model.py

#!/usr/bin/env python
# -*- coding: utf-8 -*-
# File: model.py

import tensorflow as tf
from tensorpack.tfutils import get_current_tower_context
from tensorpack.tfutils.summary import add_moving_summary
from tensorpack.tfutils.argscope import argscope
from tensorpack.tfutils.scope_utils import under_name_scope
from tensorpack.models import (
    Conv2D, FullyConnected, GlobalAvgPooling, layer_register, Deconv2D)

from utils.box_ops import pairwise_iou
import config


@under_name_scope()
def clip_boxes(boxes, window, name=None):
    """
    Args:
        boxes: nx4, xyxy
        window: [h, w]
    """
    boxes = tf.maximum(boxes, 0.0)
    m = tf.tile(tf.reverse(window, [0]), [2])    # (4,)
    boxes = tf.minimum(boxes, tf.to_float(m), name=name)
    return boxes


@layer_register(log_shape=True)
def rpn_head(featuremap, channel, num_anchors):
    """
    Returns:
        label_logits: fHxfWxNA
        box_logits: fHxfWxNAx4
    """
    with argscope(Conv2D, data_format='NCHW',
                  W_init=tf.random_normal_initializer(stddev=0.01)):
        hidden = Conv2D('conv0', featuremap, channel, 3, nl=tf.nn.relu)

        label_logits = Conv2D('class', hidden, num_anchors, 1)
        box_logits = Conv2D('box', hidden, 4 * num_anchors, 1)
        # 1, NA(*4), im/16, im/16 (NCHW)

        label_logits = tf.transpose(label_logits, [0, 2, 3, 1])  # 1xfHxfWxNA
        label_logits = tf.squeeze(label_logits, 0)  # fHxfWxNA

        shp = tf.shape(box_logits)  # 1x(NAx4)xfHxfW
        box_logits = tf.transpose(box_logits, [0, 2, 3, 1])  # 1xfHxfWx(NAx4)
        box_logits = tf.reshape(box_logits, tf.stack([shp[2], shp[3], num_anchors, 4]))  # fHxfWxNAx4
    return label_logits, box_logits


@under_name_scope()
def rpn_losses(anchor_labels, anchor_boxes, label_logits, box_logits):
    """
    Args:
        anchor_labels: fHxfWxNA
        anchor_boxes: fHxfWxNAx4, encoded
        label_logits:  fHxfWxNA
        box_logits: fHxfWxNAx4

    Returns:
        label_loss, box_loss
    """
    with tf.device('/cpu:0'):
        valid_mask = tf.stop_gradient(tf.not_equal(anchor_labels, -1))
        pos_mask = tf.stop_gradient(tf.equal(anchor_labels, 1))
        nr_valid = tf.stop_gradient(tf.count_nonzero(valid_mask, dtype=tf.int32), name='num_valid_anchor')
        nr_pos = tf.count_nonzero(pos_mask, dtype=tf.int32, name='num_pos_anchor')

        valid_anchor_labels = tf.boolean_mask(anchor_labels, valid_mask)
    valid_label_logits = tf.boolean_mask(label_logits, valid_mask)

    with tf.name_scope('label_metrics'):
        valid_label_prob = tf.nn.sigmoid(valid_label_logits)
        summaries = []
        with tf.device('/cpu:0'):
            for th in [0.5, 0.2, 0.1]:
                valid_prediction = tf.cast(valid_label_prob > th, tf.int32)
                nr_pos_prediction = tf.reduce_sum(valid_prediction, name='num_pos_prediction')
                pos_prediction_corr = tf.count_nonzero(
                    tf.logical_and(
                        valid_label_prob > th,
                        tf.equal(valid_prediction, valid_anchor_labels)),
                    dtype=tf.int32)
                summaries.append(tf.truediv(
                    pos_prediction_corr,
                    nr_pos, name='recall_th{}'.format(th)))
                precision = tf.to_float(tf.truediv(pos_prediction_corr, nr_pos_prediction))
                precision = tf.where(tf.equal(nr_pos_prediction, 0), 0.0, precision, name='precision_th{}'.format(th))
                summaries.append(precision)
        add_moving_summary(*summaries)

    label_loss = tf.nn.sigmoid_cross_entropy_with_logits(
        labels=tf.to_float(valid_anchor_labels), logits=valid_label_logits)
    label_loss = tf.reduce_mean(label_loss, name='label_loss')

    pos_anchor_boxes = tf.boolean_mask(anchor_boxes, pos_mask)
    pos_box_logits = tf.boolean_mask(box_logits, pos_mask)
    delta = 1.0 / 9
    box_loss = tf.losses.huber_loss(
        pos_anchor_boxes, pos_box_logits, delta=delta,
        reduction=tf.losses.Reduction.SUM) / delta
    box_loss = tf.div(
        box_loss,
        tf.cast(nr_valid, tf.float32), name='box_loss')

    add_moving_summary(label_loss, box_loss, nr_valid, nr_pos)
    return label_loss, box_loss


@under_name_scope()
def decode_bbox_target(box_predictions, anchors):
    """
    Args:
        box_predictions: (..., 4), logits
        anchors: (..., 4), floatbox. Must have the same shape

    Returns:
        box_decoded: (..., 4), float32. With the same shape.
    """
    orig_shape = tf.shape(anchors)
    box_pred_txtytwth = tf.reshape(box_predictions, (-1, 2, 2))
    box_pred_txty, box_pred_twth = tf.split(box_pred_txtytwth, 2, axis=1)
    # each is (...)x1x2
    anchors_x1y1x2y2 = tf.reshape(anchors, (-1, 2, 2))
    anchors_x1y1, anchors_x2y2 = tf.split(anchors_x1y1x2y2, 2, axis=1)

    waha = anchors_x2y2 - anchors_x1y1
    xaya = (anchors_x2y2 + anchors_x1y1) * 0.5

    wbhb = tf.exp(tf.minimum(
        box_pred_twth, config.BBOX_DECODE_CLIP)) * waha
    xbyb = box_pred_txty * waha + xaya
    x1y1 = xbyb - wbhb * 0.5
    x2y2 = xbyb + wbhb * 0.5    # (...)x1x2
    out = tf.concat([x1y1, x2y2], axis=-2)
    return tf.reshape(out, orig_shape)


@under_name_scope()
def encode_bbox_target(boxes, anchors):
    """
    Args:
        boxes: (..., 4), float32
        anchors: (..., 4), float32

    Returns:
        box_encoded: (..., 4), float32 with the same shape.
    """
    anchors_x1y1x2y2 = tf.reshape(anchors, (-1, 2, 2))
    anchors_x1y1, anchors_x2y2 = tf.split(anchors_x1y1x2y2, 2, axis=1)
    waha = anchors_x2y2 - anchors_x1y1
    xaya = (anchors_x2y2 + anchors_x1y1) * 0.5

    boxes_x1y1x2y2 = tf.reshape(boxes, (-1, 2, 2))
    boxes_x1y1, boxes_x2y2 = tf.split(boxes_x1y1x2y2, 2, axis=1)
    wbhb = boxes_x2y2 - boxes_x1y1
    xbyb = (boxes_x2y2 + boxes_x1y1) * 0.5

    # Note that here not all boxes are valid. Some may be zero
    txty = (xbyb - xaya) / waha
    twth = tf.log(wbhb / waha)  # may contain -inf for invalid boxes
    encoded = tf.concat([txty, twth], axis=1)  # (-1x2x2)
    return tf.reshape(encoded, tf.shape(boxes))


@under_name_scope()
def generate_rpn_proposals(boxes, scores, img_shape):
    """
    Args:
        boxes: nx4 float dtype, decoded to floatbox already
        scores: n float, the logits
        img_shape: [h, w]

    Returns:
        boxes: kx4 float
        scores: k logits
    """
    assert boxes.shape.ndims == 2, boxes.shape
    if get_current_tower_context().is_training:
        PRE_NMS_TOPK = config.TRAIN_PRE_NMS_TOPK
        POST_NMS_TOPK = config.TRAIN_POST_NMS_TOPK
    else:
        PRE_NMS_TOPK = config.TEST_PRE_NMS_TOPK
        POST_NMS_TOPK = config.TEST_POST_NMS_TOPK

    topk = tf.minimum(PRE_NMS_TOPK, tf.size(scores))
    topk_scores, topk_indices = tf.nn.top_k(scores, k=topk, sorted=False)
    topk_boxes = tf.gather(boxes, topk_indices)
    topk_boxes = clip_boxes(topk_boxes, img_shape)

    topk_boxes_x1y1x2y2 = tf.reshape(topk_boxes, (-1, 2, 2))
    topk_boxes_x1y1, topk_boxes_x2y2 = tf.split(topk_boxes_x1y1x2y2, 2, axis=1)
    # nx1x2 each
    wbhb = tf.squeeze(topk_boxes_x2y2 - topk_boxes_x1y1, axis=1)
    valid = tf.reduce_all(wbhb > config.RPN_MIN_SIZE, axis=1)  # n,
    topk_valid_boxes_x1y1x2y2 = tf.boolean_mask(topk_boxes_x1y1x2y2, valid)
    topk_valid_scores = tf.boolean_mask(topk_scores, valid)

    topk_valid_boxes_y1x1y2x2 = tf.reshape(
        tf.reverse(topk_valid_boxes_x1y1x2y2, axis=[2]),
        (-1, 4), name='nms_input_boxes')
    nms_indices = tf.image.non_max_suppression(
        topk_valid_boxes_y1x1y2x2,
        topk_valid_scores,
        max_output_size=POST_NMS_TOPK,
        iou_threshold=config.RPN_PROPOSAL_NMS_THRESH)

    topk_valid_boxes = tf.reshape(topk_valid_boxes_x1y1x2y2, (-1, 4))
    final_boxes = tf.gather(
        topk_valid_boxes,
        nms_indices, name='boxes')
    final_scores = tf.gather(topk_valid_scores, nms_indices, name='scores')
    tf.sigmoid(final_scores, name='probs')  # for visualization
    return final_boxes, final_scores


@under_name_scope()
def proposal_metrics(iou):
    """
    Add summaries for RPN proposals.

    Args:
        iou: nxm, #proposal x #gt
    """
    # find best roi for each gt, for summary only
    best_iou = tf.reduce_max(iou, axis=0)
    mean_best_iou = tf.reduce_mean(best_iou, name='best_iou_per_gt')
    summaries = [mean_best_iou]
    with tf.device('/cpu:0'):
        for th in [0.3, 0.5]:
            recall = tf.truediv(
                tf.count_nonzero(best_iou >= th),
                tf.size(best_iou, out_type=tf.int64),
                name='recall_iou{}'.format(th))
            summaries.append(recall)
    add_moving_summary(*summaries)


@under_name_scope()
def sample_fast_rcnn_targets(boxes, gt_boxes, gt_labels):
    """
    Sample some ROIs from all proposals for training.

    Args:
        boxes: nx4 region proposals, floatbox
        gt_boxes: mx4, floatbox
        gt_labels: m, int32

    Returns:
        sampled_boxes: tx4 floatbox, the rois
        sampled_labels: t labels, in [0, #class-1]. Positive means foreground.
        fg_inds_wrt_gt: #fg indices, each in range [0, m-1].
            It contains the matching GT of each foreground roi.
    """
    iou = pairwise_iou(boxes, gt_boxes)     # nxm
    proposal_metrics(iou)

    # add ground truth as proposals as well
    boxes = tf.concat([boxes, gt_boxes], axis=0)    # (n+m) x 4
    iou = tf.concat([iou, tf.eye(tf.shape(gt_boxes)[0])], axis=0)   # (n+m) x m
    # #proposal=n+m from now on

    def sample_fg_bg(iou):
        fg_mask = tf.reduce_max(iou, axis=1) >= config.FASTRCNN_FG_THRESH

        fg_inds = tf.reshape(tf.where(fg_mask), [-1])
        num_fg = tf.minimum(int(
            config.FASTRCNN_BATCH_PER_IM * config.FASTRCNN_FG_RATIO),
            tf.size(fg_inds), name='num_fg')
        fg_inds = tf.random_shuffle(fg_inds)[:num_fg]

        bg_inds = tf.reshape(tf.where(tf.logical_not(fg_mask)), [-1])
        num_bg = tf.minimum(
            config.FASTRCNN_BATCH_PER_IM - num_fg,
            tf.size(bg_inds), name='num_bg')
        bg_inds = tf.random_shuffle(bg_inds)[:num_bg]

        add_moving_summary(num_fg, num_bg)
        return fg_inds, bg_inds

    fg_inds, bg_inds = sample_fg_bg(iou)
    # fg,bg indices w.r.t proposals

    best_iou_ind = tf.argmax(iou, axis=1)   # #proposal, each in 0~m-1
    fg_inds_wrt_gt = tf.gather(best_iou_ind, fg_inds)   # num_fg

    all_indices = tf.concat([fg_inds, bg_inds], axis=0)   # indices w.r.t all n+m proposal boxes
    ret_boxes = tf.gather(boxes, all_indices, name='sampled_proposal_boxes')

    ret_labels = tf.concat(
        [tf.gather(gt_labels, fg_inds_wrt_gt),
         tf.zeros_like(bg_inds, dtype=tf.int64)], axis=0, name='sampled_labels')
    # stop the gradient -- they are meant to be ground-truth
    return tf.stop_gradient(ret_boxes), tf.stop_gradient(ret_labels), fg_inds_wrt_gt


@under_name_scope()
def crop_and_resize(image, boxes, box_ind, crop_size):
    """
    Better-aligned version of tf.image.crop_and_resize, following our definition of floating point boxes.

    Args:
        image: NCHW
        boxes: nx4, x1y1x2y2
        box_ind: (n,)
        crop_size (int):
    Returns:
        n,C,size,size
    """
    assert isinstance(crop_size, int), crop_size

    @under_name_scope()
    def transform_fpcoor_for_tf(boxes, image_shape, crop_shape):
        """
        The way tf.image.crop_and_resize works (with normalized box):
        Initial point (the value of output[0]): x0_box * (W_img - 1)
        Spacing: w_box * (W_img - 1) / (W_crop - 1)
        Use the above grid to bilinear sample.

        However, what we want is (with fpcoor box):
        Spacing: w_box / W_crop
        Initial point: x0_box + spacing/2 - 0.5
        (-0.5 because bilinear sample assumes floating point coordinate (0.0, 0.0) is the same as pixel value (0, 0))

        This function transform fpcoor boxes to a format to be used by tf.image.crop_and_resize

        Returns:
            y1x1y2x2
        """
        x0, y0, x1, y1 = tf.split(boxes, 4, axis=1)

        spacing_w = (x1 - x0) / tf.to_float(crop_shape[1])
        spacing_h = (y1 - y0) / tf.to_float(crop_shape[0])

        nx0 = (x0 + spacing_w / 2 - 0.5) / tf.to_float(image_shape[1] - 1)
        ny0 = (y0 + spacing_h / 2 - 0.5) / tf.to_float(image_shape[0] - 1)

        nw = spacing_w * tf.to_float(crop_shape[1] - 1) / tf.to_float(image_shape[1] - 1)
        nh = spacing_h * tf.to_float(crop_shape[0] - 1) / tf.to_float(image_shape[0] - 1)

        return tf.concat([ny0, nx0, ny0 + nh, nx0 + nw], axis=1)

    image_shape = tf.shape(image)[2:]
    boxes = transform_fpcoor_for_tf(boxes, image_shape, [crop_size, crop_size])
    image = tf.transpose(image, [0, 2, 3, 1])   # 1hwc
    ret = tf.image.crop_and_resize(
        image, boxes, box_ind,
        crop_size=[crop_size, crop_size])
    ret = tf.transpose(ret, [0, 3, 1, 2])   # ncss
    return ret


@under_name_scope()
def roi_align(featuremap, boxes, output_shape):
    """
    Args:
        featuremap: 1xCxHxW
        boxes: Nx4 floatbox
        output_shape: int

    Returns:
        NxCxoHxoW
    """
    boxes = tf.stop_gradient(boxes)  # TODO
    # sample 4 locations per roi bin
    ret = crop_and_resize(
        featuremap, boxes,
        tf.zeros([tf.shape(boxes)[0]], dtype=tf.int32),
        output_shape * 2)
    ret = tf.nn.avg_pool(ret, [1, 1, 2, 2], [1, 1, 2, 2], padding='SAME', data_format='NCHW')
    return ret


@layer_register(log_shape=True)
def fastrcnn_head(feature, num_classes):
    """
    Args:
        feature (NxCx7x7):
        num_classes(int): num_category + 1

    Returns:
        cls_logits (Nxnum_class), reg_logits (Nx num_class-1 x 4)
    """
    feature = GlobalAvgPooling('gap', feature, data_format='NCHW')
    classification = FullyConnected(
        'class', feature, num_classes,
        W_init=tf.random_normal_initializer(stddev=0.01))
    box_regression = FullyConnected(
        'box', feature, (num_classes - 1) * 4,
        W_init=tf.random_normal_initializer(stddev=0.001))
    box_regression = tf.reshape(box_regression, (-1, num_classes - 1, 4))
    return classification, box_regression


@under_name_scope()
def fastrcnn_losses(labels, label_logits, fg_boxes, fg_box_logits):
    """
    Args:
        labels: n,
        label_logits: nxC
        fg_boxes: nfgx4, encoded
        fg_box_logits: nfgx(C-1)x4
    """
    label_loss = tf.nn.sparse_softmax_cross_entropy_with_logits(
        labels=labels, logits=label_logits)
    label_loss = tf.reduce_mean(label_loss, name='label_loss')

    fg_inds = tf.where(labels > 0)[:, 0]
    fg_labels = tf.gather(labels, fg_inds)
    num_fg = tf.size(fg_inds)
    indices = tf.stack(
        [tf.range(num_fg),
         tf.to_int32(fg_labels) - 1], axis=1)  # #fgx2
    fg_box_logits = tf.gather_nd(fg_box_logits, indices)

    with tf.name_scope('label_metrics'), tf.device('/cpu:0'):
        prediction = tf.argmax(label_logits, axis=1, name='label_prediction')
        correct = tf.to_float(tf.equal(prediction, labels))  # boolean/integer gather is unavailable on GPU
        accuracy = tf.reduce_mean(correct, name='accuracy')
        fg_label_pred = tf.argmax(tf.gather(label_logits, fg_inds), axis=1)
        num_zero = tf.reduce_sum(tf.to_int32(tf.equal(fg_label_pred, 0)), name='num_zero')
        false_negative = tf.truediv(num_zero, num_fg, name='false_negative')
        fg_accuracy = tf.reduce_mean(
            tf.gather(correct, fg_inds), name='fg_accuracy')

    box_loss = tf.losses.huber_loss(
        fg_boxes, fg_box_logits, reduction=tf.losses.Reduction.SUM)
    box_loss = tf.truediv(
        box_loss, tf.to_float(tf.shape(labels)[0]), name='box_loss')

    add_moving_summary(label_loss, box_loss, accuracy, fg_accuracy, false_negative)
    return label_loss, box_loss


@under_name_scope()
def fastrcnn_predictions(boxes, probs):
    """
    Generate final results from predictions of all proposals.

    Args:
        boxes: n#catx4 floatbox in float32
        probs: nx#class
    """
    assert boxes.shape[1] == config.NUM_CLASS - 1
    assert probs.shape[1] == config.NUM_CLASS
    boxes = tf.transpose(boxes, [1, 0, 2])  # #catxnx4
    probs = tf.transpose(probs[:, 1:], [1, 0])  # #catxn

    def f(X):
        """
        prob: n probabilities
        box: nx4 boxes

        Returns: n boolean, the selection
        """
        prob, box = X
        output_shape = tf.shape(prob)
        # filter by score threshold
        ids = tf.reshape(tf.where(prob > config.RESULT_SCORE_THRESH), [-1])
        prob = tf.gather(prob, ids)
        box = tf.gather(box, ids)
        # NMS within each class
        selection = tf.image.non_max_suppression(
            box, prob, config.RESULTS_PER_IM, config.FASTRCNN_NMS_THRESH)
        selection = tf.to_int32(tf.gather(ids, selection))
        # sort available in TF>1.4.0
        # sorted_selection = tf.contrib.framework.sort(selection, direction='ASCENDING')
        sorted_selection = -tf.nn.top_k(-selection, k=tf.size(selection))[0]
        mask = tf.sparse_to_dense(
            sparse_indices=sorted_selection,
            output_shape=output_shape,
            sparse_values=True,
            default_value=False)
        return mask

    masks = tf.map_fn(f, (probs, boxes), dtype=tf.bool,
                      parallel_iterations=10)     # #cat x N
    selected_indices = tf.where(masks)  # #selection x 2, each is (cat_id, box_id)
    probs = tf.boolean_mask(probs, masks)

    # filter again by sorting scores
    topk_probs, topk_indices = tf.nn.top_k(
        probs,
        tf.minimum(config.RESULTS_PER_IM, tf.size(probs)),
        sorted=False)
    filtered_selection = tf.gather(selected_indices, topk_indices)
    filtered_selection = tf.reverse(filtered_selection, axis=[1], name='filtered_indices')
    return filtered_selection, topk_probs


@layer_register(log_shape=True)
def maskrcnn_head(feature, num_class):
    """
    Args:
        feature (NxCx7x7):
        num_classes(int): num_category + 1

    Returns:
        mask_logits (N x num_category x 14 x 14):
    """
    with argscope([Conv2D, Deconv2D], data_format='NCHW',
                  W_init=tf.variance_scaling_initializer(
                      scale=2.0, mode='fan_in', distribution='normal')):
        l = Deconv2D('deconv', feature, 256, 2, stride=2, nl=tf.nn.relu)
        l = Conv2D('conv', l, num_class - 1, 1)
    return l


@under_name_scope()
def maskrcnn_loss(mask_logits, fg_labels, fg_target_masks):
    """
    Args:
        mask_logits: #fg x #category x14x14
        fg_labels: #fg, in 1~#class
        fg_target_masks: #fgx14x14, int
    """
    num_fg = tf.size(fg_labels)
    indices = tf.stack([tf.range(num_fg), tf.to_int32(fg_labels) - 1], axis=1)  # #fgx2
    mask_logits = tf.gather_nd(mask_logits, indices)  # #fgx14x14
    mask_probs = tf.sigmoid(mask_logits)

    # add some training visualizations to tensorboard
    with tf.name_scope('mask_viz'):
        viz = tf.concat([fg_target_masks, mask_probs], axis=1)
        viz = tf.expand_dims(viz, 3)
        viz = tf.cast(viz * 255, tf.uint8, name='viz')
        tf.summary.image('mask_truth|pred', viz, max_outputs=10)

    loss = tf.nn.sigmoid_cross_entropy_with_logits(
        labels=fg_target_masks, logits=mask_logits)
    loss = tf.reduce_mean(loss, name='maskrcnn_loss')

    pred_label = mask_probs > 0.5
    truth_label = fg_target_masks > 0.5
    accuracy = tf.reduce_mean(
        tf.to_float(tf.equal(pred_label, truth_label)),
        name='accuracy')
    pos_accuracy = tf.logical_and(
        tf.equal(pred_label, truth_label),
        tf.equal(truth_label, True))
    pos_accuracy = tf.reduce_mean(tf.to_float(pos_accuracy), name='pos_accuracy')
    fg_pixel_ratio = tf.reduce_mean(tf.to_float(truth_label), name='fg_pixel_ratio')

    add_moving_summary(loss, accuracy, fg_pixel_ratio, pos_accuracy)
    return loss