from __future__ import division import math import tensorflow as tf from mi_gru_cell import MiGRUCell from mi_lstm_cell import MiLSTMCell from config import config eps = 1e-20 inf = 1e30 ####################################### variables ######################################## ''' Initializes a weight matrix variable given a shape and a name. Uses random_normal initialization if 1d, otherwise uses xavier. ''' def getWeight(shape, name = ""): with tf.variable_scope("weights"): initializer = tf.contrib.layers.xavier_initializer() # if len(shape) == 1: # good? # initializer = tf.random_normal_initializer() W = tf.get_variable("weight" + name, shape = shape, initializer = initializer) return W ''' Initializes a weight matrix variable given a shape and a name. Uses xavier ''' def getKernel(shape, name = ""): with tf.variable_scope("kernels"): initializer = tf.contrib.layers.xavier_initializer() W = tf.get_variable("kernel" + name, shape = shape, initializer = initializer) return W ''' Initializes a bias variable given a shape and a name. ''' def getBias(shape, name = ""): with tf.variable_scope("biases"): initializer = tf.zeros_initializer() b = tf.get_variable("bias" + name, shape = shape, initializer = initializer) return b ######################################### basics ######################################### ''' Multiplies input inp of any depth by a 2d weight matrix. ''' # switch with conv 1? def multiply(inp, W): inDim = tf.shape(W)[0] outDim = tf.shape(W)[1] newDims = tf.concat([tf.shape(inp)[:-1], tf.fill((1,), outDim)], axis = 0) inp = tf.reshape(inp, (-1, inDim)) output = tf.matmul(inp, W) output = tf.reshape(output, newDims) return output ''' Concatenates x and y. Support broadcasting. Optionally concatenate multiplication of x * y ''' def concat(x, y, dim, mul = False, extendY = False): if extendY: y = tf.expand_dims(y, axis = -2) # broadcasting to have the same shape y = tf.zeros_like(x) + y if mul: out = tf.concat([x, y, x * y], axis = -1) dim *= 3 else: out = tf.concat([x, y], axis = -1) dim *= 2 return out, dim ''' Adds L2 regularization for weight and kernel variables. ''' # add l2 in the tf way def L2RegularizationOp(l2 = None): if l2 is None: l2 = config.l2 l2Loss = 0 names = ["weight", "kernel"] for var in tf.trainable_variables(): if any((name in var.name.lower()) for name in names): l2Loss += tf.nn.l2_loss(var) return l2 * l2Loss ######################################### attention ######################################### ''' Transform vectors to scalar logits. Args: interactions: input vectors [batchSize, N, dim] dim: dimension of input vectors sumMod: LIN for linear transformation to scalars. SUM to sum up vectors entries to get scalar logit. dropout: dropout value over inputs (for linear case) Return matching scalar for each interaction. [batchSize, N] ''' sumMod = ["LIN", "SUM"] def inter2logits(interactions, dim, sumMod = "LIN", dropout = 1.0, name = "", reuse = None): with tf.variable_scope("inter2logits" + name, reuse = reuse): if sumMod == "SUM": logits = tf.reduce_sum(interactions, axis = -1) else: # "LIN" logits = linear(interactions, dim, 1, dropout = dropout, name = "logits") return logits ''' Transforms vectors to probability distribution. Calls inter2logits and then softmax over these. Args: interactions: input vectors [batchSize, N, dim] dim: dimension of input vectors sumMod: LIN for linear transformation to scalars. SUM to sum up vectors entries to get scalar logit. dropout: dropout value over inputs (for linear case) Return attention distribution over interactions. [batchSize, N] ''' def inter2att(interactions, dim, dropout = 1.0, name = "", reuse = None): with tf.variable_scope("inter2att" + name, reuse = reuse): logits = inter2logits(interactions, dim, dropout = dropout) attention = tf.nn.softmax(logits) return attention ''' Sums up features using attention distribution to get a weighted average over them. ''' def att2Smry(attention, features): return tf.reduce_sum(tf.expand_dims(attention, axis = -1) * features, axis = -2) ####################################### activations ######################################## ''' Performs a variant of ReLU based on config.relu PRM for PReLU ELU for ELU LKY for Leaky ReLU otherwise, standard ReLU ''' def relu(inp): if config.relu == "PRM": with tf.variable_scope(None, default_name = "prelu"): alpha = tf.get_variable("alpha", shape = inp.get_shape()[-1], initializer = tf.constant_initializer(0.25)) pos = tf.nn.relu(inp) neg = - (alpha * tf.nn.relu(-inp)) output = pos + neg elif config.relu == "ELU": output = tf.nn.elu(inp) # elif config.relu == "SELU": # output = tf.nn.selu(inp) elif config.relu == "LKY": # output = tf.nn.leaky_relu(inp, config.reluAlpha) output = tf.maximum(inp, config.reluAlpha * inp) elif config.relu == "STD": # STD output = tf.nn.relu(inp) return output activations = { "NON": tf.identity, # lambda inp: inp "TANH": tf.tanh, "SIGMOID": tf.sigmoid, "RELU": relu, "ELU": tf.nn.elu } # Sample from Gumbel(0, 1) def sampleGumbel(shape): U = tf.random_uniform(shape, minval = 0, maxval = 1) return -tf.log(-tf.log(U + eps) + eps) # Draw a clevr_sample from the Gumbel-Softmax distribution def gumbelSoftmaxSample(logits, temperature): y = logits + sampleGumbel(tf.shape(logits)) return tf.nn.softmax(y / temperature) def gumbelSoftmax(logits, temperature, train): # hard = False # Sample from the Gumbel-Softmax distribution and optionally discretize. # Args: # logits: [batch_size, n_class] unnormalized log-probs # temperature: non-negative scalar # hard: if True, take argmax, but differentiate w.r.t. soft clevr_sample y # Returns: # [batch_size, n_class] clevr_sample from the Gumbel-Softmax distribution. # If hard=True, then the returned clevr_sample will be one-hot, otherwise it will # be a probabilitiy distribution that sums to 1 across classes y = gumbelSoftmaxSample(logits, temperature) # k = tf.shape(logits)[-1] # yHard = tf.cast(tf.one_hot(tf.argmax(y,1),k), y.dtype) yHard = tf.cast(tf.equal(y, tf.reduce_max(y, 1, keep_dims = True)), y.dtype) yNew = tf.stop_gradient(yHard - y) + y if config.gumbelSoftmaxBoth: return y if config.gumbelArgmaxBoth: return yNew ret = tf.cond(train, lambda: y, lambda: yNew) return ret def softmaxDiscrete(logits, temperature, train): if config.gumbelSoftmax: return gumbelSoftmax(logits, temperature = temperature, train = train) else: return tf.nn.softmax(logits) def parametricDropout(name, train): var = tf.get_variable("varDp" + name, shape = (), initializer = tf.constant_initializer(2), dtype = tf.float32) dropout = tf.cond(train, lambda: tf.sigmoid(var), lambda: 1.0) return dropout ###################################### sequence helpers ###################################### ''' Casts exponential mask over a sequence with sequence length. Used to prepare logits before softmax. ''' def expMask(seq, seqLength): maxLength = tf.shape(seq)[-1] mask = (1 - tf.cast(tf.sequence_mask(seqLength, maxLength), tf.float32)) * (-inf) masked = seq + mask return masked ''' Computes seq2seq loss between logits and target sequences, with given lengths. ''' def seq2SeqLoss(logits, targets, lengths): mask = tf.sequence_mask(lengths, maxlen = tf.shape(targets)[1]) loss = tf.contrib.seq2seq.sequence_loss(logits, targets, tf.to_float(mask)) return loss ''' Computes seq2seq loss between logits and target sequences, with given lengths. acc1: accuracy per symbol acc2: accuracy per sequence ''' def seq2seqAcc(preds, targets, lengths): mask = tf.sequence_mask(lengths, maxlen = tf.shape(targets)[1]) corrects = tf.logical_and(tf.equal(preds, targets), mask) numCorrects = tf.reduce_sum(tf.to_int32(corrects), axis = 1) acc1 = tf.to_float(numCorrects) / (tf.to_float(lengths) + eps) # add small eps instead? acc1 = tf.reduce_mean(acc1) acc2 = tf.to_float(tf.equal(numCorrects, lengths)) acc2 = tf.reduce_mean(acc2) return acc1, acc2 ########################################### linear ########################################### ''' linear transformation. Args: inp: input to transform inDim: input dimension outDim: output dimension dropout: dropout over input batchNorm: if not None, applies batch normalization to inputs addBias: True to add bias bias: initial bias value act: if not None, activation to use after linear transformation actLayer: if True and act is not None, applies another linear transformation on top of previous actDropout: dropout to apply in the optional second linear transformation retVars: if True, return parameters (weight and bias) Returns linear transformation result. ''' # batchNorm = {"decay": float, "train": Tensor} # actLayer: if activation is not non, stack another linear layer # maybe change naming scheme such that if name = "" than use it as default_name (-->unique?) def linear(inp, inDim, outDim, dropout = 1.0, batchNorm = None, addBias = True, bias = 0.0, act = "NON", actLayer = True, actDropout = 1.0, retVars = False, name = "", reuse = None): with tf.variable_scope("linearLayer" + name, reuse = reuse): W = getWeight((inDim, outDim) if outDim > 1 else (inDim, )) b = getBias((outDim, ) if outDim > 1 else ()) + bias if batchNorm is not None: inp = tf.contrib.layers.batch_norm(inp, decay = batchNorm["decay"], center = True, scale = True, is_training = batchNorm["train"], updates_collections = None) # tf.layers.batch_normalization, axis -1 ? inp = tf.nn.dropout(inp, dropout) if outDim > 1: output = multiply(inp, W) else: output = tf.reduce_sum(inp * W, axis = -1) if addBias: output += b output = activations[act](output) # good? if act != "NON" and actLayer: output = linear(output, outDim, outDim, dropout = actDropout, batchNorm = batchNorm, addBias = addBias, act = "NON", actLayer = False, name = name + "_2", reuse = reuse) if retVars: return (output, (W, b)) return output ''' Computes Multi-layer feed-forward network. Args: features: input features dims: list with dimensions of network. First dimension is of the inputs, final is of the outputs. batchNorm: if not None, applies batchNorm dropout: dropout value to apply for each layer act: activation to apply between layers. NON, TANH, SIGMOID, RELU, ELU ''' # no activation after last layer # batchNorm = {"decay": float, "train": Tensor} def FCLayer(features, dims, batchNorm = None, dropout = 1.0, act = "RELU"): layersNum = len(dims) - 1 for i in range(layersNum): features = linear(features, dims[i], dims[i+1], name = "fc_%d" % i, batchNorm = batchNorm, dropout = dropout) # not the last layer if i < layersNum - 1: features = activations[act](features) return features ###################################### cnns ###################################### ''' Computes convolution. Args: inp: input features inDim: input dimension outDim: output dimension batchNorm: if not None, applies batchNorm on inputs dropout: dropout value to apply on inputs addBias: True to add bias kernelSize: kernel size stride: stride size act: activation to apply on outputs NON, TANH, SIGMOID, RELU, ELU ''' # batchNorm = {"decay": float, "train": Tensor, "center": bool, "scale": bool} # collections.namedtuple("batchNorm", ("decay", "train")) def cnn(inp, inDim, outDim, batchNorm = None, dropout = 1.0, addBias = True, kernelSize = None, stride = 1, act = "NON", name = "", reuse = None): with tf.variable_scope("cnnLayer" + name, reuse = reuse): if kernelSize is None: kernelSize = config.stemKernelSize kernelH = kernelW = kernelSize kernel = getKernel((kernelH, kernelW, inDim, outDim)) b = getBias((outDim, )) if batchNorm is not None: inp = tf.contrib.layers.batch_norm(inp, decay = batchNorm["decay"], center = batchNorm["center"], scale = batchNorm["scale"], is_training = batchNorm["train"], updates_collections = None) inp = tf.nn.dropout(inp, dropout) output = tf.nn.conv2d(inp, filter = kernel, strides = [1, stride, stride, 1], padding = "SAME") if addBias: output += b output = activations[act](output) return output ''' Computes Multi-layer convolutional network. Args: features: input features dims: list with dimensions of network. First dimension is of the inputs. Final is of the outputs. batchNorm: if not None, applies batchNorm dropout: dropout value to apply for each layer kernelSizes: list of kernel sizes for each layer. Default to config.stemKernelSize strides: list of strides for each layer. Default to 1. act: activation to apply between layers. NON, TANH, SIGMOID, RELU, ELU ''' # batchNorm = {"decay": float, "train": Tensor, "center": bool, "scale": bool} # activation after last layer def CNNLayer(features, dims, batchNorm = None, dropout = 1.0, kernelSizes = None, strides = None, act = "RELU"): layersNum = len(dims) - 1 if kernelSizes is None: kernelSizes = [config.stemKernelSize for i in range(layersNum)] if strides is None: strides = [1 for i in range(layersNum)] for i in range(layersNum): features = cnn(features, dims[i], dims[i+1], name = "cnn_%d" % i, batchNorm = batchNorm, dropout = dropout, kernelSize = kernelSizes[i], stride = strides[i], act = act) return features ######################################## location ######################################## ''' Computes linear positional encoding for h x w grid. If outDim positive, casts positions to that dimension. ''' # ignores dim # h,w can be tensor scalars def locationL(h, w, dim, outDim = -1, addBias = True): dim = 2 grid = tf.stack(tf.meshgrid(tf.linspace(-config.locationBias, config.locationBias, w), tf.linspace(-config.locationBias, config.locationBias, h)), axis = -1) if outDim > 0: grid = linear(grid, dim, outDim, addBias = addBias, name = "locationL") dim = outDim return grid, dim ''' Computes sin/cos positional encoding for h x w x (4*dim). If outDim positive, casts positions to that dimension. Based on positional encoding presented in "Attention is all you need" ''' # dim % 4 = 0 # h,w can be tensor scalars def locationPE(h, w, dim, outDim = -1, addBias = True): x = tf.expand_dims(tf.to_float(tf.linspace(-config.locationBias, config.locationBias, w)), axis = -1) y = tf.expand_dims(tf.to_float(tf.linspace(-config.locationBias, config.locationBias, h)), axis = -1) i = tf.expand_dims(tf.to_float(tf.range(dim)), axis = 0) peSinX = tf.sin(x / (tf.pow(10000.0, i / dim))) peCosX = tf.cos(x / (tf.pow(10000.0, i / dim))) peSinY = tf.sin(y / (tf.pow(10000.0, i / dim))) peCosY = tf.cos(y / (tf.pow(10000.0, i / dim))) peSinX = tf.tile(tf.expand_dims(peSinX, axis = 0), [h, 1, 1]) peCosX = tf.tile(tf.expand_dims(peCosX, axis = 0), [h, 1, 1]) peSinY = tf.tile(tf.expand_dims(peSinY, axis = 1), [1, w, 1]) peCosY = tf.tile(tf.expand_dims(peCosY, axis = 1), [1, w, 1]) grid = tf.concat([peSinX, peCosX, peSinY, peCosY], axis = -1) dim *= 4 if outDim > 0: grid = linear(grid, dim, outDim, addBias = addBias, name = "locationPE") dim = outDim return grid, dim locations = { "L": locationL, "PE": locationPE } ''' Adds positional encoding to features. May ease spatial reasoning. (although not used in the default model). Args: features: features to add position encoding to. [batchSize, h, w, c] inDim: number of features' channels lDim: dimension for positional encodings outDim: if positive, cast enhanced features (with positions) to that dimension h: features' height w: features' width locType: L for linear encoding, PE for cos/sin based positional encoding mod: way to add positional encoding: concatenation (CNCT), addition (ADD), multiplication (MUL), linear transformation (LIN). ''' mods = ["CNCT", "ADD", "LIN", "MUL"] # if outDim = -1, then will be set based on inDim, lDim def addLocation(features, inDim, lDim, outDim = -1, h = None, w = None, locType = "L", mod = "CNCT", name = "", reuse = None): # h,w not needed with tf.variable_scope("addLocation" + name, reuse = reuse): batchSize = tf.shape(features)[0] if h is None: h = tf.shape(features)[1] if w is None: w = tf.shape(features)[2] dim = inDim if mod == "LIN": if outDim < 0: outDim = dim grid, _ = locations[locType](h, w, lDim, outDim = outDim, addBias = False) features = linear(features, dim, outDim, name = "LIN") features += grid return features, outDim if mod == "CNCT": grid, lDim = locations[locType](h, w, lDim) # grid = tf.zeros_like(features) + grid grid = tf.tile(tf.expand_dims(grid, axis = 0), [batchSize, 1, 1, 1]) features = tf.concat([features, grid], axis = -1) dim += lDim elif mod == "ADD": grid, _ = locations[locType](h, w, lDim, outDim = dim) features += grid elif mod == "MUL": # MUL grid, _ = locations[locType](h, w, lDim, outDim = dim) if outDim < 0: outDim = dim grid = tf.tile(tf.expand_dims(grid, axis = 0), [batchSize, 1, 1, 1]) features = tf.concat([features, grid, features * grid], axis = -1) dim *= 3 if outDim > 0: features = linear(features, dim, outDim) dim = outDim return features, dim # config.locationAwareEnd # H, W, _ = config.imageDims # projDim = config.stemProjDim # k = config.stemProjPooling # projDim on inDim or on out # inDim = tf.shape(features)[3] ''' Linearize 2d image to linear vector. Args: features: batch of 2d images. [batchSize, h, w, inDim] h: image height w: image width inDim: number of channels projDim: if not None, project image to that dimension before linearization outDim: if not None, project image to that dimension after linearization loc: if not None, add positional encoding: locType: L for linear encoding, PE for cos/sin based positional encoding mod: way to add positional encoding: concatenation (CNCT), addition (ADD), multiplication (MUL), linear transformation (LIN). pooling: number to pool image with before linearization. Returns linearized image: [batchSize, outDim] (or [batchSize, (h / pooling) * (w /pooling) * projDim] if outDim not supported) ''' # loc = {"locType": str, "mod": str} def linearizeFeatures(features, h, w, inDim, projDim = None, outDim = None, loc = None, pooling = None): if pooling is None: pooling = config.imageLinPool if loc is not None: features = addLocation(features, inDim, lDim = inDim, outDim = inDim, locType = loc["locType"], mod = loc["mod"]) if projDim is not None: features = linear(features, dim, projDim) features = relu(features) dim = projDim if pooling > 1: poolingDims = [1, pooling, pooling, 1] features = tf.nn.max_pool(features, ksize = poolingDims, strides = poolingDims, padding = "SAME") h /= pooling w /= pooling dim = h * w * dim features = tf.reshape(features, (-1, dim)) if outDim is not None: features = linear(features, dim, outDim) dim = outDim return features, dim ################################### multiplication ################################### # specific dim / proj for x / y ''' "Enhanced" hadamard product between x and y: 1. Supports optional projection of x, and y prior to multiplication. 2. Computes simple multiplication, or a parametrized one, using diagonal of complete matrix (bi-linear) 3. Optionally concatenate x or y or their projection to the multiplication result. Support broadcasting Args: x: left-hand side argument [batchSize, dim] y: right-hand side argument [batchSize, dim] dim: input dimension of x and y dropout: dropout value to apply on x and y proj: if not None, project x and y: dim: projection dimension shared: use same projection for x and y dropout: dropout to apply to x and y if projected interMod: multiplication type: "MUL": x * y "DIAG": x * W * y for a learned diagonal parameter W "BL": x' W y for a learned matrix W concat: if not None, concatenate x or y or their projection. mulBias: optional bias to stabilize multiplication (x * bias) (y * bias) Returns the multiplication result [batchSize, outDim] when outDim depends on the use of proj and cocnat arguments. ''' # proj = {"dim": int, "shared": bool, "dropout": float} # "act": str, "actDropout": float ## interMod = ["direct", "scalarW", "bilinear"] # "additive" # interMod = ["MUL", "DIAG", "BL", "ADD"] # concat = {"x": bool, "y": bool, "proj": bool} def mul(x, y, dim, dropout = 1.0, proj = None, interMod = "MUL", concat = None, mulBias = None, extendY = True, name = "", reuse = None): with tf.variable_scope("mul" + name, reuse = reuse): origVals = {"x": x, "y": y, "dim": dim} x = tf.nn.dropout(x, dropout) y = tf.nn.dropout(y, dropout) # projection if proj is not None: x = tf.nn.dropout(x, proj.get("dropout", 1.0)) y = tf.nn.dropout(y, proj.get("dropout", 1.0)) if proj["shared"]: xName, xReuse = "proj", None yName, yReuse = "proj", True else: xName, xReuse = "projX", None yName, yReuse = "projY", None x = linear(x, dim, proj["dim"], name = xName, reuse = xReuse) y = linear(y, dim, proj["dim"], name = yName, reuse = yReuse) dim = proj["dim"] projVals = {"x": x, "y": y, "dim": dim} proj["x"], proj["y"] = x, y if extendY: y = tf.expand_dims(y, axis = -2) # broadcasting to have the same shape y = tf.zeros_like(x) + y # multiplication if interMod == "MUL": if mulBias is None: mulBias = config.mulBias output = (x + mulBias) * (y + mulBias) elif interMod == "DIAG": W = getWeight((dim, )) # change initialization? b = getBias((dim, )) activations = x * W * y + b elif interMod == "BL": W = getWeight((dim, dim)) b = getBias((dim, )) output = multiply(x, W) * y + b else: # "ADD" output = tf.tanh(x + y) # concatenation if concat is not None: concatVals = projVals if concat.get("proj", False) else origVals if concat.get("x", False): output = tf.concat([output, concatVals["x"]], axis = -1) dim += concatVals["dim"] if concat.get("y", False): output = ops.concat(output, concatVals["y"], extendY = extendY) dim += concatVals["dim"] return output, dim ######################################## rnns ######################################## ''' Creates an RNN cell. Args: hdim: the hidden dimension of the RNN cell. reuse: whether the cell should reuse parameters or create new ones. cellType: the cell type RNN, GRU, LSTM, MiGRU, MiLSTM, ProjLSTM act: the cell activation NON, TANH, SIGMOID, RELU, ELU projDim: if ProjLSTM, the dimension for the states projection Returns the cell. ''' # tf.nn.rnn_cell.MultiRNNCell([cell(hDim, reuse = reuse) for _ in config.encNumLayers]) # note that config.enc params not general def createCell(hDim, reuse, cellType = None, act = None, projDim = None): if cellType is None: cellType = config.encType activation = activations.get(act, None) if cellType == "ProjLSTM": cell = tf.nn.rnn_cell.LSTMCell if projDim is None: projDim = config.cellDim cell = cell(hDim, num_proj = projDim, reuse = reuse, activation = activation) return cell cells = { "RNN": tf.nn.rnn_cell.BasicRNNCell, "GRU": tf.nn.rnn_cell.GRUCell, "LSTM": tf.nn.rnn_cell.BasicLSTMCell, "MiGRU": MiGRUCell, "MiLSTM": MiLSTMCell } cell = cells[cellType](hDim, reuse = reuse, activation = activation) return cell ''' Runs an forward RNN layer. Args: inSeq: the input sequence to run the RNN over. [batchSize, sequenceLength, inDim] seqL: the sequence matching lengths. [batchSize, 1] hDim: hidden dimension of the RNN. cellType: the cell type RNN, GRU, LSTM, MiGRU, MiLSTM, ProjLSTM dropout: value for dropout over input sequence varDp: if not None, state and input variational dropouts to apply. dimension of input has to be supported (inputSize). Returns the outputs sequence and final RNN state. ''' # varDp = {"stateDp": float, "inputDp": float, "inputSize": int} # proj = {"output": bool, "state": bool, "dim": int, "dropout": float, "act": str} def fwRNNLayer(inSeq, seqL, hDim, cellType = None, dropout = 1.0, varDp = None, name = "", reuse = None): # proj = None with tf.variable_scope("rnnLayer" + name, reuse = reuse): batchSize = tf.shape(inSeq)[0] cell = createCell(hDim, reuse, cellType) # passing reuse isn't mandatory if varDp is not None: cell = tf.contrib.rnn.DropoutWrapper(cell, state_keep_prob = varDp["stateDp"], input_keep_prob = varDp["inputDp"], variational_recurrent = True, input_size = varDp["inputSize"], dtype = tf.float32) else: inSeq = tf.nn.dropout(inSeq, dropout) initialState = cell.zero_state(batchSize, tf.float32) outSeq, lastState = tf.nn.dynamic_rnn(cell, inSeq, sequence_length = seqL, initial_state = initialState, swap_memory = True) if isinstance(lastState, tf.nn.rnn_cell.LSTMStateTuple): lastState = lastState.h # if proj is not None: # if proj["output"]: # outSeq = linear(outSeq, cell.output_size, proj["dim"], act = proj["act"], # dropout = proj["dropout"], name = "projOutput") # if proj["state"]: # lastState = linear(lastState, cell.state_size, proj["dim"], act = proj["act"], # dropout = proj["dropout"], name = "projState") return outSeq, lastState ''' Runs an bidirectional RNN layer. Args: inSeq: the input sequence to run the RNN over. [batchSize, sequenceLength, inDim] seqL: the sequence matching lengths. [batchSize, 1] hDim: hidden dimension of the RNN. cellType: the cell type RNN, GRU, LSTM, MiGRU, MiLSTM dropout: value for dropout over input sequence varDp: if not None, state and input variational dropouts to apply. dimension of input has to be supported (inputSize). Returns the outputs sequence and final RNN state. ''' # varDp = {"stateDp": float, "inputDp": float, "inputSize": int} # proj = {"output": bool, "state": bool, "dim": int, "dropout": float, "act": str} def biRNNLayer(inSeq, seqL, hDim, cellType = None, dropout = 1.0, varDp = None, name = "", reuse = None): # proj = None, with tf.variable_scope("birnnLayer" + name, reuse = reuse): batchSize = tf.shape(inSeq)[0] with tf.variable_scope("fw"): cellFw = createCell(hDim, reuse, cellType) with tf.variable_scope("bw"): cellBw = createCell(hDim, reuse, cellType) if varDp is not None: cellFw = tf.contrib.rnn.DropoutWrapper(cellFw, state_keep_prob = varDp["stateDp"], input_keep_prob = varDp["inputDp"], variational_recurrent = True, input_size = varDp["inputSize"], dtype = tf.float32) cellBw = tf.contrib.rnn.DropoutWrapper(cellBw, state_keep_prob = varDp["stateDp"], input_keep_prob = varDp["inputDp"], variational_recurrent = True, input_size = varDp["inputSize"], dtype = tf.float32) else: inSeq = tf.nn.dropout(inSeq, dropout) initialStateFw = cellFw.zero_state(batchSize, tf.float32) initialStateBw = cellBw.zero_state(batchSize, tf.float32) (outSeqFw, outSeqBw), (lastStateFw, lastStateBw) = tf.nn.bidirectional_dynamic_rnn( cellFw, cellBw, inSeq, sequence_length = seqL, initial_state_fw = initialStateFw, initial_state_bw = initialStateBw, swap_memory = True) if isinstance(lastStateFw, tf.nn.rnn_cell.LSTMStateTuple): lastStateFw = lastStateFw.h # take c? lastStateBw = lastStateBw.h outSeq = tf.concat([outSeqFw, outSeqBw], axis = -1) lastState = tf.concat([lastStateFw, lastStateBw], axis = -1) # if proj is not None: # if proj["output"]: # outSeq = linear(outSeq, cellFw.output_size + cellFw.output_size, # proj["dim"], act = proj["act"], dropout = proj["dropout"], # name = "projOutput") # if proj["state"]: # lastState = linear(lastState, cellFw.state_size + cellFw.state_size, # proj["dim"], act = proj["act"], dropout = proj["dropout"], # name = "projState") return outSeq, lastState # int(hDim / 2) for biRNN? ''' Runs an RNN layer by calling biRNN or fwRNN. Args: inSeq: the input sequence to run the RNN over. [batchSize, sequenceLength, inDim] seqL: the sequence matching lengths. [batchSize, 1] hDim: hidden dimension of the RNN. bi: true to run bidirectional rnn. cellType: the cell type RNN, GRU, LSTM, MiGRU, MiLSTM dropout: value for dropout over input sequence varDp: if not None, state and input variational dropouts to apply. dimension of input has to be supported (inputSize). Returns the outputs sequence and final RNN state. ''' # proj = {"output": bool, "state": bool, "dim": int, "dropout": float, "act": str} # varDp = {"stateDp": float, "inputDp": float, "inputSize": int} def RNNLayer(inSeq, seqL, hDim, bi = None, cellType = None, dropout = 1.0, varDp = None, name = "", reuse = None): # proj = None with tf.variable_scope("rnnLayer" + name, reuse = reuse): if bi is None: bi = config.encBi rnn = biRNNLayer if bi else fwRNNLayer if bi: hDim = int(hDim / 2) return rnn(inSeq, seqL, hDim, cellType = cellType, dropout = dropout, varDp = varDp) # , proj = proj # tf counterpart? # hDim = config.moduleDim def multigridRNNLayer(featrues, h, w, dim, name = "", reuse = None): with tf.variable_scope("multigridRNNLayer" + name, reuse = reuse): featrues = linear(featrues, dim, dim / 2, name = "i") output0 = gridRNNLayer(featrues, h, w, dim, right = True, down = True, name = "rd") output1 = gridRNNLayer(featrues, h, w, dim, right = True, down = False, name = "r") output2 = gridRNNLayer(featrues, h, w, dim, right = False, down = True, name = "d") output3 = gridRNNLayer(featrues, h, w, dim, right = False, down = False, name = "NON") output = tf.concat([output0, output1, output2, output3], axis = -1) output = linear(output, 2 * dim, dim, name = "o") return outputs # h,w should be constants def gridRNNLayer(features, h, w, dim, right, down, name = "", reuse = None): with tf.variable_scope("gridRNNLayer" + name): batchSize = tf.shape(features)[0] cell = createCell(dim, reuse = reuse, cellType = config.stemGridRnnMod, act = config.stemGridAct) initialState = cell.zero_state(batchSize, tf.float32) inputs = [tf.unstack(row, w, axis = 1) for row in tf.unstack(features, h, axis = 1)] states = [[None for _ in range(w)] for _ in range(h)] iAxis = range(h) if down else (range(h)[::-1]) jAxis = range(w) if right else (range(w)[::-1]) iPrev = -1 if down else 1 jPrev = -1 if right else 1 prevState = lambda i,j: states[i][j] if (i >= 0 and i < h and j >= 0 and j < w) else initialState for i in iAxis: for j in jAxis: prevs = tf.concat((prevState(i + iPrev, j), prevState(i, j + jPrev)), axis = -1) curr = inputs[i][j] _, states[i][j] = cell(prevs, curr) outputs = [tf.stack(row, axis = 1) for row in states] outputs = tf.stack(outputs, axis = 1) return outputs # tf seq2seq? # def projRNNLayer(inSeq, seqL, hDim, labels, labelsNum, labelsDim, labelsEmb, name = "", reuse = None): # with tf.variable_scope("projRNNLayer" + name): # batchSize = tf.shape(features)[0] # cell = createCell(hDim, reuse = reuse) # projCell = ProjWrapper(cell, labelsNum, labelsDim, labelsEmb, # config.wrdEmbDim # feedPrev = True, dropout = 1.0, config, # temperature = 1.0, clevr_sample = False, reuse) # initialState = projCell.zero_state(batchSize, tf.float32) # if config.soft: # inSeq = inSeq # # outputs, _ = tf.nn.static_rnn(projCell, inputs, # # sequence_length = seqL, # # initial_state = initialState) # inSeq = tf.unstack(inSeq, axis = 1) # state = initialState # logitsList = [] # chosenList = [] # for inp in inSeq: # (logits, chosen), state = projCell(inp, state) # logitsList.append(logits) # chosenList.append(chosen) # projCell.reuse = True # logitsOut = tf.stack(logitsList, axis = 1) # chosenOut = tf.stack(chosenList, axis = 1) # outputs = (logitsOut, chosenOut) # else: # labels = tf.to_float(labels) # labels = tf.concat([tf.zeros((batchSize, 1)), labels], axis = 1)[:, :-1] # ,newaxis # inSeq = tf.concat([inSeq, tf.expand_dims(labels, axis = -1)], axis = -1) # outputs, _ = tf.nn.dynamic_rnn(projCell, inSeq, # sequence_length = seqL, # initial_state = initialState, # swap_memory = True) # return outputs #, labelsEmb ############################### variational dropout ############################### ''' Generates a variational dropout mask for a given shape and a dropout probability value. ''' def generateVarDpMask(shape, keepProb): randomTensor = tf.to_float(keepProb) randomTensor += tf.random_uniform(shape, minval = 0, maxval = 1) binaryTensor = tf.floor(randomTensor) mask = tf.to_float(binaryTensor) return mask ''' Applies the a variational dropout over an input, given dropout mask and a dropout probability value. ''' def applyVarDpMask(inp, mask, keepProb): ret = (tf.div(inp, tf.to_float(keepProb))) * mask return ret