)
 $ python run_lfads.py --kind=train \

research/lfads/lfads.py

@@ -365,7 +365,12 @@ class LFADS(object):
 
       if datasets and 'alignment_matrix_cxf' in datasets[name].keys():
         dataset = datasets[name]
-        print("Using alignment matrix provided for dataset:", name)
+        if hps.do_train_readin:
+            print("Initializing trainable readin matrix with alignment matrix" \
+                  " provided for dataset:", name)
+        else:
+            print("Setting non-trainable readin matrix to alignment matrix" \
+                  " provided for dataset:", name)
         in_mat_cxf = dataset['alignment_matrix_cxf'].astype(np.float32)
         if in_mat_cxf.shape != (data_dim, factors_dim):
           raise ValueError("""Alignment matrix must have dimensions %d x %d
@@ -374,7 +379,12 @@ class LFADS(object):
                             in_mat_cxf.shape[1]))
       if datasets and 'alignment_bias_c' in datasets[name].keys():
         dataset = datasets[name]
-        print("Using alignment bias provided for dataset:", name)
+        if hps.do_train_readin:
+          print("Initializing trainable readin bias with alignment bias " \
+                "provided for dataset:", name)
+        else:
+          print("Setting non-trainable readin bias to alignment bias " \
+                "provided for dataset:", name)
         align_bias_c = dataset['alignment_bias_c'].astype(np.float32)
         align_bias_1xc = np.expand_dims(align_bias_c, axis=0)
         if align_bias_1xc.shape[1] != data_dim:
@@ -387,12 +397,22 @@ class LFADS(object):
           # So b = -alignment_bias * W_in to accommodate PCA style offset.
           in_bias_1xf = -np.dot(align_bias_1xc, in_mat_cxf)
 
-      in_fac_lin = init_linear(data_dim, used_in_factors_dim, do_bias=True,
+      if hps.do_train_readin:
+          # only add to IO transformations collection only if we want it to be
+          # learnable, because IO_transformations collection will be trained
+          # when do_train_io_only
+          collections_readin=['IO_transformations']
+      else:
+          collections_readin=None
+
+      in_fac_lin = init_linear(data_dim, used_in_factors_dim,
+                               do_bias=True,
                                mat_init_value=in_mat_cxf,
                                bias_init_value=in_bias_1xf,
                                identity_if_possible=in_identity_if_poss,
                                normalized=False, name="x_2_infac_"+name,
-                               collections=['IO_transformations'])
+                               collections=collections_readin,
+                               trainable=hps.do_train_readin)
       in_fac_W, in_fac_b = in_fac_lin
       fns_in_fac_Ws[d] = makelambda(in_fac_W)
       fns_in_fac_bs[d] = makelambda(in_fac_b)
@@ -417,7 +437,7 @@ class LFADS(object):
         out_mat_fxc = None
         out_bias_1xc = None
         if in_mat_cxf is not None:
-          out_mat_fxc = np.linalg.pinv(in_mat_cxf)
+            out_mat_fxc = in_mat_cxf.T
         if align_bias_1xc is not None:
           out_bias_1xc = align_bias_1xc
 
@@ -895,13 +915,25 @@ class LFADS(object):
       return
 
     # OPTIMIZATION
-    if not self.hps.do_train_io_only:
+    # train the io matrices only
+    if self.hps.do_train_io_only:
       self.train_vars = tvars = \
-        tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
+        tf.get_collection('IO_transformations',
                           scope=tf.get_variable_scope().name)
+    # train the encoder only
+    elif self.hps.do_train_encoder_only:
+      tvars1 = \
+        tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
+                          scope='LFADS/ic_enc_*')
+      tvars2 = \
+        tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
+                          scope='LFADS/z/ic_enc_*')
+      
+      self.train_vars = tvars = tvars1 + tvars2
+    # train all variables
     else:
       self.train_vars = tvars = \
-        tf.get_collection('IO_transformations',
+        tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
                           scope=tf.get_variable_scope().name)
     print("done.")
     print("Model Variables (to be optimized): ")

research/lfads/run_lfads.py

@@ -23,6 +23,8 @@ import os
 import tensorflow as tf
 import re
 import utils
+import sys
+MAX_INT = sys.maxsize
 
 # Lots of hyperparameters, but most are pretty insensitive.  The
 # explanation of these hyperparameters is found below, in the flags
@@ -35,7 +37,7 @@ OUTPUT_FILENAME_STEM = ""
 DEVICE = "gpu:0" # "cpu:0", or other gpus, e.g. "gpu:1"
 MAX_CKPT_TO_KEEP = 5
 MAX_CKPT_TO_KEEP_LVE = 5
-PS_NEXAMPLES_TO_PROCESS = 1e8 # if larger than number of examples, process all
+PS_NEXAMPLES_TO_PROCESS = MAX_INT # if larger than number of examples, process all
 EXT_INPUT_DIM = 0
 IC_DIM = 64
 FACTORS_DIM = 50
@@ -51,8 +53,10 @@ LEARNING_RATE_STOP = 0.00001
 LEARNING_RATE_N_TO_COMPARE = 6
 INJECT_EXT_INPUT_TO_GEN = False
 DO_TRAIN_IO_ONLY = False
+DO_TRAIN_ENCODER_ONLY = False
 DO_RESET_LEARNING_RATE = False
 FEEDBACK_FACTORS_OR_RATES = "factors"
+DO_TRAIN_READIN = True
 
 # Calibrated just above the average value for the rnn synthetic data.
 MAX_GRAD_NORM = 200.0
@@ -60,7 +64,7 @@ CELL_CLIP_VALUE = 5.0
 KEEP_PROB = 0.95
 TEMPORAL_SPIKE_JITTER_WIDTH = 0
 OUTPUT_DISTRIBUTION = 'poisson' # 'poisson' or 'gaussian'
-NUM_STEPS_FOR_GEN_IC = np.inf # set to num_steps if greater than num_steps
+NUM_STEPS_FOR_GEN_IC = MAX_INT # set to num_steps if greater than num_steps
 
 DATA_DIR = "/tmp/rnn_synth_data_v1.0/"
 DATA_FILENAME_STEM = "chaotic_rnn_inputs_g1p5"
@@ -209,9 +213,9 @@ flags.DEFINE_float("co_prior_var_scale", CO_PRIOR_VAR_SCALE,
                    "Variance of control input prior distribution.")
 
 
-flags.DEFINE_float("prior_ar_atau",  PRIOR_AR_AUTOCORRELATION, 
+flags.DEFINE_float("prior_ar_atau",  PRIOR_AR_AUTOCORRELATION,
                    "Initial autocorrelation of AR(1) priors.")
-flags.DEFINE_float("prior_ar_nvar", PRIOR_AR_PROCESS_VAR, 
+flags.DEFINE_float("prior_ar_nvar", PRIOR_AR_PROCESS_VAR,
                    "Initial noise variance for AR(1) priors.")
 flags.DEFINE_boolean("do_train_prior_ar_atau", DO_TRAIN_PRIOR_AR_ATAU,
                      "Is the value for atau an init, or the constant value?")
@@ -254,13 +258,13 @@ flags.DEFINE_boolean("do_causal_controller",
 # Strictly speaking, feeding either the factors or the rates to the controller
 # violates causality, since the g0 gets to see all the data. This may or may not
 # be only a theoretical concern.
-flags.DEFINE_boolean("do_feed_factors_to_controller", 
-                     DO_FEED_FACTORS_TO_CONTROLLER, 
+flags.DEFINE_boolean("do_feed_factors_to_controller",
+                     DO_FEED_FACTORS_TO_CONTROLLER,
                      "Should factors[t-1] be input to controller at time t?")
 flags.DEFINE_string("feedback_factors_or_rates", FEEDBACK_FACTORS_OR_RATES,
                     "Feedback the factors or the rates to the controller? \
                      Acceptable values: 'factors' or 'rates'.")
-flags.DEFINE_integer("controller_input_lag", CONTROLLER_INPUT_LAG, 
+flags.DEFINE_integer("controller_input_lag", CONTROLLER_INPUT_LAG,
                      "Time lag on the encoding to controller t-lag for \
                      forward, t+lag for reverse.")
 
@@ -312,10 +316,30 @@ flags.DEFINE_boolean("do_train_io_only", DO_TRAIN_IO_ONLY,
                      "Train only the input (readin) and output (readout) \
                      affine functions.")
 
+# This flag is used for an experiment where one wants to know if the dynamics
+# learned by the generator generalize across conditions. In that case, you might
+# train up a model on one set of data, and then only further train the encoder on 
+# another set of data (the conditions to be tested) so that the model is forced
+# to use the same dynamics to describe that data.
+# If you don't care about that particular experiment, this flag should always be
+# false.
+flags.DEFINE_boolean("do_train_encoder_only", DO_TRAIN_ENCODER_ONLY,
+                     "Train only the encoder weights.")
+
 flags.DEFINE_boolean("do_reset_learning_rate", DO_RESET_LEARNING_RATE,
                      "Reset the learning rate to initial value.")
 
 
+# for multi-session "stitching" models, the per-session readin matrices map from
+# neurons to input factors which are fed into the shared encoder. These are
+# initialized by alignment_matrix_cxf and alignment_bias_c in the input .h5
+# files. They can be fixed or made trainable.
+flags.DEFINE_boolean("do_train_readin", DO_TRAIN_READIN, "Whether to train the \
+                     readin matrices and bias vectors. False leaves them fixed \
+                     at their initial values specified by the alignment \
+                     matrices and vectors.")
+
+
 # OVERFITTING
 # Dropout is done on the input data, on controller inputs (from
 # encoder), on outputs from generator to factors.
@@ -429,7 +453,9 @@ def build_model(hps, kind="train", datasets=None):
                 "write_model_params"]:
       print("Possible error!!! You are running ", kind, " on a newly \
       initialized model!")
-      print("Are you sure you sure ", ckpt.model_checkpoint_path, " exists?")
+      # cant print ckpt.model_check_point path if no ckpt
+      print("Are you sure you sure a checkpoint in ", hps.lfads_save_dir,
+            " exists?")
 
     tf.global_variables_initializer().run()
 
@@ -451,7 +477,7 @@ def jsonify_dict(d):
   Creates a shallow-copied dictionary first, then accomplishes string
   conversion.
 
-  Args: 
+  Args:
     d: hyperparameter dictionary
 
   Returns: hyperparameter dictionary with bool's as strings
@@ -535,7 +561,9 @@ def build_hyperparameter_dict(flags):
   d['max_grad_norm'] = flags.max_grad_norm
   d['cell_clip_value'] = flags.cell_clip_value
   d['do_train_io_only'] = flags.do_train_io_only
+  d['do_train_encoder_only'] = flags.do_train_encoder_only
   d['do_reset_learning_rate'] = flags.do_reset_learning_rate
+  d['do_train_readin'] = flags.do_train_readin
 
   # Overfitting
   d['keep_prob'] = flags.keep_prob
@@ -659,7 +687,7 @@ def write_model_parameters(hps, output_fname=None, datasets=None):
   fname = os.path.join(hps.lfads_save_dir, output_fname)
   print("Writing model parameters to: ", fname)
   # save the optimizer params as well
-  model = build_model(hps, kind="write_model_params", datasets=datasets) 
+  model = build_model(hps, kind="write_model_params", datasets=datasets)
   model_params = model.eval_model_parameters(use_nested=False,
                                              include_strs="LFADS")
   utils.write_data(fname, model_params, compression=None)
@@ -775,4 +803,3 @@ def main(_):
 
 if __name__ == "__main__":
     tf.app.run()
- 

research/lfads/synth_data/generate_itb_data.py

@@ -39,7 +39,7 @@ flags.DEFINE_integer("C", 800, "Number of conditions")
 flags.DEFINE_integer("N", 50, "Number of units for the RNN")
 flags.DEFINE_float("train_percentage", 4.0/5.0,
                    "Percentage of train vs validation trials")
-flags.DEFINE_integer("nspikifications", 5,
+flags.DEFINE_integer("nreplications", 5,
                      "Number of spikifications of the same underlying rates.")
 flags.DEFINE_float("tau", 0.025, "Time constant of RNN")
 flags.DEFINE_float("dt", 0.010, "Time bin")
@@ -90,8 +90,8 @@ u_rng = np.random.RandomState(seed=FLAGS.synth_data_seed+1)
 T = FLAGS.T
 C = FLAGS.C
 N = FLAGS.N  # must be same N as in trained model (provided example is N = 50)
-nspikifications = FLAGS.nspikifications
-E = nspikifications * C  # total number of trials
+nreplications = FLAGS.nreplications
+E = nreplications * C  # total number of trials
 train_percentage = FLAGS.train_percentage
 ntimesteps = int(T / FLAGS.dt)
 batch_size = 1  # gives one example per ntrial
@@ -144,7 +144,7 @@ with tf.Session() as sess:
     outputs_t_bxn = np.squeeze(np.asarray(outputs_t_bxn))
     r_sxt = np.dot(P_nxn, states_nxt)
 
-    for s in xrange(nspikifications):
+    for s in xrange(nreplications):
       data_e.append(r_sxt)
       u_e.append(u_1xt)
       outs_e.append(outputs_t_bxn)
@@ -154,7 +154,7 @@ with tf.Session() as sess:
 spiking_data_e = spikify_data(truth_data_e, rng, dt=FLAGS.dt,
                               max_firing_rate=FLAGS.max_firing_rate)
 train_inds, valid_inds = get_train_n_valid_inds(E, train_percentage,
-                                                nspikifications)
+                                                nreplications)
 
 data_train_truth, data_valid_truth = split_list_by_inds(truth_data_e,
                                                         train_inds,
@@ -188,7 +188,7 @@ data = { 'train_truth': data_train_truth,
          'train_data' : data_train_spiking,
          'valid_data' : data_valid_spiking,
          'train_percentage' : train_percentage,
-         'nspikifications' : nspikifications,
+         'nreplications' : nreplications,
          'dt' : FLAGS.dt,
          'u_std' : FLAGS.u_std,
          'max_firing_rate': FLAGS.max_firing_rate,

research/lfads/synth_data/generate_labeled_rnn_data.py

@@ -40,7 +40,7 @@ flags.DEFINE_integer("C", 400, "Number of conditions")
 flags.DEFINE_integer("N", 50, "Number of units for the RNN")
 flags.DEFINE_float("train_percentage", 4.0/5.0,
                    "Percentage of train vs validation trials")
-flags.DEFINE_integer("nspikifications", 10,
+flags.DEFINE_integer("nreplications", 10,
                      "Number of spikifications of the same underlying rates.")
 flags.DEFINE_float("g", 1.5, "Complexity of dynamics")
 flags.DEFINE_float("x0_std", 1.0,
@@ -56,8 +56,8 @@ rnn_rngs = [np.random.RandomState(seed=FLAGS.synth_data_seed+1),
 T = FLAGS.T
 C = FLAGS.C
 N = FLAGS.N
-nspikifications = FLAGS.nspikifications
-E = nspikifications * C
+nreplications = FLAGS.nreplications
+E = nreplications * C
 train_percentage = FLAGS.train_percentage
 ntimesteps = int(T / FLAGS.dt)
 
@@ -77,8 +77,8 @@ condition_labels = []
 condition_number = 0
 for c in range(C):
   x0 = FLAGS.x0_std * rng.randn(N, 1)
-  x0s.append(np.tile(x0, nspikifications))
-  for ns in range(nspikifications):
+  x0s.append(np.tile(x0, nreplications))
+  for ns in range(nreplications):
     condition_labels.append(condition_number)
   condition_number += 1
 x0s = np.concatenate(x0s, axis=1)
@@ -107,7 +107,7 @@ for trial in xrange(E):
 
 # split into train and validation sets
 train_inds, valid_inds = get_train_n_valid_inds(E, train_percentage,
-                                                nspikifications)
+                                                nreplications)
 
 rates_train, rates_valid = split_list_by_inds(rates, train_inds, valid_inds)
 spikes_train, spikes_valid = split_list_by_inds(spikes, train_inds, valid_inds)
@@ -129,7 +129,7 @@ data = {'train_truth': rates_train,
         'train_ext_input' : np.array(ext_input_train),
         'valid_ext_input': np.array(ext_input_valid),
         'train_percentage' : train_percentage,
-        'nspikifications' : nspikifications,
+        'nreplications' : nreplications,
         'dt' : FLAGS.dt,
         'P_sxn' : P_nxn,
         'condition_labels_train' : condition_labels_train,

research/lfads/synth_data/run_generate_synth_data.sh

@@ -19,22 +19,22 @@
 SYNTH_PATH=/tmp/rnn_synth_data_v1.0/
 
  echo "Generating chaotic rnn data with no input pulses (g=1.5) with spiking noise"
- python generate_chaotic_rnn_data.py --save_dir=$SYNTH_PATH --datafile_name=chaotic_rnn_no_inputs --synth_data_seed=5 --T=1.0 --C=400 --N=50 --S=50 --train_percentage=0.8 --nspikifications=10 --g=1.5 --x0_std=1.0 --tau=0.025 --dt=0.01 --input_magnitude=0.0 --max_firing_rate=30.0 --noise_type='poisson'
+ python generate_chaotic_rnn_data.py --save_dir=$SYNTH_PATH --datafile_name=chaotic_rnn_no_inputs --synth_data_seed=5 --T=1.0 --C=400 --N=50 --S=50 --train_percentage=0.8 --nreplications=10 --g=1.5 --x0_std=1.0 --tau=0.025 --dt=0.01 --input_magnitude=0.0 --max_firing_rate=30.0 --noise_type='poisson'
 
 echo "Generating chaotic rnn data with no input pulses (g=1.5) with Gaussian noise"
-python generate_chaotic_rnn_data.py --save_dir=$SYNTH_PATH --datafile_name=chaotic_rnn_no_inputs_gaussian --synth_data_seed=5 --T=1.0 --C=400 --N=50 --S=50 --train_percentage=0.8 --nspikifications=10 --g=1.5 --x0_std=1.0 --tau=0.025 --dt=0.01 --input_magnitude=0.0 --max_firing_rate=30.0 --noise_type='gaussian'
+python generate_chaotic_rnn_data.py --save_dir=$SYNTH_PATH --datafile_name=gaussian_chaotic_rnn_no_inputs --synth_data_seed=5 --T=1.0 --C=400 --N=50 --S=50 --train_percentage=0.8 --nreplications=10 --g=1.5 --x0_std=1.0 --tau=0.025 --dt=0.01 --input_magnitude=0.0 --max_firing_rate=30.0 --noise_type='gaussian'
 
  echo "Generating chaotic rnn data with input pulses (g=1.5)"
- python generate_chaotic_rnn_data.py --save_dir=$SYNTH_PATH --datafile_name=chaotic_rnn_inputs_g1p5 --synth_data_seed=5 --T=1.0 --C=400 --N=50 --S=50 --train_percentage=0.8 --nspikifications=10 --g=1.5 --x0_std=1.0 --tau=0.025 --dt=0.01 --input_magnitude=20.0 --max_firing_rate=30.0 --noise_type='poisson'
+ python generate_chaotic_rnn_data.py --save_dir=$SYNTH_PATH --datafile_name=chaotic_rnn_inputs_g1p5 --synth_data_seed=5 --T=1.0 --C=400 --N=50 --S=50 --train_percentage=0.8 --nreplications=10 --g=1.5 --x0_std=1.0 --tau=0.025 --dt=0.01 --input_magnitude=20.0 --max_firing_rate=30.0 --noise_type='poisson'
 
  echo "Generating chaotic rnn data with input pulses (g=2.5)"
- python generate_chaotic_rnn_data.py --save_dir=$SYNTH_PATH --datafile_name=chaotic_rnn_inputs_g2p5 --synth_data_seed=5 --T=1.0 --C=400 --N=50 --S=50 --train_percentage=0.8 --nspikifications=10 --g=2.5 --x0_std=1.0 --tau=0.025 --dt=0.01 --input_magnitude=20.0 --max_firing_rate=30.0 --noise_type='poisson'
+ python generate_chaotic_rnn_data.py --save_dir=$SYNTH_PATH --datafile_name=chaotic_rnn_inputs_g2p5 --synth_data_seed=5 --T=1.0 --C=400 --N=50 --S=50 --train_percentage=0.8 --nreplications=10 --g=2.5 --x0_std=1.0 --tau=0.025 --dt=0.01 --input_magnitude=20.0 --max_firing_rate=30.0 --noise_type='poisson'
 
  echo "Generate the multi-session RNN data (no multi-session synth example in paper)"
- python generate_chaotic_rnn_data.py --save_dir=$SYNTH_PATH --datafile_name=chaotic_rnn_multisession --synth_data_seed=5 --T=1.0 --C=150 --N=100 --S=20 --npcs=10 --train_percentage=0.8 --nspikifications=40 --g=1.5 --x0_std=1.0 --tau=0.025 --dt=0.01 --input_magnitude=0.0 --max_firing_rate=30.0 --noise_type='poisson'
+ python generate_chaotic_rnn_data.py --save_dir=$SYNTH_PATH --datafile_name=chaotic_rnn_multisession --synth_data_seed=5 --T=1.0 --C=150 --N=100 --S=20 --npcs=10 --train_percentage=0.8 --nreplications=40 --g=1.5 --x0_std=1.0 --tau=0.025 --dt=0.01 --input_magnitude=0.0 --max_firing_rate=30.0 --noise_type='poisson'
 
  echo "Generating Integration-to-bound RNN data"
- python generate_itb_data.py --save_dir=$SYNTH_PATH --datafile_name=itb_rnn --u_std=0.25 --checkpoint_path=SAMPLE_CHECKPOINT --synth_data_seed=5 --T=1.0 --C=800 --N=50 --train_percentage=0.8 --nspikifications=5 --tau=0.025 --dt=0.01 --max_firing_rate=30.0
+ python generate_itb_data.py --save_dir=$SYNTH_PATH --datafile_name=itb_rnn --u_std=0.25 --checkpoint_path=SAMPLE_CHECKPOINT --synth_data_seed=5 --T=1.0 --C=800 --N=50 --train_percentage=0.8 --nreplications=5 --tau=0.025 --dt=0.01 --max_firing_rate=30.0
 
  echo "Generating chaotic rnn data with external input labels (no external input labels example in paper)"
- python generate_labeled_rnn_data.py --save_dir=$SYNTH_PATH --datafile_name=chaotic_rnns_labeled --synth_data_seed=5 --T=1.0 --C=400 --N=50  --train_percentage=0.8 --nspikifications=10 --g=1.5 --x0_std=1.0 --tau=0.025 --dt=0.01 --max_firing_rate=30.0
+ python generate_labeled_rnn_data.py --save_dir=$SYNTH_PATH --datafile_name=chaotic_rnns_labeled --synth_data_seed=5 --T=1.0 --C=400 --N=50  --train_percentage=0.8 --nreplications=10 --g=1.5 --x0_std=1.0 --tau=0.025 --dt=0.01 --max_firing_rate=30.0

research/lfads/synth_data/synthetic_data_utils.py

@@ -176,13 +176,13 @@ def gaussify_data(data_e, rng, dt=1.0, max_firing_rate=100):
 
 
 
-def get_train_n_valid_inds(num_trials, train_fraction, nspikifications):
+def get_train_n_valid_inds(num_trials, train_fraction, nreplications):
   """Split the numbers between 0 and num_trials-1 into two portions for
   training and validation, based on the train fraction.
   Args:
     num_trials: the number of trials
     train_fraction: (e.g. .80)
-    nspikifications: the number of spiking trials per initial condition
+    nreplications: the number of spiking trials per initial condition
   Returns:
     a 2-tuple of two lists: the training indices and validation indices
     """
@@ -192,7 +192,7 @@ def get_train_n_valid_inds(num_trials, train_fraction, nspikifications):
     # This line divides up the trials so that within one initial condition,
     # the randomness of spikifying the condition is shared among both
     # training and validation data splits.
-    if (i % nspikifications)+1 > train_fraction * nspikifications:
+    if (i % nreplications)+1 > train_fraction * nreplications:
       valid_inds.append(i)
     else:
       train_inds.append(i)

research/lfads/utils.py

@@ -84,14 +84,15 @@ def linear(x, out_size, do_bias=True, alpha=1.0, identity_if_possible=False,
 
 def init_linear(in_size, out_size, do_bias=True, mat_init_value=None,
                 bias_init_value=None, alpha=1.0, identity_if_possible=False,
-                normalized=False, name=None, collections=None):
+                normalized=False, name=None, collections=None, trainable=True):
   """Linear (affine) transformation, y = x W + b, for a variety of
   configurations.
 
   Args:
     in_size: The integer size of the non-batc input dimension. [(x),y]
     out_size: The integer size of non-batch output dimension. [x,(y)]
-    do_bias (optional): Add a learnable bias vector to the operation.
+    do_bias (optional): Add a (learnable) bias vector to the operation,
+      if false, b will be None
     mat_init_value (optional): numpy constant for matrix initialization, if None
       , do random, with additional parameters.
     alpha (optional): A multiplicative scaling for the weight initialization
@@ -131,21 +132,22 @@ def init_linear(in_size, out_size, do_bias=True, mat_init_value=None,
     if collections:
       w_collections += collections
     if mat_init_value is not None:
-      w = tf.Variable(mat_init_value, name=wname, collections=w_collections)
+      w = tf.Variable(mat_init_value, name=wname, collections=w_collections,
+                      trainable=trainable)
     else:
       w = tf.get_variable(wname, [in_size, out_size], initializer=mat_init,
-                          collections=w_collections)
+                          collections=w_collections, trainable=trainable)
     w = tf.nn.l2_normalize(w, dim=0) # x W, so xW_j = \sum_i x_bi W_ij
   else:
     w_collections = [tf.GraphKeys.GLOBAL_VARIABLES]
     if collections:
       w_collections += collections
     if mat_init_value is not None:
-      w = tf.Variable(mat_init_value, name=wname, collections=w_collections)
+      w = tf.Variable(mat_init_value, name=wname, collections=w_collections,
+                      trainable=trainable)
     else:
       w = tf.get_variable(wname, [in_size, out_size], initializer=mat_init,
-                          collections=w_collections)
-
+                          collections=w_collections, trainable=trainable)
   b = None
   if do_bias:
     b_collections = [tf.GraphKeys.GLOBAL_VARIABLES]
@@ -155,11 +157,12 @@ def init_linear(in_size, out_size, do_bias=True, mat_init_value=None,
     if bias_init_value is None:
       b = tf.get_variable(bname, [1, out_size],
                           initializer=tf.zeros_initializer(),
-                          collections=b_collections)
+                          collections=b_collections,
+                          trainable=trainable)
     else:
       b = tf.Variable(bias_init_value, name=bname,
-                      collections=b_collections)
-
+                      collections=b_collections,
+                      trainable=trainable)
 
   return (w, b)
 

research/object_detection/g3doc/installation.md

@@ -4,14 +4,15 @@
 
 Tensorflow Object Detection API depends on the following libraries:
 
-*   Protobuf 2.6
-*   Pillow 1.0
-*   lxml
-*   tf Slim (which is included in the "tensorflow/models/research/" checkout)
-*   Jupyter notebook
-*   Matplotlib
-*   Tensorflow
-*   cocoapi
+* Protobuf 2.6
+* Python-tk
+* Pillow 1.0
+* lxml
+* tf Slim (which is included in the "tensorflow/models/research/" checkout)
+* Jupyter notebook
+* Matplotlib
+* Tensorflow
+* cocoapi
 
 For detailed steps to install Tensorflow, follow the [Tensorflow installation
 instructions](https://www.tensorflow.org/install/). A typical user can install
@@ -27,7 +28,7 @@ pip install tensorflow-gpu
 The remaining libraries can be installed on Ubuntu 16.04 using via apt-get:
 
 ``` bash
-sudo apt-get install protobuf-compiler python-pil python-lxml
+sudo apt-get install protobuf-compiler python-pil python-lxml python-tk
 sudo pip install jupyter
 sudo pip install matplotlib
 ```

research/tcn/data_providers.py

@@ -165,7 +165,7 @@ def parse_sequence_example(serialized_example, num_views):
   views = tf.stack([sequence_parse[v] for v in view_names])
   lens = [sequence_parse[v].get_shape().as_list()[0] for v in view_names]
   assert len(set(lens)) == 1
-  seq_len = tf.shape(sequence_parse[v])[0]
+  seq_len = tf.shape(sequence_parse[view_names[-1]])[0]
   return context_parse, views, seq_len
 
 

samples/core/get_started/iris_data.py

@@ -6,7 +6,7 @@ TEST_URL = "http://download.tensorflow.org/data/iris_test.csv"
 
 CSV_COLUMN_NAMES = ['SepalLength', 'SepalWidth',
                     'PetalLength', 'PetalWidth', 'Species']
-SPECIES = ['Sentosa', 'Versicolor', 'Virginica']
+SPECIES = ['Setosa', 'Versicolor', 'Virginica']
 
 def maybe_download():
     train_path = tf.keras.utils.get_file(TRAIN_URL.split('/')[-1], TRAIN_URL)
@@ -35,8 +35,8 @@ def train_input_fn(features, labels, batch_size):
     # Shuffle, repeat, and batch the examples.
     dataset = dataset.shuffle(1000).repeat().batch(batch_size)
 
-    # Return the read end of the pipeline.
-    return dataset.make_one_shot_iterator().get_next()
+    # Return the dataset.
+    return dataset
 
 
 def eval_input_fn(features, labels, batch_size):
@@ -55,8 +55,8 @@ def eval_input_fn(features, labels, batch_size):
     assert batch_size is not None, "batch_size must not be None"
     dataset = dataset.batch(batch_size)
 
-    # Return the read end of the pipeline.
-    return dataset.make_one_shot_iterator().get_next()
+    # Return the dataset.
+    return dataset
 
 
 # The remainder of this file contains a simple example of a csv parser,
@@ -89,5 +89,5 @@ def csv_input_fn(csv_path, batch_size):
     # Shuffle, repeat, and batch the examples.
     dataset = dataset.shuffle(1000).repeat().batch(batch_size)
 
-    # Return the read end of the pipeline.
-    return dataset.make_one_shot_iterator().get_next()
+    # Return the dataset.
+    return dataset
\ No newline at end of file

tutorials/image/cifar10_estimator/generate_cifar10_tfrecords.py

@@ -24,10 +24,10 @@ from __future__ import division
 from __future__ import print_function
 
 import argparse
-import cPickle
 import os
 
 import tarfile
+from six.moves import cPickle as pickle
 from six.moves import xrange  # pylint: disable=redefined-builtin
 import tensorflow as tf
 
@@ -49,7 +49,7 @@ def _int64_feature(value):
 
 
 def _bytes_feature(value):
-  return tf.train.Feature(bytes_list=tf.train.BytesList(value=[str(value)]))
+  return tf.train.Feature(bytes_list=tf.train.BytesList(value=[value]))
 
 
 def _get_file_names():
@@ -62,8 +62,8 @@ def _get_file_names():
 
 
 def read_pickle_from_file(filename):
-  with tf.gfile.Open(filename, 'r') as f:
-    data_dict = cPickle.load(f)
+  with tf.gfile.Open(filename, 'rb') as f:
+    data_dict = pickle.load(f)
   return data_dict
 
 

tutorials/rnn/ptb/ptb_word_lm.py

@@ -225,15 +225,15 @@ class PTBModel(object):
 
     self._initial_state = cell.zero_state(config.batch_size, data_type())
     state = self._initial_state
-    # Simplified version of tensorflow_models/tutorials/rnn/rnn.py's rnn().
+    # Simplified version of tf.nn.static_rnn().
     # This builds an unrolled LSTM for tutorial purposes only.
-    # In general, use the rnn() or state_saving_rnn() from rnn.py.
+    # In general, use tf.nn.static_rnn() or tf.nn.static_state_saving_rnn().
     #
     # The alternative version of the code below is:
     #
-    # inputs = tf.unstack(inputs, num=num_steps, axis=1)
-    # outputs, state = tf.contrib.rnn.static_rnn(cell, inputs,
-    #                            initial_state=self._initial_state)
+    # inputs = tf.unstack(inputs, num=self.num_steps, axis=1)
+    # outputs, state = tf.nn.static_rnn(cell, inputs,
+    #                                   initial_state=self._initial_state)
     outputs = []
     with tf.variable_scope("RNN"):
       for time_step in range(self.num_steps):