FM / python

/usr/local/lib64/python3.6/site-packages/caffe2/python/

/usr/local/lib64/python3.6/site-packages/caffe2/python

Name	Size	Mode	Actions
docs/	-	0755	rm
examples/	-	0755	rm
fakelowp/	-	0755	rm
helpers/	-	0755	rm
ideep/	-	0755	rm
layers/	-	0755	rm
mint/	-	0755	rm
mkl/	-	0755	rm
modeling/	-	0755	rm
models/	-	0755	rm
onnx/	-	0755	rm
operator_test/	-	0755	rm
predictor/	-	0755	rm
rnn/	-	0755	rm
serialized_test/	-	0755	rm
test/	-	0755	rm
trt/	-	0755	rm
__pycache__/	-	0755	rm
allcompare_test.py	2255	0644	edit dl rm
attention.py	12359	0644	edit dl rm
benchmark_generator.py	4912	0644	edit dl rm
binarysize.py	5521	0644	edit dl rm
brew.py	4762	0644	edit dl rm
brew_test.py	11739	0644	edit dl rm
build.py	153	0644	edit dl rm
cached_reader.py	4394	0644	edit dl rm
caffe2_pybind11_state.cpython-36m-x86_64-linux-gnu.so	48299712	0755	edit dl rm
caffe2_pybind11_state_gpu.cpython-36m-x86_64-linux-gnu.so	49048144	0755	edit dl rm
caffe_translator.py	35227	0644	edit dl rm
caffe_translator_test.py	3553	0644	edit dl rm
checkpoint.py	32101	0644	edit dl rm
checkpoint_test.py	13405	0644	edit dl rm
cnn.py	7626	0644	edit dl rm
context.py	2841	0644	edit dl rm
context_test.py	1792	0644	edit dl rm
control.py	19309	0644	edit dl rm
control_ops_grad.py	28893	0644	edit dl rm
control_ops_grad_test.py	1752	0644	edit dl rm
control_ops_util.py	10863	0644	edit dl rm
control_test.py	12276	0644	edit dl rm
convert.py	55	0644	edit dl rm
convert_test.py	201	0644	edit dl rm
convnet_benchmarks.py	20533	0644	edit dl rm
convnet_benchmarks_test.py	839	0644	edit dl rm
core.py	119400	0644	edit dl rm
core_gradients_test.py	38022	0644	edit dl rm
core_test.py	47683	0644	edit dl rm
crf.py	13250	0644	edit dl rm
crf_predict.py	1159	0644	edit dl rm
crf_viterbi_test.py	1663	0644	edit dl rm
dataio.py	23532	0644	edit dl rm
dataio_test.py	17575	0644	edit dl rm
dataset.py	12886	0644	edit dl rm
data_parallel_model.py	83100	0644	edit dl rm
data_parallel_model_test.py	56145	0644	edit dl rm
data_workers.py	15941	0644	edit dl rm
data_workers_test.py	6561	0644	edit dl rm
db_file_reader.py	6608	0644	edit dl rm
db_test.py	1110	0644	edit dl rm
device_checker.py	5157	0644	edit dl rm
dyndep.py	1533	0644	edit dl rm
embedding_generation_benchmark.py	5256	0644	edit dl rm
experiment_util.py	3625	0644	edit dl rm
extension_loader.py	744	0644	edit dl rm
fakefp16_transform_lib.py	322	0644	edit dl rm
filler_test.py	748	0644	edit dl rm
functional.py	4415	0644	edit dl rm
functional_test.py	4204	0644	edit dl rm
fused_8bit_rowwise_conversion_ops_test.py	3945	0644	edit dl rm
gradient_checker.py	15377	0644	edit dl rm
gradient_check_test.py	20729	0644	edit dl rm
gru_cell.py	5129	0644	edit dl rm
hip_test_util.py	405	0644	edit dl rm
hsm_util.py	2259	0644	edit dl rm
hypothesis_test.py	105762	0644	edit dl rm
hypothesis_test_util.py	26853	0644	edit dl rm
ideep_test_util.py	998	0644	edit dl rm
layers_test.py	92931	0644	edit dl rm
layer_model_helper.py	29340	0644	edit dl rm
layer_model_instantiator.py	3935	0644	edit dl rm
layer_parameter_sharing_test.py	9148	0644	edit dl rm
layer_test_util.py	4875	0644	edit dl rm
lazy.py	277	0644	edit dl rm
lazy_dyndep.py	2562	0644	edit dl rm
lazy_dyndep_test.py	3914	0644	edit dl rm
lengths_reducer_fused_8bit_rowwise_ops_test.py	7575	0644	edit dl rm
lengths_reducer_rowwise_8bit_ops_test.py	5710	0644	edit dl rm
lstm_benchmark.py	10649	0644	edit dl rm
memonger.py	34041	0644	edit dl rm
memonger_test.py	36910	0644	edit dl rm
mkl_test_util.py	1142	0644	edit dl rm
model_device_test.py	4777	0644	edit dl rm
model_helper.py	23492	0644	edit dl rm
model_helper_test.py	2336	0644	edit dl rm
modifier_context.py	1772	0644	edit dl rm
muji.py	8131	0644	edit dl rm
muji_test.py	3058	0644	edit dl rm
net_builder.py	27679	0644	edit dl rm
net_builder_test.py	11382	0644	edit dl rm
net_drawer.py	14264	0644	edit dl rm
net_printer.py	12704	0644	edit dl rm
net_printer_test.py	3190	0644	edit dl rm
nomnigraph.py	4216	0644	edit dl rm
nomnigraph_test.py	15427	0644	edit dl rm
nomnigraph_transformations.py	3787	0644	edit dl rm
nomnigraph_transformations_test.py	5767	0644	edit dl rm
normalizer.py	1411	0644	edit dl rm
normalizer_context.py	1007	0644	edit dl rm
normalizer_test.py	487	0644	edit dl rm
numa_benchmark.py	2230	0644	edit dl rm
numa_test.py	1663	0644	edit dl rm
observer_test.py	5316	0644	edit dl rm
operator_fp_exceptions_test.py	1248	0644	edit dl rm
optimizer.py	78813	0644	edit dl rm
optimizer_context.py	1462	0644	edit dl rm
optimizer_test.py	30705	0644	edit dl rm
optimizer_test_util.py	9187	0644	edit dl rm
parallelize_bmuf_distributed_test.py	9908	0644	edit dl rm
parallel_workers.py	7682	0644	edit dl rm
parallel_workers_test.py	3501	0644	edit dl rm
pipeline.py	17283	0644	edit dl rm
pipeline_test.py	2542	0644	edit dl rm
predictor_constants.py	198	0644	edit dl rm
python_op_test.py	9169	0644	edit dl rm
queue_util.py	4459	0644	edit dl rm
record_queue.py	4453	0644	edit dl rm
recurrent.py	13297	0644	edit dl rm
regularizer.py	21120	0644	edit dl rm
regularizer_context.py	1013	0644	edit dl rm
regularizer_test.py	10266	0644	edit dl rm
rnn_cell.py	68233	0644	edit dl rm
schema.py	45621	0644	edit dl rm
schema_test.py	15754	0644	edit dl rm
scope.py	3623	0644	edit dl rm
scope_test.py	5249	0644	edit dl rm
session.py	7642	0644	edit dl rm
session_test.py	2078	0644	edit dl rm
sparse_to_dense_mask_test.py	6565	0644	edit dl rm
sparse_to_dense_test.py	3556	0644	edit dl rm
task.py	24274	0644	edit dl rm
task_test.py	870	0644	edit dl rm
test_util.py	3524	0644	edit dl rm
text_file_reader.py	1990	0644	edit dl rm
timeout_guard.py	4054	0644	edit dl rm
toy_regression_test.py	2822	0644	edit dl rm
transformations.py	1832	0644	edit dl rm
transformations_test.py	11960	0644	edit dl rm
tt_core.py	9349	0644	edit dl rm
tt_core_test.py	2518	0644	edit dl rm
utils.py	14181	0644	edit dl rm
utils_test.py	1399	0644	edit dl rm
visualize.py	6315	0644	edit dl rm
workspace.py	25263	0644	edit dl rm
workspace_test.py	34844	0644	edit dl rm
_import_c_extension.py	2250	0644	edit dl rm
__init__.py	3925	0644	edit dl rm

Edit: /usr/local/lib64/python3.6/site-packages/caffe2/python/recurrent.py (13297B)

## @package recurrent
# Module caffe2.python.recurrent

from caffe2.python import core, workspace
from future.utils import viewitems, viewkeys

def recurrent_net(
        net, cell_net, inputs, initial_cell_inputs,
        links, timestep=None, scope=None, outputs_with_grads=(0,),
        recompute_blobs_on_backward=None, forward_only=False,
):
    '''
    net: the main net operator should be added to

cell_net: cell_net which is executed in a recurrent fasion

inputs: sequences to be fed into the recurrent net. Currently only one input
    is supported. It has to be in a format T x N x (D1...Dk) where T is lengths
    of the sequence. N is a batch size and (D1...Dk) are the rest of dimentions

initial_cell_inputs: inputs of the cell_net for the 0 timestamp.
    Format for each input is:
        (cell_net_input_name, external_blob_with_data)

links: a dictionary from cell_net input names in moment t+1 and
    output names of moment t. Currently we assume that each output becomes
    an input for the next timestep.

timestep: name of the timestep blob to be used. If not provided "timestep"
    is used.

scope: Internal blobs are going to be scoped in a format
    <scope_name>/<blob_name>
    If not provided we generate a scope name automatically

outputs_with_grads : position indices of output blobs which will receive
    error gradient (from outside recurrent network) during backpropagation

recompute_blobs_on_backward: specify a list of blobs that will be
                 recomputed for backward pass, and thus need not to be
                 stored for each forward timestep.

forward_only: if True, only forward steps are executed
    '''
    assert len(inputs) == 1, "Only one input blob is supported so far"

input_blobs = [str(i[0]) for i in inputs]
    initial_input_blobs = [str(x[1]) for x in initial_cell_inputs]
    op_name = net.NextName('recurrent')

def s(name):
        # We have to manually scope due to our internal/external blob
        # relationships.
        scope_name = op_name if scope is None else scope
        return "{}/{}".format(str(scope_name), str(name))

# determine inputs that are considered to be references
    # it is those that are not referred to in inputs or initial_cell_inputs
    known_inputs = [str(b) for b in input_blobs + initial_input_blobs]
    known_inputs += [str(x[0]) for x in initial_cell_inputs]
    if timestep is not None:
        known_inputs.append(str(timestep))
    references = [
        core.BlobReference(b) for b in cell_net.Proto().external_input
        if b not in known_inputs]

inner_outputs = list(cell_net.Proto().external_output)
    # These gradients are expected to be available during the backward pass
    inner_outputs_map = {o: o + '_grad' for o in inner_outputs}

# compute the backward pass of the cell net
    if not forward_only:
        backward_ops, backward_mapping = core.GradientRegistry.GetBackwardPass(
            cell_net.Proto().op, inner_outputs_map)
        backward_mapping = {str(k): v for k, v in viewitems(backward_mapping)}

backward_cell_net = core.Net("RecurrentBackwardStep")
        del backward_cell_net.Proto().op[:]

if recompute_blobs_on_backward is not None:
            # Insert operators to re-compute the specified blobs.
            # They are added in the same order as for the forward pass, thus
            # the order is correct.
            recompute_blobs_on_backward = {str(b) for b in
                                           recompute_blobs_on_backward}

for op in cell_net.Proto().op:
                if not recompute_blobs_on_backward.isdisjoint(set(op.output)):
                    backward_cell_net.Proto().op.extend([op])
                    # This fires if other outputs than the declared
                    # are computed by the ops that are recomputed
                    assert set(op.output).issubset(recompute_blobs_on_backward)

backward_cell_net.Proto().op.extend(backward_ops)
        # compute blobs used but not defined in the backward pass
        backward_ssa, backward_blob_versions = core.get_ssa(
            backward_cell_net.Proto())
        undefined = core.get_undefined_blobs(backward_ssa)

# also add to the output list the intermediate outputs of fwd_step that
        # are used by backward.
        ssa, blob_versions = core.get_ssa(cell_net.Proto())
        scratches = [
            blob
            for blob, ver in viewitems(blob_versions)
            if (ver > 0 and
                blob in undefined and
                blob not in cell_net.Proto().external_output)
        ]
        backward_cell_net.Proto().external_input.extend(scratches)
        backward_cell_net.Proto().type = 'simple'
    else:
        backward_cell_net = None

all_inputs = [i[1] for i in inputs] + [
        x[1] for x in initial_cell_inputs] + references
    all_outputs = []

cell_net.Proto().type = 'simple'

# Internal arguments used by RecurrentNetwork operator

# Links are in the format blob_name, recurrent_states, offset.
    # In the moment t we know that corresponding data block is at
    # t + offset position in the recurrent_states tensor
    forward_links = []
    backward_links = []

# Aliases are used to expose outputs to external world
    # Format (internal_blob, external_blob, offset)
    # Negative offset stands for going from the end,
    # positive - from the beginning
    aliases = []

# States held inputs to the cell net
    recurrent_states = []

for cell_input, _ in initial_cell_inputs:
        cell_input = str(cell_input)
        # Recurrent_states is going to be (T + 1) x ...
        # It stores all inputs and outputs of the cell net over time.
        # Or their gradients in the case of the backward pass.
        state = s(cell_input + "_states")
        states_grad = state + "_grad"
        cell_output = links[str(cell_input)]
        forward_links.append((cell_input, state, 0))
        forward_links.append((cell_output, state, 1))

aliases.append((state, cell_output + "_all", 1))
        aliases.append((state, cell_output + "_last", -1))
        all_outputs.extend([cell_output + "_all", cell_output + "_last"])

recurrent_states.append(state)

if backward_cell_net is not None:
            backward_links.append((cell_output + "_grad", states_grad, 1))
            backward_cell_net.Proto().external_input.append(
                str(cell_output) + "_grad")

recurrent_input_grad = cell_input + "_grad"
            if not backward_blob_versions.get(recurrent_input_grad, 0):
                # If nobody writes to this recurrent input gradient, we need
                # to make sure it gets to the states grad blob after all.
                # We do this by using backward_links which triggers an alias
                # This logic is being used for example in a SumOp case
                backward_links.append(
                    (backward_mapping[cell_input], states_grad, 0))
            else:
                backward_links.append((recurrent_input_grad, states_grad, 0))

for input_t, input_blob in inputs:
        forward_links.append((str(input_t), str(input_blob), 0))

if backward_cell_net is not None:
        for input_t, input_blob in inputs:
            backward_links.append((
                backward_mapping[str(input_t)], str(input_blob) + "_grad", 0
            ))
        backward_cell_net.Proto().external_input.extend(
            cell_net.Proto().external_input)
        backward_cell_net.Proto().external_input.extend(
            cell_net.Proto().external_output)

def unpack_triple(x):
        if x:
            a, b, c = zip(*x)
            return a, b, c
        return [], [], []

# Splitting to separate lists so we can pass them to c++
    # where we ensemle them back
    link_internal, link_external, link_offset = unpack_triple(forward_links)
    alias_src, alias_dst, alias_offset = unpack_triple(aliases)

recurrent_inputs = [str(x[1]) for x in initial_cell_inputs]

# Make sure that recurrent gradients accumulate with internal gradients
    # (if a blob in the backward_cell_net receives gradient from both an
    # external connection as well as from within the backward_cell_net,
    # those gradients need to be added together, rather than one overwriting
    # the other)
    if backward_cell_net is not None:
        proto = backward_cell_net.Proto()
        operators = []
        while len(proto.op) > 0:
            op = proto.op[-1]
            proto.op.remove(op)
            operators.append(op)
        for op in operators[::-1]:
            proto.op.extend([op])
            for j, output_blob in enumerate(op.output):
                if output_blob in proto.external_input:
                    # In place operation won't cause issues because it takes
                    # existing value of a blob into account
                    if output_blob in op.input:
                        continue
                    output_blob = core.BlobReference(output_blob)
                    accum_blob = output_blob + "_accum"
                    proto.op[-1].output[j] = str(accum_blob)
                    backward_cell_net.Sum(
                        [output_blob, accum_blob],
                        [output_blob],
                    )

def map_to_dual_list(m):
        return [str(x) for x in list(m.keys())] + \
               [str(x) for x in list(m.values())]

backward_args = {}
    if backward_cell_net is not None:
        backward_mapping_keys = set(viewkeys(backward_mapping))
        backward_link_internal, backward_link_external, backward_link_offset = \
            unpack_triple(backward_links)
        params = [x for x in references if x in backward_mapping_keys]
        param_grads = [
            str(backward_mapping[x])
            for x in references
            if x in backward_mapping_keys
        ]
        if recompute_blobs_on_backward is None:
            recompute_blobs_on_backward = set()
        backward_args = {
            'param': [all_inputs.index(p) for p in params],
            'backward_link_internal': [str(l) for l in backward_link_internal],
            'backward_link_external': [str(l) for l in backward_link_external],
            'backward_link_offset': backward_link_offset,
            'outputs_with_grads': outputs_with_grads,
            'recompute_blobs_on_backward': [
                str(b) for b in recompute_blobs_on_backward
            ],
            'param_grads': param_grads,
        }
        if len(backward_cell_net.Proto().op) != 0:
            backward_args['backward_step_net'] = backward_cell_net.Proto()

results = net.RecurrentNetwork(
        all_inputs,
        all_outputs + [s("step_workspaces")],
        alias_src=alias_src,
        alias_dst=[str(a) for a in alias_dst],
        alias_offset=alias_offset,
        recurrent_states=recurrent_states,
        initial_recurrent_state_ids=[
            all_inputs.index(i) for i in recurrent_inputs
        ],
        link_internal=[str(l) for l in link_internal],
        link_external=[str(l) for l in link_external],
        link_offset=link_offset,
        enable_rnn_executor=1,
        step_net=cell_net.Proto(),
        timestep="timestep" if timestep is None else str(timestep),
        **backward_args
    )

# Restore net type since 'rnn' is not recognized outside RNNs
    cell_net.Proto().type = 'simple'

# The last output is a list of step workspaces,
    # which is only needed internally for gradient propogation
    return results[:-1]

def set_rnn_executor_config(rnn_op, num_threads=None, max_cuda_streams=None):
    from caffe2.proto import caffe2_pb2
    assert rnn_op.type in {'RecurrentNetwork', 'RecurrentNetworkGradient'}

def add_arg(s, v):
        a = caffe2_pb2.Argument()
        a.name = "rnn_executor." + s
        a.i = v
        rnn_op.arg.extend([a])

if num_threads is not None:
        add_arg('num_threads', num_threads)
    if max_cuda_streams is not None:
        add_arg('max_cuda_streams', max_cuda_streams)

def retrieve_step_blobs(net, prefix='rnn'):
    '''
    Retrieves blobs from step workspaces (which contain intermediate recurrent
    network computation for each timestep) and puts them in the global
    workspace. This allows access to the contents of this intermediate
    computation in python. Returns the list of extracted blob names.

net: the net from which the step workspace blobs should be extracted

prefix: prefix to append to extracted blob names when placing them in the
    global workspace
    '''
    count = 1
    output_list = []
    for op in net.Proto().op:
        if op.type == "RecurrentNetwork":
            blob_name = prefix + "_" + str(count)
            count = count + 1
            scratch_workspaces_blob_name = op.output[-1]
            workspace.RunOperatorOnce(
                core.CreateOperator(
                    "RecurrentNetworkBlobFetcher",
                    [scratch_workspaces_blob_name],
                    [blob_name],
                    prefix=prefix
                )
            )
            output_list += workspace.FetchBlob(blob_name).tolist()
    return output_list

cmd: