def adadelta(ips,cost,fupdate,names,parameters,gradients,lr,norm_lim,rho=0.95,eps=1e-6): zipped_grads = [theano.shared(p.get_value() * numpy.float32(0.), name='%s_grad'%k) for k, p in zip(names, parameters)] running_up2 = [theano.shared(p.get_value() * numpy.float32(0.), name='%s_rup2'%k) for k, p in zip(names, parameters)] running_grads2 = [theano.shared(p.get_value() * numpy.float32(0.), name='%s_rgrad2'%k) for k, p in zip(names, parameters)] zgup = [(zg, g) for zg, g in zip(zipped_grads, gradients)] rg2up = [(rg2, rho * rg2 + (1. - rho) * (g ** 2)) for rg2, g in zip(running_grads2, gradients)] update_map = fupdate if fupdate else OrderedUpdates() for kk, vv in zgup: update_map[kk] = vv for kk, vv in rg2up: update_map[kk] = vv f_grad_shared = theano.function(ips, cost, updates=update_map, on_unused_input='ignore') updir = [-T.sqrt(ru2 + eps) / T.sqrt(rg2 + eps) * zg for zg, ru2, rg2 in zip(zipped_grads, running_up2, running_grads2)] ru2up = [(ru2, rho * ru2 + (1. - rho) * (ud ** 2)) for ru2, ud in zip(running_up2, updir)] param_up = [(p, p + ud) for p, ud in zip(parameters, updir)] if norm_lim > 0: param_up = clipGradient(param_up, norm_lim, names) #update_map = fupdate if fupdate else OrderedUpdates() #for kk, vv in ru2up: update_map[kk] = vv #for kk, vv in param_up: update_map[kk] = vv f_param_update = theano.function([lr], [], updates=ru2up+param_up, on_unused_input='ignore') return f_grad_shared, f_param_update
def test_updates_add(self): up1 = OrderedUpdates() up2 = OrderedUpdates() a = theano.shared('a') b = theano.shared('b') assert not up1 + up2 up1[a] = 5 # test that addition works assert up1 assert up1 + up2 assert not up2 assert len(up1 + up2) == 1 assert (up1 + up2)[a] == 5 up2[b] = 7 assert up1 assert up1 + up2 assert up2 assert len(up1 + up2) == 2 assert (up1 + up2)[a] == 5 assert (up1 + up2)[b] == 7 assert a in (up1 + up2) assert b in (up1 + up2) # this works even though there is a collision # because values all match assert len(up1 + up1 + up1) == 1 up2[a] = 8 # a gets different value in up1 and up2 try: up1 + up2 assert 0 except KeyError: pass # reassigning to a key works fine right? up2[a] = 10
def test_updates_setitem(self): up = OrderedUpdates() # keys have to be SharedVariables self.assertRaises(TypeError, up.__setitem__, 5, 7) self.assertRaises(TypeError, up.__setitem__, T.vector(), 7) up[theano.shared(88)] = 7
def test_updates_setitem(self): up = OrderedUpdates() # keys have to be SharedVariables with pytest.raises(TypeError): up.__setitem__(5, 7) with pytest.raises(TypeError): up.__setitem__(tt.vector(), 7) up[theano.shared(88)] = 7
def sgd(ips,cost,fupdate,names,parameters,gradients,lr,norm_lim): gshared = [theano.shared(p.get_value() * 0., name='%s_grad'%k) for k, p in zip(names, parameters)] gsup = [(gs, g) for gs, g in zip(gshared, gradients)] update_map = fupdate if fupdate else OrderedUpdates() for kk, vv in gsup: update_map[kk] = vv f_grad_shared = theano.function(ips, cost, updates=update_map, on_unused_input='ignore') pup = [(p, p - lr * g) for p, g in zip(parameters, gshared)] if norm_lim > 0: pup = clipGradient(pup, norm_lim, names) #update_map = fupdate if fupdate else OrderedUpdates() #for kk, vv in pup: update_map[kk] = vv f_param_update = theano.function([lr], [], updates=pup, on_unused_input='ignore') return f_grad_shared, f_param_update
def set_objective(self, loss, params, inputs=None, updts=None, grads=None, compilation_mode=None, **kwargs): ''' Changes the objective function to be optimized @param loss theano graph representing the loss to be optimized @param params theano shared variables representing the parameters to be optimized @param inputs theano variables representing the inputs required to compute the loss, other than params @param updts dictionary of list of theano updates to be applied after every evaluation of the loss function @param grads gradients of the loss function. If not provided, will be computed here ''' if inputs is None: inputs = [] if updts is not None: updts = OrderedUpdates(updts) if grads is None: utils.print_with_stamp('Building computation graph for gradients', self.name) grads = theano.grad(loss, params) utils.print_with_stamp('Compiling function for loss', self.name) self.loss_fn = theano.function( inputs, loss, updates=updts, allow_input_downcast=True, mode=compilation_mode) utils.print_with_stamp('Compiling function for loss+gradients', self.name) self.grads_fn = theano.function( inputs, [loss, ]+grads, updates=updts, allow_input_downcast=True, mode=compilation_mode) self.n_evals = 0 self.start_time = 0 self.iter_time = 0 self.params = params
def finish_scan(fn_outputs, local_vars): n_fixed_steps = local_vars["n_fixed_steps"] return_steps = local_vars["return_steps"] non_seqs = local_vars["non_seqs"] dummy_args = local_vars["dummy_args"] args = local_vars["args"] outs_info = local_vars["outs_info"] n_outs = local_vars["n_outs"] mit_sot_inner_outputs = local_vars["mit_sot_inner_outputs"] sit_sot_inner_outputs = local_vars["sit_sot_inner_outputs"] sit_sot_scan_inputs = local_vars["sit_sot_scan_inputs"] sit_sot_inner_inputs = local_vars["sit_sot_inner_inputs"] actual_n_steps = local_vars["actual_n_steps"] sit_sot_rightOrder = local_vars["sit_sot_rightOrder"] strict = local_vars["strict"] non_sequences = local_vars["non_sequences"] inner_seqs = local_vars["inner_seqs"] mit_mot_inner_inputs = local_vars["mit_mot_inner_inputs"] mit_sot_inner_inputs = local_vars["mit_sot_inner_inputs"] mit_mot_inner_outputs = local_vars["mit_mot_inner_outputs"] mit_sot_tap_array = local_vars["mit_sot_tap_array"] allow_gc = local_vars["allow_gc"] n_seqs = local_vars["n_seqs"] n_mit_mot_outs = local_vars["n_mit_mot_outs"] mit_mot_out_slices = local_vars["mit_mot_out_slices"] truncate_gradient = local_vars["truncate_gradient"] name = local_vars["name"] mode = local_vars["mode"] profile = local_vars["profile"] scan_seqs = local_vars["scan_seqs"] mit_mot_scan_inputs = local_vars["mit_mot_scan_inputs"] mit_sot_scan_inputs = local_vars["mit_sot_scan_inputs"] n_mit_mot = local_vars["n_mit_mot"] mit_sot_return_steps = local_vars["mit_sot_return_steps"] n_mit_sot = local_vars["n_mit_sot"] sit_sot_return_steps = local_vars["sit_sot_return_steps"] mit_sot_rightOrder = local_vars["mit_sot_rightOrder"] condition, outputs, updates = scan_utils.get_updates_and_outputs( fn_outputs) ################################################################## P2> if condition is not None: as_while = True else: as_while = False ## # Step 3. Check if we actually need scan and remove it if we don't ## if n_fixed_steps in [1, -1]: # We do not need to use the scan op anymore, so we can just return # the outputs and updates we have if condition is not None: _logger.warning(('When the number of steps is fixed and equal ' 'to 1, the provided stopping condition, ', str(condition), ' is ignored')) for pos, inner_out in enumerate(outputs): # we need to see if we need to pad our sequences with an # unbroadcastable dimension; case example : we return an # output for which we want all intermediate. If n_steps is 1 # then, if we return the output as given by the innner function # this will represent only a slice and it will have one # dimension less. if (isinstance(inner_out.type, tensor.TensorType) and return_steps.get(pos, 0) != 1): outputs[pos] = tensor.unbroadcast( tensor.shape_padleft(inner_out), 0) if len(outputs) == 1: outputs = outputs[0] return (outputs, updates) ## # Step 4. Compile the dummy function ## # We can now compile a dummy function just to see what shared variable # we have and what are their update rules (note that the user has # the option not to pass the shared variable to scan, so we need to # pick them manually and add them to scan) # make the compilation as fast as possible by not applying any # optimization or conversion to C [ note this region is not important # for performance so we can do stuff as unoptimal as we wish ] # extract still missing inputs (there still might be so) and add them # as non sequences at the end of our args fake_nonseqs = [x.type() for x in non_seqs] fake_outputs = scan_utils.clone(outputs, replace=OrderedDict( izip(non_seqs, fake_nonseqs))) all_inputs = ifilter( lambda x: (isinstance(x, gof.Variable) and not isinstance( x, SharedVariable) and not isinstance(x, gof.Constant)), gof.graph.inputs(fake_outputs)) extra_inputs = [x for x in all_inputs if x not in args + fake_nonseqs] non_seqs += extra_inputs # Note we do not use all_inputs directly since the order of variables # in args is quite important dummy_args += extra_inputs dummy_outs = outputs if condition is not None: dummy_outs.append(condition) dummy_f = function(dummy_args, dummy_outs, updates=updates, mode=compile.mode.Mode(linker='py', optimizer=None), on_unused_input='ignore', profile=False) ## # Step 5. Re-arange inputs of scan into a more strict order ## # Step 5.0 Check the outputs of the dummy function to see if they # match with user provided data # if the number of outputs to the function does not match the number of # assumed outputs until now (provided by the user) there can be # only one explanation: No information is provided for any of the # outputs (i.e. we are dealing with a map) tmp_dummy_f_outs = len(dummy_f.maker.outputs) if as_while: tmp_dummy_f_outs -= 1 if not (tmp_dummy_f_outs == n_outs or outs_info == []): raise ValueError('Please provide None as outputs_info for ' 'any output that does not feed back into ' 'scan (i.e. it behaves like a map) ') if outs_info == []: n_outs = len(dummy_f.maker.outputs) if as_while: n_outs = n_outs - 1 outs_info = [OrderedDict() for x in xrange(n_outs)] # Step 5.1 Outputs with taps different then -1 for i, out in enumerate(outs_info): if 'taps' in out and out['taps'] != [-1]: mit_sot_inner_outputs.append(outputs[i]) # Step 5.2 Outputs with tap equal to -1 for i, out in enumerate(outs_info): if 'taps' in out and out['taps'] == [-1]: sit_sot_inner_outputs.append(outputs[i]) # Step 5.3 Outputs that correspond to update rules of shared variables givens = OrderedDict() n_shared_outs = 0 shared_scan_inputs = [] shared_inner_inputs = [] shared_inner_outputs = [] sit_sot_shared = [] for input in dummy_f.maker.expanded_inputs: if isinstance(input.variable, SharedVariable) and input.update: new_var = safe_new(input.variable) if getattr(input.variable, 'name', None) is not None: new_var.name = input.variable.name + '_copy' if isinstance(new_var.type, ops.expandable_types): sit_sot_inner_inputs.append(new_var) sit_sot_scan_inputs.append( scan_utils.expand_empty( tensor.unbroadcast( tensor.shape_padleft(input.variable), 0), actual_n_steps)) tensor_update = tensor.as_tensor_variable(input.update) sit_sot_inner_outputs.append(tensor_update) # Not that pos is not a negative index. The sign of pos is used # as a flag to indicate if this output should be part of the # update rules or part of the standard outputs of scan. # If `pos` is positive than it corresponds to the standard # outputs of scan and it refers to output of index `pos`. If `pos` # is negative that it corresponds to update rules of scan and it # refers to update rule of index -1 - `pos`. sit_sot_rightOrder.append(-1 - len(sit_sot_shared)) sit_sot_shared.append(input.variable) givens[input.variable] = new_var else: shared_inner_inputs.append(new_var) shared_scan_inputs.append(input.variable) shared_inner_outputs.append(input.update) givens[input.variable] = new_var n_shared_outs += 1 n_sit_sot = len(sit_sot_inner_inputs) # Step 5.4 Outputs with no taps used in the input n_nit_sot = 0 nit_sot_inner_outputs = [] nit_sot_return_steps = OrderedDict() nit_sot_rightOrder = [] for i, out in enumerate(outs_info): if not 'taps' in out: nit_sot_inner_outputs.append(outputs[i]) if i in return_steps: nit_sot_return_steps[n_nit_sot] = return_steps[i] nit_sot_rightOrder.append(i) n_nit_sot += 1 # Step 5.5 all other arguments including extra inputs other_scan_args = [] other_inner_args = [] other_scan_args += [ arg for arg in non_seqs if (not isinstance(arg, SharedVariable) and not isinstance(arg, tensor.Constant)) ] # Step 5.6 all shared variables with no update rules other_inner_args += [ safe_new(arg, '_copy') for arg in non_seqs if (not isinstance(arg, SharedVariable) and not isinstance(arg, tensor.Constant)) ] givens.update(OrderedDict(izip(other_scan_args, other_inner_args))) if strict: non_seqs_set = set(non_sequences if non_sequences is not None else []) other_shared_scan_args = [ arg.variable for arg in dummy_f.maker.expanded_inputs if (isinstance(arg.variable, SharedVariable) and not arg.update and arg.variable in non_seqs_set) ] other_shared_inner_args = [ safe_new(arg.variable, '_copy') for arg in dummy_f.maker.expanded_inputs if (isinstance(arg.variable, SharedVariable) and not arg.update and arg.variable in non_seqs_set) ] else: other_shared_scan_args = [ arg.variable for arg in dummy_f.maker.expanded_inputs if (isinstance(arg.variable, SharedVariable) and not arg.update) ] other_shared_inner_args = [ safe_new(arg.variable, '_copy') for arg in dummy_f.maker.expanded_inputs if (isinstance(arg.variable, SharedVariable) and not arg.update) ] givens.update( OrderedDict(izip(other_shared_scan_args, other_shared_inner_args))) ## # Step 6. Re-order the outputs and clone them replacing things # using the givens ## inner_inputs = (inner_seqs + mit_mot_inner_inputs + mit_sot_inner_inputs + sit_sot_inner_inputs + shared_inner_inputs + other_shared_inner_args + other_inner_args) inner_outs = (mit_mot_inner_outputs + mit_sot_inner_outputs + sit_sot_inner_outputs + nit_sot_inner_outputs + shared_inner_outputs) if condition is not None: inner_outs.append(condition) # Cuda and Gpuarray are imported here, instead of being imported on top of # the file because that would force on the user some dependencies that we # might do not want to. Currently we are working on removing the # dependencies on sandbox code completeley. from theano.sandbox import cuda, gpuarray if cuda.cuda_available or gpuarray.pygpu_activated: # very often we end up in this situation when we want to # replace w with w_copy, where w is a GPU variable # and w_copy is TensorType. This is caused because shared # variables are put on GPU right aways >:| , new_givens = OrderedDict() for w, w_copy in iteritems(givens): if ((isinstance(w.type, cuda.CudaNdarrayType) or isinstance(w.type, gpuarray.GpuArrayType)) and isinstance(w_copy.type, tensor.TensorType)): for o in inner_outs: new_givens = traverse(o, w, w_copy, new_givens) else: new_givens[w] = w_copy else: new_givens = givens new_outs = scan_utils.clone(inner_outs, replace=new_givens) ## # Step 7. Create the Scan Op ## tap_array = mit_sot_tap_array + [[-1] for x in xrange(n_sit_sot)] if allow_gc is None: allow_gc = config.scan.allow_gc info = OrderedDict() info['tap_array'] = tap_array info['n_seqs'] = n_seqs info['n_mit_mot'] = n_mit_mot info['n_mit_mot_outs'] = n_mit_mot_outs info['mit_mot_out_slices'] = mit_mot_out_slices info['n_mit_sot'] = n_mit_sot info['n_sit_sot'] = n_sit_sot info['n_shared_outs'] = n_shared_outs info['n_nit_sot'] = n_nit_sot info['truncate_gradient'] = truncate_gradient info['name'] = name info['mode'] = mode info['destroy_map'] = OrderedDict() info['gpu'] = False info['as_while'] = as_while info['profile'] = profile info['allow_gc'] = allow_gc info['strict'] = strict local_op = scan_op.Scan(inner_inputs, new_outs, info) ## # Step 8. Compute the outputs using the scan op ## _scan_inputs = (scan_seqs + mit_mot_scan_inputs + mit_sot_scan_inputs + sit_sot_scan_inputs + shared_scan_inputs + [actual_n_steps for x in xrange(n_nit_sot)] + other_shared_scan_args + other_scan_args) scan_inputs = [] for arg in [actual_n_steps] + _scan_inputs: try: arg = tensor.as_tensor_variable(arg) except TypeError: # This happens for Random States for e.g. but it is a good way # to make sure no input is a cuda ndarrays pass scan_inputs += [arg] scan_outs = local_op(*scan_inputs) if type(scan_outs) not in (list, tuple): scan_outs = [scan_outs] ## # Step 9. Figure out which outs are update rules for shared variables # and so on ... ## update_map = OrderedUpdates() def remove_dimensions(outs, steps_return, offsets=None): out_ls = [] for idx, out in enumerate(outs): if idx in steps_return: if steps_return[idx] > 1: out_ls.append(out[-steps_return[idx]:]) else: out_ls.append(out[-1]) else: if offsets is None: out_ls.append(out) else: out_ls.append(out[offsets[idx]:]) return out_ls offset = n_mit_mot offsets = [abs(numpy.min(x)) for x in mit_sot_tap_array] mit_sot_outs = remove_dimensions(scan_outs[offset:offset + n_mit_sot], mit_sot_return_steps, offsets) offset += n_mit_sot offsets = [1 for x in xrange(n_sit_sot)] sit_sot_outs = remove_dimensions(scan_outs[offset:offset + n_sit_sot], sit_sot_return_steps, offsets) offset += n_sit_sot nit_sot_outs = remove_dimensions(scan_outs[offset:offset + n_nit_sot], nit_sot_return_steps) offset += n_nit_sot for idx, update_rule in enumerate(scan_outs[offset:offset + n_shared_outs]): update_map[shared_scan_inputs[idx]] = update_rule _scan_out_list = (mit_sot_outs + sit_sot_outs + nit_sot_outs) # Step 10. I need to reorder the outputs to be in the order expected by # the user rightOrder = (mit_sot_rightOrder + sit_sot_rightOrder + nit_sot_rightOrder) scan_out_list = [None] * len(rightOrder) for idx, pos in enumerate(rightOrder): if pos >= 0: scan_out_list[pos] = _scan_out_list[idx] else: # Not that pos is not a negative index. The sign of pos is used # as a flag to indicate if this output should be part of the # update rules or part of the standard outputs of scan. # If `pos` is positive than it corresponds to the standard # outputs of scan and it refers to output of index `pos`. If `pos` # is negative that it corresponds to update rules of scan and it # refers to update rule of index -1 - `pos`. update_map[sit_sot_shared[abs(pos) - 1]] = _scan_out_list[idx][-1] scan_out_list = [x for x in scan_out_list if x is not None] ################################################################## P2< return (scan_out_list, update_map)
def scan(fn, sequences=None, states=None, params=None, n_steps=None, mode=None, name=None, profile=False, allow_gc=None): """ Similar to Theano's official scan, this function gives the user more control over the scan op, avoiding certain difficulties that arose from missing optimizations. :param fn: lambda function that describes one step of scan (see the official Theano scan function) :param sequences: similar to the official Theano's scan. This version of scan does not support taps for the sequences (it can only be a list of tensor). Scan assumes that sequences have the right length and it does not check for this. :param states: similar to outputs_info of the official scan function. There is one crucial difference though, namely that the `initial` key in the dictionary has been replace by 'membuf' key. This reflects the change of meaning. Instead of passing to scan just the initial steps misisng, one has now to pass a memory buffer in which scan will try to store its output. In this memory buffer the first entries should be set to the initial states of the corresponding states. Providing a memory buffer that has less entries then the number of steps, mneans scan will only use that amount of memory. The user has to match the memory buffer size with the number of steps, otherwise scan will produce wrong results. Also if gradients are to be computed through the scan, the memory buffer should have the same length as the number of steps. For states that do not require a initial state, one has to provide a dictionary with a single key 'steps' that says how many intermediate results to store. See examples below for more insight. :param n_steps: This parameter is mandatory and it will represent the number of steps scan will do (scan will not check sequences or any other source of information to figure out how many steps it needs to do). :param mode: Same as for the official scan :param name: Same as for the official scan :param profile: Same as for the official scan Note: - there is no truncate / go_backwards anymore ! - the outputs returned by scan contain the initial states as well (i.e. if I loop over k steps, with my smallest tap for an output -3 and keep al intermediate results, my output will be of length k+3 Examples: (a) if you do not want to store any intermediate results (just the last one) # The memory buffer can be the initial state, just that we need to # add one extra dimension in front of it state = TT.unbroadcast(TT.shape_padleft(x0),0) out,_ = scan(lambda x:x+1, states = state, n_steps = 5) # Once we got our result we need to remove the extra dimension out = out[0] (b) if you want to keep every intermediate results state = TT.alloc(TT.constant(0), 6, x0.shape[0]) state = TT.set_subtensor(state[0], x0) out,_ = scan(lambda x:x+1, states = state, n_steps = 5) out = out[1:] """ def wrap_into_list(x): ''' Wrap the input into a list if it is not already a list ''' if x is None: return [] elif not isinstance(x, (list, tuple)): return [x] else: return list(x) seqs = wrap_into_list(sequences) outs_info = wrap_into_list(states) if allow_gc is None: allow_gc = config.scan.allow_gc # Make sure we get rid of numpy arrays or ints or anything like that # passed as inputs to scan non_seqs = [] for elem in wrap_into_list(params): if not isinstance(elem, gof.Variable): non_seqs.append(tensor.as_tensor_variable(elem)) else: non_seqs.append(elem) # If we provided a known number of steps ( before compilation) # and if that number is 1 or -1, then we can skip the Scan Op, # and just apply the inner function once # To do that we check here to see the nature of n_steps n_fixed_steps = None if isinstance(n_steps, (float, int)): n_fixed_steps = int(n_steps) else: try: n_fixed_steps = opt.get_scalar_constant_value(n_steps) except tensor.basic.NotScalarConstantError: n_fixed_steps = None # Check n_steps is an int if (hasattr(n_steps, 'dtype') and str(n_steps.dtype)[:3] not in ('uin', 'int')): raise ValueError(' n_steps must be an int. dtype provided ' 'is %s' % n_steps.dtype) # compute number of sequences and number of outputs n_seqs = len(seqs) n_outs = len(outs_info) return_steps = OrderedDict() # wrap outputs info in a dictionary if they are not already in one for i in xrange(n_outs): if outs_info[i] is not None: if not isinstance(outs_info[i], dict): # by default any output has a tap value of -1 outs_info[i] = dict(membuf=outs_info[i], taps=[-1]) elif (not outs_info[i].get('membuf', None) and outs_info[i].get('taps', None)): # ^ no initial state but taps provided raise ValueError(('If you are using slices of an output ' 'you need to provide a memory buffer for ' 'the state '), outs_info[i]) elif (outs_info[i].get('membuf', None) and not outs_info[i].get('taps', None)): # ^ initial state but taps not provided if 'taps' in outs_info[i]: # ^ explicitly provided a None for taps _logger.warning( 'Output %s (index %d) has a memory ' 'buffer but taps is explicitly set to None ', getattr(outs_info[i]['membuf'], 'name', 'None'), i) outs_info[i]['taps'] = [-1] else: # if a None is provided as the output info we replace it # with an dict(steps=n_steps) to simplify handling outs_info[i] = dict(steps=n_steps) ## # Step 2. Generate inputs and outputs of the inner functions # for compiling a dummy function (Iteration #1) ## # create theano inputs for the recursive function # note : this is a first batch of possible inputs that will # be compiled in a dummy function; we used this dummy # function to detect shared variables and their updates # and to construct a new and complete list of inputs and # outputs n_seqs = 0 scan_seqs = [] # Variables passed as inputs to the scan op inner_seqs = [] # Variables passed as inputs to the inner function inner_slices = [] # Actual slices if scan is removed from the picture # go through sequences picking up time slices as needed for i, seq in enumerate(seqs): if isinstance(seq, dict): seq = seq['input'] actual_slice = seq[0] _seq_val = tensor.as_tensor_variable(seq) _seq_val_slice = _seq_val[0] nw_slice = _seq_val_slice.type() # Try to transfer test_value to the new variable if config.compute_test_value != 'off': try: nw_slice.tag.test_value = gof.Op._get_test_value( _seq_val_slice) except AttributeError as e: if config.compute_test_value != 'ignore': # No need to print a warning or raise an error now, # it will be done when fn will be called. _logger.info(('Cannot compute test value for ' 'the inner function of scan, input value ' 'missing %s'), e) if seq.name: nw_slice.name = seq.name + '[t]' scan_seqs.append(_seq_val) inner_seqs.append(nw_slice) inner_slices.append(actual_slice) n_seqs += 1 actual_n_steps = tensor.as_tensor(n_steps) # Conventions : # mit_mot = multiple input taps, multiple output taps ( only provided # by the gradient function ) # mit_sot = multiple input taps, single output tap (t + 0) # sit_sot = single input tap, single output tap (t + 0) # nit_sot = no input tap, single output tap (t + 0) # MIT_MOT -- not provided by the user only by the grad function n_mit_mot = 0 n_mit_mot_outs = 0 mit_mot_scan_inputs = [] mit_mot_inner_inputs = [] mit_mot_inner_outputs = [] mit_mot_out_slices = [] mit_mot_rightOrder = [] # SIT_SOT -- provided by the user n_mit_sot = 0 mit_sot_scan_inputs = [] mit_sot_inner_inputs = [] mit_sot_inner_slices = [] mit_sot_inner_outputs = [] mit_sot_return_steps = OrderedDict() mit_sot_tap_array = [] mit_sot_rightOrder = [] n_sit_sot = 0 sit_sot_scan_inputs = [] sit_sot_inner_inputs = [] sit_sot_inner_slices = [] sit_sot_inner_outputs = [] sit_sot_return_steps = OrderedDict() sit_sot_rightOrder = [] nit_sot_steps = [] # go through outputs picking up time slices as needed for i, init_out in enumerate(outs_info): # Note that our convention dictates that if an output uses # just the previous time step, as a initial state we will only # provide a tensor of the same dimension as one time step; This # makes code much cleaner for those who do not use taps. Otherwise # they would always had to shape_padleft the initial state .. # which is ugly # Note, 'taps' might not be in the dictionary if 'taps' in init_out and init_out['taps'] == [-1]: actual_arg = init_out['membuf'] arg = safe_new(init_out['membuf'][0]) if isinstance(arg, tensor.Constant): # safe new returns a clone of the constants, but that is not # what we need for initial states arg = arg.type() # Try to transfer test_value to the new variable if config.compute_test_value != 'off': try: arg.tag.test_value = gof.Op._get_test_value(actual_arg) except AttributeError as e: if config.compute_test_value != 'ignore': # No need to print a warning or raise an error now, # it will be done when fn will be called. _logger.info(('Cannot compute test value for the ' 'inner function of scan, input value missing %s'), e) if getattr(init_out['membuf'], 'name', None) is not None: arg.name = init_out['membuf'].name + '[t-1]' # We need now to allocate space for storing the output and copy # the initial state over. We do this using the expand function # defined in scan utils sit_sot_scan_inputs.append(actual_arg) sit_sot_inner_slices.append(actual_arg[0]) if i in return_steps: sit_sot_return_steps[n_sit_sot] = return_steps[i] sit_sot_inner_inputs.append(arg) sit_sot_rightOrder.append(i) n_sit_sot += 1 elif init_out.get('taps', None): if numpy.any(numpy.array(init_out.get('taps', [])) > 0): # Make sure we do not have requests for future values of a # sequence we can not provide such values raise ValueError('Can not use future taps of outputs', init_out) # go through the taps mintap = abs(numpy.min(init_out['taps'])) mit_sot_tap_array.append(init_out['taps']) idx_offset = abs(numpy.min(init_out['taps'])) # Sequence mit_sot_scan_inputs.append(init_out['membuf']) if i in return_steps: mit_sot_return_steps[n_mit_sot] = return_steps[i] mit_sot_rightOrder.append(i) n_mit_sot += 1 for k in init_out['taps']: # create a new slice actual_nw_slice = init_out['membuf'][k + mintap] _init_out_var = tensor.as_tensor_variable(init_out['membuf']) _init_out_var_slice = _init_out_var[k + mintap] nw_slice = _init_out_var_slice.type() # Try to transfer test_value to the new variable if config.compute_test_value != 'off': try: nw_slice.tag.test_value = gof.Op._get_test_value( _init_out_var_slice) except AttributeError as e: if config.compute_test_value != 'ignore': # No need to print a warning or raise an error now, # it will be done when fn will be called. _logger.info(('Cannot compute test value for ' 'the inner function of scan, input value ' 'missing. %s'), e) # give it a name or debugging and pretty printing if getattr(init_out['membuf'], 'name', None) is not None: if k > 0: nw_slice.name = (init_out['membuf'].name + '[t+%d]' % k) elif k == 0: nw_slice.name = init_out['membuf'].name + '[t]' else: nw_slice.name = (init_out['membuf'].name + '[t%d]' % k) mit_sot_inner_inputs.append(nw_slice) mit_sot_inner_slices.append(actual_nw_slice) else: pass # Re-order args max_mit_sot = numpy.max([-1] + mit_sot_rightOrder) + 1 max_sit_sot = numpy.max([-1] + sit_sot_rightOrder) + 1 n_elems = numpy.max([max_mit_sot, max_sit_sot]) _ordered_args = [[] for x in xrange(n_elems)] offset = 0 for idx in xrange(n_mit_sot): n_inputs = len(mit_sot_tap_array[idx]) if n_fixed_steps == 1: _ordered_args[mit_sot_rightOrder[idx]] = \ mit_sot_inner_slices[offset:offset + n_inputs] else: _ordered_args[mit_sot_rightOrder[idx]] = \ mit_sot_inner_inputs[offset:offset + n_inputs] offset += n_inputs for idx in xrange(n_sit_sot): if n_fixed_steps == 1: _ordered_args[sit_sot_rightOrder[idx]] = \ [sit_sot_inner_slices[idx]] else: _ordered_args[sit_sot_rightOrder[idx]] = \ [sit_sot_inner_inputs[idx]] ordered_args = [] for ls in _ordered_args: ordered_args += ls if n_fixed_steps == 1: args = (inner_slices + ordered_args + non_seqs) else: args = (inner_seqs + ordered_args + non_seqs) # add only the non-shared variables and non-constants to the arguments of # the dummy function [ a function should not get shared variables or # constants as input ] dummy_args = [arg for arg in args if (not isinstance(arg, SharedVariable) and not isinstance(arg, tensor.Constant))] # when we apply the lambda expression we get a mixture of update rules # and outputs that needs to be separated lambda_result = fn(*args) condition, outputs, updates = scan_utils.get_updates_and_outputs( lambda_result) if condition is not None: as_while = True else: as_while = False ## # Step 3. Check if we actually need scan and remove it if we don't ## if n_fixed_steps == 1: # We do not need to use the scan op anymore, so we can just return # the outputs and updates we have if condition is not None: _logger.warning(('When the number of steps is fixed and equal ' 'to 1, the provided stopping condition, ', str(condition), ' is ignored')) for pos, inner_out in enumerate(outputs): # we need to see if we need to pad our sequences with an # unbroadcastable dimension; case example : we return an # output for which we want all intermediate. If n_steps is 1 # then, if we return the output as given by the innner function # this will represent only a slice and it will have one # dimension less. if (isinstance(inner_out.type, tensor.TensorType) and return_steps.get(pos, 0) != 1): outputs[pos] = tensor.unbroadcast( tensor.shape_padleft(inner_out), 0) if len(outputs) == 1: outputs = outputs[0] return (outputs, updates) ## # Step 4. Compile the dummy function ## # We can now compile a dummy function just to see what shared variable # we have and what are their update rules (note that the user has # the option not to pass the shared variable to scan, so we need to # pick them manually and add them to scan) # make the compilation as fast as possible by not applying any # optimization or conversion to C [ note this region is not important # for performance so we can do stuff as unoptimal as we wish ] # extract still missing inputs (there still might be so) and add them # as non sequences at the end of our args fake_nonseqs = [x.type() for x in non_seqs] fake_outputs = scan_utils.clone(outputs + updates.values(), replace=dict(zip(non_seqs, fake_nonseqs))) all_inputs = itertools.ifilter( lambda x: (isinstance(x, gof.Variable) and not isinstance(x, SharedVariable) and not isinstance(x, gof.Constant)), gof.graph.inputs(fake_outputs)) extra_inputs = filter(lambda x: x not in args + fake_nonseqs, all_inputs) non_seqs += extra_inputs # Note we do not use all_inputs directly since the order of variables # in args is quite important dummy_args += extra_inputs dummy_outs = outputs if condition is not None: dummy_outs.append(condition) # If we use a regular dict here, the results are non-deterministic if not isinstance(updates, (list, tuple)): if isinstance(updates, dict) and \ not isinstance(updates, OrderedDict): warnings.warn("Using non-deterministic dictionary.") dummy_f = function(dummy_args, dummy_outs, updates=updates, mode=compile.mode.Mode(linker='py', optimizer=None), on_unused_input='ignore') ## # Step 5. Re-arange inputs of scan into a more strict order ## # Step 5.0 Check the outputs of the dummy function to see if they # match with user provided data # if the number of outputs to the function does not match the number of # assumed outputs until now (provided by the user) there can be # only one explanation: No information is provided for any of the # outputs (i.e. we are dealing with a map) tmp_dummy_f_outs = len(dummy_f.maker.outputs) if as_while: tmp_dummy_f_outs -= 1 if not (tmp_dummy_f_outs == n_outs or outs_info == []): raise ValueError('Please provide None as output_info for ' 'any output that does not feed back into ' 'scan (i.e. it behaves like a map) ') if outs_info == []: n_outs = len(dummy_f.maker.outputs) if as_while: n_outs = n_outs - 1 outs_info = [dict(steps=n_steps) for x in xrange(n_outs)] # Step 5.1 Outputs with taps different then -1 for i, out in enumerate(outs_info): if 'taps' in out and out['taps'] != [-1]: mit_sot_inner_outputs.append(outputs[i]) # Step 5.2 Outputs with tap equal to -1 for i, out in enumerate(outs_info): if 'taps' in out and out['taps'] == [-1]: sit_sot_inner_outputs.append(outputs[i]) # Step 5.3 Outputs that correspond to update rules of shared variables givens = OrderedDict() n_shared_outs = 0 shared_scan_inputs = [] shared_inner_inputs = [] shared_inner_outputs = [] for input in dummy_f.maker.expanded_inputs: if isinstance(input.variable, SharedVariable) and input.update: new_var = safe_new(input.variable) if getattr(input.variable, 'name', None) is not None: new_var.name = input.variable.name + '_copy' shared_inner_inputs.append(new_var) shared_scan_inputs.append(input.variable) shared_inner_outputs.append(input.update) givens[input.variable] = new_var n_shared_outs += 1 # Step 5.4 Outputs with no taps used in the input n_nit_sot = 0 nit_sot_inner_outputs = [] nit_sot_return_steps = OrderedDict() nit_sot_rightOrder = [] for i, out in enumerate(outs_info): if not 'taps' in out: nit_sot_inner_outputs.append(outputs[i]) if i in return_steps: nit_sot_return_steps[n_nit_sot] = return_steps[i] nit_sot_rightOrder.append(i) nit_sot_steps.append(out['steps']) n_nit_sot += 1 # Step 5.5 all other arguments including extra inputs other_scan_args = [] other_inner_args = [] other_scan_args += [arg for arg in non_seqs if (not isinstance(arg, SharedVariable) and not isinstance(arg, tensor.Constant))] # Step 5.6 all shared variables with no update rules other_inner_args += [safe_new(arg, '_copy') for arg in non_seqs if (not isinstance(arg, SharedVariable) and not isinstance(arg, tensor.Constant))] givens.update(dict(zip(other_scan_args, other_inner_args))) other_shared_scan_args = [arg.variable for arg in dummy_f.maker.expanded_inputs if (isinstance(arg.variable, SharedVariable) and not arg.update)] other_shared_inner_args = [safe_new(arg.variable, '_copy') for arg in dummy_f.maker.expanded_inputs if (isinstance(arg.variable, SharedVariable) and not arg.update)] givens.update(dict(zip(other_shared_scan_args, other_shared_inner_args))) ## # Step 6. Re-order the outputs and clone them replacing things # using the givens ## inner_inputs = (inner_seqs + mit_mot_inner_inputs + mit_sot_inner_inputs + sit_sot_inner_inputs + shared_inner_inputs + other_shared_inner_args + other_inner_args) inner_outs = (mit_mot_inner_outputs + mit_sot_inner_outputs + sit_sot_inner_outputs + nit_sot_inner_outputs + shared_inner_outputs) if condition is not None: inner_outs.append(condition) new_givens = OrderedDict() for w, w_copy in givens.iteritems(): new_givens[w] = w.type.filter_variable(w_copy) new_outs = scan_utils.clone(inner_outs, replace=new_givens) ## # Step 7. Create the Scan Op ## tap_array = mit_sot_tap_array + [[-1] for x in xrange(n_sit_sot)] info = OrderedDict() info['tap_array'] = tap_array info['n_seqs'] = n_seqs info['n_mit_mot'] = n_mit_mot info['n_mit_mot_outs'] = n_mit_mot_outs info['mit_mot_out_slices'] = mit_mot_out_slices info['n_mit_sot'] = n_mit_sot info['n_sit_sot'] = n_sit_sot info['n_shared_outs'] = n_shared_outs info['n_nit_sot'] = n_nit_sot info['truncate_gradient'] = -1 info['name'] = name info['mode'] = mode info['destroy_map'] = OrderedDict() info['inplace'] = False info['gpu'] = False info['as_while'] = as_while info['profile'] = profile info['_scan_savemem_visited'] = True info['allow_gc'] = allow_gc local_op = scan_op.Scan(inner_inputs, new_outs, info) ## # Step 8. Compute the outputs using the scan op ## _scan_inputs = (scan_seqs + mit_mot_scan_inputs + mit_sot_scan_inputs + sit_sot_scan_inputs + shared_scan_inputs + nit_sot_steps + other_shared_scan_args + other_scan_args) scan_inputs = [] for arg in [actual_n_steps] + _scan_inputs: if not isinstance(arg, gof.Variable): arg = tensor.as_tensor_variable(arg) scan_inputs += [arg] scan_outs = local_op(*scan_inputs) if type(scan_outs) not in (list, tuple): scan_outs = [scan_outs] ## # Step 9. Figure out which outs are update rules for shared variables # and so on ... ## update_map = OrderedUpdates() offset = n_mit_mot offsets = [abs(numpy.min(x)) for x in mit_sot_tap_array] mit_sot_outs = scan_outs[offset:offset + n_mit_sot] offset += n_mit_sot offsets = [1 for x in xrange(n_sit_sot)] sit_sot_outs = scan_outs[offset:offset + n_sit_sot] offset += n_sit_sot nit_sot_outs = scan_outs[offset:offset + n_nit_sot] offset += n_nit_sot for idx, update_rule in enumerate( scan_outs[offset:offset + n_shared_outs]): update_map[shared_scan_inputs[idx]] = update_rule _scan_out_list = (mit_sot_outs + sit_sot_outs + nit_sot_outs) # Step 10. I need to reorder the outputs to be in the order expected by # the user rightOrder = (mit_sot_rightOrder + sit_sot_rightOrder + nit_sot_rightOrder) scan_out_list = [None] * len(rightOrder) for idx, pos in enumerate(rightOrder): scan_out_list[pos] = _scan_out_list[idx] if len(scan_out_list) == 1: scan_out_list = scan_out_list[0] elif len(scan_out_list) == 0: scan_out_list = None assert isinstance(update_map, OrderedDict) return (scan_out_list, update_map)
def set_objective(self, loss, params, inputs=None, updts=None, grads=None, polyak_averaging=None, clip=None, trust_input=True, compilation_mode=None, **kwargs): ''' Changes the objective function to be optimized @param loss theano graph representing the loss to be optimized @param params theano shared variables representing the parameters to be optimized @param inputs theano variables representing the inputs required to compute the loss, other than params @param updts dictionary of list of theano updates to be applied after every evaluation of the loss function @param grads gradients of the loss function. If not provided, will be computed here @param kwargs arguments to pass to the lasagne.updates function ''' if inputs is None: inputs = [] if updts is not None: updts = OrderedUpdates(updts) if grads is None: utils.print_with_stamp('Building computation graph for gradients', self.name) grads = theano.grad(loss, params) if clip is not None: utils.print_with_stamp( "Clipping gradients to norm %s" % (str(clip)), self.name) grads = lasagne.updates.total_norm_constraint(grads, clip) else: utils.print_with_stamp("No gradient clipping", self.name) utils.print_with_stamp("Computing parameter update rules", self.name) min_method_updt = LASAGNE_MIN_METHODS[self.min_method] grad_updates = min_method_updt(grads, params, **kwargs) outputs = [loss] + grads grad_updates = grad_updates + updts if polyak_averaging and polyak_averaging > 0.0: # create copy of params params_avg = [ theano.shared(p.get_value(borrow=False, return_internal_type=True), broadcastable=p.broadcastable, name=p.name + '_copy') for p in params ] # prepare updates for polyak averaging t = theano.shared(np.array(1, dtype=floatX)) b = polyak_averaging replace_dict = OrderedDict() for p, pp in zip(params, params_avg): grad_updates[pp] = ((b - b**t) * pp + (1 - b) * grad_updates[p]) / (1 - b**t) replace_dict[p] = pp grad_updates[t] = t + 1 outputs[0] = theano.clone(loss, replace=replace_dict, strict=True) self.params_avg = params_avg else: if hasattr(self, 'params_avg'): delattr(self, 'params_avg') utils.print_with_stamp('Compiling function for loss', self.name) # converts inputs to shared variables to avoid repeated gpu transfers self.shared_inpts = [ theano.shared(np.empty([1] * inp.ndim, dtype=inp.dtype), name=inp.name) for inp in inputs ] givens_dict = dict(zip(inputs, self.shared_inpts)) self.loss_fn = theano.function([], loss, updates=updts, on_unused_input='ignore', allow_input_downcast=True, givens=givens_dict, mode=compilation_mode) self.loss_fn.trust_input = trust_input utils.print_with_stamp("Compiling parameter updates", self.name) self.update_params_fn = theano.function([], outputs, updates=grad_updates, on_unused_input='ignore', allow_input_downcast=True, givens=givens_dict, mode=compilation_mode) self.update_params_fn.trust_input = trust_input self.n_evals = 0 self.start_time = 0 self.iter_time = 0 self.params = params self.optimizer_state = [s for s in grad_updates.keys()]
def get_output_for(self, inputs, accumulate_updates="warn", recurrence_flags={}, **kwargs): """ returns history of agent interaction with environment for given number of turns. parameters: inputs - [state init] + [input_nonsequences] + [input_sequences] Each part is a list of theano expressions for layers in the order they were provided when creating this layer. recurrence_flags - a set of flags to be passed to the one step agent (anything that lasagne supports) e.g. {deterministic=True} returns: [state_sequences] + [output sequences] - a list of all states and all outputs sequences Shape of each such sequence is [batch, tick, shape_of_one_state_or_output...] """ #aliases n_states = len(self.state_variables) n_state_inits = len(self.state_init) n_input_nonseq = len(self.input_nonsequences) n_input_seq = len(self.input_sequences) n_outputs = len(self.tracked_outputs) #slice inputs if self.mask_input is not None: mask, inputs = inputs[0], inputs[1:] initial_states_provided, nonsequences, sequences = unpack_list( inputs, [n_state_inits, n_input_nonseq, n_input_seq]) # infer batch size if self.batch_size is not None: batch_size = self.batch_size elif len(inputs) != 0: batch_size = inputs[0].shape[0] else: raise ValueError( "Need to set batch_size explicitly for recurrence") # reshape sequences from [batch, time, ...] to [time,batch,...] to fit scan sequences = [seq.swapaxes(1, 0) for seq in sequences] #here we create outputs_info for scan ## initial states that are given as input initial_states_provided = OrderedDict( list(zip(self.state_init, initial_states_provided))) def get_initial_state(state_out_layer, batch_size=batch_size): """Pick dedicated initial state or create zeros of appropriate shape and dtype""" # if we have a dedicated init, use it if state_out_layer in initial_states_provided: initial_state = initial_states_provided[state_out_layer] # otherwise initialize with zeros else: dtype = get_layer_dtype(state_out_layer) initial_state = T.zeros( (batch_size, ) + tuple(state_out_layer.output_shape[1:]), dtype=dtype) #cast to non-broadcastable tensortype t_state = T.TensorType(dtype, (False, ) * initial_state.ndim) initial_state = t_state.convert_variable(initial_state) assert initial_state is not None #if None, conversion failed. report ASAP return initial_state initial_states = list(map(get_initial_state, self.state_variables)) #dummy values for initial outputs. They have no role in computation, but if nonsequences are present, # AND scan is not unrolled, the step function will not receive prev outputs as parameters, while # if unroll_scan, these parameters are present. we forcibly initialize outputs to prevent # complications during parameter parsing in step function below. initial_output_fillers = list( map(get_initial_state, self.tracked_outputs)) outputs_info = initial_states + initial_output_fillers # recurrent step function def step(*args): sequence_slices, prev_states, prev_outputs, nonsequences = \ unpack_list(args, [n_input_seq, n_states, n_outputs, n_input_nonseq]) # make dicts of prev_states and inputs prev_states_dict = OrderedDict( zip(list(self.state_variables.keys()), prev_states)) input_layers = list( chain(self.input_nonsequences.keys(), self.input_sequences.keys())) assert len(input_layers) == len(nonsequences + sequence_slices) inputs_dict = OrderedDict( zip(input_layers, nonsequences + sequence_slices)) # call one step recurrence new_states, new_outputs = self.get_one_step( prev_states_dict, inputs_dict, **recurrence_flags) #make sure output variable is of exactly the same type as corresponding input get_type = lambda tensor: T.TensorType( tensor.dtype, tensor.broadcastable, sparse_grad=getattr(tensor.type, "sparse_grad", False)) new_states = [ get_type(prev_state).convert_variable( state.astype(prev_state.dtype)) for (prev_state, state) in zip(prev_states, new_states) ] assert None not in new_states, "Some state variables has different dtype/shape from init ." new_outputs = [ get_type(prev_out).convert_variable(out.astype(prev_out.dtype)) for (prev_out, out) in zip(prev_outputs, new_outputs) ] assert None not in new_outputs, "Some of the tracked outputs has shape/dtype changing over time. Please report this." return new_states + new_outputs ###handling mask_input### #a step function that utilizes a mask def step_masked(mask_t, *args): #unpack arrays sequence_slices, prev_states, prev_outputs, nonsequences = \ unpack_list(args, [n_input_seq, n_states, n_outputs, n_input_nonseq]) #get regular step new_states_and_outputs = step(*args) old_states_and_outputs = prev_states + prev_outputs #if mask_t, return new ones, else return old ones def apply_mask(mask_t, new_state, old_state): assert new_state.ndim == old_state.ndim ndim = new_state.ndim #append dims to mask pattern = list(range( mask_t.ndim)) + ['x'] * (ndim - mask_t.ndim) return T.switch(mask_t.dimshuffle(pattern), new_state, old_state) next_states_and_outputs = [ apply_mask(mask_t, new_state, old_state) for new_state, old_state in zip( new_states_and_outputs, old_states_and_outputs) ] return next_states_and_outputs if self.mask_input is not None: sequences = [mask.swapaxes(1, 0)] + sequences step_function = step_masked else: step_function = step #scan itself if self.unroll_scan: # call scan itself history = unroll_scan(step_function, sequences=sequences, outputs_info=outputs_info, non_sequences=nonsequences, n_steps=self.n_steps) #if explicitly asked to reset updates, do so if accumulate_updates == False: self.updates = OrderedUpdates() else: history, updates = theano.scan(step_function, sequences=sequences, outputs_info=outputs_info, non_sequences=nonsequences, n_steps=self.n_steps) if accumulate_updates in (True, 'warn'): self.updates += updates else: #replace updates self.updates = updates #check if user received last updates if not self._updates_received and accumulate_updates == 'warn': warn( "You called get_output from recurrence several times without gathering the updates.\n" "(A) If you wanted to get two outputs from recurrence, use NOT\n" ">>>out1 = get_output(rec[layer1])\n" ">>>out2 = get_output(rec[layer2])\n" "but instead:\n" ">>>out1,out2 = get_output((rec[layer1],rec[layer2])) #or rec[layer1,layer2].\n" "(B) If you want to run recurrence several times and accumulate updates from all runs," "use get_output(...,accumulate_updates=True) to silence the warning.\n" "(C) If you want to get rid of old updates, use get_output(...,accumulate_updates=False)\n" ) if len(self.updates) != 0: self._updates_received = False warn( "Recurrent loop without unroll_scan got nonempty random state updates list. That happened" " because there is some source of randomness (e.g. dropout) inside recurrent step graph." " To compile such graph, one must either call .get_automatic_updates() right after .get_output" " and pass these updates to a function when compiling theano.function.", verbosity_level=2) # reordering from [time,batch,...] to [batch,time,...] history = [(var.swapaxes(1, 0) if var.ndim > 1 else var) for var in check_list(history)] assert len(history) == n_states + n_outputs state_seqs, output_seqs = unpack_list(history, [n_states, n_outputs]) # handle delayed_states # selectively shift state sequences by 1 tick into the past, padding with their initialisations for i in range(len(state_seqs)): if list(self.state_variables.keys())[i] in self.delayed_states: state_seq = state_seqs[i] state_init = initial_states[i] state_seq = T.concatenate( [insert_dim(state_init, 1), state_seq[:, :-1]], axis=1) state_seqs[i] = state_seq #keys corresponding to output sequences. Note that we do not use self.keys() to correctly # handle cases where some variable is present in both state_variables and tracked_outputs output_keys = list(self.state_variables.keys()) + list( self.tracked_outputs) output_values = state_seqs + output_seqs assert len(output_keys) == len(output_values) return OrderedDict(zip(output_keys, output_values))
def scan(fn, sequences=None, outputs_info=None, non_sequences=None, n_steps=None, truncate_gradient=-1, go_backwards=False, mode=None, name=None, profile=False, allow_gc=None, strict=False): """ This function constructs and applies a Scan op to the provided arguments. :param fn: ``fn`` is a function that describes the operations involved in one step of ``scan``. ``fn`` should construct variables describing the output of one iteration step. It should expect as input theano variables representing all the slices of the input sequences and previous values of the outputs, as well as all other arguments given to scan as ``non_sequences``. The order in which scan passes these variables to ``fn`` is the following : * all time slices of the first sequence * all time slices of the second sequence * ... * all time slices of the last sequence * all past slices of the first output * all past slices of the second otuput * ... * all past slices of the last output * all other arguments (the list given as `non_sequences` to scan) The order of the sequences is the same as the one in the list `sequences` given to scan. The order of the outputs is the same as the order of ``outputs_info``. For any sequence or output the order of the time slices is the same as the one in which they have been given as taps. For example if one writes the following : .. code-block:: python scan(fn, sequences = [ dict(input= Sequence1, taps = [-3,2,-1]) , Sequence2 , dict(input = Sequence3, taps = 3) ] , outputs_info = [ dict(initial = Output1, taps = [-3,-5]) , dict(initial = Output2, taps = None) , Output3 ] , non_sequences = [ Argument1, Argument2]) ``fn`` should expect the following arguments in this given order: #. ``Sequence1[t-3]`` #. ``Sequence1[t+2]`` #. ``Sequence1[t-1]`` #. ``Sequence2[t]`` #. ``Sequence3[t+3]`` #. ``Output1[t-3]`` #. ``Output1[t-5]`` #. ``Output3[t-1]`` #. ``Argument1`` #. ``Argument2`` The list of ``non_sequences`` can also contain shared variables used in the function, though ``scan`` is able to figure those out on its own so they can be skipped. For the clarity of the code we recommend though to provide them to scan. To some extend ``scan`` can also figure out other ``non sequences`` (not shared) even if not passed to scan (but used by `fn`). A simple example of this would be : .. code-block:: python import theano.tensor as TT W = TT.matrix() W_2 = W**2 def f(x): return TT.dot(x,W_2) The function is expected to return two things. One is a list of outputs ordered in the same order as ``outputs_info``, with the difference that there should be only one output variable per output initial state (even if no tap value is used). Secondly `fn` should return an update dictionary (that tells how to update any shared variable after each iteration step). The dictionary can optionally be given as a list of tuples. There is no constraint on the order of these two list, ``fn`` can return either ``(outputs_list, update_dictionary)`` or ``(update_dictionary, outputs_list)`` or just one of the two (in case the other is empty). To use ``scan`` as a while loop, the user needs to change the function ``fn`` such that also a stopping condition is returned. To do so, he/she needs to wrap the condition in an ``until`` class. The condition should be returned as a third element, for example: .. code-block:: python ... return [y1_t, y2_t], {x:x+1}, theano.scan_module.until(x < 50) Note that a number of steps (considered in here as the maximum number of steps ) is still required even though a condition is passed (and it is used to allocate memory if needed). = {}): :param sequences: ``sequences`` is the list of Theano variables or dictionaries describing the sequences ``scan`` has to iterate over. If a sequence is given as wrapped in a dictionary, then a set of optional information can be provided about the sequence. The dictionary should have the following keys: * ``input`` (*mandatory*) -- Theano variable representing the sequence. * ``taps`` -- Temporal taps of the sequence required by ``fn``. They are provided as a list of integers, where a value ``k`` impiles that at iteration step ``t`` scan will pass to ``fn`` the slice ``t+k``. Default value is ``[0]`` Any Theano variable in the list ``sequences`` is automatically wrapped into a dictionary where ``taps`` is set to ``[0]`` :param outputs_info: ``outputs_info`` is the list of Theano variables or dictionaries describing the initial state of the outputs computed recurrently. When this initial states are given as dictionary optional information can be provided about the output corresponding to these initial states. The dictionary should have the following keys: * ``initial`` -- Theano variable that represents the initial state of a given output. In case the output is not computed recursively (think of a map) and does not require an initial state this field can be skipped. Given that (only) the previous time step of the output is used by ``fn``, the initial state **should have the same shape** as the output and **should not involve a downcast** of the data type of the output. If multiple time taps are used, the initial state should have one extra dimension that should cover all the possible taps. For example if we use ``-5``, ``-2`` and ``-1`` as past taps, at step 0, ``fn`` will require (by an abuse of notation) ``output[-5]``, ``output[-2]`` and ``output[-1]``. This will be given by the initial state, which in this case should have the shape (5,)+output.shape. If this variable containing the initial state is called ``init_y`` then ``init_y[0]`` *corresponds to* ``output[-5]``. ``init_y[1]`` *correponds to* ``output[-4]``, ``init_y[2]`` corresponds to ``output[-3]``, ``init_y[3]`` coresponds to ``output[-2]``, ``init_y[4]`` corresponds to ``output[-1]``. While this order might seem strange, it comes natural from splitting an array at a given point. Assume that we have a array ``x``, and we choose ``k`` to be time step ``0``. Then our initial state would be ``x[:k]``, while the output will be ``x[k:]``. Looking at this split, elements in ``x[:k]`` are ordered exactly like those in ``init_y``. * ``taps`` -- Temporal taps of the output that will be pass to ``fn``. They are provided as a list of *negative* integers, where a value ``k`` implies that at iteration step ``t`` scan will pass to ``fn`` the slice ``t+k``. ``scan`` will follow this logic if partial information is given: * If an output is not wrapped in a dictionary, ``scan`` will wrap it in one assuming that you use only the last step of the output (i.e. it makes your tap value list equal to [-1]). * If you wrap an output in a dictionary and you do not provide any taps but you provide an initial state it will assume that you are using only a tap value of -1. * If you wrap an output in a dictionary but you do not provide any initial state, it assumes that you are not using any form of taps. * If you provide a ``None`` instead of a variable or a empty dictionary ``scan`` assumes that you will not use any taps for this output (like for example in case of a map) If ``outputs_info`` is an empty list or None, ``scan`` assumes that no tap is used for any of the outputs. If information is provided just for a subset of the outputs an exception is raised (because there is no convention on how scan should map the provided information to the outputs of ``fn``) :param non_sequences: ``non_sequences`` is the list of arguments that are passed to ``fn`` at each steps. One can opt to exclude variable used in ``fn`` from this list as long as they are part of the computational graph, though for clarity we encourage not to do so. :param n_steps: ``n_steps`` is the number of steps to iterate given as an int or Theano scalar. If any of the input sequences do not have enough elements, scan will raise an error. If the *value is 0* the outputs will have *0 rows*. If the value is negative, ``scan`` will run backwards in time. If the ``go_backwards`` flag is already set and also ``n_steps`` is negative, ``scan`` will run forward in time. If n_steps is not provided, ``scan`` will figure out the amount of steps it should run given its input sequences. :param truncate_gradient: ``truncate_gradient`` is the number of steps to use in truncated BPTT. If you compute gradients through a scan op, they are computed using backpropagation through time. By providing a different value then -1, you choose to use truncated BPTT instead of classical BPTT, where you go for only ``truncate_gradient`` number of steps back in time. :param go_backwards: ``go_backwards`` is a flag indicating if ``scan`` should go backwards through the sequences. If you think of each sequence as indexed by time, making this flag True would mean that ``scan`` goes back in time, namely that for any sequence it starts from the end and goes towards 0. :param name: When profiling ``scan``, it is crucial to provide a name for any instance of ``scan``. The profiler will produce an overall profile of your code as well as profiles for the computation of one step of each instance of ``scan``. The ``name`` of the instance appears in those profiles and can greatly help to disambiguate information. :param mode: It is recommended to leave this argument to None, especially when profiling ``scan`` (otherwise the results are not going to be accurate). If you prefer the computations of one step of ``scan`` to be done differently then the entire function, you can use this parameter to describe how the computations in this loop are done (see ``theano.function`` for details about possible values and their meaning). :param profile: Flag or string. If true, or different from the empty string, a profile object will be created and attached to the inner graph of scan. In case ``profile`` is True, the profile object will have the name of the scan instance, otherwise it will have the passed string. Profile object collect (and print) information only when running the inner graph with the new cvm linker ( with default modes, other linkers this argument is useless) :param allow_gc: Set the value of allow gc for the internal graph of scan. If set to None, this will use the value of config.scan.allow_gc. :param strict: If true, all the shared variables used in ``fn`` must be provided as a part of ``non_sequences`` or ``sequences``. :rtype: tuple :return: tuple of the form (outputs, updates); ``outputs`` is either a Theano variable or a list of Theano variables representing the outputs of ``scan`` (in the same order as in ``outputs_info``). ``updates`` is a subclass of dictionary specifying the update rules for all shared variables used in scan This dictionary should be passed to ``theano.function`` when you compile your function. The change compared to a normal dictionary is that we validate that keys are SharedVariable and addition of those dictionary are validated to be consistent. """ # General observation : this code is executed only once, at creation # of the computational graph, so we don't yet need to be smart about # anything (to speed things up) ## # Step 1. Wrap all inputs in dictionaries and add default values ## # check if inputs are just single variables instead of lists def wrap_into_list(x): ''' Wrap the input into a list if it is not already a list ''' if x is None: return [] elif not isinstance(x, (list, tuple)): return [x] else: return list(x) seqs = wrap_into_list(sequences) outs_info = wrap_into_list(outputs_info) # Make sure we get rid of numpy arrays or ints or anything like that # passed as inputs to scan non_seqs = [] for elem in wrap_into_list(non_sequences): if not isinstance(elem, gof.Variable): non_seqs.append(tensor.as_tensor_variable(elem)) else: non_seqs.append(elem) # If we provided a known number of steps ( before compilation) # and if that number is 1 or -1, then we can skip the Scan Op, # and just apply the inner function once # To do that we check here to see the nature of n_steps n_fixed_steps = None if isinstance(n_steps, (float, int)): n_fixed_steps = int(n_steps) else: try: n_fixed_steps = opt.get_scalar_constant_value(n_steps) except tensor.basic.NotScalarConstantError: n_fixed_steps = None # Check n_steps is an int if (hasattr(n_steps, 'dtype') and str(n_steps.dtype)[:3] not in ('uin', 'int')): raise ValueError(' n_steps must be an int. dtype provided ' 'is %s' % n_steps.dtype) # compute number of sequences and number of outputs n_seqs = len(seqs) n_outs = len(outs_info) return_steps = OrderedDict() # wrap sequences in a dictionary if they are not already dictionaries for i in xrange(n_seqs): if not isinstance(seqs[i], dict): seqs[i] = OrderedDict([('input', seqs[i]), ('taps', [0])]) elif seqs[i].get('taps', None) is not None: seqs[i]['taps'] = wrap_into_list(seqs[i]['taps']) elif seqs[i].get('taps', None) is None: # seqs dictionary does not have the ``taps`` key seqs[i]['taps'] = [0] # wrap outputs info in a dictionary if they are not already in one for i in xrange(n_outs): if outs_info[i] is not None: if isinstance(outs_info[i], dict): # DEPRECATED : if outs_info[i].get('return_steps', None) is not None: raise ValueError( "Using `return_steps` has been deprecated. " "Simply select the entries you need using a " "subtensor. Scan will optimize memory " "consumption, so do not worry about that.") # END if not isinstance(outs_info[i], dict): # by default any output has a tap value of -1 outs_info[i] = OrderedDict([('initial', outs_info[i]), ('taps', [-1])]) elif (outs_info[i].get('initial', None) is None and outs_info[i].get('taps', None) is not None): # ^ no initial state but taps provided raise ValueError(('If you are using slices of an output ' 'you need to provide a initial state ' 'for it'), outs_info[i]) elif (outs_info[i].get('initial', None) is not None and outs_info[i].get('taps', None) is None): # ^ initial state but taps not provided if 'taps' in outs_info[i]: # ^ explicitly provided a None for taps _logger.warning('Output %s ( index %d) has a initial ' 'state but taps is explicitly set to None ', getattr(outs_info[i]['initial'], 'name', 'None'), i) outs_info[i]['taps'] = [-1] else: # if a None is provided as the output info we replace it # with an empty OrdereDict() to simplify handling outs_info[i] = OrderedDict() ## # Step 2. Generate inputs and outputs of the inner functions # for compiling a dummy function (Iteration #1) ## # create theano inputs for the recursive function # note : this is a first batch of possible inputs that will # be compiled in a dummy function; we used this dummy # function to detect shared variables and their updates # and to construct a new and complete list of inputs and # outputs n_seqs = 0 scan_seqs = [] # Variables passed as inputs to the scan op inner_seqs = [] # Variables passed as inputs to the inner function inner_slices = [] # Actual slices if scan is removed from the picture # go through sequences picking up time slices as needed for i, seq in enumerate(seqs): # Note that you can have something like no taps for # a sequence, though is highly unlikely in practice if 'taps' in seq: # go through the indicated slice mintap = numpy.min(seq['taps']) maxtap = numpy.max(seq['taps']) for k in seq['taps']: # create one slice of the input # Later on, if we decide not to use scan because we are # going for just one step, it makes things easier if we # compute the correct outputs here. This way we can use # the output of the lambda expression directly to replace # the output of scan. # If not we need to use copies, that will be replaced at # each frame by the corresponding slice actual_slice = seq['input'][k - mintap] _seq_val = tensor.as_tensor_variable(seq['input']) _seq_val_slice = _seq_val[k - mintap] nw_slice = _seq_val_slice.type() # Try to transfer test_value to the new variable if config.compute_test_value != 'off': try: nw_slice.tag.test_value = gof.Op._get_test_value( _seq_val_slice) except AttributeError as e: if config.compute_test_value != 'ignore': # No need to print a warning or raise an error now, # it will be done when fn will be called. _logger.info(('Cannot compute test value for ' 'the inner function of scan, input value ' 'missing %s'), e) # Add names to slices for debugging and pretty printing .. # that is if the input already has a name if getattr(seq['input'], 'name', None) is not None: if k > 0: nw_name = seq['input'].name + '[t+%d]' % k elif k == 0: nw_name = seq['input'].name + '[t]' else: nw_name = seq['input'].name + '[t%d]' % k nw_slice.name = nw_name # We cut the sequence such that seq[i] to correspond to # seq[i-k] if maxtap < 0: offset = abs(maxtap) else: offset = 0 if maxtap == mintap and maxtap != 0: if maxtap < 0: nw_seq = seq['input'][:maxtap] else: nw_seq = seq['input'][maxtap:] elif maxtap - k != 0: nw_seq = seq['input'][offset + k - mintap: -(maxtap - k)] else: nw_seq = seq['input'][offset + k - mintap:] if go_backwards: nw_seq = nw_seq[::-1] scan_seqs.append(nw_seq) inner_seqs.append(nw_slice) inner_slices.append(actual_slice) n_seqs += 1 # Since we've added all sequences now we need to level them up based on # n_steps or their different shapes lengths_vec = [] for seq in scan_seqs: lengths_vec.append(seq.shape[0]) if not scan_utils.isNaN_or_Inf_or_None(n_steps): # ^ N_steps should also be considered lengths_vec.append(tensor.as_tensor(n_steps)) if len(lengths_vec) == 0: # ^ No information about the number of steps raise ValueError(' No information about the number of steps ' 'provided. Either provide a value for ' 'n_steps argument of scan or provide an input ' 'sequence') # If the user has provided the number of steps, do that regardless ( and # raise an error if the sequences are not long enough ) if scan_utils.isNaN_or_Inf_or_None(n_steps): actual_n_steps = lengths_vec[0] for contestant in lengths_vec[1:]: actual_n_steps = tensor.minimum(actual_n_steps, contestant) else: actual_n_steps = tensor.as_tensor(n_steps) # Add names -- it helps a lot when debugging for (nw_seq, seq) in zip(scan_seqs, seqs): if getattr(seq['input'], 'name', None) is not None: nw_seq.name = seq['input'].name + '[%d:]' % k scan_seqs = [seq[:actual_n_steps] for seq in scan_seqs] # Conventions : # mit_mot = multiple input taps, multiple output taps ( only provided # by the gradient function ) # mit_sot = multiple input taps, single output tap (t + 0) # sit_sot = single input tap, single output tap (t + 0) # nit_sot = no input tap, single output tap (t + 0) # MIT_MOT -- not provided by the user only by the grad function n_mit_mot = 0 n_mit_mot_outs = 0 mit_mot_scan_inputs = [] mit_mot_inner_inputs = [] mit_mot_inner_outputs = [] mit_mot_out_slices = [] mit_mot_rightOrder = [] # SIT_SOT -- provided by the user n_mit_sot = 0 mit_sot_scan_inputs = [] mit_sot_inner_inputs = [] mit_sot_inner_slices = [] mit_sot_inner_outputs = [] mit_sot_return_steps = OrderedDict() mit_sot_tap_array = [] mit_sot_rightOrder = [] n_sit_sot = 0 sit_sot_scan_inputs = [] sit_sot_inner_inputs = [] sit_sot_inner_slices = [] sit_sot_inner_outputs = [] sit_sot_return_steps = OrderedDict() sit_sot_rightOrder = [] # go through outputs picking up time slices as needed for i, init_out in enumerate(outs_info): # Note that our convention dictates that if an output uses # just the previous time step, as a initial state we will only # provide a tensor of the same dimension as one time step; This # makes code much cleaner for those who do not use taps. Otherwise # they would always had to shape_padleft the initial state .. # which is ugly if init_out.get('taps', None) == [-1]: actual_arg = init_out['initial'] if not isinstance(actual_arg, tensor.Variable): actual_arg = tensor.as_tensor_variable(actual_arg) arg = safe_new(actual_arg) if isinstance(arg, tensor.Constant): # safe new returns a clone of the constants, but that is not # what we need for initial states arg = arg.type() # Try to transfer test_value to the new variable if config.compute_test_value != 'off': try: arg.tag.test_value = gof.Op._get_test_value(actual_arg) except AttributeError as e: if config.compute_test_value != 'ignore': # No need to print a warning or raise an error now, # it will be done when fn will be called. _logger.info(('Cannot compute test value for the ' 'inner function of scan, input value missing %s'), e) if getattr(init_out['initial'], 'name', None) is not None: arg.name = init_out['initial'].name + '[t-1]' # We need now to allocate space for storing the output and copy # the initial state over. We do this using the expand function # defined in scan utils sit_sot_scan_inputs.append( scan_utils.expand( tensor.unbroadcast( tensor.shape_padleft(actual_arg), 0), actual_n_steps )) sit_sot_inner_slices.append(actual_arg) if i in return_steps: sit_sot_return_steps[n_sit_sot] = return_steps[i] sit_sot_inner_inputs.append(arg) sit_sot_rightOrder.append(i) n_sit_sot += 1 elif init_out.get('taps', None): if numpy.any(numpy.array(init_out.get('taps', [])) > 0): # Make sure we do not have requests for future values of a # sequence we can not provide such values raise ValueError('Can not use future taps of outputs', init_out) # go through the taps mintap = abs(numpy.min(init_out['taps'])) mit_sot_tap_array.append(init_out['taps']) idx_offset = abs(numpy.min(init_out['taps'])) # Sequence mit_sot_scan_inputs.append( scan_utils.expand(init_out['initial'][:mintap], actual_n_steps)) if i in return_steps: mit_sot_return_steps[n_mit_sot] = return_steps[i] mit_sot_rightOrder.append(i) n_mit_sot += 1 for k in init_out['taps']: # create a new slice actual_nw_slice = init_out['initial'][k + mintap] _init_out_var = tensor.as_tensor_variable(init_out['initial']) _init_out_var_slice = _init_out_var[k + mintap] nw_slice = _init_out_var_slice.type() # Try to transfer test_value to the new variable if config.compute_test_value != 'off': try: nw_slice.tag.test_value = gof.Op._get_test_value( _init_out_var_slice) except AttributeError as e: if config.compute_test_value != 'ignore': # No need to print a warning or raise an error now, # it will be done when fn will be called. _logger.info(('Cannot compute test value for ' 'the inner function of scan, input value ' 'missing. %s'), e) # give it a name or debugging and pretty printing if getattr(init_out['initial'], 'name', None) is not None: if k > 0: nw_slice.name = (init_out['initial'].name + '[t+%d]' % k) elif k == 0: nw_slice.name = init_out['initial'].name + '[t]' else: nw_slice.name = (init_out['initial'].name + '[t%d]' % k) mit_sot_inner_inputs.append(nw_slice) mit_sot_inner_slices.append(actual_nw_slice) # NOTE: there is another case, in which we do not want to provide # any previous value of the output to the inner function (i.e. # a map); in that case we do not have to do anything .. # Re-order args max_mit_sot = numpy.max([-1] + mit_sot_rightOrder) + 1 max_sit_sot = numpy.max([-1] + sit_sot_rightOrder) + 1 n_elems = numpy.max([max_mit_sot, max_sit_sot]) _ordered_args = [[] for x in xrange(n_elems)] offset = 0 for idx in xrange(n_mit_sot): n_inputs = len(mit_sot_tap_array[idx]) if n_fixed_steps in [1, -1]: _ordered_args[mit_sot_rightOrder[idx]] = \ mit_sot_inner_slices[offset:offset + n_inputs] else: _ordered_args[mit_sot_rightOrder[idx]] = \ mit_sot_inner_inputs[offset:offset + n_inputs] offset += n_inputs for idx in xrange(n_sit_sot): if n_fixed_steps in [1, -1]: _ordered_args[sit_sot_rightOrder[idx]] = \ [sit_sot_inner_slices[idx]] else: _ordered_args[sit_sot_rightOrder[idx]] = \ [sit_sot_inner_inputs[idx]] ordered_args = [] for ls in _ordered_args: ordered_args += ls if n_fixed_steps in [1, -1]: args = (inner_slices + ordered_args + non_seqs) else: args = (inner_seqs + ordered_args + non_seqs) # add only the non-shared variables and non-constants to the arguments of # the dummy function [ a function should not get shared variables or # constants as input ] dummy_args = [arg for arg in args if (not isinstance(arg, SharedVariable) and not isinstance(arg, tensor.Constant))] # when we apply the lambda expression we get a mixture of update rules # and outputs that needs to be separated condition, outputs, updates = scan_utils.get_updates_and_outputs(fn(*args)) if condition is not None: as_while = True else: as_while = False ## # Step 3. Check if we actually need scan and remove it if we don't ## if n_fixed_steps in [1, -1]: # We do not need to use the scan op anymore, so we can just return # the outputs and updates we have if condition is not None: _logger.warning(('When the number of steps is fixed and equal ' 'to 1, the provided stopping condition, ', str(condition), ' is ignored')) for pos, inner_out in enumerate(outputs): # we need to see if we need to pad our sequences with an # unbroadcastable dimension; case example : we return an # output for which we want all intermediate. If n_steps is 1 # then, if we return the output as given by the innner function # this will represent only a slice and it will have one # dimension less. if (isinstance(inner_out.type, tensor.TensorType) and return_steps.get(pos, 0) != 1): outputs[pos] = tensor.unbroadcast( tensor.shape_padleft(inner_out), 0) if len(outputs) == 1: outputs = outputs[0] return (outputs, updates) ## # Step 4. Compile the dummy function ## # We can now compile a dummy function just to see what shared variable # we have and what are their update rules (note that the user has # the option not to pass the shared variable to scan, so we need to # pick them manually and add them to scan) # make the compilation as fast as possible by not applying any # optimization or conversion to C [ note this region is not important # for performance so we can do stuff as unoptimal as we wish ] # extract still missing inputs (there still might be so) and add them # as non sequences at the end of our args fake_nonseqs = [x.type() for x in non_seqs] fake_outputs = scan_utils.clone(outputs, replace=OrderedDict(zip(non_seqs, fake_nonseqs))) all_inputs = itertools.ifilter( lambda x: (isinstance(x, gof.Variable) and not isinstance(x, SharedVariable) and not isinstance(x, gof.Constant)), gof.graph.inputs(fake_outputs)) extra_inputs = [x for x in all_inputs if x not in args + fake_nonseqs] non_seqs += extra_inputs # Note we do not use all_inputs directly since the order of variables # in args is quite important dummy_args += extra_inputs dummy_outs = outputs if condition is not None: dummy_outs.append(condition) dummy_f = function(dummy_args, dummy_outs, updates=updates, mode=compile.mode.Mode(linker='py', optimizer=None), on_unused_input='ignore', profile=False) ## # Step 5. Re-arange inputs of scan into a more strict order ## # Step 5.0 Check the outputs of the dummy function to see if they # match with user provided data # if the number of outputs to the function does not match the number of # assumed outputs until now (provided by the user) there can be # only one explanation: No information is provided for any of the # outputs (i.e. we are dealing with a map) tmp_dummy_f_outs = len(dummy_f.maker.outputs) if as_while: tmp_dummy_f_outs -= 1 if not (tmp_dummy_f_outs == n_outs or outs_info == []): raise ValueError('Please provide None as outputs_info for ' 'any output that does not feed back into ' 'scan (i.e. it behaves like a map) ') if outs_info == []: n_outs = len(dummy_f.maker.outputs) if as_while: n_outs = n_outs - 1 outs_info = [OrderedDict() for x in xrange(n_outs)] # Step 5.1 Outputs with taps different then -1 for i, out in enumerate(outs_info): if 'taps' in out and out['taps'] != [-1]: mit_sot_inner_outputs.append(outputs[i]) # Step 5.2 Outputs with tap equal to -1 for i, out in enumerate(outs_info): if 'taps' in out and out['taps'] == [-1]: sit_sot_inner_outputs.append(outputs[i]) # Step 5.3 Outputs that correspond to update rules of shared variables givens = OrderedDict() n_shared_outs = 0 shared_scan_inputs = [] shared_inner_inputs = [] shared_inner_outputs = [] sit_sot_shared = [] for input in dummy_f.maker.expanded_inputs: if isinstance(input.variable, SharedVariable) and input.update: new_var = safe_new(input.variable) if getattr(input.variable, 'name', None) is not None: new_var.name = input.variable.name + '_copy' if isinstance(new_var.type, ops.expandable_types): sit_sot_inner_inputs.append(new_var) sit_sot_scan_inputs.append( scan_utils.expand( tensor.unbroadcast( tensor.shape_padleft(input.variable), 0), actual_n_steps)) tensor_update = tensor.as_tensor_variable(input.update) sit_sot_inner_outputs.append(tensor_update) # Not that pos is not a negative index. The sign of pos is used # as a flag to indicate if this output should be part of the # update rules or part of the standard outputs of scan. # If `pos` is positive than it corresponds to the standard # outputs of scan and it refers to output of index `pos`. If `pos` # is negative that it corresponds to update rules of scan and it # refers to update rule of index -1 - `pos`. sit_sot_rightOrder.append(-1 - len(sit_sot_shared)) sit_sot_shared.append(input.variable) givens[input.variable] = new_var else: shared_inner_inputs.append(new_var) shared_scan_inputs.append(input.variable) shared_inner_outputs.append(input.update) givens[input.variable] = new_var n_shared_outs += 1 n_sit_sot = len(sit_sot_inner_inputs) # Step 5.4 Outputs with no taps used in the input n_nit_sot = 0 nit_sot_inner_outputs = [] nit_sot_return_steps = OrderedDict() nit_sot_rightOrder = [] for i, out in enumerate(outs_info): if not 'taps' in out: nit_sot_inner_outputs.append(outputs[i]) if i in return_steps: nit_sot_return_steps[n_nit_sot] = return_steps[i] nit_sot_rightOrder.append(i) n_nit_sot += 1 # Step 5.5 all other arguments including extra inputs other_scan_args = [] other_inner_args = [] other_scan_args += [arg for arg in non_seqs if (not isinstance(arg, SharedVariable) and not isinstance(arg, tensor.Constant))] # Step 5.6 all shared variables with no update rules other_inner_args += [safe_new(arg, '_copy') for arg in non_seqs if (not isinstance(arg, SharedVariable) and not isinstance(arg, tensor.Constant))] givens.update(OrderedDict(zip(other_scan_args, other_inner_args))) if strict: non_seqs_set = set(non_sequences if non_sequences != None else []) other_shared_scan_args = [arg.variable for arg in dummy_f.maker.expanded_inputs if (isinstance(arg.variable, SharedVariable) and not arg.update and arg.variable in non_seqs_set)] other_shared_inner_args = [safe_new(arg.variable, '_copy') for arg in dummy_f.maker.expanded_inputs if (isinstance(arg.variable, SharedVariable) and not arg.update and arg.variable in non_seqs_set)] else: other_shared_scan_args = [arg.variable for arg in dummy_f.maker.expanded_inputs if (isinstance(arg.variable, SharedVariable) and not arg.update)] other_shared_inner_args = [safe_new(arg.variable, '_copy') for arg in dummy_f.maker.expanded_inputs if (isinstance(arg.variable, SharedVariable) and not arg.update)] givens.update(OrderedDict(zip(other_shared_scan_args, other_shared_inner_args))) ## # Step 6. Re-order the outputs and clone them replacing things # using the givens ## inner_inputs = (inner_seqs + mit_mot_inner_inputs + mit_sot_inner_inputs + sit_sot_inner_inputs + shared_inner_inputs + other_shared_inner_args + other_inner_args) inner_outs = (mit_mot_inner_outputs + mit_sot_inner_outputs + sit_sot_inner_outputs + nit_sot_inner_outputs + shared_inner_outputs) if condition is not None: inner_outs.append(condition) # Cuda is imported here, instead of being imported on top of the file # because forces on the user some dependencies that we might do not want # to. Currently we are working on removing the dependencies on sandbox # code completeley. from theano.sandbox import cuda if cuda.cuda_available: # very often we end up in this situation when we want to # replace w with w_copy, where w is CudaNdarray # and w_copy is TensorType. This is caused because shared # variables are put on GPU right aways >:| , new_givens = OrderedDict() for w, w_copy in givens.iteritems(): if (isinstance(w.type, cuda.CudaNdarrayType) and isinstance(w_copy.type, tensor.TensorType)): for o in inner_outs: new_givens = traverse(o, w, w_copy, new_givens) else: new_givens[w] = w_copy else: new_givens = givens new_outs = scan_utils.clone(inner_outs, replace=new_givens) ## # Step 7. Create the Scan Op ## tap_array = mit_sot_tap_array + [[-1] for x in xrange(n_sit_sot)] if allow_gc is None: allow_gc = config.scan.allow_gc info = OrderedDict() info['tap_array'] = tap_array info['n_seqs'] = n_seqs info['n_mit_mot'] = n_mit_mot info['n_mit_mot_outs'] = n_mit_mot_outs info['mit_mot_out_slices'] = mit_mot_out_slices info['n_mit_sot'] = n_mit_sot info['n_sit_sot'] = n_sit_sot info['n_shared_outs'] = n_shared_outs info['n_nit_sot'] = n_nit_sot info['truncate_gradient'] = truncate_gradient info['name'] = name info['mode'] = mode info['destroy_map'] = OrderedDict() info['gpu'] = False info['as_while'] = as_while info['profile'] = profile info['allow_gc'] = allow_gc info['strict'] = strict if strict: warnings.warn('In the strict mode, all neccessary shared variables ' 'must be passed as a part of non_sequences', Warning) local_op = scan_op.Scan(inner_inputs, new_outs, info) ## # Step 8. Compute the outputs using the scan op ## _scan_inputs = (scan_seqs + mit_mot_scan_inputs + mit_sot_scan_inputs + sit_sot_scan_inputs + shared_scan_inputs + [actual_n_steps for x in xrange(n_nit_sot)] + other_shared_scan_args + other_scan_args) scan_inputs = [] for arg in [actual_n_steps] + _scan_inputs: try: arg = tensor.as_tensor_variable(arg) except TypeError: # This happens for Random States for e.g. but it is a good way # to make sure no input is a cuda ndarrays pass scan_inputs += [arg] scan_outs = local_op(*scan_inputs) if type(scan_outs) not in (list, tuple): scan_outs = [scan_outs] ## # Step 9. Figure out which outs are update rules for shared variables # and so on ... ## update_map = OrderedUpdates() def remove_dimensions(outs, steps_return, offsets=None): out_ls = [] for idx, out in enumerate(outs): if idx in steps_return: if steps_return[idx] > 1: out_ls.append(out[-steps_return[idx]:]) else: out_ls.append(out[-1]) else: if offsets is None: out_ls.append(out) else: out_ls.append(out[offsets[idx]:]) return out_ls offset = n_mit_mot offsets = [abs(numpy.min(x)) for x in mit_sot_tap_array] mit_sot_outs = remove_dimensions( scan_outs[offset:offset + n_mit_sot], mit_sot_return_steps, offsets) offset += n_mit_sot offsets = [1 for x in xrange(n_sit_sot)] sit_sot_outs = remove_dimensions( scan_outs[offset:offset + n_sit_sot], sit_sot_return_steps, offsets) offset += n_sit_sot nit_sot_outs = remove_dimensions( scan_outs[offset:offset + n_nit_sot], nit_sot_return_steps) offset += n_nit_sot for idx, update_rule in enumerate( scan_outs[offset:offset + n_shared_outs]): update_map[shared_scan_inputs[idx]] = update_rule _scan_out_list = (mit_sot_outs + sit_sot_outs + nit_sot_outs) # Step 10. I need to reorder the outputs to be in the order expected by # the user rightOrder = (mit_sot_rightOrder + sit_sot_rightOrder + nit_sot_rightOrder) scan_out_list = [None] * len(rightOrder) for idx, pos in enumerate(rightOrder): if pos >= 0: scan_out_list[pos] = _scan_out_list[idx] else: # Not that pos is not a negative index. The sign of pos is used # as a flag to indicate if this output should be part of the # update rules or part of the standard outputs of scan. # If `pos` is positive than it corresponds to the standard # outputs of scan and it refers to output of index `pos`. If `pos` # is negative that it corresponds to update rules of scan and it # refers to update rule of index -1 - `pos`. update_map[sit_sot_shared[abs(pos) - 1]] = _scan_out_list[idx][-1] scan_out_list = [x for x in scan_out_list if x is not None] if len(scan_out_list) == 1: scan_out_list = scan_out_list[0] elif len(scan_out_list) == 0: scan_out_list = None return (scan_out_list, update_map)
try: arg = tensor.as_tensor_variable(arg) except TypeError: # This happens for Random States for e.g. but it is a good way # to make sure no input is a cuda ndarrays pass scan_inputs += [arg] scan_outs = local_op(*scan_inputs) if type(scan_outs) not in (list, tuple): scan_outs = [scan_outs] ## # Step 9. Figure out which outs are update rules for shared variables # and so on ... ## update_map = OrderedUpdates() def remove_dimensions(outs, steps_return, offsets=None): out_ls = [] for idx, out in enumerate(outs): if idx in steps_return: if steps_return[idx] > 1: out_ls.append(out[-steps_return[idx]:]) else: out_ls.append(out[-1]) else: if offsets is None: out_ls.append(out) else: out_ls.append(out[offsets[idx]:]) return out_ls
def test_updates_init(self): self.assertRaises(TypeError, OrderedUpdates, dict(d=3)) sv = theano.shared('asdf') OrderedUpdates({sv:3})
def train(dim_word=100, # word vector dimensionality dim=1000, # the number of LSTM units encoder='gru', decoder='gru_cond', n_words_src=30000, n_words=30000, max_epochs=5000, finish_after=10000000, # finish after this many updates dispFreq=100, decay_c=0., # L2 regularization penalty alpha_c=0., # alignment regularization clip_c=-1., # gradient clipping threshold lrate=1., # learning rate maxlen=100, # maximum length of the description optimizer='rmsprop', batch_size=16, valid_batch_size=80, saveto='model.npz', saveFreq=1000, # save the parameters after every saveFreq updates validFreq=2500, dev_bleu_freq=20000, datasets=('/data/lisatmp3/chokyun/europarl/europarl-v7.fr-en.en.tok', '/data/lisatmp3/chokyun/europarl/europarl-v7.fr-en.fr.tok'), valid_datasets=('./data/dev/dev_en.tok', './data/dev/dev_fr.tok'), small_train_datasets=('./data/train/small_en-fr.en','./data/train/small_en-fr.fr', './data/train/small_en-fr.fr'), use_dropout=False, reload_=False, overwrite=False, preload='', # Options below are from v-yanfa dump_before_train=True, plot_graph=None, vocab_filenames=('./data/dic/filtered_dic_en-fr.en.pkl', './data/dic/filtered_dic_en-fr.fr.pkl'), map_filename='./data/dic/mapFullVocab2Top1MVocab.pkl', lr_discount_freq=80000, # Options of deeper encoder and decoder n_encoder_layers=1, n_decoder_layers=1, encoder_many_bidirectional=True, attention_layer_id=0, unit='gru', residual_enc=None, residual_dec=None, use_zigzag=False, initializer='orthogonal', given_embedding=None, dist_type=None, dist_recover_lr_iter=False, unit_size=2, cond_unit_size=2, given_imm=False, dump_imm=False, shuffle_data=False, decoder_all_attention=False, average_context=False, task='en-fr', fine_tune_patience=8, nccl = False, src_vocab_map_file = None, tgt_vocab_map_file = None, trg_attention_layer_id=None, fix_dp_bug = False, temperature = 1.0, scale=1.0, gate_dropout=0.0, ): model_options = locals().copy() # Set distributed computing environment worker_id = 0 if dist_type == 'mv': try: import multiverso as mv except ImportError: from . import multiverso_ as mv worker_id = mv.worker_id() elif dist_type == 'mpi_reduce': from mpi4py import MPI mpi_communicator = MPI.COMM_WORLD worker_id = mpi_communicator.Get_rank() workers_cnt = mpi_communicator.Get_size() if nccl: nccl_comm = init_nccl_env(mpi_communicator) print 'Use {}, worker id: {}'.format('multiverso' if dist_type == 'mv' else 'mpi' if dist_recover_lr_iter else 'none', worker_id) sys.stdout.flush() # Set logging file set_logging_file('log/complete/e{}d{}_res{}_att{}_worker{}_task{}_{}.txt'.format( n_encoder_layers, n_decoder_layers, residual_enc, attention_layer_id, worker_id, task, time.strftime('%m-%d-%H-%M-%S'), )) log('''\ Start Time = {} '''.format( time.strftime('%c'), )) # Model options: load and save message('Top options:') pprint(model_options) pprint(model_options, stream=get_logging_file()) message('Done') sys.stdout.flush() #load_options(model_options, reload_, preload, src_vocab_map_file and tgt_vocab_map_file) check_options(model_options) model_options['cost_normalization'] = 1 ada_alpha = 0.95 if dist_type == 'mpi_reduce': model_options['cost_normalization'] = workers_cnt message('Model options:') pprint(model_options) pprint(model_options, stream=get_logging_file()) message() print 'Loading data' log('\n\n\nStart to prepare data\n@Current Time = {}'.format(time.time())) sys.stdout.flush() dataset_src, dataset_tgt = datasets[0], datasets[1] if shuffle_data: text_iterator_list = [None for _ in range(10)] text_iterator = None else: text_iterator_list = None text_iterator = TextIterator( dataset_src, dataset_tgt, vocab_filenames[0], vocab_filenames[1], batch_size,n_words_src, n_words,maxlen ) valid_iterator = TextIterator( valid_datasets[0], valid_datasets[1], vocab_filenames[0], vocab_filenames[1], valid_batch_size, n_words_src, n_words ) small_train_iterator = TextIterator( small_train_datasets[0], small_train_datasets[1], vocab_filenames[0], vocab_filenames[1], valid_batch_size, n_words_src, n_words ) print 'Building model' model = NMTModel(model_options) params = model.initializer.init_params() # Reload parameters if reload_ and os.path.exists(preload): print 'Reloading model parameters' load_params(preload, params, src_map_file = src_vocab_map_file, tgt_map_file = tgt_vocab_map_file) sys.stdout.flush() # Given embedding if given_embedding is not None: print 'Loading given embedding...', load_embedding(params, given_embedding) print 'Done' print_params(params) model.init_tparams(params) # Build model, stochastic_mode = 0(soft), 1(stochastic), 2(hard) trng, use_noise, stochastic_mode, hyper_param,\ x, x_mask, y, y_mask, \ opt_ret, \ cost, test_cost, x_emb, stochastic_updates,_ = model.build_model() inps = [x, x_mask, y, y_mask] all_stochastic_updates = OrderedDictUpdates() for item1 in stochastic_updates: for item2 in item1: all_stochastic_updates.update(item2) print 'Building sampler' f_init, f_next = model.build_sampler(trng=trng, use_noise=use_noise, batch_mode=True, stochastic_mode=stochastic_mode, hyper_param=hyper_param) stochastic_mode.set_value(1) # before any regularizer print 'Building f_log_probs...', f_log_probs = theano.function(inps, cost, profile=profile, updates=all_stochastic_updates) print 'Done' sys.stdout.flush() test_cost = test_cost.mean() #FIXME: do not regularize test_cost here cost = cost.mean() cost = l2_regularization(cost, model.P, decay_c) cost = regularize_alpha_weights(cost, alpha_c, model_options, x_mask, y_mask, opt_ret) print 'Building f_cost...', f_cost = theano.function(inps, test_cost, profile=profile, updates=all_stochastic_updates) print 'Done' if plot_graph is not None: print 'Plotting post-compile graph...', theano.printing.pydotprint( f_cost, outfile='pictures/post_compile_{}'.format(plot_graph), var_with_name_simple=True, ) print 'Done' print 'Computing gradient...', grads = tensor.grad(cost, wrt=itemlist(model.P)) clip_shared = theano.shared(np.array(clip_c, dtype=fX), name='clip_shared') if dist_type != 'mpi_reduce': #build grads clip into computational graph grads, g2 = clip_grad_remove_nan(grads, clip_shared, model.P) else: #do the grads clip after gradients aggregation g2 = None # compile the optimizer, the actual computational graph is compiled here lr = tensor.scalar(name='lr') print 'Building optimizers...', given_imm_data = get_adadelta_imm_data(optimizer, given_imm, preload) if optimizer == 'adadelta': f_grad_shared, f_update, grads_shared, imm_shared = Optimizers[optimizer]( lr, model.P, grads, inps, cost, g2=g2, given_imm_data=given_imm_data, alpha = ada_alpha, all_stochastic_updates=all_stochastic_updates) if optimizer == 'adam': f_grad_shared, f_update, grads_shared, imm_shared = Optimizers[optimizer]( lr, model.P, grads, inps, cost, g2=g2, given_imm_data=given_imm_data, all_stochastic_updates=all_stochastic_updates) print 'Done' if dist_type == 'mpi_reduce': f_grads_clip = make_grads_clip_func(grads_shared = grads_shared, mt_tparams= model.P, clip_c_shared = clip_shared) print 'Optimization' log('Preparation Done\n@Current Time = {}'.format(time.time())) if dist_type == 'mv': mv.barrier() elif dist_type == 'mpi_reduce': #create receive buffers for mpi allreduce rec_grads = [np.zeros_like(p.get_value()) for p in model.P.itervalues()] estop = False history_errs = [] best_bleu = -1.0 best_valid_cost = 1e6 best_p = None bad_counter = 0 uidx = search_start_uidx(reload_, preload) epoch_n_batches = 0 start_epoch = 0 pass_batches = 0 print 'worker', worker_id, 'uidx', uidx, 'l_rate', lrate, 'ada_alpha', ada_alpha, 'n_batches', epoch_n_batches, 'start_epoch', start_epoch, 'pass_batches', pass_batches start_uidx = uidx if dump_before_train: print 'Dumping before train...', saveto_uidx = '{}.iter{}.npz'.format( os.path.splitext(saveto)[0], uidx) np.savez(saveto_uidx, history_errs=history_errs, uidx=uidx, **unzip(model.P)) save_options(model_options, uidx, saveto) print 'Done' sys.stdout.flush() stochastic_mode.set_value(0) valid_cost = validation(valid_iterator, f_cost, use_noise) small_train_cost = validation(small_train_iterator, f_cost, use_noise) message('Soft Valid cost {:.5f} Small train cost {:.5f}'.format(valid_cost, small_train_cost)) stochastic_mode.set_value(1) #new_bleu = translate_dev_get_bleu(model, f_init, f_next, trng, use_noise, 5, 1.0) #best_bleu = new_bleu #message('BLEU = {:.2f} at uidx {}'.format(new_bleu, uidx)) sys.stdout.flush() commu_time_sum = 0.0 cp_time_sum =0.0 reduce_time_sum = 0.0 start_time = time.time() finetune_cnt = 0 for eidx in xrange(start_epoch, max_epochs): if shuffle_data: text_iterator = load_shuffle_text_iterator( eidx, worker_id, text_iterator_list, datasets, vocab_filenames, batch_size, maxlen, n_words_src, n_words ) n_samples = 0 if dist_type == 'mpi_reduce': mpi_communicator.Barrier() for i, (x, y) in enumerate(text_iterator): if eidx == start_epoch and i < pass_batches: #ignore the first several batches when reload continue n_samples += len(x) uidx += 1 use_noise.set_value(1.) x, x_mask, y, y_mask = prepare_data(x, y, maxlen=maxlen) if x is None: print 'Minibatch with zero sample under length ', maxlen uidx -= 1 continue effective_uidx = uidx - start_uidx ud_start = time.time() # compute cost, grads if dist_type != 'mpi_reduce': cost, g2_value = f_grad_shared(x, x_mask, y, y_mask) else: cost = f_grad_shared(x, x_mask, y, y_mask) if dist_type == 'mpi_reduce': reduce_start = time.time() commu_time = 0 gpucpu_cp_time = 0 if not nccl: commu_time, gpucpu_cp_time = all_reduce_params(grads_shared, rec_grads) else: commu_time, gpucpu_cp_time = all_reduce_params_nccl(nccl_comm, grads_shared) reduce_time = time.time() - reduce_start commu_time_sum += commu_time reduce_time_sum += reduce_time cp_time_sum += gpucpu_cp_time g2_value = f_grads_clip() print '@Worker = {}, Reduce time = {:.5f}, Commu time = {:.5f}, Copy time = {:.5f}'.format(worker_id, reduce_time, commu_time, gpucpu_cp_time) curr_lr = lrate if not dist_type or dist_recover_lr_iter < effective_uidx else lrate * 0.05 + effective_uidx * lrate / dist_recover_lr_iter * 0.95 if curr_lr < lrate: print 'Curr lr {:.3f}'.format(curr_lr) # do the update on parameters f_update(curr_lr) ud = time.time() - ud_start if np.isnan(g2_value) or np.isinf(g2_value): message('gradient NaN detected') sys.stdout.flush() if np.isnan(cost) or np.isinf(cost): message('cost NaN detected') model.save_model(saveto, history_errs, uidx) save_minibatch(x, y, saveto, uidx, vocab_filenames) sys.stdout.flush() return 1., 1., 1. # discount learning rate # FIXME: Do NOT enable this and fine-tune at the same time if lr_discount_freq > 0 and np.mod(effective_uidx, lr_discount_freq) == 0: lrate *= 0.5 message('Discount learning rate to {} at iteration {}'.format(lrate, uidx)) # sync batch if dist_type == 'mv' and np.mod(uidx, dispFreq) == 0: comm_start = time.time() model.sync_tparams() message('@Comm time = {:.5f}'.format(time.time() - comm_start)) # verbose if np.mod(effective_uidx, dispFreq) == 0: message('Worker {} Epoch {} Update {} Cost {:.5f} G2 {:.5f} UD {:.5f} Time {:.5f} s'.format( worker_id, eidx, uidx, float(cost), float(g2_value), ud, time.time() - start_time, )) sys.stdout.flush() if np.mod(effective_uidx, saveFreq) == 0 and worker_id == 0: # save with uidx if not overwrite: print 'Saving the model at iteration {}...'.format(uidx), model.save_model(saveto, history_errs, uidx) print 'Done' sys.stdout.flush() # save immediate data in adadelta saveto_imm_path = '{}_latest.npz'.format(os.path.splitext(saveto)[0]) dump_adadelta_imm_data(optimizer, imm_shared, dump_imm, saveto_imm_path) if np.mod(effective_uidx, validFreq) == 0: stochastic_mode.set_value(0) valid_cost = validation(valid_iterator, f_cost, use_noise) small_train_cost = validation(small_train_iterator, f_cost, use_noise) message('Soft Valid cost {:.5f} Small train cost {:.5f}'.format(valid_cost, small_train_cost)) #new_bleu = translate_dev_get_bleu(model, f_init, f_next, trng, use_noise, 5, 1.0) #message('BLEU = {:.2f} at uidx {}'.format(new_bleu, uidx)) sys.stdout.flush() #if new_bleu > best_bleu: # print 'Saving the model at iteration {}...'.format(uidx), # model.save_model(saveto, history_errs, uidx) # print 'Done' # best_bleu = new_bleu # sys.stdout.flush() stochastic_mode.set_value(1) # Fine-tune based on dev cost if fine_tune_patience > 0: if valid_cost < best_valid_cost: bad_counter = 0 best_valid_cost = valid_cost #dump the best model so far, including the immediate file if worker_id == 0: message('Dump the the best model so far at uidx {}'.format(uidx)) model.save_model(saveto, history_errs) #dump_adadelta_imm_data(optimizer, imm_shared, dump_imm, saveto) else: bad_counter += 1 if bad_counter >= fine_tune_patience: print 'Fine tune:', if finetune_cnt % 2 == 0: lrate = np.float32(lrate * 0.5) message('Discount learning rate to {} at iteration {}'.format(lrate, uidx)) if lrate <= 0.025: message('Learning rate decayed to {:.5f}, task completed'.format(lrate)) return 1., 1., 1. else: clip_shared.set_value(np.float32(clip_shared.get_value() * 0.25)) message('Discount clip value to {} at iteration {}'.format(clip_shared.get_value(), uidx)) finetune_cnt += 1 bad_counter = 0 # finish after this many updates if uidx >= finish_after: print 'Finishing after {} iterations!'.format(uidx) estop = True break print 'Seen {} samples'.format(n_samples) if estop: break if best_p is not None: zipp(best_p, model.P) use_noise.set_value(0.) return 0.
def test_updates_init(self): with pytest.raises(TypeError): OrderedUpdates(dict(d=3)) sv = theano.shared("asdf") OrderedUpdates({sv: 3})
def get_output_for(self, inputs, accumulate_updates="warn",recurrence_flags={}, **kwargs): """ returns history of agent interaction with environment for given number of turns. parameters: inputs - [state init] + [input_nonsequences] + [input_sequences] Each part is a list of theano expressions for layers in the order they were provided when creating this layer. recurrence_flags - a set of flags to be passed to the one step agent (anything that lasagne supports) e.g. {deterministic=True} returns: [state_sequences] + [output sequences] - a list of all states and all outputs sequences Shape of each such sequence is [batch, tick, shape_of_one_state_or_output...] """ #aliases n_states = len(self.state_variables) n_state_inits = len(self.state_init) n_input_nonseq = len(self.input_nonsequences) n_input_seq = len(self.input_sequences) n_outputs = len(self.tracked_outputs) #slice inputs if self.mask_input is not None: mask,inputs = inputs[0],inputs[1:] initial_states_provided, nonsequences, sequences = unpack_list(inputs, [n_state_inits, n_input_nonseq, n_input_seq]) # infer batch size if self.batch_size is not None: batch_size = self.batch_size elif len(inputs) != 0: batch_size = inputs[0].shape[0] else: raise ValueError("Need to set batch_size explicitly for recurrence") #here we create outputs_info for scan, basically initial values for states and outputs ## initial states that are given as input initial_states_provided = OrderedDict(list(zip(self.state_init, initial_states_provided))) def get_initial_state(layer, batch_size=batch_size): """Pick dedicated initial state or create zeros of appropriate shape and dtype :param layer: layer for new hidden state (key of self.state_variables) :param batch_size: symbolic batch_size """ # if we have a dedicated init, use it if layer in initial_states_provided: initial_state = initial_states_provided[layer] # otherwise initialize with zeros else: assert None not in layer.output_shape[1:],\ "Some of your state layers ({}) has undefined shape along non-batch dimension. (shape: {}) " \ "Therefore, it's initial value can't be inferred. Please set explicit initial value via state_init" \ "".format(layer.name or layer, layer.output_shape) dtype = get_layer_dtype(layer) initial_state = T.zeros((batch_size,) + tuple(layer.output_shape[1:]), dtype=dtype) #disable broadcasting along all axes (lasagne outputs are non-broadcastable) initial_state = T.unbroadcast(initial_state, *range(initial_state.ndim)) return initial_state initial_states = list(map(get_initial_state, self.state_variables)) # dummy initial values for tracked_outputs. # We need to provide them for step_masked to be able to backtrack to them. Also unroll scan requires them. # Initial shapes for outputs are inferred by calling get_one_step and taking shapes from it. # Theano optimizes shape computation without computing get_out_step outputs themselves # the resulting graph would be like (var1.shape[0],var1.shape[2]*3,10) so this operation is zero-cost. state_feed_dict = dict(zip(self.state_variables.keys(),initial_states)) input_feed_dict = dict(zip(list(chain(self.input_nonsequences.keys(), self.input_sequences.keys())), list(chain(nonsequences,[seq[:,0] for seq in sequences])))) initial_output_fillers = self.get_one_step(state_feed_dict,input_feed_dict,**recurrence_flags)[1] # disable broadcasting of zeros_like(v) along all axes (since lasagne outputs are non-broadcastable) initial_output_fillers = [T.unbroadcast(T.zeros_like(v),*range(v.ndim)) for v in initial_output_fillers] #/end of that nonsense #stack all initializers together outputs_info = initial_states + initial_output_fillers # reshape sequences from [batch, time, ...] to [time,batch,...] to fit scan sequences = [seq.swapaxes(1, 0) for seq in sequences] # recurrent step function def step(*args): sequence_slices, prev_states, prev_outputs, nonsequences = \ unpack_list(args, [n_input_seq, n_states, n_outputs, n_input_nonseq]) # make dicts of prev_states and inputs prev_states_dict = OrderedDict(zip(list(self.state_variables.keys()), prev_states)) input_layers = list(chain(self.input_nonsequences.keys(), self.input_sequences.keys())) assert len(input_layers) == len(nonsequences + sequence_slices) inputs_dict = OrderedDict(zip(input_layers, nonsequences + sequence_slices)) # call one step recurrence new_states, new_outputs = self.get_one_step(prev_states_dict, inputs_dict, **recurrence_flags) #make sure new state variables are of exactly the same type as their initial value state_names = [layer.name or str(layer) for layer in list(self.state_variables.keys())] for i in range(len(state_names)): try: if self.force_cast_types: new_states[i] = new_states[i].astype(prev_states[i].dtype) new_states[i] = cast_to_type(new_states[i],get_type(prev_states[i])) except: raise ValueError("Could not convert new state {}, of type {}, to it's previous/initial state type " "{}. Cast type manually or set force_cast_types=True on creation." "".format(state_names[i],get_type(new_states[i]),get_type(prev_states[i]))) #make sure output variables are of exactly the same type as their initial value output_names = [layer.name or str(layer) for layer in self.tracked_outputs] for i in range(len(output_names)): try: if self.force_cast_types: new_outputs[i] = new_outputs[i].astype(prev_outputs[i].dtype) new_outputs[i] = cast_to_type(new_outputs[i],get_type(prev_outputs[i])) except: raise ValueError("Could not convert output of {}, of type {}, to it's previous/initial state type " "{}. Cast type manually or set force_cast_types=True on creation." "".format(output_names[i],get_type(new_outputs[i]),get_type(prev_outputs[i]))) return new_states + new_outputs ###handling mask_input### #a step function that utilizes a mask def step_masked(mask_t,*args): #unpack arrays sequence_slices, prev_states, prev_outputs, nonsequences = \ unpack_list(args, [n_input_seq, n_states, n_outputs, n_input_nonseq]) #get regular step new_states_and_outputs = step(*args) old_states_and_outputs = prev_states+prev_outputs #if mask_t, return new ones, else return old ones def apply_mask(mask_t,new_state,old_state): assert new_state.ndim == old_state.ndim ndim = new_state.ndim #append dims to mask pattern = list(range(mask_t.ndim)) + ['x'] * (ndim - mask_t.ndim) return T.switch(mask_t.dimshuffle(pattern), new_state, old_state) next_states_and_outputs = [apply_mask(mask_t,new_state,old_state) for new_state,old_state in zip(new_states_and_outputs, old_states_and_outputs)] return next_states_and_outputs if self.mask_input is not None: sequences = [mask.swapaxes(1, 0)]+sequences step_function = step_masked else: step_function = step #scan itself if self.unroll_scan: # call scan itself history = unroll_scan(step_function, sequences=sequences, outputs_info=outputs_info, non_sequences=nonsequences, n_steps=self.n_steps ) #if explicitly asked to reset updates, do so if accumulate_updates == False: self.updates=OrderedUpdates() else: history,updates = theano.scan(step_function, sequences=sequences, outputs_info=outputs_info, non_sequences=nonsequences, n_steps=self.n_steps ) if accumulate_updates in (True,'warn'): self.updates += updates else:#replace updates self.updates = updates #check if user received last updates if not self._updates_received and accumulate_updates=='warn': warn("You called get_output from recurrence several times without gathering the updates.\n" "(A) If you wanted to get two outputs from recurrence, use NOT\n" ">>>out1 = get_output(rec[layer1])\n" ">>>out2 = get_output(rec[layer2])\n" "but instead:\n" ">>>out1,out2 = get_output((rec[layer1],rec[layer2])) #or rec[layer1,layer2].\n" "(B) If you want to run recurrence several times and accumulate updates from all runs," "use get_output(...,accumulate_updates=True) to silence the warning.\n" "(C) If you want to get rid of old updates, use get_output(...,accumulate_updates=False)\n" ) if len(self.updates) !=0: self._updates_received=False warn("Recurrent loop without unroll_scan got nonempty random state updates list. That happened" " because there is some source of randomness (e.g. dropout) inside recurrent step graph." " To compile such graph, one must either call .get_automatic_updates() right after .get_output" " and pass these updates to a function when compiling theano.function.",verbosity_level=2) # reordering from [time,batch,...] to [batch,time,...] history = [(var.swapaxes(1, 0) if var.ndim > 1 else var) for var in check_list(history)] assert len(history) == n_states+n_outputs state_seqs, output_seqs = unpack_list(history, [n_states, n_outputs]) # handle delayed_states # selectively shift state sequences by 1 tick into the past, padding with their initialisations for i in range(len(state_seqs)): if list(self.state_variables.keys())[i] in self.delayed_states: state_seq = state_seqs[i] state_init = initial_states[i] state_seq = T.concatenate([insert_dim(state_init, 1), state_seq[:, :-1]], axis=1) state_seqs[i] = state_seq #keys corresponding to output sequences. Note that we do not use self.keys() to correctly # handle cases where some variable is present in both state_variables and tracked_outputs output_keys = list(self.state_variables.keys()) + list(self.tracked_outputs) output_values = state_seqs + output_seqs assert len(output_keys) == len(output_values) return OrderedDict(zip(output_keys,output_values))