FM / operator

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

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

Name	Size	Mode	Actions
__pycache__/	-	0755	rm
activation_ops_test.py	9691	0644	edit dl rm
adadelta_test.py	7932	0644	edit dl rm
adagrad_test.py	7586	0644	edit dl rm
adagrad_test_helper.py	5181	0644	edit dl rm
adam_test.py	21559	0644	edit dl rm
affine_channel_op_test.py	3784	0644	edit dl rm
alias_with_name_test.py	938	0644	edit dl rm
apmeter_test.py	2738	0644	edit dl rm
arg_ops_test.py	1917	0644	edit dl rm
assert_test.py	797	0644	edit dl rm
async_net_barrier_test.py	946	0644	edit dl rm
atomic_ops_test.py	4104	0644	edit dl rm
basic_rnn_test.py	4720	0644	edit dl rm
batch_box_cox_test.py	5080	0644	edit dl rm
batch_bucketize_op_test.py	3730	0644	edit dl rm
batch_moments_op_test.py	2795	0644	edit dl rm
batch_sparse_to_dense_op_test.py	4199	0644	edit dl rm
bbox_transform_test.py	12258	0644	edit dl rm
bisect_percentile_op_test.py	6227	0644	edit dl rm
blobs_queue_db_test.py	3240	0644	edit dl rm
boolean_mask_test.py	16389	0644	edit dl rm
boolean_unmask_test.py	1711	0644	edit dl rm
box_with_nms_limit_op_test.py	8750	0644	edit dl rm
bucketize_op_test.py	930	0644	edit dl rm
cast_op_test.py	1600	0644	edit dl rm
ceil_op_test.py	888	0644	edit dl rm
channel_backprop_stats_op_test.py	2131	0644	edit dl rm
channel_shuffle_test.py	1794	0644	edit dl rm
channel_stats_op_test.py	2639	0644	edit dl rm
checkpoint_test.py	1500	0644	edit dl rm
clip_op_test.py	1984	0644	edit dl rm
clip_tensor_op_test.py	2076	0644	edit dl rm
collect_and_distribute_fpn_rpn_proposals_op_test.py	11269	0644	edit dl rm
concat_op_cost_test.py	2858	0644	edit dl rm
concat_split_op_test.py	7266	0644	edit dl rm
conditional_test.py	995	0644	edit dl rm
conftest.py	1446	0644	edit dl rm
conv_test.py	32473	0644	edit dl rm
conv_transpose_test.py	15945	0644	edit dl rm
copy_ops_test.py	7374	0644	edit dl rm
copy_rows_to_tensor_op_test.py	2526	0644	edit dl rm
cosine_embedding_criterion_op_test.py	1953	0644	edit dl rm
counter_ops_test.py	3348	0644	edit dl rm
crf_test.py	5315	0644	edit dl rm
cross_entropy_ops_test.py	10085	0644	edit dl rm
ctc_beam_search_decoder_op_test.py	5197	0644	edit dl rm
ctc_greedy_decoder_op_test.py	4743	0644	edit dl rm
cudnn_recurrent_test.py	5817	0644	edit dl rm
dataset_ops_test.py	23847	0644	edit dl rm
data_couple_op_test.py	858	0644	edit dl rm
decay_adagrad_test.py	2694	0644	edit dl rm
deform_conv_test.py	19276	0644	edit dl rm
dense_vector_to_id_list_op_test.py	2044	0644	edit dl rm
depthwise_3x3_conv_test.py	1863	0644	edit dl rm
detectron_keypoints.py	7973	0644	edit dl rm
distance_op_test.py	4351	0644	edit dl rm
dropout_op_test.py	2971	0644	edit dl rm
duplicate_operands_test.py	734	0644	edit dl rm
elementwise_linear_op_test.py	1382	0644	edit dl rm
elementwise_logical_ops_test.py	4617	0644	edit dl rm
elementwise_ops_test.py	33354	0644	edit dl rm
elementwise_op_broadcast_test.py	17466	0644	edit dl rm
emptysample_ops_test.py	1977	0644	edit dl rm
enforce_finite_op_test.py	1286	0644	edit dl rm
ensure_clipped_test.py	1505	0644	edit dl rm
ensure_cpu_output_op_test.py	1244	0644	edit dl rm
erf_op_test.py	749	0644	edit dl rm
expand_op_test.py	2109	0644	edit dl rm
fc_operator_test.py	3720	0644	edit dl rm
feature_maps_ops_test.py	21492	0644	edit dl rm
filler_ops_test.py	8476	0644	edit dl rm
find_op_test.py	1316	0644	edit dl rm
flatten_op_test.py	922	0644	edit dl rm
flexible_top_k_test.py	2609	0644	edit dl rm
floor_op_test.py	894	0644	edit dl rm
fused_nbit_rowwise_conversion_ops_test.py	14077	0644	edit dl rm
fused_nbit_rowwise_test_helper.py	2693	0644	edit dl rm
gather_ops_test.py	9216	0644	edit dl rm
gather_ranges_op_test.py	9125	0644	edit dl rm
given_tensor_byte_string_to_uint8_fill_op_test.py	1392	0644	edit dl rm
given_tensor_fill_op_test.py	1503	0644	edit dl rm
glu_op_test.py	1212	0644	edit dl rm
group_conv_test.py	2870	0644	edit dl rm
group_norm_op_test.py	5252	0644	edit dl rm
gru_test.py	12932	0644	edit dl rm
heatmap_max_keypoint_op_test.py	4770	0644	edit dl rm
histogram_test.py	3097	0644	edit dl rm
hsm_test.py	9456	0644	edit dl rm
hyperbolic_ops_test.py	1472	0644	edit dl rm
im2col_col2im_test.py	4311	0644	edit dl rm
image_input_op_test.py	17345	0644	edit dl rm
index_hash_ops_test.py	2885	0644	edit dl rm
index_ops_test.py	4597	0644	edit dl rm
instance_norm_test.py	9917	0644	edit dl rm
integral_image_ops_test.py	3419	0644	edit dl rm
jsd_ops_test.py	1044	0644	edit dl rm
key_split_ops_test.py	1289	0644	edit dl rm
lars_test.py	1354	0644	edit dl rm
layer_norm_op_test.py	14983	0644	edit dl rm
leaky_relu_test.py	5639	0644	edit dl rm
learning_rate_adaption_op_test.py	2837	0644	edit dl rm
learning_rate_op_test.py	8652	0644	edit dl rm
lengths_pad_op_test.py	1625	0644	edit dl rm
lengths_reducer_fused_nbit_rowwise_ops_test.py	15495	0644	edit dl rm
lengths_tile_op_test.py	1332	0644	edit dl rm
lengths_top_k_ops_test.py	2371	0644	edit dl rm
length_split_op_test.py	4868	0644	edit dl rm
listwise_l2r_operator_test.py	8740	0644	edit dl rm
load_save_test.py	33241	0644	edit dl rm
locally_connected_op_test.py	7761	0644	edit dl rm
loss_ops_test.py	902	0644	edit dl rm
lpnorm_op_test.py	2725	0644	edit dl rm
map_ops_test.py	2249	0644	edit dl rm
margin_ranking_criterion_op_test.py	1816	0644	edit dl rm
math_ops_test.py	1603	0644	edit dl rm
matmul_op_test.py	10096	0644	edit dl rm
mean_op_test.py	1469	0644	edit dl rm
merge_id_lists_op_test.py	2989	0644	edit dl rm
mkl_conv_op_test.py	1547	0644	edit dl rm
mkl_packed_fc_op_test.py	2647	0644	edit dl rm
mod_op_test.py	1459	0644	edit dl rm
moments_op_test.py	1722	0644	edit dl rm
momentum_sgd_test.py	6480	0644	edit dl rm
mpi_test.py	8154	0644	edit dl rm
mul_gradient_benchmark.py	1509	0644	edit dl rm
negate_gradient_op_test.py	1518	0644	edit dl rm
ngram_ops_test.py	2327	0644	edit dl rm
normalize_op_test.py	1679	0644	edit dl rm
numpy_tile_op_test.py	1924	0644	edit dl rm
one_hot_ops_test.py	7478	0644	edit dl rm
onnx_while_test.py	3070	0644	edit dl rm
order_switch_test.py	1306	0644	edit dl rm
pack_ops_test.py	12634	0644	edit dl rm
pack_rnn_sequence_op_test.py	2891	0644	edit dl rm
pad_test.py	1377	0644	edit dl rm
partition_ops_test.py	6838	0644	edit dl rm
percentile_op_test.py	4427	0644	edit dl rm
piecewise_linear_transform_test.py	6187	0644	edit dl rm
pooling_test.py	16508	0644	edit dl rm
prepend_dim_test.py	1505	0644	edit dl rm
python_op_test.py	1312	0644	edit dl rm
quantile_test.py	3276	0644	edit dl rm
rand_quantization_op_speed_test.py	3128	0644	edit dl rm
rank_loss_operator_test.py	5752	0644	edit dl rm
rebatching_queue_test.py	9047	0644	edit dl rm
record_queue_test.py	3125	0644	edit dl rm
recurrent_network_test.py	14048	0644	edit dl rm
recurrent_net_executor_test.py	10922	0644	edit dl rm
reduce_ops_test.py	17341	0644	edit dl rm
reduction_ops_test.py	4664	0644	edit dl rm
reshape_ops_test.py	8211	0644	edit dl rm
resize_op_test.py	9417	0644	edit dl rm
rmac_regions_op_test.py	3178	0644	edit dl rm
rms_norm_op_test.py	1325	0644	edit dl rm
rnn_cell_test.py	59707	0644	edit dl rm
roi_align_rotated_op_test.py	7567	0644	edit dl rm
rowwise_counter_test.py	2205	0644	edit dl rm
scale_op_test.py	2177	0644	edit dl rm
segment_ops_test.py	25745	0644	edit dl rm
self_binning_histogram_test.py	12915	0644	edit dl rm
selu_op_test.py	3232	0644	edit dl rm
sequence_ops_test.py	16000	0644	edit dl rm
shape_inference_test.py	25708	0644	edit dl rm
sinusoid_position_encoding_op_test.py	2308	0644	edit dl rm
softmax_ops_test.py	23685	0644	edit dl rm
softplus_op_test.py	516	0644	edit dl rm
sparse_dropout_with_replacement_op_test.py	2885	0644	edit dl rm
sparse_gradient_checker_test.py	1294	0644	edit dl rm
sparse_itemwise_dropout_with_replacement_op_test.py	2913	0644	edit dl rm
sparse_lengths_sum_benchmark.py	4159	0644	edit dl rm
sparse_lp_regularizer_test.py	2553	0644	edit dl rm
sparse_normalize_test.py	3136	0644	edit dl rm
sparse_ops_test.py	3469	0644	edit dl rm
sparse_to_dense_mask_op_test.py	3693	0644	edit dl rm
spatial_bn_op_test.py	20182	0644	edit dl rm
specialized_segment_ops_test.py	11775	0644	edit dl rm
split_op_cost_test.py	8645	0644	edit dl rm
square_root_divide_op_test.py	2179	0644	edit dl rm
stats_ops_test.py	1789	0644	edit dl rm
stats_put_ops_test.py	6596	0644	edit dl rm
storm_test.py	6507	0644	edit dl rm
string_ops_test.py	4154	0644	edit dl rm
text_file_reader_test.py	2517	0644	edit dl rm
thresholded_relu_op_test.py	2323	0644	edit dl rm
tile_op_test.py	3887	0644	edit dl rm
top_k_test.py	9113	0644	edit dl rm
torch_integration_test.py	39941	0644	edit dl rm
transpose_op_test.py	2722	0644	edit dl rm
trigonometric_op_test.py	1715	0644	edit dl rm
unique_ops_test.py	2255	0644	edit dl rm
unique_uniform_fill_op_test.py	1335	0644	edit dl rm
unsafe_coalesce_test.py	2940	0644	edit dl rm
upsample_op_test.py	7308	0644	edit dl rm
utility_ops_test.py	15054	0644	edit dl rm
video_input_op_test.py	10503	0644	edit dl rm
weighted_multi_sample_test.py	1997	0644	edit dl rm
weighted_sample_test.py	2739	0644	edit dl rm
weighted_sum_test.py	3052	0644	edit dl rm
weight_scale_test.py	2057	0644	edit dl rm
wngrad_test.py	8279	0644	edit dl rm
__init__.py	0	0644	edit dl rm

Edit: /usr/local/lib64/python3.6/site-packages/caffe2/python/operator_test/softmax_ops_test.py (23685B)

from caffe2.python import core, workspace
from hypothesis import given, settings
import caffe2.python.hypothesis_test_util as hu
import caffe2.python.serialized_test.serialized_test_util as serial
import hypothesis.strategies as st
import numpy as np

import unittest

class TestSoftmaxOps(serial.SerializedTestCase):

@serial.given(n=st.sampled_from([0, 2, 4, 71, 103]),
                  D=st.sampled_from([0, 4, 8, 64, 79, 256, 333]),
                  engine=st.sampled_from([None, 'CUDNN']),
                  **hu.gcs)
    def test_softmax(self, n, D, engine, gc, dc):
        # n = number of examples, D = |labels|
        # Initialize X and add 1e-2 for numerical stability
        X = np.random.rand(n, D).astype(np.float32)
        X = X + 1e-2

# Reference implementation of cross entropy with soft labels
        def label_softmax(X):
            probs = np.zeros((n, D))
            rowmax = np.zeros(n)

if D == 0:
                return [probs]

for i in range(n):
                rowmax[i] = max(X[i, ])
                # We need to subtract the max to avoid numerical issues
                probs[i] = X[i] - rowmax[i]
                exps = np.exp(probs[i, ])
                norm = sum(exps)
                probs[i, ] = exps / norm

return [probs]

op = core.CreateOperator(
            "Softmax",
            ["X"],
            ["probs"],
            engine=engine
        )

self.assertReferenceChecks(
            device_option=gc,
            op=op,
            inputs=[X],
            reference=label_softmax,
        )

@given(n=st.sampled_from([0, 2, 4, 71, 103, 555, 751, 1201]),
                  D=st.sampled_from([0, 4, 8, 64, 79, 256, 333, 1000]),
                  engine=st.sampled_from([None, 'CUDNN']),
                  **hu.gcs)
    @settings(deadline=10000)
    def test_softmax_grad(self, n, D, engine, gc, dc):
        # n = number of examples, D = |labels|
        # Initialize X and add 1e-2 for numerical stability
        Y = np.random.rand(n, D).astype(np.float32)
        dY = np.random.rand(n, D).astype(np.float32)
        Y = Y + 1e-2

# Reference implementation of cross entropy with soft labels
        def label_softmax_grad(X, dY):
            dX = Y * 0.0
            for i in range(n):
                d = np.dot(Y[i, :], dY[i, :])
                dX[i, :] = Y[i, :] * (dY[i, :] - d)
            return [dX]

op = core.CreateOperator(
            "SoftmaxGradient",
            ["Y", "dY"],
            ["dX"],
            engine=engine
        )

self.assertReferenceChecks(
            device_option=gc,
            op=op,
            inputs=[Y, dY],
            reference=label_softmax_grad,
        )

@given(axis=st.integers(min_value=1, max_value=4),
           engine=st.sampled_from([None, 'CUDNN']),
           **hu.gcs)
    def test_softmax_axis(self, axis, engine, gc, dc):
        np.random.seed(1)
        X = np.random.randn(1, 2, 3, 2, 1).astype(np.float32)
        X = X + 1e-2

def prod(xs):
            p = 1
            for x in xs:
                p *= x
            return p

N = prod(list(X.shape)[:axis])
        D = prod(list(X.shape)[axis:])

# Reference implementation of cross entropy with soft labels
        def label_softmax(X):
            X_ = X.reshape(N, D)
            probs = np.zeros((N, D))
            rowmax = np.zeros(N)
            for i in range(N):
                rowmax[i] = max(X_[i, ])
                # We need to subtract the max to avoid numerical issues
                probs[i] = X_[i] - rowmax[i]
                exps = np.exp(probs[i, ])
                norm = sum(exps)
                probs[i, ] = exps / norm

return [probs.reshape(*X.shape)]

op = core.CreateOperator(
            "Softmax",
            ["X"],
            ["probs"],
            axis=axis,
            engine=engine,
        )

self.assertReferenceChecks(
            device_option=gc,
            op=op,
            inputs=[X],
            reference=label_softmax,
        )

self.assertGradientChecks(
            gc, op, [X], 0, [0], stepsize=1e-4, threshold=1e-2)

@given(n=st.integers(2, 10), D=st.integers(4, 16),
           only_loss=st.booleans(), **hu.gcs)
    @settings(deadline=10000)
    def test_softmax_with_loss(self, n, D, gc, only_loss, dc):
        # n = number of examples, D = |labels|
        # Initialize X and add 1e-2 for numerical stability
        np.random.seed(2603)
        X = np.random.rand(n, D).astype(np.float32)
        X = X + 1e-2

# Initialize label
        label = (np.random.rand(n) * D).astype(np.int32)

# Reference implementation of cross entropy with soft labels
        def label_softmax_crossent(X, label):
            probs = np.zeros((n, D))
            rowmax = np.zeros(n)
            for i in range(n):
                rowmax[i] = max(X[i, ])
                # We need to subtract the max to avoid numerical issues
                probs[i] = X[i] - rowmax[i]
                exps = np.exp(probs[i, ])
                norm = sum(exps)
                probs[i, ] = exps / norm

label_xent = [-np.log(max(probs[i][label[i]], 1e-20))
                          for i in range(n)]
            avgloss = np.sum(label_xent) / float(n)
            return (probs, avgloss)

op = core.CreateOperator(
            "SoftmaxWithLoss",
            ["X", "label"],
            ["probs", "avgloss"],
            only_loss=only_loss,
        )

self.assertReferenceChecks(
            device_option=gc,
            op=op,
            inputs=[X, label],
            reference=label_softmax_crossent,
        )

self.assertGradientChecks(
            gc, op, [X, label], 0, [1], stepsize=1e-4, threshold=1e-2)

@given(
        n=st.integers(2, 5),
        D=st.integers(4, 16),
        only_loss=st.booleans(),
        label_prob=st.booleans(),
        **hu.gcs
    )
    @settings(deadline=10000)
    def test_softmax_with_loss_axis_2(
        self, n, D, only_loss, label_prob,
        gc, dc
    ):
        np.random.seed(2603)
        X = np.random.rand(n, n, D).astype(np.float32)
        X = X + 1e-2

if label_prob:
            label = np.random.rand(n, n, D).astype(np.float32)
            label /= label.sum(axis=2, keepdims=True)
        else:
            label = (np.random.rand(n, n) * D).astype(np.int32)

# Reference implementation of cross entropy with soft labels
        def label_softmax_crossent(X, label):
            probs = np.zeros((n, n, D))
            rowmax = np.zeros((n, n))
            for i in range(n):
                for j in range(n):
                    rowmax[i, j] = max(X[i, j, ])
                    # We need to subtract the max to avoid numerical issues
                    probs[i, j] = X[i, j] - rowmax[i, j]
                    exps = np.exp(probs[i, j, ])
                    norm = sum(exps)
                    probs[i, j, ] = exps / norm
            label_xent = 0
            for i in range(n):
                for j in range(n):
                    if label_prob:
                        for k in range(D):
                            label_xent += (
                                -np.log(max(probs[i, j, k], 1e-20)) *
                                label[i, j, k]
                            )
                    else:
                        label_xent += -np.log(max(probs[i, j, label[i, j]], 1e-20))

avgloss = label_xent / float(n * n)
            return (probs, avgloss)

op = core.CreateOperator(
            "SoftmaxWithLoss",
            ["X", "label"],
            ["probs", "avgloss"],
            only_loss=only_loss,
            label_prob=label_prob,
            axis=2,
        )

self.assertReferenceChecks(
            device_option=gc,
            op=op,
            inputs=[X, label],
            reference=label_softmax_crossent,
        )

self.assertGradientChecks(
            gc, op, [X, label], 0, [1], stepsize=1e-4, threshold=1e-2)

@unittest.skipIf(not workspace.has_gpu_support, "No gpu support")
    @given(**hu.gcs_gpu_only)
    def test_softmax_with_loss_large(self, gc, dc):
        np.random.seed(2603)
        for n in [32]:
            for D in [1000, 2000, 20000]:
                # n = number of examples, D = |labels|
                # Initialize X and add 1e-2 for numerical stability
                X = np.random.rand(n, D).astype(np.float32)
                X = X + 1e-2

# Initialize label
                label = (np.random.rand(n) * D).astype(np.int32)

# Reference implementation of cross entropy with soft labels
                def label_softmax_crossent(X, label):
                    probs = np.zeros((n, D))
                    rowmax = np.zeros(n)
                    for i in range(n):
                        rowmax[i] = max(X[i, ])
                        # We need to subtract the max to avoid numerical issues
                        probs[i] = X[i] - rowmax[i]
                        exps = np.exp(probs[i, ])
                        norm = sum(exps)
                        probs[i, ] = exps / norm

label_xent = [-np.log(max(probs[i][label[i]], 1e-20))
                                  for i in range(n)]
                    avgloss = np.sum(label_xent) / float(n)
                    return (probs, avgloss)

op = core.CreateOperator(
                    "SoftmaxWithLoss",
                    ["X", "label"],
                    ["probs", "avgloss"]
                )

self.assertReferenceChecks(
                    device_option=gc,
                    op=op,
                    inputs=[X, label],
                    reference=label_softmax_crossent,
                )

@given(n=st.integers(2, 10), D=st.integers(4, 16), **hu.gcs)
    @settings(deadline=None)
    def test_softmax_with_loss_label_prob(self, n, D, gc, dc):
        # n = number of examples, D = |labels|
        # Initialize X and add 1e-2 for numerical stability
        np.random.seed(2603)
        X = np.random.rand(n, D).astype(np.float32)
        X = X + 1e-2

# Initialize label
        label = np.random.rand(D, n).astype(np.float32)

# normalize labels to sum to 1
        label /= np.sum(label, axis=0)
        label = label.transpose()

label_xent = np.zeros(X.shape)
            for i in range(n):
                for j in range(D):
                    label_xent[i][j] = -np.log(
                        max(probs[i, j], 1e-20)) * label[i, j]
            avgloss = np.sum(label_xent) / float(n)
            return (probs, avgloss)

op = core.CreateOperator(
            "SoftmaxWithLoss",
            ["X", "label"],
            ["probs", "avgloss"],
            label_prob=1
        )

self.assertReferenceChecks(
            device_option=gc,
            op=op,
            inputs=[X, label],
            reference=label_softmax_crossent,
        )

self.assertGradientChecks(
            gc, op, [X, label], 0, [1], stepsize=1e-4, threshold=1e-2)

@given(
        n=st.integers(2, 10),
        D=st.integers(4, 16),
        only_loss=st.booleans(),
        **hu.gcs)
    @settings(deadline=None)
    def test_softmax_with_loss_weighted(self, n, D, only_loss, gc, dc):
        # n = number of examples, D = |labels|
        # Initialize X and add 1e-2 for numerical stability
        np.random.seed(2603)
        X = np.random.rand(n, D).astype(np.float32)
        X = X + 1e-2

# Initialize label
        label = (np.random.rand(n) * D).astype(np.int32)

# Init weights (weight by sample)
        weights = np.random.rand(n).astype(np.float32)

# Reference implementation of cross entropy with soft labels
        def label_softmax_crossent_weighted(X, label, weights):
            probs = np.zeros((n, D))
            rowmax = np.zeros(n)
            for i in range(n):
                rowmax[i] = max(X[i, ])
                # We need to subtract the max to avoid numerical issues
                probs[i] = X[i] - rowmax[i]
                exps = np.exp(probs[i, ])
                norm = sum(exps)
                probs[i, ] = exps / norm

label_xent = [-weights[i] * np.log(max(probs[i][label[i]], 1e-20))
                          for i in range(n)]
            avgloss = np.sum(label_xent) / sum(weights)
            return (probs, avgloss)

op = core.CreateOperator(
            "SoftmaxWithLoss",
            ["X", "label", "weights"],
            ["probs", "avgloss"],
            only_loss=only_loss,
        )

self.assertReferenceChecks(
            device_option=gc,
            op=op,
            inputs=[X, label, weights],
            reference=label_softmax_crossent_weighted,
        )

self.assertGradientChecks(
            gc, op, [X, label, weights], 0, [1], stepsize=1e-4, threshold=1e-2)

@given(n=st.integers(2, 10), D=st.integers(4, 16), **hu.gcs)
    @settings(deadline=None)
    def test_softmax_with_loss_label_prob_weighted(self, n, D, gc, dc):
        # n = number of examples, D = |labels|
        # Initialize X and add 1e-2 for numerical stability
        X = np.random.rand(n, D).astype(np.float32)
        X = X + 1e-2

# Initialize label
        label = np.random.rand(D, n).astype(np.float32)

# normalize labels to sum to 1
        label /= np.sum(label, axis=0)
        label = label.transpose()

# Init weights (weight by sample)
        weights = np.random.rand(n).astype(np.float32)

label_xent = np.zeros(X.shape)
            for i in range(n):
                for j in range(D):
                    label_xent[i][j] = -np.log(
                        max(probs[i, j], 1e-20)) * label[i, j] * weights[i]
            avgloss = np.sum(label_xent) / sum(weights)
            return (probs, avgloss)

op = core.CreateOperator(
            "SoftmaxWithLoss",
            ["X", "label", "weights"],
            ["probs", "avgloss"],
            label_prob=1,
        )

self.assertReferenceChecks(
            device_option=gc,
            op=op,
            inputs=[X, label, weights],
            reference=label_softmax_crossent_weighted,
        )

self.assertGradientChecks(
            gc, op, [X, label, weights], 0, [1], stepsize=1e-4, threshold=1e-2)

@given(n=st.integers(2, 5), D=st.integers(2, 4),
           weighted=st.booleans(), **hu.gcs)
    @settings(deadline=None, max_examples=50)
    def test_spatial_softmax_with_loss(self, n, D, weighted, gc, dc):
        # n = number of examples, D = |labels|
        # Initialize X and add 1e-2 for numerical stability
        W = 18
        H = 12
        np.random.seed(2603)
        X = np.random.rand(n, D, H, W).astype(np.float32)
        X = X + 1e-2

weighted = True
        weights = None
        if weighted:
            weights = np.random.rand(n, H, W).astype(np.float32)

# Initialize label. Some of the labels are (-1), i.e "DONT CARE"
        label = (np.random.rand(n, H, W) * (D + 1)).astype(np.int32) - 1

def label_softmax_crossent_spatial(X, label, weights=None):
            probs = np.zeros((n, D, H, W))
            rowmax = np.zeros((n, H, W))
            label_xent = np.zeros((n, H, W))
            for i in range(n):
                for x in range(W):
                    for y in range(H):
                        rowmax[i, y, x] = max(X[i, :, y, x])
                        # We need to subtract the max to avoid numerical issues
                        probs[i, :, y, x] = X[i, :, y, x] - rowmax[i, y, x]
                        exps = np.exp(probs[i, :, y, x])
                        probs[i, :, y, x] = exps / sum(exps)

label_xent[:, y, x] = \
                            [-np.log(max(probs[j, label[i, y, x], y, x], 1e-20))
                             for j in range(n)]

total_xent = 0.0
            total_weight = 0.0
            for y in range(H):
                for x in range(W):
                    for i in range(n):
                        l = label[i, y, x]
                        if (l != (-1)):
                            w = 1.0 if weights is None else weights[i, y, x]
                            total_xent += \
                                -np.log(max(probs[i, l, y, x], 1e-20)) * w
                            total_weight += w
            print("Total weight {}".format(total_weight))

return (probs, total_xent / total_weight)

op = core.CreateOperator(
            "SpatialSoftmaxWithLoss",
            ["X", "label"] + ([] if weights is None else ["weights"]),
            ["probs", "avgloss"],
        )

inputs = [X, label] + ([] if weights is None else [weights])
        self.assertReferenceChecks(
            device_option=gc,
            op=op,
            inputs=inputs,
            reference=label_softmax_crossent_spatial,
        )

self.assertGradientChecks(
            gc, op, inputs, 0, [1], stepsize=1e-4, threshold=1e-2)

@given(n=st.integers(4, 5), D=st.integers(3, 4),
           weighted=st.booleans(), **hu.gcs)
    def test_spatial_softmax_with_loss_allignore(self, n, D, weighted, gc, dc):
        # n = number of examples, D = |labels|
        # Initialize X and add 1e-2 for numerical stability
        W = 18
        H = 12
        np.random.seed(2603)
        X = np.random.rand(n, D, H, W).astype(np.float32)
        X = X + 1e-2

weighted = True
        weights = None
        if weighted:
            weights = np.random.rand(n, H, W).astype(np.float32)

# Initialize label. All labels as "DONT CARE"
        label = np.zeros((n, H, W)).astype(np.int32) - 1
        print(label)

label_xent[:, y, x] = \
                            [-np.log(max(probs[j, label[i, y, x], y, x], 1e-20))
                            for j in range(n)]

return (probs, 0.0)

op = core.CreateOperator(
            "SpatialSoftmaxWithLoss",
            ["X", "label"] + ([] if weights is None else ["weights"]),
            ["probs", "avgloss"],
        )

@given(n=st.integers(4, 5), D=st.integers(3, 4),
           weighted=st.booleans(), **hu.gcs)
    def test_softmax_with_loss_zero_weight(self, n, D, weighted, gc, dc):
        # n = number of examples, D = |labels|
        # Initialize X and add 1e-2 for numerical stability
        np.random.seed(2603)
        X = np.random.rand(n, D).astype(np.float32)
        X = X + 1e-2

weights = np.zeros(n).astype(np.float32)

# Initialize label
        label = (np.random.rand(n) * D).astype(np.int32)

def label_softmax_crossent(X, label, weights=None):
            probs = np.zeros((n, D))
            rowmax = np.zeros((n))
            for i in range(n):
                rowmax[i] = max(X[i, ])
                # We need to subtract the max to avoid numerical issues
                probs[i] = X[i] - rowmax[i]
                exps = np.exp(probs[i, ])
                norm = sum(exps)
                probs[i, ] = exps / norm
            return (probs, 0.0)

op = core.CreateOperator(
            "SoftmaxWithLoss",
            ["X", "label", "weights"],
            ["probs", "avgloss"]
        )

inputs = [X, label] + ([] if weights is None else [weights])
        self.assertReferenceChecks(
            device_option=gc,
            op=op,
            inputs=inputs,
            reference=label_softmax_crossent,
        )

@unittest.skipIf(not workspace.has_gpu_support, "No gpu support")
    def test_compare_cpugpu(self):
        '''
        Additional test that checks CPU and GPU returns same values
        with larger examples. This is mainly to test the more complex
        GPU implementation is correct.
        '''
        from caffe2.proto import caffe2_pb2

for _j in range(3):
            gpuop = core.CreateOperator(
                "SpatialSoftmaxWithLoss",
                ["X_gpu", "label_gpu"],
                ["probs_gpu", "avgloss_gpu"],
                device_option=core.DeviceOption(workspace.GpuDeviceType, 0)
            )

cpuop = core.CreateOperator(
                "SpatialSoftmaxWithLoss",
                ["X_cpu", "label_cpu"],
                ["probs_cpu", "avgloss_cpu"],
                device_option=core.DeviceOption(caffe2_pb2.CPU)
            )

n = 8
            D = 4
            W = 64 + int(np.random.rand(1) * 1024)
            H = 64 + int(np.random.rand(1) * 1024)

print("W: {} H: {}".format(W, H))

X = np.random.rand(n, D, H, W).astype(np.float32)
            X = X + 1e-2

# Initialize label. Some of the labels are (-1), i.e "DONT CARE"
            label = (np.random.rand(n, H, W) * (D + 1)).astype(np.int32) - 1

gpu0 = core.DeviceOption(workspace.GpuDeviceType, 0)
            workspace.FeedBlob("X_cpu", X)
            workspace.FeedBlob("label_cpu", label)
            workspace.FeedBlob("X_gpu", X, device_option=gpu0)
            workspace.FeedBlob("label_gpu", label, device_option=gpu0)

workspace.RunOperatorOnce(gpuop)
            workspace.RunOperatorOnce(cpuop)

probs_gpu = workspace.FetchBlob("probs_gpu")
            probs_cpu = workspace.FetchBlob("probs_cpu")
            loss_gpu = workspace.FetchBlob("avgloss_gpu")
            loss_cpu = workspace.FetchBlob("avgloss_cpu")

np.testing.assert_allclose(probs_gpu, probs_cpu, rtol=1e-4)
            np.testing.assert_allclose(loss_gpu, loss_cpu, rtol=1e-1)

if __name__ == "__main__":
    import unittest
    import random
    random.seed(2603)
    unittest.main()

cmd: